Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.

org/terms-privacy

Advanced

Methods and

Computational
Applications in

Fluid Dynamics
Reduced Order
 Computational Science & Engineering
The SIAM series on Computational Science and Engineering publishes research monographs, advanced
undergraduate- or graduate-level textbooks, and other volumes of interest to an interdisciplinary CS&E community of
computational mathematicians, computer scientists, scientists, and engineers. The series includes introductory volumes
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

aimed at a broad audience of mathematically motivated readers interested in understanding methods and applications
within computational science and engineering, monographs reporting on the most recent developments in the field,
and volumes addressed to specific groups of professionals whose work relies extensively on computational science and
engineering.
SIAM created the CS&E series to support access to the rapid and far-ranging advances in computer modeling and
simulation of complex problems in science and engineering, to promote the interdisciplinary culture required to meet
these large-scale challenges, and to provide the means to the next generation of computational scientists and engineers.

Editor-in-Chief
Donald Estep, Simon Fraser University

Editorial Board
Ben Adcock Serkan Gugercin David Keyes
Simon Fraser University Virginia Tech Columbia University
Daniela Calvetti Jan S. Hesthaven Ralph C. Smith
Case Western Reserve University Ecole Polytechnique Fédérale North Carolina State University
de Lausanne
Omar Ghattas Karen Willcox
University of Texas at Austin Johan Hoffman University of Texas at Austin
KTH Royal Institute of Technology
Chen Greif
University of British Columbia

Series Volumes
Rozza, Gianluigi, Stabile, Giovanni, and Ballarin, Rostamian, Rouben, Programming Projects in C for
Francesco, Advanced Reduced Order Methods and Students of Engineering, Science, and Mathematics
Applications in Computational Fluid Dynamics Smith, Ralph C., Uncertainty Quantification: Theory,
Gatto, Paolo, Mathematical Foundations of Finite Implementation, and Applications
Elements and Iterative Solvers Dankowicz, Harry and Schilder, Frank, Recipes for
Adcock, Ben, Brugiapaglia, Simone, and Continuation
Webster, Clayton G., Sparse Polynomial Approximation Mueller, Jennifer L. and Siltanen, Samuli, Linear and
of High-Dimensional Functions Nonlinear Inverse Problems with Practical Applications
Hoffman, Johan, Methods in Computational Science Shapira, Yair, Solving PDEs in C++: Numerical Methods
Da Veiga, Sébastien, Gamboa, Fabrice, Iooss, Bertrand, in a Unified Object-Oriented Approach, Second Edition
and Prieur, Clémentine, Basics and Trends in Sensitivity Borzì, Alfio and Schulz, Volker, Computational
Analysis: Theory and Practice in R Optimization of Systems Governed by Partial
Vidyasagar, M., An Introduction to Compressed Sensing Differential Equations
Antoulas, A. C., Beattie, C. A., and Güğercin, S., Ascher, Uri M. and Greif, Chen, A First Course in
Interpolatory Methods for Model Reduction Numerical Methods
Sipahi, Rifat, Mastering Frequency Domain Techniques Layton, William, Introduction to the Numerical Analysis of
for the Stability Analysis of LTI Time Delay Systems Incompressible Viscous Flows
Bardsley, Johnathan M., Computational Uncertainty Ascher, Uri M., Numerical Methods for Evolutionary
Quantification for Inverse Problems Differential Equations
Hesthaven, Jan S., Numerical Methods for Conservation Zohdi, T. I., An Introduction to Modeling and Simulation
Laws: From Analysis to Algorithms of Particulate Flows
Sidi, Avram, Vector Extrapolation Methods with Biegler, Lorenz T., Ghattas, Omar, Heinkenschloss,
Applications Matthias, Keyes, David, and van Bloemen Waanders,
Borzì, A., Ciaramella, G., and Sprengel, M., Formulation Bart, editors, Real-Time PDE-Constrained Optimization
and Numerical Solution of Quantum Control Problems Chen, Zhangxin, Huan, Guanren, and Ma, Yuanle,
Benner, Peter, Cohen, Albert, Ohlberger, Mario, Computational Methods for Multiphase Flows in Porous
and Willcox, Karen, editors, Model Reduction and Media
Approximation: Theory and Algorithms Shapira, Yair, Solving PDEs in C++: Numerical Methods
Kuzmin, Dmitri and Hämäläinen, Jari, Finite Element in a Unified Object-Oriented Approach
Methods for Computational Fluid Dynamics: A Practical
Guide
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Advanced
Reduced Order
Methods and
Applications in
Computational
Fluid Dynamics
GIANLUIGI ROZZA
SISSA, International School for Advanced Studies
Trieste, Italy

GIOVANNI STABILE
SISSA, International School for Advanced Studies
Trieste, Italy

FRANCESCO BALLARIN
Catholic University of the Sacred Heart
Brescia, Italy

Society for Industrial and Applied Mathematics

Philadelphia
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Copyright © 2023 by the Society for Industrial and Applied Mathematics

10 9 8 7 6 5 4 3 2 1

All rights reserved. Printed in the United States of America. No part of this book may be reproduced, stored, or transmitted in any manner
without the written permission of the publisher. For information, write to the Society for Industrial and Applied Mathematics, 3600 Market
Street, 6th Floor, Philadelphia, PA 19104-2688 USA.

No warranties, express or implied, are made by the publisher, authors, and their employers that the programs contained in this volume
are free of error. They should not be relied on as the sole basis to solve a problem whose incorrect solution could result in injury to person
or property. If the programs are employed in such a manner, it is at the user’s own risk and the publisher, authors, and their employers
disclaim all liability for such misuse.

Trademarked names may be used in this book without the inclusion of a trademark symbol. These names are used in an editorial context
only; no infringement of trademark is intended.

Publications Director Kivmars H. Bowling

Executive Editor Elizabeth Greenspan
Managing Editor Kelly Thomas
Production Editor David Riegelhaupt
Copy Editor Julia Cochrane
Production Manager Donna Witzleben
Production Coordinator Cally A. Shrader
Compositor Cheryl Hufnagle
Graphic Designer Doug Smock

Library of Congress Cataloging-in-Publication Data

Names: Rozza, Gianluigi, author. | Stabile, Giovanni, author. | Ballarin,
Francesco, author.
Title: Advanced reduced order methods and applications in computational
fluid dynamics / Gianluigi Rozza, Giovanni Stabile, Francesco Ballarin.
Description: Philadelphia, PA : Society for Industrial and Applied
Mathematics, 2022. | Series: Computational science and engineering ; 27
| Includes bibliographical references and index. | Summary: “This is the
first book dedicated to reduced order methods in computational fluid
dynamics. The book focuses on complex parametrization of shapes for
their optimization and applications. Advanced topics, such as
turbulence, stability of flows, inverse problems, optimization, and flow
control are included”-- Provided by publisher.
Identifiers: LCCN 2022029194 (print) | LCCN 2022029195 (ebook) | ISBN
9781611977240 (paperback) | ISBN 9781611977257 (ebook)
Subjects: LCSH: Computational fluid dynamics. | Fluid
dynamics--Mathematics. | Estimation theory. | AMS: Partial differential
equations. | Fluid mechanics. | Partial differential equations --
Equations of mathematical physics and other areas of application -- PDEs
in connection with fluid mechanics.
Classification: LCC TA357.5.D37 R69 2022 (print) | LCC TA357.5.D37
(ebook) | DDC 620.1/064--dc23/eng/20220919
LC record available at https://lccn.loc.gov/2022029194
LC ebook record available at https://lccn.loc.gov/2022029195

is a registered trademark.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy


To our families, to our younger collaborators,
to little Petra and Clara representing the future,
and to all people who contributed to and inspired this book
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Contents

List of Contributors xvii

List of Figures xix

List of Tables xxxiii

List of Algorithms xxxv

Preface xxxvii

1 Overview and Motivation 1

Martin W. Hess, Marco Tezzele, Gianluigi Rozza
1.1 Reduced Order Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Intrusive Approaches . . . . . . . . . . . . . . . . . . . . . . 1
1.1.2 Nonintrusive Approaches . . . . . . . . . . . . . . . . . . . . 2
1.2 Basic Principles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.1 Offline-Online Decomposition . . . . . . . . . . . . . . . . . . 2
1.2.2 Affine Parameter Dependency . . . . . . . . . . . . . . . . . . 3
1.2.3 POD and the Greedy Algorithm . . . . . . . . . . . . . . . . . 3
1.2.4 Stabilization . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 More Advanced Principles . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3.1 Parameter Space Reduction . . . . . . . . . . . . . . . . . . . 9
1.3.2 Advanced Geometrical Morphing . . . . . . . . . . . . . . . . 10
1.3.3 New Deep Learning Approaches . . . . . . . . . . . . . . . . 11
1.3.4 Digital Twins . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

I Finite Element–Based ROMs 13

2 Finite Element–Based Reduced Basis Method in Computational Fluid Dynamics 15

Federico Pichi, Maria Strazzullo, Francesco Ballarin, Gianluigi Rozza
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 The FEM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2.1 The Model Problem: The Poisson Equation . . . . . . . . . . . 16
2.2.2 Galerkin Projection . . . . . . . . . . . . . . . . . . . . . . . 17
2.2.3 A First Example in Fluid Dynamics: Stokes Equations . . . . . 21
2.2.4 A Nonlinear Model: Navier–Stokes Equations . . . . . . . . . 24

vii
 viii Contents

2.3 ROMs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.3.1 RB Approximation . . . . . . . . . . . . . . . . . . . . . . . . 26
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

2.3.2 Model Problem: Parametrized Elliptic PDE . . . . . . . . . . . 27

2.3.3 Space Generation: POD . . . . . . . . . . . . . . . . . . . . . 29
2.3.4 Space Generation: Greedy Algorithm . . . . . . . . . . . . . . 30
2.3.5 Comparison between FE and RB Spaces . . . . . . . . . . . . 32
2.3.6 Model Problem: Geometrical Parametrization in Heat Transfer 35
2.3.7 Reduction of CFD Problems . . . . . . . . . . . . . . . . . . . 37
2.3.8 Supremizer Enrichment of the Velocity Space . . . . . . . . . 38
2.3.9 Extension to the Parabolic Case . . . . . . . . . . . . . . . . . 41
2.4 A Posteriori Error Estimation for Certified RB Methods . . . . . . . . . . . 44
2.4.1 Error Estimator for an Elliptic Problem . . . . . . . . . . . . . 44
2.4.2 The min-θ Approach . . . . . . . . . . . . . . . . . . . . . . . 48
2.4.3 The Successive Constraint Method . . . . . . . . . . . . . . . 48
2.4.4 Error Estimator for Stokes Equations . . . . . . . . . . . . . . 51
2.4.5 Error Estimator for Parabolic Problems . . . . . . . . . . . . . 52
2.5 Nonaffine and Nonlinear Problems . . . . . . . . . . . . . . . . . . . . . . 53
2.5.1 The EIM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
2.5.2 Model Problem: Elliptic Problem with Gaussian Source . . . . 56

3 Certified Smagorinsky Reduced Basis Turbulence Model 59

Enrique Delgado Ávila, Francesco Ballarin, Gianluigi Rozza
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.2 RB Smagorinsky Turbulence Model . . . . . . . . . . . . . . . . . . . . . . 60
3.2.1 FE Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.2.2 Greedy Algorithm . . . . . . . . . . . . . . . . . . . . . . . . 62
3.2.3 Approximation of the Eddy Viscosity Term . . . . . . . . . . . 63
3.2.4 Well-Posedness Analysis . . . . . . . . . . . . . . . . . . . . 64
3.2.5 A Posteriori Error Estimator . . . . . . . . . . . . . . . . . . . 65
3.2.6 Numerical Results for the Smagorinsky Model . . . . . . . . . 66
3.3 RB VMS-Boussinesq Model . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.3.1 FE Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.3.2 RB Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.3.3 Well-Posedness Analysis . . . . . . . . . . . . . . . . . . . . 72
3.3.4 A Posteriori Error Bound Estimator . . . . . . . . . . . . . . . 74
3.3.5 Numerical Results for the VMS-Boussinesq Model . . . . . . . 74
3.4 Geometrical Parametrization . . . . . . . . . . . . . . . . . . . . . . . . . 77
3.4.1 Problem Setting . . . . . . . . . . . . . . . . . . . . . . . . . 77
3.4.2 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . 79
3.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

4 Finite Element–Based Reduced Basis Method for Optimal Flow Control 83

Maria Strazzullo, Francesco Ballarin, Gianluigi Rozza
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.2 Linear OCP(µ)s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.2.1 ROMs for Linear OCP(µ)s . . . . . . . . . . . . . . . . . . . 86
4.2.2 Numerical Test: Pollutant Control in the Gulf of Trieste . . . . 87
4.3 ROMs for Nonlinear OCP(µ)s . . . . . . . . . . . . . . . . . . . . . . . . . 88
4.3.1 Numerical Test: Weather Prediction through Quasi-geostrophic
Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
 Contents ix

4.4 ROMs for Space-Time OCP(µ)s . . . . . . . . . . . . . . . . . . . . . . . . 91

4.4.1 Numerical Test: Graetz Flow Boundary Control . . . . . . . . 93
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

4.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

5 Reduced Basis Approaches to Bifurcating Nonlinear Parametrized Partial

Differential Equations 97
Federico Pichi, Francesco Ballarin, Gianluigi Rozza
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
5.2 Nonlinear Analysis of PDEs and Bifurcation Problems . . . . . . . . . . . . 98
5.2.1 Mathematical Formulation . . . . . . . . . . . . . . . . . . . . 98
5.2.2 An Overview of Bifurcation Theory . . . . . . . . . . . . . . . 100
5.3 Numerical Approximation of Bifurcating Phenomena . . . . . . . . . . . . 102
5.3.1 High-Fidelity Approximation . . . . . . . . . . . . . . . . . . 103
5.3.2 Reduced Basis Approximation . . . . . . . . . . . . . . . . . 105
5.4 Von Kármán Equations for Structural Buckling of Plates . . . . . . . . . . . 109
5.4.1 Von Kármán Model . . . . . . . . . . . . . . . . . . . . . . . 109
5.4.2 Weak Formulation and Its Approximation . . . . . . . . . . . . 110
5.4.3 Numerical Investigation of Buckling Phenomena . . . . . . . . 110
5.5 Gross–Pitaevskii Equations in Bose–Einstein Condensates . . . . . . . . . . 112
5.5.1 Gross–Pitaevskii Model . . . . . . . . . . . . . . . . . . . . . 112
5.5.2 Weak Formulation and Its Approximation . . . . . . . . . . . . 112
5.5.3 Numerical Investigation of Quantum States . . . . . . . . . . . 113
5.6 Hyperelastic Models for Bending Beams . . . . . . . . . . . . . . . . . . . 115
5.6.1 A Continuum Mechanics Framework . . . . . . . . . . . . . . 116
5.6.2 Weak Formulation and Its Approximation . . . . . . . . . . . . 117
5.6.3 Numerical Investigation of Compressed Beams . . . . . . . . . 117
5.7 Navier–Stokes Flow and the Coandă Effect . . . . . . . . . . . . . . . . . . 119
5.7.1 NS Model in a Channel . . . . . . . . . . . . . . . . . . . . . 120
5.7.2 Weak Formulation and Its Approximation . . . . . . . . . . . . 120
5.7.3 Numerical Investigation of Wall-Hugging Behavior . . . . . . 121

6 Reduced Basis Stabilization for Convection-Dominated Problems 125

Enrique Delgado Ávila, Francesco Ballarin, Gianluigi Rozza
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
6.2 Advection-Diffusion Problem . . . . . . . . . . . . . . . . . . . . . . . . . 126
6.3 Steady Stokes Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
6.3.1 Stabilized FE Problem . . . . . . . . . . . . . . . . . . . . . . 128
6.3.2 Stabilized RB Problem . . . . . . . . . . . . . . . . . . . . . . 129
6.4 Steady Navier–Stokes Equations . . . . . . . . . . . . . . . . . . . . . . . 129
6.4.1 Stabilized FE Problem . . . . . . . . . . . . . . . . . . . . . . 130
6.4.2 Stabilized RB Problem . . . . . . . . . . . . . . . . . . . . . . 130
6.5 Stabilized VMS-Smagorinsky Turbulence Model . . . . . . . . . . . . . . . 131
6.5.1 FE Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
6.5.2 RB Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
6.5.3 Well-Posedness Analysis . . . . . . . . . . . . . . . . . . . . 135
6.5.4 A Posteriori Error Bound Estimator . . . . . . . . . . . . . . . 137
6.5.5 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . 138
6.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
 x Contents

II Finite Volume, Spectral Element, and Discontinuous Galerkin-Based Reduced

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Order Models 141

7 Finite Volume–Based Reduced Order Models for Laminar Flows 143

Matteo Zancanaro, Saddam Hijazi, Umberto Morelli, Giovanni Stabile, Gianluigi Rozza
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
7.1.1 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . 143
7.2 Computational Fluid Dynamics—Laminar NSEs . . . . . . . . . . . . . . . 144
7.2.1 Full-Order Model . . . . . . . . . . . . . . . . . . . . . . . . 144
7.2.2 Segregated Pressure-Based Solvers for the Incompressible NSEs146
7.2.3 Nonsegregated Finite Volume–Based ROMs . . . . . . . . . . 150
7.2.4 Segregated Finite Volume–Based ROMs (SIMPLE-ROM) . . . 153
7.2.5 Treatment of Nonhomogeneous Dirichlet Boundary Conditions 154
7.2.6 Offline-Online Computation of Lift and Drag Forces . . . . . . 156
7.3 Numerical Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
7.3.1 Lid-Driven Cavity Problem . . . . . . . . . . . . . . . . . . . 157
7.3.2 Flow around a Circular Cylinder . . . . . . . . . . . . . . . . 160
7.3.3 Comments on the Results . . . . . . . . . . . . . . . . . . . . 164

8 Finite Volume–Based Reduced Order Models for Turbulent Flows 165

Matteo Zancanaro, Saddam Hijazi, Michele Girfoglio, Andrea Mola, Giovanni Stabile,
Gianluigi Rozza
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
8.2 RANS Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
8.2.1 Closure Problem and Reynolds Averaging . . . . . . . . . . . 165
8.2.2 EVMs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
8.3 LES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
8.4 Hybrid Projection-Based/Data-Driven ROM for Turbulent Flows . . . . . . 171
8.4.1 Hybrid ROM with RBF Interpolation on the Time-Parameter
Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
8.4.2 Hybrid ROM with RBF Interpolation on the Velocity Projec-
tion Coefficient Values . . . . . . . . . . . . . . . . . . . . . . 175
8.5 Turbulent ROMs Based on the Uniform-ROM and the PPE-ROM . . . . . . 178
8.5.1 Hybrid ROM Based on the PPE-ROM . . . . . . . . . . . . . 180
8.6 Application of the H-SUP-ROM to Turbulent Problems . . . . . . . . . . . 181
8.6.1 Steady Case . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
8.6.2 Unsteady Case . . . . . . . . . . . . . . . . . . . . . . . . . . 187

9 Nonintrusive Data-Driven Reduced Order Models in Computational Fluid

Dynamics 203
Marco Tezzele, Nicola Demo, Giovanni Stabile, Gianluigi Rozza
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
9.2 A General Framework for Nonintrusive Parametric ROMs . . . . . . . . . . 204
9.2.1 Database Creation . . . . . . . . . . . . . . . . . . . . . . . . 204
9.2.2 Linear Dimensionality Reduction . . . . . . . . . . . . . . . . 204
9.2.3 Solution Manifold Approximation . . . . . . . . . . . . . . . . 206
9.2.4 Computational Considerations . . . . . . . . . . . . . . . . . . 207
9.3 DMD for Time-Dependent Problems . . . . . . . . . . . . . . . . . . . . . 207
9.3.1 Classical DMD Algorithm . . . . . . . . . . . . . . . . . . . . 207
9.3.2 DMD Extensions . . . . . . . . . . . . . . . . . . . . . . . . . 209
 Contents xi

9.4 Applications in CFD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212

9.4.1 A ROM Pipeline for Shape Optimization Problems . . . . . . . 212
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

9.4.2 Comparison between DMD and PODI for Hydroacoustics Prob-

lems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
9.4.3 A Parametric View of DMD . . . . . . . . . . . . . . . . . . . 217
9.5 Conclusions and Future Perspectives . . . . . . . . . . . . . . . . . . . . . 222

10 Spectral Element Method–Based Model Order Reduction 223

Martin W. Hess, Gianluigi Rozza
10.1 Basic Notions and Functions of the SEM . . . . . . . . . . . . . . . . . . . 223
10.1.1 Polynomials and Quadrature Rules . . . . . . . . . . . . . . . 223
10.1.2 Expansion Functions . . . . . . . . . . . . . . . . . . . . . . . 225
10.2 Assembly of System Matrices . . . . . . . . . . . . . . . . . . . . . . . . . 227
10.2.1 Method of Weighted Residuals . . . . . . . . . . . . . . . . . 227
10.2.2 Elemental Properties . . . . . . . . . . . . . . . . . . . . . . . 227
10.3 Reduced Order Modeling with the SEM . . . . . . . . . . . . . . . . . . . 228
10.3.1 Examples in Computational Fluid Dynamics . . . . . . . . . . 229
10.3.2 Reduced Order Model . . . . . . . . . . . . . . . . . . . . . . 230

11 Discontinuous Galerkin-Based Reduced Order Models 233

Andrea Lario, Francesco Romor, Gianluigi Rozza
11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
11.2 Nodal DGM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
11.2.1 Mapping from Physical to Master Elements . . . . . . . . . . . 236
11.2.2 Interface Fluxes . . . . . . . . . . . . . . . . . . . . . . . . . 238
11.3 Reduced Order Methods for DGM . . . . . . . . . . . . . . . . . . . . . . 239
11.3.1 Offline Phase . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
11.3.2 Online Phase . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
11.4 Projection of the Governing Equations . . . . . . . . . . . . . . . . . . . . 241
11.4.1 Incompressible Navier–Stokes . . . . . . . . . . . . . . . . . . 241
11.4.2 Compressible Navier–Stokes . . . . . . . . . . . . . . . . . . 242
11.5 Test Case and Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244
11.5.1 Incompressible Flow around a Square-Based Cylinder . . . . . 244
11.5.2 Parametric Simulation of a Compressible Cavity Flow in a Duct 244

III Advances in Reduced Order Models for Computational Fluid Dynamics 249

12 Weighted Reduced Order Methods for Uncertainty Quantification 251

Davide Torlo, Maria Strazzullo, Francesco Ballarin, Gianluigi Rozza
12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
12.2 Stochastic PDEs and Discretized Approximations . . . . . . . . . . . . . . 252
12.3 Weighted ROMs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
12.3.1 WRB Method . . . . . . . . . . . . . . . . . . . . . . . . . . 253
12.3.2 wPOD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
12.4 Sampling Strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
12.5 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
12.5.1 Averaging Uncertain Parameter Simulations for Heat Equation
and Stokes Problems . . . . . . . . . . . . . . . . . . . . . . . 256
 xii Contents

12.5.2 Stabilization of Advection-Dominated Problems Conditioned

to Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

12.5.3 wPOD for Optimal Control for Environmental Sciences . . . . 261

12.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263

13 Reduced Basis, Embedded Methods, and Parametrized Level-Set Geometry 265

Efthymios N. Karatzas, Giovanni Stabile, Francesco Ballarin, Gianluigi Rozza
13.1 Introduction and Overview . . . . . . . . . . . . . . . . . . . . . . . . . . 265
13.1.1 Heat Exchange Model Problem . . . . . . . . . . . . . . . . . 265
13.1.2 Shifted Nitsche Boundary Weak Formulation . . . . . . . . . . 266
13.1.3 The Parametrized Thermal-Heat Exchange Model . . . . . . . 267
13.1.4 Model Reduction Methodology . . . . . . . . . . . . . . . . . 267
13.1.5 POD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268
13.1.6 The Projection Stage and ROM Generation . . . . . . . . . . . 268
13.1.7 Numerical Experiments (Heat Exchange/SBM) . . . . . . . . . 269
13.2 Parametrized Steady Stokes Equations . . . . . . . . . . . . . . . . . . . . 271
13.2.1 POD Adapted to Flows . . . . . . . . . . . . . . . . . . . . . 272
13.2.2 Steady Stokes Numerical Experiments (SBM) . . . . . . . . . 272
13.3 Looking for a Better ROM with CutFEM . . . . . . . . . . . . . . . . . . . 276
13.4 ROM and a Fourth-Order Evolutionary Nonlinear System . . . . . . . . . . 280
13.5 Conclusions and Future Developments . . . . . . . . . . . . . . . . . . . . 282

14 Reduced Order Methods for Fluid-Structure Interaction Problems 283

Monica Nonino, Francesco Ballarin, Gianluigi Rozza
14.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
14.2 Dynamics of FSI Problems . . . . . . . . . . . . . . . . . . . . . . . . . . 284
14.2.1 The ALE Formulation . . . . . . . . . . . . . . . . . . . . . . 284
14.3 Approaches to FSI Problems . . . . . . . . . . . . . . . . . . . . . . . . . 286
14.3.1 Partitioned Algorithms . . . . . . . . . . . . . . . . . . . . . . 286
14.3.2 Monolithic Algorithms . . . . . . . . . . . . . . . . . . . . . 287
14.4 Partitioned RBM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
14.4.1 Test Case: Leaflets Bending under the Influence of a Fluid . . . 287
14.4.2 Offline Computational Phase . . . . . . . . . . . . . . . . . . 290
14.4.3 Online Phase . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
14.4.4 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . 295
14.5 A Monolithic RBM for a Transport-Dominated FSI Problem . . . . . . . . 296
14.5.1 The Kolmogorov n-Width . . . . . . . . . . . . . . . . . . . . 297
14.5.2 Nonlinear Model Reduction by Transport Maps . . . . . . . . 298
14.5.3 A Multiphysics Problem . . . . . . . . . . . . . . . . . . . . . 299
14.5.4 Offline Phase . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
14.5.5 Online Phase . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
14.5.6 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . 308
14.5.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 310

15 Reduced Order Models for Bifurcating Phenomena in Fluid-Structure

Interaction Problems 311
Moaad Khamlich, Federico Pichi, Gianluigi Rozza
15.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
15.2 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312
15.2.1 Variational Formulation . . . . . . . . . . . . . . . . . . . . . 313
 Contents xiii

15.2.2 Lagrange Multipliers Method . . . . . . . . . . . . . . . . . . 315

15.2.3 Branchwise Reduced Basis Approximation . . . . . . . . . . . 315
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

15.3 Application to the Coandă Effect . . . . . . . . . . . . . . . . . . . . . . . 317

15.3.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
15.3.2 FSI Problem with Linear Elasticity . . . . . . . . . . . . . . . 318
15.3.3 RBM Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 320
15.4 Model Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322

IV Further Advances, Perspectives, and Applications 325

16 Reduction in Parameter Space 327

Marco Tezzele, Francesco Romor, Gianluigi Rozza
16.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327
16.2 AS-Based Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328
16.2.1 Active Subspaces . . . . . . . . . . . . . . . . . . . . . . . . 328
16.2.2 Kernel-Based Active Subspaces . . . . . . . . . . . . . . . . . 329
16.2.3 Local Active Subspaces . . . . . . . . . . . . . . . . . . . . . 333
16.2.4 Multifidelity Regression . . . . . . . . . . . . . . . . . . . . . 334
16.2.5 Gradient-Free Extensions . . . . . . . . . . . . . . . . . . . . 336
16.3 Other Nonlinear Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . 337
16.3.1 Nonlinear Level-Set Learning . . . . . . . . . . . . . . . . . . 337
16.3.2 Active Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . 337
16.4 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 338
16.4.1 Response Surface Design . . . . . . . . . . . . . . . . . . . . 338
16.4.2 Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . 338
16.4.3 Sensitivity Analysis . . . . . . . . . . . . . . . . . . . . . . . 340
16.4.4 Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . 340
16.4.5 Inverse Problems . . . . . . . . . . . . . . . . . . . . . . . . . 341
16.4.6 Coupling with Other Reduced Order Methods . . . . . . . . . 342
16.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 342

17 Geometrical Parametrization and Morphing Techniques with Applications 345

Andrea Mola, Nicola Demo, Marco Tezzele, Gianluigi Rozza
17.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345
17.2 Parametrization Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . 347
17.2.1 General Purpose Shape Parametrization Algorithms . . . . . . 347
17.2.2 Object-Specific Shape Parametrization Algorithms: Some
Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
17.3 Application to Different Geometry Specifications . . . . . . . . . . . . . . 356
17.3.1 Geometric Model Parametrization . . . . . . . . . . . . . . . . 356
17.3.2 Computational Domain Parametrization . . . . . . . . . . . . . 361
17.3.3 The Use of RBF to Extend Boundary Deformation to Internal
Grid Nodes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 361
17.4 Conclusions and Future Perspectives . . . . . . . . . . . . . . . . . . . . . 363

18 Reduced Order Methods for Hemodynamics Applications 365

Zakia Zainib, Pierfrancesco Siena, Michele Girfoglio, Martin W. Hess,
Francesco Ballarin, Gianluigi Rozza
18.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
 xiv Contents

18.2 CABGs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366

18.2.1 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . 366
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

18.2.2 ROM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367

18.2.3 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . 367
18.3 Coandă Effect in Mitral Valves . . . . . . . . . . . . . . . . . . . . . . . . 370
18.3.1 Physical Principles . . . . . . . . . . . . . . . . . . . . . . . . 370
18.3.2 Mathematical Modeling . . . . . . . . . . . . . . . . . . . . . 372
18.3.3 ROM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373
18.4 LVAD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 374
18.4.1 High-Fidelity Problem . . . . . . . . . . . . . . . . . . . . . . 374
18.4.2 ROM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375
18.4.3 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . 375

19 Scientific Software Development and Packages for Reduced Order Models in

Computational Fluid Dynamics 379
Nicola Demo, Marco Tezzele, Giovanni Stabile, Gianluigi Rozza
19.1 Scientific Open-Source Software for Reduced Order Models . . . . . . . . . 379
19.1.1 RBniCS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379
19.1.2 ITHACA-FV . . . . . . . . . . . . . . . . . . . . . . . . . . . 381
19.1.3 ITHACA-SEM . . . . . . . . . . . . . . . . . . . . . . . . . . 381
19.1.4 ITHACA-DG . . . . . . . . . . . . . . . . . . . . . . . . . . 382
19.1.5 EZyRB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 382
19.1.6 PyDMD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383
19.1.7 ATHENA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384
19.1.8 PyGeM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384
19.1.9 BladeX . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385
19.2 Best Practices for Scientific Programming . . . . . . . . . . . . . . . . . . 385
19.2.1 Code Quality . . . . . . . . . . . . . . . . . . . . . . . . . . . 386
19.2.2 Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 387
19.2.3 Documentation . . . . . . . . . . . . . . . . . . . . . . . . . . 387

20 A Deep Learning Approach to Improving Reduced Order Models 389

Laura Meneghetti, Nirav Shah, Michele Girfoglio, Nicola Demo, Marco Tezzele,
Andrea Lario, Giovanni Stabile, Gianluigi Rozza
20.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 389
20.2 ANNs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 390
20.3 NN Topologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393
20.3.1 Feedforward Neural Network . . . . . . . . . . . . . . . . . . 393
20.3.2 Recurrent Neural Network . . . . . . . . . . . . . . . . . . . . 393
20.3.3 Convolutional Neural Network . . . . . . . . . . . . . . . . . 394
20.3.4 Autoencoder . . . . . . . . . . . . . . . . . . . . . . . . . . . 396
20.4 NNs to Improve POD-Based Models . . . . . . . . . . . . . . . . . . . . . 397
20.4.1 The POD-ANN Approach . . . . . . . . . . . . . . . . . . . . 397
20.4.2 The NNsPOD Approach . . . . . . . . . . . . . . . . . . . . . 399
20.5 A PINN for Solving Parametric PDEs . . . . . . . . . . . . . . . . . . . . . 401
20.6 A Reduced Approach for CNNs . . . . . . . . . . . . . . . . . . . . . . . . 402
20.6.1 Reduction Strategies for CNNs . . . . . . . . . . . . . . . . . 403
20.6.2 Reduced ANNs . . . . . . . . . . . . . . . . . . . . . . . . . 403
20.6.3 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . 407
 Contents xv

20.7 A Future Perspective on ROMs for Digital Twin . . . . . . . . . . . . . . . 410

20.7.1 Introduction to Digital Twins . . . . . . . . . . . . . . . . . . 410
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

20.7.2 The Role of Modeling and Simulation in Digital Twin . . . . . 410

20.7.3 ROMs for Digital Twin . . . . . . . . . . . . . . . . . . . . . 411
20.7.4 Next Developments and Challenges . . . . . . . . . . . . . . . 413

Bibliography 415

Index 461
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

List of Contributors

Francesco Ballarin
Department of Mathematics and Physics, Catholic University of the Sacred Heart,
Brescia, Italy

Enrique Delgado Ávila

Department of Differential Equations and Numerical Analysis, University of Seville,
Spain

Nicola Demo
mathLab, Mathematics Area, SISSA, Scuola Internazionale Superiore di Studi
Avanzati, Trieste, Italy

Michele Girfoglio
mathLab, Mathematics Area, SISSA, Scuola Internazionale Superiore di Studi
Avanzati, Trieste, Italy

Martin W. Hess
mathLab, Mathematics Area, SISSA, Scuola Internazionale Superiore di Studi
Avanzati, Trieste, Italy

Saddam Hijazi
Institute of Mathematics, University of Potsdam, Germany

Efthymios N. Karatzas
Department of Mathematics, Aristotle University of Thessaloniki, Greece

Moaad Khamlich
mathLab, Mathematics Area, SISSA, Scuola Internazionale Superiore di Studi
Avanzati, Trieste, Italy

Andrea Lario
mathLab, Mathematics Area, SISSA, Scuola Internazionale Superiore di Studi
Avanzati, Trieste, Italy

Laura Meneghetti
mathLab, Mathematics Area, SISSA, Scuola Internazionale Superiore di Studi
Avanzati, Trieste, Italy

xvii
 xviii List of Contributors

Andrea Mola
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

mathLab, Mathematics Area, SISSA, Scuola Internazionale Superiore di Studi

Avanzati, Trieste, Italy

Umberto Morelli
Departamento de Matemática Aplicada, Universidade de Santiago de Compostela,
Spain

Monica Nonino
Department of Mathematics, University of Vienna, Austria

Federico Pichi
Chair of Computational Mathematics and Simulation Science, École Polytechnique
Fédérale de Lausanne, Switzerland

Francesco Romor
mathLab, Mathematics Area, SISSA, Scuola Internazionale Superiore di Studi
Avanzati, Trieste, Italy

Gianluigi Rozza
mathLab, Mathematics Area, SISSA, Scuola Internazionale Superiore di Studi
Avanzati, Trieste, Italy

Nirav Shah
mathLab, Mathematics Area, SISSA, Scuola Internazionale Superiore di Studi
Avanzati, Trieste, Italy

Pierfrancesco Siena
mathLab, Mathematics Area, SISSA, Scuola Internazionale Superiore di Studi
Avanzati, Trieste, Italy

Giovanni Stabile
mathLab, Mathematics Area, SISSA, Scuola Internazionale Superiore di Studi
Avanzati, Trieste, Italy

Maria Strazzullo
Department of Mathematical Sciences “G.L. Lagrange”, Politecnico di Torino, Italy

Marco Tezzele
Oden Institute for Computational Engineering and Sciences, University of Texas at
Austin, USA

Davide Torlo
mathLab, Mathematics Area, SISSA, Scuola Internazionale Superiore di Studi
Avanzati, Trieste, Italy

Zakia Zainib
Technische Universität Dortmund, Germany

Matteo Zancanaro
mathLab, Mathematics Area, SISSA, Scuola Internazionale Superiore di Studi
Avanzati, Trieste, Italy
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

List of Figures

2.1 Left: example of a triangulation over the unit square resulting from Listing
2.1. Right: highlighted edges for the application of boundary conditions. . . . 18
2.2 P1 and P2 element types for two- and three-dimensional cells in (a), (b) and
(c), (d), respectively. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.3 Basis function φj and its support highlighted in pink. . . . . . . . . . . . . . . 19
2.4 Solution of the system (2.10), with F = sin(µ[0]x) cos(µ[1]y), for several
values of µ. Here, x and y indicate the spatial coordinates of the spatial do-
main. Left: µ = (2π, 2π). Center: µ = (2π, π). Right: µ = (π, π). . . . . . 20
2.5 Flow past a cylinder: solution of the system (2.23) with a parabolic inlet on the
left boundary. Velocity and pressure fields from left to right, respectively. No-
slip conditions on the bottom and top walls and on the cylinder and Neumann
boundary conditions on the right outflow. . . . . . . . . . . . . . . . . . . . . 23
2.6 Flow past a cylinder: solution of the system (2.25) with a parabolic inlet on the
left boundary and velocity and pressure fields from left to right, respectively.
No-slip conditions on the bottom and top walls and on the cylinder and Neu-
mann boundary conditions on the right outflow. The solution is represented
for several values of Re. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.7 Solution manifold at the discrete level represented by the snapshots. . . . . . . 27
2.8 Two-dimensional domain, composed of two subdomains, for the steady heat
conduction problem. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.9 Snapshots and basis for the thermal block problem. . . . . . . . . . . . . . . . 33
2.10 FE (369 nonzero elements) and RB (225 nonzero elements) structures of the
assembled systems, respectively left and right. . . . . . . . . . . . . . . . . . 35
2.11 Left: parametric domain Ωo (µ). Right: reference domain Ω. . . . . . . . . . . 36
2.12 Top: Stokes flow in deformable channels (RBniCS tutorial 12). Bottom:
Backward-facing step for moderate Reynolds, Navier–Stokes equations (RB-
niCS tutorial 17). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.13 Velocity and pressure reduced solution for the Navier–Stokes equations in the
backward-facing step channel, varying the number of supremizers Ns . . . . . . 42
2.14 Reduced manifold MN ,K for time-dependent problem. . . . . . . . . . . . . 43
2.15 Left: RB temperature solution. Right: computational mesh for Ω = (−1, 1)2
with EIM interpolation points. . . . . . . . . . . . . . . . . . . . . . . . . . . 57
2.16 Convergence of the RB method for the elliptic problem with Gaussian source,
with respect to the number of basis functions, for EIM (left) and RB (right). . . 58

3.1 Domain Ω with the different boundaries identified. . . . . . . . . . . . . . . . 67

3.2 Convergence of the EIM algorithm. . . . . . . . . . . . . . . . . . . . . . . . 67
3.3 Convergence of the greedy algorithm (left) and the value of ∆Nmax (µ) and the
exact error between the FE solution and the RB solution (right). . . . . . . . . 68

xix
 xx List of Figures

3.4 FE solution (left) and RB solution (right) for µ = 4521. . . . . . . . . . . . . 68

3.5 Unit square cavity domain Ω, with the different boundaries identified, for
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

problem (3.30). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
3.6 FE solution, velocity magnitude (top) and temperature (bottom), for µ =
4523, µ = 55732, and µ = 642639 (left to right). . . . . . . . . . . . . . . . . 76
3.7 Error evolution for the EIM for µ ∈ [103 , 105 ] (left) and µ ∈ [105 , 106 ] (right). 76
3.8 Evolution of the a posteriori error bound in the greedy algorithm for µ ∈
[103 , 105 ] (left) and µ ∈ [105 , 106 ] (right). . . . . . . . . . . . . . . . . . . . . 77
3.9 A posteriori error bound for N = Nmax for µ ∈ [103 , 105 ] (left) and µ ∈
[105 , 106 ] (right). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
3.10 FE snapshots for µg = 0.5 (top left), µg = 1 (bottom left), µg = 1.5 (middle),
and µg = 2 (right). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.11 Error evolution for the EIM and for the Boussinesq VMS-Smagorinsky model
with µ ∈ [103 , 104 ] × [0.5, 2]. . . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.12 Evolution of the a posteriori error bound in the greedy algorithm for only the
geometrical parameter (left) and for both physical and geometrical parameters
(right) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.13 A posteriori error bound for only the geometrical parameter (left) and for both
physical and geometrical parameters (right), for N = Nmax . . . . . . . . . . . 81

4.1 Graphical representation of the OCP(µ) pipeline. . . . . . . . . . . . . . . . . 84

4.2 Left: zoom of the mesh overlapping satellite images (Trieste harbor). Right:
subdomains and boundaries. Orange: observation domain Ωobs . Green: con-
trol domain Ωu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4.3 Left and center: optimal FE and ROM state pollutant concentration for µ =
(1., −1., 1.), representing the Bora wind action. Right: averaged relative log-
error for the variables. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4.4 Left: zoom of the mesh overlapping satellite images (Florida peninsula). Right:
triangulation of the spatial domain Ω, representing the North Atlantic Ocean,
from the Florida peninsula to northern Europe. . . . . . . . . . . . . . . . . . 90
4.5 Top: comparison between desired state, FE and ROM solutions for µ =
(10−4 , 0.073, 0.0452), representing the Gulf Stream. Bottom: averaged rel-
ative log-error for all the variables. . . . . . . . . . . . . . . . . . . . . . . . 91
4.6 Domain Ω. Observation domain: Ωd (µ3 ) = Ω3 (µ3 ) ∪ Ω4 (µ3 ). Control do-
main: ΓC (µ3 ) (red solid line). Blue dashed line: Dirichlet boundary condi-
tions. The reference domain Ω is given by µ3 = 1. . . . . . . . . . . . . . . . 93
4.7 Left: FE state solutions for t = 1 s, 2 s, 3 s for µ = (12.0, 2.5, 2.0). Right:
ROM state solutions for t = 1 s, 2 s, 3 s for µ = (12.0, 2.5, 2.0). . . . . . . . 94
4.8 Averaged relative log-error for the variables. . . . . . . . . . . . . . . . . . . 94

5.1 A beam subject to a load and its buckled configuration. . . . . . . . . . . . . . 100

5.2 An elastic and rectangular plate compressed along the edges parallel to the y
direction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
5.3 Left: high-fidelity bifurcation diagram for the rectangular plate with L = 2.
Right: high-fidelity displacements linked to the five branches at µ = 65. . . . . 111
5.4 Left: RB error on P for the first branch. Right: RB contour error plot of the
displacement u at µ = 65. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.5 Left: RB bifurcation diagram for the multiparameter test case with infinity
norm of the density as output. The black dotted line shows the critical values
of µ. Right: RB error on P for density in the L2 - and H01 -norms for τ = 0.18. 114
 List of Figures xxi

5.6 Left: RB density functions at µ = 0.8 with τ ∈ {0.15, 0.25} for the first three
branches. Right: RB bifurcation diagram for the multiparameter test case with
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

infinity norm of the density as output. The blue dashed lines show the critical
values of µ for both branching phenomena. . . . . . . . . . . . . . . . . . . . 115
5.7 Left: error between the branches of the bifurcation diagram computed with
FE and RB with EIM/DEIM in the µ-NB plane. Right: high-fidelity and RB
real part φ of the solution X for the 1-dark soliton stripe branch, left and right,
respectively, at µ = 0.5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
5.8 Left: RB bifurcation diagrams for the SVK unforced beam. Right: high-
fidelity zeroth and second mode displacement u for the SVK beam with B =
(0, 0) at µ = 0.2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
5.9 Left: RB bifurcation plot for SVK beam with B = (0, −1000) for five random
pairs (E, ν) ∈ [105 , 107 ] × [0.25, 0.42]. Right: RB bifurcation plot for SVK
beam with B = (0, −1000) and l ∈ [0.5, 1]. . . . . . . . . . . . . . . . . . . . 119
5.10 Left: RB bifurcation diagram for tubular unforced geometries with l ∈ [10, 20].
Right: representative solutions of the three-dimensional SVK model with l ∈
{2, 20}. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
5.11 Domain Ω, which represents a straight channel with a narrow inlet. . . . . . . 120
5.12 Left: RB bifurcation diagram for the NS system. Right: velocity magnitude
of the asymmetric branch varying the viscosity µ. . . . . . . . . . . . . . . . . 122
5.13 Representative solutions for the NS system at µ = 0.5, velocity and pressure
fields, lower and middle branch, top and bottom respectively. . . . . . . . . . . 122
5.14 RB errors with respect to µ ∈ P for the velocity and pressure of the asymmet-
ric branch. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
5.15 Eigenvalues of the state eigenproblem in the complex plane for the NSE: stable
and unstable solutions, left and right panels, respectively. . . . . . . . . . . . . 123

6.1 Comparison of the stabilization with/without supremizer (left) and evolution

of the error in the EIM (right). . . . . . . . . . . . . . . . . . . . . . . . . . . 138
6.2 Evolution of the error in the greedy algorithm (left) and a posteriori error esti-
mator for N = 16, with error (right). . . . . . . . . . . . . . . . . . . . . . . 138
6.3 FE (left) and RB (right) solution for µ = 2751. . . . . . . . . . . . . . . . . . 139

7.1 Sketch of a finite volume in two dimensions. . . . . . . . . . . . . . . . . . . 145

7.2 Sketch of the mesh for the lid-driven cavity problem together with the bound-
ary subdivisions and boundary conditions. . . . . . . . . . . . . . . . . . . . . 157
7.3 Error analysis for the velocity field. The L2 -norm of the relative error is plot-
ted over time for three different models: with supremizer stabilization (USUP,
continuous red line), with PPE stabilization (UPPE, dotted blue line), and
without stabilization (PNOS, dashed green line). The ROMs are obtained with
10 modes for velocity, pressure, and supremizer. . . . . . . . . . . . . . . . . 158
7.4 Comparison of the velocity and pressure fields for high-fidelity (UHF, column
1; PHF, column 4), SUP-ROM (USUP, column 2; PSUP, column 5), and PPE-
ROM (UPPE, column 3; PPEE, column 6). The fields are depicted for different
time instants equal to t = 0.2 s, 0.5 s, 1 s, and 5 s, respectively, and increasing
in the image from top to bottom. The ROM models are obtained with 10
modes for velocity and pressure and for the SUP-ROM only with 10 additional
supremizer modes. The velocity and pressure magnitudes are shown in the
image legends. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
 xxii List of Figures

7.5 The figure shows a general overview of the mesh with dimension and bound-
aries (upper left), a table with the imposed values at the boundaries (bottom),
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

and a zoom of the mesh near the cylinder (upper right). . . . . . . . . . . . . . 160
7.6 Kinetic energy relative error in the cylinder example for ν = 0.005 and
ν = 0.005625. The kinetic energy relative error is plotted over time for the
two values of viscosity and for the two different models: with supremizer sta-
bilization (USUP and PSUP) and PPE stabilization (UPPE and PPPE). The
ROM solutions are obtained with 15 modes for velocity, 10 modes for pres-
sure, and 12 modes for supremizers. The time window in this case is wider
than the one used to generate the RB spaces (∆T = 10 s). . . . . . . . . . . . 161
7.7 First four basis functions for velocity (first row), pressure (second row), and
supremizers (third row). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
7.8 Comparison of the velocity field for high-fidelity (UHF, first row), supremizer
stabilized ROM (USUP, second row), and PPE stabilized ROM (UPPE, third
row). The fields are depicted for different time instants equal to t = 195 s,
200 s, 230 s, and 270 s and increasing from left to right. The ROM solu-
tions are obtained with 15 modes for velocity and 10 modes for pressure, and
for the SUP-ROM only with 12 additional supremizer modes. The velocity
magnitude is shown in the image legends. . . . . . . . . . . . . . . . . . . . . 162
7.9 Comparison of the pressure field for high-fidelity (PHF, first row), supremizer
stabilized ROM (PSUP, second row), and PPE stabilized ROM (PPPE, third
row). The fields are depicted for different time instants equal to t = 195 s,
200 s, 230 s, and 270 s and increasing from left to right. The ROM solu-
tions are obtained with 15 modes for velocity and 10 modes for pressure, and
for the SUP-ROM only with 12 additional supremizer modes. The pressure
magnitude is shown in the image legends. . . . . . . . . . . . . . . . . . . . . 162
7.10 Error analysis for the velocity (left plot) and pressure (right plot) fields in
the cylinder example with ν = 0.005625. The L2 -norm of the relative error
is plotted over time for the two different models: with supremizer stabiliza-
tion (USUP and PSUP, continuous red line) and PPE stabilization (UPPE and
PPPE, dot-dashed blue line). The ROM solutions are obtained with 15 modes
for velocity, 10 modes for pressure, and 12 modes for supremizers. . . . . . . 163
7.11 Error analysis for the velocity (left plot) and pressure (right plot) fields in the
cylinder example with ν = 0.005 and ν = 0.005625. The L2 -norm of the
relative error is plotted over time for the two values of viscosity and for the
two different models: with supremizer stabilization (USUP and PSUP) and
PPE stabilization (UPPE and PPPE). The ROMs are obtained with 15 modes
for velocity, 10 modes for pressure, and 12 modes for supremizers. The time
window is in this case wider than the one used to generate the RB spaces
(∆T = 10 s). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163

8.1 The computational domain used in the numerical simulations; all lengths are
described in terms of the characteristic length D, which is equal to 1 m. . . . . 181
8.2 Cumulative ignored eigenvalue decay. In the plot, the solid black line refers to
the velocity eigenvalues, the dashed magenta line indicates the pressure eigen-
values, and the dash-dotted blue line refers to the eddy viscosity eigenvalues. . 182
8.3 k- turbulence model case, velocity fields for the value of the parameter U =
7.0886 m/s: (a) shows the FOM velocity, while in (b) one can see the U-ROM
velocity, and finally in (c) we have the H-SUP-ROM velocity. . . . . . . . . . 183
 List of Figures xxiii

8.4 k- turbulence model case, pressure fields for the value of the parameter U =
7.0886 m/s: (a) shows the FOM pressure, while in (b) one can see the U-ROM
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

pressure, and finally in (c) we have the H-SUP-ROM pressure. . . . . . . . . . 183

8.5 k- turbulence model case, eddy viscosity fields: (a) shows the FOM eddy
viscosity, while in (b) one can see the U-ROM eddy viscosity, and finally in
(c) we have the H-SUP-ROM eddy viscosity. . . . . . . . . . . . . . . . . . . 184
8.6 SST k-ω turbulence model case, velocity fields for the value of the parameter
U = 7.0886 m/s: (a) shows the FOM velocity, while in (b) one can see the
U-ROM velocity, and finally in (c) we have the H-SUP-ROM velocity. . . . . . 184
8.7 SST k-ω turbulence model case, pressure fields for the value of the parameter
U = 7.0886 m/s: (a) shows the FOM pressure, while in (b) one can see the
U-ROM pressure, and finally in (c) we have the H-SUP-ROM pressure. . . . . 185
8.8 SST k-ω turbulence model case, eddy viscosity fields: (a) shows the FOM
eddy viscosity, while in (b) one can see the U-ROM eddy viscosity, and finally
in (c) we have the H-SUP-ROM eddy viscosity. . . . . . . . . . . . . . . . . . 185
8.9 The pressure fields obtained using both k- and SST k-ω turbulence models
and the H-SUP-ROM ones for the value of the parameter U = 7.0886 m/s.
The plot is for the pressure value along the x1 direction, keeping the value of
x2 fixed at half the maximum height. . . . . . . . . . . . . . . . . . . . . . . 186
8.10 The mean of the L2 relative errors for all the online samples versus the num-
ber of modes used in the online stage. The convergence analysis is done for
both H-SUP-ROM models obtained with two different turbulence models at
the full-order level, which are k- and SST k-ω. The errors are reported in
percentages—in (a) we have the mean error of the velocity fields, while in (b)
we have the mean error of the pressure fields. . . . . . . . . . . . . . . . . . . 186
8.11 (a) The OpenFOAM mesh used in the simulations for the unsteady case of the
flow around a circular cylinder. (b) A picture of the mesh zoomed near the
cylinder. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
8.12 The lift coefficient curve for parameter sample Uin = 10 m/s. . . . . . . . . . 188
8.13 Cumulative ignored eigenvalue decay. In the plot, the solid black line refers to
the velocity eigenvalues, the dashed magenta line indicates the pressure eigen-
values, and the dash-dotted blue line refers to the eddy viscosity eigenvalues. . 189
8.14 Velocity fields for the parameter value Uin = 7.75 m/s at t = 2.8 s: (a)
shows the FOM velocity, while in (b) one can see the U-ROM velocity with
Nr = 14, and finally in (c) we have the H-SUP-ROM velocity with Nu = 20
and Np = NS = Nνt = 10. . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
8.15 Pressure fields for the parameter value Uin = 7.75 m/s at t = 2.8 s: (a)
shows the FOM pressure, while in (b) one can see the U-ROM pressure with
Nr = 14, and finally in (c) we have the H-SUP-ROM pressure with Nu = 20
and Np = NS = Nνt = 10. . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
8.16 Eddy viscosity fields for the parameter value Uin = 7.75 m/s at t = 2.8 s:
(a) shows the FOM eddy viscosity, while in (b) one can see the U-ROM eddy
viscosity with Nr = 14, and finally in (c) we have the H-SUP-ROM eddy
viscosity with Nu = 20 and Np = NS = Nνt = 10. . . . . . . . . . . . . . . 192
8.17 The time evolution of the L2 relative errors of the velocity-reduced approx-
imations for both the U-ROM and the H-SUP-ROM. The curves correspond
to the case run with the parameter value Uin = 7.75 m/s: (a) shows the er-
ror curve for the U-ROM, while (b) depicts the case of the H-SUP-ROM. The
error values in both graphs are in percentages. . . . . . . . . . . . . . . . . . . 193
 xxiv List of Figures

8.18 The time evolution of the L2 relative errors of the pressure-reduced approx-
imations for both the U-ROM and the H-SUP-ROM. The curves correspond
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

to the case run with the parameter value Uin = 7.75 m/s: (a) shows the er-
ror curve for the U-ROM, while (b) depicts the case of the H-SUP-ROM. The
error values in both graphs are in percentages. . . . . . . . . . . . . . . . . . . 193
8.19 Lift coefficient curves for the cross-validation test done for the parameter value
Uin = 7.75 m/s for the time range [0, 8] s; the figure shows the FOM, the
U-ROM, and the H-SUP-ROM lift coefficient histories: (a) the full range is
shown, and (b) the last 2 s of Cl is shown. . . . . . . . . . . . . . . . . . . . . 194
8.20 The graph of the L2 relative errors for the lift coefficient curve versus the
number of modes used in the online stage in the U-ROM and the H-SUP-ROM.
The curves correspond to the case run with the parameter value Uin = 7.75
m/s. The error is computed between the lift coefficient curve obtained by the
FOM solver and the one reconstructed from both the U-ROM and the H-SUP-
ROM for the time range [0, 8] s: (a) shows the error curve for the U-ROM,
where Nr is the number of modes used in the online stage for all variables (by
construction of the U-ROM, it is not possible to choose a different number of
online modes for the reduced variables). (b) depicts the case of the H-SUP-
ROM, where one can see the error values by varying the number of modes used
for the pure velocity, with different fixed settings for the three other variables
(the pressure, the supremizers, and the eddy viscosity). The error values in
both graphs are in percentages. . . . . . . . . . . . . . . . . . . . . . . . . . . 195
8.21 The graph of the peak relative errors for the lift coefficient curves for varied
values of the number of modes used in the online stage in the U-ROM and the
H-SUP-ROM. The curves correspond to the case run with the parameter value
Uin = 7.75 m/s. The error is computed between the peak values of the lift
coefficient curve obtained by the FOM solver and the ones reconstructed from
both the U-ROM and the H-SUP-ROM for the time range [0, 8] s: (a) shows
the error curve for the U-ROM, where Nr is the number of modes used in the
online stage for all variables (by construction of the U-ROM it is not possible
to choose a different number of online modes for the reduced variables). (b)
depicts the case of the H-SUP-ROM. The error values in both graphs are in
percentages. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
8.22 Lift coefficient curves for the cross-validation test done for the parameter value
Uin = 7.75 m/s for the time range [0, 8] s. The figure shows the FOM, the SU-
PPE-ROM, and the H-PPE-ROM lift coefficient histories: (a) the full range is
shown, and (b) the last 2 s of Cl is shown. . . . . . . . . . . . . . . . . . . . . 197
8.23 The graph of the L2 relative errors for the lift coefficient curve versus number
of modes used in the online stage in the SU-PPE-ROM and the H-PPE-ROM.
The curves correspond to the case run with the parameter value Uin = 7.75
m/s. The error is computed between the lift coefficient curve obtained by the
FOM solver and the one reconstructed from both the U-ROM and the H-PPE-
ROM for the time range [0, 8] s: (a) shows the error curve for the SU-PPE-
ROM, where Nu is the number of modes used in the online stage for both the
velocity and the eddy viscosity, while Np is the number of modes used for the
pressure field. (b) depicts the case of the H-PPE-ROM, where one can see the
error values varying the number of modes used for the velocity (including the
lifting velocity mode) with different fixed settings for the two other variables
(the pressure and the eddy viscosity). The error values in both graphs are in
percentages. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
 List of Figures xxv

8.24 The graph of the peak relative errors for the lift coefficient curves for varied
values of the number of modes used in the online stage in the SU-PPE-ROM
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

and the H-PPE-ROM. The curves correspond to the case run with the param-
eter value Uin = 7.75 m/s. The error is computed between the peak values
of the lift coefficient curve obtained by the FOM solver and the ones recon-
structed from both the SU-PPE-ROM and the H-PPE-ROM for the time range
[0, 8] s: (a) shows the error curve for the U-ROM, where Nu is the number
of modes used in the online stage for both the velocity and the eddy viscosity,
while Np is the number of modes used for the pressure field. (b) depicts the
case of the H-PPE-ROM. The error values in both graphs are in percentages. . 199
8.25 The time evolution of the L2 relative errors of the velocity-reduced approxima-
tions for both the SU-PPE-ROM and the H-PPE-ROM. The curves correspond
to the case run with the parameter value Uin = 7.75 m/s: (a) shows the er-
ror curve for the SU-PPE-ROM. (b) depicts the case of the H-PPE-ROM. The
error values in both graphs are in percentages. . . . . . . . . . . . . . . . . . . 199
8.26 The time evolution of the L2 relative errors of the pressure-reduced approxi-
mations for both the SU-PPE-ROM and the H-PPE-ROM. The curves corre-
spond to the case run with the parameter value Uin = 7.75 m/s: (a) shows the
error curve for the SU-PPE-ROM. (b) depicts the case of the H-PPE-ROM.
The error values in both graphs are in percentages. . . . . . . . . . . . . . . . 200
8.27 Lift coefficient curves for the cross-validation test done for the parameter value
Uin = 11.75 m/s for the time range [0, 10] s. The figure shows the FOM and
the H-SUP-ROM lift coefficient histories: (a) the full range is shown; (b) the
last 3-s history of Cl is shown. . . . . . . . . . . . . . . . . . . . . . . . . . . 201
8.28 The lift coefficient curves obtained using both the k- and SST k-ω turbulence
models and the H-SUP-ROM ones. The case considered is nonparametrized,
with Uin = 10 m/s corresponding to Re = 105 . The plot is for the time
range t ∈ [6, 8]; the H-SUP-ROM achieved relative L2 errors (over the range
t ∈ [0, 8]) less than 5% in both cases. . . . . . . . . . . . . . . . . . . . . . . 202

9.1 First three POD modes obtained from the resolution of a parametrized Navier–
Stokes problem, describing the flow past a cylinder. . . . . . . . . . . . . . . . 206
9.2 In the left column are the eigenvalue positions with respect to the unit circle,
and in the right column are the corresponding DMD modes. From top to
bottom we have a stable, a convergent, and a divergent mode, corresponding
to an eigenvalue on the unit circle, inside it, and outside it, respectively. . . . . 209
9.3 On the left is the initial dataset containing features at different time scales.
In the middle is the DMD reconstruction, and on the right is the mrDMD
reconstruction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
9.4 On the left is the initial dataset, in the middle is the DMD reconstruction, and
on the right is the cDMD reconstruction. . . . . . . . . . . . . . . . . . . . . 211
9.5 Computational savings of cDMD with respect to the exact DMD for an in-
creasing dimension of the snapshot matrix. . . . . . . . . . . . . . . . . . . . 211
9.6 Representation of the computational pipeline used for the shape optimization
problem. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
9.7 Example of bulbous bow deformations using the FFD method. . . . . . . . . . 213
9.8 Comparison of the average relative L2 error of the ROM in a naval shape opti-
mization problem as a function of the number of modes (left) and the number
of snapshots used (right). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
9.9 Flowchart of the FOM/ROM procedure. . . . . . . . . . . . . . . . . . . . . . 214
 xxvi List of Figures

9.10 Sketch of the computational domain used for the FOM. The sphere diameter is
D = 0.01 m. Successive mesh refinement layers (R2, R3, R4) are performed
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

using the cell splitting approach until the finest grid spacing 0.001D is reached
in the region R5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
9.11 Normalized singular values for streamwise velocity and pressure snapshots. . . 216
9.12 Isosurfaces of the DMD modes (left) at a value of −15 × 10−5 m/s versus
the isosurfaces of POD modes (right) at a value of 6.5 m/s for the streamwise
velocity field. Modes: 1, 2, 8, and 36. . . . . . . . . . . . . . . . . . . . . . . 216
9.13 Isosurfaces of the DMD modes (left) at a value of −5 × 10−4 Pa versus the
isosurfaces of POD modes (right) at a value of 300 Pa for the pressure field.
Modes: 1, 2, 8, and 36. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
9.14 Relative error percentage in the Frobenius norm for velocity and pressure (top
and bottom, respectively) fields, where half-sample-rate snapshots are used to
train the reduced model, DMD (left) or PODI (right), for a particular trun-
cation (r) while predicting the intermediate snapshots. Plots show only the
snapshotwise prediction error, while disregarding errors in the training set,
which are almost null. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
9.15 Linear dipole (top) and nonlinear quadrupole (bottom) terms of FWH equa-
tion evaluated from LES data and compared to corresponding DMD (left) and
PODI (right) reduced models at different truncation ranks (r). . . . . . . . . . 218
9.16 Flowchart representing the proposed computational pipeline. . . . . . . . . . . 218
9.17 Sketch of the computational domain used to solve the fluid dynamics problem
in its reference configuration. The left plot is a zoom on the mesh near the
wing. The right picture is a schematic view of the domain with the main
geometrical dimensions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
9.18 The temporal evolution of the lift coefficient from 1 s to 30 s for nine different
parameters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
9.19 Airfoil shape functions with respect to the profile abscissa. The leading edge
corresponds to x = 0. The bump functions are rescaled by a factor of 0.2 for
illustrative reasons. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
9.20 On the left is the sufficiency summary plot for the lift coefficient at times
t = 10.0, 14.0, 18.0 s (top to bottom). On the right are the first eigenvector
components at the corresponding parameters. . . . . . . . . . . . . . . . . . . 221
9.21 The relative error of the approximated outputs at different times. The relative
error is computed on 100 test samples, using the high-fidelity lift coefficient
to train the regression for t ≤ 20 s, while for t > 20 s the DMD forecasted
states are used for the training. . . . . . . . . . . . . . . . . . . . . . . . . . . 222

10.1 Full-order, steady-state solution for ν = 10: velocity in the x direction (top)
and y direction (bottom). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
10.2 Full-order, steady-state solution for ν = 0.05: velocity in the x direction (top)
and the y direction (bottom). . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
10.3 Mean and maximum relative L2 (Ω) error in the velocity with increasing basis
size. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232

11.1 Mapping from the physical to the master element. . . . . . . . . . . . . . . . . 236

11.2 Quadrangular master element with nodes. . . . . . . . . . . . . . . . . . . . . 237
11.3 Flux exchange between contiguous elements. . . . . . . . . . . . . . . . . . . 238
11.4 Mesh. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244
11.5 First four velocity modes for the square cylinder test case. . . . . . . . . . . . 245
 List of Figures xxvii

11.6 Reduced order solution for Nu = 5, Nu = 8, Nu = 10, and Nu = 12. . . . . . 245

11.7 Absolute error (velocity). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

11.8 Absolute error (pressure). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245

11.9 Mesh of the domain. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
11.10 First four modes (velocity). . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
11.11 Comparison between FOM (top) and ROM (bottom) solutions. . . . . . . . . . 247

12.1 Comparison of multivariate quadrature rules with order of accuracy 15. . . . . 256
12.2 Heat transfer problem: error decay with respect to the number of reduced ba-
sis functions, comparison between weighted and not weighted algorithms and
between uniform and Beta(20,10) distributed training samples. Left: POD;
right: greedy algorithm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
12.3 Reference domain D. The reference parameter is µ = (1., 1.5, 1., 1.5, µ4 ). . . 258
12.4 Stokes problem: error decay with respect to the number of RB functions,
comparison between unweighted (left) and weighted (right) algorithms and
between the Smolyak, tensor product, uniform, and Beta(75,75) distributed
training samples. For the weighted algorithm, the Smolyak rule, the tensor
product rule, and Beta(75,75) sampling coincide. . . . . . . . . . . . . . . . . 259
12.5 Graetz problem: error decay with respect to the number of RB functions,
and comparison between weighted and unweighted algorithms and between
uniform and Beta(5,3) distributed training samples. With online stabilization
(left), without online stabilization (right). POD algorithm (top), greedy algo-
rithm (bottom). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
12.6 Graetz problem: weighted algorithms and Beta(5,3) distributed training sam-
ples. Error of stabilized and nonstabilized online with respect to the advection
parameter µ. Comparison with selective online strategy. Left: POD; right:
Greedy. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
12.7 Domain D, the Gulf of Trieste. Orange: Miramare reserve Dobs . Pollutant
spill Du . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262
12.8 Optimal control problem: error decay with respect to the number of RB func-
tions; comparison between state variable (left) and control variable (right) with
an unweighted POD algorithm between uniform, Beta(20, 5), Beta(5, 5), and
Beta(75,75) distributed training samples. . . . . . . . . . . . . . . . . . . . . 263

13.1 Embedded FEM geometrical tools for (i) CutFEM and (ii) SBM. . . . . . . . 267
13.2 (i) Parametrized geometry. (ii) Some RB components for µ ∈ [−0.5, 0.5]
parametrized geometry. (iii) The full/reduced order solution and the absolute
error (µ = −0.015). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
13.3 Heat exchange problem: (i) eigenvalue decay and (ii) mean relative errors. . . 270
13.4 Stokes with SBM experiment: sketch of the mesh, the embedded domain, and
the parameters considered in the numerical examples. . . . . . . . . . . . . . 272
13.5 Stokes system ROM-SBM POD velocity and pressure components for (i) µ1
geometry and (ii) (µ0 , µ1 ) geometry. . . . . . . . . . . . . . . . . . . . . . . 273
13.6 Stokes SBM: the full and the reduced solution and the absolute error for
velocity and pressure for the (i) one-dimensional and (ii) two-dimensional
parametrization case. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274
13.7 Stokes (SBM) parametrization: (i) µ1 geometry, with the relative errors for
velocity and pressure in (i)a and the execution times in (i)b ; (ii) (µ0 , µ1 ) ge-
ometry, with the velocity and pressure fields with and without supremizer sta-
bilization. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
 xxviii List of Figures

13.8 Shape parametrization with large deformations and a zoom into the embedded
cylinder visualizing the extended solution. . . . . . . . . . . . . . . . . . . . . 277
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

13.9 Six classical (i) and six improved (ii) POD modes. . . . . . . . . . . . . . . . 278
13.10 (i) Poisson system (CutFEM): eigenvalue decay and error analysis between
reduced order and high-fidelity approximations with and without transport;
(ii) steady Stokes (CutFEM): relative errors with and without transport. . . . . 279
13.11 Circular embedded geometry (parametrized). . . . . . . . . . . . . . . . . . . 280
13.12 Cahn–Hilliard (CutFEM): ROM basis results, the first six modes. . . . . . . . 281
13.13 Results for the embedded circle geometrical parametrization µ = 0.4261 and
t = [28, 36, 46, 100]dt. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
13.14 Cahn–Hilliard (CutFEM): the FOM mass evolution with respect to time, to-
gether with its RB approximation and the conservation of mass. . . . . . . . . 282

14.1 Example: domain reference configuration Ω̂ (left) and domain configuration

at time t, Ω(t) (right). In blue we have the fluid domain; in orange the solid
domain. In green is the fluid-structure interface Γ̂F SI in the reference config-
uration. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
14.2 Original configuration at time t. Blue domain: the original fluid configuration
Ωf (t). Orange leaflets: the original solid configuration Ωs (t). ΓDs : the part of
the leaflets that does not move. . . . . . . . . . . . . . . . . . . . . . . . . . . 287
14.3 Domains: reference configuration Ω̂ (top left), parametrized reference config-
uration Ω̃(µg ) (top right), and original configuration Ω(t; µg ) (bottom). . . . . 295
14.4 Reduced order solid displacement dN s (µ). Comparison of different behaviors
of the material, for different values of the geometrical and physical parameters.
From left to right: same leaflet length µg = 0.8 cm and increased shear mod-
ulus (µs = 100000, 800000); same leaflet length µg = 1 cm and increased
shear modulus (µs = 100000, 800000); increased leaflet length (µg = 0.8,
1.0 cm), and same shear modulus µs = 100000. . . . . . . . . . . . . . . . . 296
14.5 Physical domain (reference configuration): fluid subdomain (blue) and struc-
ture subdomain (orange). The fluid-structure interface coincides with the struc-
ture in our case; the structure has been magnified for visualization purposes. . . 300
14.6 Fluid pressure behavior: the solution is pictured here at times t = 0.001,
t = 0.005, and t = 0.015. The peak of the wave is propagating into the
domain, creating a transport phenomenon. . . . . . . . . . . . . . . . . . . . . 303
14.7 Vertical component of fluid displacement behavior: again the solution is pic-
tured at times t = 0.001, t = 0.005, and t = 0.015. The peak of the wave is
still very small at the beginning, it grows for some time, and then it starts to
propagate. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
14.8 Decay of the eigenvalues for the POD on the fluid pressure (blue line), fluid
displacement (green line), and fluid velocity (magenta line). . . . . . . . . . . 304
14.9 Top: comparison between the rate of decay of the eigenvalues for the pressure
with and without preprocessing. Bottom: retained energy as a function of the
number N of POD modes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306
14.10 Top: comparison between the rate of decay of the eigenvalues for the displace-
ment with and without preprocessing. Bottom: retained energy as a function
of the number N of POD modes. . . . . . . . . . . . . . . . . . . . . . . . . 306
14.11 Original snapshots for pf at times t = 0.001, t = 0.005, and t = 0.015 (left
column) and corresponding preprocessed snapshots (right column). . . . . . . 307
 List of Figures xxix

14.12 Pressure snapshots (left column) at time t = 0.005 (top) and final time t =
0.015 (bottom). Reduced order pressure simulation (right column) at time
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

t = 0.005 and time t = 0.015. The reduced simulation was obtained with
N = 4 basis functions for each component of the solution of the FSI problem. 309
14.13 Displacement snapshots (left column) at time t = 0.005 (top) and final time
t = 0.015 (bottom). Reduced order displacement simulation (right column)
at time t = 0.005 and time t = 0.015. The reduced simulation was obtained
with N = 4 basis functions for each component of the solution of the FSI
problem. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
14.14 Analysis of the behavior of the relative error for the fluid pressure approxima-
tion. Dashed lines were obtained by employing the preprocessing procedure,
continuous lines were obtained using a standard model order reduction. . . . . 309
14.15 Behavior of the mean relative error for the fluid pressure, depending on the
number N of basis functions employed, with preprocessing (red) and without
preprocessing (blue). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 310

15.1 Reference configuration for the FSI problem. . . . . . . . . . . . . . . . . . . 314

15.2 Computational domain. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
15.3 Velocity and pressure snapshots. . . . . . . . . . . . . . . . . . . . . . . . . . 319
15.4 Solid’s deformation snapshots: the solutions are given in a spatial frame using
the df map. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
15.5 Global displacement (df , ds ) for µ = 2.0. . . . . . . . . . . . . . . . . . . . 319
15.6 Bifurcation diagram for the fluid phase. . . . . . . . . . . . . . . . . . . . . . 320
15.7 Bifurcation diagram for the solid phase. . . . . . . . . . . . . . . . . . . . . . 320
15.8 Retained energy. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
15.9 Decay of the normalized eigenvalues. . . . . . . . . . . . . . . . . . . . . . . 321
15.10 Average reconstruction error in logarithmic scale. . . . . . . . . . . . . . . . . 321
15.11 Average projection error in logarithmic scale. . . . . . . . . . . . . . . . . . . 321
15.12 Relative reconstruction error with Nrb = 16. . . . . . . . . . . . . . . . . . . 322
15.13 Maximum displacement for the SVK and the linear models. . . . . . . . . . . 323
15.14 Solid’s bifurcation diagram for the linear and the SVK model. . . . . . . . . . 323
15.15 Comparison of the bifurcation diagram for the fluid phase for the different test
cases. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323

16.1 Illustration of the inactive variable sampling strategy. Successive steps are
depicted all at once. We highlight the Chebyshev center, the selection of the
next sample using the hit and run method, and the polytope defined by (16.8). . 329
16.2 Illustration of the design of a response surface using KAS and Gaussian pro-
cess regression (GPR). The factorization through an RKHS is emphasized. . . 331
16.3 Illustration of the global AS direction, highlighted in orange in the left panel.
On the right is the corresponding sufficient summary plot. . . . . . . . . . . . 333
16.4 Comparison between K-means, K-medoids, and hierarchical top-down clus-
tering using the AS-induced distance metric for the quartic example of Fig-
ure 16.3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334
16.5 Illustrative scheme of a multifidelity regression procedure that employs AS
as a low-fidelity model. The function to be approximated is the hyperbolic
paraboloid from (16.16). Starting from 10 high-fidelity data points (depicted
in blue and in white) we construct as a low-fidelity model a response surface
which is constant along the inactive subspace. . . . . . . . . . . . . . . . . . . 336
 xxx List of Figures

16.6 NLL application two-dimensional example. On the left we have the target
function against the first transformed coordinate, in the middle the loss func-
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

tion decay, and on the right the transformed parameter space. . . . . . . . . . . 337
16.7 On the y-axis: basic reproduction number R0 of the SEIR model for the Ebola
epidemic. Left: GPR on the one-dimensional AS. Middle: GPR on the one-
dimensional KAS. Right: GPR on the one-dimensional active variable of the
NLL method. In the plots are also represented the 200 training points used to
find the reduced parameter space and build the GPRs. . . . . . . . . . . . . . . 339
16.8 Each one of the six plots represents the GPR on the reduced space of one
out of six partitions of the parameter space found with LAS. On the y-axis is
the basic reproduction number R0 of the SEIR model for the Ebola epidemic.
On the x-axis is the reduced one-dimensional AS of the relative cluster. In
the plots are also represented the 200 training points scattered among all six
clusters and used to find the partition of the parameter space with hierarchical
top-down clustering and build the GPRs. . . . . . . . . . . . . . . . . . . . . 339
16.9 ASGA scheme. The main steps of the classical GA are depicted from top to
bottom. Projections onto and from a lower-dimension AS are highlighted with
yellow boxes, which are specific to ASGA. . . . . . . . . . . . . . . . . . . . 341

17.1 Sketch of the three deformation maps which compose the FFD morphing. . . 348
17.2 Interpolations corresponding to six different RBFs. The exact function is de-
picted with a red dashed line. The parameter ε is equal to the inverse of the
average distance between the sampling points. . . . . . . . . . . . . . . . . . 349
17.3 Interpolation error for different types of RBFs. Refer to Figure 17.2 for the
actual interpolations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349
17.4 Front view (top) and side view (bottom) of a planing hull featuring a double
chine line and substantially straight walls above the chine line. . . . . . . . . 352
17.5 Vertical section of a planing hull perpendicular to its longitudinal axis. The
bottom vertex of the section is the intersection of the section plane and the keel
line. The intersections of the two chine lines are also visible on the bottom left
side of the section. The first step of the parametrization strategy is a rigid
rotation of the bottom part of the section around the keel intersection. This is
combined with a rigid translation of the chine and wall surface sections needed
to keep the top part of the section in contact with the displaced bottom part. . 353
17.6 Front view (left) and lateral view (right) of a ship propeller. In both images,
one of the five propeller blades’ surface is transparent, and a set of airfoil cylin-
drical sections is indicated in red to show how the blade surface is generated
as the envelope surface of the single sections. . . . . . . . . . . . . . . . . . 354
17.7 The airfoil section pitch as a function of relative radial coordinate. The plot
shows both an initial pitch distribution (red continuous line) and the control
points associated with its B-spline interpolation (red piecewise linear line). A
new, smooth pitch radial distribution (blue continuous line) is then obtained
using a displaced set of pitch control points (blue piecewise linear line). . . . 355
17.8 FFD application to the STL triangulation representing the bulbous bow of a
cruise ship hull. The left panel, which represents the original bow geometry,
also displays the FFD undeformed control point lattice. On the right, the vis-
ible displacement of the control point results in a deformed STL geometry.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357
 List of Figures xxxi

17.9 Two examples of a triangulated surface of a carotid artery deformed with RBF.
The original geometry is emphasized by the blue dots, while the RBF control
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

points are depicted in red (the deformed ones) and in green (the undeformed
ones). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357
17.10 A simplified two-dimensional example of the application of FFD to a NURBS
curve. On the left, the procedure starts by placing an FFD lattice (red dots)
on top of a NURBS curve. On the right, the displacement of the FFD control
points is used to modify the position of the NURBS control points (black dots).
The modified control polygon results in the morphed NURBS curve displayed
on the right. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 358
17.11 An application of the FFD algorithm to the NURBS surface specified by an
IGES file. A front view of the DTMB-5415 navy combatant hull featuring a
modified and enlarged bow is displayed on the left; the original hull surface is
displayed on the right for reference. . . . . . . . . . . . . . . . . . . . . . . . 359
17.12 A side view of the IGES geometry of the DTMB-5415 navy combatant hull
shown in Figure 17.11. The top picture depicts the CAD surface of the hull,
modified via FFD. The bottom picture shows the original hull surface. . . . . 359
17.13 Representation of a family of ship propellers generated with bottom-up con-
struction. The different propellers are obtained in this case by scaling the skew
angle curve. In particular, in the image an original propeller blade shape is pre-
sented along with two modifications resulting from scaling the skew curve by
factors of 0.8 and 1.1, respectively. . . . . . . . . . . . . . . . . . . . . . . . 360
17.14 Examples of computational domain deformation using FFD: the first picture
(left) shows the edges of the original mesh, while the others illustrate two
deformed domains. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 361
17.15 The application of RBF shape parametrization to the Ahmed body and the
surrounding volumetric mesh for CFD simulations. In the test illustrated, the
inclination of the diagonal surface at the rear end of the Ahmed body is modi-
fied to investigate its effect on aerodynamic performance. Two configurations,
including a vertical cut of the volumetric mesh, are shown in the top images.
For reference, path lines of mean velocity field resulting from the correspond-
ing unsteady RANS simulations are shown in the bottom pictures. . . . . . . 362
17.16 Example of computational domain deformation applied to a propeller design
problem, seen at the radial plane. The blue edges refer to the original mesh,
whereas the red ones show the deformed grid. . . . . . . . . . . . . . . . . . 363
17.17 A sectional view of a volumetric grid deformation carried out using RBFs in
a naval engineering problem. The blue lines denote the original volumetric
mesh edges, while the red lines indicate the deformed configuration. . . . . . 364

18.1 CABG: Boundary control magnitude mm2 /s2 (left) and eigenvalue reduc-

tion (right). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 369

18.2 CABG: Simulation velocity v: comparison between FE approximation (left)
and ROM approximation (right). . . . . . . . . . . . . . . . . . . . . . . . . . 369
18.3 CABG: Average error between FE and POD-Galerkin approximation of δ =
v, p, u, w, q (top left); total error between Galerkin FE and POD-Galerkin
approximations (top right); and error between FE and POD-Galerkin reduction
of J (bottom). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 370
18.4 The horizontal velocity of a jet stream entering a channel. . . . . . . . . . . . 371
18.5 The velocity in the channel. . . . . . . . . . . . . . . . . . . . . . . . . . . . 371
18.6 The velocity in the channel, zoomed in to the wall attachment region. . . . . . 371
 xxxii List of Figures

18.7 The mitral valve between the left atrium and the left ventricle produces the
Coandă effect. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 371
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

18.8 Scheme for the LVAD. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 374

18.9 LVAD: comparison of the FOM/ROM p (mmHg) at P F = 3.45 l/min. . . . . 376
18.10 LVAD: comparison of the FOM/ROM WSS (Pa) at P F = 3.45 l/min. . . . . 377
18.11 LVAD: comparison of the FOM/ROM u (m/s) at P F = 3.45 l/min. . . . . . 377

19.1 The logos of the ROM packages. . . . . . . . . . . . . . . . . . . . . . . . . . 380

19.2 Example of EZyRB prediction on an automotive benchmark. . . . . . . . . . . 382
19.3 Example of PyDMD application to a toy problem. . . . . . . . . . . . . . . . 383
19.4 Sketch of the techniques implemented in ATHENA for parameter space reduc-
tion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384
19.5 Example of PyGeM deformation on an automotive test case. . . . . . . . . . . 385
19.6 Example of blade deformation using BladeX. . . . . . . . . . . . . . . . . . . 385

20.1 Schematic structure of an ANN. . . . . . . . . . . . . . . . . . . . . . . . . . 390

20.2 The forward pass on the left calculates y as a function f (x1 , x2 ) using the
input variables x1 and x2 . The right side of the figure shows the backward
pass. Receiving dL/dy, the gradient of the loss function with respect to y
from above, the gradients of x1 and x2 on the loss function can be calculated
by applying the chain rule, as shown in the figure. . . . . . . . . . . . . . . . . 392
20.3 Schematic structure of an FNN. . . . . . . . . . . . . . . . . . . . . . . . . . 393
20.4 Schematic structure of an RNN. . . . . . . . . . . . . . . . . . . . . . . . . . 394
20.5 Schematic structure of a CNN. . . . . . . . . . . . . . . . . . . . . . . . . . . 394
20.6 Illustration of a single convolutional layer. . . . . . . . . . . . . . . . . . . . . 395
20.7 Illustration of a pooling layer. . . . . . . . . . . . . . . . . . . . . . . . . . . 396
20.8 Schematic structure of an AE. . . . . . . . . . . . . . . . . . . . . . . . . . . 396
20.9 Example of POD enhancement thanks to the shifting of the snapshots. A Gaus-
sian impulse advancing in time is a simple advection problem which shows
the limitations of POD in these contexts: the singular values obtained by the
decomposition of the snapshots are not able to individuate any reduced di-
mension (top). By shifting the snapshots, POD is able to detect the dimension
needed to linearly represent the solution manifold, as highlighted by the single
nonzero singular value (bottom). . . . . . . . . . . . . . . . . . . . . . . . . . 400
20.10 Example of NNsPOD applied to an advection problem in a two-dimensional
domain. The green lines represent the contour levels of the field, the blue ones
refer to one of the snapshots of the database, and the red ones refer to the same
snapshot after the shift. The plots at incremental epochs show that the network
is able to self-learn the optimal transformation. . . . . . . . . . . . . . . . . . 400
20.11 Example of a PINN solution for a Poisson problem with parametric forcing
term. The two rows show the field of interest for two different parametric
configurations, while in the columns the analytical solution, the PINN approx-
imation, and the absolute error are sketched (from left to right). . . . . . . . . 402
20.12 Graphical representation of the reduction method proposed for an ANN. . . . . 404
20.13 Graphical representation of the VGG-16 architecture. . . . . . . . . . . . . . . 407
20.14 Example of a structural model in the naval engineering field. On the left is the
entire hull, while on the right is a sectional view. . . . . . . . . . . . . . . . . 412
20.15 A modern cruise ship example of displacement field for the hogging loading
condition. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 412
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

List of Tables

2.1 Approximation accuracy for the ROM with respect to the number of suprem-
izers exploited. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

3.1 Computational time for FE solution and RB online phase, with the speedup
and the relative error. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.2 Computational time for FE and RB solutions, with the speedup and the error,
for problem (3.30): Ra ∈ [103 , 105 ] (top) and Ra ∈ [105 , 106 ] (bottom) . . . . 78
3.3 Computational time for FE and RB solutions, with the speedup and the error,
for the Boussinesq VMS-Smagorinsky model with µg ∈ [0.5, 2]. . . . . . . . . 81
3.4 Computational time for FE and RB solutions, with the speedup and the error,
for the Boussinesq VMS-Smagorinsky model with µ ∈ [103 , 104 ] × [0.5, 2]. . 81

6.1 Computational time for FE solution and RB online phase, with the speedup
and the error. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139

7.1 The table contains the cumulative eigenvalues for the lid-driven cavity test.
The first, second, and third columns report the cumulative eigenvalues for the
velocity, pressure, and supremizer fields, respectively. The last column con-
tains the value of the inf-sup constant, in the supremizer stabilization case, for
different numbers of supremizer modes and with a fixed number of velocity
and pressure modes (10 modes for velocity and 10 modes for pressure). . . . . 158
7.2 The table contains the cumulative eigenvalues for the cylinder problem. In
the first, second, and third columns are reported the cumulative eigenvalues
for the velocity, pressure, and supremizer fields, respectively, as a function of
the number of modes. In the last column is reported the value of the inf-sup
constant for the supremizer stabilization case for different numbers of suprem-
izer modes with a fixed number of velocity and pressure modes (15 modes for
velocity and 10 modes for pressure). . . . . . . . . . . . . . . . . . . . . . . . 161
7.3 The table contains the computational time for the supremizer (SUP) and the
PPE stabilization techniques. In the cavity experiment, the SUP-ROM is ob-
tained with 10 modes for velocity, pressure, and supremizers, while the PPE-
ROM is obtained with 10 modes for pressure and supremizers. In the cylinder
experiment, the SUP-ROM is obtained with 15 modes for velocity, 10 for pres-
sure, and 12 for supremizers, while the PPE-ROM is obtained with 15 modes
for velocity and 10 for pressure. . . . . . . . . . . . . . . . . . . . . . . . . . 164

8.1 Offline parameter samples and the corresponding snapshot data. . . . . . . . . 190
8.2 Summary of the accuracy and the efficiency results for the ROMs considered
in the problem of flow around a cylinder. . . . . . . . . . . . . . . . . . . . . 200

xxxiii
 xxxiv List of Tables

12.1 Tables for the selective stabilization approach. Given a certain threshold ad-
vection coefficient ε, we obtain the mean error by computing only a percentage
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

of all the stabilization terms. . . . . . . . . . . . . . . . . . . . . . . . . . . . 261

13.1 Heat exchange problem: relative error results. . . . . . . . . . . . . . . . . . . 270

13.2 Heat exchange problem: execution time, savings, and speedup. . . . . . . . . . 271
13.3 Stokes (SBM): relative errors between the full-order solution and the RB so-
lution, one-dimensional geometrical parametrization. . . . . . . . . . . . . . . 273
13.4 Execution time, at the reduced order level, for the case with one-dimensional
geometrical parametrization. . . . . . . . . . . . . . . . . . . . . . . . . . . . 274
13.5 Stokes SBM (µ0 , µ1 ) geometry: Supremizer basis enrichment and the relative
error. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 276

14.1 Physical and geometrical constants and parameters for the geometrically and
physically parametrized leaflets test case. . . . . . . . . . . . . . . . . . . . . 296
14.2 Problem data for the test case of a flow in a channel with deformable walls. . . 303

15.1 Energy retained by the first mode. . . . . . . . . . . . . . . . . . . . . . . . . 320

15.2 Values of Young’s modulus and Poisson’s ratio used for the test cases. . . . . . 323

16.1 Parameter ranges for the Ebola model. Data taken from [167]. . . . . . . . . . 339
16.2 Comparison of the R2 scores of the response surfaces for the Ebola model
built with the AS, KAS, LAS, and NLL methods. . . . . . . . . . . . . . . . . 340

18.1 L2 -norm relative errors for pressure p, WSS, and velocity components ux , uy ,
and uz for P F = 3.45 l/min. . . . . . . . . . . . . . . . . . . . . . . . . . . 376
18.2 L2 -norm relative errors for pressure p, WSS, and velocity components ux , uy ,
and uz for P F = 4.35 l/min. . . . . . . . . . . . . . . . . . . . . . . . . . . 376

19.1 Summary table for all the packages. . . . . . . . . . . . . . . . . . . . . . . . 380

20.1 Results obtained for the reduced net POD+FNN (7) trained on CIFAR10 with
different structures for the FNN. . . . . . . . . . . . . . . . . . . . . . . . . . 408
20.2 Results obtained with CIFAR10 dataset. . . . . . . . . . . . . . . . . . . . . . 409
20.3 Results obtained with a custom dataset. . . . . . . . . . . . . . . . . . . . . . 409
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

List of Algorithms

Algorithm 2.1 The prototype greedy algorithm . . . . . . . . . . . . . . . . . . . . 31

Algorithm 2.2 The actual greedy algorithm . . . . . . . . . . . . . . . . . . . . . . 31
Algorithm 2.3 The Gram–Schmidt orthonormalization . . . . . . . . . . . . . . . . 32
Algorithm 2.4 Greedy algorithm for Stokes . . . . . . . . . . . . . . . . . . . . . . 38
Algorithm 2.5 POD algorithm for Stokes . . . . . . . . . . . . . . . . . . . . . . . 38
Algorithm 2.6 Greedy algorithm with supremizer enrichment for Stokes . . . . . . . 40
Algorithm 2.7 POD algorithm with supremizer enrichment for Stokes . . . . . . . . 40
Algorithm 2.8 POD-greedy algorithm for parabolic problems . . . . . . . . . . . . . 42
Algorithm 2.9 Greedy selection for CJ . . . . . . . . . . . . . . . . . . . . . . . . . 51
Algorithm 2.10 EIM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
Algorithm 5.1 A pseudocode for the branchwise procedure . . . . . . . . . . . . . . 105
Algorithm 7.1 The SIMPLE algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 149
Algorithm 7.2 The PISO algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 150
Algorithm 7.3 The reduced order SIMPLE algorithm . . . . . . . . . . . . . . . . . 153
Algorithm 12.1 (Deterministic) greedy algorithm . . . . . . . . . . . . . . . . . . . . 253
Algorithm 12.2 Weighted greedy algorithm . . . . . . . . . . . . . . . . . . . . . . . 254
Algorithm 12.3 POD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
Algorithm 12.4 wPOD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
Algorithm 15.1 Branchwise reduced order reconstruction . . . . . . . . . . . . . . . 317

xxxv
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Preface

Reduced order modeling is an important and fast-growing research field in computational sci-
ence and engineering, motivated by several reasons, of which we mention just a few: parametric
computing, repetitive computational environments, real-time computing, increasing complexity
of scenarios with uncertainties, and better computational performance in optimization, control,
and inverse problems. This book deals with recent developments in model order reduction in
fluid dynamics, including challenging topics such as turbulence, optimal control, flow stability,
aerodynamics, shape optimization, inverse problems, multiphysics, and uncertainty quantifica-
tion. We present here an integrated, reduced, parametric computational framework characterized
by a wide portability to be embedded in the most modern computational pipelines, including
automatic learning and digital twin developments.
After this preface and an overview and motivation in Chapter 1, the first part is concerned
with finite element–based reduced order modeling with a focus on laminar computational fluid
dynamics (Chapter 2), then with the introduction of a simple turbulent pattern (Chapter 3), and
then with the optimal flow control framework (Chapter 4). In Chapter 5, bifurcation problems
are studied, while in Chapter 6 the focus is on transport-dominated problems.
The second part of the book deals with finite volume and spectral element methods and
discontinuous Galerkin-based reduced order modeling. Chapters 7 and 8 deal with finite volume–
based reduced order modeling in computational fluid dynamics (CFD) from laminar to turbulent
flows. In Chapter 9 we deal with nonintrusive data-driven reduced order models. Chapter 10
introduces spectral element method–based reduced order modeling, while Chapter 11 deals with
discontinuous Galerkin–based reduced order modeling.
The third part deals with advanced reduced order modeling in CFD. Chapter 12 deals with
weighted reduced order modeling for uncertainty quantification, Chapter 13 with model order
reduction for embedded methods and level-set geometries, Chapter 14 with multiphysics prob-
lems (fluid-structure interaction), and Chapter 15 with bifurcations in parametric multiphysics
settings.
The fourth part is concerned with perspectives and applications. In Chapter 16 we deal with
reduction in parameter space, and Chapter 17 deals with geometrical parametrization and appli-
cations. Hemodynamics applications are introduced as examples in Chapter 18. In Chapter 19
we introduce our scientific computing open-source libraries and Python tools. Last, but not least,
Chapter 20 provides perspectives and current preliminary development to improve model reduc-
tion by automatic learning, concluding with the digital twin concept.
We accompany this book with our open-source software collection and worked problems,
available at mathlab.sissa.it/cse-software.
We acknowledge all the chapter contributors (SISSA mathLab current members and past
members) as listed in the frontmatter and at the beginning of each chapter. We acknowledge our
long-lasting national and international research collaborations, without which this work would
not exist.

xxxvii
 xxxviii Preface

We would like to acknowledge the support provided by SIAM, in particular by Elizabeth

Greenspan, as well as by the book reviewers.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

We also acknowledge our several industrial partners for inspiration, trust, and challenging
ideas, since they motivated these research developments. We acknowledge Fincantieri, Danieli,
Electrolux Professional, MICAD, Monte Carlo Yachts, CETENA, ie-fluids, Engys, Optimad, and
Bormioli.
The book is based on preliminary lecture notes and slides used at the first Summer School on
Advanced Reduced Order Methods and Applications in CFD (AROMA-CFD), held in Trieste at
SISSA in July 2019; see https://indico.sissa.it/event/34/.
This project has been carried out during lockdowns and the pandemic spread of COVID-19,
and it has represented for all of us a good way to remain optimistic about the future. This is the
reason the project’s “secret” name was “Decameron,” based on Boccaccio’s manuscript.
We acknowledge the strong support provided by the European Research Council Executive
Agency through the Consolidator Grant AROMA-CFD (grant 681447), as well as FARE-X-
AROMA-CFD project and PRIN project NA-FROM-PDEs by the Italian Ministry for Univer-
sities, INdAM GNCS national group, H2020 MSCA EID ROMSOC (grant 765374), H2020
MSCA RISE ARIA (grant 872442), and regional research programs supported by the European
Social Fund (FSE FVG) and POR-FESR Friuli Venezia-Giulia (Regional Development European
Funds).
We hope that you take some inspiration from this book and that this work will be known as
the AROMA book.
Trieste, Italy, 20 April 2022 Gianluigi Rozza, Giovanni Stabile, Francesco Ballarin
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Chapter 1

Overview and Motivation

Martin W. Hess, Marco Tezzele, Gianluigi Rozza

1.1 Reduced Order Modeling

Reduced order modeling (ROM) is a quickly emerging field in applied mathematics and compu-
tational science and engineering. It fulfills a growing demand for efficient computational tools for
many-query and real-time computations. Parametric operator formulations allow the simulation
of complex systems even in multiphysics, for coupled and uncertain scenarios. Of course, ROM
can only truly achieve this task in conjunction with other techniques from high-performance
computing (HPC), adaptive discretization methods, and various other scientific and technical
disciplines.
The ROM methods presented here focus on different fields in computational fluid dynam-
ics (CFD), in particular aeronautical, mechanical, naval, offshore, wind, sport, biomedical, and
cardiovascular applications. More demanding applications in industrial, medical, and applied
sciences require overcoming the current limitations for parametric ROM in CFD. Methodologi-
cal advances in numerical analysis and mathematical modeling are crucial to understanding and
verifying the quality of ROM simulations.
The focus is on ROM methods, which are derived from a numerical analysis approach as in
[281] and usually referred to as reduced basis (RB) methods. There is also a system-theoretic
approach to model reduction, using state-space models and a transfer function as a starting point
for the model reduction. An introduction to system-theoretic methods is given in [53] and [55].
Both approaches are considered in [257].
This motivating chapter introduces some terminology to provide the reader with an overview
and put the following chapters into perspective. It provides a first glance at intrusive and non-
intrusive model reduction, introduces a few basic concepts common to many model reduction
methods, and takes a preliminary look at applications of interest.

1.1.1 Intrusive Approaches

An intrusive model reduction approach requires knowledge of the underlying equations and a
handle on the internal data structures of the solver. A common technique is the Galerkin projec-

1
 2 Chapter 1. Overview and Motivation

tion or Petrov–Galerkin projection. Assume a parametrized and discretized system

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

A(µ)x = b, (1.1)
with matrix A ∈ RN ×N , solution vector x ∈ RN , and load vector b ∈ RN . The system
dimension N is typically so large that it is prohibitive to solve the system for many parameter
values µ ∈ D. The parameter domain D depends on the application. It can denote a range of
physical parameters, different geometry configurations, realizations of statistical quantities, or
any combination, which can be posed in a rigorous mathematical form.
In order to solve for many parameter values, the system is projected onto a low-dimensional
ROM space of dimension N . Given a trial basis V ∈ RN ×N , which serves as a projection
matrix, the Galerkin projection of (1.1) is
V T A(µ)V xN = V T b. (1.2)
The resulting solution will be of dimension N , but it can be reprojected to the full-order
space as x = V xN . A Petrov–Galerkin projection uses a test space basis W ∈ RN ×N that is
distinct from the trial space V :
W T A(µ)V xN = W T b. (1.3)
The following sections and chapters will explain much more clearly how to obtain the basis
spaces and how to apply them in different settings.

1.1.2 Nonintrusive Approaches

Nonintrusive model reduction and data-driven approaches typically use only the solution vectors
at different parameter values or even only derived quantities in the ROM method. Since the terms
intrusive and nonintrusive are not clearly defined, the distinction is fluid.
For an example of a nonintrusive approach, assume that a few high-order solutions {x1 , . . . ,
x } have been computed and collected in a snapshot matrix S ∈ RN ×L . The best-approximating
L

N -dimensional subspace to S can be computed by the singular value decomposition (SVD) or

proper orthogonal decomposition (POD), which essentially compute the same quantities. The
name principal component analysis (PCA) is also used. Denote by U ∈ RN ×N the dominant
POD modes, which span the best-approximating N -dimensional subspace. The high-order solu-
tions are then projected with U as
xiN = U T xi ∀i = 1, . . . , L. (1.4)
Define a mapping G : D → RN : µ 7→ xN which maps the parameter vector to the
POD coefficients. There are L known points of G, and any interpolation technique can be used
to evaluate it at other parameter points. The full-order solution can then be approximated as
x = U xN .

1.2 Basic Principles

This section introduces some basic concepts which are common to many ROM methods and
applications in an abstract setting. The following chapters will show how these concepts are
applied explicitly.

1.2.1 Offline-Online Decomposition

Offline-online decomposition allows us to separate the computational load into two stages. The
offline stage requires a big computational investment but only needs to be run once. A typical
 1.2. Basic Principles 3

scenario would be to run the offline stage on an HPC cluster or overnight on a workstation. The
result of the offline stage is the reduced order model (also abbreviated as ROM). The reduced
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

model can be a projected system or an approximated response surface, or it can have any other
form, depending on which ROM method was used.
The reduced model can be evaluated with low computational effort and typically requires
low storage. It can thus be used by embedded systems, portable systems, or even phones. The
evaluation of the ROM for parameters of interest is called the online stage. The online stage
refers to any evaluation of the ROM, whether it is a parameter sweep directly after completion
of the offline stage or whether it is done at a later stage on an embedded device for particular
parameters of interest.
The offline-online decomposition was crucial for the success of RB methods. Early work
[199, 198, 459, 421, 424, 423, 422, 460, 368, 41] could not achieve large speedup factors due to
a dependence on the large-scale dimension in the computations.

1.2.2 Affine Parameter Dependency

An affine function is a linear function plus a translation. That means an affine function f : V →
W with vector spaces V and W of dimension m and n respectively is affine if there exists a
matrix A of dimension n × m such that

f (v) = Av + b (1.5)

for some vector b of dimension n.

An affine parameter dependency for a parameter µ is given when a parametrized function
allows a decomposition into parameter-dependent scalar coefficient functions Θi and parameter-
independent parts fi such that
Q
X
f (µ) = Θi (µ)fi . (1.6)
i=1

The parameter must not be present explicitly, so Θi = 1 is allowed. See [505] for some explicit
examples.
The affine parameter dependency is often important for the offline-online decomposition.
When the fi in (1.6) can be precomputed approximately in the offline phase and stored as low-
order quantities, then the ROM can solve a reduced system which has the same affine parameter
dependency.
The actual computation of the affine dependency can be intricate, e.g., in the case of ge-
ometric parameter dependency. Some software packages invoke symbolic computations in the
parameter to find a minimal affine expansion. An overview of current RB and more general ROM
software can be found in [258] or on the website www.modelreduction.org.

1.2.3 POD and the Greedy Algorithm

This section introduces a frequently used algorithm based on POD and a greedy procedure to
compute ROMs for time-dependent and parametrized partial differential equations (PDEs).
Although the POD has been mentioned in the context of nonintrusive approaches (see Sub-
section 1.1.2), it is introduced here as an intrusive approach.
Assume a discretized and parametrized system

A(µ)x = b, (1.7)

with matrix A ∈ RN ×N , solution vector x ∈ RN , load vector b ∈ RN , and parameter vector µ.

 4 Chapter 1. Overview and Motivation

Assume also an affine parameter dependency

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Q
!
X
A(µ)x = Θi (µ)Ai x = b. (1.8)
i=1

If the parameter dependency is nonaffine, an affine representation can be approximated using the
empirical interpolation method (EIM); see [245, 502].

POD
Collect a few snapshot solutions to (1.7) for a chosen set of parameter vectors. The chosen set
of parameters is often taken to be uniformly distributed or uses a Latin hypercube sampling. The
number of snapshot solutions actually requires an informed decision and needs some understand-
ing of how many snapshots are needed to cover all features of solutions in the parameter domain.
In the case of strongly varying solutions in part of the parameter domain, the snapshot sampling
can also happen nonuniformly.
The snapshot solution vectors {x(µ1 ), . . . , x(µL )} are collected in a snapshot matrix S ∈
R N ×L
by simply arranging them columnwise. The SVD of S computes the matrix factorization
S = U ΣW T , (1.9)
where U is an N × N orthogonal matrix, Σ is a diagonal matrix with nonnegative entries, and
V is an L × L orthogonal matrix. It is assumed that the SVD is computed such that the entries
of Σ are sorted in descending order. The diagonal entries of Σ are the singular values of S.
The columns of U are called the left singular vectors and the columns of W are called the right
singular vectors. The Schmidt–Eckart–Young–Mirsky theorem [178, 529, 400] guarantees that
the N -dimensional, best-approximating subspace to S is given by the first N left singular vectors.
See [556, 236] for a textbook reference. Thus, define a projection matrix V ∈ RN ×N as the first
N columns of U .
Computationally, it is often beneficial to compute the first N left singular vectors with the
method of snapshots. Here, the correlation matrix M = S T S is computed and the eigendecom-
position of M is given with M zi = λi zi for 1 ≤ i ≤ N . The ith left singular vector is then
−1
given by ui = λi 2 Szi .
The reduced dimension N is often chosen as a percentage of the POD energy. The total POD
energy is given by the sum of all singular values. The value of N is then computed as the lowest
number of singular values whose sum reaches a prescribed percentage of the total POD energy.
Given V , the system (1.8) can be projected onto the low-order space as
Q
!
X
V T A(µ)V xr = Θi (µ)V T Ai V xr = V T b, (1.10)
i=1

which solves for a reduced order solution xr . The approximate full-order solution is then given
by x ≈ V xr .
During the offline stage, the quantities Ari = V T Ai V and br = V T b are computed such that
the ROM is given by
Q
!
X
Θi (µ)Ari xr = br . (1.11)
i=1

Greedy Algorithm
The greedy algorithm or greedy sampling was introduced for RB methods in [595]. Like the
POD, it also computes a projection space V , which is applied as shown in (1.10) and (1.11).
 1.2. Basic Principles 5

The parameter domain is replaced with a discrete sample set Ξ, which can be much more dense
than the sample set for the POD, since it is not necessary to compute the snapshot solution at
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

every parameter location. Instead, an a posteriori error indicator or error estimator is evaluated
over the discrete sample set and the snapshot solution is only computed where the current error
is assumed to be at its maximum. This snapshot is then added to the projection space, and the
procedure is iterated until the maximum estimator falls below a prescribed threshold or the ratio
of initially estimated error to current error falls below a prescribed threshold.
The most common and easy-to-evaluate a posteriori error indicator is the residual. The resid-
ual r(µ) is given by (dropping the µ-dependency for more clear notation)
r = b − AV xr . (1.12)
Since only the norm of the residual is required in the greedy algorithm, this computation can be
recast into a form which is independent of N as
krk2 = rT r (1.13)
T
= (b − AV xr ) (b − AV xr ) (1.14)
T T
= b b − 2b AV xr + xTr V T AT AV xr . (1.15)
Using the affine parameter dependency, the terms occurring in (1.15) can be largely precomputed,
except for the ROM solution xr ; see [281], [505].
Either the greedy iteration can be initialized with an empty V , in which case the snapshot
solution will be added where b = b(µ) attains its maximum norm, or an initial snapshot location
is prescribed. The next iterate is found by finding
µmax = arg max ∆(µ) = arg max kr(µ)k2 . (1.16)
µ∈Ξ µ∈Ξ

The optimality of the greedy approximation can sometimes be established; see [81, 60].
There are more elaborate error indicators and even error estimators which allow one to rig-
orously bound the error. These methods rely on estimates of the stability constant, i.e., the coer-
civity or inf-sup constant. A common technique is the successive constraint method; see [301],
where the method was introduced, and for an extension [120, 121, 591, 278]; interpolation tech-
niques for the stability constant are investigated in [271, 388, 302]. More general time-dependent
and inf-sup stable problems are treated, e.g., in [594, 164, 631, 387, 107].

POD-Greedy
The POD-Greedy method is the standard approach to treating time-dependent problems [180,
260, 346]. The time trajectory is treated with the POD method, while parameters in parameter
space are identified with the greedy algorithm. Theoretical considerations on the convergence
rates of the method are given in [256].
The POD-greedy algorithm is initialized with a basis VN , e.g., computed from the POD of
the trajectory at an initially predefined parameter location of dimension N . An error estimator
or error indicator is then evaluated over a fine grid of the parameter domain to determine the
currently least well approximated parameter location as defined by the ROM associated with
the projection space VN . This is an analogous procedure to the greedy algorithm for stationary
problems, except that the error estimators are more technical; see, e.g., [260].
The trajectory at the greedy-determined parameter location is computed, and its orthogonal
complement with respect to the current projection space VN is determined. The POD of the com-
plement is computed, and the dominant modes are added to the projection space, thus updating
VN and the dimension N . This procedure is continued until the maximum error in the greedy
procedure falls below some predefined tolerance threshold.
 6 Chapter 1. Overview and Motivation

1.2.4 Stabilization
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

There are many different concepts of stability in mathematics. This section considers mainly
solvability at a parameter µ. This means that the system matrix A(µ) is invertible, which means
that it has no zero eigenvalues. In particular, it is important that the projected system matrix be
stable when the large-scale system matrix is.

Coercive Problems
Coercive problems carry the stability of the large-scale model over to the ROM. Examples of
coercive problems are heat conduction and mass convection; see [474] for elliptic and parabolic
parametric PDEs. A typical PDE of interest is the Poisson equation, where the system matrix A
assembles to the discrete variational form of the Laplace operator.
In particular, assume there exists a parameter-dependent bilinear form a(·, ·; µ) and a linear
form f (·; µ) defined over a function space X associated with a domain Ω, such that the solution
u(µ) satisfies
a(u(µ), v; µ) = f (v; µ) ∀v ∈ X(Ω). (1.17)
When the function space X is a discrete space, then this is an instance of the parametrized
equation A(µ)x = b, as introduced in (1.1).
The bilinear form is coercive over X and for all parameters µ if

a(w, w; µ)
∃α0 > 0 : α(µ) = inf ≥ α0 ∀µ. (1.18)
w∈X kwk2X

The property (1.18) translates on the matrix level as a positive definite system matrix A(µ)
whose smallest eigenvalue is bounded away from zero for all µ. In particular, this implies that
A(µ) is invertible and there exists a unique solution.
To cover negative definite models, the definition is sometimes altered as

|a(w, w; µ)|
∃α0 > 0 : α(µ) = inf ≥ α0 ∀µ (1.19)
w∈X kwk2X

by taking the modulus of the bilinear form.

When the PDE is projected onto the low-order space VN , then this corresponds to a restriction
of X to the space spanned by the modes or functions of VN . The coercivity constant of the
reduced system αN (µ) fulfills

|a(w, w; µ)| |a(w, w; µ)|

α(µ) = inf ≤ inf = αN (µ), (1.20)
w∈X kwk2X w∈VN kwk2X

since VN ⊂ X. This implies that the coercivity constant of the reduced system is also bounded
away from zero and the ROM is thus considered stable.

Inf-Sup Stable Problems

Inf-sup stability is more general than coercivity, in the sense that a coercive model is also inf-sup
stable but not vice versa. Examples of inf-sup stable problems are CFD, e.g., the Navier–Stokes
equations, and electromagnetics, e.g., Maxwell’s equations.
The bilinear form is inf-sup stable over X and for all parameters µ if

a(w, v; µ)
∃β0 > 0 : β(µ) = inf sup ≥ β0 ∀µ, (1.21)
w∈X v∈X kwkX kvkX
 1.2. Basic Principles 7

where β(µ) is called the inf-sup stability constant at µ. At the matrix level, inf-sup stability
corresponds to an invertible, indefinite system matrix A(µ) and the value of β(µ) corresponds to
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

the smallest-magnitude eigenvalue.

When projecting onto the reduced space VN , an analogous argument to the coercive case
(1.20) does not hold and the inf-sup constant of the ROM can be smaller or even zero, in which
case the ROM would be unstable and approximations close to the instability would be highly
inaccurate.
Following [595] or [271], it is possible to define a particular Petrov–Galerkin projection that
guarantees stability. Define the operator T µ : X → X through the relation

(T µ w, v)X = a(w, v; µ) ∀v ∈ X. (1.22)

It holds that
a(w, v; µ)
T µ w = arg sup , (1.23)
v∈X kvkX
and T µ is called the “supremizing” operator. The inf-sup constant can then be expressed as
kT µ wkX
β(µ) = inf . (1.24)
w∈X kwkX
The restriction of X to VN will then upperbound the large-scale inf-sup constant as
kT µ wkX kT µ wkX
inf ≤ inf . (1.25)
w∈X kwkX w∈VN kwkX

By choosing a Petrov–Galerkin projection with the test space VN and trial space T µ VN , the
ROM is stable, with the inf-sup constant of the reduced system βN (µ) given as
kT µ wkX
βN (µ) = inf . (1.26)
w∈VN kwkX
The stabilizing Petrov–Galerkin projection solves the normal equations known from the
residual minimization approach minx∈VN kb − Axk and linear least squares in the projection
space VN since the linear system being solved is

VNT AT X −1 AVN x = VNT AT X −1 b, (1.27)

with X also denoting the inner product matrix associated with the space X for notational conve-
nience.

CFD
A particular issue when computing stable ROMs in fluid dynamics arises from different stability
concepts usually called the Brezzi inf-sup condition and the Babuška inf-sup condition. Both
concepts are introduced in [360], while here we will only discuss the Brezzi inf-sup condition
since it is typically used to define stable ROMs.
Viscous flows are described by the incompressible Navier–Stokes equations, which are typ-
ically written in terms of the velocity u : Ω × [0, T ) → Rd and pressure p : Ω × [0, T ) → R
as
∂u 1
+ (u · ∇) u + ∇p − ∆u = 0, Ω × (0, T ), (1.28)
∂t Re
∇ · u = 0, Ω × (0, T ), (1.29)
u(x, 0) = u0 (x), (1.30)
 8 Chapter 1. Overview and Motivation

where Ω ⊂ Rd denotes the physical domain, Re denotes the Reynolds number, and appropriate
boundary conditions are supposed to be provided; see [227].
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Assume that there exists a steady-state solution such that ∂u

∂t vanishes. Then, the parametrized
variational form reads as follows: find (u, p) = (u(µ), p(µ)) ∈ V × Q such that

a(u, w; µ) + b(p, w; µ) + c(u, u, w; µ) = F (w; µ) ∀w ∈ V, (1.31)

b(q, u; µ) = G(q; µ) ∀q ∈ Q, (1.32)

with linear forms F (·; µ) and G(·; µ) as well as parametrized bilinear and trilinear forms

∂v ∂w
Z
a(v, w; µ) = νij (µ) dΩ, (1.33)
Ω ∂xi ∂xi
∂wj
Z
b(q, w; µ) = − qχij (µ) dΩ, (1.34)
∂xi
Z Ω
∂wm
c(v, w, z; µ) = vi χij (µ) zm dΩ, (1.35)
Ω ∂xj

where the parametrized tensors ν(µ) and χ(µ) might depend on the parameters as well as spatial
coordinates. Einstein summation over repeated indices is used.
Given discrete spaces Vh and Qh to V and Q, define the full-order (Brezzi) inf-sup constant

b(q, v; µ)
βh (µ) = inf sup . (1.36)
q∈Qh v∈Vh kqkQ kvkV

The condition βh (µ) > 0 ensures stability. Note that this only takes the bilinear form b(·, ·; µ)
into account and not the Navier–Stokes operator. To have a stable ansatz space in the sense that
βh (µ) > 0, Taylor–Hood elements can be used, e.g., [74], but there are also other options [468].
Given reduced order spaces VN ⊂ Vh and QN ⊂ Qh , the reduced order Brezzi inf-sup
constant is
b(q, v; µ)
βN (µ) = inf sup , (1.37)
q∈QN v∈VN kqkQ kvkV

where βN (µ) > 0 is a necessary and sufficient condition for ROM stability.
It does not hold that βh (µ) > 0 implies βN (µ) > 0. Instead, the velocity space is enriched in
a particular way so that stability carries over to the ROM. Define for each pressure basis function
Φj ∈ QN
b(Φj , v; µ)
T µ Φj = arg sup , (1.38)
v∈Vh kvkV

which is computed by solving

(T µ Φj , v)V = b(Φj , v; µ) ∀v ∈ Vh . (1.39)

The supremizers then enrich the velocity RB space as

VN∗ = VN ⊕ span (T µ Φj : j = 1, . . . , N ) , (1.40)

 1.3. More Advanced Principles 9

such that
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

b(q, v; µ)
0 < βh (µ) = inf sup (1.41)
q∈Qh v∈Vh kqkQ kvkV
b(q, v; µ)
≤ inf sup (1.42)
q∈QN v∈Vh kqkQ kvkV
b(q, T µ q; µ)
= inf (1.43)
q∈QN kqkQ kT µ qkV

b(q, v; µ)
≤ inf sup (1.44)
q∈QN v∈V ∗ kqkQ kvkV
N
∗
= βN (µ). (1.45)

With this velocity space enrichment, the stability of the large-scale system carries over to the
ROM.

1.3 More Advanced Principles

This section introduces some advanced topics which will be further discussed in the last parts of
this book.

1.3.1 Parameter Space Reduction

Thanks to modern high-performance computational facilities and to advances in mathematical
modeling, nowadays, we are able to simulate complex phenomena described by discretized para-
metric PDEs with a huge number of degrees of freedom. This has also led to an increase in the
number of input parameters, especially in outer-loop many-query applications such as optimiza-
tion tasks, e.g. To fight the curse of dimensionality resulting from high-dimensional parameter
spaces, gradient-based reduction techniques have proved very useful in the last decade, especially
in the broader context of model order reduction [134, 503].
In the following we briefly present some linear and nonlinear methods in the framework of
ridge recovery. A generalized ridge function is any real-valued function F : Rn → R such that

F (x) = f (AT x), (1.46)

where f : Rd → R and A ∈ Rn×d is a tall matrix, with 1 ≤ d < n. The aim is to estimate the
matrix A using only point queries of f or of its gradient.
For the actual implementation of parameter space reduction methods we recommend the
open-source Python package ATHENA [496].

Active Subspaces

Active subspaces (ASs) [134] are one of the most used techniques for linear reduction in input
spaces. ASs have also been used in a multifidelity setting to enhance the accuracy of Gaussian
process regression [494]. More information can also be found in Chapter 16.
To estimate the matrix A described above, ASs construct the uncentered covariance matrix
Σ of the gradients of the function of interest f with respect to the input parameters x:
Z
Σ := Eρ [∇x f ∇x f T ] = (∇x f )(∇x f )T ρ dL, (1.47)
 10 Chapter 1. Overview and Motivation

where Eρ denotes the expected value with respect to the probability density function ρ. By
decomposing the symmetric positive definite matrix Σ as W ΛW T , we can retain the first r
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

eigenvectors and store them in W1 . Then we project the inputs over the AS defined by the span
of W1 and we obtain the reduced inputs: y = W1T x.

Kernel-Based ASs
The kernel-based extension of active subspaces (KAS) [493] is a nonlinear technique which
consists of a ridge approximation of f factorized through a projection onto a reproducing kernel
Hilbert space (RKHS). This is done by introducing a feature map φ which maps the inputs into a
higher-dimensional space. Then we seek an AS in this intermediate space.
The new covariance matrix for this case is
Z
Σ := (∇φ(x) f (φ(x)))(∇φ(x) f (φ(x)))T dµ(φ(x)), (1.48)
φ(x)

where µ := φ ◦ L. As we can see, the choice of the feature map is crucial. A common choice
in the machine learning community is the following, which has the advantage that the gradients
can be computed analytically:
r
2
φ(x) = σf cos(Qx + b), (1.49)
D
where D is the number of features, σf is a hyperparameter to be tuned, Q is a matrix whose
rows are sampled from a given probability distribution, and b is a vector whose components are
sampled independently and uniformly in [0, 2π].

Local ASs
Following a completely different strategy, to achieve good linear dimensionality reduction we can
exploit domain decomposition. Local active subspaces (LASs) [495] do so by using a clustering
technique equipped with a supervised distance metric. Then on each cluster a ridge approxima-
tion is performed. Given a pair of inputs, we define the AS-induced distance as
q
kxi − xj kΛ = (xi − xj )T W Λ2 W T (xi − xj ), (1.50)

where Λ and W are the matrices obtained from the decomposition of the covariance matrix
described above. We can observe that by using this formulation we incorporate information
from the global trend given by the global AS. Effective clustering techniques are K-medoids and
hierarchical top-down. The latter is particularly versatile since it allows the selection of more
sophisticated divisions; e.g., it is possible to automatically select the optimal AS dimension on
each cluster.

1.3.2 Advanced Geometrical Morphing

Affine geometrical transformations are usually not sufficient to describe the variety and complex-
ity of possible shapes needed for a proper shape optimization study, especially in an industrial
context. More advanced geometrical morphing techniques allow for a richer design shape while
losing the affinity property exploited by intrusive ROMs. Here we briefly describe free-form
deformation (FFD) [533] and radial basis function (RBF) interpolation [82]. Other techniques
exist, such as inverse distance weighting (IDW) interpolation [537]. For a deeper description
of these methods we suggest [503] and Chapter 17. For the implementation of the geometrical
morphing techniques presented, we recommend the open-source Python package PyGeM [570].
 1.3. More Advanced Principles 11

FFD
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

FFD consists of creating a lattice of control points around the object of interest and propagating
the deformation of this lattice to the object. This is done by mapping ψ from the physical domain
Ω to a reference domain Ω̂. The FFD parameters µ ∈ P are the displacements of the control
points along the axes. The propagation map T̂ (·, µ) : Ω̂ → Ω̂(µ) is usually composed by
Bernstein polynomials, but other choices are also possible. The FFD map M is the composition
of the three different geometric deformations, i.e.,

M(·, µ) := ψ −1 ◦ T̂ ◦ ψ(·, µ), µ ∈ P. (1.51)

The FFD-induced deformation does not depend on the topology of the object to be morphed,
so it is considered very versatile and nonintrusive, especially in industrial contexts [508, 517,
571, 157, 154].

RBFs
A smooth real-valued function whose value depends only on the distance between the input
and some reference point is called a radial basis function (RBF). This class of functions can be
exploited to induce a deformation by moving some control points placed over the parametrized
object and interpolating them with a new surface.
Let {xCi }N NC
i=1 ∈ Ω be such control points, ϕ be a radial basis function, and {γi }i=1 be the
C

NC
weights associated with the RBF centered in {xCi }i=1 . We define the deformation map M for
the RBF interpolation technique as
NC
X
M(x, µ) := s(x, µ) + γi (µ)ϕ(kx − xCi k) ∀x ∈ Ω, µ ∈ P, (1.52)
i=1

where s(·, µ) is a polynomial term of degree one, and µ is the parameter vector composed of
the displacements of the RBF control points. To compute the weights γi we need to impose the
interpolatory constraints

M(xCi , µ) = yCi (µ) ∀i ∈ [1, . . . , NC ], (1.53)

where yCi are the new locations of the control points after the displacement of xCi . For the
remaining unknowns, we need to impose the conservation of the total force and momentum.
For applications of RBF interpolation we suggest among others [568, 390] for a carotid artery
and [573, 161] for a NACA airfoil profile.

1.3.3 New Deep Learning Approaches

The surge of artificial neural networks (ANNs) [237, 348] and their successful application to
real-world problems has represented a great opportunity for the computational science and en-
gineering community. Thanks to the ANN approximation’s properties, many applications are
possible, including the reconstruction of the solution manifold for POD-based nonintrusive re-
duction [447], physics-informed neural networks (PINNs) [480], and physics-reinforced neural
networks [119].
A particularly interesting application for problems with slow Kolmogorov n-width decay,
such as advection-dominated problems, is the neural network shifted proper orthogonal decom-
position (NNsPOD) algorithm [436]. It represents an extension of the shifted POD [484] that
automatically learns from the solution snapshots the optimal shifting to a reference configuration
 12 Chapter 1. Overview and Motivation

without making assumptions on the underlying model. This is done using two different neural
networks: the shifting net calculates the spatial coordinates of the shifted snapshots, while the
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

interpolation net evaluates the reference configuration in the coordinates of the shifted snapshots.
It is worth noting that along with the expansion of ANNs to solving complex problems, the
average architecture size increased as well, in terms of both number of layers and number of
neurons. To overcome this issue it is possible to apply reduced order methods, such as POD and
AS, to compress a trained neural network [394]. This leads to remarkable memory savings while
retaining a high level of accuracy. This method has been successfully applied to convolutional
nets, but its generality allows application to other types of architectures.

1.3.4 Digital Twins

The convergence of technologies generating a huge amount of data, advanced algorithms to
process them and extract key information, and the computing capabilities enabling large-scale
systems simulations, is revolutionizing every aspect of computational sciences. This has led
to the spread of digital twins [481] in many communities, integrating big data, computational
models, and decision support, in a dynamical framework evolving in time as the physical asset
mimics changes.
Within this framework ROM plays an important role thanks to the offline-online paradigm.
It enables real-time onboard evaluation of the quantities of interest and a fast and accurate up-
date of the digital twin state given the stream of data from the physical asset. Data assimilation
is thus a crucial step to guarantee optimal management of the computational resources that can
be placed onboard. It is also important to guarantee long-time consistency; i.e., the digital twin
must remain consistent with the physical assets for the entire lifespan, taking into consideration
degradation and fatigue phenomena. A mathematical foundation for digital twins using a prob-
abilistic graphical model can be found in [321]. For a view of the challenges and opportunities
for scaling digital twins, see [416].
The application of digital twins is not limited to engineering artifacts, but it extends to the
human body and to personalized medical treatments. This paradigm shift is particularly evident
for precision cancer care in oncology [269]. Chapter 20 provides a more detailed discussion of
digital twins.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Part I

Finite Element–Based ROMs

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy
 Chapter 2
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Finite Element–Based
Reduced Basis Method
in Computational Fluid
Dynamics

Federico Pichi, Maria Strazzullo, Francesco Ballarin,

Gianluigi Rozza

2.1 Introduction
The goal of this book is to give a comprehensive overview of reduced order models (ROMs)
and their applications in fluid dynamics. The main idea of ROMs is to identify a basis able to
capture the parametrized behavior of the system under investigation [281]. ROMs aim at rapidly
obtaining many solutions, corresponding to several values of parameters that can represent both
physical and geometrical features. This technique is important in several fields, from applied
sciences to industry. This chapter focuses on reduced order modeling (ROM) based on the fi-
nite element method (FEM), a classical projection-based approach to solving partial differential
equations (PDEs) [477]. In Section 2.2 we introduce the FEM for solving elliptic PDEs and
some fluid dynamics problems, dealing with both linear and nonlinear equations. Then, the
main ideas behind ROM techniques are described in Section 2.3, where we propose the most
common algorithms exploited in model reduction, i.e., proper orthogonal decomposition (POD)
[34, 85, 114, 281] and the greedy algorithm [81, 281]. The latter procedure uses error estimators
between the FEM and the ROM solutions and is the topic of Section 2.4. Finally, some advanced
techniques involving more complicated problems are treated in Section 2.5.

2.2 The FEM

This section focuses on the numerical approximation of PDEs using the FEM. We will take the
first steps to approximate the Poisson problem and, thus, we will describe a first application in
fluids given by the Stokes equations. Furthermore, this contribution includes some implementa-
tion tips from the FEniCS library [378]. FEniCS is open-source software that deals with PDEs
using the FEM. The code boxes in the chapter will help the reader reproduce all the presented
examples.

15
 16 Chapter 2. Finite Element–Based Reduced Basis Method in CFD

2.2.1 The Model Problem: The Poisson Equation

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Our description starts from a simple Poisson problem setting, which paves the way for more
complex applications. The system we introduce is a milestone in the numerical approximation
of PDEs and it represents a very useful tool for understanding basic concepts of numerical dis-
cretization, such as weak formulation, well-posedness results, discrete polynomial spaces, and
estimates of the approximation error [477].
Let us consider a domain Ω ⊂ Rd with d = 1, 2, 3, and, furthermore, let us assume that Ω
has a sufficiently regular boundary ∂Ω. The aim is to find an unknown solution u ∈ C 2 (Ω) such
that
in Ω,

−∆u = f
(2.1)
u=0 on ∂Ω,

where f = f (x) ∈ C 0 (Ω) is a given forcing term acting on the system at hand. The problem
(2.1) is characterized by homogeneous Dirichlet boundary conditions. However, all the strategies
we will present easily adapt to more complicated settings and several variants of the boundary
conditions (such as nonhomogeneous Dirichlet boundary conditions and homogeneous and non-
homogeneous Neumann boundary conditions) and more general elliptic operators.
To solve this prototypical differential problem, we analyze the weak formulation of the sys-
tem. Namely, we move from a second-order description of the operators to an integral form
of the first order of differentiation. To reach this goal, we multiply the Poisson equation by a
smooth and arbitrary function v. Thus, applying the divergence theorem, it is a matter of simple
computations to show that the problem (2.1) can be recast in the form
Z Z Z
∇u · ∇v dΩ − ∇u · vn dΓ = f vdΩ, (2.2)
Ω ∂Ω Ω

where n is the outward unit normal vector to ∂Ω. Combining this integral version of the system
and the homogeneous boundary condition, we can weaken the regularity assumption and find the
.
solution u ∈ V = H01 (Ω), considering the forcing term f ∈ L2 (Ω). Furthermore, thanks to the
assumption u = 0 ∈ ∂Ω, the boundary term of (2.2) vanishes, and by defining the bilinear form
a : V × V → R and the functional F : V → R as
Z Z
a(u, v) = ∇u · ∇v dΩ, F (v) = f v dΩ,
Ω Ω

the weak formulation (2.2) reads as follows: find the solution u ∈ V such that

a(u, v) = F (v) ∀v ∈ V. (2.3)

The well-posedness of the problem, i.e., the proof of the existence, uniqueness, and stability of
the solution, is related to the following theorem.

Theorem 2.1 (Lax–Milgram). Let V be a Hilbert space; a : V × V → R a bilinear, continuous

and coercive form; and F : V → R a linear and continuous functional. Then there exists a
unique solution to the problem (2.3). Furthermore, the solution u is bounded by the data, i.e.,

M
kukV ≤ ,
α
 2.2. The FEM 17

.
where M = kV 0 k is the continuity constant of the functional F and
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

. a(v, v)
α = inf
v∈V kvk2 V

is the coercivity constant of the bilinear form a.

In the next subsection, we will analyze the FEM as a numerical approach to dealing with the
elliptic problem (2.3) at a finite-dimensional level.

2.2.2 Galerkin Projection

To solve the variational formulation (2.3) in a finite-dimensional framework, we define a family
of finite-dimensional spaces V N ⊂ V such that dim(V N ) = N < +∞. The family V N con-
verges to the original space V as N approaches infinity. There are several strategies for building
V N . As already specified, in this section, we will focus on the FEM, a very common discretiza-
tion approach for PDEs. Let us briefly introduce the basic idea behind FE approximation.
Given a bounded polygonal domain Ω, we define the partition T N of the considered domain
in cells, called simplices, such that
• the triangulation covers the domain, i.e.,
[
Ω= K;
K∈T N

• the cells are full dimensional: K̊ 6= ∅ ∀K ∈ T N ;

• different cells do not overlap: K̊1 ∩ K̊2 = ∅ ∀K1 , K2 ∈ T N and K1 6= K2 ;
• the triangulation is conforming; namely, if K1 ∩ K2 6= ∅ ∀K1 , K2 ∈ T N and K1 6= K2 ,
then the intersection is either a whole face, an edge, or a vertex of the grid.
In this setting, we can relate the grid to the dimension of the function spaces built on it. Indeed,
if we denote by hK the diameter of a cell K and by h the maximum diameter of the whole
triangulation, then the FE space dimension N will grow as h tends to zero.
1 from dolfin import *
2 from mshr import *
3
4 # Create mesh
5 domain = Rectangle ( Point (0. , 0.) , Point (1. , 1.))
6 mesh = generate_mesh (domain , 15)
7
8 # Create boundaries
9 class Left( SubDomain ):
10 def inside (self , x, on_boundary ):
11 return on_boundary and abs(x[0]) < EPS
12
13 class Right ( SubDomain ):
14 ...
15 class Bottom ( SubDomain ):
16 ...
17 class Top( SubDomain ):
18 ...
19
20 D = mesh. topology ().dim ()
21 subdomains = MeshFunction (" size_t ", mesh , D)
 18 Chapter 2. Finite Element–Based Reduced Basis Method in CFD

22 boundaries = MeshFunction (" size_t ", mesh , D - 1)

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

23
24 Bottom ().mark( boundaries , 1)
25 Left ().mark( boundaries , 2)
26 Top ().mark( boundaries , 3)
27 Right ().mark( boundaries , 4)
28
29 dx = Measure ("dx")( subdomain_data = subdomains )
30 ds = Measure ("ds")( subdomain_data = boundaries )
Listing 2.1. Mesh generation in FEniCS.

Figure 2.1. Left: example of a triangulation over the unit square resulting from Listing 2.1.
Right: highlighted edges for the application of boundary conditions.

In Figure 2.1, we represent a conformal grid over a unit square. The triangulation is obtained
though the reported FEniCS code; see Listing 2.1. Once the physical domain is discretized, the
next step of the approximation relies on the function space definition. The FE space is
V N (r) = {v N ∈ C 0 (Ω) | v|NK ∈ Pr ∀K ∈ T N }, (2.4)
where Pr is the space of polynomials up to degree r ≥ 1. For the sake of notation, we will
drop the degree dependence from the space dimension and we will write V N instead of V N (r) .
However, it is clear that the higher r is, the higher N will be. Indeed, to represent the solution
v N over each cell of T N , its value must be known on a number of properly chosen nodes, i.e.,
specific points over the cell element. Figure 2.2 shows the required cell points to represent an FE
function with P1 and P2 elements, in both two and three dimensions. We also report the FEniCS
code that creates the related FE function spaces in Listing 2.2. The set of nodes will be denoted
by Nj for j = 1, . . . , N . A basis of the FE space V N can be set as the characteristic Lagrangian
functions φj ∈ V N for j = 1, . . . , N such that

if i 6= j,

0
φj (Ni ) = δij = (2.5)
1 otherwise.

The assumption φj ∈ V N is verified since the bases, locally, are polynomials of degree at most
r. The structure of the bases is represented in Figure 2.3: φj has a local support Sj centered in
Nj , i.e., it is nonzero only in a neighborhood of Nj .
1 V1 = FunctionSpace (mesh , "CG", 1)
2 V2 = FunctionSpace (mesh , "CG", 2)
Listing 2.2. FE function space definitions in FEniCS.
 2.2. The FEM 19

P1 P2 P1 P2
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 2.2. P1 and P2 element types for two- and three-dimensional cells in (a), (b) and (c), (d),
respectively.

φj

Nj

Figure 2.3. Basis function φj and its support highlighted in pink.

Thanks to the bases definition, a generic FE function v N ∈ V N can be expressed as the linear
combination of the {φi }N
i=1 as
XN
vN = vi φi (2.6)
i=1

for some unique real coefficients i=1 .

{vi }N In a similar way, the discrete solution uN can be
expressed as
N
X
uN = ui φi (2.7)
i=1

for some unique, yet unknown, real coefficients {uj }N . Now, imposing that the FE solution
satisfy the weak formulation for every basis function, we get the following linear system of N
equations in N unknowns:
N
X
uj a(φj , φi ) = F (φi ) ∀i = 1, . . . , N . (2.8)
j=1

The equations (2.8) can be recast in an algebraic setting. Indeed, let us define the symmetric
positive definite stiffness matrix A ∈ RN ×N as
Z Z
Aij = a(φj , φi ) = ∇φj · ∇φi dΩ = ∇φj · ∇φi dΩ ∀i, j = 1, . . . , N (2.9)
Ω Sj ∩Si

and the forcing vector f ∈ RN where fi = F (φi ). Furthermore, we indicate by u the column
vector of the unknown coefficients of the solution. Then, the equations (2.8) can be recast as the
following linear system of dimension N :

Au = f . (2.10)
 20 Chapter 2. Finite Element–Based Reduced Basis Method in CFD

We underline that the linear system (2.10) is sparse because the bases are nonzero only in the
local support. In FEniCS, the linear system (2.10) can be built and, thus, solved for different
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

parameter values (see Figure 2.4), as is done in Listings 2.3 and 2.4, respectively.
1 # Define a function space
2 V = FunctionSpace (mesh , "CG", 1)
3 u = TrialFunction (V)
4 v = TestFunction (V)
5
6 # Create stiffness matrix
7 a = inner (grad(u), grad(v))*dx
8 A = assemble (a)
9
10 # Create forcing vector
11 f = 5*v*dx
12 F = assemble (f)
Listing 2.3. Linear system definition in FEniCS.

A value which gives information about the solvability of the discrete system is the condition
number, and it is related to the Euclidean norm of the matrix A. In this context, we will always
work with the 2-norm, defined as
v
uN N
uX X
kAk2 = t |aij |2 .
i=1 j=1

In the case of a symmetric positive definite matrix, the condition number can be computed as

λmax (A)
K2 (A) = , (2.11)
λmin (A)

where λmax (A) and λmin (A) are the maximum and the minimum singular values of A, re-
spectively. Furthermore, under regularity assumptions on the mesh T N , the condition number
verifies K2 (A) ≤ Ch−2 , where C is a constant which does not depend on h yet may depend on
the polynomial degree r. We stress that the condition number deteriorates as h goes to zero.

Figure 2.4. Solution of the system (2.10), with F = sin(µ[0]x) cos(µ[1]y), for several values
of µ. Here, x and y indicate the spatial coordinates of the spatial domain. Left: µ = (2π, 2π). Center:
µ = (2π, π). Right: µ = (π, π).

1 # Create stiffness matrix

2 a = inner (grad(u), grad(v))*dx
3 A = assemble (a)
4
5 # Create forcing vector
6 mu = (pi , pi)
 2.2. The FEM 21

7 (x, y) = SpatialCoordinate (mesh)

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

8 f = sin(mu [0]*x)*cos(mu [1]*y)*v*dx

9 F = assemble (f)
10
11 # Apply boundary conditions
12 for b in (1, 2, 3, 4):
13 bc = DirichletBC (V, 0., boundaries , b)
14 bc. apply(A)
15 bc. apply(F)
16
17 # Solve the linear system
18 U = Function (V)
19 solve (A, U. vector () , F)
Listing 2.4. System assembly and solve in FEniCS.

Namely, using the FE method, we are dealing with a finite-dimensional approximation of the
continuous problem solution u. The natural question that arises concerns the error one is com-
mitting with respect to the continuous solution of the problem at hand. How far am I from u?
How is the error related to such features as the regularity of the function and the refinement of
the mesh? Answers can be found in the next theorem [477].

Theorem 2.2. Let u ∈ V be the exact solution of the weak formulation of the Poisson equation
(2.3) and uN ∈ V N its approximated solution using an FE discretization of dimension N and
degree r. If u ∈ C 0 (Ω) ∩ H p+1 (Ω) for p > 0, then the following a priori error estimate holds
for a positive constant C:
ku − uN kV ≤ Chs |u|H s+1 (Ω) , (2.12)
where s = min{r, p}.

The proof of the theorem relies on a combination of several fundamental lemmas in the
numerical analysis of the FEM for elliptic problems. As expected, the result states that the
accuracy increases polynomially as h decreases, up to the maximum regularity allowed by u ∈
C 0 (Ω) ∩ H p+1 (Ω).
In the next section, we will deal with a first example of the FEM applied to fluid dynamics.
Indeed, we will show some results and specific FE techniques to deal with the Stokes equations,
the model that describes the flow of incompressible fluids under high diffusivity action.

2.2.3 A First Example in Fluid Dynamics: Stokes Equations

In this section we deal with the FE approximation of the basic model in fluid dynamics, the
Stokes equations. These equations describe the motion of an incompressible fluid with very
small convection effect with a constant density ρ. This model can be applied to several scientific
fields, from fluid dynamics applications to natural sciences and biological flows.
The Stokes equations are

in Ω,

−ν∆u + ∇p = f
(2.13)
div(u) = 0 on ∂Ω,

where u is the vector fluid velocity, p is the scalar pressure variable divided by the density ρ,
ν = µ/ρ is the kinematic viscosity, and µ is the dynamic viscosity. Here, f is the forcing term
per unit volume. From these quantities, we can define the Reynolds number as

. u∞ L
Re = , (2.14)
ν
 22 Chapter 2. Finite Element–Based Reduced Basis Method in CFD

where L is a representative length of the spatial domain and u∞ is the velocity field. The
Reynolds number typically expresses the relation between inertial and viscous forces. Namely,
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Re is large when inertial forces are dominant. This is not the case for Stokes flows, where inertial
forces are negligible with respect to viscous forces, i.e., Re → 0.
The boundary conditions will be of either Dirichlet type, i.e., u = g on ΓD ⊂ ∂Ω, or
∂n − pn = h. Also, for these specific equations, we can write (2.13) in a
Neumann type, i.e., ν ∂u
weak form. To this purpose, we multiply the equation
−ν∆u + ∇p = f
by a smooth and arbitrary vector function v, and, exploiting the divergence theorem, we obtain
Z  
∂u
Z Z Z
ν∇u : ∇v dΩ − pdiv(v) dΩ = f · v dΩ + ν − pn · vdΓ. (2.15)
Ω Ω Ω ΓN ∂n
The same strategy can be applied to div(u) = 0 as well. The latter expression is known as
the continuity equation. Thus, exploiting the multiplication for a smooth and arbitrary scalar
function q, the continuity constraint is
Z
qdiv(u)dΩ = 0. (2.16)
Ω
To guarantee that the integral formulation is meaningful, let us consider the velocity fields in
. .
V = HΓ1D (Ω) and the scalar field in Q = L2 (Ω), where V is the space of vector functions of
H 1 (Ω) regularity which vanish over ΓD . With a similar argument with respect to the Poisson
equation, we define the bilinear forms a : V × V → R and b : V × Q → R, where
Z Z
a(u, v) = ν∇u : ∇v dΩ, b(u, q) = − qdiv(u)dΩ, (2.17)
Ω Ω
and the functional F : V → R, where
Z Z
F (v) = f · v dΩ + h · vdΓ. (2.18)
Ω ΓN

Namely, the weak formulation of the Stokes problem is to find the pair (u, p) ∈ V × Q such that

a(u, v) + b(v, p) = F (v) ∀v ∈ V,
(2.19)
b(u, q) = 0 ∀q ∈ Q.
The problem (2.19) is, then, a saddle-point problem, which is well posed thanks to the following
theorem [74].

Theorem 2.3 (Brezzi). The weak formulation of the Stokes problem admits a unique solution if
and only if
1. a : V × V → R is bilinear; coercive on the space
Vdiv = {v ∈ V : b(v, q) = 0 ∀q ∈ Q},
with coercivity constant α; and continuous with continuity constant γ;
2. b : V × Q → R is bilinear and continuous, and there exists a constant β > 0 such that
the inf-sup condition is verified, i.e.,
b(v, q)
inf sup ≥ 0. (2.20)
q∈Q v∈V kvkV kqkQ
Furthermore, the following stability estimates hold:
1 1 γ
kukV ≤ kF kV 0 , kpkQ ≤ 1+ kF kV 0 . (2.21)
α β α
 2.2. The FEM 23

However, when the system is discretized with the FEM, the problem may not satisfy the inf-
sup condition (2.21). In this case, there exists a spurious pressure mode η ∈ Q, η 6= 0, such that
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

b(v, η) = 0 for all v ∈ V . This means that, if the pair (u, p) ∈ V × Q is a solution, then the
pair (u, p + τ η) ∈ V × Q would also be a solution for any scaling factor τ ∈ R. In this case, no
stability estimate holds for the pressure field p; indeed, its norm could explode for very large τ .
However, the stability estimate for u is still valid.
To avoid this issue, the velocity and pressure spaces cannot be independently chosen; in
particular, the polynomials’ degrees must verify some features. One of the main FE approaches
used for Stokes equations relies on Taylor–Hood elements. Namely, the approximation consists
of exploiting the polynomial space pair Pr -Pr−1 with r ≥ 2 for V and Q, respectively. From now
on, we will always use this conventional choice for the FE approximations of Stokes systems.
Now we can focus on the algebraic formulation of this fluid dynamics system. Let us denote
Np
by {ϕj }Nj=1 ∈ V
u N
and {ψk }k=1 ∈ QN the Lagrangian basis functions of the spaces V and Q,
.
with N = Nu + Np , representing the global dimension of the space. Thus, the discrete solutions
can be expressed as
Nu Np
X X
N N
u = uj ϕj , p = uk ψk . (2.22)
j=1 k=1

Now, we define the matrices A ∈ R Nu ×Nu

as Aij = a(ϕj , ϕi ) and B ∈ RNu ×Np as Bij =
b(ϕm , ψk ) and the two vectors consisting of the solution coefficients U ∈ RNu and P ∈ RNp
for velocity and pressure fields, respectively. With these ingredients, the system (2.19) reads

AU + B T P = f ,

(2.23)
BP = 0,

as a result of the saddle-point algebraic structure

A BT
    
U f
= . (2.24)
B 0 P 0

In this algebraic context, the equivalent of the inf-sup condition (2.20) is ker(B T ) = {0}; when
it holds, no spurious pressure modes are present. In Figure 2.5 we present the solution of a flow
past a cylinder, a classical benchmark in fluid dynamics. We also show the FEniCS code to solve
it in Listing 2.5. In the next subsection, we will focus our attention on how to approximate a
more complicated model, the nonlinear Navier–Stokes equations.

Figure 2.5. Flow past a cylinder: solution of the system (2.23) with a parabolic inlet on the left
boundary. Velocity and pressure fields from left to right, respectively. No-slip conditions on the bottom and
top walls and on the cylinder and Neumann boundary conditions on the right outflow.

1 # Function spaces
2 V_element = VectorElement (" Lagrange ", mesh. ufl_cell () , 2)
3 Q_element = FiniteElement (" Lagrange ", mesh. ufl_cell () , 1)
4 W_element = MixedElement (V_element , Q_element ) # Taylor -Hood
5 W = FunctionSpace (mesh , W_element )
6
7 # Test and trial functions
 24 Chapter 2. Finite Element–Based Reduced Basis Method in CFD

8 (v, q) = TestFunctions (W)

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

9 (u, p) = TrialFunctions (W)

10
11 # Variational forms and boundary conditions
12 lhs = nu* inner (grad(u), grad(v))*dx - div(v)*p*dx - div(u)*q*dx
13 rhs = inner (f, v)*dx
14 bc = ...
15
16 # Solve the problem
17 up = Function (W)
18 solve (lhs == rhs , up , bc)
19 (u, p) = up. split ()
Listing 2.5. Stokes problem in FEniCS.

2.2.4 A Nonlinear Model: Navier–Stokes Equations

In this section, we describe the incompressible Navier–Stokes equations and their discretization.
This model has an additional convection term that was neglected in the Stokes case. Indeed, the
Navier–Stokes equations are

in Ω,

−ν∆u + (u · ∇)u + ∇p = f
(2.25)
div(u) = 0 on ∂Ω,

where (u · ∇)u is a nonlinear term describing the convective action of the flow. The correspond-
ing weak formulation requires the introduction of an additional trilinear form C : V ×V ×V → R
defined as Z
c(u, w, v) = [(u · ∇)w] · v dΩ. (2.26)
Ω

Thus, the weak formulation of the Navier–Stokes equations is to find the pair (u, p) ∈ V × Q
such that 
a(u, v) + c(u, u, v) + b(v, p) = F (v) ∀v ∈ V,
(2.27)
b(u, q) = 0 ∀q ∈ Q.
In Listing 2.6, we show how to deal with such a system in FEniCS, from the function space
definition to the solution of the equations. In Figure 2.6 we show some parametric solutions of
the problem at hand, varying the Reynolds number. As expected, the solution for Re = 1 is quite
similar to the Stokes one presented in Figure 2.5.
1 # Test and trial functions
2 (v, q) = TestFunction (W)
3
4 # Solution
5 up = Function (W)
6 (u, p) = split (up)
7
8 # Residual
9 F = (nu * inner (grad(u), grad(v)) * dx
10 + inner (grad(u) * u, v) * dx
11 - div(v) * p * dx
12 + div(u) * q * dx
13 - inner (f, v) * dx)
14
15 # Jacobian
16 J = derivative (F, up)
17
18 # Boundary conditions
 2.2. The FEM 25

19 bc = ...
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

20
21 # Solve the problem
22 problem = NonlinearVariationalProblem (F, up , bc , J)
23 solver = NonlinearVariationalSolver ( problem )
24 solver . solve ()
Listing 2.6. Navier–Stokes problem in FEniCS.

The existence and the uniqueness of the solution are guaranteed by the Banach contraction
theorem when
kf kV 0 ≤ C,
where the constant C depends on the continuity constant of the trilinear form and on the size and
shape of the domain Ω, represented by the constant of the Poincaré inequality. This assumption,
i.e., the small data assumption, is verified when either f is controlled in norm or the Reynolds
number Re is not large. However, this is not always the case in practice, and we may be interested
in cases lacking uniqueness [446, 452, 454]. From now on, we will always assume that the
problem at hand is well posed. In order to discretize it, one can use the same Taylor–Hood
spaces as in the Stokes case. Furthermore, to deal with the nonlinearity of the system related
to the convective trilinear term, a Newton method can be used. It results in the solution of the
following iterative scheme: given (un , pn ) ∈ V × Q, find the pair (un+1 , pn+1 ) ∈ V × Q such
that

a(un+1 , v) + c(un , un+1 , v) + c(un+1 , un , v) + b(v, pn+1 ) = F (v) ∀v ∈ V,
b(un+1 , q) = 0 ∀q ∈ Q.
(2.28)
The iterations shall be repeated until a convergence criterion is reached. In this chapter we will
not consider time dependency and time evolution. However, the extension is straightforward
and the simulations can be performed through the standard finite difference method in time,
e.g., backward Euler. In the next section, we will introduce the basic ideas behind ROMs and
the principal algorithms related to such an approximation. The different procedures will be
complemented by some numerical examples.

Figure 2.6. Flow past a cylinder: solution of the system (2.25) with a parabolic inlet on the left
boundary and velocity and pressure fields from left to right, respectively. No-slip conditions on the bottom
and top walls and on the cylinder and Neumann boundary conditions on the right outflow. The solution is
represented for several values of Re.
 26 Chapter 2. Finite Element–Based Reduced Basis Method in CFD

2.3 ROMs
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Dealing with the approximation of parametrized problems can be a critical task from the compu-
tational viewpoint. Thus, the idea is to replace the original high-dimensional problem, i.e., the
FE approximation, with a reduced one that is easy to manage. In this section, we provide a short
mathematical and practical introduction to ROMs, and in particular to the reduced basis (RB)
method, by which we can reduce the computational costs of our problems. We will present some
preliminary applications to elliptic PDEs and Stokes system test cases, which were discussed in
the previous section. We will have a look at the building blocks of the RB technique: (i) param-
eter separability and (ii) RB space generation via POD and the greedy algorithm. Moreover, we
will highlight the differences with respect to the FE method case and we will show examples of
implementation within RBniCS [38].

2.3.1 RB Approximation
As we saw in the previous section, the preliminary step for the discretization of a PDE is the pro-
jection of its weak formulation in a finite-dimensional setting, which results in a Galerkin system
with N degrees of freedom. In a parametrized many-query and real-time context, the compu-
tational burden can become unaffordable; for this reason, here we focus on the RB method.
Roughly speaking, this method consists of a projection of the high-fidelity problem onto a sub-
space of smaller dimension spanned by some properly chosen basis functions; thus it shares
many features with the Galerkin-FE method.
More precisely, the mathematical setting and the goals of this approximation are as follows.
Let µ denote parameters which affect the solution of a problem of interest (e.g., due to uncer-
tainty, optimization, or physical or geometrical features). Let P ⊂ Rp be the parameter domain;
given a tolerance  > 0, we want to build an efficient approximation of the FE solution for any
given parameter µ ∈ P. This consists of evaluating the map µ ∈ P → uN (µ) ≈ uN (µ),
possibly achieving the desired accuracy
kuN (µ) − uN (µ)kN
V ≤ ,

but more efficiently than the FEM; i.e., the RB evaluation µ → uN (µ) should be much faster
than its FE counterpart, or ∂tN
comp  ∂tcomp .
N

As we have already mentioned, two typical scenarios are as follows:

• In the real-time context, we deal with parameter estimation, often in connection with in-
dustrial/clinical applications, requiring a rapid evaluation of the solution
t0 : µ → t0 + ∂tN
comp : uN (µ).
|{z} | {z }
“need” “response”

• In the many-query context, we deal with optimization problems, or we explore the param-
eter domain, requiring multiple evaluations of the solution
tcomp (µj → uN (µj ), j = 1, . . . , J) = ∂tN
comp J as J → ∞.

We can thus introduce the notion of solution manifold in the discrete context, comprising all
solutions of the parametric problem under variation of the parameters, i.e.,
MN = {u(µ) | µ ∈ P ⊂ V N },
where each uN (µ) ∈ V N corresponds to the solution of the parametric FE problem. The key
assumption for the applicability of any reduced model is that the solution manifold in Figure 2.7
 2.3. ROMs 27
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

uN (µntrain )
N
M

uN (µ)
N 1
u (µ )

Figure 2.7. Solution manifold at the discrete level represented by the snapshots.

can be represented by the span of a low number of basis functions, i.e., N  N , with a small
error. Thus, we start obtaining some information about the behavior of the solution with respect
to the parameter, i.e., the computation of FE solutions for prescribed parameter values (the so-
called snapshots), and then we extract the RB.
As we can see in the following example, to obtain good approximation properties, a cru-
cial assumption is the smooth parametric dependence of uN (µ) on µ. We refer the reader to
Chapter 5 for complex applications concerning the lack of smoothness related to bifurcating
phenomena.

Example 2.4 (linear superposition as RB method). Consider the second-order PDE on the in-
terval I = (0, 1):
(1 + µ)u00 = 1 in I,

(2.29)
u(0) = u(1) = 1.
For any µ ∈ [0, 1] we want to find its solution u(µ) ∈ C 2 (I).
Within this simple example we can compute the two snapshots given by
(
u0 = u(µ = 0) = 12 x2 − 21 x + 1,
(2.30)
u1 = u(µ = 1) = 14 x2 − 41 x + 1,

and by taking them as an RB, we can represent the solution manifold by means of VN =
span{u0 , u1 }.
Therefore, the reduced solution for any µ ∈ [0, 1] is
   
2 2
uN (µ) = − 1 u0 − − 2 u1 , (2.31)
1+µ 1+µ
which is actually equal to the exact solution u(µ) in the continuous counterpart of the solution
manifold MN , contained in a two-dimensional subspace of the infinite-dimensional space C 2 (I)
thanks to the linear superposition property.

2.3.2 Model Problem: Parametrized Elliptic PDE

Let us now explain, in the simple setting given by a two-dimensional parametrized elliptic PDE,
the basic steps for obtaining an RB approximation. Here, we consider steady heat conduction in
the domain Ω = Ω1 ∪ Ω2 defined in Figure 2.8.
 28 Chapter 2. Finite Element–Based Reduced Basis Method in CFD
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 2.8. Two-dimensional domain, composed of two subdomains, for the steady heat conduc-
tion problem.

The conductivity κ is assumed to be constant on Ω1 and Ω2 , as follows:

κ|Ω1 = µ1 and κ|Ω2 = 1. (2.32)

A parametric heat flux µ2 occurs over Γbase , while the left and right boundaries Γside are in-
sulated, and the top boundary Γtop is kept at a reference temperature, which for the sake of
simplicity is taken as zero. The formulation of this test case, called the thermal block problem
and reported as the first tutorial in the RBniCS library, reads as follows: for a given parameter
µ = (µ1 , µ2 ) ∈ P = [0.1, 10] × [−1, 1] find u(µ) such that

 −div(κ(µ1 )∇u(µ)) = 0 in Ω,


u(µ) = 0 on Γtop ,

 κ(µ1 )∇u(µ) · n = 0 on Γside ,
on Γbase ,

κ(µ1 )∇u(µ) · n = µ2


where n denotes the outer normal to the boundaries Γside and Γbase and the conductivity is
defined as in (2.32).
As we said, in order to approximate numerically the solution of the model under considera-
tion, we need to derive its weak formulation. If we define the function space V as

V = {v ∈ H 1 (Ω) : v|Γtop = 0},

the problem reads as follows: given a parameter µ ∈ P, find u(µ) ∈ V such that

a (u(µ), v; µ) = f (v; µ) ∀v ∈ V, (2.33)

where we have defined the parametrized bilinear form a(·, ·; µ) : V × V → R as

Z
a(u, v; µ) = κ(µ1 )∇u · ∇v dΩ.
Ω

A fundamental requirement to obtain an efficient evaluation of the reduced solution is the

so-called affine decomposition assumption. The latter ensures that we can write all the forms
 2.3. ROMs 29

considered as the expansions

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Qa Qf
θfq (µ)fq (v),
X X
a(u, v; µ) = θaq (µ)aq (u, v), f (v; µ) = (2.34)
q=1 q=1

where each form aq : V × V → R and fq : V → R is independent of the parameter value µ and

the coefficients θaq , θfq : P → R are scalar quantities which are independent of u and v.
For the thermal block problem we have Qa = 2 and Qf = 1, and the forms can be decom-
posed as Z Z
a(u, v; µ) = µ1 ∇u · ∇v dΩ + |{z}
1 ∇u · ∇v dΩ
|{z} Ω1 Ω2
θ0a (µ) | {z } θ2a (µ) | {z }
a1 (u,v) a2 (u,v)

and Z
f (v; µ) = µ2 v dΓ .
|{z} Γbase
θ1f (µ) | {z }
f1 (v)

Therefore, the standard pipeline for the discretization of a general weak formulation of the form
(2.33), via the Galerkin methodology, for a given parameter µ ∈ P reads as follows:
• Continuous problem: Find u(µ) ∈ V such that

a (u(µ), v; µ) = f (v; µ) ∀v ∈ V.

• FE problem: Find uN (µ) ∈ V N such that

a uN (µ), v N ; µ = f (v N ; µ) ∀v N ∈ V N .

• RB problem: Find uN (µ) ∈ V N such that

a uN (µ), v N ; µ = f (v N ; µ) ∀v N ∈ V N .

These considerations bring us to the following questions: How can we construct the RB space
V N ? Is it possible to choose a number of basis functions N  N that can approximate the
high-fidelity manifold? In the next subsections we will provide some answers to build the RB
approximation.

2.3.3 Space Generation: POD

POD is an explore-and-compress strategy in which one samples the parameter space at the points
of a training set Ξtrain ⊂ P, with |Ξtrain | = ntrain  1; computes the corresponding high-
fidelity solutions for all the sample points; and, following compression, retains only the essential
information.
Thus, given Ξtrain , we can define the N -dimensional POD space as the one that minimizes
the quantity
 
VNP OD = arg inf ||uN − ΠXN uN ||L2 (Ξtrain ;V ) ,
VN ⊂span{uN (µ) | µ∈Ξtrain }

which expresses the approximation accuracy of the solution manifold by an N -dimensional

space.
 30 Chapter 2. Finite Element–Based Reduced Basis Method in CFD

We define the symmetric and linear correlation operator C P OD ∈ Rntrain × Rntrain as

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

1  N i N
CP ij
OD
= u (µtrain ), uN (µjtrain ) , 1 ≤ i, j ≤ ntrain .
ntrain V

Then, we want to search for the eigenpairs (ψ k ∈ Rntrain , λk ∈ R+,0 ) of the eigenproblem

C P OD ψ k = λk ψ k , 1 ≤ k ≤ ntrain .
Arranging the eigenvalues in descending order as λ1 ≥ λ2 ≥ · · · ≥ λntrain ≥ 0, we can now
identify Ψk ∈ V for 1 ≤ k ≤ ntrain as
nX
train
k k
Ψ = ψm u(µm
train ).
m=1

Moreover, we can define Nmax as the smallest N such that

v
u nX
P OD
u train
N =t λk ≤ min ,
k=N +1

to be used as a criterion to select the dimension of the POD basis by looking at the retained energy
from the discarded snapshots. Thus, the POD-RB spaces are give by XN P OD
= span{Ψn , 1 ≤
n ≤ N }, 1 ≤ N ≤ Nmax , with orthonormalized basis functions, i.e., (Ψn , Ψm )X = δn,m , 1 ≤
n, m ≤ ntrain .
It is clear that one of the disadvantages of this technique is the huge number of FE solutions
that have to be computed to obtain a fairly accurate representation of the high-fidelity mani-
fold. Another one concerns the computation of the eigenproblem of dimension ntrain × ntrain .
Despite this, this methodology is still widely used in many contexts, including time-dependent
problems and applications in CFD.
We conclude this section with an example of the implementation in RBniCS of the offline
phase regarding the POD technique (Listing 2.7).
1 def offline (self):
2 # Loop over training set
3 for (mu_index , mu) in enumerate (self. training_set ):
4 # Choose parameter mu
5 self. truth_problem . set_mu (mu)
6 # Truth solve for mu
7 snapshot = self. truth_problem . solve ()
8 # Update snapshots matrix
9 self. update_snapshots_matrix ( snapshot )
10 # Perform POD
11 self. compute_basis_functions ()
12 # Build reduced operators
13 self. reduced_problem . build_reduced_operators ()
Listing 2.7. rbnics/reduction_methods/base/pod_galerkin_reduction.py.

2.3.4 Space Generation: Greedy Algorithm

As an alternative to POD, one can implement a greedy algorithm, an iterative procedure where
at each iteration one new basis function is added and the overall precision of the basis set is
improved.
Indeed, during the N th step of the basis selection process, the next basis function to be chosen
is the one that maximizes the model order reduction error given the (N − 1)-dimensional space
built upon Ξtrain .
 2.3. ROMs 31

Thus, the so-called proto version of the greedy algorithm over Ξtrain reads as follows.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

A LGORITHM 2.1. The prototype greedy algorithm.

.
1: Given Ξtrain define u0 (µ) = 0
2: for N = 1, . . . , Nmax : . Build spaces hierarchy
3: µN = arg maxµ∈Ξtrain ||uN (µ) − uN −1 (µ)||N
V . Error computation
4: SN = SN −1 ∪ µN . Chosen parameters
5: VN = VN −1 ∪ span{uN (µN )} . RB enrichment
6: end

The bottlenecks of this methodology are (again) the need to compute ntrain FE solutions and
the suboptimality in the sense that the greedy strategy aims to be optimal in the maximum norm
over Ξtrain rather than L2 for the POD basis. To overcome the first issue, the “actual” method,
instead of building a huge dataset, requires only one high-fidelity solution per iteration, and a
total of N solutions for a basis of dimension N .
Hence, the key ingredient is the selection of the parameters for which the solutions are com-
puted. This choice is usually guided by an error estimator ∆N (µ), which is used at the generic
N th iteration to find the worst approximated solution by the RB space of dimension N in the
whole parameter space Ξtrain . Thus, this version reads as follows.

A LGORITHM 2.2. The actual greedy algorithm.

1: Given Ξtrain
2: for N = 1, . . . , Nmax : . Build spaces hierarchy
3: µN = arg maxµ∈Ξtrain ∆N −1 (µ) . Error estimator
4: SN = SN −1 ∪ µN . Chosen parameters
5: VN = VN −1 ∪ span{uN (µN )} . RB enrichment
6: end

In order for this to be an improvement of the prototype greedy algorithm, we need ∆N (µ) to
be a sharp, inexpensive1 a posteriori error bound for the ROM error, i.e., for N = 1, . . . , Nmax
we want
||uN (µ) − uN (µ)||V ≤ ∆N (µ) ∀µ ∈ P.
This way, Algorithm 2.2 only computes at most N snapshots, in contrast to ntrain of POD and
proto-greedy. Given the fundamental importance of the a posteriori error estimator ∆N (µ), let
us now detail further the approximation capabilities of the greedy technique. Indeed, the explicit
expressions for the error estimators are postponed to the next section.

Proposition 2.5 ([81]). Assume that there exists an optimal space VN∗ and an optimal u∗N (µ) ∈
VN∗ such that
||uN (µ) − u∗N (µ)||V ≤ ce−αN for some c, α > 0.
Then the RB VN constructed by the greedy method still guarantees exponential convergence

||uN (µ) − uN (µ)||V ≤ Ce−βN for some C, β > 0.

A natural question that arises is how reasonable the assumption required in the previous
proposition is. As concerns elliptic problems, one can explicitly construct a (possibly subop-
timal) space VN∗∗ which features exponential convergence for the model problem described in
1 Marginal cost (= average asymptotic cost) is independent of N .
 32 Chapter 2. Finite Element–Based Reduced Basis Method in CFD

Subsection 2.2.1 [81, 386]. Moreover, in the past decades the scientific community has sup-
ported its numerical evidence by means of a great number of different applications.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

It can still be applied when dealing with fluid dynamics problems if one considers moderate
Re numbers, while it is not necessarily true for convection-dominated problems. Indeed, in the
latter case, i.e., Re  1, linear RB approximation techniques suffer from the slow decay of the
Kolmogorov n-width [133, 393], where the exponential decay assumption is not well suited. For
this reason, there are several ongoing efforts to restore this property in that case.
As we said, by construction, the POD basis is orthonormal, while the same does not hold
for the greedy technique, which in principle only provides the basis as the computed snapshots
VN = span{uN (µn ), 1 ≤ n ≤ N }, which are not (necessarily) orthogonal. Thus, we will apply
the Gram–Schmidt orthonormalization during the greedy strategy,

{ζN }1≤n≤Nmax = GS({uN (µn )}1≤n≤N ; (·, ·)V ),

to obtain a more suitable basis for the RB space. The Gram–Schmidt orthonormalization proce-
dure is as follows:

A LGORITHM 2.3. The Gram–Schmidt orthonormalization.

1: Given {uN (µn )}N 1
max

2: for n = 1, do ζ = uN (µ1 )/||uN (µn )||V

1

3: for n = 2, . . . , NmaxP:
n−1
4: z n = uN (µn ) − m=1 (uN (µn ), ζm )X ζm . Orthogonal projection onto Vn−1
5: ζn = z /||z ||V
n n
. Normalization
6: end

As a result of this process we obtain the orthogonality condition (ζn , ζm )X = δn,m for 1 ≤
n, m ≤ Nmax , where δn,m is the Kronecker delta symbol.
Again, let us conclude this subsection with an example of the implementation in RBniCS of
the offline phase concerning the greedy algorithm (Listing 2.8).
1 def offline (self):
2 # Greedy loop with stopping criteria
3 while self. reduced_problem .N < self.Nmax and relative_error_estimator_max
>= self.tol:
4 # Truth solve for mu
5 snapshot = self. truth_problem . solve ()
6 # Update snapshots matrix
7 self. update_snapshots_matrix ( snapshot )
8 # Build reduced operators
9 self. reduced_problem . build_reduced_operators ()
10 # Build error estimation operators
11 self. reduced_problem . build_error_estimation_operators ()
12 # Choose next mu value with Greedy
13 self. greedy ()
Listing 2.8. rbnics/reduction_methods/base/rb_reduction.py.

2.3.5 Comparison between FE and RB Spaces

Let us now examine the main differences between the FE and the RB approximations. In fact,
even if the two methodologies have common features, both relying on the projection of the weak
formulation into a discrete space (of different dimensions), they are used for different purposes,
exploiting the particular properties of each approximation level.
 2.3. ROMs 33
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 2.9. Snapshots and basis for the thermal block problem.

For example, as regards the rate of convergence, for the FE we have the standard polynomial
decay [477] depending on both the approximation degree r and the regularity q:
M
ku − uN kV ≤ Chs |u|H s+1 (Ω) for s = min{r, q}. (2.35)
α
On the contrary, as we observed in the previous subsection, for the RB technique we have an
exponential order, due to the estimate
||uN − uN ||V ≤ Ce−βN for some C, β > 0,
where C and β usually depend very mildly on r, q, and h.
Another fundamental difference is about the meaning of the bases which span the discretized
space. Indeed, the FE basis functions have local support and are not necessarily orthonormal,
while the RB ones have by construction global support and are orthonormal (directly for POD,
or by Gram–Schmidt postprocessing for greedy; see Figure 2.9 for an example regarding the
thermal block problem).
Let us now describe the system assembly of the offline and online procedures. Within the RB
methodology, we can express every element of the space V N in terms of its basis {ζn } ∈ V N as
the expansion
XN
uN (µ) = uN
j (µ)ζj . (2.36)
j=1

Then, if a(u, v; µ) satisfies the affine parameter dependence assumption (and assuming the
external force f (·) to be nonparametric), we can insert equation (2.36) into (2.34), obtaining
N
X
a(ζj , ζi ; µ)uN
j = f (ζi ), 1 ≤ i ≤ N,
j=1

N Qa
!⇓
X X
θaq (µ)aq (ζj , ζi ) uN
j = f (ζi ), 1 ≤ i ≤ N.
j=1 q=1
 34 Chapter 2. Finite Element–Based Reduced Basis Method in CFD

In this view, the offline-online decomposition for an efficient evaluation of the RB solution uN
j
corresponding to the parameter value µ, reads as follows for 1 ≤ j ≤ N :
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

• Offline: With computational cost O(N ):

– Form and store aq (ζj , ζi ) for 1 ≤ i, j ≤ N, 1 ≤ q ≤ Qa .

– Form and store f (ζj ) for 1 ≤ j ≤ N .

• Online: With computational cost O(Qa N 2 + N 3 ):

PQa q
– Form q=1 θa (µ)aq (ζj , ζi ) for 1 ≤ i, j ≤ N .

j (µ) for 1 ≤ j ≤ N .
– Solve for uN

In the evaluation process of uN j for 1 ≤ j ≤ N , we note that from the jth column of the
matrix Z N = {(ζj )k }j,k ∈ R N ×N
, denoting the coefficients when the jth basis function ζj is
expressed in terms of the basis functions {ϕN
k }k , we can write

N
X
ζj = (ζj )k ϕN
k ,
k=1

and thus we obtain the following expression for the RB block corresponding to the parameter-
independent form aq (·, ·):
P 
N PN
[AN,q ]ij = aq (ζj , ζi ) = aq k=1 (ζj )k ϕN
k , l=1 (ζ i ) l ϕN
l
PN PN q N N (2.37)


= k=1 l=1 (ζj )k a ϕk , ϕl (ζi )l
= Z TN AN ,q Z N .

Of course, having considered a new (reduced) system, one can investigate its algebraic sta-
bility, and in particular its connection to the orthonormality assumption on Z N .

Proposition 2.6. Assume that a(·, ·; µ) is symmetric and that Z N is orthonormal. Then the
condition number of
Qa
X
AN (µ) = θaq (µ)AN,q
q=1

is bounded independently on N .

Proof. For the sake of simplicity, let us remove µ from the notation. Let vN = [vjN ]N
j=1 be an
N
eigenvector associated with λmax (A ), and define
N
X
N
v = vjN ζj . (2.38)
j=1

Due to orthonormality,
 
X N N
X N
X N
X
kv N k2V =  viN ζi , vjN ζj  = viN vjN (ζi , ζj ) = kvN k2 . (2.39)
i=1 j=1 i,j=1 i=1
 2.3. ROMs 35
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 2.10. FE (369 nonzero elements) and RB (225 nonzero elements) structures of the assem-
bled systems, respectively left and right.

Now the continuity of a(·, ·), with constant M , implies

T
λmax (AN )kvN k2 = vN AN vN = a(v N , v N ) ≤ M kv N k2V = M kvN k2 . (2.40)

Similarly, one obtains that λmin (AN ) ≥ α, where α is the coercivity constant. We can conclude
by taking the ratio K2 ((AN )) = λmax (AN )/λmin (AN ), which is bounded independently by the
stability constants of the system.

Finally, let us summarize what we have understood in this section concerning the properties
of the FE and the RB assembled systems.

• Dimension:

– FE: N , associated with polynomial degree p and mesh size h;

– RB: N , set either a priori or by a tolerance. Usually N  N thanks to exponential
convergence.

• Sparsity (see Figure 2.10):

– FE: sparse (usually banded) systems;

– RB: dense systems.

• Conditioning:

– FE: deteriorates as h−2 when h decreases.

– RB: may depend on N , but with a provable N -independent upper bound.

We are now ready to show an application of the RB methodology within the context of
geometrical parametrization.

2.3.6 Model Problem: Geometrical Parametrization in Heat Transfer

As a model problem, we can consider a geometrically parametrized configuration to study the
steady-state temperature distribution. The physics is governed by the Laplace equation, and
the goal here is to investigate the behavior of interesting outputs while varying the geometry of
the cooling component. We consider the parametric domain (on the left in Figure 2.11) given by
 36 Chapter 2. Finite Element–Based Reduced Basis Method in CFD
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 2.11. Left: parametric domain Ωo (µ). Right: reference domain Ω.

Ωo (µ) = [−2, 2] × [−2, 2] \ [µ0 , µ1 ] × [µ0 , µ1 ], where the geometrical parameter is µ = (µ0 , µ1 )
in the parameter space P = [0.5, 1.5]2 .
From the original parametrized domain Ωo (µ), which can be decomposed as
K[
dom
k
Ωo (µ) = Ωo (µ),
k=1

we set the problem on a reference domain Ω such that

K[
dom
k
Ω= Ω ,
k=1

where Ω = Ωo (µref ) for µref ∈ P.

We require that
k k
Ωo (µ) = T aff,k (Ω ; µ), 1 ≤ k ≤ Kdom ,
where
T aff,k (x; µ) = C aff,k (µ) + G aff,k (µ)x
k k
is an invertible affine mapping from Ω onto Ωo (µ).
We further require that for all µ ∈ P,
0 k k0
T aff,k (x; µ) = T aff,k (x; µ) ∀x ∈ Ω ∩ Ω , 1 ≤ k, k 0 ≤ Kdom ,

to ensure a continuous piecewise-affine global mapping T aff (·; µ) from Ω onto Ωo (µ).
Thus, we can now describe the weak formulation on the original parametrized domain. For
v, w ∈ H 1 (Ωo (µ)), we consider
 ∂w 
K
Xdom Z ∂xo1
∂v ∂v k ∂w
 
ao (v, w; µ) = ∂xo1 ∂xo2 v Koij (µ)  ∂xo2
,
Ωk
o (µ)
k=1 w
 2.3. ROMs 37

where Kok : P → R3×3 is symmetric and positive definite for 1 ≤ k ≤ Kdom ; note that Kok being
affine in xo is also permissible.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Of course, in order to make the parameters explicit and to be able to run the reduction ma-
chinery, we exploit the change of variables formula, resulting in the following problem. For
v, w ∈ H 1 (Ω) we consider
 ∂w 
KXdom Z ∂x1
 ∂v ∂v
 k ∂w 
a(v, w; µ) = ∂x1 ∂x2 v Kij (µ)  ∂x ,
k 2
Ω
k=1 w
where
   
aff,k −1 0
(G )
Kk (µ) = |detG aff,k (µ)|D(µ)Kok (µ)DT (µ) and D(µ) =    0 .
0 0 1
Finally, we can make a few remarks and point out some issues when dealing with problems
characterized by geometrical parametrization.
• One of the requirements of the above methodology is the existence of an analytical map-
ping between subdomains, so that snapshots are defined in a common mesh. For recent
extensions to the removal of this assumption we refer the reader to [324, 325].
• An affine mapping is used in order to ensure the availability of a parameter-separable
formulation, but more general mappings are possible and very useful in CFD [508, 569].
• The pullback procedure (change of variables in integrals) required for the assembly on the
reference domain is usually tedious and error-prone but can be carried out automatically
by symbolic processing tools once the expression of the mapping is provided. For this
reason, RBniCS offers automatic pullback capabilities.

2.3.7 Reduction of CFD Problems

In this section, we recall the basic structures and methodology to apply model order reduction
to the CFD systems that we presented in Subsections 2.2.3 and 2.2.4. As concerns the former,
we can start from the weak formulation in (2.19), project the linear system onto the high-fidelity
space, and obtain the FE approximation of the Stokes problem by solving
AU + B T P = F,

(2.41)
BP = 0,
which can be seen as a saddle-point structure by writing the system as
A BT
    
U F
= , (2.42)
B 0 P 0
where we have introduced the quantities
A = [aij ] = [a(ϕj , ϕi )] ∈ RNu ×Nu , B = [bkm ] = [b(ϕm , ψk )] ∈ RNp ×Nu ,
U = [uj ] ∈ RNu , P = [pk ] ∈ RNp .
Once we have built the reduced spaces, we can further project the above system onto the
low-dimensional manifold spanned by the RB. This way, we obtain the reduced system
T

AN U N + B N P N = FN ,
(2.43)
BN PN = 0,
 38 Chapter 2. Finite Element–Based Reduced Basis Method in CFD

which, as we can see, still has the saddle-point structure

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

T
    
AN B N UN FN
= , (2.44)
BN 0 PN 0

where we have introduced the reduced quantities, by means of the RB matrix, as follows:
uT u
AN = Z N AZN ,
pT u
BN = ZN BZN ,
uT
FN = ZN F,
p
and ZNu
, ZN are built using either a greedy or a POD training.
As an example, we present the space generation algorithms corresponding to Listings 2.7 and
2.8 for the Stokes problem.

A LGORITHM 2.4. Greedy algorithm for Stokes.

1: Given Ξtrain
2: for N = 1, . . . , Nmax : . Build spaces hierarchy
3: µN = arg maxµ∈Ξtrain ∆N −1 (µ) . Error estimator
4: (uN (µN ), pN (µN )) = Stokes(µN ) . Snapshot computation
5: VN = VN −1 ∪ span{uN (µN )} . RB enrichment for velocity space
6: QN = QN −1 ∪ span{pN (µN )} . RB enrichment for pressure space
7: end

A LGORITHM 2.5. POD algorithm for Stokes.

1: Given Ξtrain
2: for all µ ∈ Ξtrain : . Loop over training set
3: (uN (µN ), pN (µN )) = Stokes(µN ) . Snapshot computation
4: S u = S u ∪ span{uN (µN )} . Store velocity snapshot in a matrix S u
5: S p = S p ∪ span{pN (µN )} . Store pressure snapshot in a matrix S p
6: VN = P OD(S u , V, N ) . Perform POD on velocity space
7: QN = P OD(S p , Q, N ) . Perform POD on pressure space
8: end

We remark that, in order to deal with a well-posed problem, the assumptions of Theorem
2.3 have to be satisfied. In particular, the algebraic equivalent of the inf-sup condition (2.20) for
the FE discretization, which reads as ker(B T ) = {0}, must be met. Otherwise, all elements in
ker(B T ) are spurious pressure modes. Even if we assume that ker(B T ) = {0} holds at the FE
level, it cannot be the case at the reduced level. Let us call ϕm N an element of the basis of VN
(i.e., represented by a column of ZN u
) and ψN
k
an element of the basis of QN (i.e., represented
p
by a column of ZN ). Then, we obtain BN = [b(ϕm N , ψN )] = 0, because ϕN is obtained as a
k m

linear combination (POD or Gram–Schmidt) of weakly divergence-free snapshots {uN (µN )}.
Thus, ker(BN T
) is definitely larger than {0} and is actually the whole reduced pressure space
QN . Unfortunately, this means that every pressure basis function is a spurious pressure mode. In
the next subsection, we will describe how to handle the inf-sup issue at the reduced level, relying
on a space enrichment approach.

2.3.8 Supremizer Enrichment of the Velocity Space

In view of what we discussed at the end of the last subsection, it is important to control the
behavior of the inf-sup constants, both at the FE and the reduced order approximation levels.
 2.3. ROMs 39

Thus, we can define the following quantities:

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

(q, B(µ)v)
FE inf-sup constant: β N (µ) = inf sup ; (2.45)
q6=0 v6=0 kvkV kqkQ

(qN , BN (µ)vN )
RB inf-sup constant: βN (µ) = inf sup . (2.46)
q 6=0 v 6=0
N N
kvN kVN kqN kQN
As we have seen, we currently have βN (µ) = 0 due to the spurious pressure modes. Therefore,
we aim at changing the reduced space VN so that βN (µ) ≥ β N (µ).
Starting from β N (µ), and thanks to the fact that QN ⊂ QN , we can restrict the inf to a
smaller subspace, possibly increasing the infimum value, obtaining

(q, B(µ)v) (Zp qN , B(µ)v)

β N (µ) = inf sup ≤ inf sup .
q6=0 v6=0 kvkV kqkQ q 6=0 v6=0 kvkV kZp q kQ
N N

We would like to do the same for the sup argument, owing to the fact that the inclusion still holds
for the velocity spaces VN ⊂ V N . Unfortunately, in general, restricting a sup to a smaller space
might decrease the supremum, unless we were sure that the supremizer sµ (q) belongs the space
VN for every q ∈ QN , i.e.,

(q, B(µ)v) (B(µ)T q, v)

VN 3 sµ (q) = arg sup = arg sup ∀q ∈ QN . (2.47)
v6=0 kvkV v6=0 (Xu v, v)

To retrieve the quantities defined above, it is equivalent to solve

Xu sµ (q) = B T (µ)q (2.48)

for every q ∈ QN (i.e., by linearity, for every ψN k

, k = 1, . . . , N ), where Xu is the matrix
associated with the inner product V . Equivalently, the strong form of the problem, if using the
H 1 seminorm as the V norm (at least one wall/inlet), can be expressed as follows:

given q ∈ QN , find sµ (q) ∈ V such that − ∆sµ (q) = −∇q, (2.49)

which is useful for the extension to finite volume approximation [550].

At the reduced order level, the exact supremizer enrichment reads as follows: given QN =
span{ψNk
, k = 1, . . . , N }, compute

Xu sµ (ψ k ) = B T (µ)ψ k for k = 1, . . . , N (2.50)

and enrich VN as
k
VeN = VN ⊕ span{sµ (ψN ), k = 1, . . . , N }. (2.51)
This allows us to continue the argument discussed earlier, reaching the desired inequality among
the inf-sup constants [511]:
(i) (Zp qN , B(µ)v) (ii) (Zp qN , B(µ)sµ (Zp qN ))
β N (µ) ≤ inf sup = inf
q 6=0 v6=0
N
kvkV kZp qN kQ q 6=0
N
ksµ (Zp qN )kV kZp qN kQ
(iii) (B T (µ)Zp qN , Zu vN )
≤ inf sup
q 6=0 v 6=0
N N
kZu vN kV kZp qN kQ
(iv) (ZuT B T (µ)Zp qN , vN ) (qN , BN (µ)vN )
= inf sup = inf sup = βN (µ),
q 6=0 v 6=0
N N
kZu vN kV kZp qN kQ q 6=0 v 6=0
N N
kvN kVN kqN kQN
 40 Chapter 2. Finite Element–Based Reduced Basis Method in CFD

where we have exploited the following facts: (i) QN ⊂ QN ; (ii) the supremizer’s definition; (iii)
the exact enrichment of the velocity space by supremizer solutions, resulting in the space VeN ;
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

and (iv) the definition of BN T

(µ) = ZuT B T (µ)Zp .
It is clear that, for the sake of simplicity, an inexact version of the supremizer enrichment is
more practical, via either greedy or POD, as we see from the following algorithms.

A LGORITHM 2.6. Greedy algorithm with supremizer enrichment for Stokes.

1: Given Ξtrain
2: for N = 1, . . . , Nmax : . Build spaces hierarchy
3: µN = arg maxµ∈Ξtrain ∆N −1 (µ) . Error estimator
4: (uN (µN ), pN (µN )) = Stokes(µN ) . Snapshot computation
5: sN (µN ) = Supremizer(µN , pN (µN )) . Supremizer computation
6: VN = VN −1 ∪ span{uN (µN ), sN (µN )} . RB enrichment for velocity space
7: QN = QN −1 ∪ span{pN (µN )} . RB enrichment for pressure space
8: end

A LGORITHM 2.7. POD algorithm with supremizer enrichment for Stokes.

1: Given Ξtrain
2: for all µ ∈ Ξtrain : . Loop over training set
3: (uN (µN ), pN (µN )) = Stokes(µN ) . Snapshot computation
4: sN (µN ) = Supremizer(µN , pN (µN )) . Supremizer computation
5: S u = S u ∪ span{uN (µN )} . Store velocity snapshot in a matrix S u
6: S p = S p ∪ span{pN (µN )} . Store pressure snapshot in a matrix S p
7: S = S ∪ span{s (µ )}
s s N N
. Store supremizer snapshot in a matrix S s
8: VN = P OD(S , V, N ) ⊕ P OD(S , V, N )
u s
. Perform POD on velocity space
9: QN = P OD(S p , Q, N ) . Perform POD on pressure space
10: end

Examples of CFD applications are the flow in deformable channels governed by the Stokes
system or the behavior of the Navier–Stokes nonlinear model in a backward-facing step channel,
as in the ones shown in Figure 2.12. Let us focus on the latter model, where we set the Reynolds
number to Re = 150, and we want to study the stability of the online problem with respect to
the number of supremizers Ns considered [34].
We report in Table 2.1 the approximation property of the reduced model, considering both
the inf-sup constant and the norm of the reduced pressure.

Table 2.1. Approximation accuracy for the ROM with respect to the number of supremizers exploited.

Ns k kpm spurious pressure modes βN (µ)2 kpN k2

0 5 4 yes < tol 1.1653e+09
1 4 2 yes < tol 7.5223e+06
5 1 −5 no 1.8201e-09 4.3631e+02
7 1 −7 no 3.3185e-08 2.6174e+01
10 1 −10 no 1.0258e-06 1.1228e+01
15 1 −15 no 1.6581e-05 1.1078e+01
30 1 −30 no 1.7305e-02 1.1061e+01
FOM – – no 1.7312e-02 –
 2.3. ROMs 41
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 2.12. Top: Stokes flow in deformable channels (RBniCS tutorial 12). Bottom: Backward-
facing step for moderate Reynolds, Navier–Stokes equations (RBniCS tutorial 17).

As we can see from the first row in Figure 2.13, if we do not enrich the RB space with the
supremizers, we cannot avoid spurious pressure modes. Numerically, this is also confirmed by
the large value of kpN k2 and the incorrect approximation of the velocity.
If too few supremizers are considered, say 0 < Ns < Np /3, the velocity is correctly approx-
imated, yet the pressure is not recovered accurately (second row in Figure 2.13).
However, if enough supremizers are considered (say, Ns > Np /3), the ROM allows us to get
a better qualitative agreement (both for velocity and pressure) and is inf-sup stable (βN (µ) >
0) (third row in Figure 2.13). It is therefore, in general, not necessary to assume βN (µ) ≥
βh (µ) to obtain good results. As in the previous proof, even considering POD-approximate
supremizer enrichment, the inequality βN (µ) ≥ βh (µ) > 0 holds as long as Ns ≥ Np (last row
in Figure 2.13).

2.3.9 Extension to the Parabolic Case

In this subsection we want to extend the previous examination to time-dependent problems gov-
erned by parabolic PDEs. Consider a domain Ω ⊂ Rd , d = 1, 2, 3, and a time interval T = [0, T ].
For every time t ∈ T find a solution u(t) ∈ C 2 (Ω) such that
∂u


 − ∆u = g f in Ω × T,
∂t (2.52)
 u=0 on ∂Ω × T,
u = u0 in Ω × {0}.

 42 Chapter 2. Finite Element–Based Reduced Basis Method in CFD

Ns = 0 Ns = 1
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Ns = 5 Ns = 7

Ns = 10 Ns = 15

Ns = 30

Figure 2.13. Velocity and pressure reduced solution for the Navier–Stokes equations in the
backward-facing step channel, varying the number of supremizers Ns .

.
We apply a semidiscretization approach, obtaining the following formulation: ∀k ∈ K =
{1, . . . , K}, find u (µ) ∈ V such that
k

 k
u (µ) − uk−1 (µ)

m , v; µ + a(uk (µ), v; µ) = g(tk )f (v, µ) ∀v ∈ V,
∆t

where T has been divided in K subintervals of length ∆t. Then, the discrete problem is recovered
by exploiting a finite difference approximation of the temporal derivative and a space discretiza-
tion via FE. So the finite-dimensional problem reads as follows: ∀k ∈ K, find uN ,k (µ) ∈ V N
such that
 N ,k
u (µ) − uN ,k−1 (µ)

m , v; µ + a(uN ,k (µ), v; µ) = g(tk )f (v, µ) ∀v ∈ V N .
∆t

Finally, we can apply model order reduction to the parabolic system, and the reduced problem
reads as follows: ∀k ∈ K, find uN,k (µ) ∈ V N such that
 N,k
u (µ) − uN,k−1 (µ)

m , v; µ + a(uN,k (µ), v; µ) = g(tk )f (v, µ) ∀v ∈ V N .
∆t

Let us remark that when dealing with parabolic problems, if we assume that the form a(·, ·)
is stable and that the θm,a
q
(µ), 1 ≤ q ≤ Qm,a , are smooth, then the reduced manifold MN ,K =
{u (µ)| ∀k ∈ K, ∀µ ∈ P} lies on a smooth (P + 1)-dimensional manifold, where P is the
N ,k

dimension of the parameter space P; see Figure 2.14.

A LGORITHM 2.8. POD-greedy algorithm for parabolic problems.

1: Z = ∅, S∗ = {µ∗ }
2: while N ≤ Nmax,0 . Build space hierarchy
3: {χm , 1 ≤ m ≤ M1 } = P OD({uN (tk , µ∗ ) ∀k ∈ K}, M1 ) . POD on time
4: Z ←− {Z, {χm , 1 ≤ m ≤ M1 }} . Store M1 basis
 2.3. ROMs 43

5: N ←− N + M2 . Update dimension RB
6: . Perform POD on Z
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

{ζn , 1 ≤ n ≤ N } = P OD(Z, N )
7: XN = span{ζn , 1 ≤ n ≤ N } . Build reduced space
8: µ∗ = arg maxµ∈Ξtrain ∆KN (µ) . Error estimator
9: S∗ ←− S∗ ∪ µ∗ . Chosen parameters
10: end

uN ,k (µntrain )
N ,k
M

uN ,k (µ)
k=0
uN ,k (µ1 )

k=K
P

Figure 2.14. Reduced manifold MN ,K for time-dependent problem.

Note how time t and parameters µ are treated in a different way in the sampling procedure.
Other model reduction techniques may treat them as one “combined” parameter µ e = (µ, t), as
in the proper generalized decomposition (PGD) approach [123].
Let us conclude by highlighting the main differences between the greedy and the POD-greedy
procedure, as implemented in RBniCS (Listings 2.9 and 2.10).
1 def update_basis_matrix (self , snapshot ):
2 basis_functions = self. reduced_problem . basis_functions
3 new_basis_function = self.GS. apply(snapshot , basis_functions )
4 basis_functions . basis_functions . enrich ( new_basis_function )
5 self. reduced_problem .N += 1
Listing 2.9. rbnics/reduction_methods/base/rb_reduction.py.

1 def update_basis_matrix (self , snapshot_over_time ):

2 N1 , N2 = ... # set by the user
3 N = self. reduced_problem .N
4 basis_functions = self. reduced_problem . basis_functions
5
6 POD_time_trajectory . clear ()
7 POD_time_trajectory . store_snapshot ( snapshot_over_time )
8 (eig1 , _, basis_functions1 ) = POD_time_trajectory .apply(N1)
9
10 POD_basis . store_snapshot ( basis_functions1 ,\
11 weight =[ sqrt(e) for e in eig1 ])
12 (_, _, basis_functions2 ) = POD_basis .apply(N + N2)
13
14 basis_functions . clear ()
15 basis_functions . enrich ( basis_functions2 )
16 self. reduced_problem .N += N2

Listing 2.10. rbnics/reduction_methods/base/time_dependent_rb_reduction.py.

 44 Chapter 2. Finite Element–Based Reduced Basis Method in CFD

2.4 A Posteriori Error Estimation for Certified RB Methods

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

This section aims at providing explicit expressions for the error estimation employed in the
greedy Algorithms 2.2 and 2.4. Namely, we would like to employ a sharp and inexpensive a
posteriori error bound for kuN (µ) − uN (µ)kV . This is crucial in order to make the greedy
building phase efficient and nondependent on the high-fidelity dimension N . Exploiting an error
estimator in the greedy offline phase, indeed, leads to several advantages, such as the following:

• The time needed to evaluate a new parametric instance decreases in a reduced setting
without losing numerical accuracy, i.e., ∂tcomp (µ → uN (µ)) is small with a rigorous
assessment in the error kuN (µ) − uN (µ)kV for all µ ∈ P.

• A larger training set Σtrain can be used to better explore the solutions of the parametric
problem MN .

• A faster building phase can be performed, which will result in rapid convergence to VN .

In the following we provide explicit expressions of the error estimator ∆N (µ) for several
problems based on different equations, including elliptic, Stokes, and parabolic problems.

2.4.1 Error Estimator for an Elliptic Problem

Let us start with the elliptic case. We assume we are dealing with the following problem: for a
given µ ∈ P and f ∈ V 0 find u(µ) ∈ V N such that

a (u(µ), v; µ) = f (v; µ) ∀v ∈ V, (2.53)

where a : V × V → R is a symmetric, continuous, and coercive bilinear form. In order to

derive a proper definition for the error bound in this case, we need to define the reduced residual
0 0
r : P → V N , where r(v; µ) = f (v; µ) − a (uN (µ), v; µ). Here, V N represents the dual space
of V . Thus, for the Riesz representation theorem, there exists ê(µ) an element of V such that
N

(ê(µ), v)V = r(v; µ) ∀v ∈ V N . (2.54)

Furthermore,
r(v, µ)
||r(·; µ)||V N 0 = sup = ||ê(µ)||V . (2.55)
v∈V N ||v||V
In order to achieve our goal of making ∆N (µ) explicit, we still need the following parametric
quantities:

• The continuous and the discrete coercivity constants of the bilinear form a(·, ·; µ) are
defined, respectively, as
a(v, v; µ)
α(µ) = inf (2.56)
v∈V ||v||2V
and
a(v, v; µ)
αN (µ) = inf . (2.57)
v∈V N ||v||2V
For an elliptic coercive problem such as (2.53) the following relation holds:

αN (µ) ≥ α(µ) ≥ α > 0 ∀µ ∈ P.

Moreover, we assume that a lower bound αLB

N
: P → R of the coercivity constant is
 2.4. A Posteriori Error Estimation for Certified RB Methods 45

provided that, by definition, verifies

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

N
0 < αLB (µ) ≤ αN (µ) ∀µ ∈ P

such that ∂tcomp (µ → αLB

N
(µ)) is independent of N .
Besides its coercivity, the considered bilinear form is also continuous, i.e., for all v, w ∈ V ,

a(v, w; µ) ≤ γ(µ)kvkV kwkV ,

where
a(v, w; µ)
γ(µ) = sup sup .
v∈V w∈V ||v||V ||w||V

We now have all the needed quantities to allow us to define the error estimator

||ê(µ)||V
∆en
N (µ) = q . (2.58)
N (µ)
αLB

Concerning the quality of the considered error estimator, we can define the effectivity as

en ∆en
N (µ)
ηN (µ) = ,
||uN (µ)
− uN (µ)||µ

where kvk2µ = a(v, v; µ) is called the energy norm. The effectivity gives information about how
far we are from the exact error we commit by approximating the high-fidelity solution with a
reduced one. Before explaining its role in detail, we want to specify the relation between the
effectivity, coercivity, and continuity constants.

Proposition 2.7. For N = 1, 2, . . ., the following relation holds:

s
en γ(µ)
1 ≤ ηN (µ) ≤ N (µ)
∀µ ∈ P. (2.59)
αLB

Proof. The proof is straightforward once we recall the symmetric property over the bilinear form
a(·, ·; µ). Indeed, it is a matter of simple computations to show that

a(e(µ), v; µ) = (ê(µ), v)V . (2.60)

Taking v = e(µ) = uN (µ) − uN (µ), applying the Cauchy–Schwarz inequality to a(e(µ), e(µ),
µ), and recalling the definition of the energy norm, we obtain the bilinear form

||e(µ)||2µ ≤ ||ê(µ)||V ||e(µ)||V . (2.61)

In addition, exploiting the coercivity property of the bilinear form a(·, ·; µ), we have
p q
||e(µ)||µ = a(e(µ), e(µ); µ) ≥ αN (µ))||e(µ)||V , (2.62)

and hence, combining (2.61) and (2.62), we easily obtain

q ||e(µ)||2µ
αN (µ) ≤ ||e(µ)||µ . (2.63)
||ê(µ)||V
 46 Chapter 2. Finite Element–Based Reduced Basis Method in CFD

Relation (2.63) can be read as

||e(µ)||µ ≤ ∆en
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

N (µ)

or
en
ηN (µ) ≥ 1 (2.64)
thanks to the definition of the error estimator and the effectivity. We now want to show the other
side of the inequality (2.59). To this end, let us plug v = ê(µ) into (2.60). Thus, by applying
once again the Cauchy–Schwarz inequality, we have

||ê(µ)||2V ≤ ||ê(µ)||µ ||e(µ)||µ . (2.65)

Furthermore, by the continuity of the bilinear form and (2.65), we obtain

||ê(µ)||2µ = a(ê(µ), ê(µ); µ) ≤ γ(µ)kê(µ)k2V ≤ γ(µ)||ê(µ)||µ ||e(µ)||µ , (2.66)

and thus
||ê(µ)||µ ≤ γ(µ)||e(µ)||µ . (2.67)
It is now clear, exploiting (2.66) and (2.67), that the following relation holds:

∆en
N (µ)
2
||ê(µ)||µ ||e(µ)||µ γ(µ)||e(µ)||2µ
N (µ)
≤ N (µ)
≤ N (µ)
,
αLB αLB αLB
and thus, by definition, we obtain the bound
s p
en ∆en
N (µ)
2 γ(µ)
ηN (µ) = N (µ)||e(µ)||2
≤ q .
αLB µ N
αLB (µ)

The effectivity measures the quality of the considered error estimator. Inequality (2.64) guar-
antees the rigor, while the second inequality guarantees its sharpness: in order to achieve better
performance, the effectivities are desired to be close to unity. It remains to answer to the fol-
lowing question: is the error estimator inexpensive to compute? The main feature to guarantee
an efficient way to build and compute the error estimator is its parameter affine decomposition.
Indeed, the residual can be written as

r(v, µ) = f (v) − a(uN (µ), v; µ)

XN 
= f (v) − a (uN (µ))n ζn , v; µ
n=1
N
X (2.68)
= f (v) − (uN (µ))n a(ζn , v; µ)
n=1
XN Qa
X
= f (v) − (uN (µ))n θaq (µ)aq (ζn , v).
n=1 q=1

Moreover, if we consider the equivalent expression exploiting the related Riesz representation
together with the linear superposition property, we obtain
Qa X
X N
(ê(µ), v)V = f (v) − θaq (µ)(uN (µ))n aq (ζn , v).
q=1 n=1
 2.4. A Posteriori Error Estimation for Certified RB Methods 47

This translates into the relation

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Qa X
X N
ê(µ) = C + θaq (µ)(uN (µ))n Lqn ,
q=1 n=1

where
(C, v)V = f (v) ∀v ∈ V N ,

while
(Lqn , v)V = −aq (ζn , v) ∀v ∈ V N , 1 ≤ n ≤ N, 1 ≤ q ≤ Qa .

Thus, we obtain that

 Qa X
X N Qa X
X N 
kê(µ)k2V = C+ θaq (µ)(uN (µ))n Lqn , C + θaq (µ)(uN (µ))n Lqn
q=1 n=1 q=1 n=1 V
Qa X
N
(
X
= (C, C)V + θaq (µ)(uN (µ)) 2(C, Lqn )V
q=1 n=1
Qa X N
)
X
+ θar (µ)(uN (µ))m (Lqn , Lrm )V .
r=1 m=1
(2.69)
Namely, the aforementioned arguments allow us to employ an offline-online decomposition, and
the residual can be computed efficiently. Indeed, the norm of the residual, i.e., the numerator of
the error estimator (2.58), is built in the offline phase, where the complexity depends on N , Qa
and N . In this stage,

• C, Lqn are evaluated for 1 ≤ n ≤ Nmax and 1 ≤ q ≤ Qa , and

• (C, C)V , (C, Lqn )V , and (Lqn , Lrm )V are assembled and stored for 1 ≤ n, m ≤ Nmax and
1 ≤ q, r ≤ Qa .

We recall that the offline phase still depends on the high-fidelity dimension N ; however, it is
performed only once. After this possibly costly phase, the online stage is performed for each
new parametric instance µ. In this specific stage of the reduction process, the complexity de-
pends on N and Qa but not on N , so that its computational cost scales as O(Q2a N 2 ), and, thus,
many evaluations of the residual norm can be performed in a small amount of time. However, it
still remains to take care of the lower bound of the coercivity constant, which does not depend
on N . Indeed, we now discuss some strategies to compute the lower bound for the coercivity
constant. The lower bound is necessary since, by definition (2.57), the discrete version of the
coercivity constant is equivalent to the solution of the following generalized eigenvalue problem:
find (λ, w) ∈ R × V N such that

a(w, v; µ) = λ(w, v)V ∀v ∈ V N . (2.70)

Eigenvalue problems might be computationally expensive to solve in a high-fidelity setting, slow-

ing both the offline and the online phases of the ROMs. This is the main reason behind the in-
troduction of techniques able to approximate the stability constant needed in the error estimator
definition.
 48 Chapter 2. Finite Element–Based Reduced Basis Method in CFD

2.4.2 The min-θ Approach

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

The first approach we present is the min-θ approach. It applies to parametrically coercive prob-
lems. Indeed, we first recall that the bilinear form we are dealing with is affine decomposed and
can be written as
XQa
a(v, w; µ) = θaq (µ)aq (v, w),
q=1

where > 0 for all µ ∈ P and a (v, v) ≥ 0 for all v ∈ V and 1 ≤ q ≤ Qa . Thus, taking
θaq (µ) q

w = v, we obtain
Qa
X
a(v, v; µ) = θaq (µ)aq (v, v)
q=1
Qa
X θaq (µ) q 0 q
= θ (µ )a (v, v)
q=1
θaq (µ0 ) a
Qa
θaq (µ) X q 0 q (2.71)
≥ min q θa (µ )a (v, v)
q∈[1,Qa ] θa (µ0 )
q=1
θq (µ)
≥ min qa 0 a(v, v; µ0 )
q∈[1,Qa ] θa (µ )
θq (µ)
= min qa 0 ||v||2V ;
q∈[1,Qa ] θa (µ )

thus, it is a matter of simple computation to derive the lower bound

a(v, v; µ) θq (µ)
αN (µ) = inf 2 ≥ min qa 0 = αLB N
(µ).
v∈V N ||v||V q∈[1,Qa ] θa (µ )

This procedure provides a positive lower bound for the discrete coercivity constant, coinciding
with the exact coercivity constant when µ = µ0 . However, it results in an unsharp bound.

2.4.3 The Successive Constraint Method

The proposed min-θ approach can be extended to more general problems. Indeed, we will present
a local minimization technique, known as the successive constraint method (SCM), which en-
larges the class of problems for which a lower bound of the coercivity constant can be found;
see the groundbreaking work [301]. This approach is divided into offline and online phases.
While the offline phase performs generalized high-fidelity eigenvalue problems, the online part
provides a lower bound for the discrete coercivity constant at each new parametrized instance,
and its computation does not depend on N . We now address the strategy in detail.
We begin with the offline phase of the algorithm. Let us recall that
Qa
X
a(v, v; µ) = θq (µ)aq (v, v);
q=1

hence, the discrete coercivity constant is

αN (µ) = inf J obj (µ; v), (2.72)
v∈V N

where
Qa
obj
X aq (v, v)
J (µ; v) = θq (µ) .
q=1
||v||2V
 2.4. A Posteriori Error Estimation for Certified RB Methods 49

The main idea is to interpret the search for the lower bound as a minimization problem, as
described in (2.72). To this end, we write
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

αN (µ) = inf J obj (µ; y),

y∈Y

where
Qa
X
J obj (µ; y) = θq (µ)y q
q=1

and ( )
Qa N aq (vy , vy )
q
Y= y∈R | ∃vy ∈ V : y = , 1 ≤ q ≤ Qa
||vy ||2V

is the set of admissible solutions. The space of admissible solutions can be built using a lower
and an upper bound to the set. Namely, we want to find YU B ⊂ Y ⊂ YLB to restrict or enlarge
the minimization space we are dealing with. Before starting this construction argument, we still
need to define some quantities:

• the box
Qa
" #
Y aq (v, v) aq (v, v)
B= inf 2 , sup 2 ,
v∈V N ||v||V v∈V N ||v||V
q=1

given by the computation of the smallest and the largest eigenvalues of the generalized
eigenvalue problem related to each aq (·, ·), which can be performed only once, in an offline
fashion;

• a finite set of parameters for which we know the value of the coercivity constants, i.e.,

CJ = {µ1SCM , . . . , µJSCM } ⊂ P;

• the set of the closest M parameters to a given value µ, say CJM,µ .

We can now focus on building a lower bound for the admissible set. Indeed, we can define YLB
as
Qa
( )
.
θq (µ0 )y q > αN (µ) ∀µ0 ∈ CJM,µ .
X
YLB = YLB (µ; CJ , M ) = y ∈ RQa | y ∈ B,
q=1

Lemma 2.8. Given CJ ⊂ P and M ∈ N, the following holds:

Y ⊂ YLB (µ; CJ , M ) ∀µ ∈ P.

Proof. We show that if y ∈ Y, then y ∈ YLB . Indeed, since y ∈ Y, by definition, there exists
vy ∈ V N such that
aq (vy , vy )
yq = ∀µ ∈ P.
||vy ||2V
Moreover, it is straightforward to see that

aq (v, v) aq (vy , vy ) aq (v, v)

inf 2 ≤ 2 = y q ≤ sup 2 ,
v∈V N ||v||V ||vy ||V v∈V N ||v||V
 50 Chapter 2. Finite Element–Based Reduced Basis Method in CFD

namely, y ∈ B. Lastly,
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Qa
X a(vy , vy ; µ)
θq (µ)y q = ≥ αN (µ) ∀µ ∈ P,
q=1
||vy ||2V

which proves the statement.

The next step is to define a lower bound for the discrete coercivity constant. Consider the fol-
lowing linear optimization problem:
N
αLB (µ; CJ , M ) = min J obj (µ; y). (2.73)
y∈YLB

Proposition 2.9. Given CJ ⊂ P and M ∈ N, we can define the coercivity lower bound
N
αLB (µ) ≤ αN (µ) ∀µ ∈ P.

Proof. The statement directly follows by definitions and by Lemma 2.8. Indeed,
N
αLB (µ) = min J obj (µ; y)
y∈YLB (µ)
≤ min J obj
(µ; y) (2.74)
y∈Y
N
= α (µ).

Now we can define the other subset we are interested in, the upper bound for the admissible set,
i.e.,
.
YU B = YU B (µ; CJ , M ) = {y ∗ (µ0 ) ∈ Y | µ0 ∈ CJM,µ },
where
y ∗ (µ) = arg inf J obj (µ; y).
y∈Y

Clearly in this case YU B ⊂ Y. A simple enumeration exercise shows that

. N
αN (µ) = αU B (µ; CJ , M ) = min J obj (µ; y).
y∈YU B (µ;CJ ,M )

Proposition 2.10. Given CJ ⊂ P and M ∈ N,

N N
αU B (µ) ≥ α (µ) ∀µ ∈ P.

The only point left to clarify is how to build the set of parameters CJ . As specified in Algo-
rithm 2.9, a greedy approach is applied in order to iteratively enrich CJ , building an approxima-
tion of the coercivity constant which is more precise at each step of the procedure. As already
specified, the whole SCM approach can be efficiently divided into an offline and an online phase.

• Offline: We build, in a greedy fashion, the lower and upper bounds for the spaces. Here,
Jmax problems of the form (2.73) are performed, resulting in 2Qa + Jmax eigenproblems
over V N . This procedure allows us to build the box B and the set of coercivity constants
{αN (µ0 ) |µ0 ⊂ CJmax } which are the elements defining YLB . It still remains to construct
the upper bound YLB , which is given by Jmax Qa inner-product evaluations over V N .
 2.4. A Posteriori Error Estimation for Certified RB Methods 51

• Online: Given µ ∈ P, the set CJmax is sorted in order to find the M µ-nearest parameters
CJM,µ
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

max
through (M + 1)Qa evaluations of the map µ0 → θq (µ0 ) and the related M evalua-
tions of αN (µ0 ). In the end, the lower bound of the coercivity constant is computed using
(2.73), and its cost does not depend on N .

A LGORITHM 2.9. Greedy selection for CJ .

1: Ξtrain (SCM), SCM ∈ [0, 1], M ∈ N ! . Needed inputs
N N
αU B (µ; CJ ) − αLB (µ; CJ )
2: while max N (µ; C )
> SCM . Iterative selection procedure
µ∈Ξtrain αU B J
!
N N
α (µ; CJ ) − α (µ; C J )
3: µJ+1
SCM = arg max
UB LB
N (µ; C )
. New parameter selection
µ∈Ξtrain αU B J
4: CJ+1 = CJ ∪ µJ+1
SCM . CJ enrichment
5: J ← J + 1 . Update dimension RB
7: set J = Jmax
6: end

In this subsection, we focused on error estimators for linear coercive problems. However,
other results can be proved for several state equations. Next, we will extend the concept of the
error estimator to more complex problems, such as Stokes and parabolic equations.

2.4.4 Error Estimator for Stokes Equations

We now want to find a way to efficiently apply RB techniques to Stokes problem (2.13). To
achieve the goal of an explicit form for the error estimator, for this specific test case, we need the
following quantities:

• the Brezzi inf-sup constant β N (µ) and its lower bound βLB
N
(µ) defined over the continuity
equation as
. b(v, q; µ)
β N (µ) = inf sup N
≥ βLB (µ) > 0;
q∈Q v∈V kvkV kqkQ

• the Babuška inf-sup constant βBa

N
(µ) and its lower bound βBa,LB
N
(µ) defined over the
whole Stokes system as

N . a(u, v; µ) + b(v, p; µ) + b(u, q; µ) N

βBa (µ) = inf sup ≥ βBa,LB (µ) > 0;
(u,p)∈V ×Q (v,q)∈V ×Q k(u, p)kV ×Q k(v, q)kV ×Q

• the separated residuals with respect to the momentum and the continuity equations:

r1 (v; µ) = f (v; µ) − a(uN , v; µ) − b(v, pN ; µ), v ∈ V N ,

(2.75)
r2 (q; µ) = −b(uN , q; µ), q ∈ QN ;

• the global residual, which reads

r(v, q; µ) = r1 (v; µ) + r2 (q; µ).

All these ingredients combined will result in an error estimator for the Stokes problem. Indeed,
the following proposition holds [223].
 52 Chapter 2. Finite Element–Based Reduced Basis Method in CFD

Proposition 2.11. Define

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

kr(·; µ)k(V ×Q)0

∆Ba
N (µ) = N
. (2.76)
βBa,LB (µ)

Then
k(uN , pN )(µ) − (uN , pN )(µ)kV ×Q ≤ ∆Ba
N (µ). (2.77)

The proof follows from standard Banach–Nečas–Babuška theory [412], i.e., an extension of
Lax–Milgram techniques to noncoercive problems. We recall that the Stokes case is a special
case of noncoercive problems. However, the treatment of more complicated and general equa-
tions goes beyond the scope of this contribution. Focusing on Stokes equations, the estimator
∆BaN (µ) is typically used in the greedy algorithm and allows us to extend the techniques to a
very common and useful model in fluid dynamics. However, it gives no insight into the total
error, since it is divided among velocity and pressure components. Thus, an alternative can be
employed.
Let us define

• a lower bound αLB (µ) of the coercivity constant α(µ) of a(·, ·; µ), and

• an upper bound γU B (µ) of the continuity constant γ(µ) of b(·, ·; µ).

Proposition 2.12. Let us define

kr1 (·; µ)kV 0 γU B (µ) kr2 (·; µ)kQ0

 
∆uN (µ) = + 1+ ,
αLB (µ) αLB (µ) βLB (µ)
kr1 (·; µ)kV 0 γU B (µ) u
∆pN (µ) = + ∆ (µ).
βLB (µ) βLB (µ) N

Then
kuN (µ) − uN (µ)kV ≤ ∆uN (µ),
kpN (µ) − pN (µ)kQ ≤ ∆pN (µ).

We do not report the proof of this statement; however, once again, the interested reader may
refer to [223].

2.4.5 Error Estimator for Parabolic Problems

In this subsection, we present how to deal with time evolution using a greedy approach. The
problem at hand is of the form (2.52). To manage this parabolic setting, we introduce the residual

uN ,k N ,k−1
N (µ) − uN
 
(µ)
rk (v, µ) = g(tk )f (v, µ) − m , v; µ − a(uN ,k
N (µ), v; µ)
∆t
(2.78)
for all v ∈ V N and for all k ∈ K. Then, if we consider for each time instance
v
u k
u ∆t X
∆kN (µ) = t LB LB
(N (tj ; µ))2 ,
α (µ)σ (µ) j=1
 2.5. Nonaffine and Nonlinear Problems 53

it can be shown that [246]

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

||uN ,k (µ) − uN ,k k
N (µ)||L2 (Ω) ≤ ∆N (µ).

In this subsection we proposed an overview on a posteriori error bounds that are crucial quantities
in the efficient application of the greedy algorithm. Here we focus on their expression based,
essentially,

• on residual representations, and

• on lower bounds of stability factors (coercivity constants, inf-sup stability constants, . . . ).

This guarantees an efficient offline-online decomposition for several problems governed by dif-
ferent equations, such as

• elliptic problems,

• Stokes equations, and

• parabolic problems.

In the next section we address more complicated parametric problems. Indeed, many applica-
tions, from sciences to industry, rely on models which can be nonlinear or that might fail the
affine decomposition assumption (2.34). Thus, to tackle them, we present some advanced tech-
niques aimed at solving these more involved models in a faster way and adapting the classical
knowledge about ROMs to these frameworks.

2.5 Nonaffine and Nonlinear Problems

In this section, we will review the main techniques to be used when dealing with nonaffine and
nonlinear problems. As we have seen in the previous sections, a fundamental assumption to guar-
antee the offline-online efficiency of the method, and thus the independence of the online phase
from the degrees of freedom of the high-fidelity discretization, is that the following parameter
separability property holds for every parameter in the parameter space:

Q Qf
X X
a(u, v; µ) = θqa (µ)aq (u, v), F (v; µ) = θqf (µ)Fq (v).
q=1 q=1

This allows us to factor out any µ-dependent terms θq· (·) and the integrals aq (·, ·) or fq (·) on
the domain, so that, after the offline phase, a Galerkin projection of the parameter-independent
quantities aq (·, ·) or fq (·) can be performed, precomputing reduced order matrices AqN and vec-
tors FNq .
During the online stage, once given a specific value of µ, the corresponding reduced order
left-hand and right-hand sides can be obtained as

Q Qf
θqa (µ)AqN , θqf (µ)FNq .
X X
AN (µ) = FN (µ) =
q=1 q=1

Due to its importance, this form is usually assumed in many model reduction methods (e.g., the
separability requirements in PGD [123]).
 54 Chapter 2. Finite Element–Based Reduced Basis Method in CFD

As an example, we focus on the forcing term, which we can write as

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Z
F (v; µ) = g(x; µ)v(x) dΩ. (2.79)
Ω

It is clear that, if the integrand satisfies g(x; µ) = g1 (x)g2 (µ), we easily obtain an affine
(parameter-separable) term
Z
F (v; µ) = g2 (µ) g1 (x)v(x) dΩ. (2.80)
Ω

However, there are several relevant cases in which this assumption does not hold, such as non-
affine (non-parameter-separable) and nonlinear integrands, respectively:
Z
F (v; µ) = g(x; µ)v(x) dΩ for a generic g(x; µ) (2.81)
Ω

and Z
F (v; u(µ), µ) = g(x; u(x; µ), µ)v(x) dΩ. (2.82)
Ω
We remark that in some situations one can rely on a workaround to prevent the lack of affine
expansion. If we consider polynomial nonlinearities such as

g(x; u, µ) = g1 (x) up g2 (µ)

for some integer p ∈ N, we can resort to the (p + 1)-linear form

Z
t(u1 , u2 , . . . , up , v) = g1 (x) u1 u2 · · · up v,
Ω

which allows us to affine decouple the forcing term as

F (v; u(µ), µ) = g2 (µ)t(u(x; µ), u(x; µ), . . . , u(x; µ), v).

A similar formula holds for the nonlinear term of the Navier–Stokes system with p = 2 and
is commonly used to recover the computational efficiency. Let us describe this process in more
detail. Let j = [j1 , . . . , jp ] be a multi-index of size p, and ζi the ith RB function (RBF), where
each index i, j1 , . . . , jp varies in the range {1, . . . , N }. At the end of the offline stage, one can
preassemble the tensor of order p + 1, given by

Ti,j = t(ζj1 , ζj2 , . . . , ζjp , ζi ).

We remark that the tensor is dense, so it might be a challenge to store it in memory if p is very
large or if N is not small enough. Instead, during the online stage, as usual, we can expand the
reduced solution uN as
XN
uN (µ) = [uN ]j ζj
j=1

and evaluate the integrand F at the reduced solution uN as

P 
N PN
F (v; uN (µ), µ) = g2 (µ)t j1 =1 [u N ]j 1
ζ j 1
, . . . , jp =1 [uN ]j p
ζ j p
, v
PN PN
= g2 (µ) j1 =1 [uN ]j1 . . . jp =1 [uN ]jp t(ζj1 , . . . , ζjp , v) (2.83)
PN PN
= g2 (µ) j1 =1 [uN ]j1 . . . jp =1 [uN ]jp Ti,j ,
 2.5. Nonaffine and Nonlinear Problems 55

where we have exploited the multilinearity of t in the first p arguments and then the multilinearity
of t in its (p + 1)th argument to choose v = ζi for any i = 1, . . . , N . Finally, we can simply
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

recover the offline-online decomposition by using the (p + 1)th order tensor Ti,j .
Even if the case of polynomial nonlinearities occurs frequently (e.g., incompressible fluid
dynamics, optimal control problems) and allows us to obtain an exact parameter-separable rep-
resentation, it is certainly not the most general one.
Therefore, in the general case, we will resort to approximate parameter-separable represen-
tations such as
Z Qf
X
F (v; µ) = g(x; µ)v(x) dΩ ≈ θqf (µ)Fq (v).
Ω q=1

Building an approximation like the one above is an active field of research in the ROM com-
munity, which goes under the name “hyperreduction.” One of the most widespread of these
hyperreduction techniques is the empirical interpolation method (EIM) [40], which allows us to
obtain an efficient offline-online decoupling when dealing with nonaffine and nonlinear systems.
In the next subsection, we will give further details on this affine recovery technique.

2.5.1 The EIM

As discussed earlier, when dealing with nonaffine and nonlinear problems, one loses the ef-
ficiency of the ROM based on the offline-online decomposition. The EIM is a general tool,
proposed in [40], for approximating a parametrized or nonlinear function g by a sum of affine
terms as follows:
XQ
g(x; µ) ≈ cq (x)hq (µ). (2.84)
q=1

Once this representation is known, one can easily obtain an affine representation of the integral
form Z
F (v; µ) = g(x; µ)v(x) dΩ, (2.85)
Ω
i.e.,
Q
X Z
F (v; µ) = hq (µ)F q (v), where F q (v) = cq (x)v(x) dx. (2.86)
q=1 Ω

Of course, the same decomposition can be formally extended to the nonlinear case with

g(x; µ) := ge(x; u(x; µ), µ). (2.87)

In particular, the EIM approximation is based on an interpolation operator IQ g that interpo-

lates the given function g in a set of interpolation points x1 , . . . , xQ ∈ Ω. This interpolant is
constructed as a linear combination of hierarchically chosen global basis functions h1 , . . . , hQ .
In contrast with other interpolation methods, which usually work with generic and multipurpose
basis functions such as polynomial functions, the EIM works with problem-specific basis func-
tions defined in the whole domain and selected hierarchically. Having obtained the global basis
functions, the interpolant can be then expressed by
Q
X
IQ g(x; µ) = cq (µ)hq (x), x ∈ Ω, (2.88)
q=1

where cq are parameter-dependent coefficients and the basis functions hq are selected as a linear
combination of snapshots of the parametrized functions g(µ1 ), . . . , g(µQ ).
 56 Chapter 2. Finite Element–Based Reduced Basis Method in CFD

This means that, once the basis functions hq (x) are set, the problem of finding the coefficients
c (µ) is solved by imposing the interpolation condition, i.e.,
q
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Q
X
IQ g(xq ; µ) = cq (µ)hq (xq ) = g(xq ; µ), q = 1, . . . , Q. (2.89)
q=1

We remark that the above problem can be recast in matrix form as T c(µ) = g(µ), with

Tij = hj (xi ), (c(µ))j = cj (µ), (g(µ))j = g(xi ; µ), i, j = 1, . . . , Q, (2.90)

where the interpolation points xi are selected using a greedy approach.

We now present Algorithm 2.10, which implements the EIM in its abstract version.

A LGORITHM 2.10. EIM.

INPUT: Set of parametrized functions f (x; µ) : Ω × P → R, tolerance ε, maximum number of
basis functions Nmax , p order of the chosen p-norm
OUTPUT: basis functions {h1 , . . . , hQ }, interpolation points {x1 , . . . , xQ }

1: k = 1
2: I0 g(x; µ) = 0 . Initialization
3: while (k < Nmax ) . Build space hierarchy
4: µk = arg max kg(·; µ) − Ik−1 g(·; µ)kL∞ (Ω) . Greedy parameter selection
µ∈P
5: r(x) = g(x; µk ) − Ik−1 g(x; µk ) . Residual assembly
6: xk = arg max |r(x)| . Find interpolation point
x∈Ω
7: hk (x) = r(x)/r(xk ) . Residual normalization
8: εk = krkL∞ (Ω) = r(xk ) . Tolerance threshold computation
9: if (εk < ε) . Stopping criterion
10: break
11: else
12: k =k+1

Finally, in order to obtain a practical version of the EIM, it is common to replace the param-
eter space P with the discrete training set Ptrain . In the same spirit, the L∞ (Ω) computations
(three times) are replaced by evaluations of maximums over a discrete set (e.g., mesh nodes or
quadrature points).
Moreover, g(·; µ) is usually precomputed for every µ ∈ Ptrain , since this evaluation does
not change with the iteration counter k, while the right-hand sides of interpolation conditions are
usually computed on a reduced mesh that contains only the patches of xk .
Let us remark that the greedy computation of basis functions hk may be replaced by a POD
compression of snapshots g(µ1 ), . . . , g(µQ ), similarly to RB generation.
The algorithm can easily be adapted to the discrete case, where g is not an analytical function
but the result of the discretization of a linear function (vector) or a bilinear form (matrix). The
role of x will be played by an index (vector) or a pair of indices (matrix). The resulting method
is known in the literature as the discrete empirical interpolation method [116].

2.5.2 Model Problem: Elliptic Problem with Gaussian Source

As an example, we may consider the steady heat conduction problem in a two-dimensional square
domain Ω = (−1, 1)2 . The boundary ∂Ω is kept at a reference temperature (say, zero). The
 2.5. Nonaffine and Nonlinear Problems 57
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 2.15. Left: RB temperature solution. Right: computational mesh for Ω = (−1, 1)2 with
EIM interpolation points.

conductivity coefficient is fixed at one, while the heat source is characterized by the expression

g(x; µ) = exp{−2(x0 − µ0 )2 − 2(x1 − µ1 )2 } ∀x = (x0 , x1 ) ∈ Ω.

The parameter vector µ, given by µ = (µ0 , µ1 ) ∈ [−1, 1]2 , affects the center of the Gaussian
source g(x; µ), which could be located at any point in Ω. In order to more quickly (and provably
accurately) evaluate the problem, we propose to use a certified RB approximation for the prob-
lem. In order to preserve the affinity assumption (for the sake of performance), the EIM will be
used on the forcing term g(x; µ).
Thus, let u(µ) be the temperature in the domain Ω. The weak formulation for this problem
is

a (u(µ), v; µ) = f (v; µ) ∀v ∈ V,

where the function space is defined as V = v ∈ H 1 (Ω : v|∂Ω = 0 . The parametrized bilinear



form a(·, ·; µ) : V × V → R and linear form f (·; µ) : V → R are defined respectively by

Z
a(u, v; µ) = ∇u · ∇v dx
Ω

and
Z
f (v; µ) = g(µ)v dx.
Ω

In Figure 2.15 we show the computational mesh used for the high-fidelity simulation, the
solution corresponding to a particular value in the parameter space P, and the interpolation points
selected by the EIM.
Finally, we show in Figure 2.16 the behavior of the average relative error with respect to
the number of EIM and RB functions. As we can see, even in the nonaffine case we obtain
exponential convergence of the RB method.
 58 Chapter 2. Finite Element–Based Reduced Basis Method in CFD
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 2.16. Convergence of the RB method for the elliptic problem with Gaussian source, with
respect to the number of basis functions, for EIM (left) and RB (right).
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Chapter 3

Certified Smagorinsky
Reduced Basis
Turbulence Model

Enrique Delgado Ávila, Francesco Ballarin, Gianluigi Rozza

3.1 Introduction
In this chapter we start by presenting a reduced basis (RB) Smagorinsky turbulence model (cf.
[541]) for steady flows, which is the basic large eddy simulation (LES) turbulence model, in
which the effect of the subgrid scales on the resolved scales is modeled by eddy diffusion terms
(cf. [111, 515]). It is an intrinsically discrete model, since the eddy viscosity term depends
on the mesh size. Although it is well known that the Smagorinsky model is overdiffusive, the
construction of this RB Smagorinsky turbulence model represents a first step that is complex
enough and essential for the RB models presented later on.
We start by defining the finite element (FE) problem, and then, by properly selecting snap-
shots with the greedy algorithm, we obtain the low-dimensional spaces for which we define the
Smagorinsky RB model. The eddy viscosity is modeled by a nonlinear term that depends on the
mesh size and the modulus of the velocity gradient. We approximate this nonlinear term using the
empirical interpolation method (EIM; cf. [40, 245]) in order to obtain a linearized decomposition
of the RB Smagorinsky model.
Then, we present an RB Boussinesq model, with variational multiscale (VMS)-Smagorinsky
modeling for both the eddy viscosity and the eddy diffusivity. This model takes into account
the buoyancy forces present in natural convection problems. This model differs from the ones
presented in Section 3.2 because it considers the energy equation coupled with the momentum
and continuous equations. In this section, we assume that the temperature of the fluid becomes
essential in the flow of the fluid considered. We perform the RB model of Boussinesq VMS-
Smagorinsky equations with the same strategy presented in Section 3.2, developing an a poste-
riori error bound, which is essential for the greedy algorithm used in snapshot selection. Again,
we take into account the EIM in order to approximate the nonlinear terms for the eddy viscosity
and eddy diffusivity.
Finally, we consider the application of the RB Boussinesq VMS-Smagorinsky in a variable-
height cavity domain. The variability of the cavity height is considered through a geometrical
parametrization on the domain. Since we are interested in efficiently solving the parameter-
dependent problem, we need to reformulate the Boussinesq VMS-Smagorinsky model in a pa-
rameter-independent domain with a change of variables. This change of variables leads us to

59
 60 Chapter 3. Certified Smagorinsky Reduced Basis Turbulence Model

obtain operators that depend on the geometrical parameter in the same way that they depend on
the physical parameters. This setting lets us decompose affinely the operators with respect to the
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

parameters, both physical and geometrical, letting us store parameter-independent matrices and
tensors in the offline phase.
The chapter is structured as follows: in Section 3.2 we first present the Smagorinsky RB
model, and Subsection 3.2.1 is devoted to the FE problem. Then, in Subsection 3.2.2, we present
the greedy algorithm used to construct the RB spaces, whereas in Subsection 3.2.3 we describe
the EIM for the approximation of the nonlinear terms. To finish this first section, we present
the numerical analysis for the well-posedness of the FE problem in Subsection 3.2.4 with the
construction of an a posteriori error estimator in Subsection 3.2.5 and the numerical test for
this problem in Subsection 3.2.6. Then, in Section 3.3, we introduce the VMS-Smagorinsky
model, with the FE model in Subsection 3.3.1 and the RB model in Subsection 3.3.2. Again, we
present the well-posedness analysis (Subsection 3.3.3), the construction of the a posteriori error
estimator (Subsection 3.3.4), and the numerical results (Subsection 3.3.5). Finally in Section 3.4,
we describe the geometrical parametrization of the VMS-Smagorinsky model, with the problem
setting in Subsection 3.4.1 and the numerical tests in Subsection 3.4.2.

3.2 RB Smagorinsky Turbulence Model

In this section we present the Smagorinsky turbulence model, which is the basic LES turbulence
model, in which the effect of the subgrid scales on the resolved scales is modeled by eddy diffu-
sion terms (see [111, 515, 541]). We first present the FE problem, and then we derive the RB one,
by constructing low-dimensional spaces from snapshots given by the FE problem throughout the
greedy algorithm. The proofs of the main results of this section can be found in [107, 151].

3.2.1 FE Problem
To formulate the Smagorinsky turbulence model, let Ω be a bounded polyhedral domain in
Rd , (d = 2, 3). We assume that its boundary is split into Γ = ΓD ∪ ΓN , where ΓD = ΓDg ∪ ΓD0
is the boundary relative to the homogeneous and nonhomogeneous Dirichlet boundary conditions
and ΓN is the boundary relative to the Neumann conditions.
Let {Th }h>0 a family of affine-equivalent and conforming triangulations of Ω, formed by
triangles or quadrilaterals (d = 2) or tetrahedra or hexahedra (d = 3). As usual, the parameter
h is the maximum diameter hK among the elements K ∈ Th . Given an integer l ≥ 0 and an
element K ∈ Th , we denote by Rl (K) either Pl (K), the space of Lagrange polynomials of
degree ≤ l, defined on K, if the grids are formed by triangles (d = 2) or tetrahedra (d = 3),
or Ql , the space of Lagrange polynomials of degree ≤ l on each variable, defined on K, if the
family of triangulations is formed by quadrilaterals (d = 2) or hexahedra (d = 3).
Although the Smagorinsky model is intrinsically discrete, it can be interpreted as a discretiza-
tion of a continuous model. We next present this model to clarify its relationship with the Navier–
Stokes one. In this way, the “continuous” Smagorinsky turbulence model is formulated as
   
1
w · ∇w + ∇p − ∇ · + ν (w) ∇w = f in Ω,

T

Re





in Ω,

 ∇·w=0



w = gD on ΓDg , (3.1)
on ΓD0 ,

w=0





  
 1 ∂w
on ΓN ,

 −pn + + νT (w) =0


Re ∂n
 3.2. RB Smagorinsky Turbulence Model 61

where here Re is the Reynolds number, w is the velocity field, and p is the pressure, with the
latter two depending on the Reynolds number. Moreover, the eddy diffusion term is given by
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

X
νT (w) = CS2 h2K ∇w|K χK , (3.2)
K∈Th

where · denotes the Frobenius norm in Rd×d and CS is the Smagorinsky constant (cf. [512]).
For this problem, the parameter considered is the Reynolds number, denoted in the following
by µ ∈ D.
Let us consider the spaces Y = {v ∈ H 1 (Ω) : v|ΓD = 0} for velocity and M = L2 (Ω) for
pressure. We assume that there exists a lift function uD ∈ (H 1 (Ω))d such that uD |ΓDg = gD ,
uD |ΓD0 = 0, and ∇ · uD = 0 in Ω. With these conditions, we ensure that the reduced velocity
u = w − uD is still incompressible and satisfies the homogeneous Dirichlet boundary conditions
on ΓD . We will assume that f ∈ (L2 (Ω))d and gD ∈ (H 1/2 (Ω))d .
Let Yh ⊂ Y and Mh ⊂ M be two finite subspaces of Y and M . We consider the following
variational discretization of problem (3.1), actually, the “true” Smagorinsky model:

 Find (uh , ph ) = (uh (µ), ph (µ)) ∈ Yh × Mh such that






 a(uh , vh ; µ) + b(vh , ph ; µ) + aS (wh ; wh , vh ; µ)

+ c(uh , uh , vh ; µ) + c(uD , uh , vh ; µ) (3.3)

+ c(uh , uD , vh ; µ) = F (vh ; µ) ∀vh ∈ Yh ,





b(uh , qh ; µ) = 0 ∀qh ∈ Mh ,


where wh = uh + uD . The bilinear forms a(·, ·; µ) and b(·, ·; µ) are defined by

1
Z Z
a(u, v; µ) = ∇u : ∇v dΩ, b(v, q; µ) = − (∇ · v)q dΩ; (3.4)
µ Ω Ω

the trilinear form c(·, ·, ·; µ) and the nonlinear Smagorinsky term aS (·; ·, ·; µ) are given by
Z Z
c(z, u, v; µ) = (z · ∇u)v dΩ, aS (z; u, v; µ) = νT (z)∇u : ∇v dΩ. (3.5)
Ω Ω

Finally, the linear form F (·; µ) is defined by

F (v; µ) = hf, vi − a(uD , v; µ) − c(uD , uD , v; µ),

where h·, ·i stands for the duality pairing between Y 0 and Y , with Y 0 being the dual space of Y .
The solution of problem (3.3) is intended to approximate the large-scale component of the
solution of the Navier–Stokes problem (i.e., problem (3.1) with νT = 0).
Let us define the norms relative to the spaces Y and M . For the velocity space Y , we consider
a weighted inner product (·, ·)T defined as
Z  
1
(u, v)T = + νT∗ ∇u : ∇v dΩ ∀u, v ∈ Y, (3.6)
Ω µ

where νT∗ = νT (w(µ)), µ = arg minµ∈D K∈Th (CS hK )2 minx∈K |∇w(µ)|(x)χK (x), and
P
w(µ) is the velocity solution of (3.1).
1/2
This inner product induces a norm linked to the eddy diffusion term k · kT = (·, ·)T . Since
the functions of Y vanish on ΓD , this norm is equivalent to the usual H -norm. This norm will
1

turn out to be crucial to apply our error estimator in the RB construction by the greedy algorithm.
For the pressure space M , we will use the usual L2 -norm.
 62 Chapter 3. Certified Smagorinsky Reduced Basis Turbulence Model

For simplicity of notation, let us denote by X the product space X = Y × M and, by

extension, Xh = Yh × Mh ⊂ X. We also define the X-norm as
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

q
kU kX = kuk2T + kpu k20,2,Ω ∀U = (u, pu ) ∈ X. (3.7)

With this notation, we can rewrite the variational problem (3.3) as follows:

Find Uh (µ) ∈ Xh such that

(
(3.8)
A(Uh (µ), Vh ; µ) = F (Vh ; µ) ∀Vh ∈ Xh .

In this formulation, the operator A is given by

1
A(Uh , Vh ; µ) = A0 (Uh , Vh ) + A1 (Uh , Vh ) + A2 (Uh ; Vh ) + A3 (Uh ; Vh ), (3.9)
µ
where we define Vh = (vh , pvh ), and
Z
A0 (U, V ) = ∇u : ∇v dΩ,
ZΩ Z
A1 (U, V ) = (∇ · u)pv dΩ − (∇ · v)pu dΩ
Ω
Z ΩZ

+ (uD · ∇u)v dΩ + (u · ∇uD )v dΩ,

Z Ω Ω

A2 (U ; V ) = (u · ∇u)v dΩ,
ZΩ
A3 (U ; V ) = νT (u + uD ) ∇(u + uD ) : ∇v dΩ.
Ω

3.2.2 Greedy Algorithm

To select the snapshots to construct the reduced spaces, we use the greedy algorithm. To start the
greedy algorithm, we randomly select an initial parameter value µ1 ∈ Dtrain , where Dtrain ⊂ D
is a discrete set with the parameter values taken into account in the greedy algorithm. Initially, we
define the set of parameter values S1 = {µ1 }, the first reduced space M1 = span{ξkp := ph (µ1 )},
and Y1 = {ζ2k−1v
:= uh (µ1 ), ζ2k
v
:= Tpµ ξ1p }, where here we are denoting by Tpµ ξ1p the inner
pressure supremizer that assures the inf-sup stability of the RB problem (see [471] for more
details).
Then, to add a new snapshot to the reduced space, we take the parameter value that gives the
worst RB approximation with respect to the FE one, i.e., we choose µN ∈ Dtrain such that

µN = arg max kUh (µ) − UN (µ)k. (3.10)

µ∈Dtrain

In practice, the error kUh (µ) − UN (µ)k may be hard to compute numerically because we
have to compute the solution of the FE problem, uh (µ), for all µ ∈ Dtrain . Thus, rather than
computing the exact error, we consider an a posteriori error bound, ∆N (µ), that is cheaper to
compute than the exact error.
The algorithm resulting from the substitution of the real error by the a posteriori error bound
is usually called the weak greedy algorithm (cf. [60, 470]) and is the one that is commonly used
for the RB method.
In each iteration of the (weak) greedy algorithm, we actualize the set of parameter values
SN = {µ1 , . . . , µN } and the reduced spaces MN = span{ξkp := puh (µk ), k = 1, . . . , N } and
 3.2. RB Smagorinsky Turbulence Model 63

YN = span{ζ2k−1 v
:= uh (µk ), ζ2kv
:= Tpµ ξkp , k = 1, . . . , N } for N = 1, . . . , Nmax , with
Nmax the maximum number of bases that we consider in our problem. With this procedure, we
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

construct hierarchical spaces, i.e., M1 ⊂ M2 ⊂ . . . ⊂ MN , and Y1 ⊂ Y2 ⊂ . . . ⊂ YN .

In each step of the greedy algorithm, we orthonormalize, with respect to the norm in Xh ,
the resulting reduced spaces by the Gram–Schmidt orthonormalization process. This orthonor-
malization process becomes essential, since it avoids large condition numbers for the reduced
problem matrix.
We summarize the greedy algorithm for a general RB problem:
1. Set µ1 , compute Uh (µ1 ) to solve problem (3.8), and define the reduced spaces M1 and Y1 .
Orthonormalize the spaces M1 and Y1 .
2. For k ≥ 2, compute ∆k−1 (µ) ∀µ ∈ Dtrain and set µk = arg max ∆k−1 (µ).
µ∈Dtrain

3. Compute Uh (µk ), define the reduced spaces Mk and Yk by adding the newly computed
snapshots, and orthonormalize them.
4. Stop if max ∆k (µ) < εRB . If not, go back to step 2.
µ∈Dtrain

3.2.3 Approximation of the Eddy Viscosity Term

To approximate the nonlinear turbulent eddy viscosity term, we use the EIM (cf. [40, 245]).
Let us write g(µ) := g(x; wh (µ)) = |∇wh (µ)|(x). The last step in using the EIM in the
Smagorinsky RB model is decoupling the parameter dependence from the spatial dependence of
the function g(µ), i.e.,
g(µ) ≈ IM [g(µ)]. (3.11)
With the EIM, we construct an RB space WM = span{q1 (µ), . . . , qM (µ)}, selecting these
basis functions by a greedy procedure, with snapshots of g(µ). Thanks to this, we are able to
approximate the nonlinear Smagorinsky term by a trilinear form, in the following way:
aS (wN ; wN , vN ; µ) ≈ âS (wN , vN ; µ), (3.12)
where
M
X
âS (wN ; vN ; µ) = σk (µ)s(qk , wh , vh ), (3.13)
k=1
with
X Z
s(qk , wh , vh ) = (CS hK )2 qk ∇w : ∇v dΩ. (3.14)
K∈Th K

Here σk (µ) for k = 1, . . . , M is the solution of a lower-triangular linear system, where the
second member is the value of g(x; wh (µ)) at certain points xi . Furthermore,
X Z
âS (qk , wN , vN ) = (CS hK )2 qk ∇w : ∇v dΩ. (3.15)
K∈Th K

This technique allows us to linearize the eddy viscosity term. Thus, the RB problem with this
last approximation of the Smagorinsky term is defined as follows:
Find (uN , pN ) ∈ YN × MN such that






 a(uN , vN ; µ) + b(vN , pN ; µ) + âS (wN ; vN ; µ)

+ c(uD , uN , vN ; µ) + c(uN , uD , vN ; µ) (3.16)

+ c(u , u , v ; µ) = F (v ; µ) ∀v ∈ Y ,

N N N N N N




b(uN , qN ; µ) = 0 ∀qN ∈ MN .

 64 Chapter 3. Certified Smagorinsky Reduced Basis Turbulence Model

The solution (uN (µ), pN (µ)) ∈ XN of (3.16) can be expressed as a linear combination of the
basis functions
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

2N N
p
X X
uN (µ) = uN v
j (µ)ζj , pN (µ) = pN
j (µ)ξj .
j=1 j=1

Taking into account this representation, for the bilinear terms in (3.16), during the offline phase
we store the parameter-independent matrices, as in [387], defined as

(AN )ij = a(ζjv , ζiv ), i, j = 1, . . . , 2N,

(DN )ij = c(uD , ζjv , ζiv ) + c(ζjv , uD , ζiv ), i, j = 1, . . . , 2N, (3.17)
(BN )li = b(ζiv , ξlp ), i = 1, . . . , 2N, l = 1, . . . , N.

For the convective and Smagorinsky terms, we need to store a parameter-independent tensor
of order three during the offline phase, defined as

(CN (ζsv ))ij = c(ζsv , ζjv , ζiv ), i, j, s = 1, . . . , 2N,

(3.18)
(SN (qs ))ij = s(qs , ζjv , ζiv ), i, j = 1, . . . , 2N, s = 1, . . . , M.

With this tensor representation, it holds that

2N
X M
X
c(uN , ζjv , ζiv ; µ) = uN
s (µ)CN (ζs ) and
v
âS (ζjv ; ζiv ; µ) = σs (µ)SN (qs ).
s=1 s=1

Thanks to this, we are able to solve problem (3.16), linearized by a semi-implicit evolution
approach. Note that the treatment of the approximation of the eddy viscosity term in the offline-
online phase is similar to the treatment of the convective term, as a consequence of the tensoriza-
tion.

3.2.4 Well-Posedness Analysis

The well-posedness of the Smagorinsky problem is provided in [111] by the classical Brezzi
theory (cf. [74]). The boundedness of the FE solution is provided by this analysis. However,
in this section, we analyze the well-posedness of the Smagorinsky FE solution using the more
general Brezzi–Rappaz–Raviart (BRR) theory (see, e.g., [76]). The last step in using the BRR
theory instead of the Brezzi theory is to construct the error estimator provided by the BRR for
the RB problem.
Taking the directional derivative, we derive each operator term in (3.9), obtaining

∂1 A0 (U, V )(Z) = A0 (Z, V ),

∂1 A1 (U, V )(Z) = A1 (Z, V ),
Z Z
∂1 A2 (U ; V )(Z) = (u · ∇z)v dΩ + (z · ∇u)v dΩ,
Ω Ω
Z
∂1 A3 (U ; V )(Z) = νT (u + uD ) ∇z : ∇v dΩ
Ω Z
X ∇(u + uD ) : ∇z
(CS hK )2


+ ∇(u + uD ) : ∇v dΩ.
K |∇(u + uD )|
K∈T

For the well-posedness of the problem, we have to guarantee the uniform coerciveness and
boundedness of ∂1 A in the sense that, for any solution Uh (µ) of (3.8), there exist β0 > 0 and
 3.2. RB Smagorinsky Turbulence Model 65

γ0 ∈ R such that, ∀µ ∈ D,
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

∂1 A(Uh (µ), Vh ; µ)(Zh )

0 < β0 < βh (µ) ≡ inf sup ,
Zh ∈Xh Vh ∈Xh kZh kX kVh kX
(3.19)
∂1 A(Uh (µ), Vh ; µ)(Zh )
∞ > γ0 > γh (µ) ≡ sup sup .
Zh ∈Xh Vh ∈Xh kZh kX kVh kX

Then, according to the BRR theory (cf. [76, 93]), it will follow that in a neighborhood of Uh (µ)
the solution of (3.8) is unique and bounded in k · kX in terms of the data.
We have the following results.

Proposition 3.1. There exists γ0 ∈ R such that ∀µ ∈ D,

|∂1 A(Uh (µ), Vh ; µ)(Zh )| ≤ γ0 kZh kX kVh kX ∀Zh , Vh ∈ Xh .

Proposition 3.2. Let C ? = CT2 (Cµ + 1). Suppose that k∇uD k0,2,Ω < 1
C? and k∇uh k0,2,Ω ≤
˜
C ? − k∇uD k0,2,Ω . Then, there exists βh > 0 such that
1

∂1 A(Uh , Vh ; µ)(Vh ) ≥ β˜h kvh k2T ∀Vh ∈ Xh . (3.20)

Since the operator b(vh , qh ; µ) satisfies the discrete inf-sup condition

b(vh , ph ; µ)
αkqh k0,2,Ω ≤ sup ,
vh ∈Yh kvk1,2,Ω

thanks to Proposition 3.2, we can prove that the operator ∂1 A satisfies the inf-sup condition in
(3.19). See [106] for more details.
Observe, also, that since kgD k1/2,ΓD ≤ kwh k1,Ω ≤ CΩ k∇wh k0,2,Ω , the condition needed
CΩ
in Proposition 3.2, k∇wh k0,2,Ω ≤ C 2 (C1µ +1) , will only be possible if kgD k1/2,ΓD ≤ C 2 (Cµ +1)
;
T T
thus the Dirichlet boundary data should be sufficiently small.

3.2.5 A Posteriori Error Estimator

In this subsection we construct the a posteriori error bound estimator for the greedy algorithm,
which selects the snapshots for the reduced space XN . In order to obtain this a posteriori error
bound estimator, we take into account the well-posedness analysis of the reduced problem (3.8)
from the previous section.
We start by proving that the directional derivative of the operator A(·, ·; µ) is locally Lip-
schitz.

Lemma 3.3. There exists a positive constant ρT such that

∂1 A(Uh1 , Vh ; µ)(Zh ) − ∂1 A(Uh2 , Vh ; µ)(Zh ) ≤ ρT kUh1 − Uh2 kX kZh kX kVh kX (3.21)

for all Uh1 , Uh2 , Zh , Vh ∈ Xh .

We introduce the following supremizer operator TN : Xh → Xh , defined as

(TN Zh , Vh )X = ∂1 A(UN (µ), Vh ; µ)(Zh ) ∀Vh , Zh ∈ Xh , (3.22)

 66 Chapter 3. Certified Smagorinsky Reduced Basis Turbulence Model

such that
∂1 A(UN (µ), Vh ; µ)(Zh )
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

TN Zh = arg sup . (3.23)

Vh ∈Xh kVh kX
Taking into account this definition, in order to guarantee the well-posedness of the RB problem
(3.16) in the same way as in the FE problem (3.8), we define the inf-sup and continuity constants:

∂1 A(UN (µ), Vh ; µ)(Zh ) kTN Zh kX

0 < βN (µ) ≡ inf sup = inf , (3.24)
Zh ∈Xh Vh ∈Xh kZh kX kVh kX Zh ∈Xh kZh kX

∂1 A(UN (µ), Vh ; µ)(Zh ) kTN Zh kX

∞ > γN (µ) ≡ sup sup = sup . (3.25)
Zh ∈Xh Vh ∈Xh kZh kX kVh kX Zh ∈Xh kZh kX

Theorem 3.4. Let µ ∈ D, and assume that βN (µ) > 0. If problem (3.8) admits a solution Uh (µ)
such that
βN (µ)
kUh (µ) − UN (µ)kX ≤ ,
ρT
 
then this solution is unique in the ball BX UN (µ), βNρT(µ) .

At this point, we are in a position to define the a posteriori error bound estimator by

βN (µ) h p i
∆N (µ) = 1 − 1 − τN (µ) , (3.26)
2ρT

where τN (µ) is given by

N (µ) = kR(UN (µ); µ)kX 0 ,
4N (µ)ρT (3.27)
τN (µ) = 2 (µ) .
βN
The suitability of this a posteriori error bound estimator is stated by the following theorem.

Theorem 3.5. Assume that βN (µ) > 0 and τN (µ) ≤ 1 for all µ ∈ D. Then there exists a
unique solution Uh (µ) of (3.8) such that the error with respect to UN (µ), the solution of (3.16),
is bounded by the a posteriori error bound estimator, i.e.,

kUh (µ) − UN (µ)kX ≤ ∆N (µ), (3.28)

with effectivity  
2γN (µ)
∆N (µ) ≤ + τN (µ) kUh (µ) − UN (µ)kX . (3.29)
βN (µ)

3.2.6 Numerical Results for the Smagorinsky Model

In this test, we apply the reduced-order Smagorinsky turbulence model to the lid-driven cavity
problem. The numerical results were obtained using FreeFrem++ (cf. [268]).
In Figure 3.1, we represent the geometry with the boundaries described. We consider that the
nonhomogeneous Dirichlet boundary condition is given by g(x) = 1 on the lid boundary ΓDg .
For this test, we consider the Reynolds number as a parameter, ranging in the interval [1000,
5100]. The two-dimensional lid-driven cavity flow has a steady solution up to Reynolds 7500 (cf.
[111]); thus in this range, a steady solution is well known to exist. We consider the Taylor–Hood
pair FE, and we use a regular mesh with 5000 triangles and 2601 nodes.
 3.2. RB Smagorinsky Turbulence Model 67
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

ΓDg

ΓD0 ΓD0

ΓD0

Figure 3.1. Domain Ω with the different boundaries identified.

10 -1
max kg(µ) − IM [g(µ)]k∞
µ∈D

-2
10

10 -3

-4
10
0 5 10 15 20 25
M

Figure 3.2. Convergence of the EIM algorithm.

In this numerical test, we need Mmax = 22 bases in the EIM algorithm until it reaches the
tolerance for the relative error of εEIM = 5 · 10−4 . In Figure 3.2, we show the convergence of
the EIM algorithm.
For the greedy algorithm, we prescribe a tolerance of εRB = 5 · 10−5 . This tolerance is
reached for Nmax = 12 basis functions. Note that, in this case, N = 8 bases are needed in order
to ensure that τN (µ) < 1 for all µ in D. In Figure 3.3 (left) we show the convergence of the
greedy algorithm, and in Figure 3.3 (right) we show the value of the error and the a posteriori
error bound for all µ in D.
In Figure 3.4 we compare the FE velocity solution (left) and the RB velocity solution (right)
for a chosen parameter value µ = 4521. Both images are practically equal, as the error between
the solutions is of order 10−7 .
Finally, we show in Table 3.1 a summary of the results obtained for several values of µ in D.
For this test, we observe a dramatic speedup in the computation of the numerical solution. These
large speedup factors are possibly due to the highly turbulent levels of viscosity introduced by
the Smagorinsky turbulence model. The offline phase of this test took approximately two days
to complete.
 68 Chapter 3. Certified Smagorinsky Reduced Basis Turbulence Model
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

10 6 10 -3
max τN (µ) ∆N (µ)
µ∈D
kUh (µ) − UN (µ)kX
10
4 max ∆N (µ) 10 -4
µ∈D

-5
10
2
10

10 -6
0
10
-7
10

-2
10
-8
10

-4
10
10 -9

10 -6 10 -10
1 2 3 4 5 6 7 8 9 10 11 12 1000 1500 2000 2500 3000 3500 4000 4500 5000
N Re

Figure 3.3. Convergence of the greedy algorithm (left) and the value of ∆Nmax (µ) and the exact
error between the FE solution and the RB solution (right).

Figure 3.4. FE solution (left) and RB solution (right) for µ = 4521.

Table 3.1. Computational time for FE solution and RB online phase, with the speedup and the
relative error.

Data µ = 1610 µ = 2751 µ = 3886 µ = 4521 µ = 5100

TF E 638.02 s 1027.62 s 1369.49 s 1583.08 s 1699.52 s
Tonline 0.47 s 0.47 s 0.47 s 0.49 s 0.52 s
speedup 1349 2182 2899 3243 3227
kuh − uN kT 1.91 · 10−6 1.87 · 10−6 3.28 · 10−6 6.26 · 10−7 3.17 · 10−9
kph − pN k0 1.18 · 10−7 3.65 · 10−7 3.78 · 10−7 8.34 · 10−8 1.88 · 10−9
 3.3. RB VMS-Boussinesq Model 69

3.3 RB VMS-Boussinesq Model

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

In this section we present a Boussinesq model for natural convection flow problems. We deal
with a square cavity problem, for which the top and the bottom are isolated walls and the vertical
walls have a prescribed temperature. The difference in temperatures in the vertical wall produces
a circulation of the fluid, which is supposed to be viscous, Newtonian, and incompressible, and
the buoyancy forces are modeled with a Boussinesq approximation. The proof of the main results
of this section can be found in [30, 151].

3.3.1 FE Problem
We next present the Navier–Stokes equations with the Boussinesq approximation for the energy
equation (cf. [52, 66, 67, 109]), for which we also include the modeling of the eddy viscosity
and the eddy diffusivity by the Smagorinsky approach (see [512, 541]). We present a continuous
form of the Smagorinsky eddy viscosity and eddy diffusivity:



 u · ∇u − P r∆u − ∇ · (νT (u)∇u) + ∇p − P r Ra θ ed = f in Ω,
in Ω,

∇·u=0




in Ω,

 u · ∇θ − ∆θ − ∇ · (KT (u)∇θ) = Q

(3.30)

 u=0 on Γ,
on ΓD ,

 θ = θD



on ΓN .

 ∂n θ = 0


Here, u is the velocity field, p is the pressure, and θ is the temperature. In addition, ed
is the last vector of the canonical basis of Rd , while Ra and P r are the Rayleigh and Prandtl
dimensionless numbers, respectively. Both the external body forces f and the heat source term Q
are given data for the problem. In (3.30), νT (u) is the eddy viscosity term defined in (3.2), and
KT (u) is the eddy diffusivity, given by
1
KT (u) = νT (u). (3.31)
Pr
The Prandtl number is defined as
ν0
P r := ,
k0
with ν0 a reference kinematic viscosity and k0 a reference thermal diffusivity. Since the Prandtl
number depends on the physics of the fluid, we consider it as a fixed value and not as a parameter.
Thus, for this problem, we consider the Rayleigh number as the only parameter for the model,
denoted by µ ∈ D ⊂ R.
To define the variational form of problem (3.30), let us define the spaces Y = (H01 (Ω))d
for velocity, Θ = H01 (Ω) for temperature, and M = L20 (Ω) for pressure. We consider the H 1 -
seminorm for the velocity and temperature spaces and the L2 -norm for the pressure space. In
addition, let us define the Sobolev embedding constants Cu and Cθ associated with these norms,
such that
kvk0,4,Ω ≤ Cu k∇vk0,2,Ω ∀v ∈ Y (3.32)
and
kθk0,4,Ω ≤ Cθ k∇θk0,2,Ω ∀θ ∈ Θ. (3.33)
Moreover, we consider the tensor space X = Y × Θ × M , with the following associated norm:
kU k2X = k∇uk20,2,Ω + k∇θk20,2,Ω + kpk20,2,Ω ∀U = (u, θ, p) ∈ X. (3.34)
 70 Chapter 3. Certified Smagorinsky Reduced Basis Turbulence Model

To set the VMS framework, we decompose the velocity, pressure, and temperature spaces,
Y , M , and Θ, respectively, as
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Y = Yh ⊕ Y 0 , M = Mh ⊕ M 0 , Θ = Θh ⊕ Θ0 ,

where Yh , Mh , and Θh are, respectively, the large-scale finite dimensional spaces for velocity,
pressure, and temperature, while Y 0 , M 0 , and Θ0 are the small-scale complementary spaces. We
suppose that the large and small scales are separated by assuming that the sum is direct, i.e.,
Yh ∩ Y 0 = {0}, Mh ∩ M 0 = {0}, and Θh ∩ Θ0 = {0}.
The VMS modeling is a discretization of a set of macro-micro scale equations derived from
the previous decomposition (see [111] for more details), based upon the following procedure:
(i) Approximate the small-scale spaces Y 0 , M 0 , and Θ0 by finite-dimensional subspaces of
small resolved scales Yh0 , Mh0 , and Θ0 , respectively. Then, Y 0 = Yh0 ⊕Y 00 , M 0 = Mh0 ⊕M 00 ,
and Θ0 = Θ0h ⊕Θ00 , where Y 00 , M 00 , and Θ00 are complementary spaces of small unresolved
scales of finite dimension. This yields the unique decompositions

u = uh + u0h + u00 for all w ∈ Y,

p = ph + p0h + p00 for all p ∈ M,
θ = θh + θh0 + θ00 for all θ ∈ Θ,

with obvious notation.

(ii) Neglect the interaction between the large and small unresolved scales. It is assumed that
the interaction of large-small unresolved scales is weak whenever the latter lie inside the
inertial spectrum.
(iii) Model the action of small unresolved scales on small resolved scales in the Boussinesq
model by the eddy viscosity procedure.
We define Yh = Y h ⊕ Yh0 , Mh = M h ⊕ Mh0 , and Θh = Θh ⊕ Θ0h . Thus, uh = uh + u0h ∈ Yh ,
ph = ph + p0h , and θh = θh + θh0 .
Let us define the discrete space Vhl (Ω) = {r ∈ C 0 (Ω) : r|K ∈ Pl (K) ∀K ∈ Th }. Thus, we
define the velocity and pressure FE spaces as Yh = (Vhl (Ω) ∩ H01 )d , Θh = (Vhl (Ω) ∩ H01 ), and
Mh = Vhm ∩ L2 . For the subgrid eddy viscosity modeling in the VMS-Smagorinsky model, let
us consider Y h = Πh Yh , with Πh a uniformly stable interpolation operator on Y h , where

Y h = [Vhl−1 (Ω)]d . (3.35)

We can also define Y h = [VHl (Ω)]d , where VHl (Ω) is a subspace of Vhl (Ω) with a larger grid
size H > h; typically, H = 2h or H = 3h.
With that notation, we are identifying Y h as the large-scale velocity space and Yh0 = (Id −
Πh )Yh as the resolved small-scale velocity space and analogously for the temperature and pres-
sure spaces. Denoting the Rayleigh number by µ ∈ R, we obtain the following variational
formulation of problem (3.30):

 Find (uh , θh , ph ) = (uh (µ), θh (µ), ph (µ)) ∈ Xh such that

 u u u u



au (uh , vh ; µ) + b(vh , puh ; µ) + a0Su (uh ; uh , vh ; µ)




+ cu (uh , uh , vh ; µ) + f (θhu , vh ; µ) = F (vh ; µ)

∀vh ∈ Yh ,

(3.36)
v v


 b(u h , ph ; µ) = 0 ∀p h ∈ M h ,

 aθ (θhu , θhv ; µ) + cθ (uh , θhu , θhv ; µ)



 + a0 (u ; θu , θv ; µ) = Q(θv ; µ)

∀θhv ∈ Θh .
Sθ,n h h h h
 3.3. RB VMS-Boussinesq Model 71

Here, the bilinear forms au (·, ·; µ), aθ (·, ·; µ), b(·, ·; µ), and f (·, ·; µ) are defined by
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Z Z
au (u, v; µ) = P r ∇u : ∇v dΩ, aθ (θu , θv ; µ) = ∇θu · ∇θv dΩ,
Ω Ω
Z Z (3.37)
f (θu , v; µ) = −P r µ θu v2 dΩ, b(u, pv ; µ) = − (∇ · u)pv dΩ;
Ω Ω

the trilinear forms cu (·, ·, ·; µ), cθ (·, ·, ·; µ) are defined by

Z Z
cu (z, u, v; µ) = (z · ∇u)v dΩ, cθ (u, θu , θv ; µ) = (u · ∇θu )θv dΩ; (3.38)
Ω Ω

and the nonlinear VMS-Smagorinsky term for eddy viscosity, a0Su (·; ·, ·; µ), is given by
Z
a0Su (z; u, v; µ) = νT (Π∗h z)∇(Π∗h u) : ∇(Π∗h v) dΩ, (3.39)
Ω

with
νT (u) = (CS hK )2 |∇u|K |.
For the eddy diffusivity term, let us first introduce a mollifier φ ∈ Cc∞ (R), with supp(φ)⊂
B(0, 1), φ ≥ 0, kφk0,1,R > 0, and φ even, i.e., φ(−x) = φ(x). Let us consider the mollifier
sequence {φn (x)}n≥1 , with φn ∈ Cc∞ (R), supp(φn )⊂ B(0, 1/n), φn ≥ 0, kφn k0,1,R = 1,
defined by
n
φn (x) = φ(n x).
kφk0,1,R
The following result for mollifiers can be found in [73].

Proposition 3.6. Assume f ∈ Lp (Rd ) with 1 ≤ p < ∞. Then (φn ∗ f ) converges to f when
n → ∞ in Lp (Rn ).

Thus, the VMS-Smagorinsky eddy diffusivity term is defined as

Z
0 u v
aSθ,n (uh ; θh , θh ; µ) = νT,n (Π∗h uh )∇(Π∗h θhu ) · ∇(Π∗h θhv ) dΩ, (3.40)
Ω

with
νT,n (u) = (CS hK )2 (φn ∗ |∇u|K |).
Thanks to Proposition 3.6, it holds that a0Sθ,n converges uniformly to a0Sθ , with
Z
a0Sθ (uh ; θhu , θhv ; µ) = νT (Π∗h uh )∇(Π∗h θhu ) · ∇(Π∗h θhv ) dΩ.
Ω

The eddy diffusivity a0Sθ,n is considered for the well-posedness analysis and the development
of an a posteriori error bound estimator. In practice, we consider the eddy diffusivity term in the
Boussinesq VMS-Smagorinsky model as a0Sθ .
Writing Xh = Yh × Θh × Mh , we rewrite problem (3.36) as follows:

Find Uh (µ) = (uh , θhu , puh ) ∈ Xh such that

(
(3.41)
A(Uh (µ), Vh ; µ) = F (Vh ; µ) ∀Vh ∈ Xh ,
where

A(Uh , Vh ; µ) = A0 (Uh , Vh ) + µA1 (Uh , Vh ) + A2 (Uh ; Vh ) + A3 (Uh ; Vh ). (3.42)

 72 Chapter 3. Certified Smagorinsky Reduced Basis Turbulence Model

Defining Vh = (vh , θhv , pvh ) ∈ Xh , the µ-independent operators in (3.42) are given by
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Z Z
A0 (Uh , Vh ) = Pr ∇uh : ∇vh dΩ + ∇θhu · ∇θhv dΩ
Z Ω Z Ω

− (∇ · vh )ph dΩ + (∇ · uh )pvh dΩ,

u
Ω Ω
Z
A1 (Uh , Vh ) = −P r θhu v2 dΩ,
Ω
Z Z (3.43)
A2 (Uh ; Vh ) = (uh · ∇uh )vh dΩ + (uh · ∇θhu )θhv dΩ,
Ω Ω
Z
A3 (Uh ; Vh ) = νT (Π∗h uh )∇(Π∗h uh ) : ∇(Π∗h vh ) dΩ
ΩZ
1
+ νT,n (Π∗h uh )∇(Π∗h θhu ) · ∇(Π∗h θhv ) dΩ.
Pr Ω

3.3.2 RB Problem
In this section, we present the RB method for the model presented in Subsection 3.3.1. The RB
problem is as follows:

Find UN (µ) ∈ XN such that

(
(3.44)
A(UN (µ), VN ; µ) = F (VN ; µ) ∀VN ∈ XN ,

where the reduced space is defined as XN = YN × ΘN × MN ⊂ Xh . The reduced spaces for

velocity, temperature, and pressure are given by

MN = span{ξkp := puh (µk ), k = 1, . . . , N }, (3.45)

YN = span{ζ2k−1
v
:= uh (µk ), ζ2k
v
:= Tpµ ξkp , k = 1, . . . , N }, (3.46)

ΘN = span{ϕθk := θhu (µk ), k = 1, . . . , N }, (3.47)

where here, again, we are denoting by Tpµ ξkp the inner pressure supremizer that assures the inf-sup
stability of the RB problem.
With this notation, the RB solutions of problem (3.44) can be represented as

2N N N
p
X X X
uN (µ) = uN v
j (µ)ζj , θN (µ) = θjN (µ)ϕθj , pN (µ) = pN
j (µ)ξj .
j=1 j=1 j=1

The snapshots for the construction of the RB spaces are given by the greedy algorithm explained
in Subsection 3.2.2.

3.3.3 Well-Posedness Analysis

In this section, we present the well-posedness of problem (3.41) through the BRR theory [76, 93].
For this purpose, let us consider the directional derivative of A(·, ·; µ). Thus, deriving each term
 3.3. RB VMS-Boussinesq Model 73

in (3.43), we obtain
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

∂1 A0 (U, V )(Z) = A0 (Z, V ),

∂1 A1 (U, V )(Z) = A1 (Z, V ),
Z Z
∂1 A2 (U ; V )(Z) = (z · ∇u)v dΩ + (u · ∇z)v dΩ
ZΩ ΩZ
u v
+ (zh · ∇θ )θ dΩ + (u · ∇θz )θv dΩ,
Ω Ω

Z
∂1 A3 (U ; V )(Z) = νT (Π∗h u)∇(Π∗h z) : ∇(Π∗h v) dΩ
Ω
X Z ∇(Π∗h u) : ∇(Π∗h z)
+ (CS hK )2 ∇(Π∗h u) : ∇(Π∗h v) dΩ
K |∇(Π∗h u)|
K∈Th
1
Z
+ νT,n (Π∗h u) ∇(Π∗h θhz ) · ∇(Π∗h θhv ) dΩ
Pr Ω
1
Z
+ ∂1 νT,n (Π∗h u)(Π∗h z) ∇(Π∗h θu ) · ∇(Π∗h θv ) dΩ,
Pr Ω
where
∂1 νT,n (u)(z) = (CS hK )2 (φ0n ∗ |∇u|K |) : ∇z .
 

Note that we are defining U = (u, θu , pu ), V = (v, θv , pv ), Z = (z, θz , pz ) ∈ X. The

directional derivative verifies the following continuity and inf-sup conditions:

∂1 A(Uh (µ), Vh ; µ)(Zh )

∞ > γ0 < γh (µ) ≡ sup sup , (3.48)
Zh ∈Xh Vh ∈Xh kZh kX kVh kX

∂1 A(Uh (µ), Vh ; µ)(Zh )

0 < β0 < βh (µ) ≡ inf sup . (3.49)
Zh ∈Xh Vh ∈Xh kZh kX kVh kX

The existence of γ0 ∈ R and β0 > 0 satisfying (3.48) and (3.49), respectively, is given by
the following results.

Proposition 3.7. There exists γ0 ∈ R such that ∀µ ∈ D,

|∂1 A(Uh (µ), Vh ; µ)(Zh )| ≤ γ0 kZh kX kVh kX ∀Zh , Vh ∈ Xh .

Proposition 3.8. If the Dirichlet boundary data is sufficiently small, then there exists β̃(µ) > 0
such that

∂1 A(Uh , Vh ; µ)(Vh ) ≥ β̃(µ)(k∇vh k20,2,Ω + k∇θhv k20,2,Ω ) ∀Vh ∈ Xh . (3.50)

Since the operator b(vh , ph ; µ) satisfies the discrete inf-sup condition

b(vh , ph ; µ)
αkph k0,2,Ω ≤ sup ,
vh ∈Yh k∇vh k0,2,Ω

we can prove that the inf-sup (3.49) is satisfied thanks to Proposition 3.8.
 74 Chapter 3. Certified Smagorinsky Reduced Basis Turbulence Model

3.3.4 A Posteriori Error Bound Estimator

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

In this section, we present the development of the a posteriori error bound estimator, needed in
the greedy algorithm to construct the reduced spaces for velocity, temperature, and pressure. For
this purpose, we deal with the BRR theory [76].
With the following result, we prove that the directional derivative ∂1 A is Lipschitz con-
tinuous.

Lemma 3.9. There exists a positive constant ρn such that ∀Uh1 , Uh2 , Zh , Vh ∈ Xh ,

∂1 A(Uh1 , Vh ; µ)(Zh ) − ∂1 A(Uh2 , Vh ; µ)(Zh ) ≤ ρn kUh1 − Uh2 kX kZh kX kVh kX . (3.51)

Thus, we can define the a posteriori error bound estimator as

βN (µ) h p i
∆N (µ) = 1 − 1 − τN (µ) , (3.52)
2ρn

where τN (µ) is given by

4N (µ)ρn
τN (µ) = 2 (µ) , (3.53)
βN
with N (µ) the dual norm of the residual. The a posteriori error bound estimator is stated by the
following result.

Theorem 3.10. Let µ ∈ D, and assume that βN (µ) > 0. If problem (3.41) admits a solution
Uh (µ) such that
βN (µ)
kUh (µ) − UN (µ)kX ≤ ,
ρn
 
then this solution is unique in the ball BX UN (µ), βNρn(µ) .
Moreover, assume that τN (µ) ≤ 1 for all µ ∈ D. Then there exists a unique solution Uh (µ)
of (3.41) such that the error with respect to UN (µ), the solution of (3.44), is bounded by the a
posteriori error bound estimator, i.e.,

kUh (µ) − UN (µ)kX ≤ ∆N (µ), (3.54)

with effectivity  
2γN (µ)
∆N (µ) ≤ + τN (µ) kUh (µ) − UN (µ)kX . (3.55)
βN (µ)

3.3.5 Numerical Results for the VMS-Boussinesq Model

In this section, we present the numerical results for the VMS-Boussinesq RB model. The numer-
ical results were obtained using FreeFrem++ (cf. [268]). We consider a two-dimensional unit
square cavity, in which the flow is driven by a temperature gradient between the two vertical
walls. The top and the bottom of the cavity are isolated. We impose no-slip boundary condi-
tions for the velocity on all walls, and we suppose that gravity acts in the vertical direction. In
Figure 3.5 we show a schematic representation of the problem considered.
In this test, we consider two different scenarios, one for the Rayleigh number range [103 , 105 ]
and the other one for the Rayleigh number range [105 , 106 ]. For the first scenario, which cor-
responds to the lower Rayleigh number values, the heat transfer is principally in the form of
diffusion, i.e., the diffusion term in the energy equation is predominant, leading to an almost
 3.3. RB VMS-Boussinesq Model 75

(0, 1) ∂n θ = 0 u = 0 (1, 1)
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

u=0 u=0
θ=1 θ=0

(0, 0) ∂n θ = 0 u = 0 (1, 0)

Figure 3.5. Unit square cavity domain Ω, with the different boundaries identified, for problem (3.30).

vertical linear contouring for the temperature and a recirculating motion in the core of the re-
gion. As we increase the value of the Rayleigh number in D, the flow is stretched to the walls,
especially to the vertical walls, and the heat transfer becomes driven mainly by convection. The
isotherms become horizontal in a domain inside the cavity, far from the walls, that increases
as the Rayleigh number increases. When we consider the second scenario, where the Rayleigh
number range is higher, the velocity in the center of the cavity is practically zero and has large
and normal gradients near the vertical walls. The temperature isolines are horizontal in a large
domain inside the cavity, except near the vertical walls. This behavior agrees with the results
presented in several works, e.g., [109, 146, 601].
We consider different meshes depending on the scenario. For µ √ ∈ [103 , 105 ] we consider a
uniform mesh, with 50 divisions on each square side, i.e., h = 0.02 2. For µ ∈ √ [10 , 10 ] we
5 6

consider a finer mesh, with 70 divisions on each square side, i.e., h = 1/70 · 2, in order to
efficiently reproduce the eddies near the vertical walls appearing in this Rayleigh number range.
In Figure 3.6, we show some snapshots for different parameter values.
In the RB framework, we perform an EIM for both the eddy viscosity and the eddy diffusivity.
Although for the numerical analysis performed in this work we have considered a regularized
eddy diffusivity, the numerical tests are done with the eddy diffusivity defined in (3.31). Since
the eddy diffusivity is proportional to the eddy viscosity, we only need to perform one EIM. With
the EIM we are able to decouple the parameter dependence of the nonlinear eddy viscosity and
eddy diffusivity terms. For this test, we need M = 42 bases to reach a prescribed tolerance of
εEIM = 5 · 10−3 , when we consider the case when µ ∈ [103 , 105 ], and M = 150 basis functions
when we consider the second scenario with µ ∈ [105 , 106 ]. In this last case, the Smagorinsky
eddy viscosity and eddy diffusivity terms become more relevant, and for this reason we take a
lower tolerance for this test than for the previous one, considering εEIM = 10−4 . In Figure 3.7
we show the evolution of this error for both scenarios.
For the greedy algorithm we prescribe a tolerance of εRB = 10−4 for both scenarios. For
the first scenario, when µ ∈ [103 , 105 ], we need Nmax = 22 bases to reach this tolerance. When
N = 15, the condition of Theorem 3.10 holds and τN (µ) < 1 for all µ in D. In the second
scenario, when µ ∈ [105 , 106 ], we need N = Nmax = 64 basis functions to reach the tolerance
previously prescribed, making τN (µ) smaller than one when we get N = 52 basis functions. In
both cases, when τN (µ) > 1 and the a posteriori error bound is not defined, we use as a posteriori
error bound the proper τN (µ). In Figure 3.8 we show the convergence for the greedy algorithm,
 76 Chapter 3. Certified Smagorinsky Reduced Basis Turbulence Model
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 3.6. FE solution, velocity magnitude (top) and temperature (bottom), for µ = 4523,
µ = 55732, and µ = 642639 (left to right).

10 3 104

max kνT (µ) − IM [νT (µ)]k∞ max kνT (µ) − IM [νT (µ)]k∞
µ∈D
103
µ∈D

10 2

102

10 1 101

100
10 0

10-1

10
-1
10-2

10-3
10 -2

10-4

10 -3 10-5
0 5 10 15 20 25 30 35 40 45 0 10 20 30 40 50 60 70 80 90 100 110 120 130 140 150
M M

Figure 3.7. Error evolution for the EIM for µ ∈ [103 , 105 ] (left) and µ ∈ [105 , 106 ] (right).

and in Figure 3.9 we show the value of the a posteriori error estimator for N = Nmax for both
scenarios.
Finally, in Table 3.2, we compare the FE and RB solutions for several Rayleigh values in both
scenarios. We show the computational time for solving an FE solution and an RB solution in the
online phase. As can be observed, the speedup rate of the computational time is larger than three
orders of magnitude when µ ∈ [103 , 105 ], while when µ ∈ [105 , 106 ] the speedup rate is close to
300. The difference in the speedup magnitude between both cases is due to the greater number of
EIM and RB functions computed in each case. In addition, we show the errors in the H 1 -norm
for velocity and temperature and in the L2 -norm for pressure, for which we observe that the RB
solution is close enough to the FE solution, with the errors always being below of 105 in both
cases. For this test, the offline phase when µ ∈ [103 , 105 ] took approximately two days. For
the case when µ ∈ [105 , 106 ], the offline phase took approximately three weeks. In this offline
 3.4. Geometrical Parametrization 77

10 10
10 10
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

max τN (µ) max τN (µ)

µ∈D
µ∈D
max ∆N (µ)
µ∈D max ∆N (µ)
µ∈D

5 5
10 10

10 0 100

10 -5 10-5
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64
N N

Figure 3.8. Evolution of the a posteriori error bound in the greedy algorithm for µ ∈ [103 , 105 ]
(left) and µ ∈ [105 , 106 ] (right).

10 -3 10-3

-4
10
10-4

10 -5

10-5
10 -6

10 -7 10-6

10 -8
10-7

10 -9

10-8
10 -10

10 -11 10-9
10 3 10 4 10 5 105 106
Ra Ra

Figure 3.9. A posteriori error bound for N = Nmax for µ ∈ [103 , 105 ] (left) and µ ∈ [105 , 106 ] (right).

computational time we consider either the EIM or the greedy algorithm with the computation of
the a posteriori error estimator.

3.4 Geometrical Parametrization

3.4.1 Problem Setting
In this section we consider a natural convection in a cavity, as in Section 3.3, but considering
that the height of the cavity is variable. For that purpose, we consider a single geometrical
parameter for the height, denoted by µg , which modifies the aspect ratio of the cavity. Thus, let
Ωo (µg ) = [0, 1] × [0, µg ] ⊂ R2 be the original domain and let Γ = ΓD ∪ ΓN be the Lipschitz-
continuous boundary of Ωo (µg ), where ΓD is the part of the boundary with Dirichlet conditions
and ΓN is the part of the boundary with Neumann conditions.
To model this flow, we consider the Boussinesq model presented in Section 3.3, with VMS-
Smagorinsky modeling for the eddy viscosity and eddy diffusivity. Besides the geometrical
parameter, we can also consider the Rayleigh number as a parameter µph = Ra, letting µ =
(µph , µg ) ∈ D.
To be able to store parameter-independent matrices in the offline phase of the RB method,
we need to compute all the integrals in a reference domain by transforming the original domain.
 78 Chapter 3. Certified Smagorinsky Reduced Basis Turbulence Model

Table 3.2. Computational time for FE and RB solutions, with the speedup and the error, for
problem (3.30): Ra ∈ [103 , 105 ] (top) and Ra ∈ [105 , 106 ] (bottom)
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Data µ = 4060 µ = 17808 µ = 53778 µ = 93692

TF E 633.65 s 585.83 s 553.25 s 677.86 s
Tonline 0.55 s 0.5 s 0.46 s 0.49 s
speedup 1133 1151 1189 1367
kuh − uN k1 2.26 · 10−7 5.93 · 10−7 1.04 · 10−6 1.34 · 10−6
kθh − θN k1 4.57 · 10−9 5.57 · 10−9 8.9 · 10−9 8.83 · 10−9
kph − pN k0 2.49 · 10−7 3.22 · 10−7 1.11 · 10−6 2.27 · 10−6

Data µ = 16941 µ = 355402 µ = 667746 µ = 921441

TF E 3563.11 s 3675.01 s 4354.26 s 4928.37 s
Tonline 9.28 s 11.34 s 15.22 s 16.8 s
speedup 383 324 285 293
kuh − uN k1 1.74 · 10−6 2.77 · 10−6 2.28 · 10−6 5.35 · 10−6
kθh − θN k1 1.23 · 10−8 1.33 · 10−8 1.34 · 10−8 1.33 · 10−8
kph − pN k0 6.29 · 10−6 1.43 · 10−5 2.66 · 10−5 3.54 · 10−5

Thus, we set µref

g = 1, and we define the reference domain Ωr = Ωo (µref g ). The parameter-
dependent original domain can be recovered by a transformation map, T : Ωr ×D → R2 , defined
as   
1 0 x
T ((x, y); µg ) = ∀(x, y) ∈ Ωr . (3.56)
0 µg y
Because this map is linear, its Jacobian matrix and its determinant are given by
 
1 0
J((x, y); µg ) = and |J((x, y); µg )| = µg . (3.57)
0 µg
Let {Th }h>0 be a uniformly regular family of triangulations on the reference domain. We
define Yh = (H01 (Ωr ) ∩ Vhl (Ωr ))d , Mh = L20 (Ωr ) ∩ Vhm (Ωr ), Θh = H01 (Ωr ) ∩ Vhq (Ωr ),
with l, m, q ∈ N, respectively, as the discrete velocity, pressure, and temperature spaces on the
reference domain. Defining Xh = Yh × Θh × Mh , we define the problem with respect to the
reference domain, applying the change of variables of the transformation map T , as follows:
Find (uh , θhu , puh ) = (uh (µ), θhu (µ), puh (µ)) ∈ Xh such that





au,x (uh , vh ; µ) + au,y (uh , vh ; µ) + bx (vh , puh ; µ)




+ by (vh , puh ; µ) + a0Su,x (uh ; uh , vh ; µ)




0
 + aSu,y (uh ; uh , vh ; µ) + cu,x (uh , uh , vh ; µ)



+ cu,y (uh , uh , vh ; µ) + f (θhu , vh ; µ) = F̃ (vh ; µ) ∀vh ∈ Yh , (3.58)

v v v
 bx (uh , ph ; µ) + by (uh , ph ; µ) = 0 ∀ph ∈ Mh ,




 a (θu , θv ; µ) + a (θu , θv ; µ) + c (u , θu , θv ; µ)

θ,x θ,y θ,x h
 + c h(u h, θu , θv ; µ) + ah0 h (u ; θu , θv ; µ) h h





 θ,y h h h Sθ,x h h h
+ a0Sθ,y (uh ; θhu , θhv ; µ) = Q̃(θhv ; µ) ∀θhv ∈ Θh ,


where the subscripts x and y denote the addend of the corresponding operator relative to the
partial derivative with respect to x or y, respectively.
Here we suppose that we consider a uniform mesh in the reference domain Ωr , with Nh
partitions on each side. Since the mesh size, hK , in the VMS-Smagorinsky eddy viscosity and
 3.4. Geometrical Parametrization 79

eddy diffusivity terms is related to the parameter-dependent original domain, we adapt it to the
reference domain by applying the change of variables map T defined in (3.56), obtaining
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

µ2g + 1
hK = ∀K ∈ Th .
Nh2

Note that problem (3.41) is a particular case of problem (3.58), taking µg = 1. The well-
posedness analysis of this problem can be performed in an analogous way as in Section 3.3,
considering again the mollifier for the thermal eddy diffusivity in order to obtain that the deriva-
tive operator is locally Lipschitz.

3.4.2 Numerical Results

In this section we present some numerical results for the Boussinesq VMS-Smagorinsky RB
model with geometrical parametrization. We first suppose that the Rayleigh number is fixed with
Ra = 105 , and we consider the geometrical parameter µg ∈ D = [0.5, 2].
Next, we consider both the geometrical parameter and the Rayleigh number. For this test,
we consider the Rayleigh number (physical parameter), µph , in the range [103 , 104 ], and the
geometrical parameter, µg , in the range µg ∈ [0.5, 2]. Thus we are considering that the parameter
domain is D = [103 , 104 ] × [0.5, 2]. For both cases, the Prandtl number is P r = 0.71, which
corresponds to the Prandtl number of air.
The difference in the height of the cavity affects the buoyancy force, making it more relevant
when we increase the parameter value. This behavior is observed in Figure 3.10, in which we
show four solutions for different values of the geometrical parameter, with a fixed Rayleigh
number, Ra = 105 .
First, in the offline phase, we construct the RB space corresponding to the EIM, in which
we properly approximate the eddy viscosity and eddy diffusivity terms. In the case of only
geometrical parametrization, with the Rayleigh number fixed, we need M = 73 basis functions
in order to reach a prescribed tolerance of εEIM = 10−4 , while in the case of geometrical

Figure 3.10. FE snapshots for µg = 0.5 (top left), µg = 1 (bottom left), µg = 1.5 (middle), and
µg = 2 (right).
 80 Chapter 3. Certified Smagorinsky Reduced Basis Turbulence Model

10 3 10 2

max kνT (µg ) − I[νT (µg )]k∞ max kg(µ) − I[g(µ)]k∞

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

µg ∈D µ∈D
2
10 1
10

10 1

10 0

10 0

10 -1
10 -1

10 -2
10 -2

10 -3 10 -3

10 -4
10 -4
0 10 20 30 40 50 60 70 0 20 40 60 80 100 120 140
M M

Figure 3.11. Error evolution for the EIM and for the Boussinesq VMS-Smagorinsky model with
µ ∈ [103 , 104 ] × [0.5, 2].

and physical parametrization we need M = 138 basis functions with a prescribed tolerance of
εEIM = 10−3 . In Figure 3.11 we show the evolution of the infinity norm of the error between
the eddy viscosity νT (µg ) and its EIM approximation, for both cases.
For the greedy algorithm, we prescribe a tolerance for the a posteriori error bound of εRB =
10−4 in the case of the geometrical parameter only, while when we consider both parameters,
we prescribe a tolerance of εRB = 10−3 . We need N = 23 basis functions to guarantee the
conditions of Theorem 3.10 and get τN (µ) < 1. Then, we reach the prescribed tolerance when
N = Nmax = 32 for the case of a single geometrical parametrizations and N = Nmax = 46
when we consider both physical and geometrical parametrizations. In Figure 3.12, we show the
maximum value for all µ ∈ D of the a posteriori error bound estimator, as well as τN (µ) in
each iteration of the greedy algorithm for both cases. Moreover, in Figure 3.13, also for both
cases, we show the a posteriori error bound for all µ ∈ D in the last iteration of the greedy
algorithm.

10 10 10 8

max τN (µg ) max τN (µ)

µg ∈D
µ∈D
max ∆N (µg )
µg ∈D 10 6 max ∆N (µ)
µ∈D

10 5 10 4

10 2

10 0 10 0

-2
10

10 -5 10 -4
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 32 1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 54
N N

Figure 3.12. Evolution of the a posteriori error bound in the greedy algorithm for only the
geometrical parameter (left) and for both physical and geometrical parameters (right)

Finally, in Table 3.3, we summarize the results for several parameter values, for only the ge-
ometrical parametrization test. We compare the time for computing a FE solution and the online
phase computational time. We obtain a speedup rate of several hundred in the computational
time. The RB solution accuracy is fairly good, since the error is approximately of order 10−6 for
velocity, 10−8 for temperature, and 10−5 for pressure.
 3.4. Geometrical Parametrization 81
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

10 -4

10 -5

10 -6

10 -7

10 -8
0.5 1 1.5 2

Figure 3.13. A posteriori error bound for only the geometrical parameter (left) and for both
physical and geometrical parameters (right), for N = Nmax .

Moreover, in Table 3.4 we summarize some results obtained for some values of the parameter
when we consider both the geometrical and the physical parametrization. There we show that
the error between the FE solution and the RB solution is of order 10−5 for velocity, 10−7 for
temperature, and 10−5 for pressure. For this test, the speedup rate obtained in the computation
of the RB solution in the online phase with respect to the computation of the FE solution is
around 50. Again, we have obtained a good accuracy in the RB solution with respect to the FE
solution, with a decrease of the computational time that is worthy of consideration.

Table 3.3. Computational time for FE and RB solutions, with the speedup and the error, for the
Boussinesq VMS-Smagorinsky model with µg ∈ [0.5, 2].

Data µg = 0.64 µg = 1.08 µg = 1.44 µg = 1.87

TF E 808.91 s 810.16 s 866.1 s 851.82 s
Tonline 2.68 s 2.55 s 2.61 s 2.52 s
speedup 301 317 331 337
kuh − uN k1 1.13 · 10−6 1.86 · 10−6 2.82 · 10−6 3.4 · 10−6
kθh − θN k1 6.28 · 10−8 8.83 · 10−9 9.37 · 10−9 9.48 · 10−9
kph − pN k1 1.69 · 10−5 3.7 · 10−5 8.35 · 10−5 8.82 · 10−5

Table 3.4. Computational time for FE and RB solutions, with the speedup and the error, for the
Boussinesq VMS-Smagorinsky model with µ ∈ [103 , 104 ] × [0.5, 2].

Data Ra = 2143 Ra = 3506 Ra = 5922 Ra = 9618

µg = 1.95 µg = 0.71 µg = 1.13 µg = 1.63
TF E 600.96 s 914.18 s 684.95 s 630.94 s
Tonline 11.08 s 15.73 s 14.52 s 11.46 s
speedup 54 58 47 55
kuh − uN k1 1.18 · 10−5 1.27 · 10−5 5.25 · 10−6 5.65 · 10−6
kθh − θN k1 1.07 · 10−7 1.67 · 10−7 3.23 · 10−8 1.51 · 10−8
kph − pN k1 6.94 · 10−6 1.15 · 10−5 4.01 · 10−6 2.49 · 10−6
 82 Chapter 3. Certified Smagorinsky Reduced Basis Turbulence Model

3.5 Conclusion
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

In this chapter we presented two different models: the Smagorinsky model and the VMS-
Smagorinsky Boussinesq model. We constructed the RB problem associated with them and
developed numerically the a posteriori error estimator. In the case of the VMS-Smagorinsky
Boussinesq model we defined a mollifier in order to ensure that the derivative operator is locally
Lipschitz. Moreover, for the nonlinear terms, we dealt with the EIM in order to be able to store
the matrices for the RB problem. We also considered a variation of the RB VMS-Smagorinsky
Boussinesq model, introducing in the last section a geometrical parameter. We presented numer-
ical tests for all the models, in which we can observe a high speedup in the computation of the
RB solutions compared with the computational time of the FE solution.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Chapter 4

Finite Element–Based
Reduced Basis Method
for Optimal Flow Control

Maria Strazzullo, Francesco Ballarin, Gianluigi Rozza

4.1 Introduction
This chapter is a brief introduction to parametrized optimal control problems (OCP(µ)s) gov-
erned by parametrized partial differential equations (PDE(µ)s). They represent a versatile math-
ematical tool, able to bridge the gap between equations and collected data. Indeed, they have
been employed in many applications including fluid dynamics [148, 413, 461, 145], biomedicine
[33, 359, 562, 638], and environmental sciences [473, 475, 557, 562]. Despite the ubiquitous role
they play, OCP(µ)s are complex and challenging not only from a theoretical viewpoint but also
from a numerical one. Indeed, their goal is to change the expected state solution using control
variables, which steer the system toward a desired configuration, representing a more reliable
prediction of a known phenomenon or an observable realization of it.
Specifically, we will deal with a parametrized setting, which will change physical and ge-
ometrical features of the optimization processes by µ varying in a parametric space. Our goal
is to provide an overview of methodologies which can be used to treat the OCP(µ)s in a many-
query and real-time context, where several parametric instances are required in, possibly, a small
amount of time. Indeed, the applicability of optimal control strategies is limited by the computa-
tional costs they carry with them, most of all in the time-dependent or nonlinear framework. For
these kinds of systems, the needed effort for a parametric analysis can be too expensive to tackle
using standard discretization techniques. To overcome this issue, we will present reduced order
method (ROM) approaches [281, 464, 506, 505] adapted to this specific mathematical model
[25, 24, 149, 223, 302, 328, 329, 352, 413, 414, 475], moving from the well-known literature for
steady linear problems to time-dependent and nonlinear applications.
The contribution is outlined as follows. In Section 4.2, we introduce the linear problem struc-
ture in its saddle-point formulation and we show how to tackle it in a ROM fashion. Furthermore,
a first environmental application is shown: an optimal control on a pollutant release in the Gulf
of Trieste. Section 4.3 deals with the extension of the proposed structure and related ROM tech-
niques to nonlinear state equations, with a second environmental application in oceanographic
weather forecasting. Then, we move toward time dependency in Section 4.4, where the saddle-
point structure in the space-time formulation [235, 590, 631] is briefly described and space-time

83
 84 Chapter 4. Finite Element–Based Reduced Basis Method for Optimal Flow Control

ROM techniques are presented [558]. Their performance has been tested with a boundary control
for a Graetz flow. Finally, some conclusions and perspectives follow in Section 4.5.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

4.2 Linear OCP(µ)s

In this section, we present the simplest setting for OCP(µ)s. Indeed, we focus on linear quadratic
problems: namely, the optimization consists of minimizing a quadratic cost functional con-
strained to a linear PDE(µ). The proposed framework is based on Hilbert spaces, but it is also
valid for Banach spaces [290].
Let us define an uncontrolled state variable y ∈ Y , which represents the physical phe-
nomenon we want to change using unknown quantities, say u ∈ U . Specifically, Y and U are
two possibly different Hilbert spaces over the spatial domains Ω and Ωu ⊂ Ω, respectively, where
Ω denotes the closure of the spatial domain considered. The domain Ωu is called the control do-
main, and, analogously, we will denote the control variable as u. Its role is to affect the behavior
of the PDE(µ) at hand in order to recover a desired variable yd ∈ Yd ⊃ Y . The observation
space Yd is defined over Ωd ⊂ Ω. The goal is reached by minimizing the quadratic functional
J : Yd × U → R:
. 1 α
J (y, u) = ky − yd k2Yd + kuk2U , (4.1)
2 2
constrained to
F(y, u; µ) = 0, (4.2)

where F : Y ×U ×D → Y ∗ is the linear state equation, D ⊂ Rp is a parameter space with p ≥ 1,

and α ∈ (0, 1] is a penalization parameter which weights the control action in the minimization
process; i.e., a higher value of α causes a limited control effect and a larger value of J .
We remark that all the involved variables are µ dependent, but, for the sake of notation, we
will omit their explicit dependency from the parameter.
The parametrized minimization problem then reads as follows: given a parameter µ ∈ D and
an observation yd ∈ Yd , find the pair (y, u) ∈ Y × U which minimizes (4.1) and solves (4.2).
Figure 4.1 is a graphical representation of the optimization pipeline. To solve the problem we
rely on the Lagrangian formulation [290, 374]. In order to describe the Lagrangian framework,
we need to define the following quantities: the state bilinear form a : Y × Y → R, the control
bilinear form c : U × Y → R, and the forcing term f (µ) ∈ Y ∗ such that
.
hF (y, u; µ) , ξi = a (y, ξ; µ) − c (u, ξ; µ) − hf (µ) , ξi = 0 ∀ξ ∈ Y. (4.3)

y yd

F (y, u; µ) = Minimization process J (y, u)

0

optimal (y, u) which

minimizes J (y, u)

Figure 4.1. Graphical representation of the OCP(µ) pipeline.

 4.2. Linear OCP(µ)s 85

Furthermore, we consider two other bilinear forms, m : Y × Y → R and n : U × U → R,

such that the functional (12.20) can be recast as
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

1 α
J (y, u) = m(y − yd , y − yd ) + n(u, u). (4.4)
2 2
To solve the optimization problem, we define an auxiliary variable z ∈ Y , the adjoint variable,2
and the Lagrangian functional
.
L(y, u, z; µ) = J (y, u) + hF (y, u; µ) , zi . (4.5)

Namely, in this way, the constrained optimization problem at hand is equivalent (see, e.g., [290])
to the unconstrained minimization of (4.5), which can be solved thanks to the following system
of equations: given µ ∈ D, an observation yd ∈ Yd , and f (µ) ∈ Y ∗ , find the triplet (y, u, z) ∈
Y × U × Y which verifies
(adjoint equation),

 Ly (y, u, z; µ)[ζ] = 0 ∀ζ ∈ Y
Lu (y, u, z; µ)[τ ] = 0 ∀τ ∈ U (optimality equation), (4.6)
Lz (y, u, z; µ)[ξ] = 0 ∀ξ ∈ Y (state equation),


where the subscripts y, u, and z indicate the Fréchet differentiation with respect to the state,
control, and adjoint variables, respectively.
We now want to recast the system (4.6) in a saddle-point formulation. From this perspective,
.
we define the state-control variable x = (y, u) ∈ X = Y ×U , the bilinear forms A : X×X → R
and B : X × Y → R, and the functionals H ∈ Y ∗ as
.
A(x, ω; µ) = m(y, ζ) + αn(u, τ ) ∀ω = (ζ, τ ) ∈ X,
B(x, ξ; µ) = a(y, ξ; µ) − c(u, ξ) ∀ξ ∈ Y, (4.7)
hH, ωi = m(yd , ζ) ∀ζ ∈ Y.
It is a matter of computation to show that, thanks to definitions (4.7), the optimality system
(4.6) is equivalent to the following mixed problem: given µ ∈ D, an observation yd ∈ Yd , and
f (µ) ∈ Y ∗ , find the pair (x, z) ∈ X × Y which verifies

A(x, ω; µ) + B(ω, z; µ) = hH, ωi ∀ω ∈ X,
(4.8)
B(x, ξ; µ) = hf (µ), ξi ∀ξ ∈ Y.
The well-posedness of the system (4.8) relies on the following theorem.

Theorem 4.1 (Brezzi theorem). The parametrized problem (4.8) admits a unique solution
(x, z) ∈ X × Y if the following conditions hold:
• A(·, ·; µ) is continuous and weakly coercive on the set X0 which denotes the kernel of
B(·, ·; µ); i.e., for a positive constant βA ,
A(x, ω; µ) A(x, ω; µ)
inf sup ≥ βA > 0 and inf sup ≥ 0.
x∈X0 ω∈X0 kxkX kωkX ω∈X0 x∈X0 kxkX kωkX

• B(·, ·; µ) is continuous and infsup-stable; i.e., there exists a positive constant βB such that
B(x, ξ; µ)
inf sup ≥ βB > 0. (4.9)
ξ∈Y x∈X kxkX kξkY
2 The adjoint variable is chosen in the state space Y in order to guarantee the well-posedness of the problem in the

mixed formulation proposed in (4.8).

 86 Chapter 4. Finite Element–Based Reduced Basis Method for Optimal Flow Control

The choice of z ∈ Y makes Theorem 4.1 provable and thus gives us the chance to study the
discretized solution of saddle-point type OCP(µ)s. In the next subsection, we will briefly discuss
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

the finite element (FE) technique and an ad hoc reduction algorithm for OCP(µ)s that we employ
in the numerical results presented in Section 4.2.2.

4.2.1 ROMs for Linear OCP(µ)s

The first necessary step to be taken in order to simulate and study OCP(µ)s in a discrete frame-
work is to employ a high-fidelity discretization. We briefly describe the FE approximation,
which was exploited in all the results we will present in this contribution [477]. First of all,
we define a triangulation T over the spatial domain considered, say Ω. Then, we use the spaces
. .
Y Ny = Y ∩ Xry and U Nu = U ∩ Xru , where
.
Xr = {q ∈ C 0 (Ω) : q|K ∈ Pr ∀K ∈ T }, (4.10)

where Pr is the space of polynomials of degree at most r. In this discretized framework, defining
. N
X Nx = Y Ny × U Nu , the FE OCP(µ) reads as follows: given µ ∈ D, an observation yd ∈ Yd y ,
and f (µ) ∈ Y , find the pair (x , z ) ∈ X × Y
Ny Nx Ny Nx Ny
which verifies

A(xNx , ω; µ) + B(ω, z Ny ; µ) = hH, ωi ∀ω ∈ X Nx ,


(4.11)
B(xNx , ξ; µ) = hf (µ), ξi ∀ξ ∈ Y Ny .

System (4.11) is well posed, once again, thanks to Theorem 4.1. However, exploiting the FE
.
approximation of global dimension N = 2Ny + Nu might lead to high computational costs,
which can be unacceptable since the optimization process is usually performed for many different
parameter values (many-query context), and, eventually, for a given new value of µ ∈ D, we
want to compute the solution quickly (real-time context). Supposing that the system at hand is
affine decomposed [281], i.e., that all the bilinear forms considered in (4.11) can be written as a
summation of µ-dependent functions times µ-dependent forms and bilinear forms, we can build
an efficient ROM model based on two stages:
1. The first is an expensive offline procedure which is performed only once. It deals with all
the µ-independent quantities. Here Nmax snapshots, i.e., FE solutions at properly chosen
parameters, are computed and manipulated in order to build the reduced model;
2. The second is a cheap online phase performed for every new µ in a low-dimensional frame-
work dependent on a dimension value N  N .
Let us suppose we have built some low-dimensional spaces YyN , U N , and YzN for the state,
.
control, and adjoint variables, respectively. Moreover, let us define, consequently, X N = YyN ×
U N . Namely, in order to solve the optimality system more quickly, one performs a Galerkin
projection in this reduced framework. Thus, the reduced problem reads as follows: given µ ∈ D,
an observation yd ∈ YdN , and f (µ) ∈ YyN , find the pair (xN , z N ) ∈ X N × YzN which verifies

A(xN , ω; µ) + B(ω, z N ; µ) = hH, ωi ∀ω ∈ X N ,


(4.12)
B(xN , ξ; µ) = hf (µ), ξi ∀ξ ∈ YzN .

Also in this case, the well-posedness of system (4.12) relies on the Brezzi Theorem 4.1. The con-
tinuity of the involved bilinear forms and the weak coercivity of A(·, ·; µ) are directly inherited
by the high-fidelity discretization. However, this is not the case for the reduced inf-sup stability
condition
B(x, ξ; µ)
inf sup ≥ βBN > 0. (4.13)
ξ∈YzN x∈X N kxkX kξkY
 4.2. Linear OCP(µ)s 87

In order to satisfy condition (4.13), we will exploit the aggregated space technique [149, 328,
329, 413, 414]. This strategy is based on postprocessing the basis functions provided by the
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

partitioned proper orthogonal decomposition (POD)-Galerkin algorithm. For an introduction

to the POD algorithm, the interested reader may refer to [85, 114, 281]. First, we applied the
POD to the three variables separately, which led to possibly different reduced spaces for state and
.
adjoint variables. Thus, for them, a common reduced space is built: Y N = YyN ∪YzN . In this new
.
context, one can redefine the space as X N = Y N × U N , and problem (4.12) is recast as follows:
given µ ∈ D, an observation yd ∈ Yd , and f (µ) ∈ YyN , find the pair (xN , z N ) ∈ X N × Y N
N

which verifies
A(xN , ω; µ) + B(ω, z N ; µ) = hH, ωi ∀ω ∈ X N ,

(4.14)
B(xN , ξ; µ) = hf (µ), ξi ∀ξ ∈ Y N ,
which is well posed since the relative reduced inf-sup condition
B(x, ξ; µ)
inf sup ≥ βBN > 0 (4.15)
ξ∈Y N x∈X N kxkX kξkY

is provable. Despite the higher dimensionality of the reduced system, enlarging the state and
adjoint spaces assures the well-posedness of the problem in the reduced setting. In the next
subsection, we propose a first linear application in environmental sciences: control of a pollutant
release in a marine ecosystem in the reduced setting.

4.2.2 Numerical Test: Pollutant Control in the Gulf of Trieste

In this section, we validate the reduction strategy presented in Subsection 4.2.1 in an exam-
ple involving the marine environment and ecosystem monitoring. The test case deals with an
advection-diffusion pollutant control problem in the Gulf of Trieste, Italy, a physical basin char-
acterized by particularly windy behavior and rich biodiversity. In this context, ROMs can be very
effective for a parametrized analysis of the marine safeguard and, consequently, to its impact on
the local community. Indeed, Trieste overlooks the sea, and its economic and social activities
depend a great deal on its marine environment. In the presented test case, we describe a pollu-
tant loss in the city harbor. We aim at reducing the pollutant concentration y in order to achieve
a safeguard threshold, represented by the desired concentration yd , under several parametrized
instances related to the wind action on the surface currents of the gulf. The problem reads as
follows: given µ = (µ1 , µ2 , µ3 ) ∈ D = [0.5, 1] × [0., 1.] × [0., 1.], and yd = 0.2 ∈ L2 (Ωd ), find
the pair (y, u) ∈ HΓ1D (Ω) × R that solves

1 α
Z Z
min (y − yd )2 dΩd + u2 dΩu (4.16)
Y ×U 2 Ω 2 Ωu
obs

under the constraints

 µ1 ∆y + [µ2 , µ3 ] · ∇y = u in Ω,


∂y
=0 on ΓN , (4.17)
 ∂n

y=0 on ΓD ,
where the domain Ω and its divisions are represented in Figure 4.2. In this case, the Neumann
conditions are imposed in the open sea ΓN , while homogeneous Dirichlet boundary conditions
have been applied along the coastline ΓD . The control variable u is a forcing term representing
the maximum quantity of pollutant that can be released in Ωu (green subdomain in Figure 4.2) in
order to safely reach the value yd = 0.2 in Ωobs (orange subdomain in Figure 4.2), i.e., the natural
area of Miramare. The parameters represent the diffusivity and the advection action of the wind
blowing on the Gulf surface. We exploited an aggregated space POD-Galerkin algorithm over
 88 Chapter 4. Finite Element–Based Reduced Basis Method for Optimal Flow Control
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 4.2. Left: zoom of the mesh overlapping satellite images (Trieste harbor). Right: subdo-
mains and boundaries. Orange: observation domain Ωobs . Green: control domain Ωu .

Nmax = 100 uniformly distributed snapshots for a fixed α = 10−5 . The global dimension of the
FE system is 5939, where linear polynomials have been exploited for all the involved variables.
In order to build the reduced space, we used a value of Ny = 20 bases for state and adjoint
variables and Nu = 1 basis for the control variable, since U = R, leading to a global reduced
space of dimension N = 4Ny + Nu = 81. The number of bases ensures a good recovery of the
parametrized solution: this can be seen both from the plots of some representative solutions in
Figure 4.3 and from the average relative log-error of the variables with respect to the value of N ,
over 100 uniformly distributed parameters. Indeed, the FE and ROM simulations coincide and
the average log-error shows that we reach a very good approximation even with a small number
of basis functions. Furthermore, defining the speedup index as the number of reduced problems
one can perform in the time of an FE simulation, we reach values of around 250 for a reduced
space of dimension N = 81.

Figure 4.3. Left and center: optimal FE and ROM state pollutant concentration for µ =
(1., −1., 1.), representing the Bora wind action. Right: averaged relative log-error for the variables.

4.3 ROMs for Nonlinear OCP(µ)s

This section aims at extending the methodology presented in Subsection 4.2.1 to nonlinear state
equations. In the nonlinear context the constraint has the form
.
hF (y, u; µ) , ξi = a` (y, ξ; µ) + an` (y, ξ; µ) − c (u, ξ; µ) − hf (µ) , ξi = 0 ∀ξ ∈ Y, (4.18)

i.e., the action of the state equation can be divided into linear and nonlinear contributions, repre-
sented by a` (·, ·; µ) and an` (·, ·; µ), respectively. In order to achieve the goal yd ∈ Yd , we rely
once again on the Lagrangian formulation to solve the minimization problem, i.e., one can build
the Lagrangian functional (4.5) and solve the nonlinear optimality system (4.6). At the contin-
uous level, the saddle-point structure we presented in the linear case is not preserved and the
 4.3. ROMs for Nonlinear OCP(µ)s 89

Brezzi theorem cannot be exploited in this general case. For the sake of brevity, we will not fo-
cus on the well-posedness of the problem; however, the interested reader may refer to [290, 452].
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

From now on, we will always assume that the nonlinear OCP(µ) at hand is well posed.
The main goal is to recover the saddle-point structure at the discrete level.3 Indeed, once the
FE spaces Y Ny and U Nu are considered, as in Section 4.2.1, the problem to be solved is as fol-
lows: given µ ∈ D, an observation yd ∈ Yd , and f (µ) ∈ Y ∗ , find the triplet (y Ny , uNu , z Ny ) ∈
Y Ny × U Nu × Y Ny which verifies

 Ly (y Ny , uNu , z Ny ; µ)[ζ] = 0 ∀ζ ∈ Y Ny ,


Lu (y Ny , uNu , z Ny ; µ)[τ ] = 0 ∀τ ∈ U Nu , (4.19)

Lz (y Ny , uNu , z Ny ; µ)[ξ] = 0 ∀ξ ∈ Y Ny .


Algebraically, one has to deal with (4.19) through Newton’s method. Namely, we itera-
tively solve a linearized version of (4.19) until a convergence criterion is reached. It is a matter
of computation to show that the structure of the linearized system has a saddle-point nature
[289, 452, 560]. This concept is fundamental to justify the employment of the aggregated space
technique for this specific case, since the peculiar saddle-point structure is also preserved at a re-
duced level and, consequently, the technique for linear problems adapts to nonlinear ones. Thus,
applying the partitioned POD algorithm and space aggregation as in Subsection 4.2.1, the re-
duced problem reads as follows: given µ ∈ D, an observation yd ∈ YdN , and f (µ) ∈ Y N , find
the triplet (y N , uN , z N ) ∈ Y N × U N × Y N which verifies

 Ly (y N , uN , z N ; µ)[ζ] = 0 ∀ζ ∈ Y N ,


Lu (y N , uN , z N ; µ)[τ ] = 0 ∀τ ∈ U N , (4.20)
Lz (y N , uN , z N ; µ)[ξ] = 0 ∀ξ ∈ Y N .


The next subsection describes an application in environmental sciences, where the OCP(µ)
is meant as a tool to understand and forecast atmospheric oceanographic phenomena.

4.3.1 Numerical Test: Weather Prediction through Quasi-geostrophic

Equations
This section exploits optimal control strategy to make environmental predictions more reliable.
Indeed, weather forecasting might have several modeling issues:

• On the one hand, the general ocean circulation model describing large-scale flow dynamics
is a very complicated coupled system of ocean and atmosphere actions [103, Chapter 3].
Despite its completeness as an equation system, it can give results which are not compara-
ble with collected data.

• On the other, oceanography suffers from a lack of data. The data is scattered, difficult to
interpret, noisy, and, furthermore, complicated to collect.

Our aim is to add data to the model using optimal control, exploiting the following formulation:
given µ ∈ D = [0.073 , 1] × [10−4 , 1] × [10−4 , 0.0452 ], and yd ∈ L2 (Ω), find (ψ, q) ∈ Y =
H01 (Ω) × H01 (Ω) and u ∈ U = L2 (Ω) which solve

1 α
Z Z
2
min (ψ − ψd ) dΩ + u2 dΩ (4.21)
Y ×U 2 Ω 2 Ω
3 This is why we also restrict ourselves to the case z ∈ Y in the nonlinear scenario.
 90 Chapter 4. Finite Element–Based Reduced Basis Method for Optimal Flow Control

under the quasi-geostrophic constraint

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

in Ω,

 q = ∆ψ
 ∂ψ


= u − µ1 q + µ2 ∆q − µ3 Q(ψ, q) in Ω,
∂x (4.22)


 ψ = 0 on ∂Ω,
q=0 on ∂Ω,


where the nonlinear expression Q : H01 (Ω) × H01 (Ω) → R is defined as

∂ψ ∂q ∂ψ ∂q
Q(ψ, q) = − . (4.23)
∂x ∂y ∂y ∂x

Using such a data assimilation model will give an optimal solution close to the physical data
and could prevent having to collect it, a process which might be expensive, in terms of time and
money. The parametric setting is essential since the dynamic is influenced by several seasonal
phenomena and the physical parameters can help us understand and analyze them. While µ1 and
µ2 represent a diffusive action, the value of µ3 affects the local nonlinear dynamics. The model
describes both the North Atlantic Ocean large-scale currents and the location and intensity varia-
tions of the gyres which are typical of this geographic area, depicted in Figure 4.4, which shows
the Florida peninsula, where the motion of the Gulf Stream starts, eventually warming the coasts
of northern Europe. In the weather forecasting field, it is very important to run many simulations
for several parametric values in order to better study and predict climatological phenomena. In
particular, for this experiment, the desired state ψd is an FE simulation obtained by the uncon-
trolled governing equation with a right-hand side given by f = − sin(πy), with µ1 = 10−4 ,
µ2 = 0.073 , and µ3 = 0.072 , and represents the seasonal Gulf Stream. Figure 4.5 shows the
desired solution profile together with the high-fidelity and reduced solutions. They all match.
Furthermore, the same figure also represents the averaged log-error norm between FE and re-
duced variables over 100 randomly distributed parametric evaluations. It indicates the ability of
the ROM solutions to recover the field with an aggregated reduced system of global dimension
N = 180. About time performances, despite solving a much lower dimensional system than the
original one, we do not reach great speedup; it is around two. This issue is related to the nonlinear
nature of the system. Indeed, the nonlinear form still depends on the FE dimension N = 6490,
obtained by using P1 elements for all the variables, and the online phase involves the assembly
(and projection) of the high-fidelity solutions. Nevertheless, one can rely on such hyperreduc-
tion techniques as the empirical interpolation method (EIM) to recover an affine decomposed
structure; see, e.g., [40].

Figure 4.4. Left: zoom of the mesh overlapping satellite images (Florida peninsula). Right:
triangulation of the spatial domain Ω, representing the North Atlantic Ocean, from the Florida peninsula
to northern Europe.
 4.4. ROMs for Space-Time OCP(µ)s 91
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 4.5. Top: comparison between desired state, FE and ROM solutions for µ =
(10−4 , 0.073, 0.0452), representing the Gulf Stream. Bottom: averaged relative log-error for all the vari-
ables.

4.4 ROMs for Space-Time OCP(µ)s

This section deals with the extension of the already presented ROM techniques to time-dependent
problems. For the sake of clarity, we will focus on linear cases; the interested reader may refer to
[560] for a nonlinear time-dependent application. The linear time-dependent problem deals with
the state equation
Z T Z T
.
hFt (y, u; µ) , ξi = s (y, ξ; µ) dt + a (y, ξ; µ) dt
0 Z T 0 Z T (4.24)
− c (u, ξ; µ) dt − hf (µ) , ξi dt = 0 ∀ξ ∈ Y,
0 0

where s : Y × Y → R is a bilinear form describing the time evolution of the state equation (i.e.,
it depends on the time derivative yt ) in the time interval [0, T ] and
.
Y = {y ∈ L2 (0, T ; Y ) such that yt ∈ L2 (0, T ; Y ∗ )}.
The action of s(·, ·; µ) is
s(y, ζ; µ) = hS(µ)yt , ζiY ∗ Y . (4.25)
For a more detailed explanation, see, e.g., [559]. In the time-dependent setting, we have not only
spatial boundary condition but also an initial time evolution condition: in this contribution we
will restrict ourselves to the case of a state variable that verifies the spatial boundary conditions
and y(t = 0) = 0 in Ω. Here, guided by the numerical test we present in the following subsec-
tion, we deal with geometrical parametrization, and the forms involved in (4.24) are the possible
trace-backs in a reference domain that we call, with abuse of notation, Ω. The trace-back tech-
nique is used to deal with geometrical parameters in a ROM setting in [505]. We specify that
S(µ), in the nongeometrical case, is the identity map. Now, we can define the objective func-
tional and, in this specific setting of time-dependent problems, it is of the form
1 T α T
Z Z
J (y, u) = m(y − yd , y − yd )dt + n(u, u)dt, (4.26)
2 0 2 0
 92 Chapter 4. Finite Element–Based Reduced Basis Method for Optimal Flow Control

.
with yd taken in Yd = L2 (0, T ; Yd ). Also in this case, we can exploit the Lagrangian formalism,
defining the adjoint variable z ∈ Y; solve the problem by building a functional of the form
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

.
L(y, u, z; µ) = J (y, u) + hFt (y, u; µ) , zi ; (4.27)

and, differentiating with respect to the three different variables, obtain a time-dependent optimal-
ity system. The main feature of this system is that, provided we have a suitable discrete approx-
imation, the saddle-point structure is preserved even in the time-dependent setting. To this end,
we introduce the space-time approximation. Extending what was already proposed for parabolic
problems [235, 590, 631] to the optimal control context, we define the finite-dimensional function
spaces YNy and UNu . The global high-fidelity dimension is N = 2Ny +Nu , with Ny = NΩy ·Nt
and Nu = NΩu · Nt , where NΩy , NΩu , and Nt are the state-adjoint spatial, the control spatial, and
the time discretization dimensions, respectively. In the test case we will present, we used the
FE-Euler approach. Namely, the time interval [0, T ] is divided into Nt equispaced subintervals
of length ∆t. The generic discrete time instance is tk = k∆t for k = 0, . . . , Nt . Thus, the
Ny Nu Ny
discrete variables at time tk are yk Ω , uk Ω , and zk Ω . Now, given an FE basis for the spatial state
Ny Nu
space Y and the spatial control space U , say {φi }i=1
Ω
and {ϕi }i=1
Ω
, respectively, we can expand
the variables at tk as follows:
y u y
NΩ NΩ NΩ
Ny Nu Ny
X X X
yk Ω = yki φi , uk Ω = uik ϕi , and zk Ω = zki φi . (4.28)
1 1 1

In other words, each time instance is an FE variable, which drastically increases the dimension
of the algebraic problem at hand. We underline that the adjoint version of a parabolic equation
is a backward parabolic one. Indeed, the adjoint of s(y, ζ; µ) is actually −s(z, ξ; µ), exploiting
the final time condition z(t = T ) = 0. For the parabolic state equation, we used a backward
Euler method, while the adjoint equation is discretized backward in time using the forward Euler
method. In [558], the authors show that, given a common discretization technique for state and
adjoint, the whole space-time optimality system reads as follows: given µ ∈ D, an observation
N
yd ∈ Yd y , and f (µ) ∈ YNy , find the pair (xNx , z Ny ) ∈ XNx × YNy which verifies
 Z T Z T Z T
A(xNx , ω; µ)dt + B(ω, z Ny ; µ)dt = hH, ωi dt ∀ω ∈ XNx ,



Z0 T Z 0T 0 (4.29)
B(xNx , ξ; µ)dt = hf (µ), ξidt ∀ξ ∈ YNy ,



0 0

.
where XNx = YNy × UNu . Here, the bilinear forms A : X × X → R and B : X × Y → R and
the functionals H ∈ Y ∗ are defined as
.
A(x, ξ; µ) = m(y, ω; µ) + αn(u, κ; µ) ∀ξ = (ω, κ) ∈ X,
B(x, ζ; µ) = s(y, ζ) + a(y, ζ; µ) − c(u, ζ; µ) ∀ζ ∈ Y, (4.30)
H(ξ; µ) = m(yd , ω; µ) ∀ω ∈ Y,

where the involved bilinear forms were defined in Section 4.2. However, here, as already spec-
ified, they can represent their possible trace-back. This framework justifies using a partitioned
space-time POD: namely, we separately applied the POD algorithm [85, 114, 281] to the three
variables containing the solutions at all the temporal steps. After this first compression step, the
reduced spaces were built using the aggregation technique, due to the saddle-point nature of the
system at hand.
 4.4. ROMs for Space-Time OCP(µ)s 93

4.4.1 Numerical Test: Graetz Flow Boundary Control

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

In this subsection, we propose some results concerning a time-dependent OCP(µ) governed by a

Graetz flow, parametrized in both physics and geometry by µ = [µ1 , µ2 , µ3 ] ∈ D = [1/20, 1/6]×
[1/2, 3] × [1, 3]. The control acts over the boundary

ΓC = ([1, 1 + µ3 ] × {0}) ∪ ([1, 1 + µ3 ] × {1}).

The parameter µ3 stretches the length of the subdomains

.
Ω2 (µ3 ) = [1, 1 + µ3 ] × [0.2, 0.8]

and
.
Ωd (µ3 ) = [1, 1 + µ3 ] × [0, 0.2] ∪ [1, 1 + µ3 ] × [0.8, 1].
The physical domain Ω is depicted in Figure 4.6. Besides the geometrical parameter, we will
deal with µ1 and µ2 , which represent the diffusivity coefficient of the system and the desired
profile we want to reach in Ωd (µ3 ), respectively. In this specific case, we define y ∈ Y, where
Y = HΓ1D (Ω(µ3 )), and u ∈ U, where U = L2 (ΓC (µ3 )), where ΓD = ({0} × [0, 1]) ∪ ([0, 1] ×
{0}) ∪ ([0, 1] × {1}) is the portion of the domain where Dirichlet boundary conditions have been
applied.

(0, 1) (1, 1) ΓC (µ3 ) (1 + µ3 , 1)

Ω3 (µ3 )

ΓD Ω1 Ω2 (µ3 ) ΓN (µ3 )

Ω4 (µ3 )
(0, 0) (1, 0) (1 + µ3 , 0)

Figure 4.6. Domain Ω. Observation domain: Ωd (µ3 ) = Ω3 (µ3 ) ∪ Ω4 (µ3 ). Control domain:
ΓC (µ3 ) (red solid line). Blue dashed line: Dirichlet boundary conditions. The reference domain Ω is given
by µ3 = 1.

The problem we consider reads as follows: given µ ∈ D, find the state-control variable
(y, u) ∈ Y × U which solves

1 T α T
Z Z Z Z
2
min (y − yd (µ2 )) dxdt + u2 dxdt (4.31)
Y×U 2 0 Ω3 (µ) 2 0 ΓC

constrained to the equations

∂y

 yt + µ1 ∆y + x2 (1 − x2 )
 =0 in Ω(µ3 ) × (0, T ),
∂x1



y=1 on ΓD (µ3 ) × (0, T ),




 ∂y
µ1 =u on ΓC (µ3 ) × (0, T ), (4.32)

 ∂n
 ∂y
on ΓN (µ3 ) × (0, T ),

 µ1 =0


 ∂n
in Ω(µ3 ) × {0},

y = y0

 94 Chapter 4. Finite Element–Based Reduced Basis Method for Optimal Flow Control
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 4.7. Left: FE state solutions for t = 1 s, 2 s, 3 s for µ = (12.0, 2.5, 2.0). Right: ROM
state solutions for t = 1 s, 2 s, 3 s for µ = (12.0, 2.5, 2.0).

where x1 and x2 denote the spatial components, y0 is the null function in the domain which
respects the boundary conditions, yd (µ2 ) ≡ µ2 , and T = 5. With ΓN (µ3 ) = {1 + µ3 } × [0, 1],
we indicate the portion of the domain where Neumann conditions apply. First of all, in order
to solve this OCP(µ), we traced back the problem in the reference domain corresponding to
µ3 = 1, following the pullback strategy presented in [505]. We used ∆t = 1/6 over the time
interval, employing Nt = 30 time steps in an Euler-based scheme. Furthermore, we exploited
P1 elements for all the variables involved, resulting in a global space-time dimension of N =
313830. The reduced function spaces have been built by compressing the information of Nmax =
70 snapshots, retaining N = 35 bases for each variable. After the aggregation technique, we
obtained a global reduced system of dimension 5N = 175. The reliability of the approach is
shown in Figure 4.7, which depicts representative FE solutions and ROM solutions at different
times. The space-time POD is able to reproduce the fields, no matter the temporal instance
considered. Furthermore, we ensure the accuracy of the strategy used with the log-error plot
in Figure 4.8. The considered error is averaged over a testing set of 50 uniformly distributed
parameters: the minimum value reached is around 10−4 for the state and the control variable,

Figure 4.8. Averaged relative log-error for the variables.

 4.5. Conclusions 95

while the adjoint relative log-error is around 10−3 . In this case, we have a remarkable speedup
of order O(105 ), indicating the convenience of using ROMs in a time-dependent optimization
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

setting. The saved computational time can be used to solve many parametric instances and to
completely and deeply analyze the system.

4.5 Conclusions
This chapter summarized ROM approaches for OCP(µ)s and presented them as a reliable and fast
tool to deal with several state equations, including steady linear, nonlinear, and time-dependent
problems. The problems were described in their continuous and discrete FE formulation (steady
or space-time), and the ROM strategies were validated with several test cases in fluid dynamics
and environmental sciences. We presented the following examples:

• steady linear: control of a pollutant release in the Gulf of Trieste;

• steady nonlinear: a forecasting framework of the North Atlantic Ocean dynamic;
• unsteady linear: boundary control for Graetz flow.
The OCP(µ)s were tackled by a partitioned POD approach suited even for very complicated
equations, such as nonlinear PDE(µ)s. We aimed at showing how ROMs could be useful in the
context of parametrized optimization, usually characterized by demanding computational sim-
ulations. We underline that this contribution is a brief overview of the capabilities of ROM
approaches to OCP(µ)s. Indeed, we aimed to state how useful ROMs can be in the parametric
analysis of optimization problems, most of all in a time-dependent setting where large speedup
values are reached. The interested reader may refer to [560] for a complete overview of marine
coastal management applications, to [98] for an extension of the presented results in an uncer-
tainty quantification framework, or to [452] for an analysis of OCP(µ)s in a bifurcating context.
Future development might deal with nonintrusive techniques in order to lighten the POD ap-
proach, which can be costly for time-dependent and nonlinear problems. Some first steps have
been taken in this direction; see, e.g., [172] and [153] for the application of dynamic mode de-
composition and physics informed neural networks to OCP(µ)s. Moreover, there is the interest in
these nonintrusive strategies being used to better understand the input-output relation of optimal
control pipelines.
 Chapter 5
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Reduced Basis Approaches

to Bifurcating Nonlinear
Parametrized Partial
Differential Equations

Federico Pichi, Francesco Ballarin, Gianluigi Rozza

5.1 Introduction
This chapter is concerned with analyzing and developing efficient reduced order models (ROMs)
to investigate complex bifurcating phenomena represented by nonlinear parametrized partial dif-
ferential equations (PDEs) in many physical contexts, including continuum and quantum me-
chanics and fluid dynamics [446].
Indeed, reconstructing the bifurcation diagrams, which highlight the singularities of the equa-
tions and the possible coexisting states, requires a huge computational effort, especially in the
multiparameter context. To overcome this issue, we developed a reduced order branchwise al-
gorithm for the efficient investigation of such complex behavior, with a focus on the stability
properties of the solutions. We applied our approach to the Von Kámán equations for buckling
plates [451], the Gross–Pitaevskii equations in Bose–Einstein condensates [450], the hyperelastic
models for bending beams [449], and the Navier–Stokes model for flow in a channel [452, 447].
Despite the huge discrepancies among the aforementioned models, we considered a unique
abstract framework in which to analyze all the different complex physical behaviors. The pres-
ence of bifurcating phenomena usually requires fine mesh discretization of the domain and, to-
gether with the parameter dependence, leads to a many-query context that could be a bottleneck
for a deep study of the problem under investigation.
In fact, when dealing with bifurcations, one needs to compute the numerical approximation
of the solution for many instances of the parameter in order to discover the critical points of the
model and its postbifurcating behavior. Moreover, the analysis becomes even more complicated
if the goal is a complete reconstruction of all the possible solutions that the model could admit.
Finally, a key aspect in these contexts is the stability analysis of the discovered solutions, which
can be studied by means of specific eigenvalue problems.
The numerical approximation of such models requires a combination of different method-
ologies. To obtain a high-fidelity approximation of the problem, we considered as a full-order
model (FOM) the finite element method (FEM), which we combined with a continuation method,
to follow the solution by varying the parameter, and with the Newton method, which is used to
linearize the equation.

97
 98 Chapter 5. Reduced Basis Approaches to Bifurcating Nonlinear Parametrized PDEs

We tackled the huge computational cost of this investigation using ROMs. In particular, we
developed a branchwise algorithm based on the proper orthogonal decomposition (POD) tech-
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

nique, which aims to efficiently reconstruct the bifurcating behavior by projecting the governing
equations into a low-dimensional manifold spanned by global basis functions (with respect to the
parameter space).
Moreover, several issues and questions arose when dealing with the approximation and re-
duction of bifurcating phenomena. In particular, we developed a reduced order approach for
the deflated continuation method to efficiently discover new solution branches [454]. We pro-
posed and discussed different optimal control problems (OCPs) to steer the bifurcating behavior
toward desired states [452]. We investigated the evolution of bifurcations in the multiphysics
context by means of fluid-structure interaction (FSI) problems [337]. Finally, we used a neural
network approach based on the proper orthogonal decomposition (POD-NN) as an alternative to
the empirical interpolation method (EIM) to develop a reduced manifold–based algorithm for the
efficient detection of the bifurcation points [447, 448]. For the sake of exposition, we will not
treat these latest advances here, and we redirect the reader to the references reported above.

5.2 Nonlinear Analysis of PDEs and Bifurcation Problems

In this section, we will briefly introduce the mathematical setting needed for the subsequent
analysis. This will serve us later when we consider a unified framework in which to cast all the
different models that we have studied. Therefore, here we will review the main definitions for
the functional setting of PDEs in a general, parametric, and nonlinear context. Moreover, we will
focus on their singular points, which are the keystones of the bifurcation analysis.

5.2.1 Mathematical Formulation

Here, we recall the basic definitions for the standard functional setting in nonlinear analysis. We
are interested in the case of parametrized PDEs; thus, let P ⊂ RP be the parameter space, a
closed and bounded subset of the space RP with P ≥ 1. A parameter µ is a point in the space P
that can contain physical or geometrical information about the system. Then, we introduce the
(reference) domain Ω ⊂ Rd , where d = 2, 3 is the spatial dimension.
Although time-dependent problems are of great interest, we will focus on steady problems,
investigating the features of their equilibrium solutions. We can thus define the Hilbert space
.
X = X(Ω) and its dual space X0 as the space of linear and continuous functionals over X. The
1/2
Hilbert space is equipped with the norm ||V ||X = (V, V )X for any V ∈ X, induced by its inner
product (·, ·)X , and we denote the duality pairing between X0 and X by
.
hG, V i = hG, V iX0 X ∀ G ∈ X0 , V ∈ X.

Let us denote by G : X × P → X0 the parametrized mapping which represents the nonlinear

PDE and thus the model under investigation.
The strong form of the parametric PDE problem reads as follows: given µ ∈ P, find X(µ) ∈
X such that
G(X(µ); µ) = 0 ∈ X0 . (5.1)
To study the solutions of the abstract equation above, a fundamental step is to analyze its vari-
ational formulation, which will play a key role in the numerical approximation of the problem.
Thus, we introduce the parametrized variational form g(·, ·; µ) : X × X → R as

g(X, Y ; µ) = hG(X; µ), Y i ∀ X, Y ∈ X, (5.2)

 5.2. Nonlinear Analysis of PDEs and Bifurcation Problems 99

and the weak formulation of equation (5.1) reads as follows: given µ ∈ P, find X(µ) ∈ X such
that
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

g(X(µ), Y ; µ) = 0 ∀ Y ∈ X. (5.3)
In order to completely describe the problem formulation in (5.1), we need some suitable
boundary conditions (BCs) on the domain Ω. Here, we will cover the cases of both homo-
geneous and inhomogeneous Dirichlet and Neumann conditions, depending on the test case at
hand. Moreover, these BCs will automatically be embedded in (5.3) by an appropriate choice of
the space in which to seek the solution.
Having introduced the mathematical formulation of the problem, a starting point for the
analysis is to investigate its well-posedness. A PDE is said to be a well-posed problem if it
is characterized by the existence and uniqueness of the solution. Indeed, these properties are
also fundamental for the numerical approximation of the model. Although these assumptions are
often required, dealing with more realistic models can complicate the setting. In fact, one has to
find a trade-off between the complexity of the model and its effective tractability.
As we will see later on, complex nonlinear and parametric PDEs are proved to be more
accurate in describing the physics, but the lack of good mathematical properties makes them
difficult to approach, both theoretically and numerically.
Hence, in the following, we investigate when such properties are conserved and thus when
the problem admits a unique solution. On the contrary, in the next subsection, we will discuss
the presence of critical points in which the well-posedness is lost, understanding how the model
behaves and evolves in such cases.
Let us assume that the map G in equation (5.1) is continuously differentiable with respect to
X and µ. Let (X; µ) ∈ X × P be a known solution, i.e., G(X; µ) = 0. Then, we denote by
DX G(X; µ) : X → X0 and Dµ G(X; µ) : P → X0 the partial derivatives of G on a generic point
(X, µ) ∈ X × P. When dealing with the approximation of (5.1), the strong assumption usually
found in literature, which ensures the well-posedness of the problem, is that DX G(X; µ) : X →
X0 is bijective. Indeed, the following result holds.

Theorem 5.1. Let Br (X), Br (µ) be two balls of radius r and r around X and µ, respectively.
Let G : X × P → X0 be a C 1 map, and assume that
(1) G(X; µ) = 0 for the solution (X, µ) ∈ X × P, and
(2) DX G(X; µ) : X → X0 is bijective.
Then there exist r, r > 0 and a unique solution X(µ) ∈ Br (X) ∩ X such that
G(X(µ); µ) = 0 ∀µ ∈ Br (µ) ∩ P.
While this is a straightforward application of the implicit function theorem [129, 642], it
ensures the well-posedness of the problem and thus the existence of a local branch of nonsingular
solutions. Furthermore, we can rewrite the Fréchet partial derivatives of G on (Z, µ) ∈ X × P
with respect to X as
dg[Z](X, Y ; µ) = hDX G(Z; µ)X, Y i ∀X, Y ∈ X, (5.4)
where we have introduced the parametrized variational form dg[Z](·, ·; µ). We now observe that
assumption (2) in Theorem 5.1 can be reformulated in terms of the variational form dg[Z](·, ·; µ)
itself, which is said to be
(2a) continuous on X × X if there exists a continuity constant γ > 0 such that
dg[Z](X, Y ; µ)
γ(µ) = sup sup >γ ∀µ ∈ P, (5.5)
X∈X Y ∈X kXkX kY kX
 100 Chapter 5. Reduced Basis Approaches to Bifurcating Nonlinear Parametrized PDEs

(2b) inf-sup stable on X × X if there exists an inf-sup constant β > 0 such that
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

dg[Z](X, Y ; µ)
β(µ) = inf sup ≥β ∀µ ∈ P (5.6)
X∈X Y ∈X kXkX kY kX

and
dg[Z](X, Y ; µ)
inf sup >0 ∀µ ∈ P. (5.7)
Y ∈X X∈X kXkX kY kX

We remark that conditions (5.6) and (5.7) are equivalent to the injectivity and the surjec-
tivity of the Fréchet derivative DX G(Z; µ), respectively. In particular, condition (5.6) can be
reformulated as
∃β(µ) > 0 : kDX G(Z; µ)XkX0 ≥ β(µ)kXkX , (5.8)
i.e., the Fréchet derivative of G is bounded below. It is clear that (5.8) holds if and only if
DX G(Z; µ) is injective. On the other hand, condition (5.7) can be easily restated as the injec-
tivity of the adjoint operator of DX G(Z; µ), which can be expressed in terms of its variational
form dg as
Y ∈ X, dg[Z](X, Y ; µ) = 0 ∀X ∈ X ⇒ Y = 0. (5.9)
Therefore, we observe that condition (5.7) is actually stating that DX G(Z; µ) is surjective, since
the injectivity of the adjoint of an operator is equivalent to the surjectivity of the operator itself.
Moreover, when dealing with linear parametrized PDEs, it is clear that (e.g., for an unforced
problem)
dg[Z](X, Y ; µ) = g(X, Y ; µ) ∀X, Y, Z ∈ X, ∀µ ∈ P. (5.10)
Hence, the general assumptions in (5.5), (5.6), and (5.7) fall back on the standard ones required
by the well-known theorems of Lax–Milgram and Nečas [186, 520].
In conclusion, the bijectivity assumption in Theorem 5.1, and equivalently the injectivity and
surjectivity conditions (5.8) and (5.9), are the key ingredients of a well-posed problem. In the
following, we will see what happens when these hypotheses are no longer valid at some point in
the parameter space.

5.2.2 An Overview of Bifurcation Theory

We are now ready to briefly describe bifurcating behaviors and their mathematical investigation.
Before moving to the theory, let us start with a nonmathematical introductory explanation.
In the world around us, we observe many phenomena that involve either gradual or sudden
changes, and we characterize them as quantitative or qualitative, respectively. As an example,
we can think about the supported beam under compression depicted in Figure 5.1.
When the load applied to the beam, denoted by µ, is sufficiently small, we observe a slight
deformation that corresponds to a quantitative change. Of course, this deformation depends on

Figure 5.1. A beam subject to a load and its buckled configuration.

 5.2. Nonlinear Analysis of PDEs and Bifurcation Problems 101

the magnitude of the load and on the specific material properties of the beam. Hence, we expect
that a small perturbation of µ leads to a configuration with the same qualitative features.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

At the same time, we empirically know that there exists a critical value for the load, which
we call µ∗ , after which the model does not preserve the original qualitative behavior. Practically
speaking, if the load is big enough, the beam cannot sustain the compression and it buckles. This
is a clear example of a qualitative change.
We call bifurcation phenomena the qualitative changes that are not stable under small per-
turbations of a parameter µ, and bifurcation points the points µ∗ at which such changes occur
[535, 342, 355]. Usually, we can distinguish between different qualitative states by means of
their geometrical shape or pattern configuration. Indeed, two of the most relevant features that
may change in the presence of a bifurcation phenomenon are the stability and the symmetry of
the resulting states.
The aim is to investigate the solutions of a specific model by varying a one-dimensional
parameter, i.e., the bifurcation parameter, which is responsible for the bifurcating phenomena.
Correlated with the qualitative change of the solutions is the nonuniqueness behavior; in fact,
as we will see, in such situations the model admits different solutions for the same value of the
parameter. We will refer to the set of solutions with the same qualitative properties as a branch.
As we said before, when the parameter changes slightly, we expect a stable solution to evolve
continuously in a unique manner.
The situation changes drastically if the inf-sup stability of the model, ensured by (5.6) and
(5.7), is lost. This happens when the bifurcation parameter reaches a critical value µ∗ for which
the system admits the existence of a qualitatively different solution that bifurcates from the pre-
vious stable one.
Following [17, 16], we consider the problem (5.1) and assume that X = 0 is a solution for
every parameter µ ∈ P, namely

G(0; µ) = 0 ∀µ ∈ P,

which will be referred to as the trivial solution. Then, we can define the set of nontrivial solutions
of (5.1) as
S = {(X, µ) ∈ X × P | X 6= 0, G(X; µ) = 0} . (5.11)
For the sake of simplicity, we chose to present here the bifurcations that originate from the
trivial solution. Indeed, many times, and in particular among the test cases we treated, we have
a nontrivial ground state solution, which loses its uniqueness by branching to another nontrivial
state.
It is remarkable to note that this situation can happen in a nested fashion. Increasing in
complexity, one can find a bifurcation emerging from an already bifurcated configuration. This
phenomenon is usually called secondary bifurcation; an example will be shown in Subsection
5.4.3.
As we understood from the example reported at the beginning, there could exist some values
for the parameter µ such that one or more nontrivial solutions branch off from the trivial one.
These are the critical points that we have introduced before, and in the context of bifurcation
theory we define them as in [17, 16].

Definition 5.2. A parameter value µ∗ ∈ P is a bifurcation point for (5.1) if there exists a
sequence (Xn , µn ) ∈ X × P, with Xn 6= 0, such that

◦ G(Xn ; µn ) = 0,

◦ (Xn , µn ) → (0, µ∗ ).
 102 Chapter 5. Reduced Basis Approaches to Bifurcating Nonlinear Parametrized PDEs

Bifurcation is thus a paradigm for nonuniqueness in nonlinear analysis, and a necessary con-
dition is the failure of the implicit function theorem.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

In fact, we highlight that the inf-sup stability property, which should guarantee the well-
posedness of the PDE, is actually no longer valid for bifurcating phenomena. Indeed, at the bi-
furcation point µ∗ the inf-sup constant β becomes zero, and the Fréchet derivative DX G(X; µ)
fails to be invertible. Therefore, unless otherwise specified, we will talk about well-posed prob-
lems as long as µ 6= µ∗ .
The aim of bifurcation theory is thus to provide a mathematical description of the bifurcating
scenario that can be observed in physical systems and experiments. The understanding of the
global behavior, from both the theoretical and the numerical viewpoint, is usually supported by
reconstructing some scalar characteristic value, which we will denote by s(X(µ)).
Given (X(µ), µ), a solution of (5.1), for each instance of the parameter, one can draw a plot
of s(X(µ)) versus µ, which in the literature is known as a bifurcation diagram. Therefore, the
existence of different solutions for the same values of the parameter will result in the presence
of multiple branches on the diagram. Different choices of s(·) result in possibly different plots;
thus, to avoid confusion, the function has to be chosen carefully and has to represent the main
features while changing the qualitative behavior of the resulting configuration.
It is clear that there exist many different bifurcation phenomena, and usually complex non-
linear problems are characterized by one (or more) of those. In fact, the literature is full of
examples of physical systems with qualitatively changing behavior. Each one of these bifurcat-
ing phenomena has a specific peculiarity, but in the field of bifurcation theory they are usually
categorized by some common features. Among others, the most known ones are turning point
(or fold) bifurcations, transcritical bifurcations, pitchfork bifurcations, and Hopf bifurcations.
While the latter are usually of interest when dealing with time-dependent problems, here we will
focus on models that exhibit pitchfork bifurcations.

5.3 Numerical Approximation of Bifurcating Phenomena

As we said in the previous section, we are interested in nonlinear and parametric PDEs where the
solution for a given parameter µ exists, but may not be unique [76, 93]. Indeed, the local (with
respect to the parameter) well-posedness of the problem (5.1) relies on the local spectral property
of its Fréchet derivative DX G(·; ·), and if it fails to fulfill certain assumptions, the behavior of
the model has to be investigated more carefully.
In the bifurcating context, we can mimic the definition of the set of nontrivial solutions in
(5.11), introducing the terminology of solution branch as the set of solutions with the same
qualitative features for different values of the parameter µ. Hence, we will denote a solution
branch by
M = {X(µ) ∈ X | G(X(µ); µ) = 0, µ ∈ P}. (5.12)
Since we are dealing with the nonuniqueness of the solution at fixed µ, we expect the system to
be characterized by the existence of multiple solution branches. We can thus define the set of all
the existing solutions to (5.1) as the set of solution branches
k k
. [ [
X= Mi = {Xi (µ) ∈ X | G(Xi (µ); µ) = 0, µ ∈ P}. (5.13)
i=1 i=1

We remark that when µ ∈ / P, the case k = 1 can be seen as the well-posed problem, and X is
often referred to as the solution manifold [281].
Since we are interested in recovering the full bifurcation diagram for the model under consid-
eration, the numerical approximation deals with the computation of the discretized counterpart
 5.3. Numerical Approximation of Bifurcating Phenomena 103

of (5.13). To this end, we pursue a branchwise approach. In other words, we aim to reconstruct
one fixed branch M ⊂ X at time, corresponding to some i ∈ {1, . . . , k}. The global behavior is
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

thus obtained by approximating its solutions for all the values in P and then moving to the next
solution branch in X .
We will refer to these steps as the high-fidelity approximation, since we are dealing with high-
dimensional and time-consuming procedures. In order to avoid this computational complexity, in
the second part, we will present the ROMs, or low-fidelity approximations, that we have applied
to reduce the dimensionality of the system, while allowing for an efficient approximation of the
bifurcating phenomena using a reduced version of the branchwise algorithm.

5.3.1 High-Fidelity Approximation

The strategy we have developed to reconstruct and analyze the whole bifurcation diagram is
based on a combination of three well-studied methodologies: the Galerkin FEM, the Newton–
Kantorovich method, and a continuation method. These will be used to discretize, linearize, and
continue each branch. Now, we briefly introduce the standard notation for the discretization of a
generic PDE of the form (5.1) with the Galerkin FEM [477, 468].
Let XN be a finite-dimensional space such that XN ⊂ X of dimension dim(XN ) = N < ∞,
where N denotes the number of high-fidelity degrees of freedom. Then, we can go back to the
weak form (5.3) of the equation (5.1) and project it into the finite-dimensional vector space XN .
The Galerkin FEM reads as follows: given µ ∈ P, we seek XN (µ) ∈ XN that satisfies
hG(XN (µ); µ), YN i = g(XN (µ), YN ; µ) = 0 ∀ YN ∈ XN . (5.14)
The well-posedness of the discrete problem (5.14) is not inherited from the continuum one,
even assuming the inclusion property XN ⊂ X. Indeed, while it is clear that the continuity
hypothesis (5.5) is still satisfied at the finite-dimensional level, the same is not true for the inf-sup
stability, where an additional assumption is required for the discrete stability of the variational
form dg. From the algebraic point of view, we denote by {E j }N j=1 a Lagrangian basis for XN ,
so that we can write every element XN (µ) ∈ XN as
N
(j)
X
XN (µ) = XN (µ)E j (5.15)
j=1

(j)
and denote the solution vector by XN (µ) = {XN (µ)}N
j=1 . We then go back to the study of the
solution XN (µ) ∈ RN of the nonlinear system
GN (XN (µ); µ) = 0 in RN , (5.16)
where the high-fidelity residual vector GN is defined as
(GN (XN (µ); µ))i = g(XN (µ), E i ; µ) ∀ i = 1, . . . , N .
The nonlinear solver chosen to linearize the weak formulation in (5.14) is the well-known
Newton–Kantorovich method [129, 470], which reads as follows: given µ ∈ P and choosing
an initial guess XN0
(µ) ∈ XN , for every k = 0, 1, . . . , we seek the variation δXN ∈ XN such
that
k k
dg[XN (µ)](δXN , YN ; µ) = g(XN (µ), YN ; µ) ∀ YN ∈ XN . (5.17)
Then we update the solution at the iteration k + 1 as
k+1 k
XN (µ) = XN (µ) − δXN
and repeat these steps until an appropriate stopping criterion is verified.
 104 Chapter 5. Reduced Basis Approaches to Bifurcating Nonlinear Parametrized PDEs

The convergence of the Newton–Kantorovich method is ensured by the well-known theorem

reported in [642, 129]; hence, its kth step combined with the Galerkin FE discretization and
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

applied to (5.16) reads as follows: given fixed µ ∈ P, find δXN ∈ RN such that

JN (XkN (µ); µ)δXN = GN (XkN (µ); µ) , (5.18)

where the Jacobian matrix in RN ×N is defined as

(JN (XkN (µ); µ))ij = dg[XkN (µ)](E j , E i ; µ) ∀ i, j = 1, . . . , N . (5.19)

Therefore, once we have fixed µ ∈ P, we are able to approximate a solution XN (µ) and pos-
sibly discover its stability properties using some suitable eigenvalue problems [451, 452], but
our aim is to investigate the evolution of the solutions by following its branching behavior.
Moreover, detecting the solution path is the key to recovering the bifurcation diagram, since
it provides a global picture of the model behavior. In view of this need, we introduce the con-
tinuation methods [14, 170, 169, 334], which allow us to generate a sequence of solutions cor-
responding to the selected values of the parameter in order to construct branches of possible
configurations.
The main ingredients for a continuation method are (i) the selection of a good initial guess,
(ii) a rule to select the step size ∆µ, and (iii) a detection tool for the bifurcation points. In
particular, we are interested in predictor-corrector methods, which represent a wide class of
techniques that help the branch tracing by splitting the approximation in two steps, as follows.
Let us start from a solution XN (µj ) to (5.16) for the parameter value µj ; we want to approximate
the solution XN (µj+1 ) for the corresponding next value µj+1 . Instead of directly computing the
pair (XN (µj+1 ), µj+1 ), a predictor-corrector method acts as follows: in the first step the pair
(Xe N (µ e j ) is constructed from (XN (µj ), µj ), and then the former, without being a solution
e j ), µ
for µj+1 , merely serves as an initial guess in the second step.
Since we always deal with pitchfork bifurcations, we will restrict ourselves to the simple
continuation method, which can be seen as a basic predictor-corrector scheme where the pair
(Xe N (µ e j ) is given by (XN (µj ), µj + ∆µj ). We can finally present the algorithm we have
e j ), µ
developed to deal with the branchwise reconstruction of the bifurcating behavior of the models.
Algorithm 5.1 is the implementation result of the building blocks needed to linearize, discretize,
and continue each solution branch; thus, let us now review the combination of these method-
ologies. At the very beginning, one implicitly chooses the branch to approximate by choosing
the initial guess to start the iteration of the Newton method. Indeed, the simplest way to guide
the nonlinear solver to a desired branch is by the choice of the initial guess. Hence, in order
to fully recover the branching behavior, when dealing with the simple continuation method (in
which the parameter step length is fixed and already prescribed), we consider a discrete version
of the parameter space PK = [µ1 , . . . , µK ] ⊂ P of cardinality K and loop over this ordered set.
The loop serves to mimic the predictor-corrector method, assigning the solution obtained for a
given parameter µj as the initial guess for the nonlinear solver at the next iteration of µj+1 . This
allows us to follow the bifurcating behavior of the model.
For the actual linearization and discretization of equation (5.16), we then apply the Newton–
Kantorovich method (5.18) in combination with the Galerkin FEM. The former helps to linearize
the system around the approximation of the solution at the kth iteration, while the latter projects
the problem into the finite-dimensional space XN . As we said, having computed a solution Xj of
the problem for the parameter µj , we can also investigate its stability properties by solving the
eigenvalue problem, which involves the Jacobian matrix JN and the inner product matrix MN .
This will allow us to understand the physical stability of the approximated solutions and to detect
the bifurcation points connected to the qualitative changes of the model.
 5.3. Numerical Approximation of Bifurcating Phenomena 105

A LGORITHM 5.1. A pseudocode for the branchwise procedure.

1: X0 = Xguess . Initial guess
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

2: for µj ∈ PK do . Continuation loop

(0)
3: Xj = Xj−1 . Continuation guess
(k)
4: while kGN (Xj ; µj )kXN >  do . Newton method
(k) (k)
5: JN (Xj ; µj )δX = GN (Xj ; µj ) . Galerkin FEM
(k+1) (k)
6: Xj = Xj − δX
7: end while
8: JN (Xj ; µj )Xe = σµj MN Xe . Eigenproblem for stability
9: end for

5.3.2 Reduced Basis Approximation

As we have seen in the previous subsection, the preliminary step for the discretization of (5.1) is
to project the weak formulation (5.3) in a finite-dimensional setting, which results in the Galerkin
problem (5.14).
Despite the increasing computational resources available, finding the solutions to this prob-
lem is very challenging because of the potential high number of degrees of freedom N . More-
over, we recall that in addition to the many-query context and the nonlinearity of the models
considered, the existence of multiple solutions complicates the setting, requiring a deep investi-
gation of the system behavior with respect to the parameter space.
For this reason, here we focus on a specific methodology in the class of ROMs called the
reduced basis (RB) method. It consists of a projection of the high-fidelity problem on a subspace
of smaller dimension, constructed with some properly chosen basis functions, and therefore it
shares many features with the Galerkin FEM.
In practice, the key feature of the RB method is the adoption of the offline-online paradigm.
Indeed, in order to retrieve a solution to (5.14) for a given parameter µ ∈ P, the computation is
divided into two steps:
◦ In the offline phase, approximated solutions to (5.14) are computed with a high-fidelity
method (finite element, spectral element, finite volume, finite difference), corresponding to
selected representative parameter values/system configurations, and stored together with
other information about the parametrized problem. This is the computationally expen-
sive step, and it can benefit from possibly available high-performance computing (HPC)
facilities.
◦ In the online phase, the preprocessed information obtained during the offline phase is
assembled and used to compute the solution for each new instance of the parameter in a
short amount of time (ideally in real time), even on a relatively low power device such as
a laptop or a smartphone.
This split in the computational procedures is built so that new parameter-dependent quantities
can be easily and quickly computed online, while representative basis functions for selected
parameter values and more demanding quantities are precomputed offline.
Thus, the main goal of the offline phase is to construct a low-dimensional basis for a discrete
manifold XN ⊂ XN , called the reduced manifold, which we assume to be a good approximation
of the high-fidelity manifold XN .
This entails solving Ntrain times the Galerkin high-fidelity problems associated with Ntrain
values of µ in P. The obtained solutions {X(µi )}N train
i=1 are usually called snapshots. Having
recovered the information about the µ dependence of the solutions, we need to process these
snapshots in order to obtain a basis that spans the reduced manifold XN .
 106 Chapter 5. Reduced Basis Approaches to Bifurcating Nonlinear Parametrized PDEs

Among the others, the POD and the greedy techniques [281, 440, 470] are the most well
known and used ones. Let us assume for now that such a basis is already built, and let us see how
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

the reduced problem is constructed through the projection on the subspace spanned by it.
For the online phase, it is the efficient and reliable part where the solutions are computed
through the projection on XN . The complexity reduction, which enables the efficiency, is based
on two main assumptions:

(i) The affine decomposition holds, i.e., one can rewrite the weak formulation as a linear com-
bination of µ-independent forms and µ-dependent coefficients. Hence, the contribution of
the parameter is entirely encoded by the coefficients, which usually allows us to rapidly
assemble the system, relying on the precomputed parameter-independent quantities.

(ii) We can accurately approximate the discrete manifold XN with a much lower dimension
space XN , i.e., we need only a small number N  N of basis functions.

Therefore, the reduced computational cost comes mainly from avoiding projecting on the large
FE manifold, while relying on the small RB one. It is clear that, in our context, assumption
(i) is difficult to fulfill because of the µ dependence of the solution around which we linearize
the nonlinear weak formulation (5.16). Assumption (ii) is instead usually linked with the con-
cept of Kolmogorov n-width [393, 133], which expresses the ability of a reduced manifold to
approximate the high-fidelity one.
Let us now describe the details of the online phase. We consider the discrete weak formula-
tion (5.14) and, projecting it into the reduced space XN , we obtain the following problem: given
µ ∈ P, we seek XN (µ) ∈ XN such that

g(XN (µ), YN ; µ) = 0 ∀YN ∈ XN . (5.20)

As before, we can apply the Newton–Kantorovich method: choosing an initial guess XN

0
(µ) ∈
XN , for every k = 0, 1, . . ., find the variation δXN ∈ XN such that
k k
dg[XN (µ)](δXN , YN ; µ) = g(XN (µ), YN ; µ) ∀YN ∈ XN , (5.21)
k+1
and then update the solution as XN (µ) = XNk
(µ) − δXN until an appropriate stopping crite-
rion is verified.
Let {Σm }N m=1 be a given orthonormal basis for XN with respect to the inner product defined
on the space XN . Then, the reduced manifold is defined as

XN = span{Σ1 , . . . , ΣN };

therefore, we can express every element XN (µ) ∈ XN as

N
(m)
X
XN (µ) = XN (µ)Σm . (5.22)
m=1

(m)
From an algebraic standpoint, we can denote by XN (µ) = {XN (µ)}N m=1 ∈ R the reduced
N

solution vector. Plugging (5.22) into (5.20) and choosing YN = Σ ∈ XN for 1 ≤ n ≤ N , we

n

obtain the algebraic system

N
!
(m)
X
g XN (µ)Σm , Σn ; µ =0, n = 1, . . . , N . (5.23)
m=1
 5.3. Numerical Approximation of Bifurcating Phenomena 107

We will denote by GN the reduced counterpart of GN defined by

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

N
!
(m)
X
(GN (XN (µ); µ))n = g XN (µ)Σm , Σn ; µ , (5.24)
m=1

which we will refer to as the reduced residual vector. Moreover, we will denote by V ∈ RN ×N
the transformation matrix whose elements

(V)jm = Σm
(j) (5.25)

are the nodal evaluation of the mth basis function at the jth node. With this notation, we can
rewrite problem (5.23) as
VT GN (VXN (µ); µ) = 0. (5.26)
Finally, the algebraic form of the Newton method, combined with the RB technique, provides
the following formulation: at every iteration k we seek δXN ∈ RN such that

JN (XkN (µ); µ)δXN = GN (XkN (µ); µ), (5.27)

where JN is the RN ×N reduced Jacobian matrix

JN (XkN (µ); µ) = VT JN (VXkN (µ); µ)V. (5.28)

For a general nonlinear problem, equation (5.27) still involves the degrees of freedom of
the high-fidelity problem N . Because of this, the repeated assembly of the reduced Jacobian
compromises the efficiency of the reduced order method during the online phase. As we will see
later, this issue can be overcome by adopting a class of affine-recovery techniques, which allow
consistent speedups by interpolating the nonlinear part of the variational form.
Having described the online simulation, we now go back to the building process of the re-
duced manifold XN .

POD
The key point for the application of the RB method is to construct the basis. Here we focus on
POD [34, 316], which is a compression strategy, closely related to singular value decomposition
(SVD) and principal component analysis (PCA), that serves to reduce the dimensionality of a
given dataset. Indeed, starting from sufficiently rich information about the system, it allows us
to extract the main features by a lower-dimensional representation given by the first computed
modes, which ideally retain most of its energy.
In our applications, we use the POD to generate the RB space XN . Moreover, this is
proved to be optimal in the l2 (RN ) sense. Therefore, in order to construct XN , we consider
Ptrain = {µ1 , . . . , µtrain } ⊂ P and build from the corresponding solutions {XN (µj )}N train
j=1
a symmetric and linear correlation operator C : XN → XN . Studying the associated eigenvalue
problem and sorting the eigenvalue-eigenfunction pairs in descending order, we can use the first
N modes to build the reduced space XN . Moreover, we can choose the minimal integer N such
that the retained energy from the last (Ntrain −N ) snapshots is less than a fixed tolerance P OD ;
this means that we are finding the basis which minimizes, over all possible N -dimensional or-
thonormal bases ZN , the error between the snapshots and their projection through ZN .
It is clear that one of the disadvantages of this technique is the huge number of high-fidelity
solutions that have to be computed to obtain a fairly accurate representation of the high-fidelity
manifold. Indeed, if we miss some information at the first discretization level, the RB, after the
compression step given by the POD, will usually not contain it.
 108 Chapter 5. Reduced Basis Approaches to Bifurcating Nonlinear Parametrized PDEs

Hyperreduction Strategies
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

As we said, one of the main assumptions to obtain an efficient reduction method is that the
forms in (5.14) have an affine parametric dependence. This is crucial since it serves to ob-
tain an N -independent online phase, in which we can efficiently compute the solution of the
parametrized PDE at each new instance of the parameter µ. Despite this, when dealing with
nonlinear problems, this assumption is still not sufficient because the residual (5.16) depends on
the parametrized solution itself. Indeed, it is clear that we cannot recover an efficient offline-
online decomposition of

QG
q
(µ)VT GqN (VXN ),
X
GN (XN ; µ) = θG (5.29)
q=1

since its assembly and projection still involve the N degrees of freedom, compromising the
efficiency.
Among others, reliable alternatives to achieve an efficient online phase come from a hyperre-
duction approach. These techniques include EIM [40, 385] and its variant, the discrete empirical
interpolation method (DEIM) [116], which aim to make the online computations independent of
the number of degrees of freedom of the chosen full-order discretization method. The hyperre-
duction techniques serve to approximate the reduced residual vector as

QG
cqθ (XN ; µ)VT hq ,
X
GN (XN ; µ) ≈ (5.30)
q=1

where {hq }Q q
q=1 represents a suitable basis and cθ are the interpolation coefficients, which include
G

the contribution of the µ-dependent terms.

A Reduced Branchwise Algorithm

Here, we want to describe the main differences between the high-fidelity Algorithm 5.1 and its
reduced variant for the efficient reconstruction of the bifurcation diagram.
First of all, a hyperreduction technique can be applied for the affine-recovery step. Then, to
construct the reduced manifold, we implemented the POD approach described above. To help the
convergence and prevent unnecessary Newton iterations, we select the snapshot location using
the continuation method, which in its simplest variant results in an equispaced sampling of P.
A huge number of snapshots have to be computed (as usual when dealing with POD, but here
it depends also on the number of branches to be approximated). Once the offline phase is finished,
we compute a (global or branch-specific) basis for the reduced manifold; thus we combine the
projection step on XN , the Newton method (5.21), and the simple continuation method for the
reduced vector solution XN .
For the theoretical rationale behind the RB method [470, 281], it has been proved that one of
the main ingredients to have good approximation properties is the regularity of the solution as a
function of the parameter, rather than its regularity in space.
In fact, this can be an issue for bifurcation problems, where, e.g., for pitchfork phenomena
the critical points represent a discontinuity in the parametric sensitivity ∂X(µ)/∂µ. Hence, we
expect that the error analysis in the parameter space P will show higher peaks at bifurcation
points. We remark that in a black-box approach, one can utilize this statistic to forecast the
location of bifurcating phenomena.
 5.4. Von Kármán Equations for Structural Buckling of Plates 109

5.4 Von Kármán Equations for Structural Buckling of

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Plates
In the context of continuum mechanics, von Kármán proposed a model to describe all the possible
configurations that a plate under compression can assume [598].
The von Kármán model is often used to describe buckling phenomena. A buckling phe-
nomenon is an example of bifurcating behavior, where the system suddenly changes its configu-
ration. This section is mainly based on the work done in [451].

5.4.1 Von Kármán Model

Let us consider a rectangular plate Ω = [0, L] × [0, 1] in its undeformed state, subject to a
µ-parametrized external load acting on its edges, as depicted in Figure 5.2.

(0, 1) (L, 1)

(0, 0) (L, 0)

Figure 5.2. An elastic and rectangular plate compressed along the edges parallel to the y direction.

Then, the displacement from its flat state and the Airy stress potential, respectively u and φ,
satisfy the von Kármán equations

∆ u = [µh + φ, u] + f in Ω,
 2
(5.31)
∆2 φ = − [u, u] in Ω,

where h and f are the external forces acting on the plate, ∆2 is the biharmonic operator, and

. ∂ 2u ∂ 2φ ∂ 2u ∂ 2φ ∂ 2u ∂ 2φ
[u, φ] = 2 2
−2 +
∂x ∂y ∂x∂y ∂x∂y ∂ y 2 ∂ x2
is the bracket of Monge–Ampére. The von Kármán model is of fourth order, due to the presence
of the biharmonic operator; nonlinear, due to the product of second derivatives in the brackets;
and parametric, due to the buckling coefficient µ varying in a proper range of real numbers.
In order to completely describe the physics involved, we must supplement the system of
PDEs with some opportune BCs for both of the unknowns. Although they can be imposed in
many different ways (see, e.g., [128]), we present only the simply supported BCs

u = ∆u = 0 in ∂Ω,

(5.32)
φ = ∆φ = 0 in ∂Ω,

which are physically complex to reproduce but also the most used for the simulations because
of their versatility in the weak formulation. We remark that, despite the simple BCs chosen, the
goal of this work is to understand the bifurcation behavior of the von Kármán plate equation,
regardless of the numerical constraints that a conforming method for more involved boundary
conditions could impose.
 110 Chapter 5. Reduced Basis Approaches to Bifurcating Nonlinear Parametrized PDEs

5.4.2 Weak Formulation and Its Approximation

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Let us consider µ ∈ P ⊂ R, where P here is the one-dimensional bifurcation parameter space,

4
Ω ⊂ R2 is the rectangular domain that we identify with the plate, and X = X(Ω) = H01 (Ω)
is the Hilbert space in which we will seek the solution and X0 its dual space. Hence, we can
represent the nonlinear PDE given by the von Kármán system (5.31) as the parametrized mapping
G : X × P → X0 we defined in Subsection 5.2.1. Thus, its weak formulation reads as follows:
.
given µ ∈ P, find X(µ) = (u(µ), U (µ), φ(µ), Φ(µ)) ∈ X such that


 g(X(µ), Y ; µ) = a(u(µ), w) + b(U (µ), w) + a(U (µ), v)
+ µc(h, u(µ), v) + c(φ(µ), u(µ), v) + a(φ(µ), θ)

(5.33)

 + b(Φ(µ), θ) + a(φ(µ), ψ) − c(u(µ), u(µ), ψ)
∀Y ∈ X, ∀µ ∈ P,


where the following bilinear and trilinear forms have been introduced:
Z Z Z
a(η, ω) = ∇η · ∇ω dΩ , b(η, ω) = ηω dΩ , c(η, ω, ζ) = [η, ω] ζ dΩ .
Ω Ω Ω

Now we recast the weak formulation (5.33) in the finite-dimensional space XN , as we did in
general in (5.14).
Moreover, in order to embed the simply supported BCs at the discrete level, we set XN as
the space of test functions which vanish on the boundary. In this case, the matrix formulation of
the Newton method (5.27) reads as follows: given µ ∈ P and an initial guess XN 0
∈ XN , for
k = 0, 1, . . . until convergence, we seek δXN = (δuN , δUN , δφN , δΦN ) ∈ XN such that

AN ukN + BN UkN
    
AN BN 0 0 δuN
 C2 + µC0 C1N 0  δUN  k 1 k 0 k
AN  =  AN UN + C N uN + µCN uN
  
 N N  k k
,
 0 0 AN BN   δφN   AN φN + BN ΦN 
−C1N − C3N 0 0 AN δΦN AN ΦkN − C1N ukN

k+1
where, defining XkN = (ukN , UkN , φkN , ΦkN ), we set XN k
= XN − δXN . As we said, the RB
technique has the same structure as the FEM; thus at the reduced level we will have the same
formulation with the blocks obtained through the projection.

5.4.3 Numerical Investigation of Buckling Phenomena

In this section, we will show how the buckling phenomenon, i.e., the loss of uniqueness of the
solution, can be analyzed using the bifurcation diagram and the efficient reduced algorithm we
presented. We recall that we are interested in studying the solution while varying the buckling
parameter µ, which describes the compression along the edges parallel to the y-axis.
A preliminary investigation of the eigenproblem associated with the system gave us the
following necessary information: the parameters responsible for the buckling are in the inter-
val P = [35, 65], which we therefore chose as our parameter domain. The main goal is to
reconstruct the bifurcation diagram, which represents, for every value of µ ∈ PK on the x-
axis, the corresponding value of the displacement u in its point of maximum modulus, namely
s(u) = (sgnu) maxx∈Ω kuk∞ . We consider a rectangular plate with L = 2. The FE space
XN , defined on the domain Ω = [0, 2] × [0, 1], is discretized with P2 quadratic elements and
has similar dimension. We computed Ntrain = 20 snapshots for each branch during the offline
 5.4. Von Kármán Equations for Structural Buckling of Plates 111
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

s(u)

Figure 5.3. Left: high-fidelity bifurcation diagram for the rectangular plate with L = 2. Right:
high-fidelity displacements linked to the five branches at µ = 65.

phase, while for the reduced manifold we chose N = 8 bases. For the branch reconstruction, we
designed the continuation method with K = 60 equispaced points.
We can now present the bifurcation diagram in Figure 5.3, where the buckling phenomenon
from the trivial solution happens for µ1 ' 39.47.
Moreover, by properly choosing the initial guess, we were able to detect the second and the
third bifurcations for the plate, respectively, for values of µ2 ' 46.5 and µ3 ' 61.68. The
physical symmetry issue is evident in all buckling points; in fact every solution from the upper
branches is just the reflection with respect to the plane of the plate of the lower ones.
As reported in Figure 5.3, we can distinguish these solutions, and the branch to which they
belong, by their cell-like displacement. Moreover, we can observe that the buckling near µ3 has
a qualitatively different nature since two branches depart from the same point; for this reason it
is said to be a double bifurcation.
An interesting property of the von Kármán model is mode jumping. Usually, while recon-
structing a bifurcating branch, the wave number (or equivalently the number of cells) remains
constant, but as we observed in Figure 5.3 the state corresponding to the second bifurcation
undergoes a sudden change in its behavior at µ4 ' 52.25, where the buckling occurs from an al-
ready buckled state. This is an example of secondary bifurcation, and its reconstruction requires
a more involved continuation method based on arclength algorithms. Finally, we show in Fig-
ure 5.4 that the RB method works well at efficiently approximating the solution. In fact, we can
see that the RB error contour plot for the reconstruction of the first branch, and the corresponding
error on P, show the maximum peaks at the bifurcation points µ1 .

Figure 5.4. Left: RB error on P for the first branch. Right: RB contour error plot of the
displacement u at µ = 65.
 112 Chapter 5. Reduced Basis Approaches to Bifurcating Nonlinear Parametrized PDEs

5.5 Gross–Pitaevskii Equations in Bose–Einstein

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Condensates
In this section, we consider the Gross–Pitaevskii equation in quantum mechanics [456]. Of-
ten referred to as a nonlinear Schrödinger equation, the Gross–Pitaevskii equation models cer-
tain classes of Bose–Einstein condensates (BECs), a special state of matter formed by identical
bosons at ultralow temperatures. Despite the complexity of the model, here we stick to the simple
Gross–Pitaevskii equation and present a multiparameter study varying the chemical potential and
the normalized trap strength. This section is mainly based on work done with A. Quaini [450].

5.5.1 Gross–Pitaevskii Model

A BEC is a special state of matter formed by an unlimited number of bosons that condense into
the same energy state at low temperatures.
A quantum system is the environment to be studied in terms of wave-particle duality and it
involves the wave-function and its constituents, such as momentum and wavelength. The Gross–
Pitaevskii equation describes the ground state of a quantum system of identical bosons.
The Gross–Pitaevskii equation reads as follows: find the single-particle wave-function
Φ(r, t) : Ω × R+ → C such that
1
i∂t Φ = − ∆Φ + |Φ|2 Φ + W (r)Φ in Ω, (5.34)
2
where Ω ⊆ R2 is the domain, i is the imaginary unit, r = |r| = x2 + y 2 is the radial
p

coordinate, and W (r) = 21 τ 2 r2 is the external parabolic potential, with τ being the normalized
trap strength, i.e., the ratio of trappings along and transverse to the plane.
The construction of the steady solution is based on the assumption that the wave-function
Φ(r, t) = φ(r) exp(−iµt), with φ(r) : Ω → C, where µ is the chemical potential, which can be
seen as a measure of the density at the center of the trap and has to satisfy µ ≥ τ . This way, we
obtain the nonlinear problem given by the parametrized PDE
. 1
G(φ; µ) = − ∆φ + |φ|2 φ + W (r)φ − µφ = 0. (5.35)
2
The solutions of the one-dimensional problem (5.35) exhibit a not particularly rich bifurcating
behavior, while its two-dimensional version is far more interesting, including symmetry-breaking
bifurcations and vortex-bearing states [397, 398, 115, 212].
We will analyze it using its bifurcation diagram, which plots some characteristic quantities
of the solution, such as the number of bosons NB in the BEC or a different norm of the particle
density ρ, defined as Z
NB = ρ dr, ρ = |Φ|2 , (5.36)
Ω
with respect to the chemical potential µ. The critical values of the associated eigenvalue problem
.
are given by µ∗m,n = (m + n + 1)τ .
Thus, given an initial energy µ at the linear limit, we increase the chemical potential (and
therefore the number of atoms NB ) in order to approach the strongly nonlinear regime that can
lead to the discovery of new states. For the numerical characterization of the stability for each
state, we refer the reader to [115].

5.5.2 Weak Formulation and Its Approximation

Given the Gross–Pitaevskii model (5.35), here we consider its variational formulation for the
numerical investigation of the branching behavior. Let us consider µ = (µ, τ ) ∈ P, respectively,
 5.5. Gross–Pitaevskii Equations in Bose–Einstein Condensates 113

the chemical (bifurcating) parameter and the trap strength, and Ω ⊂ R2 the domain in which we
will approximate the solution.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

We recall that the solution φ to equation (5.35) is a complex function, so we will denote
its real and imaginary parts respectively by ϕ and ψ, which we will treat as the two com-
ponents of the solution. Having represented the model with the parametrized mapping G :

2
X × P → X0 with X = H01 (Ω) , its weak formulation reads as follows: given µ ∈ P,
.
find X(µ) = (ϕ(µ), ψ(µ)) ∈ X such that (omitting the µ dependency)
.
g(X, Y ; µ) = a(X, Y ) + n(X, Y ) + b(X, Y ; τ ) − d(X, Y ; µ) = 0 ∀ Y ∈ X, (5.37)

where we have introduced the variational forms

1 1
Z Z
a(X, Y ) = ∇X · ∇Y dΩ, b(X, Y ; τ ) = τ 2 |r|2 X · Y dΩ,
2 Ω 2 Ω
Z Z
d(X, Y ; µ) = µ X · Y dΩ, n(X, Y ) = |X|2 X · Y dΩ.
Ω Ω
The projection of the weak formulation (5.37) in the finite-dimensional space XN ⊂ X of di-
mension N is needed for the numerical investigation of the branching behavior. In this case,
the kth iteration of the Galerkin RB procedure combined with the Newton method reads as fol-
lows: given µ ∈ P and an initial guess XN 0
∈ XN for k = 0, 1, . . . until convergence, we seek
δXN = (δφN , δψN ) ∈ XN such that
k
a(δXN , YN ) + b(δXN , YN ; τ ) − d(δXN , YN ; µ) + c(δXN , XN , YN )
k k k k
= a(XN , YN ) + b(XN , YN ; τ ) − d(XN , YN ; µ) + n(XN , YN ) ∀ YN ∈ XN ,
k+1
and then we set XN = XN k
− δXN . Finally, the reduced Jacobian JN (XkN (µ); µ) ∈ RN ×N
can be written as
JN (XkN (µ); µ) = AN + BN (τ ) − DN (µ) + CN , (5.38)
where the reduced matrices are

AN = VT AN V, BN (τ ) = VT BN (τ )V,
N
(n)
X
DN (µ) = VT DN (µ)V, CN = XN VT CN (Σn )V,
n=1

in terms of the transformation matrix V and FE matrices

(AN )ij = a(E j , E i ), (BN (τ ))ij = b(E j , E i ; τ ),

(CN )ij = c(E j , XN

k
, E i ), (DN (µ))ij = c(E j , E i ; µ).
The time saving promised by the online-offline strategy, usually enabled by the affine decompo-
sition, cannot be obtained due to the nonlinearity, thus forcing us to adopt the empirical interpo-
lation techniques to obtain substantial savings of computational time during the online phase.

5.5.3 Numerical Investigation of Quantum States

We consider the domain Ω = (−12, 12)2 with homogeneous Dirichlet BCs on ∂Ω. The choice
of a square domain with homogeneous BCs guarantees that the solution is not influenced by the
boundary of the domain itself, since the support of the solutions is focused at the center. For the
space discretization, we use P2 FEs, which produces an FE space dimension of N = 27908.
 114 Chapter 5. Reduced Basis Approaches to Bifurcating Nonlinear Parametrized PDEs

In this case, an extensive study of the associated eigenproblem has not been taken into ac-
count given the infinite symmetry group that acts on the space of solutions. Instead, Hermite
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

guesses are used to initialize the continuation procedure, but more advanced strategies, such as
the deflation technique, can be applied in these scenarios [189, 454]. We chose the parame-
ter space as P = [0, 0.8] × [0.1, 0.3], varying the chemical potential and the concavity of the
parabolic trap. To recover the first bifurcation at µ∗ = τ , corresponding to the ground state,
we applied Algorithm 5.1 for three equispaced values of the trap strength and K = 160 eq-
uispaced values of the chemical potential. The resulting POD basis consists of only N = 12
modes, with a tolerance fixed to P OD = 10−8 . Figure 5.5 shows the evolution of this branch
in a two-parameter bifurcation diagram obtained with the RB version of Algorithm 5.1, using
∆τ = 0.01 and a continuation step of ∆µ = 1.25 · 10−3 . As expected, we see that, as τ in-
creases, the critical value of µ for the first bifurcation increases linearly. As for the efficiency,
due to the lack of a hyperreduction strategy, the speedup is only 1.52 with almost tRB = 2.73(h)
needed for the RB reconstruction of the full three-dimensional diagram in Figure 5.5, in contrast
to the tF E = 4.15(h) of the FE method. Here we were also able to build a unique ROM which
approximates with good accuracy all the first branches by varying τ ∈ [0.1, 0.3]. In fact, as we
can see from Figure 5.5, the reduced error for the density shows a maximum error of order 10−3 ,
with peak located at the current critical value of τ .

Figure 5.5. Left: RB bifurcation diagram for the multiparameter test case with infinity norm of
the density as output. The black dotted line shows the critical values of µ. Right: RB error on P for density
in the L2 - and H01 -norms for τ = 0.18.

Similarly, we recovered the reduced order bifurcation diagram for the first three emerging
branches (ground state for µ∗ = τ , single charge vortex and 1-dark soliton stripe for µ∗ = 2τ ),
shown in Figure 5.6. In this case we obtained a basis with dimension N = 41; in fact now it
has to encode different configurations. Some representative solutions for the three branches at
µ = 0.8 are depicted in Figure 5.6, where it is clear that the more one increases the trap strength,
the more the solutions remain confined in a smaller region around the center of the domain.
We remark that, while the left plot in Figure 5.5 has to be considered as an application of the
branchwise procedure, sampling offline one single qualitative configuration, the right plot in Fig-
ure 5.6 was obtained through a global approach, where all the bifurcating modes are compressed
by means of a single POD. We highlight that in these cases the speedups are still dependent on
N . For this reason we now present an application of the hyperreduction techniques to recover the
efficiency of the whole methodology. We focus on the reconstruction of a single branch, say the
single charge vortex in Figure 5.5 emerging from µ∗ = 2τ . Within the same setting as before, we
consider the EIM and DEIM techniques with greedy tolerance Gr = 10−7 together with a basis
 5.6. Hyperelastic Models for Bending Beams 115
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 5.6. Left: RB density functions at µ = 0.8 with τ ∈ {0.15, 0.25} for the first three
branches. Right: RB bifurcation diagram for the multiparameter test case with infinity norm of the density
as output. The blue dashed lines show the critical values of µ for both branching phenomena.

for the reduced manifold of dimension N = 6. We show in Figure 5.7 the reduced order error
on the number of bosons NB . These increased errors are a small price to pay for a considerable
computational speedup. In fact, while it takes tHF = 246(s) to reconstruct the whole branch
with FE, the RB with hyperreduction technique requires only tEIM = 55(s) and tDEIM = 7(s),
respectively, for EIM and DEIM. Finally, we remark that since we are dealing with bifurcating
phenomena, the reduced algorithm can converge to a different configuration of the system. This
is actually the case reported in Figure 5.7, where we show the high-fidelity real part φ of X and
its RB approximation, which is rotated by a certain angle but is still a solution for the system.

Figure 5.7. Left: error between the branches of the bifurcation diagram computed with FE and
RB with EIM/DEIM in the µ-NB plane. Right: high-fidelity and RB real part φ of the solution X for the
1-dark soliton stripe branch, left and right, respectively, at µ = 0.5.

5.6 Hyperelastic Models for Bending Beams

The aim of this section is to provide a brief description of the bifurcating (buckling) phenomena
for two- and three-dimensional beams with different constitutive relations; in particular we will
focus on the class of so-called hyperelastic models. By studying multiparametric settings, in
which the bifurcation point is investigated for parametrized geometries, we analyzed a real test
case scenario coming from the Norwegian petroleum industry [425]. This is based on work with
J. Eftang and A. T. Patera [449] in the framework of the MIT-FVG “ROM2S” project.
 116 Chapter 5. Reduced Basis Approaches to Bifurcating Nonlinear Parametrized PDEs

5.6.1 A Continuum Mechanics Framework

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Let Ω ⊂ Rd , with d = 2, 3, be a bounded domain which represents the reference configuration

of a body that undergoes deformation. We can capture the displacement u of a material point P ,
from the reference position X to the new deformed location x, using the deformation function
φ : Rd → Rd defined as x = φ(X). In fact, we can write u(X) = x − X and express the
deformation gradient and its determinant respectively as

∂φ(X)
F = = ∇u + I and J = det(F ).
∂X

Thus, the equilibrium equation derived from the equation of motion [577, 127] is

−div(P ) = B in Ω, (5.39)

where B is an external force acting on Ω and P = P (F ) is the first Piola–Kirchhoff stress

tensor, related to the Cauchy stress tensor σ by the formula P = JσF −T . Of course, we have to
characterize the relation between the stresses (forces) and the strains (displacements) by means of
constitutive laws, which highly depend on the type of elastic material considered. We focused on
hyperelastic materials [255], which are characterized by the existence of a strain energy function
ψ = ψ(F ) such that
∂ψ(F )
P (F ) = . (5.40)
∂F
We remark that these materials are characterized by the fact that the work done by stresses does
not depend on the path of deformation; thus they conserve the total energy, and in contrast
with linear elasticity, no assumptions on infinitesimal strains are required. We will analyze two
different choices for the strain energy function ψ, investigating numerically their properties with
respect to the buckling problems. The simplest constitutive relation is the Saint Venant–Kirchhoff
(SVK) model, which is an extension of the linear elastic material model to the geometrically
nonlinear regime, defining the strain energy function and the Green–Lagrange strain tensor as

1 T
ψ(F ) = λ1 E : E + λ2 (tr(E))2 /2, and E = (F F − I), (5.41)
2

where λ1 and λ2 are the Lamé constants related to the material properties through the Young
modulus E and the Poisson ratio ν as follows:

E Eν
λ1 = , λ2 = .
2(1 + ν) (1 + ν)(1 − 2ν)

The second hyperelastic constitutive relation we will consider is the neo-Hookean (NH) model,
which can be expressed using the strain energy function

λ1 λ2
ψ(F ) = (I1 − 3) − λ1 ln J + (ln J)2 , (5.42)
2 2

where I1 = tr(C) is the first principal invariant and λ1 , λ2 are the Lamé constants as before.
We remark that the words buckling and bifurcation will be used here interchangeably, but we
will be interested only in the approximation of the first buckling mode, properly following the
postbuckling behavior as the target branch.
 5.6. Hyperelastic Models for Bending Beams 117

5.6.2 Weak Formulation and Its Approximation

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Starting from the equilibrium equation (5.39), we can consider the boundary value problem, in
the reference domain Ω, given by

in Ω,

 −div(P (u)) = B
u = uD in ΓD , (5.43)
P (u)n = T in ΓN ,


where u : Ω → Rd is the in-plane displacement, B is the body force per unit reference area, and
T is the traction force per unit reference length. Thus, our aim is to study the deformation of the
domain Ω, subject to a prescribed displacement uD on the Dirichlet boundary ΓD together with
body and traction forces. We remark that, due to the Dirichlet nonhomogeneous BCs, the func-
.
tion space in which we set the problem is the space HD 1
(Ω) = v ∈ H 1 (Ω) | v = uD ∈ ΓD .
Taking the dot product with a test function v ∈ X = H0 (Ω) and integrating over the reference
1

domain Ω, the weak formulation reads as follows: find u ∈ HD 1

(Ω) such that
Z Z Z
P (u) : ∇v dΩ − B · v dΩ − T · v dΓ = 0 ∀ v ∈ X. (5.44)
Ω Ω ΓN

Furthermore, the boundary value problem in (5.43) for hyperelastic media can also be expressed
as a minimization problem by means of the theorem of virtual work [255]. In fact, we can define
the potential energy of the beam in terms of the strain energy function ψ as
Z Z Z
.
Π(u) = ψ(u) dΩ − B · u dΩ − T · u dΓ. (5.45)
Ω Ω ΓN

Thus, at the minimum point of Π(u), the directional derivative of Π with respect to the change
in u is equal to zero for all v ∈ X, i.e.,

. dΠ(u + δv)
g(u, v) = Dv Π(u) = |δ=0 , (5.46)
dδ
.
and the projection onto the RB space XN reads as follows: given µ ∈ P, we seek uN = uN (µ) ∈
XN that satisfies
JN (ukN (µ); µ)δuN = GN (ukN (µ); µ), (5.47)
updating the solution as uk+1
N = ukN − δuN until convergence.

5.6.3 Numerical Investigation of Compressed Beams

As a first application, we consider a two-dimensional beam with reference domain Ω = [0, 1] ×
[0, 0.1]. We built the FE space with P1 linear elements, resulting in a high-fidelity dimension
N = 4328. It is well known that, e.g., for B = (0, 0) and T = 0, the bifurcation diagram enjoys,
as for the von Kármán plate, a Z2 reflective symmetry and it undergoes a series of pitchfork
bifurcations as µ is increased. We consider an SVK beam, with Young modulus E = 106 and
Poisson ratio ν = 0.3, subjected to a parametrized uniform compression imposed by means of
Dirichlet BCs, as follows:

on ΓlD = {0} × [0, 0.1],


u = (0, 0)
(5.48)
u = (−µ, 0) on ΓrD = {1} × [0, 0.1],

which correspond to a clamped condition on the left end of the beam and an increasing uni-
form uniaxial compression on the other end. For the high-fidelity setting we choose Ntrain =
 118 Chapter 5. Reduced Basis Approaches to Bifurcating Nonlinear Parametrized PDEs

1000 points in the parameter space P = [0, 0.2], and an online continuation method based on
K = 2000 equispaced points in P, in order to properly detect the critical point and follow the
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

postbuckling behavior.
We obtained an RB space of dimension N = 5, from which we recover the bifurcation
diagram in Figure 5.8, with s(u) = kuy k∞ as functional output. We can clearly observe that
the beam buckles at the value µ∗ ≈ 0.03. In fact, at this point the vertical component of the
displacement u changes suddenly from being trivial to detecting the buckling. Despite the fact
that we are focusing only on the first branch, which is the most common cause of failure, different
buckling configurations still exist, such as the ones reported in Figure 5.8.

Figure 5.8. Left: RB bifurcation diagrams for the SVK unforced beam. Right: high-fidelity zeroth
and second mode displacement u for the SVK beam with B = (0, 0) at µ = 0.2.

Moreover, we also investigated the multiparameter case, defining µ = (µ, E, ν), where µ
is the bifurcating compression parameter and E and ν are respectively the Young modulus and
the Poisson ratio. For the analysis, we chose a parameter space given by P = [0.0, 0.2] ×
[105 , 107 ] × [0.25, 0.42] and a perpendicular body force B = (0, −1000). Since the parameter
space is three-dimensional, we expected that a much greater number of basis functions would be
needed; indeed, using a POD tolerance P OD = 10−8 we obtained an RB space of dimension
N = 43. The online continuation method was based on K = 201 equispaced µ-values in P,
which corresponds to a continuation step ∆µ = 10−3 , and the bifurcation diagram in Figure 5.9
was depicted for five random pairs of the elasticity constants (E, ν). We remark that we chose
the nontrivial body force was to avoid sharp gradients in the bifurcation diagrams and thus larger
errors in the reduced approximation. From the computational speedup point of view, the plot of
the reduced diagram costs exactly the same as its high-fidelity version: tHF ≈ tRB = 900(s);
despite this we were able to detect the buckling and the related postbuckling behavior for a wide
range of three-dimensional parameter space. We also considered parametrized geometries, where
we tried to understand how this influences the buckling property of the beam. Within the same
physical setting, we considered the SVK model for a beam with parametrized semilength l ∈
[0.5, 1]. We used the online continuation method to reconstruct the three-dimensional bifurcation
diagram in Figure 5.9 with N = 12 bases, using K = 201 equispaced points in P = [0, 0.2].
As we can see from Figure 5.9, the reduced manifold was also able to approximate the buckling
for nonsampled geometries with good approximation accuracy, and as expected we observed that
the longer the beam, the sooner it buckles.
Finally, we introduce a real test case scenario coming from the Norwegian petroleum indus-
try [425], investigating the deformation of a three-dimensional tubular geometry with annular
section. From the analysis of the parametrized geometries, we understood that, as before, the
 5.7. Navier–Stokes Flow and the Coandă Effect 119
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 5.9. Left: RB bifurcation plot for SVK beam with B = (0, −1000) for five random pairs
(E, ν) ∈ [105 , 107 ] × [0.25, 0.42]. Right: RB bifurcation plot for SVK beam with B = (0, −1000) and
l ∈ [0.5, 1].

length of the tube plays a fundamental role in the bifurcation, also varying the type of buckling
involved. An example is given by the tubular domain Ω isolated by the annulus section AR r with
r = 0.28 and R = 0.30. We chose SVK material and trivial body and traction forces, and we
fixed the material properties E = 2.1 · 105 and ν = 0.3. In Figure 5.10 we observe the huge dis-
crepancy of the modes’ behavior corresponding to l = 2 and l = 20. Even within this complex
multiparameter setting, we were able to recover the bifurcation diagram using the RB approach.

Figure 5.10. Left: RB bifurcation diagram for tubular unforced geometries with l ∈ [10, 20].
Right: representative solutions of the three-dimensional SVK model with l ∈ {2, 20}.

5.7 Navier–Stokes Flow and the Coandă Effect

In this section, we are interested in a fluid dynamics application, analyzing a bifurcating phe-
nomenon deriving from Navier–Stokes equations (NSEs) in a sudden-expansion channel flow
problem. We consider a simplified model of a cardiac disease, called mitral valve regurgitation,
linked to the Coandă effect [586, 447], which expresses the tendency of a fluid jet to be attracted
to a nearby surface.
In fact, a fluid characterized by a high viscosity has a symmetric jet configuration, and a pair
of vortices form downstream of the expansion. Lowering the viscosity below a critical threshold,
these two recirculation regions break the symmetry, giving rise to an asymmetric jet, which
represents the wall-hugging behavior. This part is based on the works done with M. Strazzullo
and M. Pintore in [452, 454].
 120 Chapter 5. Reduced Basis Approaches to Bifurcating Nonlinear Parametrized PDEs

5.7.1 NS Model in a Channel

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

The problem described above translates to a parametrized PDE of the form (5.1) that loses the
uniqueness of the solution while decreasing the viscosity µ below a certain critical value µ∗ .
We expect two qualitatively different configurations: (i) the symmetric solution—a physically
unstable configuration with a symmetric jet flow, and (ii) the asymmetric solution—a physically
stable configuration with a wall-hugging jet. These coexisting solution branches form the pitch-
fork bifurcation we want to approximate.
We consider a two-dimensional planar straight channel with a narrow inlet and a sudden
expansion, depicted in Figure 5.11, which represent a simplification of the left atrium and the
mitral valve, respectively. We define the inflow and outflow boundaries respectively as Γin =
{0} × [2.5, 5] and Γout = {50} × [0, 7.5], and the boundaries representing the walls as Γ0 =
∂Ω \ (Γin ∪ Γout ). The steady and incompressible NSEs for a viscous flow in Ω, with stress-
free condition on Γout , no-slip homogeneous Dirichlet on Γ0 , and nonhomogeneous Dirichlet
.
vin = [20(5 − x2 )(x2 − 2.5), 0] on Γin , is
in Ω,

 −µ∆v + v · ∇v + ∇p = 0
in Ω,

 ∇·v =0


v = vin on Γin , (5.49)
v = 0 on Γ ,

0


on Γout ,

−pn + (µ∇v)n = 0


where v is the velocity of the fluid, p is its pressure, and µ represents the kinematic viscosity.
For later convenience, we introduce the dimensionless Reynolds number as Re = U w/µ, which
represents the ratio between inertial and viscous forces, where U and w are the characteristic
velocity (maximum inlet velocity U = 31.25) and the characteristic length of the domain (length
of the inlet section w = 2.5), respectively.

40

2.5 Γin Γout 7.5

10
Γ0

Figure 5.11. Domain Ω, which represents a straight channel with a narrow inlet.

5.7.2 Weak Formulation and Its Approximation

2
Let V = H 1 (Ω) , Vin = {v ∈ V | v = vin on Γin , v = 0 on Γwall }, and V0 = {v ∈ V | v =
0 on Γin ∪ Γwall } be the function spaces for velocity. Furthermore, let Q = L2 (Ω) be the function
space for pressure. The weak formulation of (5.49) reads as follows: given µ ∈ P, find v ∈ Vin
and p ∈ Q such that

a(v, ψ; µ) + s(v, v, ψ) + b(ψ, p) = 0 ∀ ψ ∈ V0 ,
(5.50)
b(v, π) = 0 ∀ π ∈ Q,
having introduced the bilinear and trilinear forms
R
a(v, ψ; µ) =Rµ Ω ∇v · ∇ψ dΩ ∀ v, ψ ∈ V,
b(v, p) = − RΩ (∇ · v) p dΩ ∀ v ∈ V, ∀ p ∈ Q, (5.51)
s(v, v̄, ψ) = Ω (v · ∇v̄) ψ dΩ ∀ v, v̄, ψ ∈ V.
 5.7. Navier–Stokes Flow and the Coandă Effect 121

As usual, we can project the model by means of the FEM into the finite-dimensional space
XN = VNv × QNp . Thus, the combination of the Galerkin FE and the generic kth step of
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Newton’s method reads as follows: given µ ∈ P and an initial guess XN 0

= (vN0
, p0N ) ∈ XN ,
for k = 0, 1, . . . until convergence, we seek δXN = (δvN , δpN ) ∈ XN such that

k k

 a(δvN , ψN ; µ) + s(vN , δvN , ψ) + s(δvN , vN , ψN ) + b(ψN , δpN )
= a(vN , ψN ; µ) + s(vN , vN , ψ) + b(ψN , pkN ),
k k k
(5.52)
k
b(δvN , πN ) = b(vN , πN ) ∀ ψN ∈ V0,N , ∀ πN ∈ QN ,


k+1
and then we set XN = XN k
− δXN .
Equivalently, we can project (5.50) onto the RB space XN = VNv × QNp of dimension
N = Nv +Np . This way, the kth reduced step of Newton’s method reads as follows: given µ ∈ P
and an initial guess X0N = (vN 0
, p0N ) ∈ XN for k = 0, 1, . . . we seek δXN = (δvN , δpN ) ∈ XN
that satisfies (5.27), where the reduced Jacobian and residual are given respectively by
k k
) BTN
 
AN (µ) + S1,N (vN ) + S2,N (vN
JN (XkN (µ); µ) = and
BN 0
 k k k T
 (5.53)
AN (µ)vN + S1,N (vN )vN + BN pN
GN (XkN (µ); µ) = .
BN vN

Once again, we have introduced in the expression above the transformation matrices with respect
to the different components of the solution, Vv and Vp , respectively, for velocity and pressure,
defining the reduced matrices

Nv
(n)
X
AN (µ) = VvT AN (µ)Vv , BN = VpT BN Vv , Si,N = vN VvT Si,N (Σnv )Vv (5.54)
n=1

in terms of the high-fidelity matrices

j i j i
(AN (µ))ij = a(EN , EN ; µ) , (BN )ij = b(EN , EN ),
k j i j k i (5.55)
(S1,N )ij = s(vN , EN , EN ) , (S2,N )ij = s(EN , vN , EN ).

5.7.3 Numerical Investigation of Wall-Hugging Behavior

We consider a mesh on the domain Ω with 2785 cells, corresponding to N = 24301 degrees
of freedom associated with a Taylor–Hood P2 -P1 discretization of V × Q. For the high-fidelity
setting we chose Ntrain = 51 viscosity points in the parameter space P = [0.5, 2], corresponding
to Re ∈ [39.0, 156.0]. Within this setting we considered the branchwise reconstruction by using
only the snapshots coming from a single branch to build a ROM with Nv = Np = 20 for each
dynamic, the symmetric and the asymmetric one. The online continuation is based on K = 151
equispaced points in P, which corresponds to a continuation step ∆µ = 10−2 .
To plot the bifurcation diagram, we chose as output value a symmetry indicator of the ap-
proximated solution given by the point evaluation of the vertical component of the velocity on
the middle axis of the channel, namely vx2 at (x1 , x2 ) = (14, 4). In Figure 5.12 we plot the
bifurcation diagram with all the solution branches of the system (5.50) in the viscosity range
chosen. The numerical approximation clearly shows that a supercritical pitchfork bifurcation
occurs around the critical viscosity value µ∗ ≈ 0.96. It is evident that we have a unique solu-
tion for all µ > µ∗ ; thus while the fluid behaves like a Stokes one, we find three qualitatively
different solutions by increasing the Reynolds number. The bifurcation point µ∗ is also the one
 122 Chapter 5. Reduced Basis Approaches to Bifurcating Nonlinear Parametrized PDEs
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 5.12. Left: RB bifurcation diagram for the NS system. Right: velocity magnitude of the
asymmetric branch varying the viscosity µ.

responsible for the change in stability properties of the model. Indeed, the unique symmetric
solution remains stable until it encounters the critical value µ∗ , where it becomes unstable, while
at the same time the wall-hugging solution evolves as the physically stable one. Some repre-
sentative solutions for the lower and middle branches are presented in Figure 5.13, where we
show the symmetric and asymmetric velocity and pressure fields for the same viscosity param-
eter µ = 0.5. Here, the pressure for the lower branch decreases near the bottom left corner of
the expansion, causing the velocity to deflect, hugging the lower wall. Finally, thanks to the
no-slip boundary condition, the flux goes back to the midline, ending with a non-axis-symmetric
outflow. As we can see from Figure 5.14, we were able to recover online a good approximation
of the bifurcating asymmetric branch regarding the errors on the velocity and pressure fields. In
fact, we can observe once again that the RB error has its maximum at the critical value µ∗ , but it
has an average value on P of order 10−7 , confirming the accuracy of the methodology.
The stability is investigated by a standard eigenvalue analysis, analyzing the behavior of the
first Neig = 100 eigenvalues while varying the viscosity of the fluid. These eigenvalues are
plotted in Figure 5.15 for the stable lower branch and unstable middle branch. Note that since
the NS operator is not symmetric, we have both real and complex eigenvalues. Here we are just
interested in pitchfork bifurcation; thus in the zoom we follow the behavior of the biggest real
eigenvalue. Looking at the stability of the lower branch, all the eigenvalues of the NS system have
negative real part, and from this we can assert the stability of the wall-hugging branch. Indeed,
the zoom in the left plot of Figure 5.15 shows no crossing of negative real part eigenvalues. On
the contrary, the closeup in the right plot for the symmetric flow shows a sign change and thus
indicates instability.

Figure 5.13. Representative solutions for the NS system at µ = 0.5, velocity and pressure fields,
lower and middle branch, top and bottom respectively.
 5.7. Navier–Stokes Flow and the Coandă Effect 123
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 5.14. RB errors with respect to µ ∈ P for the velocity and pressure of the asymmetric branch.

Figure 5.15. Eigenvalues of the state eigenproblem in the complex plane for the NSE: stable and
unstable solutions, left and right panels, respectively.
 Chapter 6
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Reduced Basis
Stabilization for
Convection-Dominated
Problems

Enrique Delgado Ávila, Francesco Ballarin, Gianluigi Rozza

6.1 Introduction
In this chapter we present stabilized reduced basis (RB) models for convection-dominated prob-
lems. We start by defining a stabilized RB model for the advection-diffusion equation. In this
problem, the instabilities appear when the Péclet number is higher than one. Different stabiliza-
tion procedures are well known for the finite element (FE) problem [477], and we extend it for
the RB framework, following the work performed by Pacciarini and Rozza in [433].
Then we present a stabilized RB model for both the Stokes and Navier–Stokes problems.
The consideration of a nonstable pair of FE spaces for velocity and pressure leads to a nonstable
FE solution. The discrete spaces must verify a discrete inf-sup condition in order to guaran-
tee the well-posedness of the discrete solution for the Stokes and Navier–Stokes problems. For
the Stokes problem, we consider the Brezzi–Pitkäranta stabilization [75], while for the Navier–
Stokes problem we consider the streamline upwind Petrov–Galerkin (SUPG) stabilization pro-
cedure [78]. In the RB framework, different stabilization strategies are compared for both Stokes
and Navier–Stokes problems. We see that the offline-only stabilization is not consistent and the
RB solutions computed are far from the FE one. But with the offline-online stabilization for
the RB problem, the RB solution is computed accurately, either considering the inner pressure
supremizer [471] for the velocity space enrichment or not.
After that, we present a stabilized variational multiscale (VMS)-Smagorinsky model. In
this VMS-Smagorinsky model, the eddy viscosity term acts only on the resolved small scales.
We consider a local projection stabilization term for the pressure, derived from the high-order
term-by-term stabilization method (cf. [105, 108]). This methodology allows us to obtain more
accuracy than with the classic penalty stabilization procedure, with a reduced computational cost
[108]. The existence of a pressure stabilization term lets us consider a nonstable pair of FEs,
such as P2-P2, with the consequent increase of accuracy in both velocity and pressure.
The development of the a posteriori error bound estimator for the RB VMS-Smagorinsky
model with local projection stabilization on the pressure is done using the Brezzi–Rappaz–
Raviart (BRR) theory. In the online phase, we will also need the empirical interpolation method
(EIM) to approximate the nonlinear and nonaffine terms, with respect to the parameter, that
take part in the model considered in this chapter. Thus, we need to approximate using the EIM

125
 126 Chapter 6. Reduced Basis Stabilization for Convection-Dominated Problems

the eddy-viscosity term and the pressure stabilization coefficient. Finally, we present numerical
results for the two-dimensional cavity steady problems, for which we obtain speedup rates of
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

several thousand.
The chapter is structured as follows: in Section 6.2 we present the stabilized RB model for
the advection-diffusion problem. After that, we present the Stokes problem in Section 6.3, where
we define both the FE problem in Subsection 6.3.1 and the stabilized RB problem in Subsection
6.3.2. Then, in Section 6.4, we present the Navier–Stokes problem, defining both the stabilized
FE problem in Subsection 6.4.1 and the stabilized RB model in Subsection 6.4.2. Finally, we
present a stabilized VMS-Smagorinsky model in Section 6.5. We present the FE problem for this
model in Subsection 6.5.1 and the RB formulation in Subsection 6.5.2. Moreover, we present the
numerical analysis for the well-posedness of the VMS-Smagorinsky model in Subsection 6.5.3
and the development of the a posteriori error estimator in Subsection 6.5.4. Some numerical
experiments for the VMS-Smagorinsky model are presented in Subsection 6.5.5.

6.2 Advection-Diffusion Problem

In this section we present an RB model for the advection-diffusion equation. We study the
extension of the different stabilization procedures for the FE method (FEM) applied for the RB
method. For this purpose, let us introduce the advection-diffusion problem

−ε(µ)∆u + β(µ) · ∇u = f on Ω, (6.1)

where here we are considering that ε and β are functions Ω → [0, +∞) and Ω → R2 , respec-
tively, both depending on a given parameter value µ ∈ D. For problem (6.1) we consider suitable
Dirichlet, Neumann, or mixed boundary conditions.
To study problem (6.1), let us consider V ⊂ H 1 (Ω) to be a suitable Sobolev space. For
simplicity of the analysis, let us consider V = H01 (Ω), corresponding to homogeneous Dirichlet
boundary conditions. Thus, we define the variational formulation of problem (6.1) as follows:
(
Find u ∈ V such that
(6.2)
a(u, v) = F (v) ∀v ∈ V,

with
a(u, v) = (ε(µ)∇u, ∇v)Ω + (β(µ) · ∇u, v)Ω ,
F (v) = hf, vi ,
where we denote by h·, ·i the duality pairing between V and V 0 , where V 0 is the dual space of
V . Moreover, considering Vh ⊂ V as an FE internal approximation of V , we define the discrete
problem as follows:
(
Find uh (µ) ∈ Vh such that
(6.3)
a(uh , vh ; µ) = F (vh ; µ) ∀vh ∈ Vh .

It is well known that the advection-diffusion discrete problem has instabilities when the advection
term dominates the diffusive one. To analyze this issue, let Th be an admissible triangularization
of the domain Ω, and let K be an element of Th . We consider that problem (6.3) is advection
dominated in K if
|β(µ)|hK
Pe(µ) := > 1 ∀µ ∈ D, (6.4)
2ε(µ)
where here hK is the diameter of K. When we are dealing with advection-dominated problems,
we need to consider some additional terms in the discrete problem in order to avoid instabilities
 6.3. Steady Stokes Equations 127

in the numerical solutions. There are several stabilization methods for advection-dominated
problems in the literature, e.g., [78, 312, 477].
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

We consider a stabilized model, with a stabilization term in the discrete problem (6.3), de-
fined in a general way as
 
X hk
astab (uh , vh ; µ) = δK ε(µ)∆uh + β(µ) · ∇uh , (ρε(µ)∆uh + β(µ) · ∇uh ) ,
|β(µ)| K
K∈Th
(6.5)
where considering different values for ρ leads us to different stabilization methods. Here δk are
some suitable coefficients for the different stabilization procedures. The SUPG method corre-
sponds to ρ = 0, while ρ = 1 or ρ = −1 corresponds to the Galerkin least-squares (GLS) and
Douglas–Wang (DW) methods, respectively. For a detailed explanation of stabilization methods
we refer the reader to [477] and the references therein. Thus, the stabilized FE problem can be
defined as follows:
(
Find uh (µ) ∈ Vh such that
(6.6)
a(uh , vh ; µ) + astab (uh , vh ; µ) = F (vh ; µ) ∀vh ∈ Vh .

In the rest of this section, we consider the SUPG method for the stabilization procedure.
Let SN = {µ1 , . . . , µN } be a set of properly selected parameter values, e.g., with a greedy
algorithm [465, 471, 505], using a suitable a posteriori error estimator. We define the RB space
as VN = span{uh (µi ), i = 1, . . . , N }, where uh (µi ) are the high-fidelity solutions of problem
(6.6) for each µi ∈ SN .
We can consider two stabilization procedures for the RB model: the offline-only stabilization
and the offline-online stabilization. The main difference between methods is that in the offline-
only stabilization method we do not consider the stabilization term defined in (6.5) in the RB
model, while we do consider it in the offline-online stabilization model. Thus, the offline-only
stabilization method reads as follows:
(
Find uN (µ) ∈ VN such that
(6.7)
a(uN , vN ; µ) = F (vN ; µ) ∀vN ∈ VN .

The offline-online stabilization model reads as follows:

(
Find uN (µ) ∈ VN such that
(6.8)
a(uN , vN ; µ) + astab (uN , vN ; µ) = F (vN ; µ) ∀vh ∈ VN .

Recall that in both cases, the RB space VN is constructed with the solution of the stabilized
FE problem (6.6). In [433] there is a comparison of both procedures, where the authors conclude
that the consistency of the offline-online stabilization model is necessary for the good behavior
of the RB solution. They observe how with the offline-online stabilization model they are able
to obtain a reduced solution close to the FE solution, while using the offline-only stabilization
model, they are not able to approximate in a good way the RB solution with respect to the FE
solution.

6.3 Steady Stokes Equations

In this section we present a stabilized RB method for the steady Stokes equations. We study the
different possibilities of stabilization that can be performed on the reduced order model. The
 128 Chapter 6. Reduced Basis Stabilization for Convection-Dominated Problems

Stokes equations reads as follows:

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

in Ω,


 −ν∆u + ∇p = f

∇·u=0 in Ω, (6.9)
on ∂Ω,

 u = gD


where u is the velocity, p is the pressure, f is a given force, and ν is the viscosity. In the RB
framework, we can consider as a parameter, e.g., the viscosity of the fluid or a deformation of the
domain. In this last case, we should define a reference domain and perform all the computations
in this reference domain. For more details, see, e.g., [12, 30, 470]. We consider that ∂Ω is
decoupled into ∂Ω = ΓD0 ∪ ΓDg , where ΓD0 and ΓDg , respectively, are the homogeneous and
nonhomogeneous Dirichlet boundaries.
For the sake of simplicity, in the following we consider that gD = 0. In the case of nonhomo-
geneous boundary conditions, it is enough to define a lift function uD such that uD |ΓDg = gD .
To define the variational formulation of problem (6.9), let us introduce the spaces V =
(H01 (Ω))2 and M = L20 (Ω) on the reference domain. Moreover, let us consider Vh ⊂ V and
Mh ⊂ M to be two FE internal approximations. Thus, the discrete weak formulation is
a(uh , vh ; µ) + b(vh , ph ; µ) = hf, vh i ∀vh ∈ Vh ,
(
(6.10)
b(uh , qh ; µ) = 0 ∀qh ∈ Mh ,

where
a(u, v; µ) = ν(∇u, ∇v)Ω , b(v, q; µ) = −(∇v, q)Ω .
Problem (6.10) must verify the discrete inf-sup condition
b(vh , qh ; µ)
∃β0 > 0 : β0 ≤ βh (µ) = inf sup , (6.11)
qh ∈Mh vh ∈Vh kvh k1 kqh k0

where we are denoting by k · k0 and k · k1 the standard L2 -norm and H 1 seminorm, respectively.

6.3.1 Stabilized FE Problem

When we are dealing with discrete spaces Vh and Mh that do not verify the discrete inf-sup
condition (6.11), we need to consider a stabilization procedure in order to avoid instabilities in
the pressure solution. For the Stokes problem the stabilized problem can be stated as
a(uh , vh ; µ) + b(vh , ph ; µ) = hf, vh i ∀vh ∈ Vh ,
(
(6.12)
b(uh , qh ; µ) + spres (qh ; µ) = 0 ∀qh ∈ Mh .

There are different possibilities for the choice of the stabilization term in problem (6.12),
depending on the stabilization term spres . The simplest is the Brezzi–Pitkäranta stabilization,
introduced in 1984 in [75], in which the stabilization term is given by
X
spres (qh ; µ) = h2k (∇ph , ∇qh )K . (6.13)
K∈Th

A generalization of the Brezzi–Pitkäranta stabilization procedure is the Hughes, Franca, and

Balestra stabilization method, introduced in 1986 in [297], which considers the stabilization
term as X
spres (qh ; µ) = h2k (−ν∆uh + ∇ph − f, ∇qh )K . (6.14)
K∈Th
 6.4. Steady Navier–Stokes Equations 129

For a more detailed explanation of stabilization procedures for the Stokes problem, we refer the
reader to [477]. Due to its simplicity and reliability, in the following, we will consider the Brezzi–
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Pitkäranta stabilization as the stabilization procedure. The stabilized model (6.12) satisfies a new
inf-sup condition:
b(vh , qh ; µ)
0 < β0 ≤ βh = sup + h2K k∇qh k20 . (6.15)
vh ∈Vh k∇vh k0

6.3.2 Stabilized RB Problem

There are different possibilities for the RB formulation of the stabilized Stokes model. Fol-
lowing the strategy of the advection-diffusion problem of Section 6.2, we can consider either
online-offline stabilization or offline-only stabilization. Moreover, as usual in the Stokes RB for-
mulation, the inner pressure supremizer (cf. [471]) can be considered to enrich the velocity space
in order to fulfill the inf-sup stability constant in the RB problem.
Since an offline-only stabilization without inner pressure supremizer does not satisfy the RB
inf-sup stability condition, we will not consider this setting. In [13] the reader can find the
numerical study of the three cases of interest of stabilization. There, the authors observe that the
offline-only stabilization with inner pressure supremizer leads to a nonconsistent RB method that
gives a high error between the RB and FE solutions. The online-offline stabilization RB model,
which reads
a(uN , vN ; µ) + b(vN , pN ; µ) = hf, vN i ∀vN ∈ VN ,
(
(6.16)
b(uN , qN ; µ) + spres (qN ; µ) = 0 ∀qN ∈ MN ,

leads to an accurate approximation for both cases: with and without the inner pressure suprem-
izer, although the case of not considering the inner pressure supremizer gives a better approxima-
tion, with less computational effort. This better approximation might be due to the “pollution”
of the velocity space when we enrich it by adding the inner pressure supremizer.

6.4 Steady Navier–Stokes Equations

In this section we present a stabilized RB method for the steady Navier–Stokes equations. We
perform an analogous study as in Section 6.3. The Navier–Stokes equations read

1


 − ∆u + u · ∇u + ∇p = f in Ω,
 Re

∇·u=0 in Ω, (6.17)


 u=g on ∂Ω,

D

where, again, u is the velocity, p is the pressure, f is a given force, and, here, Re is the Reynolds
number. In the RB framework, we can consider as a parameter, e.g., the Reynolds number or a
deformation of the domain.
For the sake of simplicity, again in this section, we consider that gD = 0. In the case
of nonhomogeneous boundary conditions, it is enough to define a lift function uD such that
uD |ΓDg = gD , leading to a similar analysis of the problem.
To define the weak formulation of problem (6.17), let us recall the spaces V = (H01 (Ω))2
and M = L20 (Ω) on the reference domain. Moreover, let us consider Vh ⊂ V and Mh ⊂ M to
 130 Chapter 6. Reduced Basis Stabilization for Convection-Dominated Problems

be two FE internal approximations. Thus, the discrete weak formulation is

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

a(uh , vh ; µ) + b(vh , ph ; µ) + c(uh , uh , vh ; µ) = hf, vh i ∀vh ∈ Vh ,

(
(6.18)
b(uh , qh ; µ) = 0 ∀qh ∈ Mh ,

where a(·, ·; µ) and b(·, ·; µ) are defined in the same way as in the Stokes problem and

c(w, u, v; µ) = (w · ∇u, v)Ω .

In order to obtain a suitable solution for problem (6.18), the discrete spaces Vh and Mh must
verify the discrete inf-sup condition (6.11).

6.4.1 Stabilized FE Problem

In this section we define the stabilization procedure for the Navier–Stokes equations. We perform
a penalty stabilization model as with the Stokes equation. For that, we define the stabilized
Navier–Stokes model as


 a(uh , vh ; µ) + b(vh , ph ; µ) + c(uh , uh , vh ; µ),

+sconv (vh ; µ) + sdiv (vh ; µ) = hf, vh i ∀vh ∈ Vh , (6.19)

 b(uh , qh ; µ) + spres (qh ; µ) = 0 ∀qh ∈ Mh .


Here, the stabilization terms for the convective term sconv , pressure spres , and grad-div stabi-
lization sdiv are given, respectively, by
 
X 1
sconv (vh ; µ) = δ h2K − ∆uh + uh · ∇uh + ∇ph , uh · ∇vh ,
Re K
K∈Th
 
X 1
spres (qh ; µ) = δ h2K − ∆uh + uh · ∇uh + ∇ph , ∇qh ,
Re K
K∈Th

and X
sdiv (vh ; µ) = γ (∇ · uh , ∇ · vh )K ,
K∈Th

where δ and γ are suitable stabilization coefficients. We refer the reader to [78, 429, 578] and
the references therein.
This stabilized Navier–Stokes model corresponds to the SUPG model [78]. Other possibil-
ities for stabilization are the GLS [298] and the stabilization model introduced by Franca and
Frey in [205].
With this SUPG stabilization of the Navier–Stokes problem, we ensure that the spaces Vh
and Mh satisfy the discrete inf-sup condition

b(vh , qh ; µ)
0 < β0 ≤ βh = sup + h2K k∇qh k20 . (6.20)
vh ∈Vh k∇vh k0

6.4.2 Stabilized RB Problem

In the same way as for the Stokes RB problem, in the Navier–Stokes model, we can consider
different options for the RB stabilization procedure. We can consider an offline-only stabilization
procedure, with the velocity space enriched with the inner pressure supremizer (cf. [471]), or
 6.5. Stabilized VMS-Smagorinsky Turbulence Model 131

an offline-online stabilization with and without the inner pressure supremizer velocity space
enrichment.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

The RB offline-online stabilization for the Navier-Stokes equations is


 a(uN , vN ; µ) + b(vN , pN ; µ) + c(uN , uN , vN ; µ)


+sconv (vN ; µ) + sdiv (vN ; µ) = hf, vN i ∀vN ∈ VN , (6.21)

 b(uN , qN ; µ) + spres (qN ; µ) = 0 ∀qN ∈ MN .


Here, the snapshots for the construction of spaces VN and MN are selected by a weak greedy
algorithm (cf. [179]) by constructing an a posteriori error estimator based on the BRR theory.
See [162, 164, 387] for more details about the construction of the a posteriori error estimator for
the Navier–Stokes equations.
In [13] the reader can find numerical studies of this problem where the different stabilization
procedures for the RB problem are compared. There, the authors conclude that, again, the option
of an offline-only stabilization procedure is not valid since the RB solutions have a big error
with respect to the FE solutions. They conclude that considering the offline-online stabilization
preserves the consistency of the problem and leads to more accurate RB solutions. With respect
to the enrichment of the velocity space with the inner pressure supremizer, in both cases they
obtain accurate RB solutions, with the advantage in the case of not considering it of requiring
less computational effort in the computation of the RB solution.

6.5 Stabilized VMS-Smagorinsky Turbulence Model

In this section, we describe a local projection-stabilized VMS-Smagorinsky model. In this
model, we consider that the eddy viscosity only acts on small resolved scales. This leads to
a less diffusive model than the Smagorinsky one presented in Chapter 3, where the eddy vis-
cosity acts on both large and small resolved scales. We also consider a projection stabilization
coefficient for the pressure that allows us to use nonstable pairs of FEs.

6.5.1 FE Problem
Let Ω be a bounded domain of Rd (d = 2, 3), with Lipschitz-continuous boundary Γ, which we
suppose is split into Γ = ΓD ∪ ΓN . Here, ΓD is decoupled into ΓD = ΓD0 ∪ ΓDg , where ΓD0
and ΓDg , respectively, are the homogeneous and nonhomogeneous Dirichlet boundaries, while
ΓN is the Neumann boundary.
To set the VMS-Smagorinsky model, as in Chapter 3, we decompose the velocity and pres-
sure spaces, Y and M , respectively, as
Y = Yh ⊕ Y 0 , M = Mh ⊕ M 0 ,
where Yh and Mh are respectively the large-scale finite-dimensional spaces for velocity and
pressure, while Y 0 and M 0 are the small-scale complementary spaces. We suppose that the large
and small scales are separated by assuming that the sum is direct, i.e., Yh ∩ Y 0 = {0} and
Mh ∩ M 0 = {0}.
We define Yh = Y h ⊕Yh0 and Mh = M h ⊕Mh0 . Thus, uh = uh +u0h ∈ Yh and ph = ph +p0h .
Let us define the discrete space Vhl (Ω) = {r ∈ C 0 (Ω) : r|K ∈ Pl (K) ∀K ∈ Th }. Thus, we
define the velocity and pressure FE spaces as Yh = (Vhl (Ω) ∩ H01 )d and Mh = Vhm ∩ L2 . For the
subgrid eddy viscosity modeling in the VMS-Smagorinsky model, let us consider Y h = Πh Yh ,
with Πh a uniformly stable interpolation operator on Y h , where
Y h = [Vhl−1 (Ω)]d . (6.22)
 132 Chapter 6. Reduced Basis Stabilization for Convection-Dominated Problems

With this notation, we are identifying Y h as the large-scale velocity space and Yh0 = (Id−Πh )Yh
as the resolved small-scale velocity space. We denote by µ ∈ D ⊂ R the parameter considered
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

for the RB problem, which is again the Reynolds number, leading to the following variational
formulation problem:
Find (uh , ph ) = (uh (µ), ph (µ)) ∈ Yh × Mh such that





 a(uh , vh ; µ) + b(vh , ph ; µ) + a0S (wh ; wh , vh ; µ)


+ c(uh , uh , vh ; µ) + c(uD , uh , vh ; µ) (6.23)

+ c(uh , uD , vh ; µ) = F (vh ; µ) ∀vh ∈ Yh ,





b(uh , qh ; µ) = 0 ∀qh ∈ Mh ,


where here again, we are defining wh = uh + uD , with uD a lift function. The bilinear forms in
(6.23), a(·, ·; µ) and b(·, ·; µ), are defined as
a(uh , vh ; µ) = (∇uh , ∇vh )Ω , b(vh , ph ; µ) = −(ph , ∇ · vh )Ω ,
and the trilinear form c(·, ·, ·; µ), is defined as
c(wh , uh , vh ; µ) = (wh · ∇uh , vh )Ω .
The nonlinear form a0S (·; ·, ·; µ) is a multiscale Smagorinksy modeling for the eddy viscosity
term, and it is given by
Z
0
aS (zh ; uh , vh ; µ) = νT (Π∗h zh )∇(Π∗h uh ) : ∇(Π∗h vh ) dΩ, (6.24)
Ω

where Π∗h = Id − Πh . In VMS terminology, this corresponds to the small-small setting eddy
viscosity term (cf. [299]). Other possibilities are the large-small setting that models the turbulent
viscosity by a function of the whole resolved velocity, taking
Z
0
aS (zh ; uh , uh ) = νT (zh )∇(Π∗h uh ) : ∇(Π∗h vh ) dΩ. (6.25)
Ω

Depending on the stabilization procedure, we can distinguish two different kind of methods:
residual-based and penalty. The residual-based stabilization methods include GLS (cf. [298]) and
modifications of this method such as the SUPG method (cf. [78, 477]) or the adjoint-stabilized
method (cf. [204]). On the other hand, for penalty methods we particularly mention the penalty
term-by-term stabilized method (cf. [105]), which is an extension of the penalty method intro-
duced in [75]. In the latter method, the penalty term acts on the discretization of the pressure
gradient.
The stabilization procedure that we consider in this work is a local projection-stabilization
(LPS), introduced in [132], on the pressure gradient. The stabilization procedure considered
here is based on the high-order term-by-term stabilized method (cf. [108]) that stabilizes each
operator, e.g., the convection term or pressure gradient. We consider the stabilization on the latter
next.
Thus, we consider the LPS-VMS-Smagorinsky model, with stabilization on the pressure gra-
dient, as follows:
Find (uh , ph ) = (uh (µ), ph (µ)) ∈ Yh × Mh such that





 a(uh , vh ; µ) + b(vh , ph ; µ) + a0S (wh ; wh , vh ; µ)


+ c(uh , uh , vh ; µ) + c(uD , uh , vh ; µ) (6.26)

+ c(u , u , v ; µ) = F (v ; µ) ∀v ∈ Y ,

h D h h h h




b(uh , qh ; µ) + spres (ph , qh ; µ) = 0 ∀qh ∈ Mh ,

 6.5. Stabilized VMS-Smagorinsky Turbulence Model 133

where here spres (·, ·; µ) is the projection-stabilization term for the pressure, defined as
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

X
spres (ph , qh ; µ) = τp,K (µ)(σh∗ (∇ph ), σh∗ (∇qh ))K . (6.27)
K∈Th

Here, τp,K (µ) is the stabilization coefficient in (6.27), and it must verify the following hypothe-
sis: there exist two positive constants α1 , α2 , independent of h, such that

α1 h2K ≤ τK,p (µ) ≤ α2 h2K ∀µ ∈ D, ∀K ∈ Th , ∀h > 0. (6.28)

In the following, we will use the stabilization pressure coefficient τp,K (µ) proposed by Cod-
ina in [132] and used in [110, 512]:
−1
1/µ + ν T|K (µ)

UK (µ)
τp,K (µ) = c1 + c2 , (6.29)
h2K hK

where ν T|K is some local eddy viscosity, UK is a local velocity, and c1 , c2 are some positive
experimental constants. Defining τp,K (µ) in this way, we ensure (6.28).
By defining Xh = Yh × Mh , we rewrite problem (6.23) as follows:

Find Uh (µ) = (uh , ph ) ∈ Xh such that

(
(6.30)
A(Uh (µ), Vh ; µ) = F (Vh ; µ) ∀Vh ∈ Xh ,

where
1
A(Uh (µ), Vh ; µ) = A0 (Uh , Vh ) + A1 (Uh , Vh ) + A2 (Uh ; Vh )
µ (6.31)
+ A3 (Uh ; Vh ) + Apres (Uh , Vh ),
where, with U = (u, pu ), V = (v, pv ), and w = u + uD , the various terms are defined by
Z
A0 (U, V ) = ∇u : ∇v dΩ,
Ω
Z Z
A1 (U, V ) = (∇ · u)p dΩ − (∇ · v)pu dΩ
v
Ω
Z ΩZ

+ (uD · ∇u)v dΩ + (u · ∇uD )v dΩ,

Ω Ω
Z
A2 (U ; V ) = (u · ∇u)v dΩ,
Ω
Z
A3 (U ; V ) = νT (Π∗h w) ∇(Π∗h w) : ∇(Π∗h v) dΩ,
Ω
X Z
Apres (U, V ) = τp,K (µ) σh∗ (∇pu ) · σh∗ (∇pv ) dΩ.
K∈Th K

6.5.2 RB Problem
In this section we present the RB model derived from the FE problem (6.30). The reduced spaces
are computed using the greedy algorithm. The RB problem reads as follows:

Find UN (µ) ∈ XN such that

(
(6.32)
A(UN (µ), VN ; µ) = F (VN ; µ) ∀VN ∈ XN .
 134 Chapter 6. Reduced Basis Stabilization for Convection-Dominated Problems

Here, the reduced space is defined as XN = YN × MN , where the reduced velocity space YN
and the reduced pressure space MN are given by
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

YN = span{ζiv := u(µi ), i = 1, . . . , N }, (6.33)

MN = span{ξip := p(µi ), i = 1, . . . , N }. (6.34)

Both the small-small setting of the Smagorinsky eddy diffusion term defined in (6.31),
νT (Π∗h (∇w)) := νT (µ), and the pressure stabilization coefficient, τp,K (µ), defined in (6.27)
have a nonlinear representation with respect to the parameter and consequently need to be lin-
earized with the EIM.
We consider an approximation of the eddy viscosity term and the pressure stabilization term.
For this purpose, we need to build two RB spaces, WM S
1
= {q1S (µ), . . . , qM
S
1
(µ)} and WMP
2
=
{q1P (µ), . . . , qM
P
2
(µ)}, by a greedy selection procedure, with W S
M1 and W P
M2 the EIM reduced
spaces associated with the eddy viscosity term and the pressure stabilization term, respectively.
Thus, we approximate them by the following trilinear forms:

a0S (wN ; wN , vN ; µ) ≈ â0S (wN , vN ; µ),

(6.35)
spres (pN , qN ; µ) ≈ ŝpres (pN , qN ; µ),

where
M1
X
â0S (wN , vN ; µ) = σkS (µ)s(qkS , wN , vN ),
k=1
M2
(6.36)
X
ŝpres (pN , qN ; µ) = σkP (µ)r(qkP , pN , qN ),
k=1

with X
s(qkS , wN , vN ) qkS ∇(Π∗h wN ), ∇(Π∗h vN )


= K
,
K∈Th
X (6.37)
r(qkP , pN , qN ) qkP σh∗ (∇pN ), σh∗ (∇qN ) K .


=
K∈Th

Here we are considering that the approximations given by the EIM for νT (µ) and τK,p (µ) are,
respectively,
M1
X
IM1 [νT (µ)] = σkS (µ)qkS ,
k=1
M2
(6.38)
X
IM2 [τK,p (µ)] = σkP (µ)qkP .
k=1

This representation leads to a linearization of the RB problem (6.32), with an affine depen-
dence with respect to the parameter, given by the following:

Find (uN , pN ) = (uN (µ), pN (µ)) ∈ YN × MN such that






 a(uN , vN ; µ) + b(vN , pN ; µ) + â0S (wN ; vN ; µ)


+ c(uN , uN , vN ; µ) + c(uD , uN , vN ; µ) (6.39)

+ c(u , u , v ; µ) = F (v ; µ) ∀v ∈ Y ,

N D N N N N




b(uN , qN ; µ) + ŝpres (pN , qN ; µ) = 0 ∀qN ∈ MN .

 6.5. Stabilized VMS-Smagorinsky Turbulence Model 135

We can express the solution (uN (µ), pN (µ)) ∈ XN of (6.39) as a linear combination of the
basis functions:
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

N N
p
X X
uN (µ) = uN
j (µ)ζ j
v
, pN (µ) = pN
j (µ)ξj .
j=1 j=1

The tensor representations associated with the eddy viscosity term and the pressure stabiliza-
tion coefficient are defined as follows:
X
(S0N (qsS ))ij = (qsS ∇(Π∗h ζjv ), ∇(Π∗h ζiv ))K , i, j = 1 . . . , N, s = 1, . . . , M1 , (6.40)
K∈Th

(qsP σh∗ (∇ξjp ), σh∗ (∇ξip ))K ,

X
(PN (qsP ))ij = i, j = 1, . . . , N, s = 1, . . . , M2 . (6.41)
K∈Th

With this tensor representation for the nonlinear terms in (6.32) in the offline phase, it holds that

M1 M2
ŝpres (ξjp , ξip ; µ) =
X X
â0S (ζjv ; ζiv ; µ) = σsS (µ)S0N (qsS ) and σsP (µ)PN (qsP ).
s=1 s=1

Problem (6.39) is solved by a semi-implicit evolution approach. Note that thanks to that
representation, we are able to solve efficiently the online phase due to the linearization of the
eddy viscosity and pressure stabilization terms.
In this case, we no longer consider the supremizer for the velocity space enrichment as in
the previous sections, due to the offline-online stabilization proposed for the RB problem (6.32).
With this offline-online stabilization, the inner pressure supremizer operator is no longer nec-
essary to recover the pressure. Moreover, considering the supremizer operator for the velocity
space enrichment is counterproductive because the RB solution is no longer consistent with the
FE solution, i.e.,
UN (µk ) = Uh (µk ) ∀µk ∈ S = {µ1 , . . . , µN }, (6.42)
where S = {µ1 , . . . , µN } is the set of parameter values chosen in the greedy algorithm. The fact
of losing the consistency causes the greedy algorithm to rechoose a parameter value previously
chosen, making the matrix of the RB problem singular.

6.5.3 Well-Posedness Analysis

In this section we study the well-posedness of problem (6.30). The numerical analysis of the
projection-based VMS-Smagorinsky model can be found in [110]. In this section, we analyze
the projection-stabilized VMS-Smagorinsky model using the BRR theory, as done in Chapter 3
for the Smagorinsky model.
Let us introduce the scalar product
X Z
(f, g)τp = τp,K f g dΩ ∀f, g ∈ L2 (Ω) (6.43)
K∈Th K

1/2
and its associated norm kf kτp = (f, f )τp . We also consider a modification in the weight of the
scalar product. Let
Z  
1 0∗
(uh , vh )T = + νT ∇uh : ∇vh dΩ ∀uh , vh ∈ Yh , (6.44)
Ω µ
 136 Chapter 6. Reduced Basis Stabilization for Convection-Dominated Problems

where νT0∗ = νT (Π∗h wh (µ)) and

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

X
µ = arg min (CS hK )2 min |∇(Π∗h wh )(µ)|(x)χK (x),
µ∈D x∈K
K∈Th

with wh (µ) the velocity solution of (6.26). Here again, we define the norm of space Xh as
q
kUh kX = kuh k2T + kph k20,2,Ω ∀Uh = (uh , ph ) ∈ Xh . (6.45)

The analysis of the convergence and stability of the projection-stabilized VMS-Smagorinsky

can be found in [110]. Particularly, for a uniformly regular triangulation, the following result
holds.

Proposition 6.1. The following discrete inf-sup condition is verified:

(qh , ∇ · vh )Ω
 
∗ 2
kqh k0,2,Ω ≤ α sup + kσh (∇qh )kτp ∀qh ∈ Mh , (6.46)
vh ∈Yh k∇vh k0,2,Ω

with α > 0, independent of h.

The demonstration of this inf-sup condition can be derived from [108]. This result can be
extended by considering that the triangulation {Th }h>0 is regular rather than uniformly regular.
For further details, see [108].
Let us consider the directional derivative of the operator A(·, ·; µ). If we derive each operator
term in (6.31), we get

∂1 A0 (U, V )(Z) = A0 (Z, V ),

∂1 A1 (U, V )(Z) = A1 (Z, V ),
Z Z
∂1 A2 (U ; V )(Z) = (u · ∇z)v dΩ + (z · ∇u)v dΩ,
Ω Ω
Z
∂1 A3 (U ; V )(Z) = νT (Π∗h w) ∇(Π∗h z) : ∇(Π∗h v) dΩ
Ω Z
X ∇(Π∗h w) : ∇(Π∗h z)
(CS hK )2 ∇(Π∗h w) : ∇(Π∗h v) dΩ,


+ ∗
K |∇(Πh w)|
K∈Th

∂1 Apres (U, V )(Z) = Apres (Z, V ).

According to the BRR theory (cf. [76, 93]), the well-posedness problem (6.30) is guaranteed
by the following continuity and inf-sup conditions:

∂1 A(Uh (µ), Vh ; µ)(Zh )

∞ > γ0 < γh (µ) ≡ sup sup , (6.47)
Zh ∈Xh Vh ∈Xh kZh kX kVh kX
∂1 A(Uh (µ), Vh ; µ)(Zh )
0 < β0 < βh (µ) ≡ inf sup . (6.48)
Zh ∈Xh Vh ∈Xh kZh kX kVh kX
The existence of γ0 ∈ R and β0 > 0 satisfying (6.47) and (6.48), respectively, is given by
the following results, whose proofs can be found in [151].

Proposition 6.2. There exists γ0 ∈ R such that ∀µ ∈ D,

|∂1 A(Uh (µ), Vh ; µ)(Zh )| ≤ γ0 kZh kX kVh kX ∀Zh , Vh ∈ Xh .

 6.5. Stabilized VMS-Smagorinsky Turbulence Model 137

Proposition 6.3. Let C ? = CT2 (Cµ + 1). Suppose that

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

1 1
k∇uD k0,2,Ω < and k∇uh k0,2,Ω ≤ − k∇uD k0,2,Ω .
µCµ2 C ? µCµ2 C ?

Then, there exists β˜h (µ) > 0 such that

∂1 A(Uh , Vh ; µ)(Vh ) ≥ β˜h (µ)kvh k2T ∀Vh ∈ Xh . (6.49)

The inf-sup condition (6.48) is verified because the operator b(vh , qh ; µ) satisfies the discrete
inf-sup condition (6.11). Moreover, the condition of Proposition 6.3 is verified when the Dirichlet
boundary data is sufficiently small.

6.5.4 A Posteriori Error Bound Estimator

In this section we develop the a posteriori error bound estimator used in the hierarchical con-
struction of the reduced space during the greedy algorithm. We use the BRR theory [76].

Lemma 6.4. There exists a positive constant ρT such that, ∀Uh1 , Uh2 , Zh , Vh ∈ Xh ,
∂1 A(Uh1 , Vh ; µ)(Zh ) − ∂1 A(Uh2 , Vh ; µ)(Zh ) ≤ ρT kUh1 − Uh2 kX kZh kX kVh kX . (6.50)

We define the following inf-sup and continuity constants, associated with the well-posedness
of the RB problem (6.32):
∂1 A(UN (µ), Vh ; µ)(Zh )
0 < βN (µ) ≡ inf sup , (6.51)
Zh ∈Xh Vh ∈Xh kZh kX kVh kX
∂1 A(UN (µ), Vh ; µ)(Zh )
∞ > γN (µ) ≡ sup sup . (6.52)
Zh ∈Xh Vh ∈Xh kZh kX kVh kX
We define the a posteriori error bound estimator as
βN (µ) h p i
∆N (µ) = 1 − 1 − τN (µ) , (6.53)
2ρT
where τN (µ) is given by
4N (µ)ρT
τN (µ) = 2 (µ) , (6.54)
βN
with N (µ) the dual norm of the residual. The a posteriori error bound estimator is stated by the
following result.

Theorem 6.5. Let µ ∈ D, and assume that βN (µ) > 0. If problem (6.30) admits a solution
Uh (µ) such that
βN (µ)
kUh (µ) − UN (µ)kX ≤ ,
ρT
then this solution is unique in the ball BX UN (µ), βNρT(µ) .

Moreover, assume that τN (µ) ≤ 1 for all µ ∈ D. Then there exists a unique solution Uh (µ)
of (6.30) such that the error with respect UN (µ), the solution of (6.32), is bounded by the a
posteriori error bound estimator, i.e.,
kUh (µ) − UN (µ)kX ≤ ∆N (µ), (6.55)
with effectivity  
2γN (µ)
∆N (µ) ≤ + τN (µ) kUh (µ) − UN (µ)kX . (6.56)
βN (µ)
 138 Chapter 6. Reduced Basis Stabilization for Convection-Dominated Problems

6.5.5 Numerical Results

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

In this section we show the numerical results of the pressure LPS-Smagorinsky RB model
for the lid-driven cavity two-dimensional problem. The numerical results were obtained using
FreeFrem++ (cf. [268]).
We consider the Reynolds number as a parameter, ranging in the interval D = [1000, 5100].
We consider a regular mesh with 5000 triangles and 2601 nodes. The FE pair chosen is the pair
P2-P2. We impose a nonhomogeneous boundary condition uD = 1 on the top of the geometry,
while on the rest of the boundaries, we impose homogeneous Dirichlet conditions.
For this test, considering the supremizer enrichment for the velocity space leads to a reselec-
tion of a parameter value already selected in the greedy algorithm. In Figure 6.1 (left) we show
that in this case, in the fourth iteration, the greedy algorithm breaks down.
In the EIM considered in this test, we compute M1 = 20 bases for the eddy-viscosity ap-
proximation and M2 = 25 bases for the pressure stabilization constant approximation to reach
a prescribed tolerance of εEIM = 5 · 10−4 . In Figure 6.1 (right) we show the evolution of the
error for both EIMs.
For the greedy algorithm we prescribe a tolerance of εRB = 10−4 , which is reached for
N = 16 bases. In Figure 6.2 (left) we show the evolution of the a posteriori error bound estimator
during the greedy algorithm. Then, in Figure 6.2 (right) we show the comparison of the error
and the a posteriori error bound estimator. In this case, the a posteriori error bound estimator is
about two orders of magnitude greater than the exact error.

105 100
max τN (µ) without supremizer kνT (µ) − I[νT (µ)]k∞
µ∈D
max τN (µ) with supremizer kτK,p (µ) − I[τK,p (µ)]k∞
µ∈D
104
10-1

103

10-2

2
10

-3
10
101

10
0
10-4
1 2 3 4 5 6 7 8 5 10 15 20 25
N M

Figure 6.1. Comparison of the stabilization with/without supremizer (left) and evolution of the
error in the EIM (right).

105 10-2

max τN (µ) ∆N (µ)

104 µ∈D kUh (µ) − UN (µ)kX
10-3
max ∆N (µ)
3 µ∈D
10
10-4
102

10-5
101

100 10-6

10-1
10-7

10-2
10-8
10-3

10-9
10-4

10-10
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 1000 1500 2000 2500 3000 3500 4000 4500 5000
N Re

Figure 6.2. Evolution of the error in the greedy algorithm (left) and a posteriori error estimator
for N = 16, with error (right).
 6.6. Conclusions 139
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 6.3. FE (left) and RB (right) solution for µ = 2751.

In Figure 6.3, we show two solutions for a prescribed value of the Reynolds number µ =
2751, for the FE problem (right) and the RB problem (left). We can observe the similarity of
both solutions, since the error between them is of order 10−6 .
Finally, in Table 6.1 we summarize results obtained for several values of the Reynolds num-
ber. For this test, we obtain speedup rates that can reach the order of tens of thousands, with the
error between the FE and RB solutions quite small.

Table 6.1. Computational time for FE solution and RB online phase, with the speedup and the error.

Data µ = 1610 µ = 2751 µ = 3886 µ = 4521 µ = 5100

TF E 4083.19 s 6918.53 s 9278.51 s 10201.7 s 11277.8 s
Tonline 0.71 s 0.69 s 0.69 s 0.7 s 0.69 s
speedup 5750 10026 13280 14459 16248
kuh − uN kT 2.4 · 10−5 4.129 · 10−6 3.14 · 10−5 3.23 · 10−5 2.77 · 10−8
kph − pN k0 2.17 · 10−7 1.99 · 10−8 5.38 · 10−8 6.36 · 10−8 6.89 · 10−8

6.6 Conclusions
In this chapter we have presented different RB stabilization models for convection-dominated
problems. We started with an advection-diffusion problem, then we defined the different sta-
bilization procedures for the Stokes and Navier–Stokes problems, and, finally, we presented a
stabilized VMS-Smagorinsky model, with the numerical analysis for the construction of the a
posteriori error estimator used in the weak greedy algorithm for the snapshot selection. We have
seen that the offline-online procedure is the best option in the models presented in this chapter
since it is the one that preserves the consistency in the RB problem. Moreover, since the inner
pressure supremizer for the enrichment of the velocity space does not influence the Stokes and
Navier–Stokes accuracy of the solution, not considering it decreases the computational effort of
computing the solution. Finally, in the VMS-Smagorinsky model, we have seen that considering
the inner pressure supremizer is no longer valid.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Part II

Order Models
and Discontinuous
Galerkin-Based Reduced
Finite Volume, Spectral Element,
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Chapter 7

Finite Volume–Based
Reduced Order Models
for Laminar Flows

Matteo Zancanaro, Saddam Hijazi, Umberto Morelli,

Giovanni Stabile, Gianluigi Rozza

7.1 Introduction
This chapter focuses on reduced order models (ROMs) for fluid problems discretized with the
finite volume method (FVM) [408, 596]. Currently, the use of the FVM for solving industrial
problems is predominant in many different fields, e.g., fluid dynamics and continuum mechanics
applications. In fact, most of the computational fluid dynamics (CFD) solvers used to tackle
real-life applications are based on the FVM, e.g., Fluent [391] and STAR CCM+ [3] (commercial
codes) and OpenFOAM [612] (open-source code). These CFD solvers are well equipped with
the computational tools needed to tackle problems coming from various engineering fields.
The demand for ROMs which could reduce computational costs has been increasing in recent
decades. These observations and demands motivate investing in efforts to construct ROMs for fi-
nite volume–based discretization. In this work, the full-order solver is chosen to be OpenFOAM,
a choice justified by the fact that it is open-source code. In addition, OpenFOAM offers well-
established documentation and tutorials for various benchmark problems. As a consequence, the
ROMs proposed in this chapter have taken into consideration the full-order modeling techniques
utilized by OpenFOAM.

7.1.1 Literature Review

In the literature, earlier works on ROMs based on the finite volume discretization include [175,
260, 259, 261]. In these works, the reduced basis (RB) approach has been extended to work
in general linear evolution schemes such as finite volume schemes. In [380], the authors pre-
sented a proper orthogonal decomposition (POD)-Galerkin ROM for the FVM. The ROM has
been constructed to reduce the Navier–Stokes equations (NSEs) for both laminar and turbulent
flows. Another POD-Galerkin ROM is presented in [548] and is dedicated to reducing the prob-
lem of vortex shedding around a circular cylinder in a finite volume environment. Stabilization
techniques have been extended to work in finite volume ROMs [550]. In particular, the last work
presents a POD-Galerkin ROM which employs the supremizer stabilization method in order to

143
 144 Chapter 7. Finite Volume–Based Reduced Order Models for Laminar Flows

obtain a stable-pressure ROM field. The supremizer stabilization method ensures that a reduced
order version of the inf-sup condition, needed by saddle-point formulations, such as the NSEs, is
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

met; initially it was constructed for finite element (FE) POD-Galerkin ROM in [34]. The POD-
Galerkin ROMs in [380, 548, 550] differ in the methodology adopted to reduce the NSEs. Further
details on the approaches of these works are recalled in the sections related to U-ROM and PPE-
ROM. We mention works in which numerical analysis is used to study ROMs for turbulent flows
and ROMs for pressure approximations [513, 333, 91]. The last works present different ROM
frameworks for computing the reduced pressure field.
Recently, the work [551] tackled the issue of geometrical parametrization for FVM-based
POD-Galerkin ROM. The authors in [551] propose a ROM which is fully consistent with the
SIMPLE algorithm [438]. The SIMPLE algorithm is the full-order solver approach for steady
flows in OpenFOAM and is a segregated approach. Other works which deal with geometrical
parametrization in FVM-based ROM include [369, 644], which focus on inviscid Euler equa-
tions. Also addressing the issue of geometrical parametrization, the contributions [609] and
[637] deal with turbulent compressible NSEs and partial differential equation (PDE)-constrained
optimization problems, respectively. Extension of the FVM-based POD-Galerkin ROM for ther-
mal mixing problems is presented in [219]. Another POD-Galerkin ROM for the problem of
buoyancy-driven enclosed flows is developed in [552].

7.2 Computational Fluid Dynamics—Laminar NSEs

This section of the chapter is devoted to laminar CFD problems. In particular we will see how
to carry out both the full-order and the reduced order discretizations. In the end, some basic
applications are presented.

7.2.1 Full-Order Model

This paragraph presents the full-order model (FOM), starting from the governing equations of
interest, which are the incompressible NSEs. The NSEs are studied here in a parametrized form;
we start by recalling their strong form. Given a parameter vector µ ∈ P ⊂ Rq , where P is a
q-dimensional parameter space, the incompressible NSEs parametrized by µ read as follows:
 h  i
T

 ∂t
 ∂u
+ ∇ · (u ⊗ u) = ∇ · −pI + ν ∇u + (∇u) in Ω × [0, T ],
∇ · u = 0 in Ω × [0, T ],

(7.1)

 + Boundary conditions on Γ × [0, T ],
+ Initial conditions in (Ω, 0),



where Γ is the boundary of the fluid domain Ω ∈ Rd , with d = 1, 2, or 3. u is the flow velocity
vector field; t is the time; ν is the fluid kinematic viscosity; p is the normalized pressure field,
which is divided by the fluid density ρf ; and the time window under consideration is [0, T ].
We remark that in this work the parameter µ could be a physical parameter such as the fluid
kinematic viscosity or a geometrical one such as the dimension of a certain part of the domain.
We would like to emphasize that the velocity and pressure fields are functions of time, space, and
the parameter, i.e., u = u(t, x; µ), p = p(t, x; µ). These dependencies have been dropped in
the equations above for the sake of keeping the notation concise.
The incompressible NSEs in (7.1) are solved by the FVM [408]. The first step in this method
is to choose a suitable polygonal tessellation, and then the PDE system is written in integral
form over each control volume. Denote by Nh the dimension of the FOM, which is basically
the number of degrees of freedom of the discretized problem. The momentum and continuity
equations are solved with the help of a segregated approach which adapts the Rhie and Chow
 7.2. Computational Fluid Dynamics—Laminar NSEs 145
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

uP
d
uN
uf Sf

Figure 7.1. Sketch of a finite volume in two dimensions.

interpolation [486]. The discretization process starts with the momentum equation by writing it
in integral form over each control volume Vi as follows:
∂
Z Z Z   Z
T
udV + ∇ · (u ⊗ u)dV − ∇ · ν ∇u + (∇u) dV + ∇pdV = 0. (7.2)
Vi ∂t Vi Vi Vi

The discretization procedure of all terms in the momentum equation is explained in what follows.
The pressure gradient term is discretized using Gauss’s theorem:
Z Z X
∇pdV = pdS ≈ Sf pf , (7.3)
Vi Si f

where pf is the value of the pressure at the center of the faces and Sf is the area vector of each
face of the control volume.
The convective term is discretized by exploiting Gauss’s theorem as follows:
Z Z X X
∇ · (u ⊗ u)dV = (dS · (u ⊗ u)) ≈ Sf · uf ⊗ uf = Ff u f . (7.4)
Vi Si f f

In the above equation uf is the velocity vector evaluated at the center of each face of the control
volume. Note that the velocity values are initially computed at the cell centers and therefore the
values at the center of the faces have to be deduced from the ones calculated at the cell centers.
Consequently, these uf values are interpolated using the values computed at the cell centers.
Plenty of interpolation schemes can be used, such as central, upwind, second-order upwind, and
blended differencing schemes. The mass flux Ff is computed using the previous converged
velocity in the first iteration and then updated by Ff = uf · Sf to remove the nonlinearity.
As for the diffusion term, it is discretized as
Z   Z   X
T T
∇ · ν ∇u + (∇u) dV = dS · ν ∇u + (∇u) ≈ νf Sf · (∇u)f , (7.5)
Vi Si f

where (∇u)f is the gradient of u at the faces. In similar fashion to the computation of the
pressure gradient in (7.3), one may compute (∇u)f . As for the computation of the term Sf ·
(∇u)f in (7.5), this depends on one particular feature of the mesh: the orthogonality. The mesh
in Figure 7.1 is orthogonal if the line that connects two cell centers is orthogonal to the face that
divides these two cells. If the mesh is orthogonal, and assuming that the two cell centers are
equally distanced from the face, then the term Sf · (∇u)f is computed as
uN − uP
Sf · (∇u)f = |Sf | , (7.6)
|d|
 146 Chapter 7. Finite Volume–Based Reduced Order Models for Laminar Flows

where uN and uP are the velocities at the centers of two neighboring cells and d is the distance
vector connecting the two cell centers; see Figure 7.1. On the other hand, a nonorthogonal
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

correction term is needed for the case of nonorthogonal meshes. The correction term [309] is
computed by the equation
uN − uP
Sf · (∇u)f = |∆| + J · (∇u)f , (7.7)
|d|

where the relation Sf = ∆ + J holds. The first vector, ∆, is chosen parallel to Sf . The
term (∇u)f is obtained by interpolating the values of the gradient at the cell centers (∇u)N
and (∇u)P , in which the subscripts N and P indicate the values at the centers of the two
neighboring cells. In the next section, we address the discretized equations (written in matrix
form) for velocity and pressure. Additionally, we present the segregated pressure-based solver
approaches.

7.2.2 Segregated Pressure-Based Solvers for the Incompressible

NSEs
After having introduced the discretization of the different terms in the NSEs, one may proceed
to address the algorithm used to solve the discretized system. The algorithm employed in Open-
FOAM is a segregated pressure-based approach. Its version for the case of steady flows is called
the Semi-Implicit Method for Pressure-Linked Equations (SIMPLE) [438]. On the other hand,
the Pressure Implicit with Splitting of Operators (PISO) [304] algorithm is utilized for unsteady
flows; one may also use the PIMPLE [408] algorithm, which is a combination of PISO and
SIMPLE.
The starting point is the discretized system of the NSEs, which can be written in matrix form
as follows:     
[Au ] [∇(·)] u 0
= , (7.8)
∇ · (·) [0] p 0
where Au is the matrix containing the velocity coefficients, derived from the momentum equa-
tion, Au u = ∂u
∂t + ∇ · (u ⊗ u) − ∇ · ν(∇u + (∇u) ). The last system is called a saddle-point
T

system.
The momentum equation in (7.8) can be written as

u + [A−1
u ][∇(·)][p] = 0, (7.9)

where A−1u denotes the inverse of the momentum matrix in the discretized form. By applying
the divergence operator to the last equation, one obtains

[∇ · (·)]u + [∇ · (·)][A−1
u ][∇(·)][p] = 0. (7.10)

Then, by exploiting the free divergence constraint on the velocity [∇ · (·)]u = 0, one may derive
the pressure equation
[∇ · (·)][A−1
u ][∇(·)][p] = 0. (7.11)
Thus the matrix A−1 u acts as the diffusivity in the Laplace equation for the pressure. For compu-
tational reasons, solving (7.11) in this form is inconvenient. This is because A−1u is likely to be
dense. In addition, the product of three matrices could result in a dense matrix, which increases
the difficulty of solving the linear system. As a result, a different approach is used in which the
momentum matrix is decomposed into diagonal and off-diagonal matrices:

[Au ] = [Du ] + [LUu ], (7.12)

 7.2. Computational Fluid Dynamics—Laminar NSEs 147

where [Du ] is a diagonal matrix and therefore can be easily inverted and [LUu ] is the matrix
containing the off-diagonal part of [Au ]. Inserting (7.12) into (7.8) yields the modified system
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

    
[Du ] [∇(·)] u −[LUu ][u]
= . (7.13)
∇ · (·) [0] p 0

This gives the following equation for pressure:

−1 −1
[∇ · (·)][Du ][∇(·)][p] = −[∇ · (·)][Du ][LUu ][u]. (7.14)

The last equation is a Poisson equation for pressure with the diagonal part of the momentum
equation acting as a diffusivity, while the divergence of the velocity is on the right-hand side.
The derivation of the pressure equation in (7.14) is referred to as a Schur complement.
In the next part of this subsection, the same derivation will be done without having to use the
assembly of Schur’s complement. The first step is to write the discretized momentum equation
for each control volume, namely
X
au
P uP + au
N uN = rp − ∇p, (7.15)
N

where P represents a generic cell center and N is the set of neighboring points around it (7.1),
uN and uP are the velocities at the centers of two neighboring cells, au P is the vector of diagonal
coefficients of the equations, au
N is the vector that consists of off-diagonal coefficients, and rp
is a term that contains any contributions of the right-hand side of the momentum equation of the
NSEs. For the sake of simplicity, one may introduce the operator H(u) = rp − N au N uN ,
P
which contains the off-diagonal part of the momentum matrix and any right-hand side contribu-
tions. Thus,
au
P uP = H(u) − ∇p (7.16)
and
−1
uP = (au
P) (H(u) − ∇p). (7.17)
Inserting the last expression for uP into the continuity equation ∇ · u = 0 gives
−1 −1
∇ · [(au
P) ∇p] = ∇ · [(au
P) H(u)]. (7.18)

The last pressure equation is equivalent to the one in (7.14), where it can be seen that au P is
a coefficient in the diagonal matrix [Du ] and that H(u) is the product [LUu ][u]. Note that
one of the strongest assumptions in the SIMPLE algorithm is H(u) = 0. This means that this
algorithm is not consistent. If a more consistent solution is needed, the SIMPLEC algorithm is
to be preferred.
The mass flux through each face of the control volume is denoted by Ff . Equations (7.17)
and (7.18) are used together with the discretized version of the continuity equation to update the
mass fluxes Ff :
−1 −1
Ff = uf · Sf = −(au
P) Sf · ∇p + (au
P) Sf · H(u). (7.19)

The term (au

P)
−1
Sf · ∇p is computed in a similar manner to (7.6):

|Sf |
(au
P)
−1
Sf · ∇p = (au
P)
−1
(pN − pP ) = apN (pN − pP ), (7.20)
|d|
|Sf |
where pP and pN are pressure values at the centers of two neighboring cells and apN = (au
P)
−1
|d|
is the off-diagonal matrix coefficient in the pressure equation.
 148 Chapter 7. Finite Volume–Based Reduced Order Models for Laminar Flows

At this point, the SIMPLE algorithm may be introduced. The first step in this algorithm is
called the momentum predictor step, where one solves the momentum equation (7.16) for an
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

initially guessed pressure field p? . Then, using the obtained value of the velocity u? from the
last step, one may assemble the new off-diagonal vector H(u? ). This is followed by solving the
Poisson equation for pressure (7.18) to correct the pressure field. The next step is to assemble the
conservative flux fields Ff before going through the cycle again and carrying out the momentum
predictor: fluxes are needed to update both au P and H(u). This procedure is repeated till con-
vergence is achieved. However, the pressure correction equation in its current form is prone to
divergence [596]. In order to ensure convergence, underrelaxation is employed.
The pressure field which results from solving the Poisson equation in (7.18) is p? + p0 , the
sum of the initial guess p? and a correction term denoted by p0 . The underrelaxed pressure field
is modified as follows:
p?? = p? + αp p0 , (7.21)
where αp is the pressure underrelaxation factor. If αp = 1, this means that no underrelaxation
is introduced and this choice of αp is too large for stable computations, particularly when p? is
far away from the final solution. A zero value of αp means that no correction is introduced at all
and the pressure is not updated, which is obviously unwanted. Therefore, the value of αp should
be taken in the range (0, 1) to allow one to add a correction term which is a fraction of p0 . This
term should be large enough to move the algorithm forward toward convergence and at the same
time small enough to ensure stable computations.
Unlike pressure, the underrelaxation for the velocity is not explicit; instead, the momentum
underrelaxation is implicit, where the discretized momentum equation in (7.16) is modified by
adding artificial terms as follows:
1 − αu u ?? ?? ? ? 1 − αu u ?
aP uP + au
P uP = H(u ) − ∇p + aP uP . (7.22)
αu αu
The new added terms consist of the new underrelaxed velocity values, denoted by u?? P , and the
old ones, denoted by u?P . These terms cancel each other in the case of convergence. Here αu is
the velocity underrelaxation factor, and the above equation simplifies to
1 u ?? 1 − αu u ?
a u = H(u? ) − ∇p? + aP uP . (7.23)
αu P P αu
As for the guidelines for choosing underrelaxation factors, there is no straightforward way to set
their optimum values, where in fact their values depend on the type of problem and the flow.
However, in the literature it is recommended that
0 < αu < 1,
0 < αp < 1,
αu + αp ≈ 1.
Let us summarize the important aspects of the SIMPLE algorithm. In the SIMPLE algorithm,
the role of pressure in the momentum equation is to ensure that the velocity field is divergence
free. After carrying out the momentum predictor step, the velocity field does not satisfy the
divergence constraint since the pressure field is a guessed one. Consequently, the pressure field
is corrected, which gives a pressure field which consists of two parts. The first one is physical
and thus consistent with the global flow field, while the other one represents the correction term,
which guarantees the continuity. Only the first component of the pressure field is desired and used
to build the physical pressure field. To build this physical pressure field, the SIMPLE algorithm
resorts to the underrelaxing approach explained in this subsection. The full SIMPLE algorithm
is outlined in Algorithm 7.1.
 7.2. Computational Fluid Dynamics—Laminar NSEs 149

The SIMPLE algorithm is used in OpenFOAM to solve the velocity-pressure coupled system
for the case of steady flows. This algorithm may be extended to transient simulations; however,
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

the most used transient solver in CFD codes is PISO. In the remaining part of this section, the
PISO algorithm is addressed. The PISO algorithm is based on the idea that for low values of
the Courant number (small time steps), the pressure-velocity coupling is much stronger than
the nonlinear u-u coupling in the convective term. Thus it suggests that a repeated number of
pressure corrections can be carried out without having to rediscretize the momentum equation.
The PISO algorithm consists of two correction steps for the pressure field. The first yields a
conservative velocity field, while the second gives back a more physical pressure field. The use
of more than one pressure corrector implies that it is not necessary any more to underrelax the
pressure field. However, one thing that could be seen as a drawback of the PISO algorithm is
that it is assumed that the momentum discretization is not needed when the pressure correctors
are applied. This last assumption is only true for small time steps, which forces the maximum
Courant number [138, 139] to be under the value of Comax = 0.9.

A LGORITHM 7.1. The SIMPLE algorithm.

1: Start with an initial guess of the pressure field p? and the velocity field u? .
2: Momentum predictor step: assemble au ? ?
P , H(u ), and ∇p , and solve the discretized momentum equation for the guessed
pressure field p? and velocity field u? :

u ? ? ?
aP uP = H(u ) − ∇p .

3: Calculate the off-diagonal vector H(u?? ) after obtaining the velocity field u?? from u?P .
4: Correct the pressure: the new pressure field is computed based on the obtained velocity field from the last step:

u −1 u −1 ??
∇ · [(aP ) ∇p̃] = ∇ · [(aP ) H(u )].

Then correct the velocity explicitly by

u −1 ??
ũP = (aP ) (H(u ) − ∇p̃).

5: Relax the pressure field and the velocity equation with the prescribed underrelaxation factors αp and αu , respectively. The under-
relaxed fields are called pur and uur .
6: Assemble the conservative face flux Ff :

u −1 ur p ur ur
Ff = uf · Sf = (aP ) Sf · H(u ) − aN (pN − pP ).

7: Set u? = uur and p? = pur .

8: Repeat until convergence is achieved.

The PISO algorithm starts in a similar way to SIMPLE, using a guessed pressure field to carry
out the momentum predictor step. Then it assembles the H(u? ) vector and corrects the pressure
and velocities. The next step is the key difference between the two algorithms, where at this
point the velocity field has been corrected after solving the pressure equation. This means that
the H vector has been changed and so has the source term in the pressure equation, which makes
the pressure field no longer correct, so one has to update the H vector. To address this issue,
the SIMPLE algorithm goes all the way back to the momentum equation and performs another
momentum predictor, and thus it obtains a new velocity value, which will be used to update the
H vector to resolve the pressure equation. On the other hand, in the PISO loop the momentum
predictor is done just once. In more detail, rather than solving the momentum equation for the
second time, it uses the corrected velocity field to directly update the H vector and then performs
the second pressure correction step. The PISO complete procedure is structured in Algorithm 7.2.
Other transient solvers could be used to solve the coupled velocity-pressure system. We
mention here the PIMPLE algorithm, which merges PISO and SIMPLE. The PIMPLE algorithm
 150 Chapter 7. Finite Volume–Based Reduced Order Models for Laminar Flows

contains an outer loop which iterates over the PISO procedure, resulting in a more stable algo-
rithm. The number of times the outer loop is entered is known as the number of outer correctors
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

of the PIMPLE; if it is set to one then the PIMPLE replicates the PISO.

A LGORITHM 7.2. The PISO algorithm.

1: Perform steps 1–4 in the SIMPLE algorithm:

• Solve the discretized momentum equation.

• Update the off-diagonal vector H.

• Solve the pressure correction equation.

• Correct the pressure and velocities.

2: Solve the second pressure correction, and obtain the second pressure correction component p00 .
3: Correct the pressure and velocities; this step gives p??? and u??? :
??? ? 0 00
p =p +p +p .

4: Assemble the conservative face flux Ff :

u −1 ??? p ??? ???
Ff = uf · Sf = (aP ) Sf · H(u ) − aN (pN − pP ).

5: Set u? = u??? and p? = p??? .

6: Go to step 1 if convergence is not reached.
7: Proceed to the next time step if the final time is not reached.

7.2.3 Nonsegregated Finite Volume–Based ROMs

In this section we will address finite volume–based ROMs which utilize a coupled approach to
resolve the reduced order system. The ROMs discussed in this section differ from each other in
the number of equations used at the reduced level, the type of equations, and most important the
reduction assumption for the pressure field.

Momentum Equation Solely ROM (U-ROM)

After having introduced the POD method in the first chapter of this book, one may perform
the projection procedure in order to construct the reduced order system. In this contribution,
Galerkin projection is employed, which results in a POD-Galerkin ROM. At this point, different
procedures could be followed to construct POD-Galerkin ROMs designed to reduce the NSEs.
In this section, the simplest option is addressed. This approach uses just the momentum equation
of the incompressible NSEs (7.1) at the reduced order level [380].
The assumption on which projection-based ROMs sit is that the dynamics of the parametric
PDE system can be described by a few dominant modes. These modes, when suitably combined,
accurately reproduce the full-order solution. The last assumption translates to the following
decomposition approximation of the velocity and pressure fields:
PNr
u(x, t; µ) ≈ ur (x, t; µ) = i=1 ai (t; µ)φi (x),
PNr (7.24)
p(x, t; µ) ≈ pr (x, t; µ) = i=1 ai (t; µ)χi (x),

where ur and pr are the reduced velocity and pressure fields, respectively, while φi (x) and
χi (x) are the reduced modes for velocity and pressure, respectively. These modes do not depend
on µ and t. The reduced order degrees of freedom are denoted by ai (t; µ) for both velocity and
pressure; the number of degrees of freedom is Nr . The temporal coefficients ai (t; µ) depend on
time t and on the parameter vector µ. Note that in (7.24) the velocity and pressure fields have
 7.2. Computational Fluid Dynamics—Laminar NSEs 151

the same temporal coefficients. This last assumption makes the projection-based ROM simpler,
as mentioned earlier. However, this assumption has several limitations and drawbacks as may be
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

anticipated; in the next subsections these drawbacks will be addressed in greater detail.
The velocity POD modes are computed using the standard procedure. On the other hand,
the pressure modes in this formulation are computed using the eigenvectors and the eigenvalues
obtained in the SVD procedure for the velocity POD modes. Thus, the pressure modes are
expressed as
Ns
1 X
χi = S j V u. (7.25)
Ns λui j=1 p ij
The next step in building the POD-Galerkin ROM is to project the momentum equation of
(7.1) onto the velocity POD basis [φi ]N
i=1 as follows:
r

 
∂u 
T

φi , + ∇ · (u ⊗ u) − ∇ · ν ∇u + (∇u) + ∇p = 0. (7.26)
∂t L2 (Ω)

The following reduced order dynamical system is obtained after inserting the decomposition
assumptions of (7.24):
ȧ = ν(B + BT )a − aT Ca − Ha, (7.27)
where a is the vector of reduced order degrees of freedom, and each of B, BT , C, and H is
either a reduced order matrix or a tensor. These terms are computed as follows:
(B)ij = (φi , ∇ · ∇φj )L2 (Ω) ,
(BT )ij = φi , ∇ · (∇φTj ) L2 (Ω) , (7.28)

(C)ijk = (φi , ∇ · (φj ⊗ φk ))L2 (Ω) , (7.29)

(H)ij = (φi , ∇χj )L2 (Ω) . (7.30)
It is important to mention that the convective nonlinear term in the NSEs is approximated at the
reduced order level by a third order tensor C. This last approach to handling the nonlinear term
in the momentum equation could potentially increase the computational cost when the number
of reduced velocity modes grows. Other techniques for the treatment of this term include EIM-
DEIM methods [627, 40] and the gappy-POD method [100]. At this stage, (7.27) can be solved
for the reduced degrees of freedom of the fluid dynamics fields. This system can be integrated
for time values which are outside the time window in which snapshots were acquired, namely
[0, T ]. That case is referred to as extrapolation in time.
In the context of ROM, the notion of having two fully decoupled stages called the offline
and the online stages is crucial. The offline phase represents the training stage, which starts by
sampling the parameter space. Then the FOM simulations are run and snapshots are acquired.
The last step in the offline stage involves computing the POD modes, as well as any other term
which depends on the POD modes, such as the reduced order matrices and tensors. The offline
stage is characterized by being of significant computational cost since the computations depend
on the dimension of the FOM. On the other hand, the online phase involves fast computations
which should not depend on the dimension of the FOM. In the POD-Galerkin ROM developed
in this subsection, the online stage consists of solving (7.27) for parameter values which could
be new ones with respect to those used to train the model.

Momentum and Poisson Pressure Equations ROM

This section presents a POD-Galerkin ROM which is based on the use of both the momentum
equation and the Poisson pressure equation at the reduced order level. This ROM approach was
first presented in [548]. This POD-Galerkin ROM can be used only to reduce unsteady flows.
 152 Chapter 7. Finite Volume–Based Reduced Order Models for Laminar Flows

The POD-Galerkin ROM presented in the last subsection has shown a lack of accuracy in
reconstructing the pressure field at the reduced order level. An accurate pressure approximation
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

is imperative since many outputs of interest depend greatly on the pressure field. Consequently,
stabilization methods have been considered in order to accurately reproduce the pressure field.
In this work, two stabilization techniques are addressed. The first exploits the Poisson equation
for pressure (PPE) at the reduced order level, which makes it possible to separate the pressure
reduced degrees of freedom from those of the velocity. The PPE can be obtained by taking the
divergence of the momentum equation in (7.1) and then taking advantage of the fact that the
divergence of the velocity field is null. The resulting system is
∂u h 
T
i
+ ∇ · (u ⊗ u) = ∇ · −pI + ν ∇u + (∇u) in Ω × [0, T ], (7.31a)
∂t
∆p = −∇ · (∇ · (u ⊗ u)) in Ω. (7.31b)
We remark that the ROM developed in this section cannot be used to reduce steady flows [313,
376]. In fact, among the possible choices for the pressure boundary condition required to render
the PPE formulation of the NSEs equivalent to the original one, we selected one which ensures
that the velocity is divergence free only in the unsteady setting (see the remark in Section 2 in
[313]).
The reduced order decomposition assumptions in this case are
PNu
u(x, t; µ) ≈ ur (x, t; µ) = i=1 ai (t; µ)φi (x),
PNp (7.32)
p(x, t; µ) ≈ pr (x, t; µ) = i=1 bi (t; µ)χi (x),

where one can see that new temporal coefficients denoted by bi (t; µ) have been introduced to ap-
proximate the reduced pressure field. It is worth mentioning that the way the pressure modes are
computed now is different from in the previous section. Here the SVD of the pressure correlation
matrix C p ∈ RNs ×Ns has to be done. The entries of the matrix C p are
 
(C p )ij = Spi , Spj 2 , (7.33)
L (Ω)

and then the pressure modes are

Ns
1 X
χi = S j V p, (7.34)
Ns λpi j=1 p ij

where λp is the matrix containing in its diagonal the eigenvalues of the pressure correlation
matrix C p and V p is the corresponding matrix of eigenvectors. At this point one may perform
Galerkin projection of the equations. The momentum equation is projected onto the velocity
POD modes, while the PPE is projected onto the pressure POD space. These projections yield
 i
∂u h 
T
φi , + ∇ · (u ⊗ u) − ∇ · −pI + ν ∇u + (∇u) = 0, (7.35a)
∂t L2 (Ω)

(χi , ∇p + ∇ · (u ⊗ u))L2 (Ω) − ν (n × ∇χi , ∇ × u)Γ − (χi , n · Rt )Γ = 0, (7.35b)

where in the last equation, R is the initial velocity field. Substituting the reduced order approxi-
mations into the projected equations gives the reduced order dynamical system

ȧ = ν(B + BT )a − aT Ca − Hb,

(7.36)
Db + aT Ga − νN a − L = 0,
 7.2. Computational Fluid Dynamics—Laminar NSEs 153

where the new additional matrices and tensors are defined as

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

(D)ij = (∇χi , ∇χj )L2 (Ω) , (G)ijk = (∇χi , ∇ · (φj ⊗ φk ))L2 (Ω) , (7.37a)

(N )ij = (n × ∇χi , ∇ × φj )Γ , (L)i = (χi , n · Rt )Γ . (7.37b)

The reduced order system (7.36) is a differential-algebraic system of equations (DAE). This DAE
can be solved to obtain the reduced order vector degrees of freedom for velocity and pressure,
a and b, respectively. As mentioned earlier, the ROM approach presented in this section is
restricted to working with unsteady flows. Therefore, a different ROM procedure is needed to
treat steady flows.

7.2.4 Segregated Finite Volume–Based ROMs (SIMPLE-ROM)

This section focuses on a recently developed strategy for incompressible Navier–Stokes problems
that is strictly connected with the full-order SIMPLE algorithm. The main reason to develop a
new technique in order to solve the ROM NSEs is related to the fact that, recalling (7.36), one
can notice the difference between the full equations solved in the full SIMPLE algorithm (Algo-
rithm 7.2) and the projected ones. This is because the SIMPLE algorithm is not fully coherent
with the NSEs since there are some strong assumptions, simplifications, and linearizations.
The goal of this procedure is to end up with a ROM problem that is as coherent as possible
with respect to the full-order one. Since it is trivial to understand, this method is an alternative
to PPE-ROM since both momentum and continuity equations are used at the same time.

A LGORITHM 7.3. The reduced order SIMPLE algorithm.

1: Start with an initial guess of the reduced pressure field b? and the reduced velocity field a? .
2: Reconstruct both pressure and velocity fields p? and u? .
3: Momentum predictor step: assemble the momentum equation using those fields, project it using the velocity modes φi , and solve it
to obtain a new reduced velocity a?? :
u ? ? ?
(φi , aP uP − H(u ) + ∇p )L2 (Ω) = 0.

4: Reconstruct the new full velocity u?? and calculate the off-diagonal vector H(u?? ).
5: Correct the pressure: the new reduced pressure b?? field is computed after having projected the pressure equation using the pressure
modes:
u −1 u −1 ??
(χi , ∇ · [(aP ) ∇p̃] − ∇ · [(aP ) H(u )])L2 (Ω) = 0.

??
Then correct the velocity explicitly after having reconstructed the new pressure p .
6: Relax the pressure field and the velocity equation with the prescribed underrelaxation factors αp and αu , respectively. The under-
ur ur
relaxed fields are called p and u .
7: Assemble the conservative face flux Ff :
u −1 ur p ur ur
Ff = uf · Sf = (aP ) Sf · H(u ) − aN (pN − pP ).

8: Set u? = uur and p? = pur .

9: Repeat until convergence.

Convergence can be checked in different ways. The simplest is to check over the residual;
however, its order of magnitude could be very far from the full-order one and, thus, a threshold
may be difficult to set. On the contrary, it is possible to set the jump of the residual by comparing
two consecutive steps of the algorithm, but the possibility of being locked near a local minima
could become high. The best solution is to use the two aforementioned choices together so that
one achieves a convergence criterion which is as safe as possible.
Of course the reduced algorithm inherits all the instability issues affecting the full-order one.
For this reason the initial guesses a? and b? have to be properly chosen and preferably not set to
zero.
 154 Chapter 7. Finite Volume–Based Reduced Order Models for Laminar Flows

7.2.5 Treatment of Nonhomogeneous Dirichlet Boundary Conditions

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

To simulate physical fluid dynamics problems, we must often impose nonhomogeneous Dirich-
let boundary conditions on certain parts of the boundary. This might be the case in inlet-outlet
problems, where it is natural to have nonhomogeneous Dirichlet velocity conditions at the in-
let. In the context of parametric PDEs, it may very well be the case that this nonhomogeneous
boundary velocity vector is the parameter under study. These aspects make the treatment of the
nonhomogeneous Dirichlet boundary conditions essential for building an accurate ROM.
The methods employed to enforce nonhomogeneous boundary Dirichlet condition include
the penalty method [63, 23, 42, 319, 539] and the lifting function method [240, 251, 285].
Before entering into the details of both methods, we introduce the following notation: let
ΓD ⊂ Γ be the Dirichlet boundary that might be composed by separate boundaries, i.e., ΓD =
ΓD 1 ∪ΓD 2 . . .∪ΓD K . Let NBC be the number of velocity boundary conditions we would like to
impose on some parts of the Dirichlet boundary. It is important to clarify that each nonzero scalar
component value of the velocity field that has to be set at one part of the boundary is counted as
one boundary condition. As an example, in a two-dimensional problem, let UΓD 1 = (Ux1 , 0) and
UΓD 2 = (Ux2 , 0) be the velocity vectors that must be imposed at the Dirichlet boundaries ΓD 1
and ΓD 2 , respectively. In this case there are two nonhomogeneous Dirichlet boundary conditions
to set and thus NBC = 2. Let UBC,i,j be the nonzero value of the ith component of the velocity
vector to be imposed at the reduced order level at the jth part of the Dirichlet boundary ΓD j .
Denote by UBC the vector of all scalar velocities UBC,i,j —this vector has a dimension of NBC ,
and its kth element is called UBC k .

The Lifting/Control Function Method

In the lifting/control function method, the nonhomogeneous boundary condition is enforced by
introducing a lifting function which carries the nonhomogeneity. This is followed by homoge-
nizing the velocity snapshots by subtracting from each snapshot a suitably scaled version of the
lifting function. This results in a set of velocity snapshots with homogeneous boundary con-
ditions. The POD method is applied on the new set of velocity snapshots, giving at the end
homogeneous POD modes.
The nonhomogeneous Dirichlet boundary conditions are removed as follows:

ũ = u − UBC · φL , (7.38)
h
where φL ∈ RNBC ×Nu is a matrix of the lifting functions φL i,j . Each lifting function φL i,j
has homogeneous Dirichlet boundary conditions in all parts of the Dirichlet boundary except in
the ith component at ΓD j , where it has unitary value. At this point, one can apply the POD
method on the snapshot matrix
h
Sũ = {ũ(x, t1 ; µ1 ), . . . , ũ(x, tNT ; µM )} ∈ RNu ×Ns . (7.39)

In the online stage, the given boundary velocity vector is called UBC
∗
, which may contain values
different from the ones in the original velocity snapshots. We want to approximate the reduced
order velocity field corresponding to UBC
∗
. The ROM velocity field is approximated as
Nu
X
∗ ∗
u(x, ·; UBC ) ≈ UBC · φL + ai (·)φi (x). (7.40)
i=1

It can be seen above that the boundary velocity vector UBC is assumed to be the parameter
under question. However, in the presence of nonhomogeneous boundary condition(s) for the
velocity field at a part of the boundary, the same described boundary treatment procedure has to
 7.2. Computational Fluid Dynamics—Laminar NSEs 155

be followed in all of the following cases:

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

• when building a ROM for the reproduction and extrapolation in time without parameters
(nonparametrized ROM);

• with the parametrized case where the boundary velocity vector is one of the parameters;

• with the parametrized case where the boundary velocity vector is not a parameter.

The way to choose a suitable lifting function is problem dependent. To reduce unsteady
nonparametrized cases, where reduction aims to reproduce time snapshots and potentially ex-
trapolate in time, a possible choice for the lifting function is the average of the offline velocity
snapshots. A more general approach for generating appropriate lifting functions is to solve linear
potential flow problems with a unitary boundary condition for each nonhomogeneous boundary
condition to be set. These problems are steady ones; therefore, an iterative procedure with a ten-
tative velocity field is carried out till convergence. While solving each of these potential flows,
the value of the initial velocity field at the Dirichlet boundary has to be zero everywhere except
for one scalar entity where the lifting function is sought. The converged velocity field will be
considered as the lifting function corresponding to the nonhomogeneous boundary condition at
the aforementioned entity. Besides the requirement of having unitary boundary condition, the
lifting functions have to be divergence-free fields.
The lifting/control function method can be adapted to work with both the PPE-ROM and the
SUP-ROM. On the other hand, adapting this method with the ROM mentioned in Subsection
7.2.3 (U-ROM) is not straightforward. This is because the reduced order degrees of freedom of
the velocity are the same as those of the pressure. Therefore, the homogenization of the veloc-
ity snapshots in (7.38) has to be accompanied with a procedure that obtains the corresponding
pressure snapshots. This procedure is not easily defined.
As a consequence of the additional lifting function mode/modes, the dimension of the ve-
locity POD space VuP OD will be Nu + NBC , or Nu + NS + NBC for the PPE-ROMs, when a
supremizer stabilization is absent or needed, respectively.

The Penalty Method

The nonhomogeneous boundary conditions in the penalty method are enforced by presenting a
constraint in the reduced order dynamical system. This is done by adding a term in the reduced
momentum equation which has zero value everywhere except on the Dirichlet boundary. This
method was initially implemented in the context of the FEM in [23, 42]. In [240], the authors
utilized the penalty method to enforce nonhomogeneous time-dependent Dirichlet boundary con-
ditions for the case of the flow around a circular cylinder at Reynolds number 100. The POD-
Galerkin ROM in [539] introduced the penalty method in order to account for the time-dependent
Dirichlet boundary conditions—the authors presented a study of the accuracy and the stability of
the method depending on the penalty parameter values. The penalty method has also been used
in other ROMs, e.g., [318, 319, 307, 380, 548].
The method can be employed for all the ROMs mentioned in the previous sections. For
example, when a supremizer stabilization is employed for the SUP-ROM (see Subsection 1.2.4
for more details on the supremizer stabilization strategy), the architecture is modified as follows:

N
!
X BC
T
M ȧ = ν(B + BT )a − a Ca − Hb + τ k k
(UBC k D − E a) , (7.41a)
k=1

P a = 0, (7.41b)
 156 Chapter 7. Finite Volume–Based Reduced Order Models for Laminar Flows

where τ is a penalization factor whose value is set by sensitivity analysis or by automatic tun-
ing, as recently presented in [554], where an iterative penalty method is presented. Generally
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

speaking, having a higher value of τ leads to a stronger enforcement of the boundary conditions
but might ill-condition the dynamical system. The newly introduced boundary matrices E k and
vectors D k are defined as

(E k )ij = (φi , φj )L2 (ΓD k ) ∀k = 1, . . . , NBC , (7.42a)

(D k )i = (φi )L2 (ΓD k ) ∀k = 1, . . . , NBC . (7.42b)

Unlike in the lifting function method, the POD is done here directly on the nonhomogeneous
velocity snapshots.

7.2.6 Offline-Online Computation of Lift and Drag Forces

In this subsection we are interested in recovering specific performance indicators which are im-
portant in engineering problems. Here we address the forces acting on the surface of bodies
immersed in the flow.
The forces F acting on the surface of a body denoted by ∂Ωf are given by the surface integral
Z
F = (2µ∇u − pI)nds. (7.43)
∂Ωf

As mentioned earlier, having full decoupling between the offline and the online stages is a vital
feature of efficient ROM. One may approximate the reduced order forces Fr by simply comput-
ing the following integral after reconstructing the reduced fields:
Z
Fr = (2µ∇ur − pr I)nds, (7.44)
∂Ωf

but this would require accessing the original mesh, which means that the computational cost of
carrying out this integral will depend on Nh (the number of degrees of freedom of the FOM).
Therefore, an alternative approach is needed to efficiently reconstruct the fluid dynamics forces.
The approach used throughout this work involves inserting the reduced order approximations
in (7.43). These approximations could be the uniform ones in (7.24) or the nonuniform ones in
(7.32). Assuming the latter approximations, the forces are computed as follows:
 ! Np 
Z Nu
X X
Fr = 2µ∇ ai (t; µ)φi (x) − bi (t; µ)χi I  nds, (7.45)
∂Ωf i=1 i=1

Z Nu Z Np
X X
Fr = 2µ ai (t; µ)∇φi (x)nds − bi (t; µ)χi nds, (7.46)
∂Ωf i=1 ∂Ωf i=1

Nu Z Np Z
X X
Fr = ai (t; µ) 2µ∇φi (x)nds − bi (t; µ) χi nds. (7.47)
i=1 ∂Ωf i=1 ∂Ωf

Let Z
δi = 2µ∇φi (x)nds for i = 1, . . . , Nu , (7.48a)
∂Ωf
Z
θj = χj nds for j = 1, . . . , Np , (7.48b)
∂Ωf
 7.3. Numerical Experiments 157

where each of ∇φi (x) and χj can be viewed as a velocity and a pressure field, respectively. The
terms (7.48a) and (7.48b) can be precomputed during the offline stage and then stored.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

In the online stage, when new time and parameter values of t∗ and µ∗ , respectively, are
introduced, the full-order forces are computed as
Z
∗ ∗
F (t ; µ ) = (2µ∇u(x, t∗ ; µ∗ ) − p(x, t∗ ; µ∗ )I)nds, (7.49)
∂Ωf

while the reduced order approximation is

Nu Np
X X
Fr (t∗ ; µ∗ ) = ai (t∗ ; µ∗ )δi − bj (t∗ ; µ∗ )θj . (7.50)
i=1 j=1

It is important to underline the fact that the above formulas for the reduced forces are valid only
in the case of affine parameter dependency.

7.3 Numerical Experiments

In this section we will test and compare two different ROM techniques for laminar flows: the mo-
mentum and PPE-ROM presented in Subsection 7.2.3 and the supremizer stabilization method
presented in Subsection 1.2.4. The two different proposed stabilization methods are tested and
compared on two benchmark test cases. The first benchmark consists of the well-known and
much-studied lid-driven cavity problem [532]. The second benchmark consists of the flow
around a circular cylinder for moderate Reynolds number (100 < Re < 200) [525]. In the first
case, any kind of parametrization is introduced, while in the second case the kinematic viscosity
is parametrized.

7.3.1 Lid-Driven Cavity Problem

As mentioned above, the first proposed benchmark consists of the well-known lid-driven cavity
problem. The simulation is carried out on a two-dimensional square domain of length L = 0.1 m.
The boundary is subdivided into two different parts, Γ = ΓD ∪ Γ0 , and the boundary conditions
for velocity and pressure are set according to Figure 7.2. At the top of the cavity a constant
uniform and horizontal velocity equal to ux = 1 m/s is prescribed. The mesh is structured and
counts 40000 quadrilateral cells, 200 on each dimension of the square. The kinematic viscosity

ΓD Γ0
u u = (1, 0) u = (0, 0)
p ∇p · n = 0 ∇p · n = 0

Figure 7.2. Sketch of the mesh for the lid-driven cavity problem together with the boundary
subdivisions and boundary conditions.
 158 Chapter 7. Finite Volume–Based Reduced Order Models for Laminar Flows

Table 7.1. The table contains the cumulative eigenvalues for the lid-driven cavity test. The first,
second, and third columns report the cumulative eigenvalues for the velocity, pressure, and supremizer
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

fields, respectively. The last column contains the value of the inf-sup constant, in the supremizer stabi-
lization case, for different numbers of supremizer modes and with a fixed number of velocity and pressure
modes (10 modes for velocity and 10 modes for pressure).

N Modes u p s β
1 0.978946 0.975406 0.980260 9.264e-05
2 0.994184 0.991528 0.995232 9.264e-05
3 0.997737 0.995385 0.997912 7.175e-04
4 0.998990 0.998116 0.999400 7.175e-04
5 0.999483 0.999270 0.999844 7.175e-04
10 0.999971 0.999971 0.999997 1.551e-02

is equal to ν = 1 × 10−4 m2 /s, which leads to a Reynolds number of 1000. For the full-
order simulation, the time discretization is treated using a second-order backward differencing
scheme, while the discretization in space is performed with a fourth-order interpolation scheme.
The time step is kept constant and equal to ∆t = 5 × 10−4 , and the simulation is run till
T = 10 s. The snapshots are acquired every 0.01 s, giving a total number of snapshots equal
to 1000. For the ROM, the dimensions of the reduced spaces for velocity and pressure are set,
for both the presented methodologies, equal to Nur = 10 and Npr = 10. In the SUP-ROM
the reduced space for velocity is enriched with 10 additional supremizer modes. This selection
is done according to Table 7.1, where it is possible to observe that this number of modes is
sufficient to retain more than 99.9% of the energy for both velocity and pressure. We remark
that in this numerical experiment any kind of parametrization can be introduced. The ROM
is in fact used to simulate the same conditions tested in the full-order setting, and the results
are compared to the full-order simulation results. The time discretization, at the reduced order
level, is treated by using a first-order backward Newton method. Figure 7.8 depicts a comparison
between the high-fidelity simulation and the ROM one, for both velocity and pressure fields, at
different time instants. As one can see from the figure, both models are capable of reproducing
the main flow pattern for both fields. Figure 7.3 reports the evolution in time of the L2 relative
error for velocity and pressure, respectively. Figure 7.4 shows a qualitative comparison of the
velocity and pressure fields for the high-fidelity, the SUP-ROM and the PPE-ROM. The plots
also report the error without any type of stabilization. It is clear that, without stabilization,

105

104
100
||uH F − uROM ||/ ||uH F ||

||pH F − pROM ||/ ||pH F ||

103

USUP 102 PSUP

UPPE PPPE
−1 UNOS PNOS
10 101

100

10−1
10−2

0 2 4 6 8 10 0 2 4 6 8 10
Time(s) Time(s)

Figure 7.3. Error analysis for the velocity field. The L2 -norm of the relative error is plotted
over time for three different models: with supremizer stabilization (USUP, continuous red line), with PPE
stabilization (UPPE, dotted blue line), and without stabilization (PNOS, dashed green line). The ROMs are
obtained with 10 modes for velocity, pressure, and supremizer.
 7.3. Numerical Experiments 159
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 7.4. Comparison of the velocity and pressure fields for high-fidelity (UHF, column 1;
PHF, column 4), SUP-ROM (USUP, column 2; PSUP, column 5), and PPE-ROM (UPPE, column 3; PPEE,
column 6). The fields are depicted for different time instants equal to t = 0.2 s, 0.5 s, 1 s, and 5 s,
respectively, and increasing in the image from top to bottom. The ROM models are obtained with 10 modes
for velocity and pressure and for the SUP-ROM only with 10 additional supremizer modes. The velocity
and pressure magnitudes are shown in the image legends.

even though the ROM does not diverge, both the velocity and pressure fields are completely
unreliable. For this particular numerical test, the SUP-ROM produces, with respect to the PPE-
ROM, worse results for the velocity field but better results for the pressure field. This difference
can be justified by the fact that, within a supremizer stabilization technique, the POD velocity
space is enriched by unnecessary (for the correct reproduction of the velocity field) supremizer
modes. During the initial transient, both fields present a higher relative error due to the relatively
low number of snapshots acquired during the initial transient. The snapshots, as highlighted
above, are equally distributed in time, and, to enhance the performance of the ROM, one should
concentrate the snapshots in the timespan where the system exhibits the most nonlinear behavior.
For the SUP-ROM, according to [34], the number of supremizer modes is chosen equal to the
number of pressure modes. Table 7.1 also reports the value of the inf-sup constant β, obtained
by keeping the number of velocity and pressure modes (10 modes for velocity and 10 modes for
pressure) constant and varying the number of supremizer modes. As one can observe from the
table, increasing the number of supremizer modes leads to a remarkable increase in the inf-sup
constant.
 160 Chapter 7. Finite Volume–Based Reduced Order Models for Laminar Flows
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

ΓIn Γ0 Γs ΓOut
u u = (1, 0) u = (0, 0) u·n=0 ∇u · n = 0
p ∇p · n = 0 ∇p · n = 0 ∇p · n = 0 p=0

Figure 7.5. The figure shows a general overview of the mesh with dimension and boundaries
(upper left), a table with the imposed values at the boundaries (bottom), and a zoom of the mesh near the
cylinder (upper right).

7.3.2 Flow around a Circular Cylinder

The second example, which aimed to test the methodologies on a more complex flow field and
mesh structure, consists of the benchmark of the flow around a circular cylinder. In this numerical
example the physical parametrization due to the parametrized physical viscosity is introduced.
Furthermore, this numerical experiment has also been used to test the performance of both stabi-
lization methods on periodic systems for long time integrations, i.e., a wider window with respect
to time was used to create the POD bases. The mesh, which is depicted in Figure 7.5 together
with the boundary conditions, is mainly composed of quadrilateral cells. It is refined in the prox-
imity of the cylinder and counts a total number of 43762 cells. The mesh is generated starting
from a structured grid with a base resolution of 200 cells along the x direction and 80 cells along
the y direction, and it is successively refined around the cylinder, with five progressive layers of
refinement. The time step is set equal to ∆t = 0.005 s, which is sufficiently small to meet the
CFL condition in every part of the domain. Since the mesh is sufficiently fine and the time step
is sufficiently small, in the full-order simulation, for all the terms, including the convective term,
a fourth order spatial interpolation scheme is used. In this numerical experiment, the physical
parametrization given by the kinematic viscosity ν is introduced. To train the ROM, five different
values of the kinematic viscosity are used. The values of the kinematic viscosity are chosen us-
ing a uniform distribution inside the range ν ∈ [0.005, 0.01]. These values of viscosity result in
the values of the Reynolds number Re ∈ [100, 200]. In this numerical experiment the simulation
is run, for each value of the kinematic viscosity inside the training set, for 200 s. This time is
long enough to achieve a completely evolved vortex-shedding pattern. Since we were interested
in the correct reproduction of the ROM during the periodic response regime, only the last 10 s
of simulation are used to collect the snapshots for the POD basis generation. Within this time
window, the snapshots are collected every 0.05 s, returning a total of 200 snapshots for each dif-
ferent value of the kinematic viscosity. The snapshots of the five different full-order simulations
are then used to create the POD basis functions, which result in the cumulative eigenvalues of
Table 7.2. Figure 7.7 depicts the first four basis functions for velocity, pressure, and suprem-
izers. The ROM counts, for the SUP-ROM, 15 modes for velocity and 10 modes for pressure,
and 12 modes for supremizers, while the PPE-ROM has 15 modes for velocity and 10 modes for
pressure.
It is worth remarking that in this example, in the SUP-ROM, the supremizer space counts
more modes than the pressure space. This choice is made in order to improve the accuracy and
the stability of the results. For this particular case, in fact, we have experimentally observed that
an equal number of pressure and supremizer modes leads to inaccurate results. This is justified by
the fact that, using the approximated approach described in [34] for the supremizer enrichment,
 7.3. Numerical Experiments 161

Table 7.2. The table contains the cumulative eigenvalues for the cylinder problem. In the first,
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

second, and third columns are reported the cumulative eigenvalues for the velocity, pressure, and supremizer
fields, respectively, as a function of the number of modes. In the last column is reported the value of the
inf-sup constant for the supremizer stabilization case for different numbers of supremizer modes with a fixed
number of velocity and pressure modes (15 modes for velocity and 10 modes for pressure).

N Modes u p s β
1 0.390813 0.793239 0.921046 2.608e−04
2 0.598176 0.85809 0.941746 4.492e−04
3 0.802176 0.911636 0.961438 7.869e−03
4 0.879096 0.934997 0.978072 1.662e−02
5 0.949519 0.955578 0.98669 1.662e−02
10 0.986025 0.992347 0.998307 1.098e−01
15 0.995922 0.997994 0.999732 1.199e−01

010
USUP ν = 0 005
UPPE ν = 0 005
kinetic energy relative error

USUP ν = 0 005625
005 UPPE ν = 0 005625

000

−005

−010
0 10 20 30 40 50 60 70 80
Time(s)

Figure 7.6. Kinetic energy relative error in the cylinder example for ν = 0.005 and ν =
0.005625. The kinetic energy relative error is plotted over time for the two values of viscosity and for
the two different models: with supremizer stabilization (USUP and PSUP) and PPE stabilization (UPPE
and PPPE). The ROM solutions are obtained with 15 modes for velocity, 10 modes for pressure, and 12
modes for supremizers. The time window in this case is wider than the one used to generate the RB spaces
(∆T = 10 s).

Figure 7.7. First four basis functions for velocity (first row), pressure (second row), and suprem-
izers (third row).
 162 Chapter 7. Finite Volume–Based Reduced Order Models for Laminar Flows
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 7.8. Comparison of the velocity field for high-fidelity (UHF, first row), supremizer stabi-
lized ROM (USUP, second row), and PPE stabilized ROM (UPPE, third row). The fields are depicted for
different time instants equal to t = 195 s, 200 s, 230 s, and 270 s and increasing from left to right. The
ROM solutions are obtained with 15 modes for velocity and 10 modes for pressure, and for the SUP-ROM
only with 12 additional supremizer modes. The velocity magnitude is shown in the image legends.

equal dimensions for the pressure and the supremizer spaces do not automatically guarantee the
fulfillment of the inf-sup condition.
To test the ROMs, the results are compared against the full-order results for an intermediate
value of the viscosity, ν = 0.005625, which is not included in the values of viscosity (ν =
[0.005, 0.00625, 0.0075, 0.00875, 0.01]) employed to generate the snapshots used to create the
RB spaces. The results were compared on two different time windows. The first one covers 10 s
of simulation and coincides with the time window used to generate the snapshots; the second one
covers 80 s of simulation and therefore is much wider than the time window used to generate
the snapshots. Figures 7.8 and 7.9 show the comparison, for velocity and pressure, respectively,
between the results obtained with the full-order model, the SUP-ROM, and the PPE-ROM. The
fields are depicted at four different time instants equal to t = 195 s, 200 s, 230 s, and 270 s.
The first two time instants are, respectively, in the middle and at the end of the time window
used to generate the snapshots, while the other two time instants are outside of it. Figure 7.10

Figure 7.9. Comparison of the pressure field for high-fidelity (PHF, first row), supremizer stabi-
lized ROM (PSUP, second row), and PPE stabilized ROM (PPPE, third row). The fields are depicted for
different time instants equal to t = 195 s, 200 s, 230 s, and 270 s and increasing from left to right. The
ROM solutions are obtained with 15 modes for velocity and 10 modes for pressure, and for the SUP-ROM
only with 12 additional supremizer modes. The pressure magnitude is shown in the image legends.
 7.3. Numerical Experiments 163

0.040 0.20
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

USUP PSUP
0.035 UPPE PPPE
uROM ||/ ||uH F ||

0.030 0.15

pROM ||/ ||pH F ||

0.025

0.020 0.10

0.015

||pH F
||uH F

0.010 0.05

0.005

0.000 0.00
0 2 4 6 8 10 0 2 4 6 8 10
Time(s) Time(s)

Figure 7.10. Error analysis for the velocity (left plot) and pressure (right plot) fields in the cylin-
der example with ν = 0.005625. The L2 -norm of the relative error is plotted over time for the two different
models: with supremizer stabilization (USUP and PSUP, continuous red line) and PPE stabilization (UPPE
and PPPE, dot-dashed blue line). The ROM solutions are obtained with 15 modes for velocity, 10 modes
for pressure, and 12 modes for supremizers.

reports the L2 -norm of the relative error for velocity and pressure in the 10-s-wide time window
for ν = 0.005625. The figure also confirms the behavior observed in the cavity example: the
SUP-ROM produces worse results for the velocity field but better results for the pressure field.
Figure 7.11 shows the same plots in a wider time window, and also for one of the values of
viscosity (ν = 0.005) used to generate the full-order snapshots. For both ROMs, the relative
error is increasing in time. Cross-referencing the data of Figure 7.11 with the plots of Figures 7.8
and 7.9, one can deduce that the increase in the error is given, for the SUP-ROM, by the numerical
instabilities that occur for long time integrations. In the last time step (t = 270 s), it is in fact
possible to observe, for both velocity and pressure, a completely incorrect and nonphysical flow
pattern. For the PPE-ROM, instead, the flow pattern still looks regular and sufficiently similar to

1.0 3.0
USUP ν = 0.005 PSUP ν = 0.005
UPPE ν = 0.005 PPPE ν = 0.005
USUP ν = 0.005625 2.5 PSUP ν = 0.005625
0.8
uROM ||/ ||uH F ||

UPPE ν = 0.005625 PPPE ν = 0.005625

pROM ||/ ||pH F ||

2.0
0.6
1.5
0.4
1.0
||pH F
||uH F

0.2
0.5

0.0 0.0
0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80
Time(s) Time(s)

Figure 7.11. Error analysis for the velocity (left plot) and pressure (right plot) fields in the
cylinder example with ν = 0.005 and ν = 0.005625. The L2 -norm of the relative error is plotted over
time for the two values of viscosity and for the two different models: with supremizer stabilization (USUP
and PSUP) and PPE stabilization (UPPE and PPPE). The ROMs are obtained with 15 modes for velocity,
10 modes for pressure, and 12 modes for supremizers. The time window is in this case wider than the one
used to generate the RB spaces (∆T = 10 s).
 164 Chapter 7. Finite Volume–Based Reduced Order Models for Laminar Flows

the high-fidelity one but it is possible to observe a phase shift between the high-fidelity and the
ROM solution. The PPE-ROM, in fact, even though it continues to produce a regular and physical
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

pattern, has a period of vortex shedding which is slightly longer than in the HF solution. To show
a better idea about the behavior of the different ROMs for long time integrations, Figure 7.6
depicts the relative error of the total kinetic energy. It is well known that POD-Galerkin models
are affected, in fact, by a blow-up energy issue [137, 196]. It is not the objective of this work
to deal with long time integration instabilities, but it is worth checking which kind of pressure
stabilization method is likely prone to this issue. From the figure it is clear that the PPE-ROM
accurately preserves the total kinetic energy of the system. On the other hand, the SUP-ROM
exhibits an oscillating behavior with an increase in the total kinetic energy.

7.3.3 Comments on the Results

The two proposed numerical examples permit us to draw some conclusions regarding the perfor-
mances of the two different stabilization methods. The SUP-ROM produces better results for the
pressure field and worse results for the velocity field. This fact may be justified by the additional
and unnecessary (only in terms of correct representation of the velocity fields) supremizer modes
that pollute the POD velocity space. The PPE-ROM, on the other hand, is more reliable for long
time integrations.
Table 7.3 shows the results in terms of computational costs. It is possible to deduce that both
models, for both cases, permit us to achieve a considerable speedup. For the cavity example,
both the offline and the online stages are computed in serial on one processor. On the other hand,
in the cylinder example, the offline stage is performed in parallel with six processors, while the
online stage is still performed with a serial run on one processor.
The SUP-ROM is less efficient than the PPE-ROM. The SUP-ROM, in fact, in comparison
with the PPE-ROM, due to the additional supremizer modes, gives rise to a bigger reduced dy-
namical system. Both ROMs were able to capture with sufficient accuracy (especially from an
engineering standpoint) the main features of the flow field for both velocity and pressure. The
SUP-ROM is likely prone to instability issues for long time integrations.

Table 7.3. The table contains the computational time for the supremizer (SUP) and the PPE
stabilization techniques. In the cavity experiment, the SUP-ROM is obtained with 10 modes for velocity,
pressure, and supremizers, while the PPE-ROM is obtained with 10 modes for pressure and supremizers.
In the cylinder experiment, the SUP-ROM is obtained with 15 modes for velocity, 10 for pressure, and 12
for supremizers, while the PPE-ROM is obtained with 15 modes for velocity and 10 for pressure.

HF SUP-ROM PPE-ROM
Cavity Exp. 25min 7.64s 4.86s
Cylinder Exp. 18.5min × 6proc. 3.14s 0.971s
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Chapter 8

Finite Volume–Based
Reduced Order Models
for Turbulent Flows

Matteo Zancanaro, Saddam Hijazi, Michele Girfoglio,

Andrea Mola, Giovanni Stabile, Gianluigi Rozza

8.1 Introduction
This chapter focuses on turbulence modeling for incompressible flows at both the full- and the
reduced order level.
In the previous chapter, we dealt with the full-order incompressible Navier–Stokes equations
(NSEs) for both steady and unsteady flows within a laminar regime by referring to the solution
algorithms available in OpenFOAM. Here, we will follow an analogous work flow for turbulence
modeling. Several models are implemented in OpenFOAM. In this work we mainly rely on two
of them: the Reynolds-averaged Navier–Stokes (RANS) equations and the large eddy simulation
(LES).
On the other hand, at the reduced order level, we mainly propose a hybrid projection/data-
driven strategy: a classical proper orthogonal decomposition (POD)-Galerkin projection ap-
proach for the reconstruction of the velocity and the pressure fields and a data-driven reduc-
tion method based on radial basis functions (RBFs) to approximate the turbulent eddy viscosity.
However, for the sake of completeness, we also review other reduced order modeling (ROM)
approaches.

8.2 RANS Equations

In this section, we address the RANS approach. In particular, in Subsection 8.2.1 we introduce
the idea of Reynolds averaging and the closure problem in turbulence. Then, in Subsection 8.2.2,
the notion of eddy viscosity models (EVMs) is presented.

8.2.1 Closure Problem and Reynolds Averaging

One of the main features of turbulent flows is that several flow properties in this regime exhibit
high-frequency random oscillations in both the time and the space domain. For this reason,
Osborne Reynolds [485] introduced the concept of time averaging, in which all the fields are

165
 166 Chapter 8. Finite Volume–Based Reduced Order Models for Turbulent Flows

expressed as the sum of mean and fluctuating parts. In the majority of cases, the fluctuating
component appears as a vibration around an equilibrium, or average flow solution. In a common
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

case, one is interested in solving only for the time-averaged part of the fluid dynamic variables.
To achieve this, following the procedure introduced by Reynolds, after decomposing all the vari-
ables, the momentum and continuity equations are time averaged. So one obtains a new system
in which the unknowns include the mean components of all the flow fields. However, the pres-
ence of the nonlinearity associated with the convective term leads to residual terms in which the
fluctuations are still present in the new system, where the number of equations is not equal to the
number of unknowns. This is called the closure problem.
If we consider a generic fluid dynamics scalar field called σ(x, t), then the Reynolds average
operation for σ(x, t) gives
σ(x, t) = σ(x, t) + σ 0 (x, t), (8.1)
where σ(x, t) and σ 0 (x, t) are the mean and the fluctuating parts, respectively. For steady flows,
the mean part σ is a spatial field without dependence on time, i.e., σ = σ(x). In this case, the
mean field is computed as
t+T
1
Z
σ(x) = lim σ(x, τ )dτ. (8.2)
T →∞ T t

As for the unsteady case, let T1 and T2 be the time scales of the fluctuating and the mean fields,
respectively. Time averaging is computed by the integral
t+T
1
Z
σ(x, t) = σ(x, τ )dτ, T1 ≤ T ≤ T2 . (8.3)
T t

Hereinafter, we will use the overbar notation to indicate any time-averaged quantity. We
recall that the time average of the mean field σ(x, t) is the mean field itself, while the time
average of the fluctuating part σ 0 (x, t) is zero.
We consider the time average of the product of two scalar quantities named φ and ψ:

φψ = (φ + φ0 )(ψ + ψ 0 ) = φψ + φψ 0 + ψφ0 + φ0 ψ 0 = φψ + φ0 ψ 0 , (8.4)

where one exploits the fact that the product of a mean quantity and a fluctuating quantity has zero
mean. On the other hand, the quantity φ0 ψ 0 is not zero, which basically means that the product
of the means φψ is not equal to the mean of the products φψ. In fact the two quantities φ and ψ
are called uncorrelated if φ0 ψ 0 = 0; otherwise they are correlated.
Now we are ready to derive the RANS system. We start from the NSEs (7.1) written in scalar
form:
∂ui
= 0, (8.5)
∂xi
∂ui ∂ui ∂p ∂Sji
+ uj =− + 2ν , (8.6)
∂t ∂xj ∂xi ∂xj
where the Einstein summation convention has been used and where S is the strain-rate tensor,
given by  
1 ∂ui ∂uj
Sij = + . (8.7)
2 ∂xj ∂xi
The convective term can be written as
∂ui ∂ (ui uj ) ∂uj ∂ (ui uj )
uj = − ui = , (8.8)
∂xj ∂xj ∂xj ∂xj
 8.2. RANS Equations 167

where (8.5) has been employed. By using (8.8) we can recast the momentum equation (8.6) as
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

∂ui ∂ (ui uj ) ∂p ∂Sji

+ =− + 2ν . (8.9)
∂t ∂xj ∂xi ∂xj

The application of Reynolds time averaging to (8.5) and (8.9) leads to the RANS equations,
namely
∂ui
= 0, (8.10)
∂xi
 
∂ S − u0 u0
∂ui ∂ui ∂p ji i j
+ uj =− + 2ν . (8.11)
∂t ∂xj ∂xi ∂xj
We point out that the averaged continuity equation is identical to the original one, which also
means that the fluctuating velocity component u0i has zero divergence. As for the momentum
equation, not all the fluctuating terms appearing in it have vanished. In fact, the term u0i u0j still
appears in the momentum equation because of the nonlinearity of the convective term. The term
u0i u0j is often referred to as the Reynolds stress tensor R, with Rij = u0i u0j . The Reynolds stress
tensor R is a symmetric tensor with six unknown components. Therefore, in the RANS system
(8.11) and (8.10) there are six additional unknowns with respect to the Navier–Stokes system,
without any additional equation. The bottom line in turbulence modeling is to deal with the
Reynolds stress tensor in order to close the system so as to compute the mean flow fields. Such
an issue in turbulence modeling is the aforementioned closure problem.

8.2.2 EVMs
Over the years the closure problem has been addressed by several methodologies characterized
by different computational costs and accuracy of the results. Among others, we refer to the meth-
ods based on the derivation of a transport equation for the Reynolds stress tensor components.
Such closure models are obtained by multiplying the NSEs by a fluctuating quantity and then
taking the mean. However, this procedure will result in additional nonlinear terms containing
fluctuation product averages, which furthermore increase the number of unknowns. Specifically,
these additional unknowns include a correlation term involving fluctuating velocity components
coming from the convective term, and also other correlation terms from the diffusive contribution
and a term which amounts to a fluctuating velocity-pressure correlation. Overall, the procedure
brings in 22 new unknowns. In order to avoid such complexity, several algorithms in RANS tur-
bulence modeling are based on the Boussinesq eddy-viscosity assumption [71]. The Boussinesq
assumption (or hypothesis) states that the Reynolds stress tensor is proportional to the traceless
strain rate tensor S, namely
2
Rij = 2νt Sij − kδij , (8.12)
3
where k is the turbulent kinetic energy per unit mass and νt is the so-called artificial or eddy vis-
cosity. The turbulence models which employ the Boussinesq eddy viscosity assumption are often
called eddy viscosity models (EVMs). In these models, the eddy viscosity is approximated by
making use of either algebraic relations or transport-diffusion-reaction partial differential equa-
tions (PDEs) for other quantities which have an algebraic relationship with the eddy viscosity.
The first group of EVMs is the algebraic models or the zero-equation models. As an example we
mention the mixing length turbulence model [462]. The first example of a PDE-based model is
the Spalart–Allmaras model [546]. In this model one solves for a viscosity-like variable called
ν̃. Richer EVMs are based on solving two additional PDEs for specific turbulent flow variables,
 168 Chapter 8. Finite Volume–Based Reduced Order Models for Turbulent Flows

among which we mention the k- and the k-ω models [314, 347]. In these models  and ω rep-
resent the specific turbulent dissipation rate of the turbulent kinetic energy into internal thermal
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

energy. The turbulent kinetic energy from the turbulent velocity fluctuations u0i is given by
1 0 0 1 2 2 2
k= u u = (u01 + u02 + u03 ). (8.13)
2 i i 2
As for , we have
∂u0i ∂u0i
=ν , (8.14)
∂xk ∂xk
and for ω we get

ω= , (8.15)
kβ ∗
where β ∗ is a constant equal to 0.09.
The first form of a two-equation turbulence model was presented by Kolmogorov in [347] as
a k-ω model. Later on, other forms of the k-ω model were proposed with significant improve-
ments, such as in [514, 617, 618]. In [361] the standard k- model is presented.
The RANS equations have been utilized extensively in the literature for modeling turbulence
in different application fields. In [207] one may find an application of the RANS in aerodynam-
ics, in particular for modeling a high-speed aerodynamic flow transition. In [331], numerical
simulations for steady turbulent problems in automotive engineering using the Spalart–Allmaras
model are performed. The RANS equations have also been used to study wind dynamics in
civil engineering: in [405] the authors present a three-dimensional study involving the use of
steady RANS to predict mean wind pressure distributions on windward and leeward surfaces of
a medium-rise building with and without balconies. In [200], the k- model is used to describe
the hydrodynamic flow around a boat. The work [141] presents an application of the RANS
equations for the design of sailing yachts.
Another two-equation model called the SST k-ω is presented in [395]: this turbulence model
merges the k-ω model and the k- model by combining their advances. The SST k-ω is used
extensively in the computational fluid dynamics (CFD) community because it is a good compro-
mise for flows with detachment and recirculation. This model has also been used in this work,
so we have chosen to report its complete mathematical formulation.
The turbulent kinetic energy k equation is given by
 
∂k ∂k ∂ ∂k
+ uj = P − β ∗ ωk + (ν + σk νt ) , (8.16)
∂t ∂xj ∂xj ∂xj

while the specific turbulent dissipation rate ω equation yields

 
∂ω ∂ω 2 2 ∂ ∂ω σω2 ∂k ∂ω
+ uj = αΩs − βω + (ν + σω νt ) + 2(1 − F1 ) . (8.17)
∂t ∂xj ∂xj ∂xj ω ∂xj ∂xj

As for the terms and the constants which appear in the two equations above, they are defined as
follows:
   
∂ui 1 ∂ui ∂uj
, 10β ∗ ωk , Wij =
p
P = min Rij − , Ωs = 2Wij Wij ,
∂xj 2 ∂xj ∂xi
a1 k
νt = , φ = F1 φ1 + (1 − F1 )φ2 ,
max(a1 ω, Ωs F2 )
" √ ! #
4

k 500ν 4σω2 k
F1 = tanh arg1 , arg1 = min max , , ,
β ∗ ωy y 2 ω CDkω y 2
 8.3. LES 169
 
1 ∂k ∂ω −10
CDkω = max 2ρσω2 , 10 ,
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

ω ∂xj ∂xj
√ !
2

k 500ν
F2 = tanh arg2 , arg2 = max 2 ∗ , 2 ,
β ωy y ω
σk1 = 0.85, σw1 = 0.65, β1 = 0.075,
σk2 = 1.00, σw2 = 0.856, β2 = 0.0828,
β ∗ = 0.09, a1 = 0.31.
Hence the RANS equations coupled with the SST k-ω turbulence model can be written as
 h  i
T



∂u
∂t + ∇ · (u ⊗ u) = ∇ · −pI + (ν + νt ) ∇u + (∇u) in Ω × [0, T ],
∇ · = 0 in Ω × [0, T ],



 u
νt = max(aa11ω,Ω k
in Ω × [0, T ],

s F2 ) (8.18)
 ∂t + u · ∇k = P − β ∗ ωk + ∇ · [(ν + σk νt ) ∇k] in Ω × [0, T ],


∂k
∂ω
+ u · ∇ω = αΩ2s − βω 2 + ∇ · [(ν + σω νt ) ∇ω]


 ∂t


+2(1 − F1 ) σωω2 ∇ω · ∇k in Ω × [0, T ].

8.3 LES
Large eddy simulation (LES) was introduced for the first time in 1963 by Joseph Smagorinsky
to simulate atmospheric flows [541]. The basic idea of LES is to obtain a significant reduction
of the computational cost required by a DNS by neglecting the smallest length scales via low-
pass filtering of the NSEs. Then the information related to such scales is recovered by proper
additional modeling. So one can assert that in a LES framework the largest length scales are
simulated while the smallest length scales are modeled.
A LES filter can perform a spatial and/or temporal filtering operation. The filtered field
ϕ(x, t) associated with a field ϕ(x, t) is defined as
Z ∞Z ∞
ϕ(x, t) = ϕ(r, τ )G(x − r, t − τ )dτ dr, (8.19)
−∞ −∞

where G is the filter convolutional kernel, which can be expressed as

ϕ = G ? ϕ. (8.20)
The filter kernel G removes from ϕ the scales smaller than the associated cutoff length scale ∆
and cutoff time scale T . So the field ϕ can be split into a filtered and a subfiltered contribution:
ϕ = ϕ + ϕ0 . (8.21)
Now we derive the filtered incompressible NSEs. By applying the filter to the continuity equa-
tion, we obtain
∇ · u = 0. (8.22)
On the other hand, filtering the momentum equation results in
∂u 1
+ ∇ · (u ⊗ u) = − ∇p + ν∆u. (8.23)
∂t ρ
By assuming that the filtering and differentiation operations are commutative, we get
∂u 1
+ ∇ · (u ⊗ u) = − ∇p + ν∆u. (8.24)
∂t ρ
 170 Chapter 8. Finite Volume–Based Reduced Order Models for Turbulent Flows

Notice that, unlike RANS modeling, it results in u ⊗ u 6= u ⊗ u. Therefore, the evaluation

of the nonlinear filtered advection term ∇ · (u ⊗ u) would require knowledge of the unfiltered
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

velocity u. Equation (8.24) can be recast as

∂u 1
+ ∇ · (u ⊗ u) = − ∇p + ν∆u − ∇ · τ SGS , (8.25)
∂t ρ

where
τ SGS = u ⊗ u − u ⊗ u (8.26)
is the subgrid scale stress tensor that takes into account the effect of filtered (namely solved)
scales and small (namely unsolved) scales as well as their interaction and needs to be properly
modeled. By using (8.21) we obtain the following expression for τ SGS :

τ SGS = u ⊗ u − u ⊗ u + u ⊗ u0 − u0 ⊗ u + u (8.27)



0 u0 ,
| {z } | {z } |{z}
R
L C

where L is the Leonard stress tensor involving only the resolved quantities (and therefore it can
be computed). However, C, the cross-term stress tensor, and R, the subgrid scale Reynolds stress
tensor (which can be interpreted as an analogous term to the Reynolds stress tensor in the RANS
equations), are related to unresolved scales. Then the filtered NSEs are given by (8.22)–(8.25).
A LES filtering approach that has recently been successfully used in the context of numerical
simulations of incompressible flows at moderately large Reynolds numbers is the Leray model
[371]. It couples the NSEs with a differential filter, resulting in the following system:

∂u 1
+ ∇ · (u ⊗ u) = − ∇p + ν∆u, (8.28)
∂t ρ

∇ · u = 0, (8.29)
−δ 2 ∇ (a(u)∇u) + u + ∇λ = u, (8.30)
∇ · u = 0, (8.31)
where δ can be interpreted as the filtering radius (namely the radius of the neighborhood where
the filter extracts information from the unresolved scales), the variable λ is a Lagrange multiplier
to enforce the incompressibility constraint for the filtered velocity u, and a(·) is the so-called
indicator function such that a(u) ≈ 0 where the flow does not need regularization and a(u) ≈ 1
where the flow does need regularization. The indicator function plays a crucial role in the success
of the Leray model. The choice a = 1 corresponds to the classic Leray model [371] that has the
advantage of making the operator in the filter equations linear and constant in time, but its ef-
fectivity is rather limited because it introduces the same amount of regularization in each region
of the computational domain, which results in overdiffusion in many cases. Different choices of
a(·) have been proposed and compared in [72, 599, 300, 362]. One of the most mathematically
convenient indicator functions is a(u) = |∇u| [69] because of its strong monotonicity prop-
erties. With this choice, we recover a Smagorinsky-like model, which is however known to be
not sufficiently selective. Recently, a class of deconvolution-based indicator functions has been
investigated:
a = |v − F (v)|, (8.32)
where F is typically the linear Helmholtz filter operator FH defined by

F = FH = (I − α2 ∆)−1 . (8.33)
 8.4. Hybrid Projection-Based/Data-Driven ROM for Turbulent Flows 171

Notice that finding FH v = ṽ is equivalent to finding ṽ such that

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

ṽ − α2 ∆ṽ = v. (8.34)

A monolithic approach to solving the problem (8.28)–(8.31) would be associated with a high
computational cost. In order to decouple the Navier–Stokes system (8.28)–(8.29) from the filter
system (8.30)–(8.31) at time tn+1 , a solve-then-filter approach can be employed: (i) solve equa-
tions (8.28)–(8.29) first, replacing the advection field un+1 with a suitable extrapolation, and (ii)
then solve the filter problem (8.30)–(8.31). In this framework, the evolve-filter-relax (EFR) algo-
rithm has recently been proposed, providing very promising results; see, e.g., [362, 58, 229, 231].
It was shown in [430] that the EFR algorithm is equivalent to a generic viscosity model in LES.
The two main advantages of the EFR method over other LES models are (i) the modularity, which
in fact introduces a differential problem to the Navier–Stokes problem unlike, e.g., the variational
multiscale approach [49], where extra terms are added directly in the NSEs, and (ii) the ease of
implementation; in fact the filter problem can be interpreted as a generalized Stokes problem and
can therefore be employed in a legacy Navier–Stokes solver by allowing it to simulate higher
Reynolds number flows without major modifications to the software core.
For a complete validation of the EFR method for Reynolds numbers up to 6500 in fixed
domains, we refer the reader to [58, 229].

8.4 Hybrid Projection-Based/Data-Driven ROM for

Turbulent Flows
In Sections 8.2 and 8.3 we addressed turbulence modeling at the full-order level. The focus
now shifts to the treatment of turbulence at the reduced order level. We have mentioned that
turbulence is simulated using RANS or LES equations with the help of a suitable eddy viscosity
or subgrid scale model, respectively. Thus, the construction of a turbulent ROM has to take
into account the fact that the full-order model involves additional variables associated with the
specific turbulence model. We remark that hereinafter the focus will be on ROMs for RANS
equations, although the same techniques could be successfully extended to a LES framework, as
shown in some recent works [230, 233, 232, 561].
A possible methodology which could be adapted at the reduced order level involves decom-
posing each turbulence variable, such as k, , ω or ν̃, into a finite sum of spatial modes multiplied
by scalar coefficients, as is done in Chapter 7 with the velocity and the pressure field in the con-
struction of ROMs for laminar flows. This approximation needs to be followed by a Galerkin
projection of the additional transport-diffusion PDEs in order to retrieve the evolution of the
reduced coefficients. A set of reduced equations is obtained by projecting the momentum, the
continuity, and the additional turbulence model equations. These reduced equations are then
coupled, and the reduced solution for velocity, pressure, and each of the turbulence variables can
be obtained by solving the coupled system. Yet, the last approach has a number of issues. In
particular, we need to consider the following:

• The approach complicates the task of solving the reduced order system in the online stage.
This is due to the additional equations, which have to be treated at the reduced order level,
where the additional turbulent equations are characterized by high nonlinearity.

• The approach forces the creation and the maintenance of a ROM for each turbulence clo-
sure model available at the full-order level. This customization of the ROM is clearly not
practical. For instance, a popular library such as OpenFOAM is well supplied with various
closure models, all of which would require a separate and specific ROM.
 172 Chapter 8. Finite Volume–Based Reduced Order Models for Turbulent Flows

Since the goal here is to design unified reduction methodologies which work with multiple full-
order model solvers, each with its own specific implementation of several different turbulence
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

models, we prefer to opt for a different strategy. To this end, the proposed approach involves
extending the decomposition assumption only to the eddy viscosity νt without doing the same
for the other turbulence variables, such as k, , ω, or ν̃. This essentially means that a reduced
order version of the eddy viscosity (νtr ) is introduced, namely
Nνt
X
νt (x, t; µ) ≈ νtr (x, t; µ) = gi (t, µ)ηi (x), (8.35)
i=1

where ηi (x) is the ith eddy viscosity POD mode and gi (t, µ) is the ith coefficient of the POD
expansion. If one considers the above expansion as an extension to those used for the supremizer
(SUP)-ROM and the Poisson pressure equations (PPE)-ROM, then it can be seen that the reduced
eddy viscosity assumes a different set of degrees of freedom for ai (t, µ) and bi (t, µ), which are
the reduced velocity and pressure solutions, respectively. Indeed, in principle, g varies over
time, responding to variations of a and b. This reflects at the reduced order level the fact that
the turbulent viscosity in every EVM depends on the mean flow field variables. At this point,
Nνt
one has to find a suitable way to compute the eddy viscosity reduced solution [gi (t, µ)]i=1 .
Since the specific turbulence equations of each EVM will not be used, the approach chosen to
compute the reduced eddy viscosity coefficients is based on data-driven methods and in particular
interpolation with RBFs [363, 396]. In detail, the reduced turbulent model treatment starts by
computing the eddy viscosity modes by making use of the method of snapshots, as explained in
Subsection 1.2.3. First, we define the eddy viscosity snapshot matrix
h
Sνt = {νt (x, t1 ; µ1 ), . . . , νt (x, tNT ; µM )} ∈ RNνt ×Ns , (8.36)
where Sνi t is the ith column of the eddy viscosity snapshot matrix Sνt . Then the eddy viscosity
correlation matrix is computed as
(C νt )ij = Sνi t , Sνjt L2 (Ω) . (8.37)

The eddy viscosity modes are then expressed as

Ns
1 X
ηi = S j V νt , (8.38)
Ns λνi t j=1 νt ij

where the matrix λνt contains in its diagonal the eigenvalues of the matrix C νt and V νt is the
matrix whose columns are the corresponding eigenvectors of C νt .
The next step consists of training the RBF using the data of the snapshots acquired. Later,
Nνt
during the online stage, the coefficients [gi (t, µ)]i=1 are interpolated.
We will extend the SUP-ROM presented in Chapter 7 for the turbulent case with the de-
composition assumption (8.35). This means that the momentum equation is projected onto the
modes of the velocity and the continuity equation is projected onto the modes of pressure, with
the use of the supremizer stabilization approach. These projections will result in the following
differential-algebraic system of equations (DAEs):
M ȧ = ν(B + BT )a − aT Ca + g T (CT 1 + CT 2 )a − Hb,

(8.39)
P a = 0,
where g is the vector containing the reduced order degrees of freedom of the eddy viscosity and
the new terms in the DAEs above are defined as
(CT 1 )ijk = (φi , ηj ∇ · ∇φk )L2 (Ω) ,
(CT 2 )ijk = φi , ∇ · ηj (∇φTk ) L2 (Ω) .


 8.4. Hybrid Projection-Based/Data-Driven ROM for Turbulent Flows 173

The system (8.39) has Nu + Np equations, while the number of unknowns is Nu + Np + Nνt .
This problem is solved, as mentioned, by employing an interpolation technique to obtain the
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

vector g. Thus for the vector g we are using a data-driven approach known in the literature as
POD-I [608, 600, 519]. We remark that the system in (8.39) will have additional penalty terms
if the penalty method for the treatment of the nonhomogeneous boundary conditions is used.
The ROM put forward makes use of the projected momentum and continuity equations to
obtain the reduced solution of the velocity and the pressure. On the other hand, it employs data-
driven techniques to approximate the eddy viscosity solution manifold. These kinds of ROMs are
called hybrid ROMs because they merge the classical projection-based methods with data-driven
techniques. It is common to label the projection-based methods with the term intrusive, while
the term nonintrusive is used to describe methods which employ only the data without using the
governing equations. In this contribution we will refer to the ROM given by the reduced DAE in
(8.39) as H-SUP-ROM.
Data-driven methods are helpful in approximating the maps between various terms and quan-
tities which appear in the full-order model (FOM) formulation of the NSE. In this work, the
data-driven method selected is, as mentioned, interpolation using RBF. This method is used to
Nνt
approximate the coefficients [gi (t, µ)]i=1 in (8.39). In order to do this, one has to use the data
available for the eddy viscosity. This task can be done by properly choosing the maps that need
to be approximated or, in other words, by choosing a suitable independent variable of the RBF
interpolation. In the upcoming subsections, we present two different methodologies for carrying
out the interpolation step with their corresponding hybrid ROM.

8.4.1 Hybrid ROM with RBF Interpolation on the Time-Parameter

Values
In this subsection, we assume that the independent variable of the RBF interpolation is the vector
that merges the time and the parameter values. In order to render the methodology clear, we set
up the following conventions and notation. Let Xµ,t be the set given by

Xµ,t = PM × {t1 , t2 , . . . , tNT }, (8.40)

where Xµ,t is the Cartesian product of the set containing the time instants at which snapshots
were acquired and the discretized parameter set. The set Xµ,t has a cardinality Ns , and its ith
member is denoted by xiµ,t . It can be seen that there is a unique correspondence between the
elements of Xµ,t and the columns of the matrix of snapshots (for all fields) used in the offline
stage to construct the reduced basis (RB).
The parameter sample introduced in the ROM at the online stage is denoted by µ∗ . In order
to have an accurate ROM result, the value of µ∗ should be close enough to the values of the
parameter samples used in the offline phase. Let t∗ be the time instant at which the ROM solution
is sought, where t1 6 t∗ 6 tNT . We refer to z ∗ = (t∗ , µ∗ ) as the online time-parameter
combined vector.
Let gr,l be the offline L2 projection coefficient vector which results from the projection of
the rth eddy viscosity snapshot Sνrt onto the lth eddy viscosity mode ηl , as follows:

gr,l = (Sνrt , ηl )L2 (Ω) for r = 1, 2, . . . , Ns and l = 1, 2, . . . , Nνt . (8.41)

At this point, one may define the interpolation problem as follows. Given the set Xµ,t , the cor-
Ns ,Nνt
responding eddy viscosity snapshots [Sνi t ]Ni=1 , and the corresponding coefficients [gr,l ]r=1,l=1 ,
s

approximate the vector g in (8.39) for the online time-parameter vector z . In this case one
∗
Nνt
extrapolates the scalar coefficients of the reduced eddy viscosity expansion [gi (t∗ , µ∗ )]i=1 (de-
 174 Chapter 8. Finite Volume–Based Reduced Order Models for Turbulent Flows

N
noted briefly by [gi (z ∗ )]i=1
νt
), which means that one has to perform one interpolation for each of
the Nνt modes used in the online stage.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

It is worth remarking that the procedure described implies that each of the mode’s coefficients
are interpolated independently from the others. Therefore, we may fix the eddy viscosity mode
to be the Lth one, ηL , where L is an arbitrary integer index, 1 6 L 6 Nνt . The goal of the
interpolation is to approximate the map [GL ] in gL = GL (t, µ):
GL : Rq+1 → R for any L = 1, 2, . . . , Nνt . (8.42)
In order to approximate this map, firstly, we form the set of observations which consists of the
coefficients of the L2 projections of all the offline snapshots onto the eddy viscosity mode ηL ,
namely YL = [gr,L ]N r=1 ∈ R
s Ns
. The goal is to approximate the value of gL (z ∗ ) = gL (t∗ , µ∗ ).
Interpolation using RBF is based on the formula
Ns  
wL,j ζL,j kz − xjµ,t kL2 (Rq+1 )
X
GL (z) = for L = 1, 2, . . . , Nνt , (8.43)
j=1

where ζL,j for j = 1, . . . , Ns are the RBF functions; z = (t, µ), where t refers to any time
instant inside the snapshot window and µ is a parameter value which lies in the parameter space
P; and wL,j are the weights of the RBFs which have to be computed during the training stage.
In order to compute the weights, we will use the data obtained by the FOM:
GL (xiµ,t ) = gi,L , for i = 1, 2, . . . , Ns , (8.44)
which implies that
Ns  
wL,j ζL,j kxiµ,t − xjµ,t kL2 (Rq+1 ) = gi,L
X
for i = 1, 2, . . . , Ns . (8.45)
j=1

It is possible to rewrite the last equation as a linear system:

AζL wL = YL , (8.46)

where (AζL )ij = ζL,j (kxiµ,t − xjµ,t kL2 (Rq+1 ) ). The linear system is then solved to obtain the
weights wL , which are stored in order to use them during the online stage. In this work, the
linear system above is solved using a QR decomposition with column pivoting.
Nνt
In the online stage, the value of g(z ∗ ) = [gi (z ∗ )]i=1 may be computed as
Ns  
wi,j ζi,j kz ∗ − xjµ,t kL2 (Rq+1 )
X
∗ ∗
gi (z ) ≈ Gi (z ) = for i = 1, 2, . . . , Nνt . (8.47)
j=1

The strategy of the RBF interpolation in this subsection has the deficiency of not being accurate
for values of time t∗ which lie outside the snapshot window [0, T ] (the RBF interpolant is trained
with time values which lie inside the aforementioned range, and, therefore, it is not logical to
expect that it could accurately approximate values of t∗ that do not fall in [0, T ]). This could
be considered as a major issue in several applications in ROM since extrapolation is generally
regarded as a more complex task than interpolation. In addition, the strategy of treating time as a
part of the interpolation independent variable might not make physical sense. In fact, in several
applications the flow could be periodic, which could make the absolute value of the time variable
irrelevant to the dynamics of the fluid fields. For these reasons, we have tried to circumvent this
problem by adopting a different methodology for interpolating the RBF, as explained in the next
subsection.
 8.4. Hybrid Projection-Based/Data-Driven ROM for Turbulent Flows 175

8.4.2 Hybrid ROM with RBF Interpolation on the Velocity Projection

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Coefficient Values
The RBF interpolation methodology discussed in the last subsection is quite simple and straight-
forward. However, this methodology only works with steady and unsteady flows, without being
able to extrapolate over time (outside the offline time snapshot window [0, T ]). H-SUP-ROM
with the methodology presented in the last subsection is introduced in [284], where it is applied
to a steady problem. This subsection suggests a different way of carrying out the RBF inter-
polation, in which the velocity projection coefficients (or a form of them as we will see) are to
be the independent variable of the interpolation. This choice will be justified in the following
discussion. The approach carried out in this subsection is presented in [286].
Before addressing the choice of independent variable for the interpolation, it is important to
introduce a modification to the way the eddy viscosity field was decomposed at the reduced order
level in (8.35). Primarily, we suggest spliting the FOM eddy viscosity field as follows:

νt (x, t; µ) = νt (x; µ) + νt0 (x, t; µ), (8.48)

where basically we have divided the eddy viscosity into two contributions. The first term corre-
sponds to the time-averaged viscosity field for the parameter µ, while the second term represents
the time-varying part. The rationale for this proposal is based on the fact that the parameter
variation is largely seen in the mean part νt (x; µ). On the other hand, parameter changes do
not affect to a great extent the small oscillations in the field νt0 (x, t; µ). In fact, we have noticed
that, in the numerical examples discussed in this work, we may drop the dependency on the pa-
rameter for the time-varying field, i.e., νt0 (x, t; µ) ≈ νt0 (x, t). This decomposition will turn out
to be beneficial for approximating the reduced order eddy viscosity. Splitting the eddy viscosity
into two terms, one which is time dependent and the other which is parameter dependent, makes
approximating νt an easier task.
The reduced order approximation of the νt is now modified as follows:

M Nνt
X X
νt (x, t; µ) ≈ νtr (x, t; µ) = g i (µ)η i (x) + gi (t)ηi (x), (8.49)
i=1 i=1

where the averaged part νt (x; µ) is approximated by the first sum in the above decomposition,
and the second sum approximates the time-varying part νt0 (x, t). In this formulation, one may
look at the fields [η i ]M
i=1 as additional eddy viscosity modes. The number of these fields is M ,
which is the number of parameter samples used in the offline stage (or the dimension of the dis-
cretized parameter set PM ). Each of these fields is computed by taking the time average over the
offline snapshots which correspond to only one of the parameter samples inside PM . In other
words, the field η i is computed by taking the average of the eddy viscosity fields acquired at dif-
ferent time instants but corresponding to the ith parameter in PM . As for [g i ]M i=1 , they represent
the degrees of freedom corresponding to the averaged fields and are important for periodic regime
problems. These are parameter dependent, and if we are testing the ROM in the online stage for
the ith parameter inside the offline parameter set (i.e., µ∗ = µi ), which is a reproduction test
in the parameter space, then the vector g = eM i , where ei is the unit vector of dimension M
M

consisting of zeros except for the its ith component, which it has the value one. The value of g
for cross-validation tests (extrapolation in the parameter space) can be fixed by linear interpola-
tion; e.g., if µ∗ lies at the distances d∗1 and d∗2 in the parameter space from its closest two offline
parameter samples µ1 and µ2 , respectively (i.e., kµ∗ − µ1 kL2 (Rq ) = d∗1 , kµ∗ − µ2 kL2 (Rq ) =
d∗2 , d∗1 < kµ∗ − µi kL2 (Rq ) and d∗2 < kµ∗ − µi kL2 (Rq ) for all i = 3, 4, . . . , M ), then
d∗ d∗
g can be approximated as d∗
2 M
∗ e1 + d∗
1
∗ e2 .
M
Finally, the dynamical system of the H-SUP-
1 +d2 1 +d2
 176 Chapter 8. Finite Volume–Based Reduced Order Models for Turbulent Flows

ROM is modified as follows:

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

M ȧ = ν(B + BT )a − aT Ca + g T (CT 1 + CT 2 )a + g T (CT 1 + CT 2 )a − Hb,



P a = 0,
(8.50)
where the two new tensors are defined as


(CT 1 )ijk = φi , η j ∇ · ∇φk L2 (Ω) ,
(CT 2 )ijk = φi , ∇ · η j (∇φTk ) L2 (Ω) .

At this point, we may come back to tackle the main issue in this subsection, which is how we
could choose the independent variable of the interpolation in such a way that it leads to a hybrid
ROM which we could extrapolate over time. The suggested solution is to take the velocity
L2 projection coefficients as the independent variable of the RBF interpolation. This choice is
inspired by the fact that the eddy viscosity ultimately depends on the mean velocity field and its
gradient. The spirit of this choice is established on the fact that the eddy viscosity νt is a function
of the time history of the velocity field u. In other words, if we denote the FOM eddy viscosity
and the velocity at time tn by νtn and un , respectively, then we may express the eddy viscosity as
νtn = νt (u1 , u2 , . . . , un ). In order to mimic the dependency between the velocity field and the
eddy viscosity field at the reduced order level, we have to take into account the time evolution
aspect of the last equation. This is done by assuming that the eddy viscosity reduced vector g is
a function of the velocity coefficients a, along with their time derivative ȧ, i.e.,

g n = g n (tn ) ≈ g n (an , ȧn ). (8.51)

Formula (8.51) represents an approximation of the relationship that exists between the eddy
viscosity and velocity fields. This approximation has been successfully used in the numerical
examples that we have considered in this work. However, the eddy viscosity evolves through a
PDE, albeit in many cases not directly, but passing through an algebraic relationship with turbu-
lence variables, which in turn evolve through PDEs. Thus, it is possible that the approximation
(8.51) could eventually be improved by including other fluid dynamics variables in the turbulence
model PDEs.
We proceed now to the training phase of the RBF interpolation. This phase is conducted
in the ROM offline stage. First, the L2 projection coefficients of the velocity modes onto the
snapshots are computed. Then, those corresponding to the supremizer are dropped, and the time
derivatives of the projection coefficients are calculated using a backward Euler scheme. The
velocity projection coefficients with their corresponding derivatives are put together to train the
RBF. In more detail, the procedure starts from the solution snapshots, namely
     
Sµ1 ,u Sµ1 ,p Sµ1 ,νt
 Sµ2 ,u   Sµ2 ,p   Sµ2 ,νt 
Su =  ..  , Sp =  ..  , Sν t =  .. , (8.52)
     
 .   .   . 
SµM ,u SµM ,p SµM ,νt

where the snapshot matrices for all the variables have been expressed as M vertically aligned
submatrices, with each submatrix containing the time snapshots corresponding to a single pa-
rameter sample. Then one may define the L2 velocity projection coefficients arµk ,L2 ∈ RNu :

arµk ,L2 = [(Sµr k ,u , φ1 )L2 (Ω) , . . . , (Sµr k ,u , φNu )L2 (Ω) ] (8.53)
for r = 1, 2, . . . , NT , k = 1, 2, . . . , M.
 8.4. Hybrid Projection-Based/Data-Driven ROM for Turbulent Flows 177

Let
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

a1µk ,L2 a2µk ,L2

   
 a2µk ,L2   a3µk ,L2 
A1,k =  ..  ∈ R(NT −1)×Nu , A2,k =  ..  ∈ R(NT −1)×Nu . (8.54)
   
 .   . 
aN T −1
µk ,L2 aN T
µk ,L2

At this point, we can compute the time derivative vectors needed for the RBF interpolation by
simply using the backward differentiation scheme as follows:

arµk ,L2 − ar−1

µk ,L2
ȧrµk ,L2 = for r = 2, 3, . . . , NT , k = 1, 2, . . . , M, (8.55)
∆tµk
where ∆tµk is the time step at which snapshots were acquired for the parameter sample µk . This
yields the following matrix of time derivative velocity coefficients:
 2
ȧµk ,L2

3
 ȧµ ,L2 
A2,k − A1,k k
Ȧk = = ..  ∈ R(NT −1)×Nu . (8.56)
 
∆tµk  . 
ȧNT
µk ,L2

In order to build the matrix containing all the data observations needed to train the RBF, one has
to merge the L2 projection coefficients of velocity, starting from the second time snapshot, with
the time derivative coefficients. This gives the following matrix:

∈ R(NT −1)×2Nu . (8.57)

 
Ãk = A2,k Ȧk

As for the eddy viscosity, one has to compute the projection coefficients of the eddy viscosity
modes onto the snapshots corresponding to the time-varying part in (8.48). In other words, the
time-averaged part corresponding to the kth parameter in the training set PM , which is νt (x; µk ),
has to be subtracted from each eddy viscosity snapshot, as follows:
r
gµ k ,i,L
r
2 = (Sµ ,ν − νt (x; µk ), ηi )L2 (Ω)
k t
for r = 2, 3, . . . , NT , (8.58)
i = 1, 2, . . . , Nνt , and k = 1, 2, . . . , M.

Let G̃i,k ∈ R(NT −1) be the vector containing the coefficients in (8.58) for fixed i and k. The
combined matrices and vectors for all parameter samples are denoted by à and G̃i , respectively,
and are defined as follows:
   
Ã1 G̃i,1
 Ã2   G̃i,2 
à =  .  ∈ R(Ns −M )×2Nu , G̃i =  ..  ∈ R(Ns −M ) . (8.59)
   
 . .  . 
ÃM G̃i,M
N
The maps to be interpolated are [Gi ]i=1
νt
in gi = Gi (a, ȧ), where

Gi : R2Nu → R for i = 1, 2, . . . , Nνt . (8.60)

This approximation is based on the interpolation points given in each row of the matrix à and
the vector G̃i . The interpolation procedure relates the mth row of Ã, which is a vector called
ãm
L2 ∈ R
2Nu
, to the mth element of G̃i , denoted by [G̃i ]m ∈ R.
 178 Chapter 8. Finite Volume–Based Reduced Order Models for Turbulent Flows

In this formulation, the interpolation equation is

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

NX
s −M  
Gi (ãonline ) = wi,j ζi,j kãonline − ãjL2 kR2Nu for i = 1, 2, . . . , Nνt , (8.61)
j=1

where ãonline := [aonline , ȧonline ] ∈ R2Nu is a combination of the reduced velocity vector and its
vector derivative. The FOM data is used to establish the relation

Gi (ãm
L2 ) = [G̃i ]m for m = 1, 2, . . . , Ns − M, (8.62)

NX
s −M  
j
wi,j ζi,j kãm
L2 − ãL2 kR2Nu = [G̃i ]m for m = 1, 2, . . . , Ns − M. (8.63)
j=1

N
In the online stage one can approximate g(a∗ , ȧ∗ ) = [gi (a∗ , ȧ∗ )]i=1
νt
:

NX
s −M  
gi ((a∗ , ȧ∗ )) ≈ Gi ([(a∗ , ȧ∗ )]) = wi,j ζi,j k[(a∗ , ȧ∗ )] − ãjL2 kR2Nu . (8.64)
j=1

We would like to remark that the H-SUP-ROM in its different versions discussed in this section
is compatible with both the penalty method and the lifting function method. The interpolation
discussed in this subsection could be done with the velocity coefficients coming from the lifting
Nνt
modes when the lifting function approach is employed. If that is the case, then the maps [Gi ]i=1
in gi = Gi (a, ȧ) are

Gi : R2NBC +2Nu → R for i = 1, 2, . . . , Nνt . (8.65)

It is also important to underline that the homogenized set of velocity snapshots Sũ in Chapter 7
has to replace Su in (8.52) if the lifting function method is used.
The H-SUP-ROM in (8.50) can be solved by time-integrating the dynamical system in the
case of unsteady flows, or by simply solving the algebraic system in the case of steady flows.
In this work, we used a Newton method to solve the reduced dynamical system. The Newton
method computes the Jacobian numerically. In the general unsteady case, it is recommended to
choose a time step which is consistent with the one used during the FOM simulations. The time
advancement scheme for computing the derivative of the reduced vector a is chosen as a first- or
second-order backward Euler scheme.

8.5 Turbulent ROMs Based on the Uniform-ROM and the

PPE-ROM
In the last section, we introduced the H-SUP-ROM proposed for the reduction of turbulent flows.
The H-SUP-ROM is based on the projection of the momentum and continuity equations with
the supremizer approach to obtain the reduced velocity and pressure degrees of freedom, while
it employs a data-driven strategy for closing the system by approximating the reduced eddy
viscosity coefficients. In this section, we address other relevant ROM formulations which may
reduce turbulent flows under certain specific conditions.
The first turbulent ROM addressed here is an extension of the ROM addressed in Subsection
7.2.3, which we refer to as the uniform-ROM or U-ROM. In this case, the extended ROM in-
volves approximating the reduced eddy viscosity field using the velocity degrees of freedom, as
 8.5. Turbulent ROMs Based on the Uniform-ROM and the PPE-ROM 179

was done with the pressure field. In other words, the reduced eddy viscosity is given by
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Nr
X
νt (x, t; µ) ≈ νtr (x, t; µ) = ai (t, µ)ηi (x), (8.66)
i=1

where the eddy viscosity modes [ηi (x)]N

i=1 in this case are computed using the singular value
r

decomposition (SVD) of the velocity correlation matrix:

Ns
1 X
ηi = Sνjt Viju . (8.67)
Ns λu
i j=1

The corresponding DAE for this turbulent ROM, which is based only on the projection of the
momentum equation, is
ȧ = ν(B + BT )a − aT (C − CT 1 − CT 2 )a − Ha, (8.68)
where the vector a represents the degrees of freedom of the velocity, pressure, and eddy viscosity.
This ROM can be used to reduce both steady and unsteady flows. It also has the advantage of
being of low computational online cost, since the system (8.68) is relatively small. However, as
mentioned earlier, this formulation undermines the reduced pressure approximation. In addition,
numerical evidence has demonstrated that this ROM is prone to long-term instabilities when the
system (8.68) is integrated over long times (see Subsection 8.6.2). Furthermore, the U-ROM
is limited by not being compatible with the lifting function method described in Subsection
7.2.5. The lifting function method requires a homogenization process for the velocity snapshots.
The proposal of having a unique set of reduced coefficients [ai (t, µ)]N i=1 is directly affected by
r

the homogenization process of the velocity snapshots. The process of modifying the velocity
snapshots suggests that this process should be followed for the pressure and the eddy viscosity
snapshots, which is not well defined. In the next chapter, we compare the turbulent U-ROM
and the H-SUP-ROM with the penalty method used to enforce the nonhomogeneous boundary
conditions.
At this point, one may proceed to the extension of the PPE-ROM presented in Subsection
7.2.3. There are two different ways of extending this ROM. The first is based on assuming
that the reduced eddy viscosity field has the same temporal coefficients as the velocity field, as
proposed in (8.67). In this formulation the velocity and the eddy viscosity fields will have the
vector a as their reduced vector, while the pressure field will still be computed using a different
set of coefficients represented by the vector b. This ROM is called the semiuniform-PPE-ROM
or SU-PPE-ROM. We recall the reduced approximations of this ROM:
Nu Np
X X
u(x, t; µ) ≈ ai (t, µ)φi (x), p(x, t; µ) ≈ bi (t; µ)χi (x), (8.69)
i=1 i=1
Nu
X
νt (x, t; µ) ≈ ai (t, µ)ηi (x), (8.70)
i=1

with the POD modes computed as follows:

Ns Ns
1 X 1 X
φi = uj Viju , χi = pj Vijp , (8.71)
Ns λ u
i j=1 Ns λpi j=1
Ns
1 X
ηi = νtj Viju . (8.72)
Ns λ u
i j=1
 180 Chapter 8. Finite Volume–Based Reduced Order Models for Turbulent Flows

The FOM momentum and Poisson equations for the RANS turbulent modeling are as follows:
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

 h  i
T



∂u
∂t + ∇ · (u ⊗ u) = ∇ · −pI + (ν + νt ) ∇u + (∇u) in Ω × [0, T ],
 h   i
T
in Ω, (8.73)

∆p = −∇ · (∇ · (u ⊗ u)) + ∇ · ∇ · νt ∇u + (∇u)
+ Boundary conditions on Γ × [0, T ],



+ Initial conditions in (Ω, 0).



The Galerkin projections of the momentum and the Poisson equations give the following DAE:

ȧ = ν(B + BT )a − aT (C − CT 1 − CT 2 )a − Hb,

(8.74)
Db + aT (G − CT 3 − CT 4 )a − νN a − L = 0.
The additional tensors appearing in the reduced Poisson equation come from the divergence of
the eddy viscosity term in the FOM momentum equation in (8.18). Their entries are defined as

(CT 3 )ijk = (∇χi , ηj ∇ · ∇φk )L2 (Ω) ,

(CT 4 )ijk = ∇χi , ∇ · ηj (∇φTk ) L2 (Ω) .

The SU-PPE-ROM cannot be used to reduce steady flows. This is related to the boundary con-
ditions, which have to be satisfied at the full-order level. We mentioned this issue in Subsection
7.2.3, where it is related to the way the PPE is derived from the momentum and continuity equa-
tions of the NSEs or the RANS. The derivation of an equivalent system to the one formed by the
NSEs requires an additional boundary condition for the divergence of the velocity or the pressure
[313, 376]. In this work, the PPE formulation adopted involves an additional Neumann bound-
ary condition for the pressure field. This latter condition is needed in order to make sure that the
velocity field is divergence free for all time instants. This condition can be fulfilled only in the
general unsteady setting (see the remark in Section 2 in [313]).

8.5.1 Hybrid ROM Based on the PPE-ROM

In this subsection, we address the second choice for the extension of the PPE-ROM for the
goal of reducing turbulent flows. The second methodology proposes to employ the reduced
approximation of the eddy viscosity fields of the one used in the case of the H-SUP-ROM. That
means decomposing νt as done in (8.35) or (8.49), where this implies that the coefficients of
the reduced eddy viscosity will be computed by the interpolation strategy explained in the last
section. This ROM will be called H-PPE-ROM. Its DAE reads as follows:
ȧ = ν(B + BT )a − aT Ca + g T (CT 1 + CT 2 )a + g T (CT 1 + CT 2 )a − Hb,

(8.75)
Db + aT Ga − g T (CT 3 + CT 4 )a − g T (CT 3 + CT 4 )a − νN a − L = 0.

The additional tensors CT 3 and CT 4 account for the parameter contribution of the eddy viscosity
in the Poisson equation and are calculated as


(CT 3 )ijk = ∇χi , η j ∇ · ∇φk L2 (Ω) ,
(CT 4 )ijk = ∇χi , ∇ · η j (∇φTk ) L2 (Ω) .

Like the case of the PPE-ROM, its hybrid turbulent extension works only for nonstationary
flows. The H-PPE-ROM has an advantage over the H-SUP-ROM in having fewer unknowns in
the dynamical system because the supremizer reduced degrees of freedom are not present in the
H-PPE-ROM formulation. This often results in a slight advantage in terms of speedup during the
online stage.
 8.6. Application of the H-SUP-ROM to Turbulent Problems 181

8.6 Application of the H-SUP-ROM to Turbulent Problems

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

In this section, we present the results obtained by applying the proposed POD-Galerkin H-SUP-
ROM to two test cases that are very common benchmark problems in the development of tur-
bulent flow simulation tools. The first problem is that of turbulent flow past a backward-facing
step. This classical benchmark in the turbulence modeling community is used here to evaluate
the performance of the proposed ROM when applied to parametrized problems based on steady
RANS equations. The second problem is that of turbulent flow past a circular cylinder. This case
is instead able to characterize the POD-Galerkin H-SUP-ROM behavior in the presence of un-
steady turbulent flows, which also depend on physical parameters—the Reynolds number in this
case. For both test cases we will compare the H-SUP-ROM results with the ones obtained by the
ROM developed in [380]. In that work, the authors proposed a ROM which considers a single set
of reduced coefficients for the velocity, pressure, and eddy viscosity field. The main advantage
of this methodology is that it allows for a unified approach to dealing with laminar or turbulent
flows, with a consequently simple implementation. From now on, in all the comparisons of the
results of the methodology proposed here, the ROM developed in [380] will be referred to as
U-ROM.
The finite volume C++ library OpenFOAM [612] has been used to solve all the full-order
problems involved in the numerical tests presented in this section. As for the ROM simulations,
the reduced order problems shown have been reduced and resolved using the C++-based library
ITHACA-FV [549].

8.6.1 Steady Case

Figure 8.1 shows the computational domain layout of the backward step benchmark case used
to test the steady H-SUP-ROM solver. The image also depicts the computational mesh used for
the FOM simulation, along with a summary of the boundary conditions imposed on every side
of the domain. Particular attention is paid to the condition prescribed on the inflow boundary,
where the constant Dirichlet datum U for the velocity field is varied to parametrize the problem
with respect to the flow Reynolds number. Thus, the objective of this numerical experiment is to
assess the H-SUP-ROM solver’s ability to reproduce flows with high Reynolds number and their
dependence on the parameters. To this end, the H-SUP-ROM results will be compared both to
the full-order results and to the results of the U-ROM model. Moreover, to test the H-SUP-ROM
solver’s ability to deal with different turbulence models, we trained and tested the model, using
full-order solutions obtained with both k- and SST k-ω models.
The 100 snapshots required for the training during the offline phase were generated by solv-
ing the full-order problem with inlet velocity values ranging from 1 m/s to 25 m/s on an
equally spaced distribution. Given the physical viscosity ν = 10−3 m2 /s and the character-
istic length chosen as D = 1 m, this corresponds to Reynolds number variations in the range

Figure 8.1. The computational domain used in the numerical simulations; all lengths are de-
scribed in terms of the characteristic length D, which is equal to 1 m.
 182 Chapter 8. Finite Volume–Based Reduced Order Models for Turbulent Flows

[1 × 103 , 2.5 × 104 ]. As for the numerical setup of the full-order simulations, the Gauss linear
scheme was selected to approximate the gradients, and the Gauss linear scheme with nonorthog-
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

onal correction was selected to approximate the Laplacian terms. A second-order bounded Gauss
upwind scheme was instead used to approximate the convective term. Finally, a first-order
bounded Gauss upwind scheme is used to approximate all terms involving the turbulence model
parameters k, , and ω.
The modes of the velocity, pressure, and eddy viscosity fields were obtained by POD analysis
of the respective snapshot matrices. Figure 8.2 shows the cumulative eigenvalue decay for the
velocity, pressure, and eddy viscosity fields. As can be appreciated, a relatively small number of
modes is sufficient to recover most of the energy information contained in the snapshots.
Once the reduced model was trained and the POD modes were obtained, a set of H-SUP-
ROM and U-ROM simulations were executed on a set of sampling points in the parameter space,
which was not contained in the original set used for training. More specifically, the online sam-
ple values for U , denoted by Ui∗ , where i = 1, . . . , Nonline-samples , were chosen as 80 equally
distributed samples in the interval [3, 20]. This set of samples is selected to include both sam-
ples close to the offline training ones and others which lie almost midway between two offline
samples. Clearly, the test is aimed at assessing how accurate the reduced approximation is for
parameter values that were not in the training set.
The correct inflow velocity in the reduced simulations is enforced by means of the penalty
method, as in Subsection 7.2.5. In this regard, we must remark here that the simulation results
appeared quite sensitive to the penalization factor τ . A sensitivity analysis was performed to se-
lect a value of τ that would allow satisfactory results for both the k- and the SST k-ω turbulence
models considered.
The first step of the online stage is to interpolate the eddy viscosity coefficients with respect
to the values of the inflow velocity parameter considered. More specifically, the result of the
interpolation is the reduced eddy viscosity vector g, which is then introduced in the reduced
system (8.39) to close the mathematical problem and finally obtain the vectors of coefficients a
and b. The RBF interpolation methodology used in this work to obtain g based on the parameter
value is implemented in the C++ library SPLINTER [247].
Figure 8.3 depicts the velocity fields corresponding to the velocity value U ∗ = 7.0886 m/s
computed via the FOM, the U-ROM, and the H-SUP-ROM for the k- turbulence model. A
similar comparison is presented in Figure 8.4 for the pressure field and in Figure 8.5 for the

Figure 8.2. Cumulative ignored eigenvalue decay. In the plot, the solid black line refers to the
velocity eigenvalues, the dashed magenta line indicates the pressure eigenvalues, and the dash-dotted blue
line refers to the eddy viscosity eigenvalues.
 8.6. Application of the H-SUP-ROM to Turbulent Problems 183
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

(a) (b)

(c)

Figure 8.3. k- turbulence model case, velocity fields for the value of the parameter U = 7.0886
m/s: (a) shows the FOM velocity, while in (b) one can see the U-ROM velocity, and finally in (c) we have
the H-SUP-ROM velocity.

eddy viscosity field. At the online stage level, all the solutions presented in the aforementioned
figures were generated using 10 modes to approximate the velocity, pressure, supremizer, and
eddy viscosity fields, for both the H-SUP-ROM and the U-ROM. The images indicate that the
hybrid projection/data-driven-based approach allows for qualitatively accurate approximations
of the FOM solutions for velocity, pressure, and eddy viscosity fields. However, this is not
completely the case when the U-ROM approach is employed. In fact, despite being able to obtain
a qualitatively accurate result for the velocity and eddy viscosity fields, the U-ROM computes a
pressure field that does not correctly reproduce its FOM counterpart. To quantitatively measure
the performance of both ROMs, we evaluate the relative L2 error for the velocity and pressure
fields, which, respectively, read

ku − u∗ kL2 (Ω) kp − p∗ kL2 (Ω)

u = × 100%, p = × 100%, (8.76)
kukL2 (Ω) kpkL2 (Ω)

in which u∗ and p∗ are general reduced order velocity and pressure fields, respectively. The
relative L2 errors between the FOM and the H-SUP-ROM velocity and pressure fields presented

(a) (b)

(c)

Figure 8.4. k- turbulence model case, pressure fields for the value of the parameter U = 7.0886
m/s: (a) shows the FOM pressure, while in (b) one can see the U-ROM pressure, and finally in (c) we have
the H-SUP-ROM pressure.
 184 Chapter 8. Finite Volume–Based Reduced Order Models for Turbulent Flows
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

(a) (b)

(c)

Figure 8.5. k- turbulence model case, eddy viscosity fields: (a) shows the FOM eddy viscosity,
while in (b) one can see the U-ROM eddy viscosity, and finally in (c) we have the H-SUP-ROM eddy
viscosity.

in Figures 8.3 and 8.4 are respectively u = 0.4444% and p = 0.3654%, confirming the positive
result of the qualitative assessment. As for the U-ROM results, the corresponding errors are
u = 0.6522% and p = 20.9441%, respectively. The results of this test case suggest that the
H-SUP-ROM is able to overcome the main limit of the U-ROM approach, which is represented
by its inability to obtain accurate pressure field approximations.
A further simulation campaign was carried out with a different SST k-ω turbulence model
to evaluate how responsive the results of the hybrid H-SUP-ROM and of the U-ROM are with
respect to the turbulence model employed for the FOM simulations. Thus, a new set of SST
k-ω FOM simulations was executed. The inflow velocity parameter values used for this simu-
lation campaign were the same as in the k- model case previously described. The snapshots
generated were once again used for the modal decomposition–based training of both reduced
models considered. Figures 8.6, 8.7, and 8.8 show the velocity, pressure, and eddy viscosity
fields obtained by the FOM, the U-ROM, and the H-SUP-ROM for the inflow velocity value
U ∗ = 7.0886, respectively. Again, a visual inspection of the plots suggest that the H-SUP-ROM
velocity, pressure, and eddy viscosity fields appear in good qualitative agreement with their SST

(a) (b)

(c)

Figure 8.6. SST k-ω turbulence model case, velocity fields for the value of the parameter U =
7.0886 m/s: (a) shows the FOM velocity, while in (b) one can see the U-ROM velocity, and finally in (c)
we have the H-SUP-ROM velocity.
 8.6. Application of the H-SUP-ROM to Turbulent Problems 185
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

(a) (b)

(c)

Figure 8.7. SST k-ω turbulence model case, pressure fields for the value of the parameter U =
7.0886 m/s: (a) shows the FOM pressure, while in (b) one can see the U-ROM pressure, and finally in (c)
we have the H-SUP-ROM pressure.

k-ω FOM counterparts. As for the U-ROM results, although the velocity and, to some extent,
the eddy viscosity, field appear close to the k-ω FOM result, the U-ROM pressure field is visibly
different from the FOM one. From a quantitative standpoint, the L2 relative errors between the
FOM and the H-SUP-ROM velocity and pressure fields are, respectively, u = 0.8088% and
p = 0.7329%. As for the U-ROM results, the corresponding errors are u = 0.8177% and
p = 22.3972%, respectively. These results once again confirm that the H-SUP-ROM is able to
improve the pressure predictions obtained with the U-ROM approach.
The fact that the same test was run with both the k- and the SST k-ω models offers the
opportunity to measure how sensitive the H-SUP-ROM results are to the turbulence models em-
ployed at the offline FOM stage. Since the most appreciable differences between the k- and
k-ω solutions occur in the pressure fields, in Figure 8.9 we present the pressure field values on a
horizontal line cutting the domain at a vertical coordinate x2 = 5D6 located at half the maximum
height. The curves in the plot, which represent the pressure predicted by each model on the afore-
mentioned line as a function of the horizontal coordinate x1 , show a clear difference between the
k- and k-ω full-order model results. Quite remarkably, the curves referring to the H-SUP-ROM

(a) (b)

(c)

Figure 8.8. SST k-ω turbulence model case, eddy viscosity fields: (a) shows the FOM eddy
viscosity, while in (b) one can see the U-ROM eddy viscosity, and finally in (c) we have the H-SUP-ROM
eddy viscosity.
 186 Chapter 8. Finite Volume–Based Reduced Order Models for Turbulent Flows
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 8.9. The pressure fields obtained using both k- and SST k-ω turbulence models and the
H-SUP-ROM ones for the value of the parameter U = 7.0886 m/s. The plot is for the pressure value along
the x1 direction, keeping the value of x2 fixed at half the maximum height.

pressure prediction each fall very close to their FOM counterparts. The H-SUP-ROM trained
with the k- model (green dots) is in fact very close to the black line representing the FOM k-
result, while the H-SUP-ROM trained with the SST k-ω model (blue dots) is instead very close
to the red line representing the FOM SST k-ω prediction. Thus, the investigation presented in
the plots suggests that the H-SUP-ROM approach is able to adapt to the turbulence model used
in the FOM simulations without losing accuracy. This accomplishes one of the main goals of the
H-SUP-ROM development, which was to obtain a ROM model sensitive to the turbulence model
used at the FOM level, without involving its specific turbulence model equations in the Galerkin
projection.
Finally, the convergence analysis for the H-SUP-ROM results is shown in Figure 8.10. The
plots show the mean L2 relative error over all the 80 samples used in the cross-validation test
in the online stage as a function of the number of modes used. As previously mentioned, in
these preliminary tests we used a uniform number of modes for velocity (Nu ), pressure (Np ),
supremizer (NS ), and eddy viscosity (Nνt ). The plots indicate that for the problem considered,
the H-SUP-ROM results exhibit fast convergence to the FOM solution for both k- and SST k-ω.
Yet, after less than 10 modes, the convergence appears to stall, as the error settles on nonzero,
but fairly acceptable, values. This is likely because, as the number of modes grows, the gain in
accuracy becomes only marginal compared to the νt field interpolation error.

(a) (b)

Figure 8.10. The mean of the L2 relative errors for all the online samples versus the number of
modes used in the online stage. The convergence analysis is done for both H-SUP-ROM models obtained
with two different turbulence models at the full-order level, which are k- and SST k-ω. The errors are
reported in percentages—in (a) we have the mean error of the velocity fields, while in (b) we have the mean
error of the pressure fields.
 8.6. Application of the H-SUP-ROM to Turbulent Problems 187

8.6.2 Unsteady Case

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

The present subsection evaluates the performance of the reduction methodologies on an unsteady
turbulent test case. The classical computational fluid dynamics (CFD) benchmark considered
is that of unsteady flow past a cross-flow circular cylinder. We refer the interested reader to
[640, 641, 526] for a deep analysis of the main physical aspects of this phenomenon, which is
of particular relevance in the design of several engineering artifacts, such as electrical power
lines, gas and oil pipelines and risers, suspension bridge cables, and chimneys. Aside from their
physical relevance, vortex-shedding models are also of great interest among applied mathemati-
cians, because they represent a particular case of unsteady PDE problems with steady boundary
conditions. For these reasons, in the recent past several researchers in the model order reduction
community have considered vortex shedding past cylinders as an ideal playground for testing
reduction algorithms. Among others, we mention [548, 402, 628, 91, 253].

Problem Setup

Given the high cylinder length/diameter ratios encountered in all the aforementioned application
fields, tip effects typically have marginal relevance in this kind of flow. Thus, vortex shedding
past circular cylinders can be considered an inherently two-dimensional phenomenon. Although
there are of course notable exceptions to this assumption (see for instance [403]), we based the
validation campaign reported in the present subsection on a two-dimensional simulation setup.
A sketch of the flow domain and details of the computational grid composed of 11644 cells are
depicted in Figure 8.11. All the distances indicated in the diagram are expressed in terms of the
characteristic length of the problem, i.e., the cylinder diameter D = 1 m. The figure also reports
the boundary conditions imposed on the velocity and the pressure fields in correspondence with
all the domain edges. In all the simulations, the fluid kinematic viscosity ν is set equal to 10−4
m2 /s, and the horizontal velocity at the inlet is set to the uniform value Uin . Through variations
of this inlet velocity magnitude Uin , the test case is parametrized based on the Reynolds number
Re. In more detail, Uin was varied in the range [7.5, 12] m/s throughout the simulation cam-
paign, corresponding to Re ∈ [7.5 × 104 , 1.2 × 105 ]. Each simulation evolved over time until
a final periodic regime was fully developed, and a minimum of two periods were available for
postprocessing.
The full-order simulations were carried out using the nonsteady solver pimpleFoam imple-
mented in the framework of the OpenFOAM software library. This solver merges the SIM-
PLE and the PISO algorithms described in Subsection 7.2.2. The pimpleFoam solver in Open-
FOAM is able to adapt the time step length to ensure that the maximum Courant number CF L
[138, 139] does not exceed a prescribed threshold, which in this case has been set to CF Lmax =
0.9. A backward Euler time-advancing scheme was selected, resulting in first-order accurate
time derivatives of the velocity field. A similarly accurate first-order Gauss linear scheme was

(a) (b)

Figure 8.11. (a) The OpenFOAM mesh used in the simulations for the unsteady case of the flow
around a circular cylinder. (b) A picture of the mesh zoomed near the cylinder.
 188 Chapter 8. Finite Volume–Based Reduced Order Models for Turbulent Flows

employed for the spatial gradients. As for the convective term, the selected approximation was
a second-order bounded Gauss upwind divergence scheme which utilizes upwind interpolation
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

weights, with explicit correction based on the local cell gradient. The diffusive term was also
discretized with a Gauss linear approximation scheme. The values of the relaxation factors αu
and αp (see Subsection 7.2.2) were set at 0.7 and 0.3, respectively. At each time step, a single
nonorthogonal corrector iteration was used to deal with the nonorthogonality of the mesh. As for
the linear solvers, a Gauss–Seidel smoother solver was used for the velocity equation, whereas
the GAMG solver with Gauss–Seidel smoother was selected for the pressure equation. Finally,
turbulence treatment at the full-order level was carried out by means of the SST k-ω turbulence
model available in the OpenFOAM library.
A first necessary step of the offline phase was to identify the sample points in the parameter
space. Ten uniformly spaced values of inlet horizontal velocity Uin were then taken inside the
range [7.5, 12] m/s. In each of the resulting simulations, snapshots were collected at each time
step after making sure that the flow reached the final periodic regime. We emphasize that in
such a numerical example, the objective of the reduction is to reproduce the final periodic regime
solution for each field and other related quantities. Thus, for each inflow velocity considered
the time snapshots need to be collected wisely by covering 1.5–2 solution cycles in the periodic
regime. This is vital because it allows one to compute POD modes able to fully represent the fluid
dynamics field features across the periodic solution cycle. Thus, in order to properly choose the
time instants at which snapshots had to be acquired, it was necessary to obtain the time period
of the final periodic regime. This was done by means of Fourier analysis, computing the fast
Fourier transform (FFT) of the FOM time signal of lift and drag fluid dynamics forces acting on
the cylinder’s surface. To illustrate this procedure, Figure 8.12 depicts the time signal of the lift
coefficient (obtained from the lift force L as Cl = 1 ρUL2 D ) for the case where Uin = 10 m/s.
2
Since this time signal was computed with nonuniform time steps, an interpolation procedure
was used to obtain the CL values on the uniform time grid required by the FFT algorithm. The
FFT analysis resulted in a vortex-shedding period of 0.4299 s, which corresponds to a Strouhal
number [563] of St = fVUS D = 0.2326, where fV S is the frequency of vortex shedding. We
point out that the value obtained is relatively close to the well-assessed experimental one of
approximately 0.20 [64]. Once the time period was computed, it was possible to simulate enough
cycles of the solution and collect the correct number of snapshots for each inflow velocity tested.
In the case presented in Figure 8.12, 1.2 additional seconds were simulated with a fixed time step
of 0.0003 s, while snapshots of the fluid dynamics fields were saved every 0.006 s. This resulted
in a total of 200 snapshots collected for the parameter sample point discussed here.

Figure 8.12. The lift coefficient curve for parameter sample Uin = 10 m/s.
 8.6. Application of the H-SUP-ROM to Turbulent Problems 189

The procedure of acquiring the time step snapshots was then repeated for each parameter
sample in the training set. The number of snapshots taken for each parameter sample was NT =
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

200, which resulted in a grand total of Ns = 2000 snapshots. Table 8.1 reports the time steps
and the time intervals in which snapshots were taken for each value of the sampled parameter.
The table suggests that the time step and the time window in which snapshots were saved had
to be varied to adapt to the changes in the frequency of vortex shedding driven by modifications
of the asymptotic velocity Uin . The correct value of Uin was introduced in the offline full-order
simulations by means of a nonhomogeneous Dirichlet boundary condition at the inlet, enforced
by the penalty method [240, 539].

POD Basis Computation

Once the results of all the offline simulations were collected, the POD procedure was directly
applied to the snapshot matrices of the velocity, pressure, and eddy viscosity fields. Figure 8.13
shows the cumulative ignored eigenvalue decay of the correlation matrices of the three fluid
dynamics variables. Once the POD modes were obtained, the supremizer problem could be
solved for each pressure mode [34], resulting in the supremizer modes used to enrich the velocity
POD space.

The Eddy Viscosity Interpolation Strategy

A further step of the offline phase was to set up the interpolation procedure to approximate the
eddy viscosity coefficient vector g. As discussed in Subsections 8.4.1 and 8.4.2, the value of
g used at each time step is interpolated based on either the current time and parameter values
(t∗ , Uin
∗
) or on the reduced order velocity coefficient vectors of a and ȧ as in (8.51). The main
limit of the application of the former interpolation approach for the present unsteady case is that
it restricts the online time integration to the window where snapshots were acquired (for instance,
1.2 s for the case of Uin = 10 m/s or the window of 1-s length for Uin = 12 m/s). For the
present test case we selected the latter interpolation strategy (described in Subsection 8.4.2) be-
cause it is designed precisely to allow for time extrapolation in the online phase. This is based
on the assumption that as the values of the components of the vector time signal a(t) oscillate
between a minimum and a maximum over time, the replacement of t by a(t) as the RBF inter-

Figure 8.13. Cumulative ignored eigenvalue decay. In the plot, the solid black line refers to the
velocity eigenvalues, the dashed magenta line indicates the pressure eigenvalues, and the dash-dotted blue
line refers to the eddy viscosity eigenvalues.
 190 Chapter 8. Finite Volume–Based Reduced Order Models for Turbulent Flows

polation independent variable has in fact the advantage of allowing extrapolation. Nonetheless,
this is true if the values of the vector a components obtained during the ROM time integration
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

fall within the bounds of the FOM snapshots. Consequently, it is evident that the accuracy of this
interpolation outside the offline snapshot window is highly dependent on how close the current
solution vector a is to the vectors of the L2 projection coefficients (see (8.53)) used in the offline
stage to train the RBF. To further optimize the interpolation strategy, in this problem we applied
the splitting assumption presented in equation (8.48). Thus, the RBF interpolation was only used
to obtain the coefficients of the fluctuating part of the reduced eddy viscosity field. As a result,
the contribution of the parameter Uin only affects the time-averaged part νt (x; µ). At the re-
duced order level, the vector g is constant over time and is dependent on Uin . In the framework
of the present test case we had M = 10 corresponding to the number of eddy viscosity time-
averaged fields, which were computed as the average of the 200 time snapshots collected for
each parameter sample tested (see Table 8.1). At the online stage, the vector g(Uin ∗
) was com-
puted by linear interpolation, while the parameter-independent vector g(t ) was obtained from
∗

the RBF interpolation with respect to a and ȧ. Finally, the initial values for all vectors a(0, Uin
∗
),
b(0, Uin ), and g(0) were obtained from the inlet velocity parameter using a linear interpolation
∗

(based on the values of the initial L2 projection vectors of a(0, Uin ), b(0, Uin ), and g(0)). We
finally report on the tuning of the shape parameter required by the radial basis (RB) chosen for
the RBF interpolation. For all the tests in this work, the bases selected were Gaussian radial
functions, namely ζi,j (k · kR2Nu ) = exp(−r2 k · k2R2Nu ). The value of the RBF shape parameter
r was selected using cross-validation algorithm called leave-one-out cross-validation (LOOCV)
[190]. This choice resulted in satisfactory results for the tests conducted in this chapter.

Table 8.1. Offline parameter samples and the corresponding snapshot data.

Parameter sample: Uin in m/s FOM time step in s Snapshot acquiring time in s
7.5 0.0004 0.008
8 0.0004 0.008
8.5 0.00035 0.007
9 0.0003 0.006
9.5 0.0003 0.006
10 0.0003 0.006
10.5 0.0003 0.006
11 0.0003 0.006
11.5 0.00025 0.005
12 0.00025 0.005

U-ROM versus H-SUP-ROM Results

Here, we first report the results of the reduction strategy based on the supremizer approach. We
will compare the U-ROM and the H-SUP-ROM on the unsteady turbulent test case considered
here. At a later stage, we will also report the results of the turbulent ROMs based on the Pois-
son pressure equation (PPE). However, most of the results presented focus on the H-SUP-ROM
because of its general reduction formulation (it may in fact be used for both steady and unsteady
flows).
At this point, we illustrate the details of the first numerical test, which is a cross-validation or
an extrapolation test in the parameter space. The test is aimed at reconstructing the time history
of the fluid dynamics fields for the online parameter value Uin ∗
= 7.75 m/s, which was not
 8.6. Application of the H-SUP-ROM to Turbulent Problems 191

used to train the ROM. Also, one of the objectives of this test is to obtain the fluid dynamics
forces for time values which go far beyond the offline snapshot time window. After carrying out
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

the offline phase, all the reduced vectors, matrices, and tensors which appear in the dynamical
systems of the U-ROM and the H-SUP-ROM (in (8.68) and (8.50), respectively) were saved for
later use at the online stage. The online stage involved integrating the aforementioned dynamical
systems over time to obtain the vector a in the case of the U-ROM and the two vectors a and b
in the case of the H-SUP-ROM. The initial reduced vectors were approximated as the weighted
average of the ones corresponding to the first two offline parameter samples since Uin ∗
= 7.75
m/s lies between them. The fields obtained by solving the reduced systems of both ROMs
were then compared to the FOM ones. The FOM simulator was run for enough time to reach the
periodic regime, before extending the simulation at a fixed time step of 0.0004 s. FOM snapshots
were taken every 0.008 s, and the total simulation time reached was 8 s, which encompassed 13
solution periods. We emphasize that the last FOM snapshots mentioned were computed just for
the sake of comparing them with the obtained ROM fields, and these FOM fields were obviously
not used during the offline stage compute the POD modes. Given the periodic time dependence
of the variables considered, and the objective of comparing different models for cross-validation,
it was also important to pay special attention to the phase of the FOM snapshots. In particular,
the phase of the first snapshot acquired for all the training parameter samples was chosen to be
the same. This was also the case for the FOM testing snapshots computed to assess the ROM
accuracy. It was also assumed that the starting time of the online simulations was equal to zero
in all the tests considered here.
The velocity, pressure, and eddy viscosity fields computed by the FOM, the U-ROM, and
the H-SUP-ROM at the time instant t = 2.8 s are depicted in Figures 8.14, 8.15, and 8.16,
respectively. The results show clearly that the FOM velocity field is accurately approximated by
both the U-ROM and the H-SUP-ROM. This visual assessment is also confirmed by quantitative
evaluation. The L2 relative errors for both the U-ROM and the H-SUP-ROM are in fact 1.3553%
and 0.6954%, respectively. As for the pressure field, upon visual inspection the accuracy of the
U-ROM is noticeably lower than that of the hybrid ROM. This is again confirmed by quantitative
evaluation, because the L2 relative error for the U-ROM and H-SUP-ROM pressure reduced
fields are 33.0963% and 4.8085%, respectively. The modes used in the online stage in this test
case were Nr = 14 in the case of the U-ROM and Nu = 20 and Np = NS = Nνt = 10

(a) (b)

(c)

Figure 8.14. Velocity fields for the parameter value Uin = 7.75 m/s at t = 2.8 s: (a) shows the
FOM velocity, while in (b) one can see the U-ROM velocity with Nr = 14, and finally in (c) we have the
H-SUP-ROM velocity with Nu = 20 and Np = NS = Nνt = 10.
 192 Chapter 8. Finite Volume–Based Reduced Order Models for Turbulent Flows
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

(a) (b)

(c)

Figure 8.15. Pressure fields for the parameter value Uin = 7.75 m/s at t = 2.8 s: (a) shows the
FOM pressure, while in (b) one can see the U-ROM pressure with Nr = 14, and finally in (c) we have the
H-SUP-ROM pressure with Nu = 20 and Np = NS = Nνt = 10.

for the H-SUP-ROM. This online mode setup results for both ROMs in the best results in terms
of the fluid dynamics field’s L2 norm relative error with respect to the FOM solution. The
reduced pressure field plots suggest that the main limit of the U-ROM approach is its inability
to obtain accurate pressure field predictions. This lack of accuracy is most noticeable in the
region close to the surface of the cylinder, which clearly affects the accuracy of the reduced
approximation of the lift and drag forces. This shortcoming is clearly associated with the lack of
specific pressure and turbulence treatments in the U-ROM formulation. On the other hand, the
H-SUP-ROM features additional variables that are included to reproduce at the ROM level the
pressure-velocity coupling and the turbulence model used in the FOM. As a result, the accuracy
of its pressure field prediction is satisfactory.
A further assessment of the accuracy of both ROMs is given by the time evolution of the L2
relative error associated with the approximation of each field. The plot corresponding to the time

(a) (b)

(c)

Figure 8.16. Eddy viscosity fields for the parameter value Uin = 7.75 m/s at t = 2.8 s: (a)
shows the FOM eddy viscosity, while in (b) one can see the U-ROM eddy viscosity with Nr = 14, and
finally in (c) we have the H-SUP-ROM eddy viscosity with Nu = 20 and Np = NS = Nνt = 10.
 8.6. Application of the H-SUP-ROM to Turbulent Problems 193
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

(a) (b)

Figure 8.17. The time evolution of the L2 relative errors of the velocity-reduced approximations
for both the U-ROM and the H-SUP-ROM. The curves correspond to the case run with the parameter value
Uin = 7.75 m/s: (a) shows the error curve for the U-ROM, while (b) depicts the case of the H-SUP-ROM.
The error values in both graphs are in percentages.

evolution of the velocity field error is depicted in Figure 8.17, while Figure 8.18 shows the error
curve for the pressure field. In ROM, it is expected that the error is a growing function of time.
This is especially true when the ROM simulation extends well beyond the time window in which
the training snapshots were collected. Thus, ROMs which are able to contain the growth of this
error during the integration of the dynamical system are considered more accurate and favorable.
In this regard, the error plots presented suggest that the H-SUP-ROM is more effective at curbing
the error values. The velocity field error diagrams also indicate that increasing the number of
velocity modes Nu (Nr ) results in lower values of u for the H-SUP-ROM (U-ROM). As for the
pressure field errors observed in Figure 8.18, an increasing number of modes Nr does not lead
to significant error reduction in the U-ROM. Instead, increasing the number of modes associated
with all the fields generally lowers the pressure error, which can be kept to satisfactory values for
the whole time window considered.

Accuracy of H-SUP-ROM Fluid Dynamics Forces

The next result in this test is related to fluid dynamics forces, which are often one of the most
important performance indicators for engineers. Such forces are in fact in most cases the main
output of fluid dynamics simulation campaigns. Fluid dynamics forces acting on a body de-

(a) (b)

Figure 8.18. The time evolution of the L2 relative errors of the pressure-reduced approximations
for both the U-ROM and the H-SUP-ROM. The curves correspond to the case run with the parameter value
Uin = 7.75 m/s: (a) shows the error curve for the U-ROM, while (b) depicts the case of the H-SUP-ROM.
The error values in both graphs are in percentages.
 194 Chapter 8. Finite Volume–Based Reduced Order Models for Turbulent Flows

pend on the values of the velocity and pressure fields in the specific domain region surrounding
the boundary corresponding to such a body. Thus, evaluating the accuracy of the ROM using
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

the global error evaluators discussed so far only provides marginal information on how well the
reduced solution is able to lead to satisfactory force predictions. In the case of the present nu-
merical test, for instance, a considerable pressure or velocity error localized in the small region
around the cylinder might have a substantial impact on the fluid dynamics forces acting on the
cylinder, while having little effect on the global field errors. For this reason, the following anal-
ysis considers the time evolution of the lift coefficient Cl , i.e., the nondimensionalized vertical
component of the fluid dynamics force acting on the cylinder. It is important to point out that
the lift and drag forces exerted by the fluid on the cylinder are not a direct result of the H-SUP-
ROM computations. The reduced system solution consists in fact of the modal coefficients of
the velocity and pressure fields at each time instant, which are in turn used to obtain the H-SUP-
ROM approximation of the full-rank flow field. This approximation can obviously be used to
obtain—by integrating the pressure and skin friction on the cylinder surface—the reduced order
approximation of the fluid dynamics force components and the corresponding nondimensional
force coefficients. Yet, in the ROM community it is recommended to avoid possibly expensive
operations which involve evaluating the full-rank flow field. For this reason, the lift and drag
coefficients in this work are computed in a fully reduced order fashion, based on the offline com-
putation of suitable matrices which are then used in the online stage. The detailed procedure for
the online fluid dynamics forces computation was laid out in Subsection 7.2.6.
The resulting lift coefficient curves computed by both the U-ROM and the H-SUP-ROM
are compared to the FOM. Figure 8.19 shows the time signal of the lift coefficient obtained
with the FOM, the U-ROM, and the H-SUP-ROM, when the inflow velocity parameter value
is set to Uin
∗
= 7.75 m/s. Plot (a) presents the full history of the lift coefficient for the [0, 8]
s time interval of this test. Plot (b) represents instead a closeup of the same curves’ behavior
in the last 2 s of the simulation. The diagrams—plot (b) in particular—indicate that the H-
SUP-ROM predictions lead to improved force predictions with respect to the U-ROM. Thus,
the lower overall L2 error norms on pressure and velocity fields obtained by the H-SUP-ROM
indeed result in more accurate force predictions with respect to the U-ROM. Moreover, the long-
time integration used in the present test suggests that the U-ROM solution presents a slightly
unstable behavior, since its force amplitude grows as the simulation time increases, along with
the prediction error. On the other hand, the H-SUP-ROM Cl curve matches its FOM counterpart
to a good degree, even in the final time instants of the simulation. The number of reduced modes
employed in the online stage is the same as reported for the fluid dynamics field error results
presented previously.

(a) (b)

Figure 8.19. Lift coefficient curves for the cross-validation test done for the parameter value
Uin = 7.75 m/s for the time range [0, 8] s; the figure shows the FOM, the U-ROM, and the H-SUP-ROM
lift coefficient histories: (a) the full range is shown, and (b) the last 2 s of Cl is shown.
 8.6. Application of the H-SUP-ROM to Turbulent Problems 195

In order to have a better quantitative assessment of lift coefficient prediction accuracy by both
ROMs, we introduce the L2 relative percentage error
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

kCl (t) − Cl ∗ (t)kL2 (T1 ,T2 )

Cl = × 100%, (8.77)
kCl (t)kL2 (T1 ,T2 )

where Cl (t) is the time signal of the FOM lift coefficients between T1 and T2 . Cl ∗ (t) is the
corresponding lift coefficient time signal computed by the ROM—either the U-ROM or the H-
SUP-ROM. Referring back to Figure 8.19, we calculated Cl for the full time range of [0, 8] s and
also for its first half [0, 4] s. The values of the U-ROM for these time intervals are 16.7095% and
5.1252%, respectively. On the other hand, the H-SUP-ROM Cl approximation error is 3.5792%
for the full simulation time range [0, 8] s and 4.0504% for its first half [0, 4] s. The higher values
of the Cl error in the second part of the simulation confirm the U-ROM’s unsteady behavior,
which particularly affects problems requiring long-time integration. Such instability has not
been observed in the case of the hybrid model, in which instead the errors measured in the full
time interval and in its first half are very close. Figure 8.20 shows the error mCl as a function
of the number of modes used in the online stage. The plot suggests that the H-SUP-ROM force
coefficient error is minimized, considering 20 velocity modes and 10 modes each for pressure,
supremizers, and eddy viscosity fields. In this condition, the relative force coefficient error is as
low as 3%. As for the U-ROM Cl error, it is consistently above 16%, with a minimum value
obtained with 14 velocity modes.
Figure 8.21 displays the ROM results in terms of a further error metric devised in an attempt
to attribute how much of the L2 error is caused by imprecise reproduction of the amplitude rather
than the phase of the lift coefficient oscillations. The relative peak error peak is
P Kn,F OM − P Kn,∗
n,peak = × 100%, (8.78)
P Kn,F OM
where P Kn,F OM is the value of the nth FOM Cl peak and P Kn,∗ is the value of the nth U-ROM
or H-SUP-ROM Cl peak. The figure shows the relative peak error for both ROMs, computed for

(a) (b)

Figure 8.20. The graph of the L2 relative errors for the lift coefficient curve versus the number
of modes used in the online stage in the U-ROM and the H-SUP-ROM. The curves correspond to the case
run with the parameter value Uin = 7.75 m/s. The error is computed between the lift coefficient curve
obtained by the FOM solver and the one reconstructed from both the U-ROM and the H-SUP-ROM for the
time range [0, 8] s: (a) shows the error curve for the U-ROM, where Nr is the number of modes used in the
online stage for all variables (by construction of the U-ROM, it is not possible to choose a different number
of online modes for the reduced variables). (b) depicts the case of the H-SUP-ROM, where one can see
the error values by varying the number of modes used for the pure velocity, with different fixed settings for
the three other variables (the pressure, the supremizers, and the eddy viscosity). The error values in both
graphs are in percentages.
 196 Chapter 8. Finite Volume–Based Reduced Order Models for Turbulent Flows
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

(a) (b)

Figure 8.21. The graph of the peak relative errors for the lift coefficient curves for varied values
of the number of modes used in the online stage in the U-ROM and the H-SUP-ROM. The curves correspond
to the case run with the parameter value Uin = 7.75 m/s. The error is computed between the peak values
of the lift coefficient curve obtained by the FOM solver and the ones reconstructed from both the U-ROM
and the H-SUP-ROM for the time range [0, 8] s: (a) shows the error curve for the U-ROM, where Nr is
the number of modes used in the online stage for all variables (by construction of the U-ROM it is not
possible to choose a different number of online modes for the reduced variables). (b) depicts the case of the
H-SUP-ROM. The error values in both graphs are in percentages.

each of the 29 peaks in the time interval [0, 8] s. The different curves in the plot refer to the dif-
ferent combinations of modal truncation values employed for each field considered in the ROM
simulations. Figure 8.21 indicates that the U-ROM relative error increases at each peak until it
reaches values as high as 10–20%. On the other hand, the hybrid ROM presents lower values
of the relative peak errors, which are not monotonically growing at each lift coefficient peak
throughout the simulation. The peak error is lower than 3.5% for several combinations of the
modal truncation order fields tested. The cross-validation tests presented so far only involved the
U-ROM and the H-SUP-ROM. They were clearly aimed at establishing how well such reduced
order methodologies are able to perform when employed in a significant engineering application.
We now turn our focus to the performance of the two other turbulent ROMs described in Sec-
tion 8.5. To this end, we report on the results of a further cross-validation simulation campaign
carried out on the same cross-flow cylinder test case considered in the previous tests.

SU-PPE-ROM and H-PPE-ROM

The ROMs we consider now are referred to as the SU-PPE-ROM and the H-PPE-ROM. At the
online level, both ROMs use the Poisson equation for pressure to obtain a stable pressure-velocity
coupling. Given the formulation of the Poisson equation for pressure boundary conditions, these
two ROMs are only suited for the reduction of unsteady flows, and they cannot be extended
to the steady case. The SU-PPE-ROM is an extension of the U-ROM, obtained by adding a
specific Poisson equation–based treatment for the pressure field. The H-PPE-ROM is a further
extension, obtained by adding a turbulence treatment equivalent to the H-SUP-ROM, based on
the data-driven approximation of the eddy viscosity field. Thus, whereas the SU-PPE-ROM is
an intrusive ROM, by our previous definition the H-SUP-ROM is a hybrid ROM.
As mentioned above, the cross-validation test for the SU-PPE-ROM and the H-PPE-ROM is
identical to the one conducted for the U-ROM and the H-SUP-ROM. Therefore, the online pa-
rameter value selected for the comparisons was Uin ∗
= 7.75 m/s. In the online simulations, the
SU-PPE-ROM was applied by using the penalty method to treat the nonhomogeneous Dirichlet
boundary conditions on the velocity field. On the other hand, the lifting function method was
used to enforce the boundary conditions in the H-PPE-ROM online solution. The lifting function
 8.6. Application of the H-SUP-ROM to Turbulent Problems 197

chosen in this case corresponds to the solution of a potential flow problem with the inlet velocity
fixed at Uin = 1 m/s. Such an auxiliary problem was solved by running an iterative steady
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

solver. The lifting velocity field obtained was considered as the additional lifting mode/function
which was added to the velocity POD modes. The next step of the offline phase was to homoge-
nize the velocity snapshots using the velocity lifting function, which resulted in the 2000 velocity
snapshots processed to obtain the velocity POD modes. All the reduced matrices, vectors, and
tensors which appear in the dynamical systems of both ROMs and are reported in (8.75) and
(8.74) were then computed and stored. The initial velocity, pressure, and eddy viscosity reduced
vectors were set in the exact way as in the previous test.
The first results shown refer to the valuation of the lift coefficient Cl in the same [0, 8] s
time interval considered in the previous test. Equation (8.22) depicts the lift coefficient curves
obtained by the PPE ROM and the FOM. The plot indicates that both ROMs lead to acceptable
force coefficient predictions. If on one hand the accuracy of the hybrid H-PPE-ROM prediction
is once again superior to that of the SU-PPE-ROM, the latter model leads to significantly better
Cl results than the U-ROM ones presented in Figure 8.19. The main difference between the
U-ROM and the SU-PPE-ROM lies in the specific pressure treatment included in the latter ROM
online equations. Thus, its improved performance over the U-ROM indicates the accuracy gains
resulting from including suitable pressure-velocity coupling methodologies at the online stage.
In addition, the further performance improvement obtained with the H-PPE-ROM is an indirect
measure of the accuracy gains associated with the H-PPE-ROM-specific turbulence treatment,
which is the main algorithmic difference between the two ROMs.
We also point out that the combination of modal truncation levels for each fluid dynamics
field that was used to generate Figure 8.19 represents for both ROMs the best combination in
terms of accuracy measured by the error Cl (defined in (8.77)). In a similar fashion, Figure
8.22, which shows Cl as a function of time, was generated using the SU-PPE-ROM with the
number of velocity and pressure modes—nine—minimizing the error. Also, the H-PPE-ROM
curve in the plot was obtained with the most successful combination of modal truncation levels
tested, in which the number of velocity, pressure, and eddy viscosity modes is equal to 15. From
a quantitative perspective, the results in Figure 8.22 show that the SU-PPE-ROM has an error
Cl = 5.0525% for the time range [0, 4] s and Cl = 12.4255% for the full time range. On the
other hand, the corresponding H-PPE-ROM errors are 2.2445% and 4.3678%. This evaluation
confirms that, with the addition of a specific pressure-velocity coupling methodology, the SU-
PPE-ROM represents a significant improvement over the U-ROM. Nevertheless, its main source
of residual inaccuracy is associated with its turbulence treatment. In particular, the constraint of
having temporal coefficients shared between velocity and eddy viscosity could be problematic

(a) (b)

Figure 8.22. Lift coefficient curves for the cross-validation test done for the parameter value
Uin = 7.75 m/s for the time range [0, 8] s. The figure shows the FOM, the SU-PPE-ROM, and the
H-PPE-ROM lift coefficient histories: (a) the full range is shown, and (b) the last 2 s of Cl is shown.
 198 Chapter 8. Finite Volume–Based Reduced Order Models for Turbulent Flows

in specific settings. For example, if 20 and 12 modes are needed to accurately reconstruct the
velocity and the eddy viscosity fields, respectively, then this formulation would force us either to
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

work with a nonoptimal set of velocity shape functions or to include a set of eight unneeded eddy
viscosity modes in the solution of the reduced DAE. Since these additional eddy viscosity modes
are typically affected by a high signal to noise ratio, this is likely to cause stability problems.
This is further confirmed by Figure 8.23, which presents the convergence analysis for the
ROMs based on the Poisson equation for pressure. The plots, which depict the SU-PPE-ROM
(left) and the H-PPE-ROM (right) lift coefficient error as a function of the velocity modal trun-
cation level, present different curves, associated with the different truncation level combinations
selected for the other fields in the tests. The left plot clearly shows that the SU-PPE-ROM error
reaches a minimum when Nu = Np = 9, which is the case depicted in Figure 8.22, for a total
number of degrees of freedom of 18. The SU-PPE-ROM error pattern is similar to the U-ROM
one depicted in Figure 8.20. Both curves present in fact two minima located at truncation values
of 9 and 14 for the velocity and the pressure modes, and the error values remain consistently
above 10%. This can once again be explained by pointing out that after nine velocity modes
are reached, the accuracy gains granted by adding further modes are spoiled by the presence of
unneeded and noisy eddy viscosity modes in the system.
As for the H-PPE-ROM, it was instead able to fully exploit the potential accuracy gains
associated with higher velocity truncation levels. The lift coefficient error in the right plot is in
fact still decreasing when the number of velocity modes grows up to 13–15 modes. In addition,
the error values are lower than 10% for several truncation settings. For example, Cl ' 8%
when 13 modes are employed for all reduced variables. We must however remark that among the
Nu H-PPE-ROM modes, we also include the lifting function mode added to treat the boundary
conditions at the reduced order level. Figure 8.24 reports the error for each peak of the Cl
obtained for the PPE ROMs for different settings of the online modes. The left plot substantially
confirms that the SU-PPE ROM improved results with respect to the U-ROM, which speaks to
the necessity of having a separate set of reduced coefficients for the pressure field. The values of
the SU-PPE ROM peak errors observed in the left diagram are in fact as low as 2% for certain

(a) (b)

Figure 8.23. The graph of the L2 relative errors for the lift coefficient curve versus number of
modes used in the online stage in the SU-PPE-ROM and the H-PPE-ROM. The curves correspond to the
case run with the parameter value Uin = 7.75 m/s. The error is computed between the lift coefficient
curve obtained by the FOM solver and the one reconstructed from both the U-ROM and the H-PPE-ROM
for the time range [0, 8] s: (a) shows the error curve for the SU-PPE-ROM, where Nu is the number of
modes used in the online stage for both the velocity and the eddy viscosity, while Np is the number of modes
used for the pressure field. (b) depicts the case of the H-PPE-ROM, where one can see the error values
varying the number of modes used for the velocity (including the lifting velocity mode) with different fixed
settings for the two other variables (the pressure and the eddy viscosity). The error values in both graphs
are in percentages.
 8.6. Application of the H-SUP-ROM to Turbulent Problems 199
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

(a) (b)

Figure 8.24. The graph of the peak relative errors for the lift coefficient curves for varied values
of the number of modes used in the online stage in the SU-PPE-ROM and the H-PPE-ROM. The curves
correspond to the case run with the parameter value Uin = 7.75 m/s. The error is computed between the
peak values of the lift coefficient curve obtained by the FOM solver and the ones reconstructed from both the
SU-PPE-ROM and the H-PPE-ROM for the time range [0, 8] s: (a) shows the error curve for the U-ROM,
where Nu is the number of modes used in the online stage for both the velocity and the eddy viscosity, while
Np is the number of modes used for the pressure field. (b) depicts the case of the H-PPE-ROM. The error
values in both graphs are in percentages.

choices of the online modal truncation levels. The image on the right illustrates that the H-PPE-
ROM results succeeded in recovering the peaks of the Cl time signal, with errors lower than 2%
for online modal truncation orders Nu = 13, 14, 15. A further result is reported in Figure 8.25
and Figure 8.26, which depict the time evolution of the ROMs’ L2 prediction error for velocity
and pressure fields, respectively. The plots further confirm the H-PPE-ROM’s superior ability to
obtain more accurate pressure field approximation than the SU-PPE-ROM. The SU-PPE-ROM
pressure field error values at the end of the 8-s time integration interval are above 10% for all
the different modal truncation orders tested. On the other hand, the corresponding error value
for the H-PPE-ROM reaches values as low as 4% at the final time of the reduced simulations,
for the best combinations of the online mode numbers tested. A comparison with Figure 8.18
also suggests that the H-PPE-ROM has a greater ability to suppress the error growth than the H-
SUP-ROM. A possible justification for this finding is that the use of the PPE at the reduced level
has introduced a higher level of consistency between the ROM formulation and the FOM one.
In fact, the finite volume FOM solver features a pressure equation obtained from the momentum

(a) (b)

Figure 8.25. The time evolution of the L2 relative errors of the velocity-reduced approximations
for both the SU-PPE-ROM and the H-PPE-ROM. The curves correspond to the case run with the parameter
value Uin = 7.75 m/s: (a) shows the error curve for the SU-PPE-ROM. (b) depicts the case of the H-PPE-
ROM. The error values in both graphs are in percentages.
 200 Chapter 8. Finite Volume–Based Reduced Order Models for Turbulent Flows
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

(a) (b)

Figure 8.26. The time evolution of the L2 relative errors of the pressure-reduced approximations
for both the SU-PPE-ROM and the H-PPE-ROM. The curves correspond to the case run with the parameter
value Uin = 7.75 m/s: (a) shows the error curve for the SU-PPE-ROM. (b) depicts the case of the H-PPE-
ROM. The error values in both graphs are in percentages.

and the continuity equation, as explained in Subsection 7.2.2. This procedure is mimicked by
the H-PPE-ROM, possibly allowing for an improvement in accuracy. Nonetheless, we have to
mention that in general the FOM solution algorithm differs from the ROM one, in that the FOM
one is based on an iterative segregated approach, while the ROM one uses a coupled approach.
To summarize the results given by the four ROMs tested in the current problem, Table 8.2
reports the performance of all the ROMs in terms of their computational efficiency and of the
accuracy indicators discussed. In the table SU stands for the speedup, which is calculated as
tof f
SU = tonline , with tof f being the wall time needed to run the full-order simulations and export
the snapshots and tonline the corresponding online time. In the cross-validation test considered
for Uin = 7.75 m/s, the value of tof f is 1540.766 s. This corresponds to the wall time of running
the simulations in parallel on six processors, along with the time needed to reconstruct the fields.
The accuracy results reported in the table are obtained with each ROM’s best combinations of
online modal truncation orders and refer to the error CL in the full time integration range [0, 8] s.

Table 8.2. Summary of the accuracy and the efficiency results for the ROMs considered in the
problem of flow around a cylinder.

ROM CL maxn n,peak SU DoF

U-ROM 16.7095 % 12.67 % 31.80 Nr = 14
SU-PPE-ROM 12.4255 % 4.995 % 30.16 Nu = Np = 9
H-SUP-ROM 3.5792 % 3.498 % 9.34 Nu = 20, Np = NS = Nνt = 10
H-PPE-ROM 4.3678 % 1.59 % 12.86 Nu = Np = Nνt = 15

Higher Reynolds Numbers and Time Extrapolation

The objective of the next test is to measure the accuracy of the hybrid ROM (in this case the H-
SUP-ROM was chosen) for higher Reynolds number and for longer time extrapolation intervals.
Therefore, we select the inlet parameter online sample Uin ∗
= 11.75 m/s (corresponding to Re =
117500) and choose a ROM time integration interval of [0, 10] s. This time interval features 27
periodic solution cycles, which is almost double the number of solution cycles contained in the
[0, 8] s considered in the previous case. The results include the time history of the lift coefficient
Cl in addition to its corresponding L2 relative error CL and also the peak error and the reduced
approximation of the time period.
 8.6. Application of the H-SUP-ROM to Turbulent Problems 201
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

(a) (b)

Figure 8.27. Lift coefficient curves for the cross-validation test done for the parameter value
Uin = 11.75 m/s for the time range [0, 10] s. The figure shows the FOM and the H-SUP-ROM lift
coefficient histories: (a) the full range is shown; (b) the last 3-s history of Cl is shown.

As in the first test, the FOM simulator was run for enough time to reach the periodic regime.
Then the simulation was extended for another 10 s, starting from a phase which is equivalent to
the one set for the offline samples. The additional 10 s was simulated with time step equal to
0.00025 s. The formulation and the reduction strategy of the H-SUP-ROM are the same as the
ones set in the first numerical test in this section. The Cl H-SUP-ROM curve is reproduced using
12 modes for the velocity and 10 modes for each of pressure, supremizers, and eddy viscosity.
Figure 8.27 shows the results for the Cl curves over the whole time range considered in plot (a).
Diagram (b) depicts a closeup of the lift curves in the last 3 s simulated. The images clearly
suggest that the approximately 50% Reynolds number increase does not affect the accuracy of
the H-SUP-ROM. They also prove the ability of the hybrid ROM to tackle long-time integration,
since at a visual investigation the ROM solution accuracy does not appear to deteriorate over
time. The value of the overall error CL (as defined in (8.77)) is 1.9486%. The peaks of the FOM
Cl curve were recovered with maximum error of 2.0526% among all 55 peaks present in the
time integration interval. Finally, the average time period computed by the FOM solver is about
0.3641 s, while the average time period approximated by the hybrid ROM is about 0.3642 s.

Effect of Different Turbulence Models on the Accuracy of the ROMs

The results presented in this section fully characterize the accuracy and efficiency of the turbulent
ROMs developed. The final test we present is instead aimed at assessing their flexibility with
respect to the turbulent FOM. An important aim in generating these reduction algorithms was
to obtain ROMs that were sensitive to the turbulence model used at the FOM level and yet did
not require resolving the turbulent model equations at the ROM level. Thus, the present test will
evaluate the hybrid ROM’s ability to obtain accurate approximations of FOM solutions obtained
with different turbulence models. To this end, we carried out a further FOM simulation in which
the turbulence models selected were both the k- model and the SST k-ω model previously
considered. The test case was still the cylinder benchmark case. No parameter dependence was
considered in this test. So, the goal of the reduction was just to reproduce the time snapshots and
extrapolate the time evolution of the system. The Reynolds number selected is Re = 105 , and
the FOM simulation was carried out for both turbulence models until the regime solution was
fully developed. After that, snapshots were acquired covering a time window of 1.2 s and 1.6
s for the k- and SST k-ω models, respectively. The hybrid ROM chosen is the H-SUP-ROM
with the penalty method–based treatment of the boundary conditions. The H-SUP-ROM was
run for 8 s for both turbulence models. The FOM lift coefficients obtained with both closure
models are depicted in Figure 8.28, along with their ROM approximations. It is evident from the
 202 Chapter 8. Finite Volume–Based Reduced Order Models for Turbulent Flows
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 8.28. The lift coefficient curves obtained using both the k- and SST k-ω turbulence mod-
els and the H-SUP-ROM ones. The case considered is nonparametrized, with Uin = 10 m/s corresponding
to Re = 105 . The plot is for the time range t ∈ [6, 8]; the H-SUP-ROM achieved relative L2 errors (over
the range t ∈ [0, 8]) less than 5% in both cases.

graph that the H-SUP-ROM proves sensitive to the specific turbulence model used in the FOM
solver, although no additional PDEs for the turbulent quantities are solved at the reduced level.
The value of the errors defined in (8.77) for both turbulence models are well below 5%, and the
reduced approximation of the average time period is accurate in both cases, with relative errors
below 1%.
 Chapter 9
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Nonintrusive Data-Driven
Reduced Order Models
in Computational Fluid
Dynamics

Marco Tezzele, Nicola Demo, Giovanni Stabile,

Gianluigi Rozza

9.1 Introduction
In the previous chapters we discussed a wide range of aspects of the reduced basis (RB) ap-
proach, presenting different applications and different couplings with the traditional discretiza-
tion methods for the resolution of partial differential equations (PDEs). Such applications have
in common that, independent of the solver, the construction of the reduced order model (ROM)
requires knowledge of the actual mathematical model and access to the operators describing the
PDE in order to project them onto the reduced space.
This projection is often impossible when dealing with commercial software, since the source
code is not accessible by the user, making the standard intrusive ROM approach infeasible. More-
over, nonlinear terms are particularly challenging to treat with classical projection-based meth-
ods, and the computational gain obtained by dimensionality reduction is not as remarkable as for
linear problems.
To address these issues, nonintrusive data-driven methods can be used. This approach retains
the typical offline-online decomposition, but instead of computing the reduced order operators,
it computes the low-dimensional representation of the snapshots. These new latent variables are
computed in an equation-free and data-driven framework, making this approach very versatile
and suitable for model discovery, system identification, and efficient future state predictions, and
in general for dealing with experimental data. These features allowed the spread of nonintrusive
ROMs in many industrial contexts, since they can be easily integrated into existing computational
pipelines without disrupting the industrial workflow, while reducing the overall computational
burden [517, 508, 571].
This chapter focuses on such nonintrusive methods, presenting in Section 9.2 the general
framework for parametric ROMs in a data-driven fashion and in Section 9.3 the dynamic mode
decomposition (DMD) technique for reducing dynamical systems. Lastly, we provide in Sec-
tion 9.4 a summary of the numerical results obtained by applying the methods to three different
engineering problems.

203
 204 Chapter 9. Nonintrusive Data-Driven ROMs in CFD

9.2 A General Framework for Nonintrusive Parametric

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

ROMs
In this section we present a general framework for nonintrusive data-driven ROMs. We assume
we do not know the equations behind the computation of the solution snapshots, or, even if we
know the model being used, we do not have access to the solver, so we cannot perform intrusive
model order reduction.
From an abstract point of view we can identify three main steps to constructing this kind
of ROM: create the solutions database, compute a basis of reduced dimension, and choose a
regression method to reconstruct the solution manifold. As we can already see, there are multiple
choices for every one of these steps. We can use, e.g., different sampling methods to collect
the snapshots, linear or nonlinear techniques for modes selection, and finally any method from
polynomial interpolators to artificial neural networks to reconstruct the parameters into a modal
coefficients map.
The framework described in this section, with its modular structure, can be found in the
open-source Python package4 called EZyRB [158]; the reader can also refer to Chapter 19.

9.2.1 Database Creation

Let us denote the parameters by µ and the state snapshots by x. The solutions database is thus the
set of parameter-snapshot pairs defined by {µi , xi }m
i=1 , with µi ∈ P ⊂ R and xi ∈ R . Here,
p n

n represents the number of degrees of freedom of our system, while p is the number of input
parameters. We remark that the snapshots can represent data coming from experiments, from
simulations, or even from sensors in real time. We arrange the snapshots by column in X as
 
| | |
X =  x1 x2 . . . x m  . (9.1)
| | |

There are several possible initial sampling strategies for selecting the parameters’ location,
depending on the specific task for which we construct the ROM. For example, to cover as much
of the parameter space as possible the most used approaches are Latin hypercube sampling (LHS)
[555]; Sobol sequences [544], which are quasi-random low-discrepancy sequences; or other
space-filling strategies, such as dynamic propagation sampling [339]. If the parameters repre-
sent data coming from sensors or other components with an associated probability distribution, a
weighted approach is suggested (see Chapter 12). For optimization tasks, any a priori knowledge
about the possible region of interest and/or the simulated process should be incorporated within
the sampling method. We remark that some of the aforementioned sampling schemes will not
work if the parameters can assume only a finite set of possible values.

9.2.2 Linear Dimensionality Reduction

Linear dimensionality reduction is the most widespread technique for unveiling structures within
a large dataset. The main assumption is that the data can be expressed by a linear combination
of a few global basis functions, often called modes. Following the notation introduced in the
previous subsection, we seek to approximate the snapshots xi as
m
X r
X
xi = ϕk cki ≈ ϕk cki ∀i ∈ [1, . . . , m], r  m, (9.2)
k=1 k=1

4 Available at https://github.com/mathLab/EZyRB/.
 9.2. A General Framework for Nonintrusive Parametric ROMs 205

for some modes ϕk ∈ Rn and for some modal coefficients cki ∈ R. We can rewrite (9.2) in
matrix form, arranging all the data by column, as done in equation (9.1) obtaining
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

    
| | | | | | | | |
 x1 x2 . . . xm  ≈  ϕ1 ϕ2 . . . ϕr   c1 c2 . . . cm  (9.3)
| | | | | | | | |

or, equivalently, X ≈ ΦC, with X ∈ Rn×m , Φ ∈ Rn×r , and C ∈ Rr×m .

As we can see, there are many possible choices for computing the modes, such as classical
principal component analysis (PCA), also called proper orthogonal decomposition (POD), and
other nonlinear approaches such as kernel POD [531] and autoencoders [367]. In this chapter we
will review only POD, which has been proved successful in a variety of industrial applications
[571, 409, 519] and both is explainable and can be linked to energy considerations. Moreover,
many high-dimensional data can be approximated by low-rank matrices [589]. In case of miss-
ing data or sensors providing partial information, a gappy POD [83, 619] approach can be used.
We emphasize that the basis identified by the POD can also be exploited in a Galerkin projec-
tion framework [558, 230, 281] or in a hybrid framework combining data-driven methods with
projection [286, 220].
In particular, if the pair (Φ, C) minimizes the sum of the squares of the residuals X − ΦC,
i.e., in a least squares sense, we have that Φ is the matrix composed by the POD modes. Using
the l1 -norm instead of least squares leads to the so-called robust PCA [94], which is very useful
if outliers are present. To actually compute the modes Φ in the POD case, we can perform the
singular value decomposition (SVD) of the snapshot matrix X,

X = U ΣV ∗ , (9.4)

where ∗ denotes the conjugate transpose, and retain only the first r components. The columns of
U are the so-called POD modes. They are orthogonal and they span the optimal low-dimensional
subspace in the least squares sense. They can also be computed as the eigenvectors (up to a nor-
malization factor) of the Gram matrix X T X. The diagonal matrix Σ = diag(λ1 , . . . , λr ) is
composed by the singular values, arranged in descending order, which indicate the energy con-
tribution of the corresponding modes. The error introduced by the truncation r can be measured
as [470]
kX − U ΣV ∗ k22 = λ2r+1 , (9.5)
v
u m
u X
∗
kX − U ΣV kF = t λ2i , (9.6)
i=r+1

where the subscripts 2 and F refer to the Euclidean norm and the Frobenius norm, respectively.
The singular values λi are arranged in descending order. In Figure 9.1 we can see an example of
POD modes for a flow past a cylinder modeled with a parametrized Navier–Stokes problem.
We remark that given the first r modes, we only need r coefficients to reconstruct the whole
state xi . To compute the modal coefficients, we project the data onto the POD subspace:

C = U T X. (9.7)

Given these coefficients we can only reconstruct the original snapshots which make up the initial
database. For a fast and accurate prediction of the state x∗ , corresponding to an unseen parameter
µ∗ , we need to reconstruct the parameters–to–modal coefficients map. Then by performing a
matrix multiplication with the POD modes we obtain the whole field of interest. This method
is called POD with interpolation (PODI). In the next subsection we show some possible choices
for reconstructing such a map.
 206 Chapter 9. Nonintrusive Data-Driven ROMs in CFD
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 9.1. First three POD modes obtained from the resolution of a parametrized Navier–Stokes
problem, describing the flow past a cylinder.

9.2.3 Solution Manifold Approximation

In this subsection we present some possible choices for reconstructing the solution manifold
described by the modal coefficients corresponding to the initial database. In particular, we want
to construct a regression s : P → Rr which approximates the map f : µ ∈ P → c ∈ Rr , given a
i=1 , where ci = U xi (µi ). The regression model s can then
set of m input-output pairs {µi , ci }m T

be used to predict the state x for a new input parameter µ∗ by computing

∗

x∗ = U s(µ∗ ). (9.8)

To build the interpolant s we can use several techniques, such as linear interpolation [519,
213], nonlinear Gaussian process regression (GPR) [622, 252, 431, 160], multifidelity methods
[494, 575], radial basis function (RBF) interpolation [82, 571], inverse distance weighting (IDW)
[623, 31], and artificial neural networks (ANNs) [602, 448], to cite a few. Here we will not review
all the possibilities; some of them are extensively presented in Chapter 17 in the context of shape
morphing, while in Chapter 20 an ANN approach is presented. We will focus only on GPR due
to its versatility and spread and to make clear the applications in the following sections.
We start by assuming the m input-output pairs to be noise-free observations. We assign
a prior to s with mean m(µ) and covariance function k(µ, µ0 ; θ), which means that s(µ) ∼
GP(m(µ), k(µ, µ0 ; θ)). Our beliefs about the function before looking at the observed values are
expressed by the prior. We consider zero mean GP, i.e., m(µ) = 0. We define the covariance
matrix K ∈ Rm×m as Ki,j = k(µi , µj ; θ). To make predictions using the Gaussian process we
 9.3. DMD for Time-Dependent Problems 207

need to find the optimal values for the hyperparameter vector θ by maximizing the log likelihood:
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

1 1 m
log p(c|µ, θ) = − cT K −1 c − log |K| − log 2π. (9.9)
2 2 2
Let µ∗ be the test samples and Km∗ = k(µ, µ∗ ; θ) be the matrix of the covariances evaluated
at all pairs of training and test samples. In a similar way, K∗m = k(µ∗ , µ; θ) and K∗∗ =
k(µ∗ , µ∗ ; θ). The predictions s∗ are obtained by conditioning the joint Gaussian distribution on
the observed values and by sampling the posterior as
s∗ |µ∗ , µ, c ∼ N (K∗m K −1 c, K∗∗ − K∗m K −1 Km∗ ). (9.10)

9.2.4 Computational Considerations

From a computational point of view, the differences between the full-order model (FOM) and
the ROM are remarkable: whereas the FOM requires the solution of a system of dimension n,
the ROM only requires a query of the interpolation function and a matrix multiplication with
the modes. So the computational cost of a ROM is mainly in the construction phase and not
in the prediction. Since we rely on the SVD, the method shows an algorithmic complexity of
O(min(n, m)nm).

9.3 DMD for Time-Dependent Problems

Dynamic mode decomposition (DMD) is a powerful method of identifying and approximating
dynamical systems using only a few spatiotemporal coherent structures [353, 79, 527]. The
method is equation free since it works using only high-dimensional snapshot data and makes
almost no assumptions regarding the underlying system to approximate. This technique has
been applied to a diverse range of fields, with particular success in computational fluid dynamics
(CFD). We can highlight three main tasks which can be accomplished by DMD: system identifi-
cation, future state prediction, and control.
For an efficient implementation of all the algorithms presented in this section we suggest
using the open-source Python package5 called PyDMD [159]. The reader can also refer to Chap-
ter 19.

9.3.1 Classical DMD Algorithm

We start by collecting the snapshots {xi }m
i=1 representing the state of the system at each time
instant, with xi ∈ Rn . We arrange them by columns in two matrices, X and X 0 , where X 0 is the
time-shifted version of X, i.e.,
   
| | | | | |
X =  x1 x2 . . . xm−1  , X 0 =  x2 x3 . . . x m  . (9.11)
| | | | | |
We seek a regression of the data onto locally linear dynamics such as xk+1 = Axk for k =
1, . . . , m − 1. In matrix form this can be written as
X 0 ≈ AX. (9.12)
A can be found as the best-fit matrix using the Moore–Penrose pseudoinverse of X, i.e., X † , as
follows:
A = X 0X †. (9.13)
5 Available at https://github.com/mathLab/PyDMD/.
 208 Chapter 9. Nonintrusive Data-Driven ROMs in CFD

This solution minimizes the error in the Frobenius norm kX 0 − AXkF . Since the usual CFD
problem is characterized by high-dimensional state snapshots with n  m, we cannot solve this
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

problem directly with a regression method. We rely instead on a lower-dimensional operator Ã

which evolves over the POD modes of the snapshot matrix in order to approximate the eigenpairs
of the full operator A. The eigenvectors of A are called DMD modes.
The DMD algorithm starts with an SVD of the snapshot matrix X with a truncation rank r:

X ≈ U ΣV ∗ , (9.14)

where U ∈ Cn×r , Σ ∈ Cr×r , and V ∈ Cm−1×r . By ∗ we denote the conjugate transpose. We

emphasize that U and V are orthonormal, and the columns of U are the so-called POD modes.
Substituting X † in (9.13) with its SVD, we obtain

A = X 0 V Σ−1 U ∗ . (9.15)

From a computational point of view it is more efficient to project the operator A onto the POD
modes, obtaining its low-rank representation

à = U ∗ AU = U ∗ X 0 V Σ−1 . (9.16)

With this operator we can express the evolution of the low-dimensional projection of the snap-
shots as x̃k+1 = Ãx̃k , with x̃k = U ∗ xk . The final step is to compute the eigenpairs of à with its
eigendecomposition
ÃW = W Λ, (9.17)
where Λ stands for the diagonal matrix of the eigenvalues and the columns of W are the eigen-
vectors. We remark that the eigenvalues of A are given by Λ, while the eigenvectors Φ, also
called exact DMD modes [588], are given by

Φ = X 0 V Σ−1 W. (9.18)

In the literature can also be found the so-called projected DMD modes [527] expressed by Φ =
U W , which tend to converge to the exact ones if the matrices in (9.11) share the column space.
The ith eigenvalue λi provides information about the frequency ωi of the corresponding
mode, defined as
Im(log λi )
ωi = , (9.19)
2π∆t
and also about its stability. In Figure 9.2 we can see the DMD modes corresponding to different
eigenvalues for a toy problem: if the eigenvalue is on the unit circle we have a stable mode; if it
is inside the unit circle it is convergent; otherwise it diverges.
Finally, to predict the future state for every time, we use
r
X
x(t) ≈ φk eψk t bk = ΦeΨt b, (9.20)
k=1

where Ψ = diag(ψk ) = diag(ln(λk )/∆t) and bk denotes the initial amplitude of the corre-
sponding kth DMD modes φk . To compute amplitudes we use the best fit in the least-squares
sense:
b = Φ† x1 , (9.21)
where we use the first snapshot x1 . We remark that this can be done with respect to every snap-
shot, and of course it will improve the accuracy in the neighborhood of that specific time instant.
The amplitudes can also be computed by minimizing the error over all the snapshots [315].
 9.3. DMD for Time-Dependent Problems 209
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 9.2. In the left column are the eigenvalue positions with respect to the unit circle, and in the
right column are the corresponding DMD modes. From top to bottom we have a stable, a convergent, and
a divergent mode, corresponding to an eigenvalue on the unit circle, inside it, and outside it, respectively.

9.3.2 DMD Extensions

Despite its versatility and performance, DMD has some limitations, mainly inherited from the
SVD at its core. DMD is thus incapable of efficiently handling translations or rotations, and
transient phenomena are not well characterized. Moreover, the presence of an external input
acting on the state results in the system being misidentified. To overcome these issues, many
variants have been developed and proposed in the past years, among them DMD with control
[463], multiresolution DMD [354], compressed DMD [183], higher-order DMD [364], and ran-
domized DMD [184, 62]. In the following sections we briefly review some of these extensions,
emphasizing their main advantages.

DMD with Control

When dealing with actuated systems, DMD fails to properly reconstruct the underlying dynamics
because it is unable to incorporate the contribution of the forcing term. DMD with control
(DMDc) [463] has been developed to overcome this issue. It includes the actuation snapshots in
the analysis and thus it is able to provide reduced order representations for input-output systems.
The DMDc method quantifies the effect of control inputs on the state of the system and computes
the underlying dynamics without being confused by the effect of external control.
In mathematical terms, and using the notation introduced in the previous subsection, we
collect the input control snapshots in the matrix I as
 
| | |
I =  u1 u2 . . . um−1  . (9.22)
| | |
 210 Chapter 9. Nonintrusive Data-Driven ROMs in CFD

Given the state xk for the kth time instant and the corresponding control uk , we seek the follow-
ing approximation for the successive state xk+1 :
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

xk+1 = Axk + Buk , (9.23)

or, in matrix form,

X 0 = AX + BI. (9.24)
The resolution method of finding A depends on whether the matrix B is known. For a complete
exposition of this extension, please refer to the foundational paper [463]. For an application to
the data-driven resolution of optimal control problems governed by PDEs, see [172].

Multiresolution DMD
Multiresolution DMD (mrDMD) [354] integrates multiresolution analysis into the classical DMD.
This method allows one to handle transient phenomena and structures which rotate or translate
in the domain, overcoming a known weakness of every SVD-based method.
The main idea is to iteratively select the slow modes for a given time window and then split
each subdomain along the temporal axis. This results in the slowest modes being retained at the
first level, while the fastest are in the last level. We can see the different performance of DMD
and mrDMD with a simple example where the dataset contains many features at varying time
scales, like oscillating sines and cosines, one-time events, and random noise. First we apply
the classical DMD without any SVD rank truncation, and then we try to reconstruct the data.
In Figure 9.3, in the middle panel, the results clearly show that all the transient time events are
missing. Applying instead mrDMD, we are able to properly reconstruct the whole dataset thanks
to its hierarchical nature (right panel of Figure 9.3).
For an extension and generalization of this particular method, the interested reader can refer
to [177].

Figure 9.3. On the left is the initial dataset containing features at different time scales. In the
middle is the DMD reconstruction, and on the right is the mrDMD reconstruction.

Compressed DMD
In the case of very high dimensional data, DMD can be computed on a compressed version of
the data, thus achieving great computational savings. This extension is called compressed DMD
(cDMD) [183].
 9.3. DMD for Time-Dependent Problems 211

The algorithm is very similar to the classical one, with the main difference just before the
computation of the modes, where we use a measurement matrix to compress the snapshot data.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

This is very useful, e.g., when dealing with video streams or in general when we have a prescribed
set of sensors. Despite the reconstruction error being slightly higher, the computational savings
are remarkable. Thus this DMD extension is useful even when we do not have any sensors, and
the compression matrix is randomly generated. In Figure 9.4 we compare the reconstruction
performance of DMD and cDMD.

Figure 9.4. On the left is the initial dataset, in the middle is the DMD reconstruction, and on the
right is the cDMD reconstruction.

In order to access the computational savings, we perform a simple test with an increasing
dimension of the snapshot matrix and compare the time needed to compute the DMD and the
cDMD modes, respectively. The results are shown in Figure 9.5. For further details, please refer
to tutorial6 number 4 of the PyDMD Python package.

14 Exact DMD
Compressed DMD
12
10
8
Seconds

6
4
2
0
0 20000 40000 60000 80000 100000
Snapshots dimension

Figure 9.5. Computational savings of cDMD with respect to the exact DMD for an increasing
dimension of the snapshot matrix.

6 Freely available at https://mathlab.github.io/PyDMD/tutorial4cdmd.html.

 212 Chapter 9. Nonintrusive Data-Driven ROMs in CFD

9.4 Applications in CFD

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

In this section we present different engineering problems where we applied the nonintrusive
ROMs introduced in the previous sections.

9.4.1 A ROM Pipeline for Shape Optimization Problems

This subsection presents the numerical results obtained by applying the aforementioned data-
driven techniques in an industrial shape optimization context. In naval engineering, a typical
shape optimization problem is to reduce the total resistance over the hull at a certain speed and
sea condition: we explain in this section the entire computational pipeline built to achieve this
goal, mainly focusing on ROM methods, but for an exhaustive description we recommend [154]
for the optimization of the bulbous bow and [152] for the optimization of the hull. Of course the
final goal is the computational reduction of the overall optimization procedure in a nonintrusive
fashion.
In more detail, we extend a typical shape optimization procedure by adding the DMD and
PODI techniques as shown in Figure 9.6. Such double reduction aims to minimize the compu-
tational load required to compute the high-fidelity snapshots: the Reynolds-averaged Navier–
Stokes model with the volume of fluid (VoF) method indeed needs a high computational effort
to be solved, especially on industrial geometries, which could make constructing the database
(for the parametric ROM) very expensive. Dealing with transient simulations, as in the de-
scribed naval application, allows one to apply the DMD to the high-fidelity model, considering
each parametric configuration as an independent dynamical system, resulting in an approxima-
tion of the regime state with a remarkable reduction in computational effort. The final solu-
tions and their corresponding parameters are then combined by means of the PODI technique,
constructing a data-driven ROM capable of approximating the new solutions for any untested
parameter.

ORIGINAL SHAPE FINITE

FULL ORDER
FINITE VOLUME
MODEL
VOLUME SNAPSHOT REDUCED ORDER
GEOMETRY PARAMETRIZATION DATABASE MODEL
OpenFOAM
RANS OpenFOAM
model with FV by
Ω FFD by PyGeM OpenFOAM ⎡ | | ⎤ POD-GPR by EZyRB/GPy
⎢ ⎥
= 0 …
⎢ ⎥
⎣ | | ⎦
DYNAMICMODE
DYNAMIC MODE
DECOMPOSITION
FORECASTING
DECOMPOSITION OPTIMIZATION

PyDMD
DMDPyDMD
by PyDMD ⎡ | | ⎤ GA by DEAP
⎢ ⎥
= 0 …
⎢ ⎥
⎣ | | ⎦

APPROXIMATED
Validation and enrichment OPTIMUM
∗ ∗
( , )
ROM

Figure 9.6. Representation of the computational pipeline used for the shape optimization problem.

The naval hull is parametrized by applying the free-form deformation (FFD) technique [533]
(see also Chapter 17) to the region of interest of the ship. A visual example of the deformations
obtained by morphing the bulb is presented in Figure 9.7, where the preservation of surface
derivative continuity is evident.
Once the deformation map has been constructed and the parameter bounds are set, we sample
the parameter space to obtain a limited set of deformed configurations, which are used to create
the database of high-fidelity snapshots. To compute such snapshots we numerically solve the full-
order model: since for these applications a Reynolds-averaged Navier–Stokes (RANS) model
 9.4. Applications in CFD 213
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 9.7. Example of bulbous bow deformations using the FFD method.

with VoF and turbulence is used, we apply a finite volume discretization using the OpenFOAM
library [612]. As already mentioned, the computational cost of these simulations is reduced by
involving the DMD to accelerate the convergence to regime state. The finite volume simulation
is carried out for a limited set of time steps, collecting the system states and using them to feed
the DMD algorithm. Assuming convergent or periodic behavior, DMD is able to capture such
dynamics and we can exploit the linear operator to predict a future state. Again, we remark that
this procedure is applied to all parametric samples independently.
After DMD, we obtained a snapshot for all the parametric samples and for any field of inter-
est. For this particular application, we collected the pressure, velocity, and VoF variable fields,
since they are required in order to compute the objective function. We applied the PODI algo-
rithm by decomposing the snapshot matrix by means of SVD and by approximating the solution
manifold in the reduced space. We construct this approximation using an RBF interpolation
[154] or a GPR [152]. From the latter work, we report in Figure 9.8 a summarizing comparison
between the different approximation techniques we tested: linear interpolator, RBF (with two
different values of the smoothing parameter, which controls the smoothness and thus the interpo-
latory nature of the regression), and GPR. GPR and (smoothed) RBF are the methods that show
the smallest error, but a remarkable difference is the higher ROM accuracy for GPR with few
input snapshots, allowing for a computationally cheaper offline phase.
The PODI algorithm was applied for the three aforementioned fields, resulting in a faster
evaluation of the total drag for any new geometrical deformation. Finally, this parametric model
can be used within an optimization cycle: the objective function is not computed any more on

0.1 0.1
relative L2 error

8 · 10−2
8 · 10−2

6 · 10−2
6 · 10−2

0 20 40 60 80 20 40 60 80
# modes # snapshots
GP linear RBF0 RBF0.01

Figure 9.8. Comparison of the average relative L2 error of the ROM in a naval shape optimiza-
tion problem as a function of the number of modes (left) and the number of snapshots used (right).
 214 Chapter 9. Nonintrusive Data-Driven ROMs in CFD

the simulation output, but using the PODI. The computational gain is huge: instead of requiring
several hours (using a high-performance computing facility) for a single iteration, the PODI
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

lasts only a few seconds for a new query. The aforementioned references [154, 152] show the
successful use of the genetic algorithm for the optimization task. The large number of evaluations
usually required by such evolutive procedures is compensated for by the extremely fast response.
We remark again that, thanks to the generality of the model, any optimization method can be
used as well.
Due to the missing error estimation, we have validated the optimal shape resulting from the
optimization procedure using the FOM. If the discrepancy error is above a certain threshold, we
refine the solution manifold by adding to the solution database the snapshot computed during the
validation. The ROM is then enriched with additional information, resulting in a more accurate
output: iteratively, a new optimization is performed using the new model until the PODI error
results are acceptable.

9.4.2 Comparison between DMD and PODI for Hydroacoustics

Problems
In this section we present an example of the application and comparison of DMD and PODI for
hydroacoustic problems [209]. Flow fields have been hydroacoustically analyzed with a Ffowcs
Williams and Hawkings (FWH) analogy [195]. The goal of this test is to investigate the use
of these two techniques on a hydroacoustic dataset obtained by simulating the turbulent flow
past a sphere at Re = 5000. The high-fidelity solutions are obtained using wall-resolved large
eddy simulation (LES) techniques. The overall procedure is summarized in Figure 9.9. The
high-fidelity data is generated with the OpenFOAM finite volume solver [612] using an equi-
spaced distribution in the time space running finite volume simulations with LES turbulence
modeling. The obtained snapshot matrix is then factorized using the SVD, and the results are
used to construct both the POD and the DMD spaces. The snapshot matrix is divided into two
Ns /2 Ns /2
different sets, a training one, Xtrain = {x2i−1 }i=1 , and a test one, Xtest = {x2i }i=1 . In the
PODI case, the scalar coefficients resulting from the projection of the training set on the space

Data acquisition
High-fidelity fluid dynamics FOM in parameter space

Basis functions
Modal decomposition using SVD

Construct ROM
DMD/POD + continuous modal coefficients

Midcast
Interrogate intermediate snapshots in time as parameter.

Acoustic evaluation
Analyze acoustic performance from prediction data.

Figure 9.9. Flowchart of the FOM/ROM procedure.

 9.4. Applications in CFD 215

spanned by POD modes are used to construct an interpolant function. Also in the DMD case
the operators of the dynamical system are constructed using only the training set. In both the
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

DMD and the PODI cases the accuracy of the ROM procedure is checked on the testing set. The
FOM is based on the incompressible Navier–Stokes equations coupled with a LES filtering tech-
nique, and the physical problem consists of the flow around a sphere of diameter D = 0.01 m,
immersed in a water stream with constant streamwise velocity U0 = 1 m/s. The kinematic
viscosity is ν = 2.0 × 10−6 m2 /s, so that the Reynolds number based on the sphere diameter
is Re = 5000.
The computational domain, depicted in Figure 9.10, consists of a box with dimensions 16D×
16D × 16D along the x-, y-, and z-axes, respectively. The sphere is centered along the y- and
z-axes. Along the x-axis it is at distance 12D from the outlet boundary. The boundary conditions
for the pressure field are set to zero-gradient on all boundaries with the exception of the outlet
boundary, where we set a uniform null value. The velocity is set to U0 at the inlet, the stress-free
condition is set at the lateral boundaries, and the zero-gradient condition is set for the velocity
components at the outlet. The grid, unstructured and body-fitted, consists of about five million
cells. Since no wall functions are adopted, the grid spacing normal to the wall for the densest
layer of
pcells (indicated by R5) has its first cell center within a wall unit y (y = uτ y/ν with
+ +

uτ = τw /ρ0 and τw the mean shear stress). The grid has five different refinement zones, as
depicted in Figure 9.10.

Figure 9.10. Sketch of the computational domain used for the FOM. The sphere diameter is
D = 0.01 m. Successive mesh refinement layers (R2, R3, R4) are performed using the cell splitting
approach until the finest grid spacing 0.001D is reached in the region R5.

The time integration is carried out using a constant time step of ∆t = 10−5 s, which is
enough to keep the Courant number under the threshold of 0.5. The flow around the sphere is
completely developed after about 80 characteristic times D/U0 .
Once the snapshots have been acquired, the techniques reported in Sections 9.2 and 9.3 are
used for PODI and DMD, respectively. In particular, for the PODI case we used a cubic spline
interpolation approach, while for the DMD case we used the classical algorithm. We then used
the PODI and DMD approaches to reconstruct the snapshots of the testing set, and we compared
 216 Chapter 9. Nonintrusive Data-Driven ROMs in CFD
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Modal Index

Figure 9.11. Normalized singular values for streamwise velocity and pressure snapshots.

the results with the high-fidelity data by changing the dimension of the POD and DMD spaces.
The eigenvalue decay resulting from the SVD of the training snapshots matrix is reported in Fig-
ure 9.11. As one can deduce from the picture, the problem is not characterized by a sudden drop
in the eigenvalues. This suggests the presence of a large number of different spatial scales. This
feature is a particular characteristic of high-fidelity LES simulations. A visual representation of
the identified POD and DMD modes for different index modes is reported in Figures 9.12 and
9.13 for velocity and pressure, respectively. In order to assess the accuracy of the procedure for
predicting hydroacoustic analysis, we postprocessed the reconstructed results using both a dipole
and a quadrupole approach [125, 126] and we compared the results coming from full-order snap-
shots. Figure 9.14 shows a quantitative comparison of the error of changing the dimension of
the POD and DMD spaces. In the picture we depict the time evolution of the L2 relative er-
ror for both the velocity and the pressure fields. As expected, for both approaches, increasing
the number of employed modes decreases the error. For this particular case the PODI approach
performs the best for both fields. Since the aim of this test was to assure the capability of this
type of ROM to be used for hydroacoustic analysis, in Figure 9.15 we compare the spectrum of
the hydroacoustic analysis for both the dipole and the quadrupole case. Both ROMs prove to be
accurate in predicting acoustic noise.

Velocity (DMD) Velocity (POD)

Figure 9.12. Isosurfaces of the DMD modes (left) at a value of −15 × 10−5 m/s versus the
isosurfaces of POD modes (right) at a value of 6.5 m/s for the streamwise velocity field. Modes: 1, 2, 8,
and 36.
 9.4. Applications in CFD 217

Pressure (DMD) Pressure (POD)

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 9.13. Isosurfaces of the DMD modes (left) at a value of −5 × 10−4 Pa versus the isosur-
faces of POD modes (right) at a value of 300 Pa for the pressure field. Modes: 1, 2, 8, and 36.

9.4.3 A Parametric View of DMD

In this subsection we show a parametric version of the DMD applied to the reconstruction of
a parametric scalar function [573]. More precisely, the focus is on the development of a non-
intrusive computational pipeline (Figure 9.16) to approximate the lift coefficient acting on a
geometrically parametrized wing profile. The methodology is based on the DMD coupled with
dynamic active subspaces (refer to Chapter 16 for an in-depth presentation of reduction in pa-
rameter space).

Figure 9.14. Relative error percentage in the Frobenius norm for velocity and pressure (top and
bottom, respectively) fields, where half-sample-rate snapshots are used to train the reduced model, DMD
(left) or PODI (right), for a particular truncation (r) while predicting the intermediate snapshots. Plots
show only the snapshotwise prediction error, while disregarding errors in the training set, which are almost
null.
 218 Chapter 9. Nonintrusive Data-Driven ROMs in CFD
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 9.15. Linear dipole (top) and nonlinear quadrupole (bottom) terms of FWH equation
evaluated from LES data and compared to corresponding DMD (left) and PODI (right) reduced models at
different truncation ranks (r).

Figure 9.16. Flowchart representing the proposed computational pipeline.

 9.4. Applications in CFD 219

The physical problem is the one of flow around a wing profile. The geometry of the wing
is parametrized using the parameter vector µ ∈ R10 , and the undeformed configuration corre-
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

sponds to the four-digit NACA 4412 wing profile. The deformed configuration is obtained by
summing five bump functions on the upper and lower part of the profile [283]. Defining the upper
and lower y coordinates of the wing profile by yu and yl and the upper and lower coordinates in
the undeformed configuration by yu and yl , the deformed configuration is given by
5
X 5
X
yu = yu + µi ri , yl = yl − µi+5 ri , (9.25)
i=1 i=1

where ri are bump functions as defined in [283], which are also reported in Figure 9.19. The
parameters used for the training are obtained with a uniform probability distribution in the pa-
rameter space D := [0, 0.03]10 ⊂ R10 . Once the coordinates of the deformed configuration are
obtained, the mesh is deformed using an RBF interpolation technique. We refer the reader to
[573] for more details on the deformation approach. The high-fidelity computational model is
based on the incompressible Navier–Stokes equations with RANS turbulence modeling. Specifi-
cally, the Spalart–Allmaras turbulence model has been employed [546]. The discretized problem
is obtained by means of the finite volume method using the OpenFOAM [612] and the ITHACA-
FV volume packages [550]. A schematic view of the computational domain, together with a
zoom on the mesh near the wing profile, is reported in Figure 9.17. The mesh counts 46500
hexahedral cells. At the inlet we have a constant and uniform velocity uin = 1 m/s. The kine-
matic viscosity equals ν = 2 × 10−5 m2 /s, which considering a chord length of D = 1 m,
produces a Reynolds number of Re = 50000. The pressure velocity coupling was resolved using
the PIMPLE algorithm as implemented in the OpenFOAM library. To advance the simulations
in time we used a constant time step of ∆t = 10−3 s. Each configuration is simulated for a total
time of T = 30 s. In Figure 9.18 we report the temporal evolution of the lift coefficient for nine
different values of the input parameter.

7.5C C 16C

16C

24.5C

Figure 9.17. Sketch of the computational domain used to solve the fluid dynamics problem in its
reference configuration. The left plot is a zoom on the mesh near the wing. The right picture is a schematic
view of the domain with the main geometrical dimensions.

Parameter Space Reduction

The possibility of exploring the solution manifold with a limited number of simulations during
the training stage is highly correlated with the number of parameters. By increasing the number
of parameters it is in fact easy to fall victim to the curse of dimensionality. It is then crucial to
keep the number of parameters as low as possible. In this section we show a possible approach
 220 Chapter 9. Nonintrusive Data-Driven ROMs in CFD
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 9.18. The temporal evolution of the lift coefficient from 1 s to 30 s for nine different parameters.

NACA profile
0.2 r1
r2
0.1 r3
y

r4
0.0 r5

−0.1
0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4
x

Figure 9.19. Airfoil shape functions with respect to the profile abscissa. The leading edge corre-
sponds to x = 0. The bump functions are rescaled by a factor of 0.2 for illustrative reasons.

to reducing the number of parameters using the property of active subspaces (ASs) [134]. In
Figure 9.20 we show the DyAS results for the lift coefficient output that aim to show the effec-
tiveness of the AS over time. The left diagrams show, at different time instants, the lift coefficient
at each sample point as a function of the first active variable, which is obtained by a linear combi-
nation of the sample point coordinates in the parameter space. We underline that the eigenvector
components of all the time instants presented, corresponding to the coefficients c1 , c5 , d1 , and
d5 , are almost zero. This means that on average the lift coefficient is almost flat along these
directions. We will exploit this fact by freezing these parameters and constructing a GPR on a
reduced parameter space.

GPR Approximation and Prediction of the Lift Coefficient

With the previous analysis we were able to discover the presence of several input parameters
with minimal influence on the output function (the lift coefficient). This information was used
to construct a surface response which is based only on the relevant input parameters. The ap-
proximation was constructed using a GPR technique with RBF kernel exploiting the GPy library
[239]. The regression with GPR was constructed using both the full and the reduced parameter
space, and results were compared over a test set of 100 samples. The training set is the same for
both the reduced and the full parameter space case and consists of 70 samples. For the first 20 s
 9.4. Applications in CFD 221
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 9.20. On the left is the sufficiency summary plot for the lift coefficient at times t =
10.0, 14.0, 18.0 s (top to bottom). On the right are the first eigenvector components at the corresponding
parameters.

we used the high-fidelity data, while the last 10 seconds of the simulation were approximated
using DMD to speed up convergence. The DMD model was constructed using the high-fidelity
results in the temporal interval [12, 20] s. The DMD approximation was constructed using 8000
temporal pieces of information (∆t = 0.001 s) with 10 DMD modes. Figure 9.21 summarizes
the different configurations analyzed. In particular we have the GPR regression in the full param-
eter space and in the reduced parameter space, and the same results using a DMD approximation
for the last 10 s of the simulation.
 222 Chapter 9. Nonintrusive Data-Driven ROMs in CFD

·10−2
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

8 GPR
GPR reduced
GPR + DMD
6 GPR + DMD reduced
Relative Error

0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30
Time [s]

Figure 9.21. The relative error of the approximated outputs at different times. The relative error
is computed on 100 test samples, using the high-fidelity lift coefficient to train the regression for t ≤ 20 s,
while for t > 20 s the DMD forecasted states are used for the training.

9.5 Conclusions and Future Perspectives

Nonintrusive data-driven ROMs are a powerful set of techniques especially when dealing with
industrial contexts. In this chapter we presented a general framework for PODI to resolve
parametrized problems and DMD dynamical systems. We presented a diverse range of appli-
cations, including shape optimization of bulbous bows and the analysis of a parametrized NACA
airfoil profile, and we also compared these methods for hydroacoustics analysis.
Other techniques have been developed in recent years, such as sparse identification of non-
linear dynamical systems [80] or other numerical methods which are nonintrusive despite losing
their data-driven nature [645]. We also mention other approaches for differential-algebraic equa-
tions derived from lifting transformations [341], or localized ROMs in the operator inference
framework [217].
We think the field of data-driven ROMs will play a crucial role in enabling the construction
of digital twin models of industrial artifacts, without disrupting the existing numerical pipelines.
ROMs also allow fast prototyping and reduced time to market since they can be used in the
design phase in many different engineering fields.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Chapter 10

Spectral Element
Method–Based Model
Order Reduction

Martin W. Hess, Gianluigi Rozza

The spectral element method (SEM) uses high-order polynomials defined over a coarse spatial
grid to numerically resolve a function. This is in contrast to, e.g., the widely adopted finite
element method (FEM), which uses low-order polynomials but a very dense spatial discretization
grid. This chapter introduces the main concepts of model reduction with the SEM, which was
first introduced in [439]. Beginning with the polynomial spaces used as ansatz spaces, it also
shows the global assembly and points out some particularities of the technical implementations
with the partial differential equation (PDE) solver Nektar++ and the model reduction software
ITHACA-SEM. This is an introduction with several references to the literature, where the topic
is analyzed more deeply. Overview books are [96, 97, 166, 330].

10.1 Basic Notions and Functions of the SEM

The first section reviews relevant polynomials and quadrature rules and introduces technical
terms and notions used in the SEM literature. Since the high-order polynomials are used to
resolve solutions to PDEs, several adaptations, such as a boundary-interior decomposition and
quadrature points on the boundary, are required.

10.1.1 Polynomials and Quadrature Rules

The Jacobi polynomials are presented as a general class of polynomials from which many practi-
cally relevant polynomial families can be derived. Jacobi polynomials in turn can be understood
within the Askey scheme of hypergeometric orthogonal polynomials; see [330], Appendix A.

Jacobi Polynomials
Jacobi polynomials Pnα,β (x) (with α, β > 1) are the solutions to the singular Sturm–Liouville
problem
 
d 1+α 1+β d
(1 − x) (1 − x) P (x) = λn (1 − x)1+α (1 − x)1+β Pnα,β (x), (10.1)
α,β
dx dx n
λn = −n(α + β + n + 1). (10.2)

223
 224 Chapter 10. Spectral Element Method–Based Model Order Reduction

Jacobi polynomials can be computed with the Rodriguez formula and the following two
recursion relations.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Rodriguez Formula

(−1)n dn 
Pnα,β (x) = (1 − x)−α (1 + x)−β n (1 − x)α+n (1 + x)β+n . (10.3)

n
2 n! dx

Recursion Relations

P0α,β (x) = 1, (10.4)

1
P1α,β (x) = (α − β + (α + β + 2)x), (10.5)
2
α,β α,β
a1n Pn+1 (x) = (a2n + a3n x)Pnα,β (x) − a4n Pn−1 (x), (10.6)
a1n = 2(n + 1)(n + α + β + 1)(2n + α + β), (10.7)
a2n 2
= (2n + α + β + 1)(α − β ),2
(10.8)
a3n = (2n + α + β)(2n + α + β + 1)(2n + α + β + 2), (10.9)
a4n = 2(n + α)(n + β)(2n + α + β + 2), (10.10)

d α,β α,β
b1n (x) P (x) = b2n (x)Pnα,β (x) + b3n (x)Pn−1 (x), (10.11)
dx n
b1n (x) = (2n + α + β)(1 − x2 ), (10.12)
b2n (x) = n(α − β − (2n + α + β)x), (10.13)
b3n (x) = 2(n + α)(n + β). (10.14)

Orthogonality
Jacobi polynomials are orthogonal with respect to (1−x)1+α (1−x)1+β over the interval [−1, 1]:
Z 1
(1 − x)1+α (1 − x)1+β Ppα,β (x)Pqα,β (x)dx = Cδpq . (10.15)
−1

Derived Polynomials
Special cases of derived polynomials are, e.g.,
• Legendre polynomials (α = β = 0), i.e., Ln (x) = Pn0,0 (x), and
1 1
22n (n!)2 − 2 ,− 2
• Chebyshev polynomials (α = β = − 12 ), i.e., Tn (x) = (2n)! Pn (x).

Quadrature Rules
Quadrature rules approximate the integral of a function in the elemental domain (here the interval
[−1, 1]) as
Z 1 Q−1
X
u(ξ)dξ ≈ wi u(ξi ). (10.16)
−1 i=0

In contrast to the FEM, the integrand u(ξ) is a high-order polynomial and thus requires
quadrature rules which are sufficiently accurate.
There are three different types of Gauss quadrature, known as Gauss, Gauss–Radau, and
Gauss–Lobatto. The difference between the three types of quadrature lies in the choice of the
 10.1. Basic Notions and Functions of the SEM 225

zeros. Gauss quadrature uses zeros which have points that are interior to the interval, −1 < ξi <
1 for i = 0, . . . , Q − 1. In Gauss–Radau the zeros include one of the endpoints of the interval,
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

usually ξ = −1, and in Gauss–Lobatto the zeros include both endpoints of the interval, i.e.,
α,β
ξ = +1, −1. Let ξi,P denote the P zeros of the P th-order Jacobi polynomial PPα,β (x).

Gauss–Legendre Quadrature
0,0
ξi = ξi,Q , i = 0, . . . , Q − 1, (10.17)
 −2
2 d
wi0,0 = (L Q (ξ))| ξ=ξi , i = 0, . . . , Q − 1. (10.18)
1 − (ξi )2 dξ

The Gauss–Legendre quadrature is exact for polynomial orders up to 2Q − 2.

Gauss–Radau–Legendre Quadrature

−1, i = 0,

ξi = 0,1 (10.19)
ξi−1,Q−1 , i = 1, . . . , Q − 1,
1 − ξi
wi0,0 = , i = 0, . . . , Q − 1. (10.20)
Q2 [L Q−1 (ξi )]
2

The Gauss–Radau–Legendre quadrature is exact for polynomial orders up to 2Q − 2.

Gauss–Lobatto–Legendre Quadrature

 −1, i = 0,
ξi = ξ 1,1 , i = 1, . . . , Q − 2, (10.21)
 i−1,Q−2
1, i = Q − 1,
2
wi0,0 = i = 0, . . . , Q − 1. (10.22)
Q(Q − 1)[LQ−1 (ξi )]2

The Gauss–Lobatto–Legendre quadrature is exact for polynomial orders up to 2Q − 3.

Aliasing
Aliasing occurs when the integrand is of higher order than the quadrature rule can resolve. An
approximation error will thus occur. In many implementations only the linear terms are integrated
exactly. Nonlinear terms, like in the Navier–Stokes equations, are then underresolved in favor of
computational speed.

10.1.2 Expansion Functions

The term spectral/hp method is derived from two possible refinement strategies, h-type and p-
type. h-type refinement increases the number of spatial elements, i.e., increases spatial resolution.
Here, h refers to the characteristic size of an element. In complex geometries, some h-type
refinement is often required. p-type refinement keeps the spatial grid intact but increases the
polynomial order. If the whole domain is just a single element, then p-type refinement is also
known as a (global) spectral method.
To define a basis on a given discretization, elemental basis functions are defined on simplices
(triangles, quadrilaterals) and then mapped onto the actual elements. This expansion is modi-
fied in the process for desirable numerical features, i.e., an orthogonal or near-orthogonal set is
 226 Chapter 10. Spectral Element Method–Based Model Order Reduction

chosen and the numerical efficiency and the condition numbers of resulting linear systems are
considered. In particular, a boundary-interior decomposition is used in the SEM, where “bound-
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

ary” also refers to internal boundaries given by the h-refinement and not just the boundary of the
actual domain.
Two frequently used terms in the field of SEM are nodal and modal expansions.

Modal Expansions
Given a set of polynomial functions {Φp , 0 ≤ p ≤ P }, a modal expansion set refers to a hierar-
chical expansion set, i.e., the expansion set of order P contains the set of order P − 1.
An example of a modal expansion is the set of monomials {xp , p = 0, . . . , P }. It is called
the moment expansion.

Nodal Expansions
In a nodal expansion {Φp , 0 ≤ p ≤ P }, the expansion coefficients are the values at certain nodes
{xp , 0 ≤ p ≤ P }, i.e., Φp (xq ) = δpq .
An example of a nodal expansion is the Lagrange polynomials
ΠP
q=0,q6=p (x − xq )
Φp (x) = , p = 0, . . . , P. (10.23)
ΠP
q=0,q6=p (xp − xq )

Modal Expansion Bases on Quadrilaterals

The one-dimensional expansion functions can be extended to higher dimensions by a tensor
product extension

φpq (ξ1 , ξ2 ) = φp (ξ1 )φq (ξ2 ), (10.24)

φpqr (ξ1 , ξ2 , ξ3 ) = φp (ξ1 )φq (ξ2 )φr (ξ3 ). (10.25)

Let the Jacobi polynomials be denoted by Pnα,β (x). The reference quadrilateral in reference
coordinates ξ1 , ξ2 is defined with the midpoint (ξ1 , ξ2 ) = (0, 0). The vertices, i.e., corners are
(ξ1 , ξ2 ) = (−1, −1), (ξ1 , ξ2 ) = (−1, +1), (ξ1 , ξ2 ) = (+1, −1), and (ξ1 , ξ2 ) = (+1, +1).
A modal basis over the reference quadrilateral is given by vertex-related expansion bases
1 − ξ1 1 − ξ2
Φ= , (10.26)
2 2
1 − ξ1 1 + ξ2
Φ= , (10.27)
2 2
1 + ξ1 1 − ξ2
Φ= , (10.28)
2 2
1 + ξ1 1 + ξ2
Φ= ; (10.29)
2 2
edge-related expansion bases
1 − ξ1 1 + ξ1 1,1 1 − ξ2
Φ= Pp−1 (ξ1 ) , 0 < p < P1 , (10.30)
2 2 2
1 − ξ1 1 + ξ1 1,1 1 + ξ2
Φ= Pp−1 (ξ1 ) , 0 < p < P1 , (10.31)
2 2 2
1 − ξ1 1 − ξ2 1 + ξ2 1,1
Φ= Pp−1 (ξ2 ), 0 < p < P2 , (10.32)
2 2 2
1 + ξ1 1 − ξ2 1 + ξ2 1,1
Φ= Pp−1 (ξ2 ), 0 < p < P2 ; (10.33)
2 2 2
 10.2. Assembly of System Matrices 227

and interior expansion bases

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

1 − ξ1 1 + ξ1 1,1 1 − ξ2 1 + ξ2 1,1
Φ= Pp−1 (ξ1 ) Pp−1 (ξ2 ), 0 < p < P1 , 0 < q < P2 . (10.34)
2 2 2 2
More explicitly given expansion bases can be found in Appendix D of [330].

Expansion Bases on Triangles

Using a tensorial expansion on nonrectilinear elements (triangles, tetrahedrons, curved) requires
collapsed coordinates. This introduces a singularity: the whole upper boundary is mapped to
(−1, 1) on the triangle, which needs to be treated with appropriate quadrature rules. It defines
the new coordinates η1 , η2 as
1 + ξ1
η1 = − 1, (10.35)
1 + ξ2
η2 = ξ2 . (10.36)

A change to cylindrical coordinates shows that for ξ1 = −1, ξ2 = 1, we have η1 ∈ [−1, 1].

10.2 Assembly of System Matrices

Like the FEM, the SEM can also be interpreted as an instance of the method of weighted residu-
als.

10.2.1 Method of Weighted Residuals

Consider a linear differential equation in a domain Ω as

L(u) = 0, (10.37)

and consider an approximate solution

Ndof
X
uδ (x) = ûi Φi (x). (10.38)
i=1

Define the residual as L(uδ ) = R(uδ ).

Compute the weighted residuals as
Z
0 = (vj , R) = vj (x)R(x)dx, j = 1, . . . , Ndof , (10.39)
Ω

to obtain the scalar coefficients ûi .

Galerkin Orthogonality
The test space is orthogonal to the residual, which is the interpretation of (10.39).

10.2.2 Elemental Properties

In a spectral/hp discretization, the solution domain is decomposed to capture broadly the geome-
try and local scale of the problem (h-type expansion), and then increasing polynomial orders are
used (p-type expansion) for accuracy.
 228 Chapter 10. Spectral Element Method–Based Model Order Reduction

Define the discrete hp expansion space χδ as

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

χδ = {uδ |uδ ∈ H 1 , uδ (χe (ξ)) ∈ Ppe (Ωst ), e = 1, . . . , Nel } (10.40)

for discrete functions uδ in the Sobolev space H 1 , which implies continuity, i.e., uδ ∈ C 0 . The
mapping between reference simplex and actual element is denoted by χe (ξ).
Since global modes Φi (x) can be expressed in terms of local expansions φe (ξ), it holds that

Ndof Nel X
P Nel X
P
X X X
uδ (x) = ûi Φi (x) = ûep φep (ξ) = ûep φep ((χe )−1 (x)). (10.41)
i=1 e=1 p=0 e=1 p=0

The relationship between local and global degrees of freedom is expressed with a scattering
matrix A as ûl = Aûg from global to local degrees of freedom and a corresponding assembly
matrix, which maps from local to global degrees of freedom.
An important feature of ansatz functions is the sparsity pattern generated. It can be assessed
by looking at the projection problem. The projection problem reads as follows: find uδ ∈ χδ
such that
(v δ , uδ ) = (v δ , f ) ∀v δ ∈ χδ . (10.42)

This leads to the linear system M û = f with Mpg = (Φp , Φq ) and fp = (Φp , f ). For a moment
expansion, i.e., the set of monomials, the mass matrix M is 50% full; in a Lagrange expansion
the mass matrix M is full; and in a Legendre expansion the mass matrix M is diagonal.
A boundary-interior decomposition requires uδ ∈ H 1 to impose continuity across element
boundaries. But an interface-matching condition potentially couples all degrees of freedom:

P
X P
X
ûep φep (1) = ûe+1 e+1
p φp (−1). (10.43)
p=0 p=0

Instead, a division into boundary modes and interior modes, such as that given by the Lagrange
expansion

φp (−1) = 1, p = 0, (10.44)
φp (−1) = 0, p 6= 0, (10.45)
φp (1) = 1, p = P, (10.46)
φp (1) = 0, p 6= P, (10.47)

such that
ûeP φeP (1) = ûe+1 e+1
0 φ0 (−1) (10.48)

guarantees continuity, i.e., uδ ∈ C 0 .

10.3 Reduced Order Modeling with the SEM

This section focuses on SEM-based reduced order modeling (ROM) with the reduced basis
method (RBM) and proper orthogonal decomposition (POD). A recent overview of ROM meth-
ods can be found in [53, 54].
Since ROM methods can typically be adapted to work with any discretization method, the
underlying methodology does not change with respect to the SEM. The SEM has been used in
 10.3. Reduced Order Modeling with the SEM 229

ROM work, e.g., in [457, 458, 273, 270]. The SEM solver used in the work presented here is the
spectral/hp element software Nektar++ [95].
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

10.3.1 Examples in Computational Fluid Dynamics

Let Ω ∈ R2 denote the spatial computational domain. Incompressible, viscous fluid motion in a
domain Ω over a time interval (0, T ) is governed by the incompressible Navier–Stokes equations.
Denoting the velocity u, pressure p, kinematic viscosity ν, and body forcing f , the Navier–Stokes
equations read
∂u
+ u · ∇u = −∇p + ν∆u + f, (10.49)
∂t
∇ · u = 0. (10.50)

Boundary and initial conditions are given as

u=d on ΓD × (0, T ), (10.51)

∇u · n = g on ΓN × (0, T ), (10.52)
u = u0 in Ω × 0, (10.53)

with d, g, and u0 assumed known and ∂Ω = ΓD ∪ ΓN , ΓD ∩ ΓN = ∅. The Reynolds number Re

depends on the kinematic viscosity ν via the characteristic velocity U and characteristic length
L as Re = UνL .
When computing the steady states for varying viscosity ν, such that ∂u ∂t = 0, a solution
u(ν0 ) for a parameter value ν0 is used as an initial guess for the fixed-point iteration to obtain
the steady-state solution u(ν1 ) at a parameter value ν1 , provided that the solution u(ν) depends
continuously on ν in the interval [νmin , νmax ]. This is similar to continuation methods used for
following bifurcation branches; see [454].
The Oseen iteration is a secant modulus fixed-point iteration which exhibits a linear rate of
convergence. Provided with a current iterate (or initial condition) uk , the linear system

−ν∆u + (uk · ∇)u + ∇p = f in Ω, (10.54)

∇ · u = 0 in Ω, (10.55)
u=d on ΓD , (10.56)
∇u · n = g on ΓN (10.57)

is solved for the next iterate uk+1 = u. A common stopping criterion is that the relative change
between iterates in the H 1 -norm falls below a given tolerance. The initial solution u0 (ν0 ) is
computed by time advancement of (10.49)–(10.50) from zero initial conditions at a parameter
value ν0 . From this starting point, solutions on the whole parameter domain can be found.
The discretized system solved in each step of the Oseen iteration is given by (10.58) as
T
    
A −Dbnd B vbnd fbnd
 −Dbnd 0 −Dint   p  =  0  , (10.58)
T T vint fint
B̃ −Dint C

where vbnd and vint denote velocity degrees of freedom on the boundary and in the interior,
respectively. The forcing terms fbnd and fint refer to the boundary and interior, respectively.
The matrix A assembles the boundary-boundary terms, B the boundary-interior terms, B̃ the
interior-boundary terms, and C the interior-interior terms of elemental velocity ansatz functions.
In a Stokes system, B = B̃ T , but this is not the case here for the Oseen equation, since the
 230 Chapter 10. Spectral Element Method–Based Model Order Reduction

linearization term (uk · ∇)u is present in (10.54). The matrices Dbnd and Dint assemble the
pressure-velocity boundary and pressure-velocity interior couplings, respectively.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

The linear system (10.58) is assembled in local degrees of freedom, leading to block matrices
A, B, B̃, C, Dbnd , and Dint , with each block referring to a single spectral element. It follows that
the system is singular in this form. To solve the system, the local degrees of freedom need to be
gathered into the global degrees of freedom. Since C contains the interior-interior contributions,
it is invertible and the system can be statically condensed into

A − BC −1 B̃ T BC −1 Dint
T T
  
− Dbnd 0 vbnd
 Dint C −1 B̃ T − Dbnd −1 T
−Dint C Dint 0  p 
B̃ T T
−Dint C vint
fbnd − BC −1 fint
 

= Dint C −1 fint  . (10.59)

fint

Taking the top left 2×2 block and reordering the degrees of freedom, such that the mean pressure
mode of each element is inserted into the corresponding block of Â, leads to

fˆbnd
    
 B̂ b
= , (10.60)
Ĉ D̂ p̂ fˆp

with D̂ invertible. This means that a second level of static condensation can be used. It follows
that

 − B̂ D̂−1 Ĉ 0 fˆbnd − B̂ D̂−1 fˆp

    
b
= . (10.61)
Ĉ D̂ p̂ fˆp

The vector b contains the velocity boundary degrees of freedom and the mean pressure modes,
while the remaining solution components are computed by reverting the transformations of the
static condensations. The largest computational effort is in solving the final system:

(Â − B̂ D̂−1 Ĉ)b = fˆbnd − B̂ D̂−1 fˆp . (10.62)

The matrices C and D̂ also have to be inverted, which because of the elemental block struc-
ture requires inverting submatrices of the size of the degrees of freedom per element in each
submatrix.

10.3.2 Reduced Order Model

The reduced order model (ROM) should approximate the full-order solutions accurately over the
parameter domain of interest. Two ingredients are common in reduced basis (RB) modeling, a
projection onto a low-order space of snapshot solutions (i.e., full-order solutions) and an offline-
online decomposition to guarantee computational efficiency. A set of snapshots is computed
by solving (10.62) over a coarse sample of the parameter domain and used to define a projec-
tion space U of size N using the standard solver provided by Nektar++. The POD computes
a singular value decomposition (SVD) of the snapshot solutions to 99.9% of the most domi-
nant POD modes. This defines a projection matrix U ∈ RNδ ×N . The open-source software
ITHACA-SEM7 is used to compute the ROM.
7 See https://mathlab.sissa.it/ITHACA-SEM.
 10.3. Reduced Order Modeling with the SEM 231

The offline-online decomposition allows for fast input-output solves, because they are inde-
pendent of the original model size Nδ . It is an important part of efficient ROM, but since the
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

static condensation includes the inversion of the parameter-dependent matrix C, the projection is
applied to the first system, (10.58). Other options are possible by gathering part of (10.58) into
global degrees of freedom; see [276]. In the offline phase, snapshot solutions have been gathered
over the parameter domain, which now serves as a projection space to define the reduced order
setting. To have fast reduced order solves, the offline-online decomposition expands (10.58) in
the parameter of interest and stores the parameter-independent projections as small matrices of
order N × N . Since during the Oseen iteration each matrix is dependent on the previous iterate,
the submatrix corresponding to each basis function is assembled and then formed online with
the RB coordinate representation of the current iterate. This is analogous to reduced order as-
sembly of the nonlinear term in the Navier–Stokes equations. For more details and applications,
see [273].

ROM: A Numerical Example

Consider a variable kinematic viscosity in the interval ν ∈ [0.05; 10] for a channel flow, as
depicted in Figures 10.1 and 10.2. A parabolic inflow profile is prescribed on the left wall at y =
0, a natural outflow boundary is prescribed at y = 8, and the other walls are no-slip boundaries.
Figures 10.1 and 10.2 show the extreme cases considered here for very small viscosity and very
large viscosity.

Figure 10.1. Full-order, steady-state solution for ν = 10: velocity in the x direction (top) and y
direction (bottom).

Sample solutions are computed at 13 values of ν in the interval of interest. The POD is
computed and the POD energy reaches a threshold of 99% with three modes and a threshold of
99.99% with seven modes. Ideally, the exponential decay in POD energy should translate into
an exponential decay in relative approximation error. Figure 10.3 shows the mean and maximum
relative L2 (Ω) error in the velocity as an increasing number of POD modes is considered. A
mean relative error of 1% is reached with four basis functions, and a maximum relative error of
1% is reached with five basis functions.
 232 Chapter 10. Spectral Element Method–Based Model Order Reduction
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 10.2. Full-order, steady-state solution for ν = 0.05: velocity in the x direction (top) and
the y direction (bottom).

Mean error
Maximum error
relative L2 (Ω) error in velocity

−1
10

10−3

10−5

10−7

10−9
0 2 4 6 8 10 12
ROM basis size

Figure 10.3. Mean and maximum relative L2 (Ω) error in the velocity with increasing basis size.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Chapter 11

Discontinuous
Galerkin-Based Reduced
Order Models

Andrea Lario, Francesco Romor, Gianluigi Rozza

11.1 Introduction
In this chapter the theory and applications of intrusive reduced order methods based on the
discontinuous Galerkin method (DGM) are presented. We focus on presenting the numerical
method and on showing how techniques which are used for the finite volume method (FVM)
and the finite element method (FEM) can also be employed in this framework in order to design
reduced order models (ROMs).
The DGM is a numerical method for the approximation of partial differential equations
(PDEs) and it is well suited for conservative laws and, in particular, for predicting phenom-
ena in which discontinuities and steep gradients are involved. For example, convection-diffusion
problems may present boundary layers in the immediate proximity of bounding surfaces char-
acterized by steep gradients. Propagating shock waves with sharp discontinuities may appear
in the solutions of nonlinear conservation laws, e.g., in the case of the Navier–Stokes equations
describing compressible flows.
The DGM can be considered as a generalization of the FVM and the FEM since it tries to
combine the best features of both.
Two key features of the DGM are the local conservativeness when approximating conserva-
tion laws, achieved by introducing the numerical flux, and the local formulation of the method.
Each element can be thought of as autonomous from the others, so the order of the method and
the relative number of nodes may change independent of the neighboring elements; this makes
the implementation relatively simple and suitable for p-adaptivity with less effort. In fact, unlike
the FEM, continuity is not imposed on the elemental boundaries, which in the DGM formulation
can have different nodes depending on the side considered. Moreover, all the basis functions
of the discretized finite space of the solutions have support in only one single element and the
fluxes can be evaluated with just the information from the neighboring cells that constitute the
compact stencil; this property of the DGM makes it efficient to run in parallel, since the amount
of information that has to be exchanged between elements is limited. The DGM is also suit-
able for implementing solvers for problems with complex geometries, nonstructured meshes,
and curved elements.
The main drawback of the DGM is the increment of the number of degrees of freedom used
to obtain the same accuracy as the FEM, but this additional cost is well balanced by its benefits.

233
 234 Chapter 11. Discontinuous Galerkin-Based Reduced Order Models

In fact, with respect to the FEM with stabilization, the DGM introduces less numerical viscosity,
thus following the physical solution more tightly.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

In addition, FEM may suffer from the Gibbs phenomenon if no advanced stabilization tech-
nique is involved, while DGM counters this problem with techniques like flux limiters. Like
the FVM, the DGM searches for piecewise-continuous solutions defined on every element of
the computational domain. This allows a better resolution of shock waves and contact discon-
tinuities. FVM, unlike DGM, uses only low-order approximations with a consequent loss of
accuracy; on the other hand, DGM, as a high-order method, is more susceptible to numerical
instabilities: underresolved simulations typically blow up.
A complete introduction to the DGM and a dissertation on its stability properties can be
found in [282].
The first formulation of the DGM dates back to 1973, proposed by Reed and Hill [483]
to solve equations which describe neutron transport. The first applications of the DGM were
mainly related to hyperbolic systems of equations. Lasaint and Raviart [357] provided a first a
priori error analysis for linear hyperbolic systems, while Cockburn and Dong [131] later analyzed
nonlinear systems. The Euler set of equations was solved using a Runge–Kutta (RK) method for
the time integration, while the discontinuities of the solution on the elemental interfaces were
treated by evaluating the fluxes between neighboring cells with the aid of techniques originally
developed in the finite volume framework. Then, Hu et al. [294] analyzed the properties of
the DGM for linear problems (i.e., wave propagation phenomena). In the mid-nineties DGM
was used to solve problems involving compressible flows; notable were the works of Bassi and
Rebay, which considered both the Euler equations [43] and the Navier–Stokes equations [44].
A complete analysis of DGM for applications related to compressible flows can be found in
[171]. Some later works by Cockburn et al. provided error estimates, numerical stability, and
convergence analysis for the DGM for convection-diffusion problems [131]. Current efforts are
focused on efficient strategies for implementations and on the development of new algorithms in
order to reduce the computational costs of DGM-based codes [441].
Up to now, only a few works have dealt with model reduction for codes based on the DGM,
and in general techniques which were originally proposed for either FVM or FEM have been
adapted to this different framework. Nevertheless, some notable works about ROMs for DGM
can be found in the literature [632, 19].
The chapter is structured as follows: the following three sections are devoted to the presen-
tation of the nodal DGM; in Subsection 11.3.2 a reduced order method is introduced; results are
presented in the last section.

11.2 Nodal DGM

Many different formulations based on the DGM have been proposed in the literature. In this
chapter one of the most popular is described, but the main considerations reported here still hold
for other DGM-based formulations.
Let’s consider the set of equations

∂U
+ ∇ · F (U, Q) = S, (11.1)
∂t
where U is the state variable vector, Q is the vector containing the derivatives of the state vari-
ables, F is the matrix containing the terms related to the fluxes, and S is the source term.
Many different sets of equations can be written in the form reported above, including the
Euler equations, the Maxwell equations, the Navier–Stokes equations, the linearized Navier–
Stokes equations, and the linearized Euler equations.
 11.2. Nodal DGM 235

Proper initial conditions U0 and boundary conditions UΣ must be set in order to have a
well-posed problem:
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

U = UΣ on Σ,
U = U0 for t = 0.
The domain Ω is approximated by a partition of nonoverlapping elements Ωk as
M
[
Ω ≈ Ωh = Ωk , (11.2)
k=1

and within each element Ωk the solution is approximated by the following nodal expansion [280]
for i ∈ {1, 2, 3}:
N
X N
X
Ukh (x, t) uk (xl , t) i Lkl (x) i = ukl (t) i Lkl (x) i , (11.3)






i
=
l=0 l=0

where N is the number of nodes inside the element Ωk ; Lkl (x) are the multivariate Lagrange
interpolation polynomials defined by the points xl in the element Ωk ; and ukl (t), the expansion
coefficients, are the values of the solution at the nodal points.
Applying a Galerkin projection in (11.1) onto each member of the basis set Lki , the weak
form of the problem can be written for each element Ωk as
∂Ukh
Z   Z
k k k
Li · + ∇ · Fh dΩ = Lki · Skh dΩk , (11.4)
Ωk ∂t Ωk
where the source term is approximated using the basis sets, and likewise for the conservative
variables (11.3), and the flux term is approximated with a polynomial matrix basis Mlk (x) for
i, j ∈ {1, 2, 3}:
N
X N
X
Skh skl Lkl Fhk Flk Mlk (11.5)







i
= i i
, ij
= ij ij
.
l=1 l=1

Integrating (11.4) by parts, one obtains

k
k ∂Uh
Z Z Z Z
k k k k k k k
Li · dΩ − ∇Li : Fh dΩ + Li · Fh n dΣ = Lki · Skh dΩk , (11.6)
Ωk ∂t Ωk Σk Ωk

where Σk indicates the boundary limiting Ωk and n is the outward-pointing unit normal to each
edge.
With expansions (11.3) and (11.5), setting
Nedge
k
X
Fhk ·n= (F · n)l Lkl , (11.7)
l=1

where Nedge is the number of nodes on the element edges and Lkl is the restriction of the poly-
nomial basis onto the edge, the weak formulation (11.6) reads
 
N k
! N
! Nedge
du
Z X Z X Z X
Lki · l
Lk dΩk − ∇Lki · Flk Lkl dΩk + Lki ·  Flk Lkl · n dΣk
Ωk dt l Ωk Σk
l=1 l=1 l=1
N
Z !
X
= Lki · skl Lkl dΩk , (11.8)
Ωk l=1

which holds for every element of the domain.

 236 Chapter 11. Discontinuous Galerkin-Based Reduced Order Models

The standard DGM is formulated to approximate first-order derivatives of the primitive vari-
ables. The viscous components of the fluxes of the Navier–Stokes equations show a dependence
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

up to the second-order derivatives of the conservative variables. Among the several extensions
of the DGM to convection-diffusion problems with higher-order derivatives, following Yan and
Shu [630], it is possible to introduce an auxiliary Q = ∇U. The global system of equations for
the variables U and Q on the computational domain Ω is

Q − ∇U = 0 , (11.9)
∂U
+ ∇ · F (U, Q) = S . (11.10)
∂t

Similarly, (11.9) can be written in a weak formulation using the same spatial discretization
and basis functions, and one eventually ends up with the following equation:

N N
Z ! Z !
X X
Mik : qlk Mlk dΩk + k
ukl Lkl dΩk


∇ · Mi ·
Ωk l=1 Ωk l=1
 
Z Nedge
X
ukl Lkl  · Mik n dΣk = 0 . (11.11)


− 
Σk l=1

11.2.1 Mapping from Physical to Master Elements

In principle, all the integrals in (11.8) and (11.11) may be evaluated over each single element Ωk
of the domain. However, it is convenient to map every element to a master element ΩM defined
on the Cartesian reference system (ξ, η), as shown in Figure 11.1. Moreover, Bassi and Rebay
[44] and others have pointed out how the discontinuous Galerkin discretization is particularly
sensitive to the geometric approximation of the physical boundaries, and curved elements must
be introduced in order to avoid loss in accuracy and in velocity of convergence.

Figure 11.1. Mapping from the physical to the master element.

For triangular elements, the master element is the unit right triangle with vertices (ξ1 , η1 ) =
(0, 0), (ξ2 , η2 ) = (1, 0), and (ξ3 , η3 ) = (0, 1). For quadrangular elements, the master element is
a square centered at the origin of the (ξ, η) system, with sides parallel to the reference axes and
length lside = 2. The master elements are shown in Figure 11.2.
 11.2. Nodal DGM 237
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 11.2. Quadrangular master element with nodes.

The mapping between Ωk and ΩM can be expressed in the form

N NX
X −m
x(ξ, η) = ϑmn ξ m η n , (11.12)
m=0 n=0
N NX
X −m
y(ξ, η) = φmn ξ m η n , (11.13)
m=0 n=0

where the coefficients ϑmn and PφNmn P are the solution of a linear system which is always well
N −m
conditioned because the terms m=0 n=0 ξ m η n form a linearly independent set of bases.
The mapping expressions (11.12) and (11.13) enable a fast evaluation of the transformation
Jacobian J and its determinant ||J||. The mapping does not impose any limits on the geometric
shape of the physical cell, so it is suited for describing curvilinear elements.
Moreover, (11.12) permits one to evaluate simply the Jacobian J at each point of the element:
" #
∂x ∂x
∂ξ ∂η
J= ∂y ∂y , ||J|| = det(J), (11.14)
∂ξ ∂η

and, therefore,

N NX−m
∂x(ξ, η) X
= ϑmn mξ m−1 η n ,
∂ξ m=0 n=0
N NX−m
∂x(ξ, η) X
= ϑmn nξ m η n−1 ,
∂η m=0 n=0
N NX−m
∂y(ξ, η) X
= φmn mξ m−1 η n , (11.15)
∂ξ m=0 n=0
N NX−m
∂y(ξ, η) X
= φmn nξ m η n−1 ,
∂η m=0 n=0

where the coefficients δmn and γmn depend only on the physical values of the element. With
respect to a numerical evaluation of the Jacobian, the approach that has just been described is
much more efficient even if this advantage is limited to the initialization process.
 238 Chapter 11. Discontinuous Galerkin-Based Reduced Order Models

11.2.2 Interface Fluxes

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

In this section we provide a brief overview of the fluxes implemented in the full-order solver,
starting from the discretized numerical solution rather than the exact one. A complete mathe-
matical dissertation on the influence of the fluxes on the stability and accuracy of the solution is
beyond the scope of this work.
Since the numerical solution takes two distinct values in general at the interfaces between
neighboring elements, a numerical procedure must be adopted in order to evaluate flux exchanges
between contiguous cells, see Figure 11.3.
In particular, the third term of the left-hand side of (11.6) can be written as
Z Z
k k k
Lk+ k+ + k−
· Fhk− n− dΣk , (11.16)


Li · Fh n dΣ = i · Fh n + Li
Σk Σk

where + and − distinguish the values of the discontinuous quantities on either side of an inter-
face. The + sign corresponds to the element Ωk , with the normal direction consistent with the
edge orientation. Defining the jump operator [[A]] = A+ · n+ + A− · n− = (A+ − A− ) · n+
and the average operator hAi = (A+ + A− ) /2, relation (11.16) can be rewritten as
Z Z
Lki Fhk · n dΣk =
 k  k
Li hFh i + Fhk hLki i dΣk . (11.17)
 

Σk Σk

In the DGM, the argument of integral (11.17) is replaced by an interface flux function Γ depend-
ing on the test function, the local normal, and the values of the vector U and its derivatives of
both elemental edges delimiting the interface:

Γ U+ , U− , Lk+ k−
(11.18)


i , Li , n ;

it is stressed that the numerical flux Γ is evaluated from quantities that could be expanded in a
polynomial basis with different order Lk+ k−
i , Li .
To ensure consistency, Γ must satisfy the condition

Γ U, U, Lki , Lki , n = Lki hFhk i . (11.19)

 

To make the DGM elementwise conservative, the numerical flux must satisfy

Γ U+ , U− , Lk+ k− − + k− k+
(11.20)



i , Li , n = −Γ U , U , Li , Li , −n .

Figure 11.3. Flux exchange between contiguous elements.

 11.3. Reduced Order Methods for DGM 239

The convective and diffusive fluxes are discretized separately:

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Γ = Γc + Γd . (11.21)
The convective discretization is stabilized by using an approximate Riemann solver H for Γ. The
approximate Riemann solvers H are based on the ones developed in the finite volume frame-
work. Among the most common are the HLL scheme [582], the Lax–Friedrichs scheme, and the
Rusanov scheme [581]:

F (U+ ) + F (U− ) supV∈Vad |∇(F · n)(V)|
HLF U+ , U− , Lk+ i , L k−
i , n = + [[U]] ,
2 2
(11.22)
where Vad is the space of admissible states and U+ and U− are expanded in the possibly different
polynomial bases Lk+i and Lk−i , respectively.
On the other hand, the diffusive flux exchange between neighboring elements depends both
on the values of the conservative variables and on their derivatives.
Many techniques can be adopted to evaluate the diffusive fluxes; one of the most common is
the one known as interior penalty [174].
Usually, diffusive splitting schemes also provide the splitting for the evaluation of the last
term of the left-hand side of (11.11), which depends again on the values of the variables on the
edges delimiting the interface. The reader can refer to [20, 375] for extensive studies of fluxes.

11.3 Reduced Order Methods for DGM

Nowadays, parametric studies of fluid dynamics flows are performed in a huge variety of applica-
tions (e.g., naval, aeronautic, space, automotive fields) for different purposes, including design,
optimization, and sensitivity analysis.
DGM, because of its high intrinsic computational cost on standard hardware, could benefit
from adopting reduced order methods.
Regardless of the method with which the full-order solutions are obtained, ROMs are based
on the assumption that the dynamics of a given system is governed by a limited number of
dominant modes. Once a set of parameters is fixed, one wants to reconstruct the variable v as
Nv
X
v(x, t; µ) ≈ ai (t, µ)ϕi (x), (11.23)
i=1

where µ is a set of parameters which the solution depends on, ϕ(x) is a set of global basis
functions, and ai are the unknown coefficients.
As with the other numerical methods, the reduced order methods for DGM rely on an online-
offline paradigm:
• offline phase: the snapshots are collected, modes are extracted, and matrices for reduced
order computations are computed;
• online phase: solutions are obtained in near real time for a given combination of parameters
with the aid of the matrices computed in the previous phase.
In the following subsections, a brief recap of the main steps needed to obtain the reduced
models is reported for the sake of clarity.

11.3.1 Offline Phase

During the offline phase, the following steps are performed:
 240 Chapter 11. Discontinuous Galerkin-Based Reduced Order Models

• A finite-dimensional training set is defined:

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

K = µ1 , µ2 , . . . , µNk , µi ∈ P ∀i ∈ {1, . . . , Nk } , (11.24)



where µk is a set of parameters, P is the space of parameters, and Nk is the size of the
training set.

• Accurate Ns solutions of the flow field are computed for given sets of parameters µ in
order to obtain the snapshots:

Sv = v(µ1 , t1 ), v(µ2 , t2 ), . . . , v(µNk , tNt ) . (11.25)

 

The total number of snapshots to be saved is therefore equal to Nk × Nt , where Nt is

the number of snapshots saved at different time instants for a fixed combination of pa-
rameters µ. Note that for problems which present a steady-state solution, the dependence
on time can be dropped and therefore Nt = 1. In general, one cannot establish in ad-
vance a proper number for Ns in order to guarantee the creation of an accurate ROM
since this number depends heavily on the unsteadiness and the nonlinearity of the problem
considered.

• Reduced bases of dimension Nvr with Nvr  N × M (i.e., the total number of degrees of
freedom of the full order case),

(11.26)
 
Lv = ϕ1 , ϕ2 , . . . , ϕNur ,

are obtained through a proper orthogonal decomposition (POD). These bases are used to
represent the flow field state variables, and they contain information about the dominant
dynamics which characterize the problem. They are ordered according to the magnitude
of their eigenvalues, whose values are proportional to the average energy of the projection
of the ensemble
λ1 > λ2 > · · · > λn > · · · > λN , (11.27)
and only the most energetic modes are retained. It is interesting to note that we are intro-
ducing a set of global bases here, while in a DGM formulation, bases are defined only on
the single elements of the domain.
The POD modes are obtained by solving a constrained optimization problem of the form

Rφ = λφ , (11.28)

where R is a compact, self-adjoint, and nonnegative definite operator; λ is a diagonal

matrix containing nonnegative λi eigenvalues; and φ gathers the eigenvalues; a more ef-
ficient approach is based on singular value decomposition (SVD), which, starting from
the snapshot matrix, allows one to obtain the singular values and the associated singular
vectors.

• Once the modes have been extracted, the governing equations describing the problem at
hand can be projected onto the reduced space distinguished by the first modes computed
in the previous step.

The details of how the reduced matrices are obtained by projecting the governing equations onto
the reduced basis are presented in Subsection 11.4.
 11.4. Projection of the Governing Equations 241

11.3.2 Online Phase

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

During the online phase, the reduced order system is solved for a given set of parameters µ in
near real time, given the small size of the matrices involved in the computation. The fields are
approximated at each time step as a linear combination of modes multiplied by proper weights,
which are the unknowns of the reduced order method. In general, a reliable approach consists
of solving the reduced system with the same time integration scheme adopted for the full-order
set in order to guarantee consistency between the full-order method and the reduced one. Both
explicit and implicit strategies can be adopted.

11.4 Projection of the Governing Equations

In this section the procedures adopted for evaluating reduced matrices, starting from the govern-
ing equations, are reported. The implementation of these procedures can be found in the open-
source library ITHACA-DG [1], a software library based on HopeFOAM, a major extension of
OpenFOAM for high-order methods [629].

11.4.1 Incompressible Navier–Stokes

First, we focus on the solution of incompressible flows, discretized according to the Chorin
method [124] implemented in HopeFOAM. The Chorin method is a partitioned approach for
solving for pressure and velocity separately by introducing a Poisson equation. The equation for
the full-order system has the form
∂V
+ (V · ∇)V = −∇p + ν∇2 (V ) ,
∂t
∇·V =0,
where V is the two-dimensional velocity vector, p is the pressure, and ν is the viscosity. A
variant of the Chorin partitioned numerical scheme implemented is reported in [282]:
γ0 V̄ − α0 Vn − α1 Vn−1
= −β0 ∇ · (Vn ⊗ Vn ) − β1 ∇ · (Vn−1 ⊗ Vn−1 ) ,
δt
γ0
−∆pn+1 = − ∇ · V̄ ,
δt
γ 0 γ 0 ∇pn+1
−∆Vn+1 + Vn+1 = V̄ − ,
νδt νδt ν
where the coefficients γ0 , α0 , α1 , β0 , β1 are relative to a chosen multistep method for the time
discretization and δt is the time step. The first equation explicitly solves for the contribution of
the nonlinear term to the velocity V̄, the second equation is the pressure projection step necessary
to obtain an intermediate divergence-free velocity V̄ − (δt)/γ0 )∇pn+1 , and the final equation
updates the velocity with the contribution of the viscous term. The boundary conditions are not
reported for brevity and can be found in [282].
During the offline phase we end up with the modes for the flow variables p and V, i.e., φpI
and φVI ; thus the projection of the governing equations on the reduced basis has the form
D γ V̄ − α Vn − α Vn−1 E D
0 0 1
, φV
I = − β0 ∇ · (Vn ⊗ Vn )
δt E
− β1 ∇ · (Vn−1 ⊗ Vn−1 ), φV I ,
D E D γ E
0
− ∆pn+1 , φpI = − ∇ · V̄, φpI ,
δt
D γ0 n+1 V E D γ0 ∇pn+1 V E
− ∆Vn+1 + V , φI = V̄ − , φI .
νδt νδt ν
 242 Chapter 11. Discontinuous Galerkin-Based Reduced Order Models

Putting together the system of equations just written and (11.23), we obtain the following ex-
pressions
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Nv Nv
X γ0 v̄j − α0 vjn − α1 vjn−1 D V V E X
φj ,φI = −β0 vjn vkn
j
δt
jk
D E
− β1 vjn−1 vkn−1 φV
j ∇φ V
k , φV
I ,
Np Nv
D
p p
E γ0 D V E
v̄j φj , ∇φpI ,
X X
pn+1
j ∇φ j , ∇φ I = −
j j
δt
Nv Np n+1 D
D E γ0
D V VE pj E
∇φpj , φV
X X
vjn+1 V V
∇φj , ∇φI + n+1
vj − v̄j φj , φI = − I ,
j
νδt j
ν

where the scalar product is defined as follows:

D E Z D E Z D E Z
φa , φb = φa φb dΩ, φa , φb = φa ·φb dΩ, Φa , Φb = Φa : Φb dΩ , (11.29)
Ω Ω Ω

where φa and φb , φa and φb , and Φa and Φb are two generic scalar, vectorial, and matrix-valued
functions, respectively, defined over the domain Ω. An L2 -norm was chosen in this context
because this particular type of projection minimizes the energy loss in the system.

11.4.2 Compressible Navier–Stokes

ROMs for Compressible Flows
In HopeFOAM, the following set of equations is implemented to deal with compressible flows
of a perfect gas with constant temperature:
∂ρ
+ ∇ · ρV = 0 ,
∂t
∂ρV
+ ∇ · (ρV ⊗ V) = −∇p + µ∇ · (∇V + λ∇ · V) ,
∂t
where the first equation takes into account the local variations of density ρ, λ is the bulk viscosity,
and µ is the dynamic viscosity.
Unfortunately, an energy-stable scheme for convection-dominated problems, when com-
pressible phenomena are considered, is not straightforward to implement because the L2 -norm
does not correspond to a kinetic energy integral; for this reason a stabilization strategy must be
introduced. Kalashnikova and Arunajatesan [317] proposed a method for stabilization which re-
lies on the linearization of the governing equations reported above. The significant magnitude of
the flow fields is written as
v = v 0 + v̄, (11.30)
where v 0 represents the local fluctuations of the generic variable v with respect to the correspond-
ing mean flow field values v̄. Linearized equations are then obtained by substituting (11.30) into
the full set of equations; this linear set of equations is then projected onto the reduced space.
Therefore, the fluctuating flow variables for the ROMs are
Nm
X Nm
X
ρ0 ≈ ρ
cm ψ m , V0 ≈ V
am ψm . (11.31)
m=1 m=1
 11.4. Projection of the Governing Equations 243

This approach is effective for describing flows in which strong nonlinear phenomena do not
occur.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

The linearized compressible Navier–Stokes set of equations projected on the domain Ω can
be written as

RT̄ D dρ0 ρ E RT̄ D 0 ρ E D E

, ψI + V̄ ∇ρ , ψI + RT̄ ∇ · V0 , ψIρ = 0 ,
ρ dt ρ
D dV0 E D E D E D E
ρ̄ , ψIV + ρ̄V̄ ∇ · V0 , ψIV + RT̄ ∇ρ0 , ψIV = µ ∇2 V, φV I , (11.32)
dt

where the term depending on the pressure gradient is rewritten in terms of temperature T and
density using the gas equation
p = ρRT , (11.33)

where R is the gas constant. Moreover, in the mass conservation equation, the term describing
the density fluxes has been split into two components, applying the chain rule for derivatives.
The two equations have been multiplied by Rρ̄T̄ and ρ̄, respectively. This has been demon-
strated to be an effective way to further stabilize the reduced order system, and it is equivalent to
adopting a stabilized inner norm:
D E Z
φa , φb = φa Hφb dΩ , (11.34)
H Ω

where H is the stabilization matrix. Further details are reported in [317, 318].
The terms in (11.32) are evaluated in a way similar to the one adopted for the incompressible
flows, as follows:

D dρ0 E XNm D E
, ψIρ = ċj ψjρ , ψIρ ,
dt j=1
D E XNm D E
∇ρ0 , ψIρ = cj ∇ψjρ , ψIρ ,
j=1
D E XNm D E
∇ · V0 , ψIρ = ak ∇ · ψkV , ψIρ ,
k=1
D dV0 E Nm
X D E
, ψIV = ȧj ψjV , ψIV ,
dt j=1
D E XNm D E
∇ρ0 , ψIV = cj ∇ψjp , ψIV ,
j=1
D E XNm D E
∇ · V0 , ψ V = aj ∇ · ψlV , ψIV .
j=1

The terms reported above are all linear, which has beneficial effects in terms of storage and com-
putational efficiency. The integration in time is carried out with a strong stability-preserving
algorithm both for the full-order solver and the reduced order one in order to guarantee consis-
tency between full and reduced orders.
 244 Chapter 11. Discontinuous Galerkin-Based Reduced Order Models

11.5 Test Case and Results

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

In this section, two different test cases are described, one for the incompressible case and one at
a moderate Mach number, in order to show both the effectiveness and the flexibility of the ROMs
based on DGM for modeling cases characterized by very different Mach and Reynolds number.

11.5.1 Incompressible Flow around a Square-Based Cylinder

In this test case, a cylinder with a square base of 0.1 m has been placed in a two-dimensional
duct. The size of the domain varies between −10 and +25 chord lengths along the x-axis and
between −2.5 and 2.5 chord lengths along the y-axis, where the origin of the Cartesian reference
system is located in the center of the base of the cylinder.
The inlet of the domain is on the left, and the walls which delimit the cylinder and the duct
were modeled using the no-slip boundary conditions. The Reynolds number varies between 25
and 250, and the inlet velocity is constant and set equal to 1 m/s, while a pressure equal to one
is imposed at the outlet.
The domain is discretized using 1766 triangular elements (see Figure 11.4), and the polyno-
mial degree inside each element is equal to three.

Figure 11.4. Mesh.

In order to train the reduced order methods, 10 simulations were carried out with different
values of viscosity, varying between 0.004 and 0.04 to match the Reynolds number, and 24
different snapshots were saved at different time instants for each simulation. Then both pressure
and velocity modes were extracted using a procedure based on the SVD. The first few modes for
the velocity are depicted in Figure 11.5.
Different numbers of modes Nu = Np were considered, and the equations were projected
onto the reduced space created by the modes.
Simulations for different Nu were run, imposing a viscosity equal to 0.01, a value which did
not belong to the training set. The final time of simulation was set to 5 s. The results obtained
with the ROM were compared to the ones obtained with the high-fidelity model, and they are
in good agreement for both pressure and velocity. Figure 11.6 shows the fields obtained when
different numbers of modes are taken into account. As expected, one can notice that the ones
with a higher number of modes provide a better estimation of the flow field. In Figures 11.7 and
11.8, the point-by-point differences between the field obtained with the ROM with Nu = 12 and
the one computed with the discontinuous Galerkin–based solver are reported.

11.5.2 Parametric Simulation of a Compressible Cavity Flow in a Duct

The main goal of this second test case is to demonstrate the capability of the reduced order
solver described in Subsection 11.4.2 to properly capture the dominant dynamics of cases in-
volving compressible flows. In detail, in this test the formation of vortex in a two-dimensional
duct with a cavity is studied; the duct is 7 m long, and the narrowest section of the duct has a
width of 0.8 m. The cavity is located at a distance of 2.5 m from the inlet, and it is 1 m long.
 11.5. Test Case and Results 245
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 11.5. First four velocity modes for the square cylinder test case.

Figure 11.6. Reduced order solution for Nu = 5, Nu = 8, Nu = 10, and Nu = 12.

Figure 11.7. Absolute error (velocity).

Figure 11.8. Absolute error (pressure).

 246 Chapter 11. Discontinuous Galerkin-Based Reduced Order Models
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 11.9. Mesh of the domain.

Corresponding to the cavity, the width of the duct is 1.3 m. Solid wall no-slip boundary condi-
tions were imposed on the bottom wall, while on the upper part, slip boundary conditions were
used; the inlet and outlet boundaries were modeled with nonreflecting boundary conditions based
on the conservation of the Riemann invariants in order to avoid the reflection of outgoing waves
back into the computational domain. The discretization of the domain is depicted in Figure 11.9,
consisting of approximately 6200 unstructured straight-edged triangles. The polynomial order
is constant and equal to three in the whole computational domain. The Mach number of the
unperturbed flow at the inlet is fixed and set to 0.3, while the Reynolds number varies in a range
between 9000 and 12000.
During the offline phase, nine full-order simulations were run, corresponding to different
linearly spaced values of the Reynolds number, and 40 snapshots were saved for each of them
every 0.015 s; after that, the behavior became periodic. Once all the snapshots were collected,
the snapshot matrix was built and the modes were extracted with the aid of an SVD procedure;
the first ones are reported in Figure 11.10. Figure 11.11 shows a qualitative comparison between
a full-order solution and the corresponding reduced one for a Reynolds number equal to 10000,
a value which was not included in the previous set of simulations. The ROM results are in good
accordance with the full-order ones; some differences can be observed in particular where the
nonlinear dynamics are stronger, e.g., close to the corners and in the area where the cavity is
placed. These discrepancies can be explained by the fact that in this case a linearized version of
the Navier–Stokes equations was implemented in the reduced order solver.

Figure 11.10. First four modes (velocity).

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

11.5. Test Case and Results

Figure 11.11. Comparison between FOM (top) and ROM (bottom) solutions.
247
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Part III

Fluid Dynamics
Models for Computational
Advances in Reduced Order
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Chapter 12

Weighted Reduced Order

Methods for Uncertainty
Quantification

Davide Torlo, Maria Strazzullo, Francesco Ballarin,

Gianluigi Rozza

12.1 Introduction
Partial differential equations (PDEs) are an effective tool to model phenomena in applied sci-
ences. Realistic problems usually depend on several physical and geometrical parameters that
can be calibrated by using real data. In real scenarios, however, these parameters are affected by
uncertainty due to measurement errors or scattered data information. To deal with more reliable
models which take this issue into account, stochastic PDEs can be numerically approximated. In
the uncertainty quantification (UQ) context, many simulations are run to better understand the
system at hand and to compute statistics of outcomes over quantities of interest. In particular, the
input parameters of the stochastic PDEs are assumed to be random finite-dimensional variables.
Classical numerical approximations, i.e., high-fidelity solutions, can lead to untenable com-
putational costs for computing statistical momenta. These are the leading motivations of this
contribution. Indeed, we will focus on projection-based reduction techniques which can lighten
this issue: reduced order models (ROMs) [281, 505]. In many applications, the parametrized so-
lutions can be sought in a low-dimensional subset of the solution space. If one applies a Galerkin
projection in this reduced space, the problem is solved more rapidly than the original discretiza-
tion and is still accurate. These approaches, thus, might accelerate standard statistical analysis
techniques, such as the Monte Carlo methods. (For alternative methods for forward UQ with
random inputs, e.g., stochastic Galerkin, stochastic collocation, and Karhunen–Loève approxi-
mation, we refer the reader to [565].) More precisely, this chapter focuses on weighted ROMs for
forward uncertainty problems, where the reduced algorithms are modified to comply with some
previous knowledge of the distribution of the parameters and to use this information to accelerate
the reduced simulations even more; see, e.g., [117, 592] and the references therein. Moreover,
the proposed weighted ROMs are not intrusive with respect to the classical ROM approaches;
i.e., given a reduced model the weighted ROMs just weigh its outcome without changing its
form. We refer the reader to [406] for inverse problem applications and to [285] for polynomial
chaos expansion-reduced techniques.

251
 252 Chapter 12. Weighted Reduced Order Methods for Uncertainty Quantification

The contribution is outlined as follows: in Section 12.2 we present stochastic PDEs and their
approximation. Then, we move toward weighted ROMs, describing standard algorithms such
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

as weighted proper orthogonal decomposition (POD) [117] and weighted reduced basis [592]
in Section 12.3 and the related sampling strategies in Section 12.4. Then, in Section 12.5, we
validate the proposed algorithms in several contexts, including heat transfer, Stokes problems,
advection-dominated phenomena, and optimal control for environmental sciences. Conclusions
follow in Section 12.6.

12.2 Stochastic PDEs and Discretized Approximations

In this section, we introduce the stochastic setting for PDEs in their high-fidelity discretization.
Consider a complete probability space (A, F, P ), with A the set of possible outcomes, F a σ-
algebra of events, and P a probability measure. Let µ = µ(ω), µ : (A, F) → (Γ, B), where
Γ ⊂ RK is a compact set; B be the Borel measure; and µ(ω) = (µ1 (ω), . . . , µK (ω)) be a
random vector whose components are independent absolutely continuous random variables and
which parametrizes the physical problem. Denote by ρ : RK → R the probability density
function of µ. Let D ⊂ Rd with d = 1, 2, 3 be a bounded physical domain. The physical
problems we are interested in can be modeled by PDEs over D as follows:
find u : Γ → V such that P(u(µ(ω)); µ(ω)) = 0 (12.1)
for almost every (a.e.) ω ∈ A. The solution u is usually sought in a Sobolev space that we
generically define as V. Here, P represents a parametrized PDE problem in its weak formulation,
e.g., an elliptic problem or a parabolic initial-valued problem, Navier–Stokes equations, or a
shallow-water problem.
We suppose, furthermore, that a high-fidelity method, namely finite element or finite volume,
provides an approximation uN of the solution u. These solutions are characterized by an error
which is inversely proportional to a power of the number of degrees of freedom N of the dis-
cretized functional space VN and a computational cost which is proportional to a polynomial in
N . Let us define the high-fidelity solver as follows:
find uN : Γ → VN such that PN (uN (µ(ω)); µ(ω)) = 0 for a.e. ω ∈ A. (12.2)
The computation of the high-fidelity solution can be infeasible in real-time or many-query
situations, where a faster but still reliable solver is desired. In particular, we focus in this chapter
on stochastic problems, where statistical momenta are the objective of the study. Typically,
averaging algorithms, such as Monte Carlo, are used to obtain such quantities, and hence the
computation of many simulations for different parameters µ ∈ Γ is necessary.
To speed up these procedures, ROMs offer interesting strategies. The core idea of ROMs is
to search the parametrized solution inside a reduced functional space Vn ⊂ VN . This reduced
space is characterized by a dimension n  N . Model order reduction algorithms consist of two
phases: an offline phase, where the reduced space is built using some snapshots, i.e., high-fidelity
simulations for some parameters µ, and an online phase, where the reduced space structure is
exploited and a reduced solver with computational costs proportional to n is used to simulate the
solution for new parameters µ. The reduced problem is defined as follows:
find un : Γ → Vn such that Pn (un (µ(ω)); µ(ω)) = 0 for a.e. ω ∈ A. (12.3)
For the problem we will deal with, the reduced space is built in order to control the error between
the high-fidelity solutions and the reduced ones kuN (µ) − un (µ)k, and a cheap and sharp error
bound is available, i.e.,
ηn (un (µ(ω)); µ(ω)) ≥ kuN (µ(ω)) − un (µ(ω))kVN =: eVn (µ(ω)) for a.e. ω ∈ A. (12.4)
 12.3. Weighted ROMs 253

12.3 Weighted ROMs

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

In this chapter we will focus on how the offline phase of different algorithms can be modified to
take into account the knowledge of the distribution of the random variable µ. This is done mainly
by means of weighted algorithms that are derived from classical ROM algorithms [281, 505]. We
will present the weighted reduced basis (wRB) method [117] and the weighted proper orthogonal
decomposition (wPOD) [592]. Moreover, we will focus only on the offline modification of such
algorithms, because the online phase is unmodified with respect to the original algorithms. The
difference is that now we are interested not only in one evaluation for a given µ but also in the
momenta of the random variable un (µ). We refer the reader to [281, 504] for the development
of those parts of the algorithms.

12.3.1 WRB Method

The wRB [117] is an extension of the (deterministic) reduced basis (RB) method [510]. In
the deterministic setting, the reduced space Vn is constructed following the greedy Algorithm
12.1, which selects some of the high-fidelity snapshots from a training set {uN (µ)}µ∈Ξt . The
construction of the reduced space is progressive, and at every iteration the dimension of the
reduced space increases by adding the selected snapshot to the bases of the reduced space. The
chosen snapshot is the one that is worst approximated by the RB space at each iteration (in
a greedy fashion). This choice is quickly done using the error bound, which allows us not to
compute the high-fidelity solution for all the parameters µ ∈ Ξt but only for the ROM ones.

A LGORITHM 12.1. (Deterministic) greedy algorithm.

Sample Ξt ⊂ Γ
Let V0 = ∅
Pick an arbitrary µ1 ∈ Ξt
n=0
while err > tol do
n := n + 1
Solve (12.2) for µ = L
µn to get uN (µn )
Update Vn := Vn−1 span{uN (µn )}
Find µ n+1
:= arg maxµ∈Ξt ηn (un (µn+1 ); µn+1 )
err := ηn (un (µn+1 ); µn+1 )
end

In the stochastic context, we want to give different importance to different parameters µ, i.e.,
different realizations, according to the underlying probability distribution. This must be done
by modifying the error measure that we used in the deterministic setting, introducing a different
norm k · kw that modifies the search of the worst-represented snapshot in the greedy algorithm.
In particular, the new norm is defined as

ku(µ)kw := w(µ)ku(µ)kVN ∀u ∈ VN , ∀µ ∈ Γ. (12.5)

The weight function w : Γ → R+ depends on the target norm that we want to minimize. As an
example, if we want to minimize the expected value of the 2-norm of the error, i.e.,
Z
E[kuN − un k2VN ] = kuN (µ(ω)) − un (µ(ω))k2VN dP (ω) (12.6)
D
Z
= kuN (µ) − un (µ)k2VN ρ(µ)dµ, (12.7)
Γ
 254 Chapter 12. Weighted Reduced Order Methods for Uncertainty Quantification

the natural choice for the weight function is w(µ) = ρ(µ) [117]. Consequently, the error
p

estimator also reflects this modification by introducing a new error bound ηnw (µ) = w(µ)ηn (µ),
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

resulting in
Z Z
2 w 2
E[kuN − un kVN ] ≤ ηn (µ) dy = ηn (µ)2 ρ(µ)dµ. (12.8)
Γ Γ

Other choices may be more meaningful in other contexts, e.g., different norms or output error
minimization [580, 593]. The modified algorithm is reported in Algorithm 12.2.

A LGORITHM 12.2. Weighted greedy algorithm.

Properly sample Ξt ⊂ Γ
Let V0 = ∅
Pick an arbitrary µ1 ∈ Ξt
n=0
while err > tol do
n := n + 1
Solve (12.2) for µ = L
µn to get uN (µn )
Update Vn := Vn−1 span{uN (µn )}
Find µ n+1
:= arg maxµ∈Ξt ηnw (un (µn+1 ); µn+1 )
err = ηn (un (µn+1 ); µn+1 )
w

end

Another important aspect in this algorithm is the choice of the sampling strategy of the parameter
training set Ξt . In Section 12.4, we will investigate in more detail the different possibilities and
their effects on the overall results.

12.3.2 wPOD
The other well known algorithm in the ROMs community is the proper orthogonal decomposition
(POD) [281]. Its weighted version (wPOD), originally introduced in [592], is useful when an
error estimator ηn is not available.
The POD Algorithm 12.3 finds the global minimum of the mean square error
Z
kuN (µ) − un (µ)k2VN dµ (12.9)
Γ

over all the possible reduced spaces Vn ⊂ VN of dimension n. In practice, a tolerance on the
error can be set and the dimension n will be provided by the algorithm. In the discrete situation
(12.9) becomes
1 X
kuN (µ) − un (µ)k2VN , (12.10)
|Ξt |
µ∈Ξt

with Ξt being the training set. The reduced space Vn is defined by the eigenvectors correspond-
ing to the n leading eigenvalues of the operator C : VN → VN defined by
|Ξt |
X
C(v) := hv, uN (µi )iVN uN (µi ). (12.11)
i=1

In practice, this is computed by an eigenvalue analysis on the correlation matrix Ĉij := huN (µi ),
P|Ξt | i
uN (µj )iVN . The bases of the reduced space Vn are defined by ξ i = j=1 ψj uN (µj ), where ψj
are the eigenvectors of Ĉ.
 12.4. Sampling Strategies 255

A LGORITHM 12.3. POD.

Sample Ξt ⊂ Γ
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

for µ ∈ Ξt do
Solve (12.2) for µ to get uN (µ)
end
Assemble the matrix Ĉij = huN (µi ), uN (µj )iVN
Compute the biggest n eigenvalues λk and the eigenvectors ψk of Ĉ for k = 1, . . . , n
P|Ξt | i
Define Vn := span{ξ 1 , . . . , ξ n }, where ξ i = j=1 ψj uN (µj ).

The stochastic setting modifies the mean square error (12.9) as the expected value (12.7).
This leads to a modified correlation matrix
w
Ĉij := w(µi )huN (µi ), uN (µj )iVN . (12.12)

We remark that Ĉ w is not diagonalizable in the usual sense, but it is with respect to the scalar
product induced by C, which allows us to obtain n orthogonal leading eigenvectors. The modi-
fied algorithm is reported in Algorithm 12.4.

A LGORITHM 12.4. wPOD.

Properly sample Ξt ⊂ Γ
for µ ∈ Ξt do
Solve (12.2) for µ to get uN (µ)
end
Assemble the matrix Ĉ w as in (12.12)
Compute the biggest n eigenvalues λk and the eigenvectors ψk of Ĉ w for k = 1, . . . , n
P|Ξt | i
Define Vn := span{ξ 1 , . . . , ξ n }, where ξ i = j=1 ψj uN (µj ).

12.4 Sampling Strategies

In the previously proposed algorithms, the choice of the training set Ξt over which the mini-
mization occurs is fundamental. There are different strategies that one could consider. In the
stochastic setting it is very common to use Monte Carlo algorithms to approximate the momenta
of interest. In the numerical analysis setting, this algorithm can be seen as a collection of quadra-
ture points for a specific quadrature formula. These two perspectives lead to different methods
that one can use, with different pro and cons:
• Monte Carlo uniform sampling to approximate (12.7);
• Monte Carlo with sampling led by the underlying distribution of the random variable µ,
approximating (12.6) directly;
• tensor product quadrature rules of univariate quadrature rules; or
• sparse Smolyak quadrature rule.
The Monte Carlo rules, which are very common in the stochastic world, suffer from a slow decay
of the error of the integral, which scales as O(|Ξt |−1/2 ). The sampling using the underlying
distribution may lead to some advantages, in the case of very concentrated distributions, but it is
strongly not recommended for rare events prediction.
The classical quadrature rules that generalize to tensor products in more dimensions suffer
from the curse of dimensionality. In particular, according to the smoothness of the function we
want to approximate, they will behave as O(|Ξt |−n̄/K ), where n̄ is the minimum of the regularity
 256 Chapter 12. Weighted Reduced Order Methods for Uncertainty Quantification

Tensor product Smolyak rule

1 1
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

Figure 12.1. Comparison of multivariate quadrature rules with order of accuracy 15.

of the function and the accuracy degree of the quadrature rule, and K is the dimension of the
(parameter) space Γ.
Sparse quadrature rules can improve this by putting the quadrature points in different loca-
tions. The Smolyak quadrature rule [292, 48, 610] allows us to not fall victim to the curse of
dimensionality, using different refinement levels of the grids on the different parameters. Con-
(j)
sider Γ := ΠK j=1 Γj , and define by Ui the univariate quadrature rule at the refinement level i on
the interval Γj ⊂ R, introducing the differences operators on Γj as
(j) (j) (j) (j)
∆0 = 0, ∆i := Ui+1 − Ui for i ≥ 0. (12.13)
The Smolyak quadrature rule of order q in Γ is defined as
X K
O
QK
q :=
(i)
∆α i
. (12.14)
|α|1 ≤q, α∈NK i=1

The difference can be observed in Figure 12.1, where a two-dimensional example shows the
tensor product and the Smolyak rule.
In the next section, we will alternate different quadrature strategies to show the advantages of
choosing samples distributed with the underlying probability law and the cost reduction without
loss of quality of sparse quadratures.

12.5 Applications
In this section we apply the strategies presented above to different problems. In the first subsec-
tion we apply the presented strategies directly to a heat equation problem and to Stokes problems.
In Subsection 12.5.2 we use the proposed methods for a selective stabilization technique in the
context of uncertainty quantification. Finally, in Subsection 12.5.3 we apply the weighted algo-
rithms to parametrized optimal control problems OCP(µ)s in environmental sciences.

12.5.1 Averaging Uncertain Parameter Simulations for Heat Equation

and Stokes Problems
The first problem we consider is a heat transfer problem on a squared domain D = [0, 1]2 . The
equation that we solve is
−µ(x1 , x2 )∆u(x1 , x2 ) = 1, (12.15)
 12.5. Applications 257

where µ : [0, 1]2 → R is defined piecewise constant on nine subsquares defined by Dij =
[i/3, (i+1)/3]×[j/3, (j+1)/3], i, j = 0, . . . , 2, by nine parameters µij ∈ R with i, j = 0, . . . , 2.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

In particular, we consider these heat transfer coefficients to behave like some Beta(20,10) random
variables, i.e., µij ∼ Beta(20,10) rescaled on the interval [1, 10]. The Beta distribution is defined
R1
using the Beta function B(α, β) = 0 µα−1 (1 − µ)β−1 dµ with the probability density function

µα−1 (1 − µ)β−1
ρ(µ) = . (12.16)
B(α, β)

We compare the reduced algorithms presented before in their weighted and classical formu-
lations, and we compare the uniform Monte Carlo sampling strategy with the Beta(20,10) Monte
Carlo sampling one, in both cases with |Ξt | = 100.
To assess the quality of the algorithms, we compute reduced solutions on a Beta(20, 10)
distributed test set and we average the error obtained with such solutions.
Test error on Beta distributed samples

Test error on Beta distributed samples

10 2
10 3

10 3
Uniform MC, Uniform POD Uniform MC, Uniform RB
10 4 Uniform MC, Beta POD Uniform MC, Beta RB
Beta MC, Uniform POD Beta MC, Uniform RB
Beta MC, Beta POD Beta MC, Beta RB
2 4 6 8 10 12 14 2 4 6 8 10 12 14
n reduced basis n reduced basis

Figure 12.2. Heat transfer problem: error decay with respect to the number of reduced ba-
sis functions, comparison between weighted and not weighted algorithms and between uniform and
Beta(20,10) distributed training samples. Left: POD; right: greedy algorithm.

In Figure 12.2, we notice that for the POD the weighting of the algorithm does not lead to
great improvement in the decay of the error, while sampling of the training set with the underlying
distribution greatly reduces the error, by as much as a factor of 10. Conversely, for the greedy
algorithm we see a great improvement with only the weighted algorithm of factor five. Here the
sampling strategy also plays a bigger role in the training phase and gives an overall improvement
of at least a factor of 10. The combination of the two strategies slightly improves this result. We
remark that a uniform sample considers parameters which are not at all well represented in the
reduced space, leading to worse average errors.
We now focus on the stochastic steady Stokes equations results shown in [218]: here we
extend the deterministic problem presented in [211] following the strategies first discussed in
[118]. In this case, we will deal with physical and geometrical parametrization. Namely, we
consider the random domain D(µ(ω)) : A → R2 and a random inlet condition gin (x, µ(ω)) :
∂Din ×A → R2 . From now on, for the sake of notation, we will use µ and µ(ω) interchangeably.
The random boundary of the domain is given by the partition

∂D(µ) = ∂DD,0 (µ) ∪ ∂Din (µ) ∪ ∂DN (µ),

where ∂DD,0 (µ) and ∂Din (µ) represent the portions of the boundary with homogeneous and
nonhomogeneous Dirichlet conditions, respectively, while ∂DN (µ) is characterized by Neumann
boundary conditions. The stochastic Stokes problem reads as follows: given an outcome ω ∈ A,
 258 Chapter 12. Weighted Reduced Order Methods for Uncertainty Quantification

find the µ-dependent pair (u, p) : D × A → R2 × R such that

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

−∆u(x) + ∇p(x) = 0 in D(µ),



∇ · in D(µ),



 u(x) = 0
on ∂DD,0 (µ),

u(x) = 0 (12.17)

 u(x) = gin (x, µ) on ∂Din (µ),
 ν ∂u (x) − p(x)n = 0


on ∂DN (µ).

∂n
The parameter we consider is
µ = (µ0 , µ1 , µ2 , µ3 , µ4 ) ∈ (0.2, 1.9) × (0.2, 2.) × (0.2, 2.) × (0.2, 2.) × (0.2, 20.),
where the first four parameters are related to the shape of the considered computational domain
and the last one is related to the diffusivity of the flow since it changes the inlet condition; indeed,
we define gin (x, µ) = (−µ4 x1 (x1 − 3), 0).
The domain is given by
D(µ) = {[0, 6] × [0, 3]} \ {([2, 2 + µ0 ] × [0, µ1 ]) ∪ ([3 + µ0 , 3 + µ0 + µ2 ] × [0, µ3 ])}.
The boundary portions are ∂DN (µ) = {6} × [0, 3] and ∂Din (µ) = {0} × [0, 3], while ∂DD,0 (µ)
= ∂D(µ) \ {∂DN (µ) ∪ ∂Din (µ)}. In order to deal with geometrical parametrization, we need to
consider a reference domain, say D, which we represent in Figure 12.3. At each new parameter
evaluation, the system (12.17) is traced back to D; see, e.g., [505]. To build the reduced model we
applied wPOD with supremizer enrichment [34, 511], which guarantees the well-posedness of
the Stokes problem at the reduced level for each new realization of µ. Furthermore, we assume
that the parameters verify µi ∼ Beta(75,75) for i = 0, . . . , 4. The testing set has cardinality
|Ξt | = 241 when dealing with the Smolyak rule, while |Ξt | = 243 for all other test cases.
Increasing the value of n, we compare several sampling techniques: the Smolyak rule, the tensor
product rule, and uniform and Beta(75, 75) sampling over a testing set of 100 outcomes from a
Beta(75, 75) sampling. In Figure 12.4, we show results for the unweighted algorithm (left plot)
and the weighted one (right plot). For the sake of brevity we report only the velocity-averaged
log-errors since the pressure has a comparable behavior. First of all, we notice that, apart from
the uniform sampling, we obtain very good results in term of relative errors, with values around
10−4 for both algorithms. Focusing on the unweighted algorithm, we see how the sampling
strategies might strongly help in accuracy with respect to uniform Monte Carlo sampling. Indeed,
uniform Monte Carlo suffers from the curse of dimensionality. As already seen in the left plot
of Figure 12.2, the wPOD performs slightly better for uniform sampling. We remark that the
two approaches are not equivalent since the resulting RB are different. For the other sampling
strategies, the weighted algorithm gives comparable results to the unweighted one.

Figure 12.3. Reference domain D. The reference parameter is µ = (1., 1.5, 1., 1.5, µ4 ).
 12.5. Applications 259
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 12.4. Stokes problem: error decay with respect to the number of RB functions, compar-
ison between unweighted (left) and weighted (right) algorithms and between the Smolyak, tensor product,
uniform, and Beta(75,75) distributed training samples. For the weighted algorithm, the Smolyak rule, the
tensor product rule, and Beta(75,75) sampling coincide.

12.5.2 Stabilization of Advection-Dominated Problems Conditioned

to Parameters
In this section we use the weighted and distributed algorithms presented above to further reduce
computational costs in the online phase.
Let us consider the Graetz problem, an advection-diffusion problem

10µ x2 (1 − x2 )∂x u − ∆u = 0 (12.18)

on the domain D = [0, 2] × [0, 1] with Dirichlet boundary condition equal to one on the bound-
aries with x1 ≤ 1 and equal to zero on the boundaries with x1 > 1. We suppose that the
parameter µ is distributed as a Beta(5,3) rescaled on the interval [0, 6]. The considered problem
will be advection dominated as µ increases, and a standard finite element method (FEM) would
fail to provide an accurate solution because it would be led by oscillations due to the instability
of the problem. In order to stabilize the method, we introduce in the offline phase a stabilization
term, represented by streamline upwind Petrov–Galerkin (SUPG) [477], which maintains the
consistency of the method but adds a stabilization effect.
In this simple example the stabilization is not computationally expensive, but in more realistic
cases the stabilization term might include more elaborate nonlinear terms.
In the online phase we can choose whether we want to apply the stabilization again or not.
As for the high-fidelity model, it is necessary to use it in the advection-dominated regime. We
see this behavior by running an online error decay study with respect to the dimension of n in
Figure 12.5. There, it is clear that, both with POD and greedy without online stabilization, the
error has a plateau at 10−1 , which is the distance between the nonstabilized and the stabilized
solutions [580].
In Figure 12.5 for stabilized methods, we observe the same order of convergence but smaller
errors for weighted algorithms and distributed training samples. Hence, for the rest of this section
we consider only weighted algorithms with distributed training samples.
We can observe in Figure 12.6 that the unstabilized and stabilized online solutions show
differences in the error, particularly in the advection-dominated regime, while, in the diffusive
regime, the two solutions show similar errors. Hence, one may be more interested in solving
these solutions without the stabilization terms to save computational time.
 260 Chapter 12. Weighted Reduced Order Methods for Uncertainty Quantification

Uniform MC, Uniform RB Uniform MC, Uniform RB

100
Test error on Beta distributed samples

Test error on Beta distributed samples

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Uniform MC, Beta RB Uniform MC, Beta RB

Beta MC, Uniform RB Beta MC, Uniform RB
10 1 Beta MC, Beta RB Beta MC, Beta RB
100

10 2

10 3

10 4 10 1

0 5 10 15 20 25 30 0 5 10 15 20 25 30
n reduced basis n reduced basis

Uniform MC, Uniform RB 102 Uniform MC, Uniform RB

101
Test error on Beta distributed samples

Test error on Beta distributed samples

Uniform MC, Beta RB Uniform MC, Beta RB
100 Beta MC, Uniform RB Beta MC, Uniform RB
Beta MC, Beta RB Beta MC, Beta RB
10 1 101
10 2

10 3
100
10 4

10 5

10 6 10 1

0 10 20 30 40 50 0 10 20 30 40 50
n reduced basis n reduced basis

Figure 12.5. Graetz problem: error decay with respect to the number of RB functions, and
comparison between weighted and unweighted algorithms and between uniform and Beta(5,3) distributed
training samples. With online stabilization (left), without online stabilization (right). POD algorithm (top),
greedy algorithm (bottom).

Stabilized Online 100

100 Not Stabilized Online
Selective Online 10 1
10 1 Stabilization Line
10 2
Stabilized Online
10 2 10 3 Not Stabilized Online
Error

Error

Selective Online
10 4
Stabilization Line
10 3
10 5

10 4
10 6

10 5 10 7

0 1 2 3 4 5 6 0 1 2 3 4 5 6
Advection parameter Advection parameter

Figure 12.6. Graetz problem: weighted algorithms and Beta(5,3) distributed training samples.
Error of stabilized and nonstabilized online with respect to the advection parameter µ. Comparison with
selective online strategy. Left: POD; right: Greedy.

The task that we aim to solve in this section is to compute some statistical quantities of the
stochastic solution u(µ), e.g., the average. A classical strategy could be to approximate the
average via a Monte Carlo algorithm, i.e.,
N MC
u(µi )ρ(µi )
Z Z X
E(u(µ)) = u(µ(ω))dP (ω) = u(µ(ω))ρ(µ(ω))dω ≈ , (12.19)
A A i=1
NM C

where ρ is the probability density function of the underlying distribution of µ and µi are Monte
Carlo samples uniformly distributed in the domain of interest.
 12.5. Applications 261

Table 12.1. Tables for the selective stabilization approach. Given a certain threshold advection
coefficient ε, we obtain the mean error by computing only a percentage of all the stabilization terms.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

POD algorithm Greedy algorithm

ε Mean Error Stabilized Rate ε Mean Error Stabilized Rate
0.0 5.43329 10−5 100 % 0.0 2.55161 10−6 100 %
1.0 5.62446 10−5 83% 1.0 8.02889 10−6 85.5%
1.5 6.79605 10−5 78% 1.5 8.069458 10−5 75%
2.0 3.61034 10−4 71% 2.0 5.031474 10−4 68 %
2.5 1.46119 10−3 65% 2.5 3.192200 10−3 60 %
3.0 4.17433 10−3 56.5% 3.0 1.066145 10−2 52.5 %
6.0 2.42411 10−1 0.0% 6.0 3.217842 10−1 0%

Computing many high-fidelity simulations uN (µi ) is too expensive for this task, so we con-
sider the reduced solutions un (µi ) instead. Now, the stabilization term is important to guarantee
an accurate solution in the advection-dominated regime, i.e., µ  0. Conversely, in the diffusion-
dominated regime, the stabilization term is not necessary and would be time-consuming.
What we present here is a selective approach to the reduced solutions, which will be com-
puted with the stabilization term or not according to a rule led by a threshold value ε. Namely,
while computing (12.19), for every ui we will use un (µi ) with the stabilization term if µi > ε;
else we will use un (µi ) without the stabilization term. This idea is illustrated in Figure 12.6 by
the selective online plot.
In Table 12.1 we see the result of this approach of changing the value of the threshold ε. In
the POD case we see that for ε = 1.5 we save the computational costs of 22% of the stabilization
terms and we obtain a negligible worsening of the error. With ε = 2 the error is already a bit
larger though increasing to almost the 30% of stabilizations not computed. Higher thresholds
make less sense. For the greedy approach we observe similar results, even if the error increases
a bit more with small thresholds.

12.5.3 wPOD for Optimal Control for Environmental Sciences

This section focuses on the application of wROMs to parametrized optimal control problems
(OCP(µ)s) applied to environmental sciences. Namely, we extend the results of [557] in a
stochastic setting, as presented in [98]. UQ combined with OCP(µ) is of utmost interest in the
field of natural sciences and is strongly related to in situ data. Furthermore, there is a need for
reliable simulations based on PDEs which are similar to the collected measurements. OCP(µ)s
comply with this task: they aim to change the classical behavior of a system toward a desired
configuration, namely, the collected information. This scope is reached using external factors
called controls. The control, thus, acts on the system and guides it in order to lower as much as
possible a cost functional dependent on the data observations of the physical phenomenon. The
ingredients of an OCP(µ) are
• a state variable y ∈ Y,
• a control variable u ∈ U, and
• an observation yd ∈ Z ⊃ Y,
where Y, U, and Z are Hilbert spaces. The problem we are dealing with is the minimization over
Y and U of
1 2 α 2
J (y, u; µ) := ky − yd kZ + kukU , (12.20)
2 2
 262 Chapter 12. Weighted Reduced Order Methods for Uncertainty Quantification

constrained to
(12.21)
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

P(y(µ), u(µ); µ) = 0,
where P is a linear stochastic constraint in the weak formulation. The interested reader may refer
to [587] for a survey on optimal control theory. This problem formulation has been used in data
assimilation contexts, where the variable u(µ) steers the system toward the desired configuration
yd , making the solution y(u(µ)) the most similar to the observation. The stochastic parameter
still plays an important role since both the data and the system may be affected by uncertainty. In
the following, we report some results concerning this framework applied to a hypothetical loss
of pollutant in the Gulf of Trieste. Let us indicate this geographical region with D ⊂ R2 , repre-
sented in Figure 12.7. We call Du the subdomain where the pollutant loss is happening (green
subdomain in Figure 12.7). The goal is to determine the maximum loss allowed in Du to keep
harmless the pollutant concentration in Dobs , the Miramare natural reserve (orange subdomain
in Figure 12.7). The boundary is partitioned into ∂DD , the coastline, and ∂DN , the open sea,
where Dirichlet and Neumann boundary conditions have been applied, respectively. The state
concentration is y ∈ Y := H∂D 1
D
, u ∈ R, and yd = 0.2 over Dobs . The observation represents a
safe threshold for the pollutant concentration in the Miramare area.

Figure 12.7. Domain D, the Gulf of Trieste. Orange: Miramare reserve Dobs . Pollutant spill Du .

For this specific test case, P is an advection-diffusion equation of the form

Z Z Z
µ1 ∇y · ∇zdD + ([µ2 , µ3 ] · ∇y) zdD − Lu zdDu = 0 ∀z ∈ Y. (12.22)
Ω Ω Ωu

Here, the random parameter µ = (µ1 , µ2 , µ3 ) ∈ (0.5, 1) × (−1, 1) × (−1, 1) models the sea
dynamics under several meteorological phenomena. The constant L = 1000 is used to make
the system nondimensional [557]. Furthermore, we impose α = 10−7 . We want to compare
the performances of informed sampling and standard POD. We will show how exploiting data
information in this context might help monitor and solve in real time a potentially dangerous
situation. Thus, we exploit the unweighted version of the POD with aggregated space strategy
[328, 329, 414], since, as we already said in the aforementioned applications, the POD is only
slightly affected by the weights, while the sampling strategy is of great importance. In this
experiment |Ξt | = 100. In Figure 12.8 we present the average relative error for the state and
the control variables, where the sampling for the test coincides with that of the offline phase.
This ensures good performances for all the samplings. However, we would like to underline
that adding distribution information can be crucial when dealing with random variables. Indeed,
thanks to some previous knowledge about parameter distribution, one can employ a much lower
number of bases. This is due to the nature of the algorithm that focuses on a very specific
parameter setting both in creating the reduced spaces and in assessing their accuracy. This is the
 12.6. Conclusions 263
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 12.8. Optimal control problem: error decay with respect to the number of RB functions;
comparison between state variable (left) and control variable (right) with an unweighted POD algorithm
between uniform, Beta(20, 5), Beta(5, 5), and Beta(75,75) distributed training samples.

case for µ ∼ Beta(75,75) (rescaled on the respective parameter ranges), where n = 4 is sufficient
to reach error values around 10−6 for both the variables, gaining two orders of magnitude with
respect to uniform sampling. Fewer bases translate to a gain of computational time that can be
exploited in UQ analysis even for very complicated problems such as OCPs, which are usually
characterized by high computational costs due to the minimization framework.

12.6 Conclusions
This chapter focused on wROMs for a broad class of stochastic PDE-based models. First, we
introduced stochastic PDEs with random inputs and their standard discretizations. Then, we
introduced ROM strategies in order to deal with them more quickly, following the POD and the
RB model. The related weighted approaches were also described. Furthermore, we highlighted
the role of several sampling strategies based on different (possibly sparse) quadrature rules. We
validated this setting in many contexts, such as the heat transfer equation, Stokes problems, the
stabilization of advection-dominated problems, and optimal control for environmental sciences.
The results show the better performances of weighted approaches and distributed quadratures
than a standard algorithm based on a deterministic viewpoint.
 Chapter 13
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Reduced Basis,
Embedded Methods, and
Parametrized Level-Set
Geometry

Efthymios N. Karatzas, Giovanni Stabile, Francesco Ballarin,

Gianluigi Rozza

13.1 Introduction and Overview

In this chapter we examine reduced order techniques for geometrical parametrized heat exchange
systems, Poisson, and flows based on Stokes, steady and unsteady incompressible Navier–Stokes,
and Cahn–Hilliard problems. The full-order finite element methods (FEMs), employed in an em-
bedded and/or immersed geometry framework, are the shifted boundary method (SBM) and the
cut finite element method (CutFEM), with applications mainly to fluids. We start by introducing
Nitsche’s method for both SBM/CutFEM and parametrized physical problems as well as high-
fidelity approximation. We continue with the full-order parametrized Nitsche shifted boundary
variational weak formulation and the reduced order modeling (ROM) ideas based on a proper or-
thogonal decomposition (POD)-Galerkin method and geometrical parametrization, quoting the
main differences and advantages with respect to the reference domain approach used for clas-
sical FEMs, while stability issues may overcome employing supremizer enrichment method-
ologies. Numerical experiments verify the efficiency of the introduced “hello world” problems
considering reduced order results in several cases for one-, two-, three-, and four-dimensional
geometrical parametrizations. We investigate execution times, and we illustrate transport meth-
ods and improvements. Several important references related to unfitted methods and ROM are
[322, 325, 326, 327, 324, 323, 332].

13.1.1 Heat Exchange Model Problem

Let us consider the simplest problem, Poisson’s equation:

−∆T = g in D, with T = gD on Γ.

We are looking for a variational formulation that is satisfied by the weak solution T ∈ H 1 (D),
i.e., it is consistent, it is symmetric, and it has a unique solution—the bilinear form is coercive.
We start by taking the strong form of the equation, multiply by a test function v ∈ H 1 (D), and

265
 266 Chapter 13. RB, Embedded Methods, and Parametrized Level-Set Geometry

integrate by parts. Starting with the right-hand side, we add the productive zero 0 = T − gD on
the boundary and obtain
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Z Z
(g, v)D = (−∆T, v)D = (∇T, ∇v)D − n · ∇T v ds − (T − gD )n · ∇v ds.
Γ Γ

Separating linear and bilinear forms, for all v ∈ H 1 (D), the variational equation—the symmetric
bilinear formulation—becomes
Z Z Z Z
(∇T, ∇v)D − n · ∇T v ds − T n · ∇v ds = − gD n · ∇v ds + gv dx.
Γ Γ Γ D

We note that in the above formulation the symmetric bilinear form is not coercive because it
cannot be bounded from below for T = v by ckvk2H 1 (D) . So, we add a symmetric term that
vanishes for the true solution: η Γ (T − gD )v ds for some η > 0 large enough. This leads to
R

the following symmetric, consistent, coercive weak formulation: find T ∈ H 1 (D), ∀v ∈ H 1 (D)
such that
Z Z Z Z
(∇T, ∇v)D − n · ∇T v ds − T n · ∇v ds + η T v ds = − gn · ∇v ds
Γ Γ Γ Γ
Z Z
+η gD v ds + gv dx. (13.1)
Γ D

For finite element (FE) discrete approximations Th , vh ∈ Vh ⊂ H 1 (Ω), and for the Nitsche
penalty chosen as η = ch−1 with c > 0 sufficiently large, one can manage a stable discrete
problem with respect to a suitable mesh-dependent norm.

13.1.2 Shifted Nitsche Boundary Weak Formulation

Starting from the aforementioned Nitsche form (13.1) and based on a closest-point projection
(see, e.g., [325] and references therein), we derive the segmented/faceted nature of the surrogate
boundary. In particular, a smooth mapping M from points in a surrogate boundary Γ̃ to points
in the true boundary Γ is introduced—M : x̃|Γ̃ → x|Γ —where M is defined using a distance
vector function d ≡ dM (x) = x−x̃ = [M − I](x̃); see Figure 13.1(ii). Finally, we arrive at the
shifted boundary weak formulation:
Z Z Z Z
(∇T, ∇v)D̃ − ñ · ∇T v ds − T ñ · ∇v ds + η T v ds = − g̃D ñ · ∇v ds
Γ̃ Γ̃ Γ̃ Γ̃
Z Z
+η g̃D v ds + gv dx
Γ̃ D̃

or, equivalently,
Z Z Z
(∇T, ∇v)D̃ − ñ · ∇T v ds − T ñ · ∇v ds + η T v ds
ZΓ̃ Γ̃
Z Γ̃
Z
=− (T + ∇T · d)ñ · ∇v ds + η (T + ∇T · d)v ds + gv dx,
Γ̃ Γ̃ D̃

where g̃D denotes an extension of the Dirichlet boundary condition gD to the boundary Γ̃ of the
surrogate domain based on a second-order accurate Taylor expansion, T + ∇T · d ≈ g̃D , or, if
we let x = M(x̃), then T (x̃) + ∇T (x̃) · d ≈ gD (x).
 13.1. Introduction and Overview 267

(i) CutFEM
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

(ii) SBM

Figure 13.1. Embedded FEM geometrical tools for (i) CutFEM and (ii) SBM.

13.1.3 The Parametrized Thermal-Heat Exchange Model

We consider a parameter-dependent geometry with a parameter which lives in a k-dimensional
parameter space P, and we have parameter vector µ ∈ P ⊂ Rk . We denote by D(µ) a bounded
parametrized domain depending on µ with boundary Γ(µ). So, the model problem in D(µ) is as
follows: find the temperature T (µ) so that

−∆T (µ) = g(µ) in D(µ), with T (µ) = gD (µ) on Γ(µ), (13.2)

which, using an SBM weak formulation, can be transformed into the system of linear equations
(rewritten in matrix form)
A(µ)T(µ) = Fg (µ), (13.3)

where A(µ) ∈ RNh ×Nh and Fg (µ) ∈ RNh ×1 correspond to the bilinear and linear forms,
respectively.

13.1.4 Model Reduction Methodology

Offline stage. In the offline stage, which is called “training,” one performs a certain number
of full-order solves for various parameters in order to use the solutions to construct a low-
dimensional reduced basis (RB). This basis can approximate any member of the solution set
to a prescribed accuracy, while it is possible to perform a Galerkin projection of the full-order
differential operators, describing the governing equations, onto the RB space. This procedure
involves solving a possibly large number of high-dimensional problems and manipulating high-
dimensional structures. The required computational cost is high, so this operation is usually
performed on a high-performance system such as a computer cluster.

Online stage. This stage is often performed on a system with reduced computational power and
storage capacity; the reduced system can be solved for any new value of the input parameters
with predicted accuracy and reduced computational and time cost. For the interested reader we
refer to [281].
 268 Chapter 13. RB, Embedded Methods, and Parametrized Level-Set Geometry

13.1.5 POD
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Subsequently, with a POD one can generate the RB space after the full-order model (FOM) has
been solved for each µ ∈ K = {µ1 , . . . , µNs } ⊂ Rk , where K is a finite-dimensional training
set of parameters inside the parameter space P, Ns is the number of snapshots, Nh denotes the
number of degrees of freedom for the discrete full-order solution, and the snapshot matrix S is
given by Ns full-order snapshots properly extended to a fixed background mesh and defined on
µi parameter-dependent domains:

S = [T (µ1 ), . . . , T (µNs )] ∈ RNh ×Ns . (13.4)

Given a general scalar function T : D → Rd , with a certain number of realizations T1 , . . . , TNs ,

the POD problem consists of finding, for each value of the dimension of the POD space NP OD =
1, . . . , Ns , the scalar coefficients a11 , . . . , aN Ns
1 , . . . , aNs , . . . , aNs and functions ϕ1 , . . . , ϕNs that
s 1

minimize the quantity

Ns NX 2
X P OD

ENP OD = Ti − aki ϕk ∀NP OD = 1, . . . , N,

i=1 k=1 L2 (D)

with (ϕi , ϕj )L2 (D) = δij , i, j = 1, . . . , Ns . This is equivalent to solving the eigenvalue problem
CQ = Qλ for Cij = (Ti , Tj )L2 (D) , i, j = 1, . . . , Ns , where C is the correlation matrix
obtained by starting from the snapshots S, Q is a square matrix of eigenvectors, and λ is a
diagonalPmatrix of eigenvalues. The basis functions can then be obtained with the formula ϕi =
Ns Nh ×N r
j=1 Tj Qij , and the POD space L = [ϕ1 , . . . , ϕN ] ∈ R for N r < Ns is chosen
1 r
1/2
Ns λii
according to the eigenvalue decay of λ (see [505]).

13.1.6 The Projection Stage and ROM Generation

The reduced solution, then, can be approximated with
r
N
X
r
T ≈ ai (µ)ϕi (x) = La(µ), (13.5)
i=1

r
for which the reduced solution vector a ∈ RN ×1 depends only on the parameter values. The
basis functions ϕi depend only on the physical space, and the unknown vector of coefficients a
can be obtained by using a Galerkin projection of the full-order system of the equations onto the
POD RB space, resulting in the reduced algebraic system

LT A(µ)La(µ) = LT F(µ), (13.6)

which is equivalent to the algebraic reduced system

Ar (µ)a(µ) = Fr (µ), (13.7)

r r r
where Ar (µ) ∈ RN ×N and Fr (µ) ∈ RN ×1 are the reduced discretized operators and reduced
forcing vectors, respectively. We highlight that the dimension of the latter reduced system is
much smaller than the dimension of the full-order system of equations and this is much cheaper
to solve.
 13.1. Introduction and Overview 269

(i)
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

(ii)

(iii)

Figure 13.2. (i) Parametrized geometry. (ii) Some RB components for µ ∈ [−0.5, 0.5]
parametrized geometry. (iii) The full/reduced order solution and the absolute error (µ = −0.015).

13.1.7 Numerical Experiments (Heat Exchange/SBM)

We assume that the embedded domain consists of a rectangle of size 0.8 × 0.7, and its position
inside the domain is parametrized with a geometrical parameter µ. The position (0, µ) of the
rectangular embedded domain depends on its parametrized y center, while the horizontal coor-
dinate of the center of the box is not parametrized and is located at the x center of the domain.
The whole configuration is immersed in a background domain of size D = [−2, 2] × [−1, 1].
The ROM has been trained with 400 samples, which is the dimension of the offline FOM, for
µ ∈ [−0.5, 0.5] chosen randomly inside the parameter space, while the ROM results have been
compared with the FOM for 50 additional random samples. The mean relative errors, the eigen-
value decay, and the behavior of the results for one parameter sample are visualized in Fig-
ures 13.2 and 13.3 and Table 13.1. The time savings are reported in Table 13.2, where the com-
putation time includes assembling the full-order matrices, their projection, and the resolution of
the reduced problem, while we have avoided (a) using a reference domain, (b) remeshing, and
(c) adaptive refinement.
 270 Chapter 13. RB, Embedded Methods, and Parametrized Level-Set Geometry
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

(i)

1 Eigenvalues Decay
10

4
10
Eigenvalues magnitude

7
10

10
10

13
10

16
10

0 50 100 150 200 250 300

index

(ii)

10% Relative error for ROM

Relative error for L2 proj.
Mean Relative Errors

1%

0.1%

5 10 15 20 25 30 35 40 45 50
N of modes

Figure 13.3. Heat exchange problem: (i) eigenvalue decay and (ii) mean relative errors.

Table 13.1. Heat exchange problem: relative error results.

Modes L2 projection Galerkin projection

2 6.45035392e-02 7.10916700e-01
10 4.83332393e-03 2.64459969e-02
20 2.19454585e-03 5.61736266e-03
30 1.27046941e-03 3.30372025e-03
40 7.72326410e-04 2.50189079e-03
50 5.39532759e-04 1.69903034e-03
100 6.79464703e-05 3.36531580e-04
 13.2. Parametrized Steady Stokes Equations 271

Table 13.2. Heat exchange problem: execution time, savings, and speedup.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Savings Speedup
Modes Execution time(s) tFOM −tRB tFOM
tRB tRB
2 4.119470 × 10−2 96.399% 27.770
10 4.168334 × 10−2 96.356% 27.445
20 4.243647 × 10−2 96.290% 26.957
30 4.353909 × 10−2 96.194% 26.275
40 4.449359 × 10−2 96.110% 25.711
50 4.494564 × 10−2 96.071% 25.452
100 4.992923 × 10−2 95.635% 22.912
FOM 1.14540 × 100 – –

13.2 Parametrized Steady Stokes Equations

In this section, we examine a Newtonian incompressible viscous fluid flow in a parametrized
domain D(µ) when the convective forces are negligible with respect to viscous forces, namely
the Stokes system

−∇ · (2ν(u(µ)) − p(µ)I) = g(µ) in D(µ),

∇ · u(µ) = 0 in D(µ),
u(µ) = gD (µ) on Γ(µ),
(2ν(u(µ)) − p(µ)I) · n = gN (µ) on ΓN (µ),

where (u) = 1/2(∇u + ∇uT ) is the velocity strain tensor (i.e., the symmetric gradient of the
velocity), p is the pressure, g is a body force, gD is the value of the velocity on the Dirichlet
boundary, and gN is the normal stress on the Neumann boundary.
Based on [326] we define the following shifted boundary weak formulation: find u ∈ Vh
and p ∈ Qh such that, ∀w ∈ Vh and ∀q ∈ Qh ,

((w), 2ν(u))D̃(µ) − hw ⊗ ñ, 2ν(u)iΓ̃(µ) − h2ν(w), (u + (∇u) · d) ⊗ ñiΓ̃(µ)

+α h2ν/h(w + (∇w) · d), u + (∇u) · diΓ̃(µ) + β h2νh∇τ̄i w, ∇τ̄i uiΓ̃(µ) − (∇ · w, p)D̃(µ)
+ hw · ñ, piΓ̃(µ) = −(w, g)D̃(µ) + hw, gN iΓ̃N (µ) − h2ν(w), ḡD ⊗ ñiΓ̃(µ)
+α h2ν/h(w + (∇w) · d), ḡD iΓ̃(µ) + β h2νh∇τ̄i w, ∇τ̄i ḡD iΓ̃(µ)

and

−(∇ · u, q)D̃(µ) + hu · ñ, qiΓ̃(µ) + hqd ⊗ ñ, ∇uiΓ̃(µ) = hḡD · ñ, qiΓ̃(µ) ,

or, in a more abstract notation,

a(u, w; µ) + b(p, w; µ) = `g (w; µ),

b(q, u; µ) + b̂(q, u; µ) = `q (q; µ),

where n and τi are the unit normal vector and unit tangential vectors to the boundary Γ and can be
 272 Chapter 13. RB, Embedded Methods, and Parametrized Level-Set Geometry
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 13.4. Stokes with SBM experiment: sketch of the mesh, the embedded domain, and the
parameters considered in the numerical examples.

extended to the boundary Γ̃, namely ñ(x̃) ≡ n(M(x̃)) and τ̄i (x̃) ≡ τi (M((x̃)), ḡD (x̃) =
gD (M(x̃)). We clarify that µ is defined similarly in Subsection 13.1.5 and h is a characteristic
length of the elements.

13.2.1 POD Adapted to Flows

Again, we collect Ns , the number of snapshots, with Nuh and Nph the number of degrees of
freedom for the discrete full-order solution for the velocity and pressure, respectively, and we
construct separate bases for both velocity and pressure:
h r h r
Lu = [ϕ1 , . . . , ϕNur ] ∈ RNu ×Nu , Lp = [χ1 , . . . , χNpr ] ∈ RNp ×Np , (13.8)

where Nur , Npr < Ns are chosen according to the eigenvalue decay of the vectors of eigenvalues
λu and λp . Furthermore, for the best approximation we employ supremizer enrichment as in the
work of [511] and we manage a solvable and stable problem that satisfies a reduced and also
parametric version of the inf-sup condition. With this approach, the velocity supremizer basis
N h ×Nsup
r
functions Lsup , with Lsup = [η1 , . . . , ηNsup
r ] ∈ R u , are computed and added to the reduced
velocity space, which is transformed into L̃u :
h r r
N ×(Nu +Nsup )
r ] ∈ R u
L̃u = [ϕ1 , . . . , ϕNur , η1 , . . . , ηNsup . (13.9)

13.2.2 Steady Stokes Numerical Experiments (SBM)

The present experiment considers a steady Stokes flow around an embedded circular cylinder,
within the framework of a parametrized embedded domain described by the level set (x − µ0 )2 +
(y − µ1 )2 ≤ R2 (see Figure 13.4). We will consider two different geometrical parametriza-
tion test cases, a one-dimensional case with parameter µ1 , and a two-dimensional case with
two parameters, µ0 and µ1 (see Figure 13.8 for qualitative results of the simulations). The set-
tings of the problem are an embedded cylinder with constant radius 0.2, viscosity ν = 1, force
f = 1 in the x direction, and uin = 1 on the left side of the domain and an open bound-
ary condition with pout = 0 on the right. A no-slip penetration boundary condition is applied
on the top and bottom edges. On the boundary of the embedded cylinder a no-slip boundary
condition is applied. The mesh size is h = 0.0350 for the background mesh, using 15022 tri-
angles, P1/P1 finite elements in space with stabilization terms, with and without supremizer
basis enrichment in the offline stage. Some RB components derived with POD for the Stokes
system discretized by the SBM for the one- and two-dimensional geometrical parametrization
with µ1 ∈ [−0.65, 0.65] are visualized in Figure 13.5. In parallel, the FOM and ROM so-
lutions and absolute error values are visualized in Figure 13.6 for both experiments. Relative
errors (||u − ur ||L2 /||u||L2 and ||p − pr ||L2 /||p||L2 ), as well as the execution times for the one-
dimensional parametrization, are reported in Tables 13.3 and 13.4 and Figure 13.7(i), while for
 13.2. Parametrized Steady Stokes Equations 273

(i)
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

(ii)

(i)

(ii)

Figure 13.5. Stokes system ROM-SBM POD velocity and pressure components for (i) µ1 geome-
try and (ii) (µ0 , µ1 ) geometry.

the two-dimensional case the relative errors are in Table 13.5 and Figure 13.7(ii). Focusing on
the supremizer stabilization approach, the improved results for pressure are obvious and the plots
clearly show at a glance that the high-fidelity and ROM solutions cannot be easily distinguished.
Related to the execution time investigation for the one-dimensional geometrical parametrization,
we compare online stage execution times against the full-order computational times and FOM
solutions. Namely, for all 1024 snapshots and supremizers, the solutions cost 1 h 42 min 54 s.

Table 13.3. Stokes (SBM): relative errors between the full-order solution and the RB solution,
one-dimensional geometrical parametrization.

Snapshots: 1024 1024

Suprem.: No Yes
Modes rel. error u rel. error p rel. error u rel. error p
8 0.0947158 12.309881 0.2406999 22.319781
12 0.0723268 12.133591 0.2078557 5.7159319
16 0.0610052 9.6652163 0.1692787 2.6962056
20 0.0538906 6.1692750 0.1243368 1.2535779
25 0.0434925 3.2331644 0.0770726 0.5568314
30 0.0396132 1.4693532 0.0437348 0.2504069
35 0.0298269 0.7455038 0.0262345 0.1356788
40 0.0177170 0.2918072 0.0121903 0.0611154
45 0.0085905 0.0923509 0.0060355 0.0330206
50 0.0053882 0.0473412 0.0046300 0.0279857
 274 Chapter 13. RB, Embedded Methods, and Parametrized Level-Set Geometry

(i)
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

(ii)

(i)

(ii)

Figure 13.6. Stokes SBM: the full and the reduced solution and the absolute error for velocity
and pressure for the (i) one-dimensional and (ii) two-dimensional parametrization case.

Table 13.4. Execution time, at the reduced order level, for the case with one-dimensional geo-
metrical parametrization.

no supr. supr.
Modes exec. time (s) exec. time (s)
8 7.3858961 7.710907
12 7.6042165 8.091225
16 7.9584049 8.290780
20 8.0206915 9.036709
25 8.2229143 9.495323
30 8.9529275 9.972288
35 9.0867916 10.47633
40 9.6555775 11.13931
45 9.8934008 11.49422
50 10.302459 11.92024
 13.2. Parametrized Steady Stokes Equations 275

u
u, supremizers
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

p
p, supremizers

100%

relative error
(i)a
10%

0.1%

10 20 30 40 50
N

truth
reduced
reduced (supr.)
40

30
execution time

(i)b 20

10

10 20 30 40 50
N

u, 1024 snapshots
u 1024 snapshots, supremizers
u, 900 snapshots
u 900 snapshots, supremizers
100%
relative error U

(ii)a 10%

1%

20 40 60 80 100
N

p, 1024 snapshots
p, 1024 snapshots, supremizers
p, 900 snapshots
p, 900 snapshots, supremizers

100%
relative error P

(ii)b

10%

20 40 60 80 100
N

Figure 13.7. Stokes (SBM) parametrization: (i) µ1 geometry, with the relative errors for velocity
and pressure in (i)a and the execution times in (i)b ; (ii) (µ0 , µ1 ) geometry, with the velocity and pressure
fields with and without supremizer stabilization.
 276 Chapter 13. RB, Embedded Methods, and Parametrized Level-Set Geometry

Table 13.5. Stokes SBM (µ0 , µ1 ) geometry: Supremizer basis enrichment and the relative error.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Snapshots: 900 1024

Suprem.: Yes Yes
Modes rel. error u rel. error p rel. error u rel. error p
10 0.2448511 3.9240637 0.1672753 2.1243228
20 0.2175821 2.6531343 0.1353706 1.3611011
30 0.1652331 1.6234701 0.1124619 0.9680506
40 0.1340978 1.1560352 0.0696437 0.4958605
50 0.1158443 0.7777786 0.0444991 0.2958338
60 0.1013961 0.5876048 0.0244793 0.1574037
70 0.0914650 0.4335489 0.0151749 0.0928402
80 0.0822658 0.2933336 0.0097848 0.0434299
90 0.0744696 0.1355488 0.0076431 0.0257060
100 0.0660493 0.0714350 0.0037280 0.0194051
110 0.0609040 0.0675720 0.0031577 0.0174815

This is an expensive stage, but fortunately it is executed only once at the beginning. We clarify
that the online stage computation time includes assembling the full-order matrices and product-
ing the ROM and its resolution (see Table 13.4). Ten different values of the input parameter
are considered, with reference time the time execution at the FOM level, which for each one-
parameter solution is equal to ≈ 37 s. Obviously, the ROM leads to a considerable speedup
for all the different analyzed configurations and for cases both with and without supremizer en-
richment. The interest in the latter experiments was more about testing the feasibility and the
accuracy of a ROM, constructed starting from a shifted boundary FOM, and we did not employ
any hyperreduction techniques, which means that, at the reduced order level, we also assembled
the full-order discretized differential operators. Focusing on the (µ0 , µ1 ) geometry experiment
with range µ = (µ0 , µ1 ) = [−1.5, −1.0] × [−0.15, 0.15], which is a more complex and demand-
ing scenario, we can easily notice that the supremizer enrichment is necessary for convergence
stability and reliable pressure results, and without supremizer enrichment and 900 snapshots, the
best achieved—and most disappointing—relative errors were equal to 0.0263821 and 0.1854861,
respectively, for velocity and pressure. Figure 13.7 and Table 13.5 also give an overview of the
dependence on the number of snapshots for two choices of 900 and 1024 used in the offline stage.

13.3 Looking for a Better ROM with CutFEM

In embedded methods, the reduced order basis is constructed on the whole background domain,
and great care is needed to manipulate the truth geometry area outside the area of interest, the so
called ghost area. In this section, with CutFEM, we employ the solution values of the surrogate
boundary that were computed using the natural smooth extension from the truth to the surrogate
domain, which allows a smooth extension of the solution to the neighboring ghost elements with
values decreasing smoothly to zero. The latter guarantees a regular “solution” in the background
domain and constructs a good reduced order basis and a much more promising reduced order
approximation. Next we present the weak formulation for the Poisson equation by introducing
the CutFEM basics.
For any real parameter vector µ we seek u(µ) in

VgD (µ) = w ∈ H 1 (D(µ)) with w|Γ(µ) = gD (µ) ,



such that ∀v(µ) ∈ V0 (µ),

(∇u(µ), ∇υ(µ)) = (g(µ), υ(µ)) + (gN (µ), υ(µ))ΓN (µ) .

 13.3. Looking for a Better ROM with CutFEM 277
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 13.8. Shape parametrization with large deformations and a zoom into the embedded
cylinder visualizing the extended solution.

The boundary value problem is formulated on a domain T (µ) that contains D(µ) ⊂ T (µ),
whose mesh Th (µ) is not fitted to the domain boundary, where Gh (µ) := {K ∈ Th (µ) : K ∩
Γ(µ) 6= ∅} is the set of elements that are intersected by the interface, DT (µ) := {K ∈ Th (µ) :
K ∩ D(µ)} ∪ Gh (µ); the background domain is typified by B, while its corresponding mesh is
denoted by Bh , such that DT (µ) ⊂ B and Th (µ) ⊂ Bh for all µ ∈ K. See also Figure 13.1(ii).
We remark that Th (µ), Gh (µ), and DT (µ) depend on µ through D(µ) or its boundary, while the
background domain B and its mesh Bh do not depend on µ. Furthermore, the set of element faces
FG (µ) associated with Gh (µ) is defined as follows: for each face F ∈ FG (µ), there exist two
simplices K 6= K 0 such that F = K ∩K 0 and at least one of the two is a member of Gh (µ). Note
that the boundary faces of Th (µ) are excluded from FG (µ). On a face F ∈ FG (µ), F = K ∩K 0 ,
the jump of the gradient of v ∈ C 0 (DT ) is defined by [[nF · ∇v]] = nF · ∇v|K − nF · ∇v|K 0 ,
where nF denotes the outward-pointing unit normal vector with respect to K. So the CutFEM
discretization is as follows: we seek a discrete solution uh (µ) in the FE space

Vh (µ) = υ ∈ C 0 (DT (µ)) : υ|K ∈ P 1 (K) ∀K ∈ Th (µ)



such that ∀υh (µ) ∈ Vh (µ),

αh (uh (µ), vh (µ)) = `h (vh (µ)),

αh (uh (µ), vh (µ)) = (∇uh (µ), ∇vh (µ))D(µ) − (nΓ · ∇uh (µ), υh (µ))Γ(µ)
− (uh (µ), nΓ · ∇υh (µ))Γ(µ) + γD h−1 uh (µ), vh (µ) Γ(µ)

+ (γN hnΓ · ∇uh (µ), nΓ · ∇vh (µ))ΓN (µ) + j(uh (µ), vh (µ)),
h
` (vh (µ)) = (g(µ), υh (µ))D(µ) − (gD , nΓ · ∇υh (µ))Γ(µ)
+ γD h−1 gD , vh (µ) Γ(µ) + (gN (µ), vh (µ)+γN hnΓ · ∇vh (µ))ΓN (µ) ,

where the stabilization term

X
j(uh (µ), vh (µ)) = (γ1 h[[nF · ∇uh (µ)]], [[nF · ∇vh (µ)]])F
F ∈FG

extends the coercivity from the physical domain D(µ) to the µ-dependent mesh domain DT . γD ,
γN , and γ1 are positive penalty parameters. The coefficients γD and γN account for a Nitsche
weak imposition of boundary conditions. We set up some experiments again for embedded
 278 Chapter 13. RB, Embedded Methods, and Parametrized Level-Set Geometry

(i) Standard POD modes (no preprocessing)

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

(ii) POD modes with preprocessing

Figure 13.9. Six classical (i) and six improved (ii) POD modes.

FEMs and ROMs, emphasizing improvements in cases with large geometrical deformations. The
parametrized domain D(µ) ⊂ R2 is an ellipse described by the level set φ(x, y; µ1 , µ2 , µ3 , µ4 ) =
µ22 (x − µ3 )2 + µ21 (y − µ4 )2 − µ21 µ22 R, where the reference radius R = 0.05, the length of
the axes of the ellipse is (µ1 , µ2 ) ∈ [0.3, 1.8]2 , and the position of the center of the ellipse is
(µ3 , µ4 ) ∈ [−0.85, 0.85]2 . A corresponding background domain B = [−1.2, 1.2]2 is chosen so
that the ellipse is strictly contained in B for any µ = (µ1 , µ2 , µ3 , µ4 ) in the parametric range
K = [0.3, 1.8]2 × [−0.85, 0.85]2 . The value µ = (1, 1, 0, 0), corresponding to a circle of radius
R centered at the origin, is chosen for the transport method. The data of the problem described
by (13.2) comprises the force g(x, y; µ) = 20 and the Dirichlet boundary force gD (x, y; µ) =
0.5 + xy.
In Figure 13.10(i) for the CutFEM Poisson system, the reduced solution obtained from the
zero extension is inaccurate even for N = 140 modes, being affected by relative errors of the
order of 10−1 . A nonzero extension is beneficial, resulting in relative errors on the order of 10−2 ,
for the maximum value of N . The combination with inverse transportation allows us to further
improve the results, up to errors of 10−4 for N = 140 modes in the case of a POD basis obtained
 13.3. Looking for a Better ROM with CutFEM 279
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

1
10

4
10

7
10
N / max

(i) 10 10

13 zero extension
10
harmonic extension
natural smooth extension
16
10 transport and harmonic extension
transport and natural smooth extension
19
10

0 50 100 150 200 250 300 350 400

N

zero extension
harmonic extension
natural smooth extension
transport and harmonic extension
22% transport and natural smooth extension
13%
Relative error

2%

0.8%

0.2%

0.04%

0 20 40 60 80 100 120 140

N

u - natural smooth extension

p - natural smooth extension
u - transport and nat. smooth ext.
68% p - transport and nat. smooth ext.
Relative error

(ii) 35%

17%

10%

3%

0 10 20 30 40 50
N

Figure 13.10. (i) Poisson system (CutFEM): eigenvalue decay and error analysis between re-
duced order and high-fidelity approximations with and without transport; (ii) steady Stokes (CutFEM):
relative errors with and without transport.
 280 Chapter 13. RB, Embedded Methods, and Parametrized Level-Set Geometry
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 13.11. Circular embedded geometry (parametrized).

from transport and natural smooth extension. Thus, the pivotal role of snapshot transportation can
be inferred from these results, being capable of improving the accuracy of almost three orders of
magnitude compared to the simplest zero extension. Figure 13.9 shows the first six POD modes
with and without preprocessing. Nonetheless, all methods reach a plateau after which no further
improvement is shown. We claim that this is due to integration errors occurring on Γ(µ) and
Nitsche weak imposition of the Dirichlet boundary conditions, e.g., the maximum values of the
error are consistently attained on the boundary.
Further investigation has shown similar good behavior for fluid flow systems, namely steady
Stokes and RB with transportation and CutFEM [322], with relative errors without and with
transport improvements visualized in Figure 13.10(ii). Once more, the transported snapshots and
the inverse transported modes appear beneficial for both velocity and pressure, needing very few
modes; in particular it manages to reach its best accuracy with only three basis components.

13.4 ROM and a Fourth-Order Evolutionary Nonlinear

System
In this section, we consider the Cahn–Hilliard model problem describing phase flow time evo-
lution. An unknown function u indicates the perturbation of the concentration of one of the
phases of, e.g., fluid components constituting a liquid mixture which contains a binary fluid. As
first suggested by [144] and thereafter extended in [89], assuming that the mobility is equal
to one and ε is a measure of the size of the
interface of two fluids, then the mass flux is
given by J(µ) = −∇ ε12 F (u(µ)) − ε2 ∆u(µ) , where F denotes the chemical potential dif-
ference between the two species. From the latter we can derive the Ginzburg–Landau energy
2
E(u(µ)) = D (F (u(µ)) + ε2 |∇u(µ)|2 )dx.
R

The equilibrium state of the considered mixture minimizes the above Ginzburg–Landau en-
ergy subject to mass conservation: ∂t
∂
= 0. Hence, the parametrized Cahn–Hilliard
R
D(µ)
u(µ)dx
system can be described as
∂u(µ) 1
= −ε2 ∆2 u(µ) + 2 ∆F 0 (u(µ)) in D(µ) × [0, T ], (13.10)
∂t ε
1
∂n u(µ) = ∂n (−ε2 ∆u(µ) + 2 F 0 (u(µ))) = gN (µ) on Γ(µ) × [0, T ], (13.11)

u(·, 0) = u0 (·) in D(µ), (13.12)

where n is the unit outward normal vector of Γ and F is a double-well function of u, usually a
fourth-power polynomial:

u4 (µ) u3 (µ) u2 (µ)

F (u(µ)) = γ2 + γ1 + γ0 with γ2 > 0. (13.13)
4 3 2
 13.4. ROM and a Fourth-Order Evolutionary Nonlinear System 281
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 13.12. Cahn–Hilliard (CutFEM): ROM basis results, the first six modes.

Figure 13.13. Results for the embedded circle geometrical parametrization µ = 0.4261 and
t = [28, 36, 46, 100]dt.

Starting from the above initial form and based on a splitting method and Nitsche boundary
enforcement and with efficient CutFEM stabilization, we manage a proper weak form of H 1
space regularity; for more details we refer the reader to [324]. The CutFEM mesh and level-
set geometry can be seen in Figure 13.11, while some ROM basis components can be seen
in Figure 13.12, and the concentration field in the FOM and the ROM and the absolute error
level after parametrization of the embedded circle are visualized for the time instances t =
[28, 36, 46, 100]dt and for a randomly selected parameter µ = 0.4261 in Figure 13.13. The full-
order and the reduced mass evolution with respect to time verifies the conservation of mass for
the proper number of modes and for the truth solver and parameter µ = 0.4261. The experiment
took place for time instances t = ndt, n = 1, . . . , 100; see Figure 13.14. We mention that we
trained our basis with 900 snapshots in the parameter range for the diameter µtest ∈ [0.36, 0.48].
 282 Chapter 13. RB, Embedded Methods, and Parametrized Level-Set Geometry
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

1
5.38 × 10

1
5.375 × 10

1
5.37 × 10
Mass(t)

3 modes ROM
5 modes ROM
1
5.365 × 10 7 modes ROM
11 modes ROM
21 modes ROM
1
5.36 × 10
29 modes ROM
41 modes ROM
50 modes ROM
1
5.355 × 10
FOM

0 20 40 60 80 100
Time

Figure 13.14. Cahn–Hilliard (CutFEM): the FOM mass evolution with respect to time, together
with its RB approximation and the conservation of mass.

13.5 Conclusions and Future Developments

In this chapter, we introduced a POD-Galerkin ROM based on SBM/CutFEM high-fidelity sim-
ulations, for linear and nonlinear steady and unsteady PDE problems, characterized by a ge-
ometrical parametrization with possibly large deformations. The embedded boundary method
discretization naturally allows one to use a level-set description of the parametrized geometry.
In our opinion, these results were derived in a simpler and more versatile high-fidelity and ROM
method than the FE formulation with pullback to a reference domain. The transportation ap-
proach and the developed ROM accurately to reproduce the high-fidelity solution, with relative
errors of the order of 10−4 , and for the Stokes case of the order of 10−3 , employing only four
modes for both velocity and pressure. Also, efficiency in nonlinear time-dependent systems,
namely Cahn–Hilliard, has been verified.
To add some perspective, we mention, since the proposed ROM is not offline-online separable
in the usual sense, the empirical interpolation method (EIM) and application of greedy algorithms
during the generation of the RB space. As a further future development, we mention the extension
of the proposed ROM to a more general framework of nonlinear problems in fluid dynamics, as
well as fluid-structure interaction (FSI) problems, multiphase flow and Navier–Stokes coupled
systems, multiphysics, coupled processes or systems like heat transfer, stress and strain, optimal
control in hydrodynamics, chemical reaction systems, and shallow-water flow systems.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Chapter 14

Reduced Order Methods

for Fluid-Structure
Interaction Problems

Monica Nonino, Francesco Ballarin, Gianluigi Rozza

14.1 Introduction
This chapter is entirely devoted to the presentation of some reduced order model (ROM) pro-
cedures that are tailored for fluid–structure interaction (FSI) problems. Despite their intrinsic
complexity (see [185, 242]), FSI problems are a widespread topic in the applied mathematics
community: in naval engineering, FSI can be used to study and simulate the interaction between
the water and the hull of a ship (see, e.g., [379]); in a biomedical framework, multiphysics can
model the interaction between blood flow and the deformable walls of a vessel, as well as the be-
havior of blood flow in the presence of serious pathologies (as an example of the implementation
of FSI in the medical field, see [607, 466, 32, 469, 476, 384]); in aeronautical engineering, FSI
models the interaction of the air with a plane or with (parts of) a shuttle (see [478, 165, 187, 373]).
These are just some of the many applications for FSI, and each of these problems has different
characteristics that lead to different dynamics of the coupled system: e.g., the aerospace ap-
plications are in the setting of high Reynolds numbers and a turbulent regime; in addition, the
fluid is usually compressible and has a behavior that requires a three-dimensional model. On the
other hand, for biomedical applications, we are usually in a laminar regime, so the models can
be either two- or three-dimensional, and blood is usually modeled as an incompressible fluid.
Another very important difference in all these problems is represented by the physical quantities
describing the solid and the fluid, and mainly the solid density ρs and its ratio with respect to
the density ρf of the fluid under consideration: there are some algorithms for solving FSI prob-
lems that may work very well for a model that couples air with a very stiff material but cease
to work when the problem couples blood flow with biological tissue. Other difficulties come
from the different assumptions on the structure, and hence the different dynamics that one can
obtain; see, e.g., [47]. To summarize, FSI problems give rise to a wide range of different situ-
ations and different dynamics, requiring different computational tools, depending on the setting
in which we are working: the aforementioned difficulties are just some of the many peculiarities
of multiphysics problems that contribute to making it a very difficult yet extremely fascinating
topic of research.
In this chapter, according to the approach adopted to address a multiphysics problem, we
will see how we can adapt, modify, and improve the reduced basis method (RBM); all the ROMs

283
 284 Chapter 14. Reduced Order Methods for Fluid-Structure Interaction Problems

presented are based on a proper orthogonal decomposition (POD) [249, 351, 482, 620]. In the
following, we will consider two different problems of interest: in the first test case we model the
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

behavior of two leaflets that bend under the influence of a fluid jet, whereas in the second test
case the solution of the FSI problem exhibits a transport-dominated behavior.

14.2 Dynamics of FSI Problems

The dynamics of FSI problems is based on the interplay between the fluid and the solid under
consideration: this interplay is possible thanks to the coupling of the two different physics at
the fluid-structure interface, which is the portion of the physical domain that is common to the
fluid subdomain and the solid subdomain. This coupling is the consequence of three different
principles:

• Continuity of the displacement: This condition imposes continuity on the solid displace-
ment and the fluid displacement at the fluid-structure interface. This geometrical condition
is the mathematical translation of the hypothesis that the fluid and the solid domains do
not overlap.

• Continuity of the velocity: This is a kinematic condition that represents the hypothesis
that the fluid sticks to the moving fluid-structure interface. This condition resembles in
some ways the no-slip boundary condition, which is very common in viscous fluid dy-
namics. In aeroelasticity applications, the continuity of the velocity at the fluid-structure
interface is relaxed and substituted with the nonpenetrating condition, which prescribes
the motion in the direction normal to the fluid-structure interface.

• Balance of the stresses: This is a classical action-reaction principle that imposes the
balance between the fluid and the solid stresses at the fluid-structure interface.

These coupling conditions are imposed at the fluid-structure interface: one of the big challenges
of coupled problems is the fact that this interface is not fixed; instead it deforms in time, and
the deformation is not known a priori. Another difficulty is the fact that the fluid domain in
FSI applications is a moving domain: in solid mechanics it is common to deal with moving
domains, where the deformation is usually the unknown of the problem. For fluid dynamics, on
the other hand, one usually considers fixed domains. This different point of view is an intrinsic
characteristic of FSI problems, and it gives rise to a formalism that is very well known and widely
used in the community, called the arbitrary Lagrangian Eulerian (ALE) formulation.

14.2.1 The ALE Formulation

We begin by setting the notation that will be used throughout the chapter. Ω(t) ⊂ R2 is the
physical domain of the FSI problem: Ω(t) = Ωf (t) ∪ Ωs (t), where Ωf (t) ⊂ R2 and Ωs (t) ⊂ R2
are, respectively, the fluid and the solid domain at time t; an important assumption is that the
two domains do not overlap, i.e., Ωf (t) ∩ Ωs (t) = ∅. The fluid-structure interface is defined as
ΓF SI (t) := Ω̄f (t) ∩ Ω̄s (t). The arbitrary Lagrangian Eulerian (ALE) formulation [487, 173,
296, 46] is widely used to simulate FSI problems: this formulation is needed because of the two
different approaches that are usually adopted to describe the motion of a fluid and the motion of
a solid. The Lagrangian formalism, or Lagrangian point of view, is usually employed to describe
the behavior of a solid: the conservation laws (mass conservation, momentum conservation) are
formulated on the reference configuration Ω̂s = Ωs (t = 0), i.e., on the solid configuration at the
starting time t = 0. This is made possible by the fact that it is rather easy to define, and identify,
 14.2. Dynamics of FSI Problems 285

x
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Ω̂ Ω(t)

Figure 14.1. Example: domain reference configuration Ω̂ (left) and domain configuration at time
t, Ω(t) (right). In blue we have the fluid domain; in orange the solid domain. In green is the fluid-structure
interface Γ̂F SI in the reference configuration.

the “initial position” of a given point inside the solid physical domain. On the contrary, when
describing the behavior of a fluid, the definition of “domain reference configuration” or “original
position of a particle” is not clear, and therefore the Eulerian formalism is used instead: the
conservation laws are formulated on the configuration Ωf (t) at the current time t. In order to be
able to describe both the fluid and the solid behavior, a mixed formulation (the ALE formulation)
is used. The underlying idea is that we can pull back the fluid equations, rewriting them in an
arbitrary time-independent configuration Ω̂f : one possible choice for Ω̂f is Ω̂f = Ωf (t = 0),
the domain at the initial time. In Figure 14.1 we can see an example of a reference configuration
and the configuration of the domain at the current time t. Let [0, T ] be a time interval, and let Ω̂f
be a reference configuration for the fluid.

Definition 14.1. The ALE mapping Af (t) for every t ∈ [0, T ] is defined as follows:

Af (t): Ω̂f 7→ Ωf (t),

x̂ 7→ x = x̂ + dˆf (x̂, t). (14.1)

Here dˆf (t): Ω̂f 7→ Ω̂f is the mesh or fluid displacement. The definition of dˆf usually changes
depending on the kind of fluid problem we want to model, and also depending on the degree of
regularity that we would like to see in the solution of the coupled system; usually the mesh
displacement is defined as an extension of the solid displacement to the whole fluid domain.
Let Ω̂s ⊂ R2 be the reference configuration for the solid, and let dˆs (t): Ω̂s 7→ R2 be the solid
displacement: one possible way to define dˆf is by a harmonic extension of dˆs :

ˆ dˆf = 0 in Ω̂f ,

−∆
(14.2)
df = dˆs on Γ̂F SI ,
ˆ

where Γ̂F SI is the fluid-structure interface in the reference configuration and ∆ ˆ denotes the
gradient with respect to the coordinates in the reference configuration. For other alternatives on
how to define dˆf , we refer the reader to [487].
We underline that dˆf represents the displacement of the grid points, so it is not a quantity
with a real physical meaning but rather a geometrical quantity that describes the deformation of
the mesh according to the deformation of the physical domain. It is also important to underline
that ∂t dˆf 6= ûf ; in fact, while ûf represents the velocity of the fluid, ∂t dˆf is again a geometrical
quantity that can be interpreted as the velocity with which the mesh moves.
Great attention has to be paid to the definition of the mesh displacement because different
definitions for dˆf lead to different levels of regularity: if we lose regularity due to the mesh
displacement, then we lose regularity at the fluid-structure interface, which is exactly where the
coupling between the two physics takes place. It is beyond the scope of this chapter to discuss
 286 Chapter 14. Reduced Order Methods for Fluid-Structure Interaction Problems

the regularity of different definitions of the mesh displacement; we refer the interested reader to
Subsection 5.3.5 of [487].
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

14.3 Approaches to FSI Problems

To solve an FSI problem from the numerical point of view we can adopt two different approaches:
the first approach is a partitioned, or segregated, procedure [37, 193, 191, 194, 192, 104], whereas
the second approach is the monolithic procedure [36, 625]. In the following we briefly introduce
partitioned and monolithic algorithms because they will be employed in the reduction procedures
explained in this chapter.

14.3.1 Partitioned Algorithms

Existing simulation tools for fluid dynamics and for structural dynamics are well developed and
are used on a daily basis, e.g., in industrial applications. It is therefore important to combine
these computational tools in one whole procedure to address coupled problems: this is exactly
the idea behind a partitioned algorithm. Indeed, in a partitioned procedure, we solve separately
the fluid and the solid problems, and then we couple the two physics with an iterative procedure;
see, e.g., [241, 37]. Even though the advantage of a partitioned algorithm is the possibility of
combining different discretization tools for the two physics (e.g., finite volumes for the fluid and
finite elements (FEs) for the structure), the drawback of these procedures is that, under some
physical and geometrical conditions, they turn out to be unstable: this happens, for example, if
the physical domain has a slender shape, or if the fluid density ρf is close to the solid density ρs ,
which is usually the case in hemodynamics applications, where the density of the blood is quite
close to the density of the walls of the vessel. The reason for this instability is the added mass
effect: the fluid acts like an added mass to the solid, thus changing its natural behavior; we refer
the reader to [102] for a detailed derivation of the added mass effect and related consequences.
Partitioned algorithms differ from one another according to the strategy used to impose the
coupling conditions at the fluid-structure interface; indeed, we can categorize partitioned algo-
rithms as follows:

• Explicit algorithms: After having discretized the FSI problem in time, the conditions
on the continuity of the displacement and on the continuity of the velocity at the inter-
face are treated explicitly at every time step. These algorithms, also known as weakly
or loosely coupled algorithms [84], are successfully applied in aerodynamics applications
(see [188, 455]), but some studies (see [102, 305, 365]) show that they are unstable under
some physical and geometrical conditions due to the added mass effect, as we previously
mentioned.
• Implicit algorithms: In these algorithms, also known as strongly coupled algorithms,
the coupling conditions are treated implicitly at every time step (see, e.g., [605, 606]).
This implicit coupling represents a way to circumvent the instability problems due to the
added mass effect; nevertheless, an implicit treatment of the coupling conditions leads to
algorithms that are more expensive in terms of computational time.
• Semi-implicit algorithms: In these algorithms (see [27, 28, 45]), the continuity of the dis-
placement is treated explicitly, whereas the other coupling conditions are treated implicitly.
This alternative represents a trade-off between the computational cost of the algorithm and
its stability in relation to the physical and geometrical properties of the problem. In this
chapter we will see a reduced order method that is based on this kind of partitioned ap-
proach.
 14.4. Partitioned RBM 287

14.3.2 Monolithic Algorithms

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

In a monolithic algorithm (see, e.g., [36, 487, 216, 201, 29]), the fluid and the solid problem
are solved simultaneously. This usually results in algorithms that are more stable, which is
highly desirable, especially if we wish to use large time steps in our simulations. The main
drawback is that they deeply rely on the availability of ad hoc software that can be used to solve
the fluid problem and the solid problem; in this sense, monolithic algorithms are less flexible and
more tailored to the particular problem at hand. Moreover, the coupling conditions at the fluid-
structure interface are more complicated to treat; in order to pursue a Galerkin discretization of
the original problem of interest, we will use two Lagrange multipliers to impose the continuity
of the displacement and of the velocity at the interface: this results in the introduction of two
new unknowns in the coupled problem.

14.4 Partitioned RBM

We now present an algorithm that combines the RBM with a partitioned approach to solve mul-
tiphysics problems. Partitioned algorithms, as we have previously mentioned, are particularly
useful when we want to couple different discretization tools for the fluid problem and for the
solid problem. In addition, from the ROM point of view, addressing a coupled problem by
means of a partitioned procedure is convenient in terms of computational efficiency: indeed, in
the online phase of the RBM, we have to solve, separately, smaller systems. Moreover, with
some minor changes, such as change of variables and appropriate choices for the couplings, it is
possible to further reduce the dimension of the online systems, as we will see in the following.
The goal of the work presented in this section is to combine a segregated procedure with a
ROM and to test and analyze the performance of the obtained algorithm.

14.4.1 Test Case: Leaflets Bending under the Influence of a Fluid

We want to simulate the bending behavior of two leaflets under the influence of an incompress-
ible fluid in the time interval [0, T ]. The length of the leaflets is given by the parameter µg ,
whereas the physical parameter µs represents the leaflets’ shear modulus: Figure 14.2 shows the
physical domain in its original configuration. Before stating the problem formulation, we need
to introduce some notation that is necessary in order to handle a domain which has a geometrical
parametrization.

ΓD
s Γtop

Γin Γout

µs ←
}µg

ΓD
s Γbottom

Figure 14.2. Original configuration at time t. Blue domain: the original fluid configuration
Ωf (t). Orange leaflets: the original solid configuration Ωs (t). ΓD
s : the part of the leaflets that does not
move.
 288 Chapter 14. Reduced Order Methods for Fluid-Structure Interaction Problems

ALE Formulation in the Presence of Shape Parametrization

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Let us denote by µ = (µg , µs ) the parameter vector, with µ ∈ P ⊂ R2 , where P := [µming , µmax
g ]
× [µs , µs ] is the parameter domain. Let us denote by Ω(t; µg ) := Ωf (t; µg ) ∪ Ωs (t; µg )
min max

the current physical domain: we now have a time dependence and a parameter dependence.
We introduce the time-independent intermediate configuration Ω̃(µg ) := Ω̃f (µg ) ∪ Ω̃s (µg ),
where we are considering the reference configuration of both physics, still taking into account the
parameter dependence. Finally, we have the time-independent, parameter-independent reference
configuration Ω̂ := Ω̂f ∪ Ω̂s .
We call T the shape parametrization map: for every µg ∈ [µming , µmax
g ], we have a map Tµg
defined as follows:

Tµg : Ω̂ 7→ Ω̃(µg ),
x̂ 7→ x̃ = Tµg (x̂).

We then have the ALE map Af (t; µg ), introduced in Subsection 14.2.1, which is now a map from
the current parametrized fluid configuration Ω̂f (t; µg ) to the intermediate fluid configuration
Ω̃f (µg ):

Af (t; µg ): Ω̃f (µg ) 7→ Ωf (t; µg ),

x̃ 7→ x = x̃ + d˜f (x̃, t; µg ),

where d˜f is the mesh displacement defined in Subsection 14.2.1.

Let us define the gradients and the determinants of the deformation maps:

G(x̂; µg ) = ∇Tˆ µ (x̂), K(x̂; µg ) = detG(x̂; µg ),

g

˜ +∇
F̃ (x̃; µg ) = Id ˜ d˜f , ˜ x̃; µg ) = detF̃ ,
J(

where now ∇ ˜ represents the gradient with respect to the coordinates x̃ of the intermediate con-
figuration. We can pull back the gradient F̃ (x̃; µg ) to the reference domain Ω̂f , and we obtain
F (x̂; µg ) = Id + ∇ ˆ dˆf G−1 (x̂, µg ). With this notation, we can conclude that the gradient of
the deformation map from the reference configuration to the current configuration is given by
F (x̂, µg )G(x̂, µg ); for the sake of simplicity of notation let us denote by Fµg and Gµg the
gradients F (x̂, µg ) and G(x̂, µg ), respectively.

Problem Formulation
With the notation introduced in the previous paragraph we are now ready to state the strong
formulation in the current configuration: for every t ∈ [0, T ] and for every µ ∈ P, find the
fluid velocity uf (t; µ): Ωf (t; µ) 7→ R2 , the fluid pressure pf (t; µ): Ωf (t; µ) 7→ R, the mesh
displacement d˜f (t; µ): Ω̃f (µ) 7→ R2 , and the solid displacement d˜s (t; µ): Ω̃s (µ) 7→ R2 such
that
˜ d˜f = 0 in Ω̃f (µ) × [0, T ],

−∆
(14.3)
˜
df = d˜s on Γ̃F SI × [0, T ]
and

 ρf ∂t uf |x̃ + ρf (uf − ∂t df |x̃ ) · ∇uf − divσf (uf , pf ) = bf in Ωf (t; µ) × [0, T ],


divuf = 0 in Ωf (t; µ) × [0, T ],
ρs ∂tt d˜s − div
˜ P̃ (d˜s , µs ) = b̃s in Ω̃s (µ) × [0, T ],

(14.4)
 14.4. Partitioned RBM 289

where ∂t df |x̃ (x; t) = ∂t d˜f (x̃; t) is the ALE derivative (and similar for ∂t uf |x̃ ), ρf is the fluid
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

density, ρs is the solid density, and bf and b̃s are the external volume forces acting on the fluid
and on the solid, respectively, on the domains Ωf (t; µ) and Ω̃s (µ). In system (14.4), div() is the
divergence with respect to the coordinates x, whereas div() ˜ is the divergence with respect to the
intermediate coordinates x̃. σf is the fluid Cauchy stress tensor in the configuration Ωf (t; µ):

σf (uf , pf ) := µf (∇uf + ∇T uf ) − pf I, (14.5)

with I the 2 × 2 identity matrix and µf the fluid viscosity. P̃ is the second Piola–Kirchoff stress
tensor expressed in the intermediate configuration:

P̃ (d˜s , µs ) := λs trεs (d˜s ) + 2µs ε(d˜s ), µs ∈ [µmin

s , µmax
s ], (14.6)

with λs the first Lamé constant. System (14.4) is completed by the coupling conditions

 df = ds on ΓF SI (t; µg ) × [0, T ],


uf = dt d
ds on ΓF SI (t; µg ) × [0, T ], (14.7)
J σ̃f (ũf , p̃f )F̃ −T ñf = P̃ (d˜s )ñs on ΓF SI (t; µg ) × [0, T ]
 ˜

and by the boundary conditions

 σf (uf , pf )n = −pin (t)n in Γin × (0, T ],



σf (uf , pf )n = −pout (t)n in Γout × (0, T ],

(14.8)
 u f = 0 in Γtop ∪ Γbottom × (0, T ],
ds = 0 in ΓD

s × (0, T ],


where ñf and ñs are the normal vectors to the fluid-structure interface Γ̃F SI in the intermediate
configuration, out of the fluid domain and the solid domain, respectively. Moreover,

˜ ũf F̃ −1 + F̃ −T ∇
σ˜f (ũf , p̃f ) := µf (∇ ˜ T ũf ) − p̃f I (14.9)

is the Cauchy stress tensor in the fluid intermediate configuration Ω̃f .

Thanks to the pullback maps already introduced in the previous paragraph, we can formulate
the problem in the reference configuration Ω̂. For every t ∈ [0, T ], and for every µ ∈ P, find the
fluid velocity ûf (t, µ) : Ω̂f 7→ R2 , the fluid pressure p̂f (t, µ) : Ω̂f 7→ R, the mesh displacement
dˆf (t, µ): Ω̂f 7→ R2 , and the solid deformation dˆs (t, µ) : Ω̂s 7→ R2 such that

ρf JK(∂t ûf + ∇ ˆ ûf G−1 −1 ˆ

 µg Fµg (ûf − ∂t df ))


 ˆ
 − div(JK σ̂f (ûf , p̂f )Fµg Gµg ) = b̂f in Ω̂f × [0, T ],
−T −T

ˆ
div(JKG µg Fµg uf ) = 0 in Ω̂f × [0, T ],
−1 −1

− div(K ∇dˆf G−1

ˆ ˆ
µg Gµg ) = 0 in Ω̂f × [0, T ],
−T




ρs K∂tt dˆs − div(K P̂ (dˆs , µs )G−T ) = b̂s in Ω̂s × [0, T ],

 ˆ µg

where

ˆ ûf G−1
σ̂f (ûf , p̂f ) = µf (∇ −1 −T −T ˆ T
µg Fµg + Fµg Gµg ∇ ûf ),

P̂ (dˆs , µs ) = λs trε̂s (dˆs )I + 2µs ε̂s (dˆs ), µs ∈ [µmin

s , µmax
s ],
ˆ = 1 (∇ ˆ dˆs G−1 −T ˆ T ˆ
ε̂s (ds) µg + Gµg ∇ ds ).
2
 290 Chapter 14. Reduced Order Methods for Fluid-Structure Interaction Problems

The coupling conditions (14.7) are pulled back onto the reference fluid-structure interface Γ̂F SI :
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

 dˆf = dˆs on Γ̂F SI (µg ) × [0, T ],



û = ∂t dˆs on Γ̂F SI (µg ) × [0, T ], (14.10)

 f
JK σ̂f (ûf , p̂f )F −T G−T n̂f = K P̂ (dˆs )G−T n̂s on Γ̂F SI (µg ) × [0, T ].

Here n̂f and n̂s are the normals to the fluid-structure interface in the reference configuration,
out of the fluid domain and the solid domain, respectively.
We remark that, from now on, in order to ease the notation and the exposition, we will drop
the notation with the ∧, since everything is intended to be taken in the reference configura-
tion Ω̂.

14.4.2 Offline Computational Phase

We start by discretizing the time interval [0, T ] with {0 = t0 , . . . , tNT = T }, where ti = i∆T
for i = 0, . . . , NT . We discretize the partial derivative of a function f with a first backward
i+1 i
difference: Dt f i+1 = f ∆T−f and Dtt f i+1 = Dt (Dt f i+1 ), where f i+1 = f (ti+1 ). We then
discretize the parameter space P = [µming , µmax
g ] × [µmin
s , µmax
s ] with an equispaced sampling,
M
s ]. We then define the following function
and we obtain Ptrain = [µ1g , . . . , µg g ] × [µ1s , . . . , µM s

spaces for the fluid,

V (Ωf ) := [H 1 (Ωf )]2 , E f (Ωf ) := [H 1 (Ωf )]2 , Q(Ωf ) := L2 (Ωf ), (14.11)

endowed with the H 1 -norm (V (Ωf ) and E f (Ωf )) and the L2 -norm, respectively, and the func-
tion space for the solid, E s (Ωs ) = [H 1 (Ωs )]2 , endowed with the H 1 -norm. We discretize in
space the FSI problem, using second-order Lagrange FEs for the fluid velocity, the fluid dis-
placement, and the solid displacement, resulting in the discrete spaces Vh ⊂ V , Ehf ⊂ E f , and
Ehs ⊂ E s , while the fluid pressure is discretized with first-order Lagrange FEs, resulting in the
discrete space Qh ⊂ Q.
After space discretization, the problem of interest is defined as follows: for every µ =
(µg , µs ) ∈ Ptrain , and for i = 0, . . . , NT , we have the following:

f f
• Extrapolation of the mesh displacement: Find di+1
f,h ∈ Eh such that ∀ef,h ∈ Eh ,

( R
K∇di+1 −1 −1
f,h Gµg · ∇ef,h Gµg dx = 0,
Ωf
(14.12)
di+1
f,h= dis,h on ΓF SI .

• Fluid explicit step: Find ui+1

f,h ∈ Vh such that ∀vh ∈ Vh ,

  ui+1 −ui 
· vh dx + ρf Ωf JK[∇ui+1 −1 −1 i+1
R f,h
R
f,h Gµg Fµg ]uf,h · vh dx
f,h


 ρ f Ωf
JK ∆T
i+1 i+1

 − ρf Ωf JK[∇uf,h G−1 −1
R
µg Fµg ]Dt df,h · vh dx


i+1 −T −T (14.13)
R
 + µf Ωf JKε(uf,h )Fµg Gµg : ∇vh dx
−T −T i
R R
 + Ωf JKFµg Gµg ∇pf,h · vh dx = Ωf JKbf · vh dx,




i+1 i+1
uf,h = Dt df,h on ΓF SI ,


subject to the boundary conditions ui+1

f,h = 0 on Γtop ∪ Γbottom .
 14.4. Partitioned RBM 291

• Implicit step: For any j = 0, . . . until convergence do the following:

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

1. Fluid projection substep (pressure Poisson formulation): Find pi+1,j+1

f,h ∈ Qh
such that ∀qh ∈ Qh ,

ρf
Z Z
− −1 −1 i+1
div(JKGµg Fµg uf,h )qh dx + αROB pi+1,j
f,h qh ds
∆T Ωf ΓF SI
Z Z
i+1,j i+1,j+1
− ρf (Dtt ds,h ) · JKFµ−T g
G−T
µg nf qh ds = αROB pf,h qh ds
ΓF SI ΓF SI
Z
i+1,j+1
+ JKG−T −T
µg Fµg ∇pf,h · G−T −T
µg Fµg ∇qh dx,
Ωf

subject to the boundary conditions pi+1

f,h = pin (t
i+1
) on Γin and pi+1
f,h = 0 on Γout .

2. Structure projection substep: Find di+1,j+1

s,h ∈ Ehs such that ∀es,h ∈ Ehs ,
Z Z
i+1,j+1 i+1,j+1
ρs KDtt ds,h · es,h dx + KP (ds,h , µs )G−T
µg : ∇es,h dx
Ωs Ωs
Z Z
i+1,j+1 i+1,j+1
=− JKσf (uf,h , pf,h )G−T
µg F −T
µg n f · es,h dx + Kbs · es,h dx,
ΓF SI Ωs

subject to the boundary condition ds = 0 on ΓsD .

Steps 1 and 2 are iterated until a convergence criterion is satisfied. In our case, this stopping
criterion is given by

i+1,j+1 i+1,j
||Q ||di+1,j+1 − di+1,j
( )
||pf,h − pf,h s,h s,h ||E
max i+1,j+1
; i+1,j+1
< ε, (14.14)
||pf,h ||Q ||ds,h ||E

where ε is a tolerance of our choice. To enhance the stability of the fluid-solid subproblem
(implicit step), we used a Robin coupling condition in the pressure Poisson formulation:

αROB pi+1,j+1 + Fµ−T

g
G−T
µg ∇p
i+1,j+1
· JKFµ−T
g
G−T
µg nf

= αROB pi+1,j − ρf Dtt di+1,j

s · JKFµ−T
g
G−T
µg nf .

ρf
The constant αROB is defined as αROB = zp ∆T , where zp is called the solid impedance:

zp = ρs cp ,
s
λs + 2µs
cp = .
ρs

Introducing the Robin coupling between the fluid and the solid problem improves the stability
properties of the implicit step; computing the Robin coupling can be straightforward, e.g., in the
coupling of a fluid with a thin, one-dimensional structure (see, e.g., [26, 22]), or it may be more
complicated, such as in our test case, where the structure is thick and two-dimensional and the
 292 Chapter 14. Reduced Order Methods for Fluid-Structure Interaction Problems

influence of the fluid on the solid does not appear immediately from the beginning in the solid
problem formulation. For this second situation, we refer the reader to [39].
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Change of Variable for the Fluid Velocity

The main idea here is to introduce a change of variable in the fluid problem, in order to transform
condition (14.132 ) into a homogeneous boundary condition. The motivation for this choice is
that, to impose condition (14.132 ), we could use a Lagrange multiplier λ, but unfortunately
introducing a new variable leads to an increased dimension of the system to be solved in the
online phase. Therefore, in order to avoid this and design a more efficient reduced method, we
choose to transform the nonhomogeneous coupling condition into a homogeneous one. In order
i+1
to do this, we define a new variable zf,h :

i+1
zf,h := ui+1 i+1
f,h − Dt df,h . (14.15)

With this change of variable, (14.132 ) is equivalent to the homogeneous boundary condition for
the new variable:
i+1
zf,h = 0 on ΓF SI , (14.16)

for which no imposition of the Lagrange multiplier is needed. Therefore, during the offline phase
of the scheme, at every iteration i + 1, after we have computed the velocity ui+1
f,h , we compute
i+1
the change of variable zf,h . We then consider the snapshot matrix

h
1 NT
S z = [zf,h , . . . , zf,h ] ∈ RNu ×NT , (14.17)

where Nuh = dimVh .

POD and Reduced Basis Generation

For this test case, the idea is to first perform a standard POD on the set of snapshots computed
for every value of the parameter µ in the training set Ptrain ; subsequently, we take the modes
computed with the standard POD, weighted with the corresponding eigenvalue, and we perform
a final outer run of POD; for a more detailed discussion we refer the reader to [415].
We start by constructing, for each parameter µi,j = (µig , µjs ) ∈ Ptrain , the snapshot matrices
S z (µi,j ) for the fluid change of variable zf , S p (µi,j ) for the fluid pressure pf , and S ds (µi,j )
for the solid displacement ds :
h
S z (µi,j ) = {zf,h (t0 , µi,j ), . . . , zf,h (tNT , µi,j )} ∈ RNu ×NT ,
h
S p (µi,j ) = {pf,h (t0 , µi,j ), . . . , pf,h (tNT , µi,j )} ∈ RNp ×NT ,
h
S ds (µi,j ) = {ds,h (t0 , µi,j ), . . . , ds,h (tNT , µi,j )} ∈ RNds ×NT ,

where Nph = dimQh and Ndhs = dimEhs . We then perform a standard POD on each snapshot ma-
Nz Np Nd
trix and we extract the basis functions {Φkzf (µi,j )}k=1
i,j
, {Φkp (µi,j )}k=1
i,j
, and {Φkds (µi,j )}k=1
i,j
.
Nz Np Nd
Let us also call {λ(µi,j )zk }k=1
i,j
, {λ(µi,j )pk }k=1
i,j
, and {λ(µi,j )dk }k=1
i,j
the eigenvalues, in decreas-
ing order of magnitude, returned by the POD on each snapshot matrix S z (µi,j ), S p (µi,j ), and
S ds (µi,j ). Afterwards, we perform a second run of POD in the following way: we start by
 14.4. Partitioned RBM 293

building the snapshot matrices, normalizing each snapshot with the corresponding eigenvalue
given by the standard POD:
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

q z

z
q NM
λ(µ1,1 )1 Φ1zf (µ1,1 ), . . . , λ(µMg ,Ms )zN z
,M
Sz = Φzf g s (µMg ,Ms ) ,
Mg ,Ms

p
q 
p NM
λ(µ1,1 )1 Φ1p (µ1,1 ), . . . , λ(µMg ,Ms )pN p
q ,M
Sp = Φp g s (µMg ,Ms ) ,
Mg ,Ms
(q )
r d
d NM
S ds = λ(µ1,1 )1 Φ1ds (µ1,1 ), . . . , λ(µMg ,Ms )dN d Φds g ,Ms
(µMg ,Ms ) .
Mg ,Ms

We then perform a second POD on the previous snapshot matrices, and finally we obtain a set of
z
k Np k Nd
basis functions {Φkzf }N
k=1 , {Φp }k=1 , and {Φds }k=1 .

Harmonic Extension of the Fluid Displacement

In order to generate the reduced basis for the fluid displacement df (µ), we pursue the idea
d
presented in [37]. Once we have obtained the reduced basis {Φkds }Nk=1 for the solid displace-
ment, we employ a harmonic extension of each one of the reduced bases Φkds on the whole fluid
domain:
(
−∆Φkdf = 0 in Ωf ,
(14.18)
Φkdf = Φkds on ΓF SI .

We can then define the reduced space for the fluid displacement as
f
EN := span{Φkdf }N
k=1 .
d
(14.19)

This choice is extremely efficient: the first advantage is that we can avoid introducing another
Lagrange multiplier to impose the nonhomogeneous boundary condition (14.122 ) because it is
automatically satisfied by every Φkdf and hence by every linear combination of these basis func-
tions. In addition, with our method, we avoid solving the reduced system related to (14.12):
instead of solving a harmonic extension problem at every time step in the online phase, we solve
all Nd harmonic extension problems at once in the expensive offline phase.

14.4.3 Online Phase

i+1
We start by introducing the online components zf,N , pi+1 i+1
f,N , and ds,N of the solution at time step
ti+1 :
Nz
X
i+1
zf,N (µ) = z i+1 k
k (µ)Φzf , (14.20)
k=1
Np
X
pi+1
f,N (µ) = pi+1
k
(µ)Φkp , (14.21)
k=1
Nd
X
di+1
s,N (µ) = di+1 k
k (µ)Φds . (14.22)
k=1
 294 Chapter 14. Reduced Order Methods for Fluid-Structure Interaction Problems

The reduced problem then reads as follows: for every i = 0, . . . , NT and for µ = (µg , µs ) ∈ P,
do the following:
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

• Mesh displacement: Let di+1

f,N (µ) be defined by the reduced solid displacement at the
previous time step:
Nd
X
di+1
f,N (µ) = dik (µ)Φkdf . (14.23)
k=1

i+1
• Fluid explicit step (with change of variable): Find zf,N (µ) ∈ VN such that ∀vN ∈ VN ,

 z i+1 (µ) − ui
f,N (µ)
Z  Z
f,N i+1
ρf JK · vN dx + µf JKε(zf,N (µ))Fµ−T G−T
µg : ∇vN dx
Ωf ∆T Ωf
g

Z
i+1 −1 i+1
+ ρf JK∇zf,N (µ)G−1
µg Fµg zf,N (µg ) · vN dx
Ωf
Z
+ ρf JK∇Dt di+1 −1 −1 i+1
f,N (µ)Gµg Fµg zf,N (µ) · vN dx
Ωf
Z Z  Dt di+1 (µ) 
f,N
+ JKFµ−T G−T i
µg ∇pf,N (µ) · vh dx = −ρf JK · vN dx
Ωf
g
Ωf ∆T
Z Z
− µf JKε(Dt di+1 −T −T
f,N (µ))Fµg Gµg : ∇vN dx + JKbf,N · vN ,
Ωf Ωf

where bf,N is the projection of the fluid external force bf onto the reduced space VN . We
then restore the reduced fluid velocity: ui+1 i+1 i+1
f,N (µ) = zf,N (µ) + Dt df,N (µ).
• Implicit step: For any j = 0, . . . until convergence do the following:

1. Fluid projection substep: Find pi+1,j+1

f,N (µ) ∈ QN such that ∀qN ∈ QN ,

ρf
Z Z
− div(JKG−1 −1 i+1
µg Fµg uf,N (µ))qN dx + αROB pi+1,j
f,N (µ)qN ds
∆T Ωf ΓF SI
Z Z
i+1,j
− ρf Dtt ds,N (µ) · JKFµ−T
g
G−T
µg nf qN ds = αROB pi+1,j+1
f,N (µ)qN ds
ΓF SI ΓF SI
Z
i+1,j+1
+ JKFµ−T
g
G−T
µg ∇pf,N (µ) · Fµ−T
g
G−T
µg ∇qN dx,
Ωf

subject to some suitable boundary conditions for the pressure.

2. Structure projection substep: Find di+1,j+1
s,N
s
(µ) ∈ EN such that ∀es ∈ EN
s
,
Z Z
i+1,j+1
ρs KDtt ds,N (µ) · eN dx + KP (di+1,j+1
s,N (µ), µs )G−T
µg : ∇eN dx
Ωs Ωs
Z Z
i+1,j+1
=− JKσf (ui+1
f,N (µ), pf,N (µ))Fµ−T
g
G−T
µg nf · eN dx + Kbs,N · eN ,
Ωs Ωs

where bs,N is the projection of the solid external force bs onto the reduced space EN
s
.
 14.4. Partitioned RBM 295

Tµg
x
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Ω̂ Ω̃(µg )

↓ Af (t; µg )

Ω(t; µg )

Figure 14.3. Domains: reference configuration Ω̂ (top left), parametrized reference configuration
Ω̃(µg ) (top right), and original configuration Ω(t; µg ) (bottom).

14.4.4 Numerical Results

We now present some numerical results for the test case with a geometrical and physical parame-
trization. The original configuration, the intermediate configuration, and the reference configu-
ration are depicted in Figure 14.3. The channel is 2.5 cm high, and it is 10 cm long; the leaflets’
length is represented by the geometrical parameter µg ∈ [0.8, 1.0], and the leaflets are situated
1 cm downstream from the inlet boundary (left vertical boundary of the channel), and they are
0.2 cm thick. An affine mapping T is chosen in order to deform the reference configuration Ω̂,
obtained for µg = 1 cm, into the intermediate configuration Ω̃(µ). The top and bottom walls of
the blue fluid domain are assumed to be rigid; thus both the displacement df (µ) and the fluid
velocity uf (µ) are set to zero. A homogeneous Neumann condition is imposed on uf (µ) on the
outlet; a pressure profile pin (t) is described at the inlet, where
(  
5 − 5cos T2πt for t ≤ 0.1 s,
pin (t) = in (14.24)
5 for t > 0.1 s,

and Tin = 0.4 s. The tolerance chosen for the implicit step of the whole algorithm is ε = 10−6 ,
and the time step used in the problem time discretization is ∆T = 10−4 .
For the simulation, we use a time step of ∆t = 10−4 for a maximum number of time steps
NT = 800; thus T = 0.08 s. Table 14.1 summarizes the details of the offline stage and of the
finite element (FE) discretization.
Figure 14.4 shows three different examples of the behavior of the leaflets, according to the
change of the physical and/or of the geometrical parameters: all the pictures represent the online
displacement of the solid at the final time step of the simulation, namely for t = 0.08 s. The
results were obtained using Nd = 100 reduced bases for the displacement; as we can see, in-
dependent of the length of the leaflets, an increase in the shear modulus leads to a material that
is much harder to deform. On the contrary, for a fixed value of the properties of the material
under consideration, an increase in the length of the leaflets leads to an increase in the displace-
ment. As the results show, our ROM is able to capture small deformations of the solid, and it is
also able to reproduce variations and changes in the behavior of the leaflets, also at very small
magnitudes.
 296 Chapter 14. Reduced Order Methods for Fluid-Structure Interaction Problems

Table 14.1. Physical and geometrical constants and parameters for the geometrically and physi-
cally parametrized leaflets test case.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Physical constant Value

ρs 1.1 g/cm3
λs 800000
ρf 1 g/cm3
µf 0.035 Poise
Geometrical parameter Value
µg [0.5, 1.0]
FE displacement order 2
FE velocity order 2
FE pressure order 1
Ntrain 10
N 100
Physical parameter Value
µs [10000, 80000]

Same leaflet length Same leaflet length Increased leaflet length

µg = 0.8cm, but increased µg = 1cm, but increased µg = 0.8cm (left) and
shear modulus µs = 100000 shear modulus µs = 100000 µg = 1cm (right) but same
(left) and µs = 800000 (left) and µs = 800000 shear modulus
(right) (right) µs = 100000.

Figure 14.4. Reduced order solid displacement dN s (µ). Comparison of different behaviors of the
material, for different values of the geometrical and physical parameters. From left to right: same leaflet
length µg = 0.8 cm and increased shear modulus (µs = 100000, 800000); same leaflet length µg = 1 cm
and increased shear modulus (µs = 100000, 800000); increased leaflet length (µg = 0.8, 1.0 cm), and
same shear modulus µs = 100000.

14.5 A Monolithic RBM for a Transport-Dominated FSI

Problem
This section is devoted to formulating a model order reduction procedure for an FSI problem
whose solution manifold exhibits a slow decay of the Kolmogorov n-width. The introduction,
during the offline phase, of a preprocessing procedure on the set of snapshots brings significant
improvements in terms of the rate of decay of the eigenvalues returned by a POD on the solution
manifold. The advantage of adopting a reduced order method that is enriched with a preprocess-
ing procedure is that we are able to construct a reduced solution using a significantly smaller
number of basis functions, with a consequent reduction of the dimension of the system to be
solved in the online phase. Even though the capability of the RBM has been acclaimed in a large
 14.5. A Monolithic RBM for a Transport-Dominated FSI Problem 297

variety of situations, there is still a challenging task for these procedures: advection-dominated
problems; see, e.g., [428, 244, 181]. In the last years it has therefore become clear that a modifi-
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

cation of the standard RBM is necessary, especially in order to obtain a small basis set in these
more challenging situations.
Before going any further, we will try to explain the difficulties that arise when we are in the
presence of transport-dominated problems.

14.5.1 The Kolmogorov n-Width

Let Ω be the physical domain of the problem of interest, and let P ⊂ Rp , p ∈ N, be a compact
set. Denote by µ ∈ P the parameter, and let t ∈ [0, T ] be the time for some T > 0. In
the following exposition, the symbol η will stand for µ (in the case of parametrized stationary
problems), for t (in the case of nonparametric unsteady problems), or for the pair (t, µ) (in
the case of parametrized unsteady problems). We assume that the problem of interest has the
following form: for any value of η ∈ E, we seek z(η): Ω → RZ , Z ∈ N, such that

N (z(η); η) = 0, (14.25)

where N is a nonlinear operator working on functions defined over Ω and E := P, E := [0, T ],

or E := [0, T ] × P, respectively, in the three aforementioned cases. For our problem of interest
we can define the associated solution manifold, whose Kolmogorov dimension (or Kolmogorov
n-width) will become very important shortly.

Definition 14.2. Let z(η) be a component of z(η), and let Mz be the corresponding solution
manifold, embedded in some normed linear space (Xz , || · ||Xz ), defined as

Mz = {z(η), η ∈ E}. (14.26)

One fundamental assumption of the RBM is that Mz can be approximated accurately by

a sequence of finite-dimensional spaces: any element of Mz can be recovered using a linear
combination of solutions of (14.25), which are computed only once. The mathematical entity
that incorporates this concept is the Kolmogorov n-width of Mz .

Definition 14.3. The Kolmogorov n-width of the solution manifold Mz is

Dn (Mz , Xz ) = inf sup inf ||f − g||Xz , (14.27)

En ⊂Xz f ∈Mz g∈En

where En is any linear subspace of Xz of dimension n.

From this definition we understand that, given a problem of interest, the Kolmogorov n-width
Dn represents the entity of the error that we commit by approximating any element f of Mz with
an element g of a linear space En of dimension n. Of course this means that the faster Dn decays
as we let n grow, the better chance we have to find a good linear approximation space for Mz
of small dimension. In the majority of problems there is no explicit analytic formula for Dn , yet
there are some situations where we can compute good bounds on it; see, e.g., [393, 133]. Since
it is very difficult to provide bounds on Dn , one can only hope that the n-width of the solution
manifold of the problem of interest is small. A heuristic way to check that this hypothesis is
satisfied is to run a POD on a set of snapshots and check the rate of decay of the eigenvalues
{λi }i returned by the POD: if the {λi }i decay fast, then we can expect to be able to build a good
low-dimensional linear approximation space for Mz . Unfortunately, the problem is that this
assumption often fails in transport-dominated problems, which usually show a very slow decay
 298 Chapter 14. Reduced Order Methods for Fluid-Structure Interaction Problems

of the eigenvalues: it therefore becomes clear that the standard RBM is unable to reconstruct any
element of Mz by using a small number of basis functions.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

In recent years a growing number of works have appeared that focus on constructing alter-
native (nonlinear) model order reduction techniques, that can be effectively applied to transport-
dominated problems: we refer the interested reader to some examples of these works, such as
[9, 221, 222, 181, 99, 18, 442, 303, 57, 88, 87, 613, 614, 411, 567, 322, 427, 500, 59]. In
[367], the authors attempt to design a reduction procedure that is not strictly related to the phys-
ical phenomenon under investigation, e.g., transportation of a quantity or translation of a wave.
The authors propose a methodology which focuses on optimal projection of general dynamical
systems onto arbitrary nonlinear trial manifolds in the online stage, where this nonlinear trial
manifold is computed from snapshot data alone in the offline stage, thanks to the employment of
convolutional autoencoders. The reader interested in the use of machine learning techniques in
the model order reduction framework is referred to in [206, 264]. A methodology that is light,
simple, more natural, based on the definition of some transport maps is introduced and applied to
some toy problems by Cagniart [86] and Cagniart, Maday, and Stamm [88]. The goal of this sec-
tion is to present an application of model reduction based on transported snapshots for problems
in fluid dynamics and FSI, focusing in particular on comparing results of the standard offline
phase and the novel one, which relies on transport.

14.5.2 Nonlinear Model Reduction by Transport Maps

In this subsection we summarize the nonlinear approach that we will apply in the forthcoming
sections to the FSI problem of interest. We follow the notation introduced by Cagniart et al.
[88] for a parametrized viscous Burgers equation. The idea proposed therein is to “preprocess”
the solution manifold Mz by a composition with a map belonging to a family of smooth and
invertible mappings:

Fz = {Fη : Ω → Ω, Fη is smooth and invertible, η ∈ E}. (14.28)

Maps in the family Fz depend on the same η appearing in (14.25) and are essentially problem
specific in the sense that different maps (translation, dilatation) should be employed in different
settings: translation maps will be enough in some simple periodic settings, while dilatation maps
are more suited in the case of nonperiodic boundary conditions. In any case, the main goal of
these maps is to align some feature of the solution of the problem under consideration (e.g., a
shock or a peak) with a fixed point. In [88], e.g., a family of translations Fη (x) = x − γ(η),
x ∈ Ω ⊂ R, γ(η) ∈ Ω, is employed for the viscous Burgers equation, parametrized by the
parameter η: there, γ(η) is chosen so that the steepest points of the solutions coincide.
Before going any further, let us make the following assumption: the problem of interest is time
dependent, and no other parameter appears in the formulation; thus η = t and E = [0, T ]. Given
the family Fz of smooth and invertible mappings, we can introduce the preprocessed solution
manifold.

Definition 14.4. The preprocessed solution manifold MFz is

MFz = {z(Ft−1 (·); t), t ∈ [0, T ]}. (14.29)

Assuming that Fz is carefully chosen, MFz has a smaller Kolmogorov n-width than that of
Mz . The practical realization of this preprocessing procedure is incorporated in the offline phase.
Given a time discretization of the original problem [0 = t0 , . . . , tNT = T ], we compute each
solution component z(ti ) with i = 0, . . . , NT . The discrete approximations of the corresponding
 14.5. A Monolithic RBM for a Transport-Dominated FSI Problem 299

standard and preprocessed solution manifolds,

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Mtr i
z = {z(t ), i = 0, . . . , NT },
Mtr
Fz = {z(t ) ◦ Ft−1
i
i , i = 0, . . . , NT },

provide snapshots for compression by a POD. The compression is applied here to both Mtr z
and Mtr Fz to compare the standard offline phase and one with preprocessing, but in practical
computations one would neglect the compression of Mtr z because it is understood that MFz
tr

would result in a POD basis set {Φk }N k=1 of lower dimension. Assume now that an accurate
approximation zN (ti ) of the solution component z(ti ) at time ti is known as a linear combination
of a pullback of the basis functions Φk :
N
X
zN (ti ) = αki Φk ◦ F̃ti , (14.30)
k=1

where F̃ti is a map in Fz . In order to recover zN (ti+1 ) as an expansion,

N
X
zN (ti+1 ) = αki+1 Φk ◦ F̃ti+1 , (14.31)
k=1

the idea proposed in [88] is to iterate between the search for the reduced coordinates αki+1 and the
search for a suitable pullback map F̃ti+1 . This procedure is carried out by means of a minimiza-
tion problem of the L2 -norm of some residual. In this book, we do not follow a minimization
approach: we employ instead a polynomial interpolation method to learn the best-suited pullback
map F̃ti+1 to be used in the online phase.
To summarize, here is an outline of the general RBM with preprocessing of the snapshots:

• compute the snapshots z(ti ) for i = 0, . . . , NT ;

• build the family of preprocessing maps Fz = {Ft , t ∈ [0, T ]};
• preprocess the snapshots and obtain the preprocessed solution manifold Mtr
Fz ;

• run a POD on Mtr Fz and obtain the basis functions {Φ }k=1 ;

k N

PN
• during the online phase, at time step ti+1 , find zN (ti+1 ) = k=1 αki+1 Φk ◦ F̃ti+1 , where
F̃ti+1 is a suitable pullback map.

14.5.3 A Multiphysics Problem

Now that we have presented the mathematical framework and we have set the notation, we can
present the problem of interest: a multiphysics, advection-dominated problem, where the solu-
tion behaves like a traveling wave. The problem presents an ALE formulation, which was in-
troduced in Subsection 14.2.1. In this case we address the problem with a monolithic approach;
see, e.g., [36].

Problem Formulation
A two-dimensional rectangle of height hf and length L is filled with a Newtonian fluid. The
top and the bottom walls of the rectangle represent the structure, which is considered to be
deformable, and its thickness is negligible with respect to the height of the rectangle. Since the
 300 Chapter 14. Reduced Order Methods for Fluid-Structure Interaction Problems

%
Γin Γout
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Ω̂ Σ̂ &

Figure 14.5. Physical domain (reference configuration): fluid subdomain (blue) and structure
subdomain (orange). The fluid-structure interface coincides with the structure in our case; the structure
has been magnified for visualization purposes.

structure is thin, it is described by a one-dimensional model. We further assume that the displace-
ment of the walls in the horizontal direction is negligible, and hence the structure undergoes only
vertical motion: the behavior of the compliant walls is therefore described by the generalized
string model [469, 476]. We want to study the behavior of the solution in the time interval [0, T ].
Let Σ(t) and Ωf (t) be, respectively, the structure and the fluid domain at time t. Let Σ̂ be the
solid reference configuration (undeformed walls), and let Ω̂f be an arbitrary fluid reference do-
main: for convenience we take Ω̂f = Ωf (t = 0) (the blue fluid domain in Figure 14.5). Thanks
to the introduction of the ALE map Af (t) defined in Subsection 14.2.1, we can map the fluid
equations back to the reference domain Ω̂f . Let F be the gradient of Af (t), and let J be the
Jacobian. The FSI problem, formulated on the reference configuration, reads as follows: find the
fluid velocity ûf (t): Ω̂f → R2 , the fluid pressure p̂f (t): Ω̂f → R, and the structure displacement
dˆs (t): Σ̂ → R such that

 Jρf (∂t ûf + F −1 ((ûf − ∂t dˆf ) · ∇) ˆ ûf ) − div(J σ̂ f F −T ) = bf in Ω̂f × [0, T ],



div(JF ûf ) = 0 in Ω̂f × [0, T ],

−1

ρs hs ∂tt dˆs − c0 ∂xx dˆs + c1 dˆs = −σ̂ f (ûf , p̂f )n̂ · n̂ in Σ̂ × [0, T ].

(14.32)
Here ρf is the fluid density, ρs is the structure density, hs is the structure height (or thickness),
c0 and c1 are the structure constitutive parameters, and n̂ is the outward normal to Σ̂. σ̂ f is
the Cauchy stress tensor for the fluid, in the reference configuration, which was introduced in
Subsection 14.4.1.
In order to ease the notation, we drop the ∧, since everything will be formulated in the
reference configuration. The coupling conditions are

df = ds n in Σ × [0, T ],

(14.33)
uf = ∂t ds n in Σ × [0, T ].

We further assume that at time t = 0 the system is at rest; finally, the boundary conditions can
be summarized in the following system:

 σf (uf , pf )n = −pin (t)n in Γin × (0, T ],



σf (uf , pf )n = −pout (t)n in Γout × (0, T ], (14.34)

ds = 0 in ∂Σ × (0, T ],


where the third condition says that the structure is fixed at its extremities.

Weak Formulation with a Monolithic Approach

We now want to present the algebraic system related to our FSI problem, which arises from the
adoption of a monolithic approach, where, we recall, the fluid and the solid problem are solved
all together. For this reason, we introduce the function spaces for the fluid problem,

V := [H 1 (Ωf )]2 , Q := L2 (Ωf ), E f := [H01 (Ωf )]2 ; (14.35)

 14.5. A Monolithic RBM for a Transport-Dominated FSI Problem 301

for the structure problem,

E s := H01 (Σ); (14.36)
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

and for the two Lagrange multipliers,

Lu = Ld := [H01 (Σ)]2 , (14.37)

where
H01 (Ωf ) := {ef ∈ [H 1 (Ωf )]2 such that ef = 0 on Γin ∪ Γout } (14.38)
and
H01 (Σ) := {es ∈ H 1 (Σ) such that es = 0 on ∂Σ}, (14.39)
and where V is endowed with the H 1 -norm; Q with the L2 -norm; and E f , E s , Lu , and Ld with
the H1 seminorm. Let us now consider some finite-dimensional subspaces Vh ⊂ V , Qh ⊂ Q,
Ehf ⊂ E f , Ehs ⊂ E s , Luh ⊂ Lu , and Ldh ⊂ Ld of dimension N u , N p , Nfd , Nsd , Nuλ , and Ndλ ,
respectively. We discretize the time interval [0, T ] with the following partition:

0 = t0 < · · · < tNT = T , (14.40)

where every interval (ti , ti+1 ] has measure ti+1 − ti = ∆T , i = 0, . . . , NT − 1. We de-

note by zhi = zh (ti ) any of six components of the solution of our FSI problem, evaluated at
time ti ; as done in Subsection 14.4.2, we discretize ∂t zh (t) with a first backward difference
z i+1 −z i
Dt zhi+1 = h ∆T h , and we discretize ∂tt zh (t) with Dtt zhi+1 = Dt (Dt zhi+1 ). The discretized
weak formulation of the original FSI problem reads as follows: for any i = 0, . . . , NT , find
uif,h ∈ Vh , pif,h ∈ Qh , dif,h ∈ Ehf , dis,h ∈ Ehs , λiu,h ∈ Luh , and λid,h ∈ Ldh such that

JDt uif,h · vh dx + Ωf J∇uif,h FR−1 uif,h · vh dx
R R

 ΩfR
− RΩf J∇uif,h F −1 Dt dRif,h · vh dx + Ωf RJσ f F −T : ∇vh dx




+ RΓin piin vh · n dx − Σ λiu,h · vh dx = Ωf bif · vh dx




− Ωf qh divuif,h dx = 0,








∇dif,h : ∇ef,h dx − λid,h · ef,h dx = 0,

 R R

 Ωf Σ
(14.41)
ρs hs RΣ Dtt dis,h es,h dxR+ c1 Σ ∂x dis,h ∂x es,h dx
 R R



+ c0 Σ dis,h es,h dx + Σ es,h n · λiu,h dx = 0,








uif,h · mu,h dx − Dt dis,h n · mu,h dx = 0,

 R R
Σ Σ






dif,h · md,h dx − dis,h n · md,h dx = 0

 R R
Σ Σ

for all vh ∈ VH , qh ∈ Qh , ef,h ∈ Ehf , es,h ∈ Ehs , muh ∈ Luh , and mdh ∈ Ldh . To obtain a stable
approximation of the fluid pressure during the online phase of the method, we adopt a supremizer
enrichment technique. In this case, the supremizer sih ∈ Vh is defined by the following problem:
for any i = 0, . . . , Nt , find sih ∈ Vh such that
Z Z Z
∇sih : ∇vh dx + sih · vh dx = − pif,h divvh dx (14.42)
Ωf Ωf Ωf
u p N p
for all vh ∈ Vh . Let {ϕuj }N
j=1 be a set of basis functions for Vh , {ϕj }j=1 a set of basis functions
d Nfd Nd
for Qh , {ϕj f }j=1 a set of basis functions for Ehf , {ϕdj s }j=1
s
a set of basis functions for Ehs ,
 302 Chapter 14. Reduced Order Methods for Fluid-Structure Interaction Problems

λu λd N d λ
{ϕλj u }N
j=1 a set of basis functions for Lh , and {ϕj }j=1 a set of basis functions for Lh , where
u d
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

f
N u = dimVh , N p = dimQh , Nfd = dimEh , Nsd = dimEhs , N λu = dimLuh , and N λd =
dimEhd . We introduce the vector of degrees of freedom associated with the velocity components
of the solution:
u
N
(j) u u X (j)
uf := (uf,h )N
j=1 ∈ R
N
⇐⇒ uf,h = (uf,h )ϕuj ∈ Vh , (14.43)
j=1

and similarly for the other components of the solution. Then the system (14.41) is equivalent to
the following algebraic system:

Mu + A + N (uif ) BT C(uif ) −Lu1,T uif

    
0 0 Fu
 B 0 0 0 0 0  pif   0 
−Ld1,T
    

 0 0 D 0 0 
 dif  
= 0 


 0 0 0 Q + Ms1 2,T
−Lu 0 
 dis  
  0 

 L1u 0 0 −Ms2 0 0  λiu   0 
0 0 L1d −L2d 0 0 λid 0
  i−1  
−C(uif ) uf ui−2
  
Mu 0 0 0 0 0 0 0 0 0 0 f
 0 0 0 0 0 0  pi−1   0 0 0
 
0 0 0  pi−2 
  f
  f 
 0 0 0 0 0 0   di−1   0 0 0
 
0 0 0  dfi−2 
+  f +  .

 0 0 0 −Ms1 0 0   di−1
 s   0 0 0 2Ms1 0 0 
 dsi−2 

0 0 0 −Ms2 0 0   λi−1 0 0 0 0 0 0 λu i−2
    
u

0 0 0 0 0 0 λi−1 0 0 0 0 0 0 λd i−2
d

In the previous system, we have the following terms:

J u
Z
(Mu )j,k = ϕj · ϕuk dx,
Ω f ∆T
ϕdj s · ϕdks
Z
(Ms1 )j,k = ρs hs dx,
Σ ∆T 2
ϕλj u · nϕdks
Z
(Ms2 )j,k = dx,
Σ ∆T
Z
(A)j,k = J∇ϕuj : ∇ϕuk dx,
Ωf
Nu Z
(r)
X
(N (uif ))j,k = uf,h J∇ϕuk F −1 ϕur · ϕuj dx,
r=1 Ωf
Z
(B)j,k = ϕpj divϕuk dx,
Ωf
u
N
J
Z
(r) d
X
(C(uif ))j,k = uf,h ∇ϕur F −1 ϕkf · ϕuj ,
r=1 Ωf ∆T
Z
d d
(D)j,k = ∇ϕj f : ∇ϕkf dx,
Ωf
Z Z
(Q)j,k = c0 · ϕdj s
dx + c1 ϕdks ∂x ϕdj s · ∂x ϕdks dx,
Σ Σ
Z
(L1u )j,k = ϕλj u · ϕuk dx,
Σ
 14.5. A Monolithic RBM for a Transport-Dominated FSI Problem 303

ϕdj s n · ϕλk u
Z
(L2u )j,k = dx,
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

∆T
ZΣ
d
(L1d )j,k = ϕλj d · ϕkf dx,
ZΣ
(L2d )j,k =ϕλj d · nϕdks dx,
Σ
Z Z
(Fu )j = − pin (ti )n · ϕuj dx + bi · ϕuj dx. (14.44)
Ωf Ωf

14.5.4 Offline Phase

Problem data used for the simulation of our test case can be found in Table 14.2: corresponding
values are taken from the numerical results presented in [566, 410].
The behavior of the fluid pressure pf and of the extended displacement df is shown in Figure
14.6 and Figure 14.7, respectively. Our problem is transport dominated: the change in time of the
position of the peak of the pressure wave, e.g., is a difficult feature to capture at the reduced order
level with just a few modes. This expectation is confirmed at the numerical level by running a
POD on a few of the components of the solution of the FSI problem, such as pf , df , and uf ; as
we can see in Figure 14.8, the eigenvalues decay relatively slowly. For this reason, we rely on a
transformation on the set of solutions in order to compensate for this transport phenomenon.

Table 14.2. Problem data for the test case of a flow in a channel with deformable walls.

Physical constants Value

ρf 1 g/cm3
Es 0.75 × 106 dyn/cm2
µf 0.035 Poise
bf 0
pin (t) 103 × [1 − cos( T2πt
in
)]χ[0,0.0025]
pout (t) 0
νs 0.5
ρs 1.1 g/cm3
hs Es
c0 2(1+νs )
hs Es
c1 h2f (1−νs2 )
hs 0.1 cm
Time constants Value
Tin 2.5 × 10−3 s
∆T 10−4 s
NT 150 s

Figure 14.6. Fluid pressure behavior: the solution is pictured here at times t = 0.001, t = 0.005,
and t = 0.015. The peak of the wave is propagating into the domain, creating a transport phenomenon.
 304 Chapter 14. Reduced Order Methods for Fluid-Structure Interaction Problems
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 14.7. Vertical component of fluid displacement behavior: again the solution is pictured at
times t = 0.001, t = 0.005, and t = 0.015. The peak of the wave is still very small at the beginning, it
grows for some time, and then it starts to propagate.

Figure 14.8. Decay of the eigenvalues for the POD on the fluid pressure (blue line), fluid dis-
placement (green line), and fluid velocity (magenta line).

In order to make the following exposition clearer and easier to read, from now on we focus
only on a particular component of the solution of our problem, namely the fluid pressure pf . In
any case, it is important to keep in mind that, based on our simulation, all the components of the
solution to the FSI problem are subject to a transport phenomenon, and hence every consideration
that we make on pf can be easily applied to any of the other components of the solution.

Preprocessing Step
We apply the preprocessing procedure to the fluid pressure solution manifold Mpf . Figure 14.6
shows how the peak of the pressure is transported in the domain. The idea is to move, at every
time step, the peak of the wave to a fixed position of our choice; in this way, we obtain a set of
snapshots where the pressure wave is not moving at all.
Starting from this observation, we build a one-parameter family of mappings

Fpf = {Ft : Ωf → Ωf , t ∈ [0, T ]} (14.45)

such that for every t in [0, T ], the peak of pf (Ft−1 (·), t) is located not at its original position
but is instead moved to the middle of the domain. A suitable map to be used in this problem
(nonperiodic setting) is a dilatation map Ft−1 :
3xγ
Ft−1 (x) = , (14.46)
x(γ − 3) + 3(6 − γ)
where γ = γ(t) is the abscissa of the position of the peak of the wave at time t. We assume that
the abscissa of the points on the inlet boundary Γin is x = 0 and the length of the domain Ωf is
L = 6; in addition, the position to which we are moving the peak of the wave at every time t is
exactly in the center of the domain. We remark that the map Ft−1 is just a stretching in the hori-
zontal direction: this is because the solution of the original problem does not exhibit any transport
 14.5. A Monolithic RBM for a Transport-Dominated FSI Problem 305

phenomena in the vertical direction, and hence there is no need for a transformation in the y-axis.
So, with an abuse of notation, we can think of Ft−1 (x, y) as Ft−1 (x, y) = (Ft−1 (x), y).
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Our family of mappings Fpf is a one-parameter family; therefore to identify the stretching
map Ft−1 we have to identify the parameter γ(t) at time t. One simple way of doing this is to use
polynomial interpolation: we approximate the physical law governing the position of the peak of
the wave γ(t) with a polynomial qk (t) of degree k. Of course, when the physical phenomenon
under investigation becomes particularly complex, or in the presence of multiple parameters,
other techniques may become more useful and better performing, such as polynomial regression
or neural networks, presented in [579]. For this toy problem, however, a polynomial of degree
k = 1 is sufficient to obtain very good results in the online phase.
First of all we compute all the snapshots p1 , . . . , pNT , where NT = ∆t
T
, pi = pf (ti ), and
t = i∆t. Once we have the snapshots, we compute the peak γTin and γNT . The reason for
i

choosing γTin instead of a more straightforward γ0 is that usually in computational fluid dy-
namics (CFD) applications, complex phenomena such as vortex formation and propagation of a
quantity require some time to develop, unlike the wave propagation problem presented in [88] or
[579], e.g. For our problem, Tin is a reasonable starting time for the preprocessing procedure:
recall the expression of pin from Table 14.2; as we can see, pin represents a given pulse at the
inlet boundary, which is nonzero up to time t = Tin = 0.0025. So, up to time t = 0.0025
the pulse makes the wave grow; when the pulse is null, the wave starts to propagate into the
domain. At the reduced level, this translates to the fact that up to t = Tin we adopt a standard
reduction technique, using a standard reduced basis obtained with a standard POD on the first
Tin snapshots. From the snapshot Tin to the last one, on the other hand, we first perform the
preprocessing step and then perform a POD.
In Figure 14.9 (top) we can compare the decay of the eigenvalues for the POD on the pressure
with and without this preprocessing procedure. As we can see, with the preprocessing technique
we do actually get an improvement in the decay of the eigenvalues, and in fact with less than 15
modes we reach a level of 10−3 , which is one order of magnitude less than the one we get with
less than 15 modes in the standard case. We get the same results for the extended displacement
df ; see Figure 14.10 (top). As we can see from Figure 14.9 (bottom) and 14.10 (bottom), with
the preprocessing procedure, the first modes are able to capture almost 70% more energy than
the modes without preprocessing. Figure 14.11 represents some snapshots for the fluid pressure
before and after the preprocessing, in order to help the reader to better visualize the effect that
this procedure has on the transport phenomenon.

14.5.5 Online Phase

In the previous paragraph we briefly explained the preprocessing procedure adopted for the fluid
pressure, which is nevertheless just one of the many components of the solution to our coupled
problem. Therefore, we stress the fact that interpolating the position of the peak γ(t) of the
pressure wave p(t) with the polynomial q1 (t) is enough for the whole preprocessing procedure
to be applied to our multiphysics problem. Indeed, since the problem is coupled, the behavior of
each component of the solution is influenced and influences the other components. Thus, a peak
in the pressure translates to a peak in the displacement of the solid and to a peak in the velocity
of the fluid. Therefore, with q1 (t) we are able to preprocess all the other components.
Once we have performed a standard POD on the preprocessed solution manifold MzF , where
z denotes any of the components of the solution, we have at hand a set of preprocessed re-
duced spaces VNz = span{Φkz }N k=1 . Then the reduced approximation zN of the component z is
defined as
N
X
zN (ti ) = αki Φkz ◦ F̃ti , (14.47)
k=1
 306 Chapter 14. Reduced Order Methods for Fluid-Structure Interaction Problems
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 14.9. Top: comparison between the rate of decay of the eigenvalues for the pressure with
and without preprocessing. Bottom: retained energy as a function of the number N of POD modes.

Figure 14.10. Top: comparison between the rate of decay of the eigenvalues for the displacement
with and without preprocessing. Bottom: retained energy as a function of the number N of POD modes.
 14.5. A Monolithic RBM for a Transport-Dominated FSI Problem 307
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 14.11. Original snapshots for pf at times t = 0.001, t = 0.005, and t = 0.015 (left
column) and corresponding preprocessed snapshots (right column).

where F̃ti is the inverse map of the dilatation map introduced previously, with the important dif-
ference that now we are replacing the exact position of the peak γi = γ(ti ) with γ̃n , a prediction
of the position of the peak of the wave at time ti , obtained by simply evaluating q1 (ti ).
We remark that, in order to have a more stable approximation of the pressure at the reduced
order level, we enrich the reduced space associated with the fluid velocity with some supremizer
u
modes; therefore VN f = span{Φ1u , . . . , ΦN 1 Ns
u , Φs , . . . , Φs }.
u

Reduced Problem Formulation

We introduce the matrix Z p , which is a rectangular matrix whose entries are the FE degrees of
freedom of the reduced bases Φpi , returned by a POD on the preprocessed solution manifold, and
the matrix Z u,s , which is the matrix of FE degrees of freedom of the basis of the reduced space
u
VN f . Then the online system to be solved reads as follows: for every i = 0, . . . , Nt , find the
u p i Np df i N df
vector of coefficients αu (ti ) := (αku (ti ))N
k=1 , α (t ) := (αk (t ))k=1 , α (t ) := (αk (t ))k=1 ,
p i df i
ds λ u i N λu λd i N λd
αds (ti ) := (αkds (ti ))N
k=1 , α (t ) := (αk (t ))k=1 , and α (t ) := (αk (t ))k=1 such that
λu i λd i

CRB (ZF̃u αu (ti )) αu (ti )

  
MRB,u 0 0 0 0
ti
 0 0 0 0 0 0  αp (ti ) 
αdf (ti )
  
 0 0 0 0 0 0  
1 αds (ti )
  
 0 0 0 MRB,s 0 0  
 0 0 0 2
−MRB,s 0 0
 αλu (ti ) 
0 0 0 0 0 0 αλd (ti )
 
ARB + NRB (ZF̃u αu (ti )) T
BRB 0 0 −L1,T 0 αu (ti )

RB,u
ti

 BRB 0 0 0 0 0 
 αp (ti ) 
−L1,T αdf (ti )

 0 0 DR B 0 0 
+ RB,d  
−L2,T αds (ti )

 0 0 0 QRB RB,u 0  
αλu (ti )
 
L1RB,u

 0 0 0 0 0 
0 0 L1RB,d −L2RB,d 0 0 αλd (ti )
−CRB (ZF̃u αu (ti )) αu (ti−1 )
    
FRB,u MRB,u 0 0 0 0
ti
 0   0 0 0 0 0 0  αp (ti−1 ) 
0 αdf (ti−1 )
    
= 0 0 0 0 0 0
+
    
0 1
−MRB,s αds (ti−1 )
 
   0 0 0 0 0  
 0   0 0 0 2
−MRB,s 0 0
 αλu (ti−1 ) 
0 0 0 0 0 0 0 αλd (ti−1 )
αu (ti−2 )
  
0 0 0 0 0 0
 0 0 0 0 0 0  αp (ti−2 ) 
0 0 0 0 0 0 αdf (ti−2 )
  
+ . (14.48)
  
1
0 0 0 2MRB,s 0 0 αds (ti−2 )

  
 0 0 0 0 0 0  αλu (ti−2 ) 
0 0 0 0 0 0 αλd (ti−2 )
 308 Chapter 14. Reduced Order Methods for Fluid-Structure Interaction Problems

In system (14.48) we have the following:

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

(MRB,u ) = ZF̃u,T Mu ZF̃u i ,

ti t

d
1
(MRB,s ) = ZF̃u,T Ms1 ZF̃f ,
ti ti

2
(MRB,s ) = ZF̃ds ,T Ms2 ZF̃λu ,
ti ti

ARB = ZF̃u,T AZF̃u i ,

ti t

NRB (·) = ZF̃u,T N (·)ZF̃u i ,

ti t

BRB = ZF̃p,T BZF̃u i ,

ti t

d ,T
C(·)RB = ZF̃f C(·)ZF̃u i ,
ti t

d ,T d
DRB = ZF̃f DZF̃f ,
ti ti

QRB = ZF̃ds ,T QZF̃ds ,

ti ti

L1RB,u = ZF̃λu ,T L1u ZF̃u i ,

ti t

L2RB,u = ZF̃λu ,T L2u ZF̃ds ,

ti ti

d ,T
(L1RB,d ) = ZF̃f L1d ZF̃λd ,
ti ti

(L2RB,d )j,k = ZF̃ds ,T L2d ZF̃λd ,

ti ti

FRB,u = ZF̃u,T Fu . (14.49)

ti

In the expressions of (14.49), e.g., the matrix ZF̃p is defined as

ti

ZF̃p = [Φp1 ◦ F̃ti | . . . |ΦpNp ◦ F̃ti ], (14.50)

ti

d
and the other matrices ZF̃u , ZF̃f , ZF̃ds , ZF̃λu , ZF̃λd are defined similarly. Therefore in our
ti ti ti ti ti
algorithm, in order to obtain the reduced pressure at every time step, we use the basis matrix
Z p obtained by performing a POD on the pressure preprocessed solution manifold MFpf , and
we obtain a new matrix ZF̃p by a simple composition with F̃ti : ideally, the new matrix ZF̃p
ti ti

should represent a set of basis functions Φ̃pj (ti ) := Φpj ◦ F̃ti which is sufficient to express the
reduced pressure pN (ti ) in the original configuration, i.e., with the peak of the wave moving in
the domain.
It is important to underline that, following the definition of the matrix ZF̃p , it is clear that
ti
the set of basis functions used in the online step of the algorithm is time dependent: at every time
step of our simulation we compute a new set of basis functions Φ̃pj (ti ), according to the chosen
deformation map F̃ti .

14.5.6 Numerical Results

As we can see from Figure 14.12 and Figure 14.13, the online step of the method performs
very well. These results are even more striking given the fact that we were able to obtain them
 14.5. A Monolithic RBM for a Transport-Dominated FSI Problem 309
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 14.12. Pressure snapshots (left column) at time t = 0.005 (top) and final time t = 0.015
(bottom). Reduced order pressure simulation (right column) at time t = 0.005 and time t = 0.015. The
reduced simulation was obtained with N = 4 basis functions for each component of the solution of the FSI
problem.

Figure 14.13. Displacement snapshots (left column) at time t = 0.005 (top) and final time
t = 0.015 (bottom). Reduced order displacement simulation (right column) at time t = 0.005 and time
t = 0.015. The reduced simulation was obtained with N = 4 basis functions for each component of the
solution of the FSI problem.

Figure 14.14. Analysis of the behavior of the relative error for the fluid pressure approximation.
Dashed lines were obtained by employing the preprocessing procedure, continuous lines were obtained
using a standard model order reduction.

by using just N = 4 modes for each component of the solution. Since in total we have six
unknowns for our FSI problem (remember the two Lagrange multipliers used to impose the
coupling conditions), we use Ntot = 24 basis functions. To obtain similar results, but without the
preprocessing procedure, we need to use at least 10 basis functions for each component, meaning
at least 60 modes in total. Thus, during the online phase we are decreasing the dimension of
the system to be solved by at least 40. These considerations are finally confirmed by an error
behavior analysis, as shown in Figures 14.14 and 14.15. In these figures, we can see that the
overall behavior of the relative error is improved by the preprocessing procedure. In Figure
14.14 the oscillating behavior of the relative error without preprocessing shows how the set of
basis functions struggles to reproduce the exact location of the peak of each component of the
solution. On the other hand, Figure 14.15 shows that the mean relative error is very low, with
just four basis functions for each component of the solutions.
 310 Chapter 14. Reduced Order Methods for Fluid-Structure Interaction Problems
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 14.15. Behavior of the mean relative error for the fluid pressure, depending on the number
N of basis functions employed, with preprocessing (red) and without preprocessing (blue).

14.5.7 Conclusions
In this chapter, we presented two reduced order models that can be adopted when addressing
FSI problems. We first introduced the ALE formalism, adapting it to the case where a shape
parametrization is considered, and we set the formalism of multiphysics problems. We then
constructed two RBMs, both based on a POD to generate the reduced basis. The first method is
a partitioned procedure, where the coupling conditions are treated in a semi-implicit way with
respect to time, and where we added a Robin coupling condition to improve the stability of
the algorithm. At the reduced order level, we employed a change of variable and a harmonic
extension of the displacement basis functions in order to control the dimensions of the systems
to be solved online.
The second algorithm is a monolithic procedure that has been adapted in order to obtain good
performance in the case of transport-dominated problems. Here we added a preprocessing of the
snapshots during the offline phase, where we try to somehow get rid of the transport phenomenon.
At the online level, after having obtained the reduced basis functions with a standard POD, we
obtain the reduced solution as a linear combination of a pullback of these reduced bases. In this
case, a supremizer enrichment technique is used to obtain stable approximations of the pressure.
Both procedures perform very well and show the ability of the RBM to handle problems as
complex as FSI problems, even in the presence of shape parametrization.
 Chapter 15
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Reduced Order Models

for Bifurcating Phenomena
in Fluid-Structure
Interaction Problems

Moaad Khamlich, Federico Pichi, Gianluigi Rozza

15.1 Introduction
This chapter will present the development and analysis of an efficient reduced order model
(ROM) for the study of bifurcating phenomena governed by nonlinear partial differential equa-
tions (PDEs) in fluid-structure interaction (FSI) problems. In particular, the main ingredients for
this investigation are those presented in Chapter 5 on the numerical framework for the analy-
sis of the bifurcation phenomenon, and Chapter 14 for the arbitrary Lagrangian Eulerian (ALE)
formulation to be used for FSI problems [418].
We focus on a particular bifurcating phenomenon known as the Coandă effect (see Section
18.3). We deal with this phenomenon in the context of a symmetry-breaking bifurcation for the
solutions of the steady Navier–Stokes equations in a planar contraction-expansion channel. We
have chosen to deal with this particular geometry because it represents a simplified setting aimed
at analyzing a heart disease known as mitral valve regurgitation [112]. The latter occurs when the
blood flows abnormally from the left ventricle to the left atrium, due to the defective closure of
the mitral valve. It may happen that this flow, in addition to flowing in an anomalous direction,
exhibits a wall-hugging behavior, i.e., an asymmetric jet that attaches to the left atrium wall
[457]. This behavior is an example of the aforementioned Coandă effect. Moreover, from the
medical point of view, the problem is that the wall-hugging jet leads to incorrect measurements
of the blood volume through echocardiography.
Several studies have been conducted on the phenomenon. In particular, the literature is ex-
tensive in experimental [122] and numerical [543] analyses which aim at identifying the main
characteristics of the flow. The most recent investigations use reduced models [467, 457, 446,
448, 452] to alleviate the computational complexity of this type of study.
The results of these analyses show that the Coandă effect occurs even for relatively low
Reynolds numbers in the range [40, 160]. This can be partially justified by the fact that although
the Coandă effect occurs more intensely in a turbulent regime, it is amplified in two-dimensional
flows [586]. Specifically, it has been shown that, working with a fixed orifice width and decreas-
ing the kinematic viscosity, the inertial effects become progressively more important, breaking
the symmetry of the solution for a certain critical value. Therefore, below this value, one phys-
ically unstable configuration of symmetric flow and two stable asymmetric ones can be found.

311
 312 Chapter 15. ROMs for Bifurcating Phenomena in Fluid-Structure Interaction Problems

The peculiarity is that these solutions coexist for the same parameter values, as is typical of
bifurcating phenomena.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

We point out that all previous works have addressed the problem in a purely fluid dynamics
context. For this reason, in this chapter, we want to introduce a multiphysics model that allows
us to study the coupling between the fluid and an elastic solid [420] and the effect that this can
have on the bifurcation. To do this, we analyze both linear and nonlinear constitutive relations
for the structure.
Furthermore, since reconstructing the bifurcation diagrams related to a nonlinear PDE re-
quires considerable computational effort, we mitigated this cost by using a reduced order mod-
eling (ROM) technique. To do so we employed the reduced branchwise algorithm presented
in Chapter 5, which is based on a monolithic proper orthogonal decomposition (POD) coupled
with Galerkin finite elements (FEs). This approach optimizes the computational resources and
provides results that can be applied in a multiparameter context. For a more exhaustive de-
scription of the techniques and results obtained for the present investigation, we recommend
consulting [337].

15.2 Problem Formulation

In this section, we introduce the fully coupled formulation of the FSI problem. First of all, since
our model problem will be solved in a steady setting, we present its strong form.
Specifically, the equations to be solved are the momentum balance for the solid phase and
the Navier–Stokes system for the fluid phase. The considerations reported in Chapter 14 clarify
the need to use the ALE formulation [173] for the fluid phase. This involves introducing an
additional unknown field, namely the displacement of the fluid mesh, which solves a harmonic
extension problem.
We denote by Ω ⊂ R2 the physical domain in which the problem is solved. We further
assume that this domain is decomposed into a resolution domain for the solid Ωs and one for
the fluid Ωf , which do not overlap and are such that Ω = Ωf ∪ Ωs . Since we employ the ALE
description for the fluid and the Lagrangian description for the solid, we need to introduce two
distinct reference domains, i.e., one for the solid Ω̂s and one for the fluid Ω̂f (see Figure 15.1).
The unknowns are the global displacement d = (ds , df ), the global velocity u = (us , uf ), and
the pressure p. By combining the equations for the solid and fluid phases into one system, we
obtain a complete formulation for the FSI problem:

ρf (uf · ∇uf ) − div σf (uf , pf ) = bf in Ωf ,



 div u = 0 in Ω ,

f f
ˆ f = 0 in Ω̂f , (15.1)
 ∆d
ˆ P(dˆs ) = b̂s in Ω̂s ,

−div


where ρs (respectively ρf ) is the density of the solid (respectively fluid), while b̂s (respectively
bf ) represents the external volume forces for the solid (respectively fluid).
In system (15.1) the symbol ∧ is used to denote operators and fields taken with respect to the
reference variable x̂.
The Cauchy stress tensor σf and the second Piola–Kirchoff stress tensor P are prescribed by
means of the following constitutive relations:
  
T


 σf (u f , pf ) = ρ f µ ∇uf + ∇u f − pI,
ˆ (15.2)
P(ds ) = λs tr 
Es (ds )I + 2µs Es (ds ),
 Es dˆs = 1 ∇ ˆ dˆs + ∇
ˆ T dˆs ,


2
 15.2. Problem Formulation 313

where λs and µs are the Lamé parameters. The kinematic viscosity of the fluid µ has a prominent
role in the present chapter since it is the bifurcation parameter under investigation.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

In order to solve the former problem, we need to pull back the Navier–Stokes equations on
the reference configuration Ω̂f . To do so, we first introduce the gradient of the ALE map and its
determinant, namely,  
F := ∇ ˆ x̂ + dˆf and J := det F, (15.3)
and then we use the former quantities to give a reference representation of σf , i.e.,
 
σ̂f = µ ∇ûˆ f F−1 + F−T ∇ ˆ T ûf − p̂I. (15.4)

We exploit the Piola transformation rule, which leads to the following system:
  

 ρf J +∇ûˆ f F−1 (ûf ) − divˆ (J σ̂f (ûf , pf )F−T ) = Jbf in Ω̂f ,

ˆ (JF−1 uf ) = 0 in Ω̂f ,
div

(15.5)

 ˆ
∆df = 0 in Ω̂f ,


−divˆ P(dˆs ) = bs in Ω̂s .

This set of equations is then completed by means of the following conditions:

• Coupling conditions:

dˆf = dˆs on Γ̂F SI ,


(15.6)
P · n̂ = (Jσf F−T ) · n̂ on Γ̂F SI .

• Boundary conditions:

J σ̂f (ûf , p̂f ) F−T n̂ = −pin (t)n̂ on Γ̂in ,



J σ̂f (ûf , p̂f ) F−T n̂ = −pout (t)n̂ on Γ̂out ,


(15.7)
û = 0 on Γ̂0 ∪ Γ̂F SI ,
 ˆf


ds = 0 on Γ̂D s ,

where n̂ the normal outgoing vector from the fluid resolution domain. See Figure 15.1 for
a sketch of the different portions of the boundaries.

15.2.1 Variational Formulation

To solve problem (15.5) by the finite element method (FEM), we need to introduce its weak
formulation. Since everything is now expressed in the reference configuration Ω̂, there is no
need to distinguish between different coordinate systems. Therefore, we will drop the symbol ∧,
which so far has been used for both fields and operators.
We then define the following functional spaces:

Vf = H 1 (Ω̂f ; Rn ),


Qf = L2 (Ω̂f ),


(15.8)
E = H 1 (Ω̂f ; Rn ),
 f


Vs = H 1 (Ω̂s ; Rn ).
We multiply each equation of system (15.5) by an appropriate test function belonging to the
corresponding functional spaces (15.8) and apply the divergence theorem. This leads to the
following equations:
 314 Chapter 15. ROMs for Bifurcating Phenomena in Fluid-Structure Interaction Problems

Γ̂D
s
Γ̂0
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Γ̂F SI

Γ̂in Ω̂s Ω̂f Γ̂out

Γ̂F SI
Ω̂s

Γ̂D
s Γ̂0

Figure 15.1. Reference configuration for the FSI problem.

• Fluid momentum equation:

Jρf [∇uf F−1 ]uf · vf dx̂ + Ω̂f Jσf F−T : ∇vf dx̂
R R
Ω̂R
f
(15.9)
− ∂ Ω̂f Jσf F−T n · vf dAx̂ − Ω̂f Jbf · vf dx̂ = 0 for every vf ∈ Vf .
R

• Fluid incompressibility equation:

Z
− div (JF−1 uf )qf dx̂ = 0 for every qf ∈ Qf . (15.10)
Ω̂f

• Extension equation:
Z
∇df : ∇ef dx̂ = 0 for every ef ∈ Ef . (15.11)
Ω̂f

• Solid momentum equation:

Z Z Z
P : ∇vs dx̂ − Pn · vs dAx̂ − bs · vs dx̂ = 0 for every vs ∈ Vs . (15.12)
Ω̂s ∂ Ω̂s Ω̂s

A first simplification comes from summing up the two equations (15.9) and (15.12). In fact, by
doing so one can directly exploit the dynamic coupling condition in the following way:

− ∂ Ω̂f Jσf F−T n · vf dAx̂ − ∂ Ω̂s Pn · vs dAx̂

R R

= Γ̂F SI (Jf σf F−T

R R R
f luid n) · (vs − vf ) dAx̂ + Γ̂in pin n · vf dAx̂ + Γ̂out pout n · vf dAx̂ .
(15.13)
Care must be taken because Pn = (Jσf F−T f luid )n was substituted to obtain relation (15.13),
which means that the stress at the interface was assigned to the fluid subproblem, while this term
appears as forcing in the solid equation, as can be deduced from the test functions involved.
 15.2. Problem Formulation 315

Therefore, by using the former result we can join (15.9) and (15.12) in a single fluid-structure
momentum equation as follow:
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Jρf [∇uf F−1 ]uf · vf dx̂ + Ω̂f Jσf F−T : ∇vf dx̂
R R
Ω̂R
f

+ RΓ̂F SI (Jf σf F−T

R
R s − vf ) dAx̂ + Γ̂in pin n · vf dAx̂
f luid n) · (v
+ RΓ̂out pout n · vf dA (15.14)
R x̂ − Ω̂s P : ∇vs dx̂
− Ω̂s bs · vs dx̂ − Ω̂f Jbf · vf dx̂ = 0
for every (vf , vs ) ∈ Vf × Vs .

15.2.2 Lagrange Multipliers Method

We are now in a place to talk about the imposition of boundary conditions. This will be done
now using a Lagrange multipliers approach [634]. The idea is to weakly impose the constraints
uf = 0 and df = ds by introducing two new unknowns, called Lagrange multipliers. In order
to do so, let us define the following functional spaces:

Λu = H 1 (Γ̂F SI ; Rn ),

(15.15)
Λd = H 1 (Γ̂F SI ; Rn ).

The constraints will be incorporated in the following way:

• Geometric adherence constraint: Condition df = ds can be imposed by means of a

Lagrange multiplier denoted by λd ∈ Λd , which will change (15.11) in the following way:
R
Ω̂R
f
∇df : ∇ef dx̂
+ RΓ̂F SI (df − ds ) · mλd dx̂
(15.16)
− Γ̂F SI λd · (ef ) dAx̂ = 0
for every (ef , mλd ) ∈ Ef × Λd .

• Kinematic Constraint: Condition uf = 0 can be imposed by means of a Lagrange mul-

tiplier denoted by λu ∈ Λu , which will change (15.14) in the following way:
−1
Jσf F−T : ∇vf dx̂
R R
Ω̂R
f
Jρ f [∇u f F ]uf · v f dx̂ + Ω̂f
−T
R
+ RΓ̂F SI (Jf σf Ff luid n) · (vR s − vf ) dAx̂ + Γ̂in pin n · vf dAx̂
+ RΓ̂out pout n · vf dAx̂ − Ω̂Rs P : ∇vs dx̂
(15.17)
+ RΓ̂F SI (uf ) · mλuRdAx̂ − Γ̂F SI λu · (vf − vs ) dAx̂
− Ω̂s bs · vs dx̂ − Ω̂f Jbf · vf dx̂ = 0
for every (vf , vs , mλu ) ∈ Vf × Vs × Λu .

15.2.3 Branchwise Reduced Basis Approximation

The system of equations (15.16) and (15.17) represents the special case of the weak formulation
for a parametric PDE and thus can be expressed as follows: given µ ∈ P, find X(µ) ∈ X such
that
f (X(µ), Y ; µ) = 0 ∀Y ∈ X, (15.18)
where the solution X(µ) represents the collection of all unknown fields, while the generic pa-
rameter µ in our case is the aforementioned kinematic viscosity. A peculiarity of the present
case concerns the nonlinearity of (15.17), due to the convective acceleration term. For this rea-
son, in this section, we will briefly explain how to deal with the nonlinearity using the reduced
 316 Chapter 15. ROMs for Bifurcating Phenomena in Fluid-Structure Interaction Problems

basis method (RBM). To do so, we will follow the approach of [389]. The first critical observa-
tion is that we will only present a basic strategy of projecting the nonlinear governing equations
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

onto the reduced basis. This, in principle, does not guarantee substantial computational savings
because the complexity of the procedure still depends directly on N (the cardinality of the high-
fidelity model). An alternative to the presented methodology is hyperreduction strategies such
as the empirical interpolation method (EIM) [40, 245], which can be used to achieve an online
computation cost independent of N . However, such strategies require further complexity in the
offline phase and seem to lack robustness when applied to bifurcating problems [446, 450]. Let
us denote by XN the span of the reduced basis obtained via POD (see Chapter 2), where the car-
dinality of this basis is N . The reduced problem therefore reads as follows: find XN (µ) ∈ XN
such that
f (XN (µ), YN ; µ) = 0 ∀ YN ∈ XN . (15.19)
This reduces, as for the full-order model (FOM), to a system of N nonlinear equations, which
can be solved by means of the Newton method, namely, gien an initial guess XN 0
∈ XN for
k = 0, 1, . . . , do the following:
1. Solve the linear variational problem:
 k k
(15.20)


df XN (µ) (δXN , YN ; µ) = −f XN (µ); YN ; µ .

2. Update the solution:

k+1 k
XN = XN + δXN . (15.21)

From the discussion in Chapter 2 we know that problem (15.20) can be cast in matrix form as

JN XkN (µ); µ δXN = −FN XkN (µ); µ , δXN ∈ RN , JN ∈ RN × RN , (15.22)

where JN is the reduced Jacobian.

The issue is that because the reduced basis is expressed in terms of the full-order basis,
one cannot directly compute the quantities in (15.22), but one has to deal with their full-order
counterpart and transform them by employing the matrix V ∈ RN ×N , whose columns are the
coefficients of the reduced basis functions {Ψk }N k=1 in terms of the full-order basis functions
i=1 . This matrix has the following property:
{φi }N
N
X
Ψk = Vik φi , Vik = (V)ik . (15.23)
i=1

One can therefore prove that the linear system (15.22) is equivalent to

VT Jh VXkN (µ); µ VδXN = −VT Fh VXkN (µ); µ . (15.24)

At this point, we recall that our goal, as discussed in Chapter 5, is to be able to follow a single
branch while changing the value of the parameter we are dealing with. This task is not trivial
since, after the bifurcation has occurred, we lose the uniqueness of the solution, and thus the
convergence of the Newton method is strongly tied to the initial hypothesis. Consequently, we
need a strategy to follow the solution as the parameter varies. For this reason, we will implement
a continuation method, i.e., a procedure to generate a succession of solutions associated with
the same branch [170]. We refer the reader to Subsection 5.3.1 for details on the algorithm
for full-order branchwise reconstruction. In order to use the RBM, it is therefore necessary to
use Algorithm 5.1 by operating on a discrete version of the parameter space, i.e., Ptrain =
{µ1 , µ2 , . . . , µNtrain } ⊂ P.
 15.3. Application to the Coandă Effect 317

Thus, having the set of solutions {X(µj ) | µ ∈ Ptrain }, one can extract the reduced space
XN using the POD. At this point, the branch is reconstructed using the reduced order counterpart
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

of Algorithm 5.1, which is almost identical to its FOM version. The only differences lie in
substituting the FOM quantities by their reduced order equivalent and taking account of the
previous considerations for the Jacobian matrix. The result is presented in Algorithm 15.1.

A LGORITHM 15.1. Branchwise reduced order reconstruction.

1: XN,0 = XN,guess . Initial guess
2: for µj ∈ Ph . Continuation loop
3: X0N,j = XN,j−1 . Continuation guess
4: while ||FN XkN,j ; µj ||RN >  do . Newton method

5: VT Jh VXkN,j ; µj VδXN,j = −VT Fh VXkN,j ; µj

6: Xk+1 k
N,j = XN,j + δXN
7: end while
8: end for

15.3 Application to the Coandă Effect

In this section, we present the numerical results for the Coandă effect test case in an FSI setting.
We conducted this investigation by exploring different models for the constitutive equation of
the solid phase. In addition, we have also coupled the full-order resolution with the reduced
strategy discussed in the previous section, testing its effectiveness for new parametric instances
belonging to the branch we have chosen to approximate. Finally, the various proposed models
were compared by including the rigid structure case already presented in Chapter 5 to accurately
characterize the effect of the elastic solid on the bifurcation. We refer the interested reader to
[337] for a more complete description of the following results.

15.3.1 Preliminaries
Let us start by noting that we have used the centimeter-gram-second unit system for all the
equations. In the following, we omit the units of measurement of the quantities we will be
working with to simplify the notation. The dimensions of the computational domain are reported
in Figure 15.2. In particular, the leaflets (in magenta) represent the solid resolution domain Ω̂s ,
whereas the blue interstice is the fluid resolution domain Ω̂f .
For the particular case under consideration, we imposed the following boundary conditions:

J σ̂f (ûf , p̂f ) F−T n̂ = −450n̂ on Γ̂in ,


(15.25)
J σ̂f (ûf , p̂f ) F−T n̂ = 0 on Γ̂out ,

5. 1. 44.

2.5

2.5 7.5

2.5

50.

Figure 15.2. Computational domain.

 318 Chapter 15. ROMs for Bifurcating Phenomena in Fluid-Structure Interaction Problems

where 
Γ̂in = 0 × [0, 7.5],
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

(15.26)
Γ̂out = 50 × [0, 7.5].
The previous conditions are equivalent to enforcing a stress gradient between the outlet and the
inlet of the channel, which therefore animates the fluid flow, with direction from Γ̂in to Γ̂out .
The FOM problem was solved by employing the FEM. In particular, the meshes we used for the
spatial discretization are unstructured and composed of triangular elements. The discretization
performed is second order for uf , df , and ds , while it is first order for p, λd , and λu . This
choice was made to avoid known stability issues typical of saddle-point problems [65, 74].
The software employed to solve the FOM problem is multiphenics [6], which has been cou-
pled with RBniCS [38] for the reduced order phase.

15.3.2 FSI Problem with Linear Elasticity

In this subsection we present the numerical results for the FSI problem with a linear elastic
constitutive relation for the solid phase, i.e.,
ˆs ) = λs tr Es (ds )I + 2µs Es (ds ),
(
P(d    (15.27)
Es dˆs = 12 ∇ ˆ dˆs + ∇
ˆ T dˆs .

In particular, the previous constitutive relation is completely characterized by the Lamé con-
stants, which in our case assume the following values:

Physical quantity Unit of measure

λs (Lamé’s first parameter) 8 · 105 barye (Ba)
µs (Shear modulus) 105 barye (Ba)

For the bifurcation analysis, we analyzed the behavior of the solution for the kinematic viscosity
interval P = [0.5, 2.0], discretized by using 51 points. In the following, we will denote this
discretization by Ph .
The first phase was conducted by computing the set of FOM solutions for each value of
the kinematic viscosity µj in Ph . In this regard, Figure 15.3 shows the velocity and pressure
fields obtained for some of the parametric instances considered during this step. Figure 15.4
reports the displacement field of the solid df for certain values of the kinematic viscosity. Each
of these fields is associated with the fictitious displacement of the fluid mesh df , which is its
extension, as prescribed by the harmonic problem (see Figure 15.5). At this point we want to plot
the bifurcation diagram for the problem under consideration (see Chapter 5). In particular, the
significant scalar output we choose is the vertical velocity of the fluid at a point of the channel
symmetry axis upstream of the expansion, i.e., uy (x), with x = (14, 3.75). The diagram in
question is shown in Figure 15.6. The result obtained is consistent with the analysis presented
in Subsection 5.7.3, in which the problem was solved in the absence of the elastic structure, and
therefore only for the fluid phase. In that case, a pitchfork bifurcation was identified, which is
also present for the FSI problem. In particular, in our case we focused on obtaining the lower
stable branch, which was obtained by recursively decreasing the value of the kinematic viscosity.
The next step is to obtain a bifurcation diagram for the solid phase. Since the particular
bifurcation faced is caused by the nonlinearity of the Navier–Stokes system, it is interesting to
understand how the structure undergoes it. In this case, we used the scalar indicator
∆d = | max dˆs − max dˆs |, (15.28)
x̂∈Ω̂1 x̂∈Ω̂2
 15.3. Application to the Coandă Effect 319
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

f_234 Magnitude
0.0e+00 5 10 15 20 25 2.9e+01

Figure 15.3. Velocity and pressure snapshots.

Figure 15.4. Solid’s deformation snapshots: the solutions are given in a spatial frame using the df map.

Figure 15.5. Global displacement (df , ds ) for µ = 2.0.

 320 Chapter 15. ROMs for Bifurcating Phenomena in Fluid-Structure Interaction Problems

0.0
0.014
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

− 0.5
0.012

− 1.0
0.010
u y (14, 3.5)

− 1.5
0.008

Δd
− 2.0
0.006

− 2.5
0.004

− 3.0
0.002

− 3.5
0.000

0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0
μ μ

Figure 15.6. Bifurcation diagram for the Figure 15.7. Bifurcation diagram for the
fluid phase. solid phase.

where Ω̂1 is the upper leaflet and Ω̂2 is the lower one. This quantity is significant in plotting
the bifurcation diagram presented in Figure 15.7 because it represents an indirect measure of the
imbalance of the forces acting on the two leaflets, following the onset of the asymmetric flow
downstream of the expansion.

15.3.3 RBM Results

In this section, we present the results of the reduction phase for the FSI problem. In particular,
we used the snapshots coming from the continuation method presented in the previous section,
and we extracted a reduced basis for each of the fields involved in the FSI problem.
The first indicator to use in order to choose the cardinality of the reduced basis is the trend of
the POD’s singular values for the different fields. A general approach relies on a normalization
with respect to the first eigenvalue, which is also the maximum in terms of magnitude:

σi
σ̂i = . (15.29)
σ1

Furthermore, if we take into account the cumulative sum of the eigenvalues, we get another
important indicator known as the retained energy:
Pn
σi
En = PNi=1
train
, (15.30)
i=1 σi

with Ntrain = 51 being the number of snapshots computed during the offline phase. The trend
of both these indicators is presented in Figure 15.8.
We note an exponential decay of the eigenvalues as a function of the cardinality of the basis.
In particular, the two displacement fields decay faster than the pressure and the velocity of the
fluid. This is also confirmed by the trend of retained energy, whose values for the first mode are
presented in Table 15.1.

Table 15.1. Energy retained by the first mode.

uf p df ds
En 92.9% 90.1% 99.9 % 99.9%
 15.3. Application to the Coandă Effect 321

10
0
100 p
uf
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

df
9.8 × 10
−1
ds
10−2

−1
9.6 × 10

10−4
En

σn̂
−1
9.4 × 10

10−6

9.2 × 10
−1
p
uf
df 10−8
9 × 10
−1 ds

10
0
10
1
0 10 20 30 40 50
n n

Figure 15.8. Retained energy. Figure 15.9. Decay of the normalized eigenvalues.

Thus, a single mode would be sufficient to retain 99.9% of the energy (variance) of the dis-
placement’s snapshots. In fact, we recall that the problem associated with ds is that of linear
elasticity, while that for df is a harmonic extension, which is also linear. Consequently, the
problems’ linearity justifies the lower complexity of their solution manifold.
We decided to investigate the trend of the mean error by varying the number of bases used
by the RBM. The reconstruction efficiency was measured by the metric
NX
train
1 kXh (µj ) − Xrb (µj )kX
Errrb = . (15.31)
Ntrain j=1
kXh (µj )kX

In Figure 15.10 we observe that, as expected, the errors are higher than their direct projection
counterparts presented in Figure 15.11. Moreover, we can observe that the decay, although not
monotone, maintains an average exponential trend. To refine the reconstruction of the bifurcation
diagram, we decided to use 101 equispaced points within the interval P. These kinematic vis-
cosity values were employed to reconstruct the lower stable branch using Algorithm 15.1. Each
of the reduced solutions was subsequently compared to its FOM counterpart by calculating the
relative reconstruction error (Figure 15.12). In particular, despite the fluctuations due to the local
effect of approximation improvement by each of the basis functions, we observe that the average
trend is in line with what was previously observed in Figure 15.10 for the mean errors. In this
case, the reconstructed pressure worse than the other fields. Finally, we note the increase of the
error both for low-viscosity regimes, around µ = 0.5, and at the bifurcation point.

10−1 p 10−1 p
df df
ds 10−2 ds
uf uf
10−2 10−3

10−4
Errproj

10−3
Errrb

10−5

10−6
10−4
10−7

10−5 10−8

2 4 6 8 10 12 14 16 2 4 6 8 10 12
n n

Figure 15.10. Average reconstruction er- Figure 15.11. Average projection error
ror in logarithmic scale. in logarithmic scale.
 322 Chapter 15. ROMs for Bifurcating Phenomena in Fluid-Structure Interaction Problems
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 15.12. Relative reconstruction error with Nrb = 16.

15.4 Model Comparison

In this section, we compare various models. This is done in order to obtain insights into the
behavior of the bifurcation point as a function of the constitutive relation for the solid phase and
of its material-dependent parameters. For this reason, we decided to consider, in addition to the
linear case examined in the previous section, the following two test cases:

• The rigid structure case represents the incompressible fluid dynamics problem already
presented in Chapter 5. In this case, the Navier–Stokes equations are solved, removing the
structure’s description from the computational domain (see Figure 15.2).

• The Saint Venant–Kirchoff (SVK) model is a nonlinear hyperelastic model for which the
constitutive relation for the Piola tensor is

P(ds ) = F(λs tr Es (ds )I + 2µs Es (ds )),


(15.32)
Es (ds ) = 12 ∇ds + ∇T ds + +∇T ds ∇ds .

For the FSI problem, addressed with both the linear and SVK models, we decided to use as
independent parameters, within the constitutive relation, Young’s modulus (E) and Poisson’s
ratio (ν), which can be used to express the Lamé constants as

E Eν
µs = , λs = . (15.33)
2(1 + ν) (1 + ν)(1 − 2ν)

In particular, the results of the previous section for the linear model refer to the test case where
E = 2.89e5 and ν = 0.44. In that case, introducing the structure does not seem to alter the
bifurcation phenomenon, since the high stiffness results in small deformations. For this reason,
we decided to work in a regime of large bifurcations by reducing the structure stiffness under
consideration for both the linear and the SVK models, as reported in Table 15.2. In fact, both
the linear (A) and the SVK cases lead to deformations of the same order of magnitude as the
resolution domain.
 15.4. Model Comparison 323

Table 15.2. Values of Young’s modulus and Poisson’s ratio used for the test cases.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Case E ν
linear (A) 4.67e4 0.21
linear (B) 2.89e5 0.44
SVK 4.67e4 0.21

Figure 15.13. Maximum displacement Figure 15.14. Solid’s bifurcation dia-

for the SVK and the linear models. gram for the linear and the SVK model.

0 SVK
linear (A)
linear (B)
rigid body
−1
uy(14, 3.75)

−2

−3

−4

0.50 0.75 1.00 1.25 1.50 1.75 2.00

μ

Figure 15.15. Comparison of the bifurcation diagram for the fluid phase for the different test cases.

A first comparison, presented in Figure 15.13, concerns the maximum displacement of the
two leaflets for the linear (A) and SVK model. We notice that although the SVK model generates
solid deformations of higher magnitude, it reduces the asymmetry of the behavior between the
two leaflets in the postbifurcation regime, as evidenced by the trend of ∆d in Figure 15.14. As a
result, the response of the linear model in front of the bifurcation is accentuated compared to its
nonlinear counterpart, although the deformation field is of lower intensity. To understand how
this difference in behavior is reflected in the fluid phase, we finally examined the bifurcation
diagram relative to the already employed quantity uy ((14, 3, 75)). Figure 15.15 shows how the
 324 Chapter 15. ROMs for Bifurcating Phenomena in Fluid-Structure Interaction Problems

different constitutive relation for the solid has an important influence on the bifurcation point
for the fluid. In particular, we have two linear models (A and B) that relate to cases of more
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

or less rigid structure. The increase in structural stiffness in the linear case leads to a delay
in bifurcation, and in fact the linear model (B) seems to not modify the bifurcation point with
respect to the rigid structure case. The SVK model, on the other hand, delays the bifurcation with
respect to the rigid case but anticipates it when compared to its linear (A) counterpart. Finally, it
can be seen that the SVK and the linear models (A) differ most at the bifurcation point, showing
similar behavior once the bifurcation has occurred.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Part IV

and Applications
Further Advances, Perspectives,
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Chapter 16

Reduction in Parameter
Space

Marco Tezzele, Francesco Romor, Gianluigi Rozza

16.1 Introduction
Parameter space reduction has been established in recent years as a valuable tool to fight the curse
of dimensionality, which means that the complexity of the algorithms grows exponentially with
the dimension of the input parameter space. Reduced order methods [281, 517, 508, 503] have
been developed to deal with complex parametric problems, but the limit for high-dimensional
parameter spaces still remains. Nowadays the problem is even more emphasized thanks to the
spread of high-performance computing facilities which enable the study of highly parametrized
systems. On the other hand, if we consider modern industrial optimization we also find the
presence of many input parameters in order to analyze a wide range of possible designs.
The active subspace (AS) [134, 136] method has emerged as a reliable and explainable tech-
nique for linear dimensionality reduction of the input parameter space. This success is also
testified to by the range of successful applications, including shape optimization [383, 161, 157],
turbomachinery [534], chemistry [311, 310, 597], automotive engineering [432, 494], and naval
and nautical engineering [574, 571, 404, 155, 572, 575].
The methodology has been innovated through the development of several extensions, e.g.,
ASs for vector-valued functions [635], multifidelity approximations [356], localized models
[495], and kernel-based ASs [493]. Moreover, other nonlinear techniques are emerging, such
as nonlinear level-set learning [643] and the active manifold method [77]. We will present all
these methods in the following sections, highlighting their characteristics and providing a gen-
eral picture for parameter space reduction. Many of the techniques described in this chapter are
implemented in the open-source Python package8 called ATHENA [496]; the reader can also
refer to Chapter 19.
This chapter is organized as follows: in Section 16.2 we introduce a general framework for
the ridge approximation problem and we present several techniques based on the AS method,
such as kernel-based active subspaces (KASs), local active subspaces (LASs), some gradient-
free extensions, and a multifidelity setting; in Section 16.3 we briefly present other nonlinear
approaches to parameter space dimensionality reduction; Section 16.4 is devoted to showing a

8 Available at https://github.com/mathLab/ATHENA/.

327
 328 Chapter 16. Reduction in Parameter Space

wide range of possible applications of the described techniques, with some comparisons. Finally,
in Section 16.5 we draw some conclusions and describe some future perspectives for parameter
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

space reduction.

16.2 AS-Based Methods

In this section we present techniques based on the search of a linear subspace of the input space
to approximate the function of interest. This can be viewed in the context of ridge recovery. We
define a ridge function [453] as any real-valued multivariate function F : Rn → R such that

F (x1 , . . . , xn ) = f (a1 x1 + · · · + an xn ) = f (a · x), (16.1)

where a ∈ Rn \ {0} is called the direction and f is a univariate real-valued function, i.e.,
f : R → R. We note that F is constant on the hyperplanes a · x = c, where c ∈ R. This
concept can be generalized by considering more than one direction. We call the generalized
ridge function any real-valued multivariate function F : Rn → R such that

F (x) = f (AT x), (16.2)

where A ∈ Rn×d is a tall matrix, with 1 ≤ d < n, and f : Rd → R. The goal of ridge recovery
is to estimate the matrix A using only point queries of f .
The methods presented below seek to approximate a target function of interest with a gener-
alized ridge function.

16.2.1 Active Subspaces

Let X : (Ω, F, P ) → Rn be the absolutely continuous random vector representing the n-
dimensional inputs with probability density function ρ : X ⊂ Rn → R, and let f : X ⊂ Rn → R
be the model function that we assume to be continuously differentiable and Lipschitz continuous.
Let Σ be the second moment matrix of ∇f defined as
Z
Σ := Eρ [∇x f ∇x f ] = (∇x f )(∇x f )T ρ dL,
T
(16.3)

h iT
∂f ∂f
where Eρ denotes the expected value, ∇x f = ∇f (x) = ∂x 1
, . . . , ∂xn is the column vector of
partial derivatives of f , and dL is the Lebesgue measure. Since Σ is symmetric positive definite,
it has an eigendecomposition with real eigenvalues Σ = WΛWT .
We can retain only a small number r  n of eigenpairs: r becomes the AS dimension, and
the span of the corresponding eigenvectors defines the AS. The partition is
 
Λ1 0
Λ= , W = [W1 W2 ] , (16.4)
0 Λ2

where Λ1 = diag(λ1 , . . . , λr ) and W1 contains the first r eigenvectors arranged in columns.

With this matrix we can project the input parameters onto the AS, and its orthogonal complement,
i.e., the inactive subspace, is as follows:

Y = Pr (X) = W1 W1T X ∈ Rn , Z = (I − Pr )(X) = W2 W2T X ∈ Rn . (16.5)

The AS dimension r can be selected a priori, or by looking at the presence of a spectral gap
[134], or by imposing a cumulative energy threshold for the eigenvalues.
 16.2. AS-Based Methods 329

Sampling the Inactive Variable

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

We remark that the projection map onto the AS is not an injective map because W1T is defined
as a linear projection onto a subspace. So the back-mapping from the AS onto X is not trivial.
Let x∗M be a point in the AS; we can find K ∈ N+ points in the full parameter space which are
mapped onto x∗M by W1T . Following from the decomposition introduced above we have

x = W1 W1T x + W2 W2T x = W1 xM + W2 y ∀x ∈ X , (16.6)

with the additional constraint coming from the rescaling of the input parameters needed to apply
the AS: −1 ≤ x ≤ 1. By 1 we denote the vector in Rn with all elements equal to one. To
sample the inactive variable y we need to satisfy

−1 ≤ W1 x∗M + W2 y ≤ 1 (16.7)

or, equivalently,
W2 y ≤ 1 − W1 x∗M − W2 y ≤ 1 + W1 x∗M . (16.8)
These inequalities define a polytope in R from which we want to uniformly sample K points.
n−r

The inactive variables are in general sampled from the conditional distribution p(y|x∗M ); here
we show how to perform it for the uniform distribution. For more general distributions one could
use Hamiltonian Monte Carlo. Here we start with a rejection sampling scheme, which finds a
bounding hyperbox for the polytope, draws points uniformly from it, and rejects points outside
the polytope. If this method does not return enough samples, we use a hit and run method
[542, 50, 381] to sample from the polytope. This method starts from the center of the largest
hypersphere within the polytope and then selects a random direction and identifies the longest
segment lying inside the polytope. The new sample is randomly drawn along this segment. This
procedure continues by iterating the same steps starting from the last sample until K samples are
found. If that does not work, we select K copies of the Chebyshev center [70] of the polytope.
In Figure 16.1 we depict the sampling strategy at different stages.

Chebyshev center Polytope

New samples

Figure 16.1. Illustration of the inactive variable sampling strategy. Successive steps are depicted
all at once. We highlight the Chebyshev center, the selection of the next sample using the hit and run method,
and the polytope defined by (16.8).

16.2.2 Kernel-Based Active Subspaces

In this section we introduce the kernel-based extension of the AS property (KAS) [493], which
consists of a ridge approximation of the model of interest factorized through a projection onto
a reproducing kernel Hilbert space (RKHS). In this sense it is a nonlinear extension, since the
projection is nonlinear. It is based on the preliminary work [434] in the context of supervised
 330 Chapter 16. Reduction in Parameter Space

dimension reduction algorithms in machine learning. KAS is implemented in the open-source

Python library ATHENA [496]; some tutorials on basic test cases are provided there.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Sometimes the function of interest does not directly admit an accurate ridge approxima-
tion but can nevertheless be approximated by it after a nonlinear transformation. Even if, in-
stead of (16.2), the function of interest is factorized through an optimal profile [134], f (x) ≈
Eρ [f |W1 W1T ](W1 W1T x), i.e., the random variable f (X) conditioned on W1 W1T X, this may
still not be enough to provide an accurate approximation.
A fairly simple example is radial symmetric functions, with generatrix g : R+ → R, of the
form
f (x) = g(kxk) ∀x ∈ X ⊂ Rn (16.9)
in a radial domain X with respect to a fixed origin. For each chosen active direction, the optimal
profile is the average of the radial symmetric function over the orthogonal inactive direction
p(xM , y)
Z
Eρ [f |W1T ](xM ) = f (x) dy,
p(y)
but a more natural factorization would associate with each input its norm (as an encoded feature)
and then choose the generatrix as profile.
In order to overcome these problematic situations, nonlinear dimension reduction is em-
ployed. The best way to achieve nonlinear dimension reduction with a one-dimensional reduced
parameter space could be realized by discovering a curve in the space of parameters that cuts the
level sets of f transversely; this variation is presented in Subsection 16.3.2 as an active manifold.
Based on similar reasoning, nonlinear level-set learning (NLL) [643] is introduced in Subsec-
tion 16.3.1.

Factorization through an RKHS

As anticipated, we consider an immersion φ from the space of parameters X to an infinite-
dimensional Hilbert space H, obtaining the factorization
φ
X ⊂ Rd φ(X ) ⊂ H
f˜
f
R
This is a common procedure in machine learning [61]. Then AS is applied to the new model
function f˜ : φ(X ) ⊂ H → R with parameter space φ(X ) ⊂ H. In practice we consider
a discretization of the infinite-dimensional Hilbert space RD ' H with D > n. Sufficient
dimension reduction with KAS results in the choice of a d-rank projection in the much broader
set of d-rank projections in H. A schematic view of the factorization with the RKHS is shown in
Figure 16.2.
A response surface can be built, remembering to replace every occurrence of the inputs X
with their images φ(X). However, we revert to having a retrievable backward map as explained
in Subsection 16.2.1.
To stress the fact that with this factorization we introduce an element of complexity that
increases the representation capabilities of the ridge approximation, we expand the projected
inputs with KAS:
(W1T φ)(X) = w1 (w1 · φ(X)) + · · · + wd (wd · φ(X)),
with {wi } ⊂ H the active eigenvectors of KAS. Thus, the reduced latent input space is a non-
linear combination of the inputs, employing as basis of the feature space the nonlinear functions
 16.2. AS-Based Methods 331
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Feature space
mapping
Active
subspace

Regression

Projection

Active variable

Figure 16.2. Illustration of the design of a response surface using KAS and Gaussian process
regression (GPR). The factorization through an RKHS is emphasized.

{wi · φ}i∈{1,...,d} . In comparison, in the AS method we use the following linear encoding of the
random variable X:
W1T X = w1 (w1 · X) + · · · + wd (wd · X),

with {wi } ⊂ Rd the active eigenvectors (i.e., the columns of matrix W1 ).

There is no need to say that the choice of the feature map φ is crucial, but for clarity it will
be introduced later. We just require it to be an immersion φ : X ⊂ Rn → H.

Evaluation of the Correlation Matrix

Since for AS only the samples of the Jacobian matrix of the model function are employed, we
can ignore the definition of the new map f˜ : φ(X ) ⊂ H → R and focus only on the computation
of the Jacobian matrix of f˜ with respect to the new input (latent) variable y := φ(X). The
correlation matrix becomes
Z Z
H̃ = ˜ T ˜
(Dy f ) (y)(Dy f )(y) dµ(y) = (Dy f˜)T (Dy f˜) dLX (x),
φ(X ) X φ(x) φ(x)

where µ := φ ◦ LX is the push-forward probability measure of LX , i.e., the probability distribu-

tion of X, with respect to the map φ. Simple Monte Carlo can be applied by sampling from the
distribution with density ρ in the input space X :

M
1 X
Z
H̃ = (Dy f˜)T (Dy f˜) dLX (x) ≈ (Dy f˜)T (Dy f˜) .
X φ(x) φ(x) M i=1 φ(Xi ) φ(Xi )
 332 Chapter 16. Reduction in Parameter Space

The gradients of f˜ with respect to the new input random variable y are computed from the
known values DX f with the chain rule: if the Hilbert space H has finite dimension H ∼ RD ,
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

this procedure leaves us with an underdetermined linear system to solve for Dy f˜:

Dy f˜(φ(X))Dφ(X) = DX f (X) ⇒ Dy f˜(φ(X)) = DX f (X)(Dφ(X))† ,

where Dy f˜(φ(X)) and DX f (X) are row vectors and †

stands for the right Moore–Penrose
inverse with rank r, i.e.,
(Dφ(X))† = V Σ† U T ,

with the usual notation for the singular value decomposition (SVD) of Dφ(X),

Dφ(X) = U ΣV T ,

and Σ† ∈ M(r × r) equal to the diagonal matrix with the inverse of the singular values as
diagonal elements.

Choice of the Feature Map

The choice for the map φ is linked to the theory of RKHS [56] and in particular is linked to the
definition of random Fourier features [479]. It is defined as
r
2
y = φ(X) = σf cos(W X + b),
D

1
cos(W X + b) := √ (cos(W [1, :] · X + b1 ), . . . , cos(W [D, :] · X + bD ))T , (16.10)
D

where σf is a hyperparameter, W ∈ M(D × d, R) is the projection matrix whose rows are

sampled from a probability distribution µ on Rd , and b ∈ RD is a bias term whose components
are sampled independently and uniformly in the interval [0, 2π]. We remark that its Jacobian can
be computed analytically as

D
r !
∂z j 2 X
=− σf sin Wik xk + bk Wij (16.11)
∂xi D
k=1

for all i ∈ {1, . . . , d} and for all j ∈ {1, . . . , D}.

In order to guarantee the correctness of the procedure for evaluating the gradients, we have
to prove that the feature map is injective and nonsingular. In general, however, the feature
map (16.10) is not injective due to the periodicity of the cosine, but at least it is almost surely
nonsingular.
The feature map defined in (16.10) is not the only effective immersion that provides a kernel-
based extension of the ASs: possible choices consider random ReLU features and Fourier fea-
tures. Other approaches involve the use of deep neural networks [585] to learn the profile
g : Rr → Rd and an encoding function h : X → Rr in a nonlinear fashion, such that f ≈ g ◦ p.
A crucial step in the whole procedure is the tuning of the hyperparameters that appear in
the feature map, i.e., σf , bk , and the choice of the probability distribution µ. Possible im-
plementations include optimization methods that minimize the ridge approximation error with
 16.2. AS-Based Methods 333

cross-validation. For a more in depth description of the hyperparameter tuning procedure, the
reader can refer to [493].
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

16.2.3 Local Active Subspaces

In analogy with model order reduction, when the model at hand is complex enough that the
most common linear techniques do not produce accurate results, two directions can be under-
taken: usually a simpler model, which is well suited for linear methods, is reached through the
factorization via nonlinear maps or decomposition of the domain/parameter space.
Local active subspaces (LASs), proposed in [495], realize the latter possible direction, par-
titioning the domain with an AS-informed clustering algorithm and then performing ridge ap-
proximation in each subdomain. Effective supervised clustering algorithms are K-medoids and
hierarchical top-down clustering with the AS-informed distance
q
kxi − xj kΓ = (xi − xj )T W Γ2 W T (xi − xj ), (16.12)

where Γ is the diagonal matrix of the eigenvalues of the correlation matrix of the gradients of the
function of interest and xi and xj are a couple of input points.
In Figure 16.3 we consider a quartic function defined as f (x) = x41 − x42 in the domain
[0, 1]2 , where we highlight the direction of the global AS, which induces the cluster division in
Figure 16.4, depending on the chosen technique. As we can see, K-means produces uninformed
subsets, since it uses the standard Euclidean metric to find the centroids. By exploiting the
distance in (16.12), instead, the clusters are more aligned with the inactive space direction.

Function of interest: x14 x24 Global active subspace

1.0 1.00
0.9
0.75
0.8 0.6
0.50
0.3 0.25
0.6
0.0 0.00
f ( x)
x2

0.4 0.25
0.3
0.50
0.2 0.6
0.75
0.9
0.0 1.00
0.0 0.2 0.4 0.6 0.8 1.0 1.0 0.5 0.0 0.5 1.0
x1 Active variable W1Tx
Figure 16.3. Illustration of the global AS direction, highlighted in orange in the left panel. On
the right is the corresponding sufficient summary plot.

In general, applying the AS response surface design to many smaller subdomains is not
guaranteed to achieve better accuracy than employing the global AS. This is due mainly to
two reasons. When data is scarce, since the response surface design with AS is based on a
Monte Carlo approximation, the local surrogate models will be even more penalized. Secondly,
if the function of interest has a relatively high total variation, it is not guaranteed that the lo-
cal AS is a better minimizer of the local ridge approximation error with respect to the global
AS [495].
 334 Chapter 16. Reduction in Parameter Space

K-means K-medoids Hierarchical top-down

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

1.0
0.8
0.6
x2

0.4
0.2
0.0
0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00
x1 x1 x1

Figure 16.4. Comparison between K-means, K-medoids, and hierarchical top-down clustering
using the AS-induced distance metric for the quartic example of Figure 16.3.

16.2.4 Multifidelity Regression

Frequently, real-world applications are characterized by the need for ROM for computation-
ally intensive simulations and, at the same time, by the availability of low-cost models with
the same (multilevel) or different (multifidelity) structure. Multifidelity surrogate modeling is
often integrated into outer-loop applications such as optimization and forward and inverse un-
certainty quantification problems [443]. Many of these multifidelity approaches are based on
GPR in combination with the autoregressive information fusion scheme put forth by Kennedy
and O’Hagan [335].
Not every low-fidelity model can be included in a multifidelity paradigm: how to integrate the
information between models and the detection of the validity region are crucial steps. Moreover,
a model management strategy is often needed in real-world applications, providing theoretical
guarantees and establishing accuracy and/or convergence of outer-loop procedures: multisource
Bayesian optimization and experimental design are two possible paradigms.
We introduce a multifidelity regression [494, 497] scheme for low intrinsic dimensional-
ity objective functions that combines the AS property with the nonlinear autoregressive GPR
(NARGP) [445]. The main features of the methodology are recursiveness and autoregression,
nonlinear information fusion, and low intrinsic dimensionality of the objective function. The
method is purely data driven. The input is a hierarchy of data from increasing fidelity models:
from the highest p-fidelity to the lowest 1-fidelity:

N
Sq := {xiq , yqi }i=1
q
⊂ X ×R ⊂ Rd ×R, π1 (Sp ) ⊂ π1 (Sp−1 ) ⊂ · · · ⊂ π1 (S1 ) ⊂ X , (16.13)

N
where {xiq , yqi }i=1
q
represent the input-output pairs of the qth fidelity model and π1 : X × R →
X is the projection onto the d-dimensional input components. The final multifidelity model
integrates recursively all the previous ones. Since all q-fidelity models fq that we consider are
Gaussian processes, they are completely defined by the mean field (which is supposed to be the
0 constant field as before) and by their kernel

yq (x̄) −  ≈ GP(fq (x̄))|mq (x̄), Kq (x̄)) = GP(fq (x̄))|0, kq (θq )) , (16.14)

where  is a noise term, we employed the notation x̄ := (x, fq−1 (x)) ∈ Rd × R for q > 1 and
 16.2. AS-Based Methods 335

x̄ := x ∈ Rd for q = 1, and kq (θq ) is defined as

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

kq ((x, fq−1 (x)), (x0 , fq−1 (x0 )); θq ) = kqρ (x, x0 ; θqρ ) · kqf (fq−1 (x), fq−1 (x0 ); θqf )
+ kqδ (x, x0 ; θqδ ) ,

in which the hyperparameters to be tuned are represented by θq = (θqρ , θqf , θqδ ). So, apart from
the lowest fidelity, all the GP models have the input domain X ⊂ R augmented by the predic-
tions of the previous fidelity. The hyperparameters θq are optimized recursively with maximum-
likelihood estimation for each GP model GP(fq |0, k q (θq )):
1 1
arg max log p(yq |xq , θq ) ∝ − log |k q (θq )| − yqT (k q (θq ))−1 yq ; (16.15)
θq 2 2
this is why a hierarchical dataset is needed. After tuning, the predictive mean and variance for a
new input x∗ at fidelity q are integrated recursively with Monte Carlo over the distribution of the
previous fidelity Lfq−1 (x∗ ) :
Z
p(fq (x∗ )|x∗ , xq , yq ) = p(fq (x∗ )|f∗q−1 , x∗ , xq , yq )p(f∗q−1 )dz
ZX
= p(fq (x∗ )|s, x∗ , xq , yq )dLfq−1 (x∗ ) (s)
fq−1 (X )
M
X
≈ p(fq (x∗ )|si , x∗ , xq , yq ).
i=1

We now introduce how ASs are involved in the design of a two-fidelity multifidelity model.
The incorporation of prior knowledge, biasing the learning process, is inevitable for the success
of learning algorithms. In our methodology this is realized in two ways: with the kernel of the
Gaussian process (GP; type of spatial correlation, degree of smoothness of the model) and with
the low-fidelity intrinsic dimensionality assumption (presence of a dominant AS).
In our paradigm, the low-fidelity model is represented by the response surface built with the
AS (computed from the high-fidelity data) on the whole domain X , and the high-fidelity model
is the GPR of the input-output data available (usually in small quantity). Due to the low intrinsic
dimensionality assumption, we expect the multifidelity model to perform better than the high-
fidelity one, thanks to the nonlinear integration of the data with the information of the presence
of an AS.
We provide a simple example. The function of interest is a hyperbolic paraboloid

f : [−1, 1]2 → R, f (x1 , x2 ) = x21 − x22 . (16.16)

The high-fidelity inputs are sampled uniformly: we sample and evaluate 50 high-fidelity and
200 low-fidelity input-output pairs. As previously said, the AS used to build the low-fidelity
model is obtained after computing the analytic gradients for the 50 high-fidelity input-output
pairs (generally the gradients could also be approximated directly from the input-output pairs).
The low-fidelity samples are cheap, and the cost of their evaluation is not comparable with the
cost of a high-fidelity output evaluation. The low-fidelity response surface is represented in
Figure 16.5.
Usually the high-fidelity model is not accurate enough due to the relatively scarce data avail-
able, while the multifidelity model, by exploiting the inductive bias of an intrinsic low dimen-
sionality, performs better. The outputs predicted on the 1000 test inputs for the three models
(low-fidelity, high-fidelity, multifidelity) are qualitatively compared in Figure 16.5. In the high-
fidelity model, the actual input-output pairs used for training are depicted in white.
 336 Chapter 16. Reduction in Parameter Space

GPR over the AS Constant extension

1.5
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Mean
Data
1.0
Confidence

0.5

f (x)
0.0

0.5
High-fidelity data
1.0 0.5 0.0 0.5
Active variable W1Tx

High-fidelity GPR Multi -fidelity prediction

Figure 16.5. Illustrative scheme of a multifidelity regression procedure that employs AS as a low-
fidelity model. The function to be approximated is the hyperbolic paraboloid from (16.16). Starting from
10 high-fidelity data points (depicted in blue and in white) we construct as a low-fidelity model a response
surface which is constant along the inactive subspace.

The l2 error of the multifidelity model against the 1000 test samples is 0.00101, while the l2
error for the high-fidelity model is 0.00125. Even if in this test case (employed only to clarify
the procedure) the gain is only marginal, in practice there may be noticeable improvements,
especially when the dimension of the parameter space is higher and the data is still scarce for the
complexity of the problem at hand [494].

16.2.5 Gradient-Free Extensions

In this subsection we briefly present some extensions of ASs. In order to overcome the need
to compute or approximate gradients, gradient-free methods have been devised, especially in a
Bayesian context.
For example, in [584] they train a GP with built-in dimensionality reduction by learning
the projection matrix A in (16.2) as the result of tuning the GP hyperparameters. The process
comprises two steps: the projection of the data into the AS and the actual regression over the
lower-dimensional space. The GP model is

y | θ, W ∼ N (0, K(XW, XW; θ)), (16.17)

where X is the training data matrix. It is clear how it simultaneously learns the link function and
the AS projection matrix. With a similar approach, in [215] they use a fully Bayesian approach
in order to obtain full posterior probability distributions for both the hyperparameters θ and the
projection matrix W in (16.17). Thus, for a given new unseen input, the predictive uncertainty
obtained from the queried surrogate model accounts for all the uncertain model parameters. The
authors also impose that the projection matrix is an element of the Stiefel manifold.
Another approach is presented in [626], where the authors propose a closed-form expression
for the uncentered covariance matrix defining the AS for a GP for the popular covariance kernels,
a technique usually executed via Monte Carlo finite differencing. This allows one to create
acquisition functions enabling evaluation-efficient estimation of such subspaces via sequential
 16.3. Other Nonlinear Techniques 337

design, with fewer black-box evaluations. A gradient-free approach exploiting Morris screening
elementary effects can be found in [372].
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

16.3 Other Nonlinear Techniques

In this section we briefly review some recent nonlinear techniques for parameter space reduction
which do not fall under the ridge approximation with ASs framework.

16.3.1 Nonlinear Level-Set Learning

The nonlinear level-set learning (NLL) method [643] for dimensionality reduction in high-dimen-
sional function approximation is one of the most recent techniques to go beyond the linear co-
ordinate transformation of ASs. The method seeks a bijective nonlinear transformation g to
capture the geometry of level sets and parametrize them in a low-dimensional space. To this end,
in [643] they employ reversible neural networks (RevNets) [113] to learn the transformation g.
The designed loss function uses samples of the gradients of the target function to encourage the
transformed function to be sensitive to only a few active coordinates. In [248] they introduce a
variation of this function in the context of PDEs.
To construct the RevNet they use the following architecture [262], which is reversible by
definition:
T

un+1 = un + hKn,1 σ(Kn,1 vn + bn,1 ) ,
T (16.18)
vn+1 = vn − hKn,2 σ(Kn,2 un + bn,2 ) for n = 0, 1, . . . , N − 1,

where u and v are partitions of the states, h is a scalar time step, the matrices K contain the
weights, b represent the biases, and σ is the activation function. We remark that the original
coordinates and the transformed ones are split in two in u and v. In Figure 16.6 we depict a
two-dimensional example where we can actually plot the transformed input parameter space.
We implemented this technique in the ATHENA [496] Python package.

Figure 16.6. NLL application two-dimensional example. On the left we have the target func-
tion against the first transformed coordinate, in the middle the loss function decay, and on the right the
transformed parameter space.

16.3.2 Active Manifolds

Another approach is to compute active manifolds [77]. This method requires further hypotheses
on the target function of interest f , which has to be C 1 . The advantage is that it is able to identify
a one-dimensional curve in the domain, the so-called active manifold, which represents almost
 338 Chapter 16. Reduction in Parameter Space

all the values f can attain. This allows for more accurate regressions over the low-dimensional
manifold with respect to the AS method. The drawback is the computational overhead. An
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

application to multidisciplinary analysis and optimization can be found in [68].

16.4 Applications
In this section we review a wide range of possible applications of the techniques presented above.
We provide a comparison in the construction of response surfaces, we introduce the activity
scores for sensitivity analysis, and we show how to integrate ASs into optimization strategies.
Finally, we present some insights into the treatment of inverse problems and coupling with other
reduced order methods.

16.4.1 Response Surface Design

The term response surface refers to the general procedure of finding the values of a model func-
tion f for new inputs without directly computing it but exploiting regression or interpolation
from a training set {xi , f (xi )}. The presence of an AS can be exploited to design response
surfaces upon a reduced parameter space, with a decrease in the computational cost.
Regarding the error analysis, an upper bound on the ridge approximation error can be derived
[134] when applying the Monte Carlo method to approximate both the AS from the correlation
matrix (16.3) and the optimal profile g, which is a conditioned random variable:

(g ◦ Pr )(X) = Eµ f (X)|σ(W1 W1T X) ◦ Pr

 

N
1 X
≈ f (W1 W1T X + (Id − W1 W1T )Yi ) =: ĝ(W1 W1T X),
N i=1

where Y1 , . . . , YN are independent and identically distributed samples of Y, the absolutely

continuous random variable with the same probability density function ρ of the input random
variable X.

16.4.2 Regression
We show some examples of response surface design with GPR [622]: we compare four response
surfaces built by exploiting an AS, a KAS, an LAS, and the NLL method. The chosen test case
models the spread of the Ebola virus:9 the output of interest is the basic reproduction number
R0 of the susceptible-exposed-infectious-removed (SEIR) model, described in Table 16.1, which
reads
β1 + β2 ρω1 γ1 + βγ23 ψ
R0 = , (16.19)
γ1 + ψ

with parameters distributed uniformly in Ω ⊂ R8 . The parameter space Ω is defined by the lower
and upper bounds summarized in Table 16.1.
We collected 200 input-output gradient triples as the training set. They are employed to find
the AS, to tune the feature map for KAS, to cluster the parameter space with hierarchical top-
down clustering for LAS, and to train the recursive neural network (RNN) for NLL. The spectral

9 The dataset was taken from https://github.com/paulcon/as-data-sets.

 16.4. Applications 339

Table 16.1. Parameter ranges for the Ebola model. Data taken from [167].
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

β1 β2 β3 ρ1 γ1 γ2 ω ψ
Lower bound 0.1 0.1 0.05 0.41 0.0276 0.081 0.25 0.0833
Upper bound 0.4 0.4 0.2 1 0.1702 0.21 0.5 0.7

measure chosen for KAS is a Beta distribution whose two shape parameters are tuned with log-
grid search and cross-validation [493]. The clustering algorithm employed for LAS is hierarchi-
cal top-down clustering, which applies K-medoids with the AS-informed metrics (16.12) at each
level of the clustering refinements. The RNN of the NLL method has seven layers and is trained
for 1000 epochs with the Adam stochastic optimization method [344].
Figure 16.7 shows the response surfaces for the AS, KAS, and NLL methods. The local
response surfaces for each cluster of LAS are reported in Figure 16.8.

3.0
Active Subspaces Kernel-based Active Subspaces Non-linear Level-set Learning
2.5 2.5
2.5
2.0 2.0 2.0
mean
1.5
R0

1.5 1.5 68% confidence

train data
1.0 1.0 1.0
0.5 0.5
0.5
1.0 0.5 0.0 0.5 1.0 0.00 0.01 0.02 2 1 0 1
reduced parameter space reduced parameter space reduced parameter space

Figure 16.7. On the y-axis: basic reproduction number R0 of the SEIR model for the Ebola
epidemic. Left: GPR on the one-dimensional AS. Middle: GPR on the one-dimensional KAS. Right: GPR
on the one-dimensional active variable of the NLL method. In the plots are also represented the 200 training
points used to find the reduced parameter space and build the GPRs.

LAS cluster 0 LAS cluster 1 2.00

LAS cluster 2
2.5
1.2 1.75
2.0 1.0 1.50
R0

1.25
1.5 0.8 1.00
0.8 0.6 0.4 0.2 0.0 0.2 0.0 0.2 0.4 0.6 0.75 0.50 0.25 0.00 0.25 mean
LAS cluster 3 LAS cluster 4 LAS cluster 5 68% confidence
2.0 train data
1.50 3.0
1.8
1.25
1.6 2.5
1.00
R0

1.4
2.0
0.75 1.2
0.50 1.0 1.5
0.0 0.5 1.0 0.6 0.4 0.2 0.0 0.2 1.0 0.8 0.6 0.4 0.2
reduced parameter space reduced parameter space reduced parameter space

Figure 16.8. Each one of the six plots represents the GPR on the reduced space of one out of six
partitions of the parameter space found with LAS. On the y-axis is the basic reproduction number R0 of
the SEIR model for the Ebola epidemic. On the x-axis is the reduced one-dimensional AS of the relative
cluster. In the plots are also represented the 200 training points scattered among all six clusters and used
to find the partition of the parameter space with hierarchical top-down clustering and build the GPRs.
 340 Chapter 16. Reduction in Parameter Space

Each response surface is tested on an independent set of 500 samples. The R2 scores are
shown in Table 16.2. The KAS, LAS, and NLL methods present an evident gain with respect to
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

the use of the AS to design a global response surface. Usually the overhead of the KAS, LAS,
and NLL training procedures over the standard AS method is negligible with respect to the cost
for obtaining the input-output datasets.

Table 16.2. Comparison of the R2 scores of the response surfaces for the Ebola model built with
the AS, KAS, LAS, and NLL methods.

AS KAS LAS NLL

R2 0.8414 0.875 0.9447 0.991

16.4.3 Sensitivity Analysis

In [135] the authors propose two sensitivity metrics derived from the eigenpairs of the covariance
matrix (16.3). The first one uses the elements of the first eigenvector as a metric to rank the
parameters’ importance. The second sensitivity metric uses the squared eigenvector components
scaled by the eigenvalues, the so-called activity score.
The activity score αi (n) of the ith parameter depends on the dimension n of the AS, which
is defined as
Xn
2
αi = αi (n) = λj wij , i = 1, . . . , m, (16.20)
j=1

where m is the total number of parameters and wij , λj are the elements of the eigenvector matrix
and the eigenvalues, respectively.
Conversely to the first eigenvector, which provides information through the sign of its com-
ponents, the activity scores remove this information by squaring the components. This results
in a much more comparable metric with respect to other widespread sensitivity metrics such as
Sobol’s total sensitivity indices and derivative-based global sensitivity measures.

16.4.4 Optimization
A novel application of ASs for optimization tasks is the coupling with the classical genetic algo-
rithm (GA), called the active subspace-based genetic algorithm or ASGA [161]. The main idea
is to perform the typical reproduction and mutation steps on the lower-dimensional AS of the
current generation. In Figure 16.9 is depicted the scheme of the ASGA procedure, where the
yellow boxes denote the parts performed for ASGA which are not done in the classical GA. The
AS back-mapping is performed as described in Subsection 16.2.1. Of course, given a sample in
the AS, in general, there are many points in the full space which are mapped onto the same point
on the reduced space. The new algorithm allows for the selection of more than one point in the
full space in order to explore different parts of the domain simultaneously.
ASGA has been tested on classical high-dimensional benchmark functions, exhibiting great
performance in terms of both better convergence to the actual minimum f (x∗ ) for a given com-
putational budget and better spatial convergence to the location of x∗ . The method has also been
tested for shape optimization in a naval engineering context [157].
The coupling of AS with optimization algorithms is not limited to GA. We cite [226, 340]
for application of ASs to Bayesian optimization for aerostructural design of an aircraft wing for
fuel-burn minimization and for materials design, respectively.
 16.4. Applications 341

3 individuals
First generation
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

with N genes

Accordingly to
Selection best fitness

From N genes
AS Projection to M genes

Crossover
Reproduction

Random
Mutation mutation

From M genes
AS back mapping to N genes

Offsprings +
New generation best individuals

Figure 16.9. ASGA scheme. The main steps of the classical GA are depicted from top to bottom.
Projections onto and from a lower-dimension AS are highlighted with yellow boxes, which are specific to
ASGA.

16.4.5 Inverse Problems

ASs can be applied to increase the efficiency of Monte Carlo methods when sampling from the
posterior distribution of the inputs in the Bayesian framework [135, 636].
Let us consider the following inverse problem: supposing that the inputs x are distributed
according to a multivariate Gaussian distribution in X ⊂ Rm , we can compute the posterior
distribution of x with knowledge of a dataset (x, y), where y are the targets obtained after the
simulation m(x), with the addition of a noise term:

y ≈ f (x) = m(x) + , m : X ⊂ Rm → R, (16.21)

x ∼ N (0, Id ),  ∼ N (0, σId ), f ∼ N (m(x), σId ). (16.22)

Sampling from the posterior distribution of the latent variables x with Monte Carlo is costly
due to the evaluations of the model m:

plikelihood (f |x)pprior (x)

pposterior (x|f ) = ∝ plikelihood (f |x)pprior (x). (16.23)
pmarginal (f )

In order to avoid this we can approximate the model with a surrogate obtained by employing an
 342 Chapter 16. Reduction in Parameter Space

AS g : Xactive ⊂ Rr → R such that g(s) ≈ m(x) = m(W1 s + W2 t):

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

plikelihood (g|s)pprior (s)

pposterior (s|g) = ∝ plikelihood (g|s)pprior (s), (16.24)
pmarginal (g)

pprior (t) = pposterior (t). (16.25)

It is important to observe that in this way the inactive variable z has a multivariate Gaussian
distribution and is independent of the random variable f .
ASs are also employed to design a surrogate model of the parameters-to-observable map of
a PDE model in the context of optimal experimental design [624]. In a Bayesian framework, the
expected information gain is maximized in order to find the best sensor locations. In this case, the
function f of (16.21) is the output of a complex PDE model with an intrinsic low dimensionality;
thus reduction in the parameter and output spaces is beneficial for decreasing the computational
cost of the optimization.

16.4.6 Coupling with Other Reduced Order Methods

The AS method has been also coupled with other reduced order methods, both intrusive and
nonintrusive.
In the first category falls the proper orthogonal decomposition (POD)-Galerkin approach,
which was used in [568] combined with AS for the efficient numerical estimation of the pres-
sure drop in a set of deformed carotid arteries. In order to simulate a wide range of possible
occlusions after the bifurcation of the carotid, the authors used the radial basis function (RBF)
interpolation method to parametrize a reference domain. After the reduction of the parameter
space, the ROMs were created over the AS to further reduce the required computational effort
and enhance computational efficiency.
For nonintrusive approaches such as POD with interpolation, AS was proposed as a way to
enhance the modal coefficient regression accuracy in cases where the solution snapshot database
is not sufficiently rich in terms of sampled points [160]. This was proven both for structural
and computational fluid dynamics (CFD) problems with parametrized geometries. A possible
coupling with dynamic mode decomposition (DMD) was done in [573]. Recently, in [575], the
authors used the multifidelity approach introduced in Subsection 16.2.4 to improve the accuracy
of POD with interpolation with respect to using a single-fidelity GPR in the context of structural
optimization of modern cruise ships.

16.5 Conclusions
In this chapter we presented a general overview of different parameter space reduction tech-
niques, both linear and nonlinear, to fight the curse of dimensionality, together with a wide range
of possible applications.
In the first part we focused on AS-based methods in a ridge approximation framework. We
started with classical AS, and then we introduced novel extensions such as KAS and LAS. Both
techniques try to move from linear to nonlinear manifold approximation. We then presented a
multifidelity framework to enhance GPR through a low intrinsic dimensionality bias coming from
parameter space reduction. To complete the overview we briefly mentioned some gradient-free
approaches and nonlinear methods not based on AS, such as NLL and active manifolds.
In the second part we moved to the application side, showing how it is possible to use the
techniques presented in the first part to solve common numerical analysis problems. We provided
 16.5. Conclusions 343

a comparison in terms of the regression accuracy of response surfaces built over the reduced pa-
rameter space identified by AS, KAS, LAS, and NLL. We used a common benchmark dataset
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

created to model the spread of Ebola. We then briefly touched on how to exploit AS for sen-
sitivity analysis with activity scores, and how to enhance genetic algorithms by performing the
reproduction and mutation steps over the AS of the current population. Finally, we showed
with a simple example how to treat inverse problems and commented on possible coupling of
AS with other reduced order methods. Other interesting applications, not touched on in this
work, are the reduction of highly parametrized artificial neural networks (ANNs); in this direc-
tion, in [394], the authors propose an AS approach to decreasing the memory consumption of
modern ANNs for image recognition and classification. For a deeper presentation, please refer
to Chapter 20.
More work has to be done to better integrate parameter space reduction with other ROMs,
especially for highly parametrized systems and outer-loop applications such as shape optimiza-
tion. Future research lines should also focus on the development of reliable nonlinear methods,
possibly incorporating physical constraints, without losing the interpretability characterizing the
linear methods.
 Chapter 17
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Geometrical
Parametrization and
Morphing Techniques
with Applications

Andrea Mola, Nicola Demo, Marco Tezzele, Gianluigi Rozza

17.1 Introduction
As repeatedly pointed out, the present book focuses for the most part on parametrized partial
differential equation (PDE) problems. In many cases of interest, the input parameters of such
PDE problems happen to be associated with the shape of the computational domain, or of one
of its portions. For instance, virtual prototyping applications very often investigate the influence
that the shape of a given craft or artifact has on its predicted performance. In some case, through
repetitive resolution of PDE systems, such applications go as far as identifying the optimal shape
of the craft or artifact under study, or quantifying the sensitivity of its performance to variations
of its geometry. In these last cases, the virtual prototyping application pipelines will include op-
timization or uncertainty quantification tools. Naturally, such mathematical algorithms can only
be fed with numbers. Thus, the main purpose of shape parametrization is to map any geometric
variation of the object under study to variations of numbers—the parameters—which can then
be processed by any other mathematical algorithm used in the virtual prototyping pipeline.
To allow for completely automated virtual prototyping and optimization pipelines, it is very
important that the shape parametrization algorithm be able to automatically obtain the deformed
domain and computational grid once the shape parameter values are prescribed. This constraint
not only ensures that repeated solution of parametrized PDE problems is possible when shape
parameters are included among the inputs but also significantly limits the time required for mesh-
ing, which in such a framework typically represents the most time-consuming activity in terms
of human operator hours.
A further constraint for shape parametrization algorithms arises from their application in the
context of model order reduction. As discussed with greater detail in several chapters of the
present book, many modal decomposition strategies used in reduced order models (ROMs) are
based on a snapshot matrix which is obtained under the assumption that the number of degrees
of freedom and the order of the PDE problem remain unaltered as input parameter values are
modified. For this reason, the shape parametrization algorithm must not alter the topology of
the computational grid used to resolve the PDE. Thus, a simple approach consisting of creating
a new mesh after the shape of the domain has been altered is not suited for ROM applications.

345
 346 Chapter 17. Geometrical Parametrization and Morphing Techniques with Applications

A more comprehensive strategy which deforms the computational mesh along with the domain
geometry is needed.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

For this reason, in the present chapter we first analyze algorithms that are suited for deforming
codimension-one subsets of R2 or R3 , which are used to define the geometry of the computa-
tional domain by describing their boundary representation. In two dimensions, codimension-one
subsets are typically described by the points of piecewise linear curve representations or by
the control points and nodes of nonuniform rational basis spline (NURBS) curves included in
computer aided design (CAD) files. In three dimensions the most common representations of
codimension-one subsets are given by the points and connectivity of surface triangulations, or
by the control points and nodes of NURBS saved into CAD files. After this, we will focus on
the deformation of codimension-zero computational domains in R2 or R3 , which are typically
used to define computational grids. In both two and three dimensions, these domains are typi-
cally described by the nodes and the connectivity of polygonal or polyhedral tessellations. It is
important to stress that an additional advantage of deforming both the domain geometry and the
computational grid in a consistent way is that it allows for a strong interaction between design
and simulation teams. In fact, the shape modifications validated by PDE-based simulations will
be readily available for production with no further processing.
In this chapter, we first consider parametrization algorithms which attempt to come up with
a general law to deform at once the whole two- or three-dimensional space in which the mor-
phed object is embedded. Such general purpose algorithms typically operate by extending or
interpolating the displacement of a relatively small set of control points to all the points of the
two- or three-dimensional space. Thus, the displacements of such control points become the user-
prescribed shape parameters. The position of each point of the objects of interest is then displaced
by interpolating the control point displacement to obtain the modified shape. This paradigm is
common to free-form deformation (FFD), originally proposed in [533], and to the application of
radial basis function (RBF) interpolation [82] and inverse distance weighting (IDW) interpola-
tion [537] to shape parametrization. A considerable advantage of these algorithms is that they
can indifferently operate on the points of domain geometries or on the nodes of computational
grids. For this reason they have been easily applied to CAD geometries, as for instance in [574],
and to computational meshes, as in [157]. On the other hand, one of the possible drawbacks of
these tools is that the shape parameters identified are not necessarily related to the specific phys-
ical behavior of the object under study. For instance, a simple scaling parameter can affect the
aspect ratio of a ship hull and have a clear effect on its hydrodynamic performance. On the other
hand, the effect of the prescribed displacement of one of many shape control points on the hull
performance is more difficult to establish. Ideally, in several engineering applications the shape
parameter should be readily associated with qualitative variations in the performance of the de-
signed object. Moreover, in some cases general purpose algorithms might violate engineering
constraints on the shapes to be deformed. This is for instance the case for aircraft and ship pro-
pellers, which are objects with a highly engineered shape and strict design rules [101]. Thus, we
will also discuss cases in which the shape parametrization strategy is tailored to certain bodies to
alter their shape, preserving the geometric characteristics they have been designed to possess.
The content of this chapter is organized as follows. Section 17.2 introduces and discusses
the mathematical background of different algorithms for shape parametrization. After general
purpose methods (FFD, RBF, IDW) are presented, we will also describe examples of different
object-specific deformation strategies. In Section 17.3, we discuss the application of all these
strategies to the deformation of both domain geometries and computational meshes. In this
context, we will also present several applications of the proposed shape parametrization method-
ologies to industrial and real-life problems. Finally, Section 17.4 provides some concluding
remarks, as well as some interesting perspectives on the field of shape parametrization in the
next few years.
 17.2. Parametrization Techniques 347

17.2 Parametrization Techniques

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

In this section, we present a diverse range of shape parametrization tools. We divide such de-
formation techniques into two main categories based on their objective: general purpose shape
parametrizations and object-specific ones.

17.2.1 General Purpose Shape Parametrization Algorithms

In the first category we group the morphing techniques which come with a general law to deform
the whole three-dimensional space at once. They usually make use of a set of control points to
move in order to propagate the desired shape deformation.
Let us introduce a physical domain Ω ⊂ R3 and a general parametric deformation map
M(·, µ) : R3 → R3 acting on Ω, where µ ∈ P represents the parameter vector characterizing
the morphing.
In the following, we present the specific definition of M for three nonaffine techniques: FFD,
RBF interpolation, and IDW interpolation. For affine maps used in the model order reduction
context, refer to [281], while for a more general overview we suggest [503]. For an efficient
implementation of the methods presented below we recommend the open-source Python package
called PyGeM [570], which also integrates easily with industrial CAD files.

FFD
FFD is a well-known morphing technique for both local and global deformations of a given
object. Originally presented in [533], in the last decades it has gained popularity in the context
of parametric PDEs and model order reduction for shape optimization studies (see [521, 507, 35,
519, 154]), optimal flow control problems [509], and sailing boats [379], in naval engineering
[574, 155, 168, 569, 572, 156], CFD applications [160], and isogeometric analysis [518, 213],
to cite a few. Since the deformation induced by the FFD does not depend on the topology of the
object to be morphed, it is very versatile and nonintrusive, especially for complex geometries or
in industrial contexts [508, 517, 571].
FFD deforms the physical domain Ω by propagating the displacements of a lattice of points,
called FFD control points, placed in Ω. The map M is the composition of three different geo-
metric deformations, i.e.,

M(·, µ) := ψ −1 ◦ T̂ ◦ ψ(·, µ), µ ∈ P, (17.1)

where the maps ψ and T̂ are depicted on a simple bidimensional example in Figure 17.1. The
function ψ represents the affine mapping from the physical domain Ω to its reference configura-
tion Ω̂. The FFD control points P are moved, and the morphing is propagated to all the points
within the lattice, usually by using Bernstein polynomials, but other choices are also possible.
The displacements of such control points along the cardinal directions define the parameter vec-
tor µ. The map T̂ (·, µ) : Ω̂ → Ω̂(µ) for a general three-dimensional case with a lattice of
L × M × N control points is then defined as
L X
X M X
N
T̂ (x, µ) = blmn (x)Plmn (µlmn ) ∀x ∈ Ω̂, µ ∈ P, (17.2)
l=1 m=1 n=1

where blmn stands for the Bernstein polynomials of degree l, m, n in each direction, and
Plmn (µlmn ) := Plmn + µlmn , with Plmn indicating the coordinates of the FFD control point
with indices l, m, n. By exploiting the good properties of such polynomials it is also possible to
impose continuity constraints by adding control points. This is particularly useful at the interface
 348 Chapter 17. Geometrical Parametrization and Morphing Techniques with Applications
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 17.1. Sketch of the three deformation maps which compose the FFD morphing.

between the unmorphed part of the geometry and the lattice of control points, for which physical
constraints usually have to be satisfied.
In the classical offline-online framework, we emphasize that the terms blmn (x) can be com-
puted in the offline phase, while in the online phase we just perform linear combinations of
precomputed quantities depending on x- and µ-dependent control point locations. Moreover,
since ψ and ψ −1 are affine maps, the two phases remain separated.

RBF Interpolation

By RBF [82] we mean any smooth real-valued function ϕ̄ : Rd → R whose value depends only
on the distance between the input and some fixed point. This means that there exists ϕ : R+ → R
such that ϕ̄(x) = ϕ(kx − ck), where k · k stands for the Euclidean distance and c is some fixed
point.
To illustrate the many possible choices among all the RBFs, in Figure 17.2 we consider
a simple univariate example where we approximate the function f (x) = 0.5 sin(4πx) +
exp(x cos(2πx)) using six different types of RBFs:

multiquadratic biharmonic splines (also called multiquadric) [523], defined as ϕ(r) =

• p
1 + (rε)2 ;

• inverted multiquadratic biharmonic splines (abbreviated as inverse in Figure 17.2) [82],

defined as ϕ(r) = √ 1 2 ;
1+(rε)

2
• Gaussian [82], defined as ϕ(r) = e−(rε) ;

• cubic and quintic, defined as ϕ(r) = r3 and ϕ(r) = r5 , respectively; and

• thin plate [176], defined as ϕ(r) = r2 ln(r).

 17.2. Parametrization Techniques 349

Multiquadric Exact Centers Inverse Exact Centers

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

2
y

Gaussian Exact Centers Cubic Exact Centers

2
y

Quintic Exact Centers Thin plate Exact Centers

2
y

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
x x

Figure 17.2. Interpolations corresponding to six different RBFs. The exact function is depicted
with a red dashed line. The parameter ε is equal to the inverse of the average distance between the sampling
points.

0.2 Multiquadric Gaussian Quintic Centers

Inverse Cubic Thin plate
0.1
ypred yexact

0.0

0.1

0.2
0.0 0.2 0.4 0.6 0.8 1.0
x

Figure 17.3. Interpolation error for different types of RBFs. Refer to Figure 17.2 for the actual
interpolations.

We remark that the list does not cover all the possible RBFs. Figure 17.3 reports the interpolation
error over the entire domain for the different RBFs described above.
We follow [407, 390] for a brief presentation of the RBF interpolation method in order to
define M(·, µ) for this technique. Let {xCi }N i=1 be the control points placed over the surface
C

of the object we want to parametrize. We can induce a deformation by moving these control
points and interpolating them with a new surface. The parameter vector µ is composed by the
displacements of the RBF control points. Let ϕ be the selected RBF and {γi }N i=1 be the weights
C

NC
associated with the RBF centered in {xCi }i=1 , respectively. We can define the deformation map
 350 Chapter 17. Geometrical Parametrization and Morphing Techniques with Applications

M for the RBF interpolation technique as

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

NC
X
M(x, µ) := s(x, µ) + γi (µ)ϕ(kx − xCi k) ∀x ∈ Ω, µ ∈ P, (17.3)
i=1

where s(·, µ) is a polynomial term of degree one, which means s(x, µ) = S(µ)x + c(µ). The
total unknowns are d + d2 for the polynomial term and d × NC for the weights, where d = 2 or 3,
depending on the dimension of Ω. Regarding the γi , let {yCi (µ)}N i=1 be the new control points
C

obtained from xCi after applying the displacements defined by µ. The interpolation constraints
we impose are
M(xCi , µ) = yCi (µ) ∀i ∈ [1, . . . , NC ]. (17.4)
For the remaining unknowns due to the polynomial term, we impose the conservation of the total
force and momentum (see [82]) as
NC NC
(j)
X X
γi (µ) = 0, γi (µ)xCi = 0 ∀j ∈ [1, . . . , d], (17.5)
i=1 i=1

where x(j) denotes the jth component of x.

We emphasize that the RBF contributions ϕ(kx − xCi k) can be precomputed in the offline
phase, while for a given µ in the online phase we just need to solve the d × NC + d + d2 linear
system and perform the matrix-vector product in (17.3).
For an application of the RBF interpolation technique used to morph a carotid artery, we
mention [568]; moreover, Figure 17.9 later in this chapter illustrates two possible deformations.
For aeronautics fluid-structure interaction (FSI) problems, see [202]. There are also extensions of
the method such as the rescaled localized RBF interpolation method for nonconforming meshes
[163] or the combination with nonintrusive model order reduction [573, 220, 161].

IDW Interpolation
IDW is an interpolation method originally proposed in [537]. Let {(xCi , f (xCi )}N i=1 , where
C

xCi ∈ R , f (xCi ) ∈ R for i = 1, . . . , NC , be the set of input-output pairs obtained by evaluating

d

f : Rd → R. The IDW interpolant of f is defined as

NC
X
IIDW [f ](x) = wi (x)f (xCi ), x ∈ Rd , (17.6)
i=1

such that IIDW [f ](xCi ) = f (xCi ) for i = 1, . . . , NC . The functions wi : Rd → R are defined
as
kx−xCi k−p
(
if x 6= xCi ,
(17.7)
PNC
wi (x) = j=1 kx−xCj k
−p

1 if x = xCi for i = 1, . . . , NC ,
where the power −p controls the rate of influence of the ith input point and k · k is the Euclidean
norm. Such functions require one to compute all the distances between the point to interpolate
x and the input data {xCi }Ni=1 . Assuming the weight function becomes negligible after a certain
C

distance , we can rewrite (17.7) as

d(x,xCi )−p

 PNC d(x,x

−p
if d(x, xCi ) ∈ [0, ),
Cj )
(17.8)
j=1
wi (x) =
 1
 if d(x, xCi ) = 0,
0 if d(x, xCi ) ∈ [, +∞] for i = 1, . . . , NC ,
 17.2. Parametrization Techniques 351

where for the sake of notation we define the distance function d : Rd × Rd → R+ as d(x, y) =
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

kx − yk.
This interpolation can be adopted as an efficient tool for shape parametrization problems.
Among all the contributions in the literature, we suggest [623] for a parametric computational
grid by means of IDW, [202] for application of IDW to FSI problems, and [31] for an example
of domain parametrization in a ROM framework.
The main idea for shape parametrization using interpolation techniques is to extend the de-
formation we impose to a limited set of control points by interpolating their displacements. We
define this parameter-dependent displacement at point xCi as fµ (xCi ) ≡ f (xCi , µ) ∈ R for
i = 1 . . . , NC . Interpolating the displacements with IDW, we obtain
NC
X
IIDW [fµ ](x) = wi (x)f (xCi , µ), x ∈ Rd , µ ∈ P. (17.9)
i=1

We remark that we have considered until now scalar displacement, but the IDW extension to
vector functions is trivial: assuming all components are independent, we simply apply IDW to
each component. Thus, considering a d-dimensional vector as the displacements, we can derive
the deformation map M(·, µ) : Rd → Rd from (17.9):
NC
X
M(·, µ) := wi (·)f (xCi , µ), x ∈ Rd , µ ∈ P. (17.10)
i=1

Equation (17.10) shows the decoupling between the spatially dependent terms (wi (·)) and the
parameter-dependent ones (f (xCi , µ)). This implies that for any x, we can precompute (and
store) the weight functions, requiring us to compute only the parameter-dependent term for any
new evaluation of M(·, µ) in the online phase.

17.2.2 Object-Specific Shape Parametrization Algorithms: Some

Examples
The general purpose parametrization techniques just described are ideal in cases in which the
artifacts or objects to be deformed are not bound to preserve local or global geometric properties
throughout the deformation process. In the engineering field, several objects or their components
are often designed to carry out a distinct task or to deliver a specific performance. Hence, the
parameter-to-shape relationship must be built so that all the deformations considered result in
shapes which would not impair the possibility of the component carrying out its task or providing
the required performance.
There are several examples of this situation, which can also be used to illustrate different
possible constraints the geometry can be subjected to. For instance, the aerodynamic design of
an aircraft wing tip fuel tank or external drop fuel tank will probably require any shape considered
to be characterized by a certain volume. A similar situation is encountered in the design of the
ballast-filled bulb at the bottom of the keel of a sailing yacht. To provide the adequate restoring
moment, the bulb weight is fixed, and if the ballast material is not changed this results in a
volume constraint.
RBF, IDW, and FFD methodologies in the form described in the previous subsection are not
suited to automatically producing shapes that have a fixed volume. Of course, it is in principle
possible to complement such algorithms with a supplementary parameter which can be used to
scale each RBF or FFD in order to obtain the prescribed volume. This parameter, however, might
in some cases turn out to be not completely independent with respect to the original one, which
must be accounted for in the overall design process.
 352 Chapter 17. Geometrical Parametrization and Morphing Techniques with Applications

In addition, there are several situations in which general purpose algorithms cannot be em-
ployed without a more invasive and complex adaptation. As an example, consider the design
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

of ship or yacht hulls. Typically, the lateral walls of a hull are constrained to be vertical or at
least straight because this results in easier management and arrangement of internal passengers
or cargo areas. Compared to a global constraint, such as the aforementioned fixed-volume one,
this local constraint is more difficult to fulfill. Both RBF and FFD would in fact require signifi-
cant modifications to be effective for generating parametrized hulls with vertical or straight side
walls. An even more important case of shapes with local constraints limiting the effectiveness of
general purpose shape parametrization algorithms is that of aircraft wings and propeller or turbo
machinery blades. These objects were originally designed to have airfoil sections with standard
and tabulated shapes with well-documented fluid dynamics performance. Because remarkably
efficient estimates of the overall wing or blade performance can be obtained from that of its air-
foil section, it is very important that any parametrization algorithm applied on these artifacts not
lead to sections with nontabulated shape or performance. It is clear that in this case RBF and
FFD, as they have been described, also cannot generate shapes with this property.
In light of all the problems illustrated by these examples, in this subsection we will discuss
shape parametrization algorithms adapted to specific design problems. In particular, making use
of some of the aforementioned design cases (i.e., planing hull yachts and naval propeller blades),
we will illustrate possible approaches for when general purpose shape parametrization methods
cannot be directly employed.

Planing Yacht Hulls

We now focus on the case of the planing hull yacht illustrated in Figure 17.4. The geometry of
this kind of hull is typically characterized by a V-shaped lower part, surmounted by a lateral wall
which is typically straight and—except in the bow region—has a constant angle with respect to
the vertical direction. The V-shaped bottom part of the hull and its lateral walls are typically
divided by a sharp edge called the chine. The example in Figure 17.4 has a double chine line
which delimits a narrow chine surface.

Figure 17.4. Front view (top) and side view (bottom) of a planing hull featuring a double chine
line and substantially straight walls above the chine line.
 17.2. Parametrization Techniques 353

In the hydrodynamic design process of this kind of hull engineers typically want to investigate
how shape modifications affect performance in terms of lift, drag, and stability (or seakeeping).
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

It is quite clear that deforming the V-shaped bottom part of the hull, which is in contact with
the water throughout the navigation, will have the most impact on the overall hydrodynamic
behavior of the hull. For this reason, design engineers are mainly interested in modifying the
lower part of the hull and—to a less extent—the chine surface, while keeping the top part of the
hull substantially nondeformed. Under these requirements, FFD can hardly be used because its
smooth shape functions would not allow one to sharply separate what happens above and below
the chine lines. As for RBF, it is still possible to employ its control point displacements to impose
the desired deformation on the bottom part of the hull, and a somewhat rigid displacement on
the top part, to accommodate for the bottom part deformation. Yet this opens the problem of
identifying the RBF control point displacements that would result in the desired deformations of
the bottom and top hull regions. So, rather than solving the latter problem, it is convenient to deal
with the more intuitive problem of developing a parametrized law to displace any point on the hull
surface and generate admissible overall deformations. As we will see in Subsection 17.3.3, RBF
is still used in this context as an extremely valid tool to extend the hull surface deformations to
the surrounding volume, adapt the volumetric computational meshes, and carry out parametrized
fluid dynamics simulations.
The parametrized law to displace any point on the hull is in this case developed by taking
advantage of the main design features of the object at hand. As suggested in Figure 17.5, the
transverse vertical sections have the same topology for most of the hull length. Each section is
in fact characterized by the bottom vertex coinciding with the section plane intersection with the
keel line and by two vertices resulting from the section plane intersection with the chine lines.
So, the devised parametrization strategy acts by modifying each section with the same strategy
based on these easily identifiable points. More specifically, the V angle of the hull bottom is
altered by rigidly rotating all the section points located between the intersection of the internal
chine line and the keel line around the keel line intersection by the desired angle θ. This rotation

Figure 17.5. Vertical section of a planing hull perpendicular to its longitudinal axis. The bottom
vertex of the section is the intersection of the section plane and the keel line. The intersections of the two
chine lines are also visible on the bottom left side of the section. The first step of the parametrization
strategy is a rigid rotation of the bottom part of the section around the keel intersection. This is combined
with a rigid translation of the chine and wall surface sections needed to keep the top part of the section in
contact with the displaced bottom part.
 354 Chapter 17. Geometrical Parametrization and Morphing Techniques with Applications

results in a net displacement of the internal chine line intersection point. This displacement is
used to rigidly translate the section of the chine and wall surfaces and ensure that the top part
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

of the section remains connected to the rotating bottom. Under the same principle, once these
two steps are completed, it is possible to add rotations to the chine surface and, if desired, to the
wall surface. Finally, a further parametrized function can be introduced to modulate the sectional
rotations introduced along the hull’s longitudinal coordinate.
In summary, the present example suggests that when a highly engineered object such as a
planing yacht hull is to be parametrized by fulfilling some specific shape constraints, a viable
possibility is to exploit the main topological features of the original geometry. Such topologi-
cal features are often a result of the characteristics that designers want their object to possess.
So, as in the case of the keel and chine lines for a planing hull, they can be used to build a
parametrization law that is useful to engineers.

Propellers

As previously mentioned, aircraft wings and propeller or turbo machinery blades are further ex-
amples of objects for which RBF, IDW, and FFD are not ideal shape parametrization algorithms.
This is mainly because such objects are designed starting from a set of two-dimensional airfoil
sections, which are placed in the three dimensional space according to some strict design rule. If
we consider for instance a propeller blade, a specific design table first reports the list of airfoil
section shapes to be used in its generation. In such a design table, each airfoil section, originally
described in the R2 space, is also complemented with a set of parameters (radius, chord length,
pitch angle, skew angle, and rake) which indicate how the airfoil section should ultimately be
placed in the R3 space. As depicted in the left part of Figure 17.6, the airfoil sections are first
laid on cylinders with axes coinciding with that of the propeller. Thus, each airfoil section is
associated with a specific radius. The chord length parameter represents the distance between
the leading and trailing edges of the airfoil, projected on the section cylinder. The skew and

Figure 17.6. Front view (left) and lateral view (right) of a ship propeller. In both images, one of
the five propeller blades’ surface is transparent, and a set of airfoil cylindrical sections is indicated in red
to show how the blade surface is generated as the envelope surface of the single sections.
 17.2. Parametrization Techniques 355

pitch angles represent instead the rotations prescribed for each section around the blade axis and
the entire propeller axis, respectively. Finally, the rake is the translation along the propeller axis
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

prescribed for each section. For more details about propeller blade design, the interested reader
is referred to in [101].
It should at this point appear clear that the design procedure described automatically leads
to parametrized propeller blades. In fact, the chord length, pitch angle, and skew angle at each
radius corresponding to the airfoil sections available are indeed parameters of the generated
propeller blade. Moreover, several different families of parametrized airfoil sections are available
across the engineering community. For instance, the NACA airfoils [8] are parametrized based
on quantities such as maximum thickness, maximum deflection of the camber line, and location
along the chord of the maximum deflection point.
Despite the fact that the previous description suggests that the geometry of propeller blades—
as well as turbo machinery blades and aircraft wings—possesses an inherent parametrization, it
is still possible to modify the parametrization to have better control over the finalized shape. As
depicted in Figure 17.7 for the case of pitch radial distribution, it is possible to use B-splines
to interpolate the blade parameter variation along the radial coordinates. This results in the
possibility of switching the shape parameters to the control points of the single chord length,
pitch, skew, and rake radial distribution curves. Given the smoothness of B-splines, this ap-
proach results in a smoother parameter dependence, especially considering that the CAD mod-
elers used to build the three-dimensional surfaces passing through the airfoil sections are also
based on B-splines and NURBS surfaces. A further advantage of this approach is that if the
two-dimensional airfoil parameters are also interpolated with this strategy, the number of air-
foil sections can be arbitrarily increased to provide the CAD modeler with more sections and
allow them to have more control over the generated three-dimensional surfaces. For an efficient
selection of such parameters using the active subspace (AS) method, please refer to [404] and
to Chapter 16.
The standard generation of the three-dimensional propeller blade from its design table and
from B-spline interpolated radial distributions of airfoil parameters have been implemented in
the Python package BladeX [210]. Given sectional values of airfoil section, radii, chord length,
pitch and skew angle, and rake, BladeX can directly generate CAD representations of the cor-
responding blade surface. In addition, the software supports the B-spline representation of the
airfoil section parameter variations along the relative radial coordinate, which as mentioned can
be used to reparametrize the blade shape or to enhance the smoothness of the resulting CAD
surface.

1.70 Original PPTC pitch

1.65 Original ctrl points
Deformed pitch
1.60 Deformed ctrl points
1.55
[rad]

1.50
1.45
1.40
1.35
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
r/R

Figure 17.7. The airfoil section pitch as a function of relative radial coordinate. The plot shows
both an initial pitch distribution (red continuous line) and the control points associated with its B-spline
interpolation (red piecewise linear line). A new, smooth pitch radial distribution (blue continuous line) is
then obtained using a displaced set of pitch control points (blue piecewise linear line).
 356 Chapter 17. Geometrical Parametrization and Morphing Techniques with Applications

17.3 Application to Different Geometry Specifications

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

The parametrization strategies described in the previous section have been applied to different
design problems. In this section, we will describe a wide spectrum of shape parametrization ex-
amples, with particular focus on practical details such as the file format containing the geometric
information of the considered objects. In addition, we will distinguish between cases in which
the parametrization tools only act on the surfaces describing the object of interest [574, 404] and
those that are used to deform a three-dimensional computational grid [32]. In the former case,
the deformation of a codimension-one, R2 surface embedded in R3 is studied. In the latter case,
the complete morphing of a codimension-zero computational domain Ω ⊂ R3 is instead carried
out. Recall that the codimension of a subspace U of a vector space V is the dimension of the
quotient space V /U . It is indicated by the symbol codimV U and is equal to the dimension of the
orthogonal complement of subspace U in V , namely

dimU + codimV U = dimV.

17.3.1 Geometric Model Parametrization

This subsection focuses on the deformation of an object or artifact shape by parametrizing the
surfaces that delimit it. In fact, most CAD tools used for industrial manufacturing identify the
shape of a three-dimensional object by mathematically describing its codimension-one bound-
aries. For this reason, we first discuss applications in which surfaces in R3 are specified by both
triangulations and parametric equations.

Triangulated Surfaces

Triangulated surfaces describe a codimension-one surface using a finite nonoverlapping trian-

gulation obtained by discretizing the original surface manifold. Such objects are particularly
widespread in many computational and simulation contexts because of their discrete nature. Con-
trary to CAD surfaces (which will be extensively discussed in the next subsection), triangulated
surfaces are defined only by the coordinates of the triangle’s vertices and their topological con-
nections, so they are easier to handle. Usually stored in stereo lithography (STL) interface files
[5], triangulated surfaces are widely adopted in the meshing phase. In several applications, sur-
face deformation is preferred over computational mesh deformation, since it allows one to apply
the meshing procedure after surface morphing, ensuring better control of the quality of the final
mesh. Note that this approach requires creating a new grid for all of the deformations, increasing
the required computational effort.
Given the generic parametric deformation map M(·, µ), it is applied to a triangulated surface
by applying this mapping to all nodes of the triangulation. Thus, the topology is unaltered. Fig-
ure 17.8 demonstrates surface deformation applied to the bulbous bow of a naval hull by means
of FFD for a shape optimization problem [154], where ROM was used to reduce the computa-
tional load. Here the control points span around the bulb of the ship to induce a deformation only
in that region, thus reducing the wave resistance of the ship (in a calm-water environment). As
we can note from the picture, the surface derivative is kept continuous, an important feature for
many industrial applications.
Another example of deformation applied to triangulated surfaces is depicted in Figure 17.9,
where a biomedical problem is presented [568]. In this case, an RBF method is applied to
the geometry of a carotid artery, where the undeformed triangulation is morphed in order to
obtain a patient-specific geometry, thanks to the interpolatory nature of the method itself. This
parametrized problem, numerically solved by adopting a POD-Galerkin framework, aimed to
 17.3. Application to Different Geometry Specifications 357
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 17.8. FFD application to the STL triangulation representing the bulbous bow of a cruise
ship hull. The left panel, which represents the original bow geometry, also displays the FFD undeformed
control point lattice. On the right, the visible displacement of the control point results in a deformed STL
geometry.

Figure 17.9. Two examples of a triangulated surface of a carotid artery deformed with RBF. The
original geometry is emphasized by the blue dots, while the RBF control points are depicted in red (the
deformed ones) and in green (the undeformed ones).

discover possible occlusions after the bifurcation of the carotid by simulating the blood inside
the vessel.

CAD Surfaces

We now describe some examples of the application of the shape parametrization strategies dis-
cussed in Section 17.2 to NURBS surfaces [530]. NURBS are an important component of numer-
ous industry and computer graphics CAD standards, such as IGES [336], STEP [4], and ACIS.
A full mathematical description of NURBS surfaces and their application to analytic geom-
etry is beyond the scope of the present work (we refer the interested reader to [488, 489]). Yet,
for our purposes it is at least important to recall here that NURBS curves and surfaces are de-
 358 Chapter 17. Geometrical Parametrization and Morphing Techniques with Applications
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 17.10. A simplified two-dimensional example of the application of FFD to a NURBS

curve. On the left, the procedure starts by placing an FFD lattice (red dots) on top of a NURBS curve. On
the right, the displacement of the FFD control points is used to modify the position of the NURBS control
points (black dots). The modified control polygon results in the morphed NURBS curve displayed on the
right.

fined by their order, by a set of weighted control points, and by a knot vector. Control points
in particular either are always directly linked to the curve or surface or behave as if they were
connected by a spring to it. For this reason, they offer the possibility of modifying the NURBS
surface by altering the position of the control points. This very intuitive editing mechanism—
which is among the features that made this mathematical instrument very popular in the CAD
and computer animation industries—is exploited here to apply any parametrization strategy to
NURBS surfaces. As illustrated in Figure 17.10, a NURBS curve is deformed using FFD by di-
rectly applying the parametrization algorithm to the NURBS control points and then computing
the resulting modified shape.
It is important to point out here that the deformed surface obtained by moving the CAD
control points is identical to the one obtained by moving every surface point only if the trans-
formation law applied is affine [488]. We also recall that most of the parametrization algorithms
discussed in the present work do not, in general, lead to affine transformations. Thus, we must
remark that deforming a CAD shape by displacing its control points might result in slight dif-
ferences compared to deforming it by a surface triangulation. We report that, for all the exam-
ples discussed in the present chapter, the difference between applying the same deformation to
NURBS surfaces or to their triangulated equivalent is quite modest. Yet, we stress that this issue
must be taken into account if consistency between NURBS and triangulated surface deforma-
tions must be enforced up to a very strict tolerance. Ideally, in the latter case it is possible to
resort to least squares fitting to generate a novel NURBS surface passing as close as possible to
the deformed triangulation. We warn the reader, however, that this procedure would still bring in
the possibly significant source of error associated with the least squares minimization.
Figures 17.11 and 17.12 depict an example of the application to CAD geometries of the
parametrization algorithms described in Section 17.2 and presented in [155]. The shape consid-
ered is that of the DTMB-5415 hull [401], which was originally conceived for the preliminary
design of a US Navy Combatant ship. FFD is used here in the bow regions to displace the control
points of the hull NURBS surfaces and create a family of hulls which feature deformed sonar
domes. The images display the front and side views of both the original DTMB-5415 IGES
geometry (right and bottom), and of the IGES geometry of a specific modification in which the
sonar dome has been enlarged and inclined downward (left and top).
As can be appreciated in the pictures, the DTMB-5415 hull portrayed in Figures 17.11
and 17.12 is composed of three NURBS surfaces. Such faces border each other through shared
edges in which the shape is C 1 or C 0 continuous only up to a tolerance specified in the IGES
file containing the hull data. We use this very common example to remark that an additional
problem associated with deforming NURBS surfaces by displacing their control points is that
 17.3. Application to Different Geometry Specifications 359
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 17.11. An application of the FFD algorithm to the NURBS surface specified by an IGES
file. A front view of the DTMB-5415 navy combatant hull featuring a modified and enlarged bow is displayed
on the left; the original hull surface is displayed on the right for reference.

Figure 17.12. A side view of the IGES geometry of the DTMB-5415 navy combatant hull shown
in Figure 17.11. The top picture depicts the CAD surface of the hull, modified via FFD. The bottom picture
shows the original hull surface.

it can introduce discontinuities in shapes composed of several CAD patches. Despite the fact
that all parametrization strategies described in this chapter are designed to be continuous maps
from R3 to R3 , the control point distribution in two bordering faces can in principle be so dif-
ferent that the deformation affects the two surfaces unequally. This might potentially alter the
continuity tolerance of the original CAD file and result in small gaps, overlaps, or normal vec-
tor jumps appearing between the NURBS surfaces which were originally smoothly joined. This
loss of C 1 and/or C 0 continuity is more pronounced when the deformation map gradients are
high relative to the local NURBS control point density. Thus, a solution to mitigate the problem
is to add control points to the NURBS surfaces. In the vast majority of cases considered, this
procedure proved effective in obtaining small continuity tolerances in composite CAD surface
deformations, although the increased number of control points results in larger CAD files [570].
The ship propeller family shown in Figure 17.13 represents a different strategy of obtaining
parametrized IGES geometries. The algorithm for propeller design described in Section 17.2 was
used for the bottom-up generation of parametrized ship propeller blades. In this specific case,
each blade is generated as the NURBS surface passes through 12 airfoil sections, each defined
by a different curve in the xy plane and different values of radial coordinate, chord length, pitch
angle, skew angle, and rake. The procedure implemented in the library BladeX [210] allows
 360 Chapter 17. Geometrical Parametrization and Morphing Techniques with Applications
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 17.13. Representation of a family of ship propellers generated with bottom-up construc-
tion. The different propellers are obtained in this case by scaling the skew angle curve. In particular, in the
image an original propeller blade shape is presented along with two modifications resulting from scaling
the skew curve by factors of 0.8 and 1.1, respectively.

the airfoil section shape to be specified by means of a NACA four- or five-digit identifier or by
means of a CSV file. Once the radial coordinate, chord length, pitch angle, skew angle, and
rake associated with every section are specified by the user, the airfoil section points are placed
in the three-dimensional space, and a series of NURBS curves are generated, passing through
all the points of each section. Finally, the NURBS surface passing through all the sectional
curves is generated and joined to the propeller hub. The latter operations are carried out using
functions of the CAD Python library PythonOCC.10 In the example reported in Figure 17.13, the
original blade shape is portrayed along with two variants generated by altering the skew angle
distribution along the radial coordinate. We recall that the skew angle designs the position of
each radial blade section along the circumferential coordinate of the propeller disc plane. Thus,
the blade generated by scaling the skew values by a factor of 1.1 appears more curved along the
radial direction than in the original design. On the other hand, scaling the skew values by a factor
of 0.8 results in a propeller with blades that appear straighter along the radial direction.
We finally point out that the specific deformation displayed in Figures 17.11 and 17.12 is
part of a family of 130 IGES hulls generated via FFD in the framework of a CFD simulation
campaign presented in [155]. In that work, the geometric deformation was concentrated in the
region surrounding the bulbous bow. A further set of 130 hulls was obtained by deforming the
full length of the hull and was used to carry out the CFD simulations presented in [574]. A family
of 500 ship propellers like the one depicted in Figure 17.13 was also designed in the framework
of an optimization campaign presented in [404]. In all these works, active subspaces (ASs) were
successfully used to detect possible redundancies in the geometric parameters introduced with
FFD (please refer to Chapter 16). On this note, we point out that AS analysis is an extremely
useful postprocessing tool for complementing shape parametrization algorithms because it allows
a reduced number of combined parameters to be identified which have the most effect on the
output of interest. In addition, it can also be used to evaluate the output sensitivity to variations
of the original geometric parameters and eliminate inconsequential parameters.

10 Available at https://github.com/tpaviot/pythonocc-core.
 17.3. Application to Different Geometry Specifications 361
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 17.14. Examples of computational domain deformation using FFD: the first picture (left)
shows the edges of the original mesh, while the others illustrate two deformed domains.

17.3.2 Computational Domain Parametrization

In this subsection, we turn our attention to codimension-zero domains and their parametrization
using the methodologies described in Section 17.2. Since these domains include subsets of R3 in
three dimensions and of R2 in two dimensions, the examples will mainly refer to the morphing
of volumetric computational grids for the solution of PDE problems arising from fluid dynamics
or structural simulations.

FFD and RBF Applied to Computational Grids

A first example, presented in Figure 17.14, illustrates the application of FFD to deforming the
structured surface mesh surrounding a square in a two-dimensional application. The images,
which refer to the original configuration (left) and to two deformed domains (center and right),
depict a closeup of the grid region surrounding the square.
Figure 17.15 shows instead an application of RBF in the context of a CFD simulation cam-
paign [639]. The pictures show two possible modifications of the Ahmed body [10] obtained
by altering the inclination of the diagonal surface at the rear end of the body. The computa-
tional grid for the CFD simulations is deformed by distributing a series of control points on the
boundaries of the three-dimensional domain and by displacing a portion of those located on the
modified diagonal surface. The vertical cut of the resulting volumetric grids is shown in the
plots on the top. The corresponding bottom images display the path lines of the mean velocity
field resulting from the unsteady Reynolds-averaged Navier–Stokes (RANS) simulations carried
out with the deformed grid. As can be appreciated, the deformation strategy is able to alter the
original computational mesh without impairing the convergence and effectiveness of the CFD
simulations.

17.3.3 The Use of RBF to Extend Boundary Deformation to Internal

Grid Nodes
In this section, we discuss a further interesting and useful application of RBF to propagate the
boundary deformation to the internal nodes of volumetric grids for PDE-based simulations. Most
of the algorithms designed to modify the shape of specific artifacts—such as the planing hulls and
propellers described in Subsection 17.2.2—result in the deformation of codimension-one mani-
folds. Thus, if these deformation methodologies must be used in a pipeline involving volumetric
PDE resolution, they must be complemented by an adequate strategy to extend the boundary
displacements to the whole, codimension-zero, computational domain.
 362 Chapter 17. Geometrical Parametrization and Morphing Techniques with Applications
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 17.15. The application of RBF shape parametrization to the Ahmed body and the sur-
rounding volumetric mesh for CFD simulations. In the test illustrated, the inclination of the diagonal
surface at the rear end of the Ahmed body is modified to investigate its effect on aerodynamic performance.
Two configurations, including a vertical cut of the volumetric mesh, are shown in the top images. For ref-
erence, path lines of mean velocity field resulting from the corresponding unsteady RANS simulations are
shown in the bottom pictures.

Figure 17.16 shows the deformation of the CFD computational mesh past a Potsdam propeller
test case (PPTC) ship propeller, following a variation of the blade pitch in the tip region obtained
with the strategy described in Subsection 17.2.2. In the plot, the blue lines represent the edges
of a cut of the volumetric mesh obtained by means of a plane perpendicular to the propeller axis.
The red lines refer to the edges of the deformed mesh in correspondence with the same sectional
plane. The closeup of the blade tip region displayed on the right shows how the deformation of
the blade surface is smoothly propagated throughout the internal nodes of the volumetric mesh,
avoiding the presence of inverted or highly skewed cells.
Finally, we want to stress that for practical reasons the RBF extension to volumetric meshes
is often convenient for surfaces originally deformed with FFD. As previously discussed, in prin-
ciple FFD allows for simultaneous deformation of both the boundary surface and the volumetric
grid. Yet, when deforming volumetric meshes, the aim is to distribute the deformations over a
high number of cells, rather than concentrating all the displacements in a very confined region
in which cells can get distorted or even inverted. But because FFD only affects points located
within the control point lattice, the latter must extend to a bigger volume. In addition, to max-
imize the volumetric mesh quality, the user must include more control points in the lattice to
make sure that different deformation magnitudes are imposed in regions close to the deformed
surface and far from it. Such manual control over the local mesh deformation can often become
quite cumbersome.
For these reasons, in several cases it will be convenient to employ FFD only to deform the
boundary surface mesh or a triangulated format of its geometry. Then, after the surface mesh has
 17.4. Conclusions and Future Perspectives 363
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 17.16. Example of computational domain deformation applied to a propeller design prob-
lem, seen at the radial plane. The blue edges refer to the original mesh, whereas the red ones show the
deformed grid.

been modified using FFD, it is possible to resort to RBF to propagate the boundary displacements
to the internal nodes of the volumetric mesh for CFD simulations. In a broader sense, RBF is
an interpolation algorithm and is used here to estimate the node displacement function on every
point of a domain, based on the function values prescribed on the boundary. In the framework
of RBF, this is done by linear combinations of radial functions used to approximate at any point
of the domain a function with values prescribed only at a finite number of points. In the case of
interest, the displacement field function is prescribed only on the points of the boundary surface
and is interpolated in the positions corresponding to every node of the volumetric mesh. The
application of this procedure in a naval engineering test case (see [157] for further details) is
illustrated in Figure 17.17. The illustration displays a sectional view of the original (blue lines)
and deformed (red lines) volumetric mesh around the surface of a ship hull. As documented in
the original publication, the combined FFD/RBF strategy has been used to successfully generate
200 shape variations of a container ship hull and accordingly deform the volumetric grids to carry
out an optimization campaign based on the OpenFOAM [612] CFD solver library. The output
of the checkMesh utility of the OpenFOAM library confirms that the mesh deformation does not
significantly alter either the minimum face area or the minimum cell volume, while it only leads
to a modest—7.2% at most—increase in the average mesh nonorthogonality index. This increase
is well below a level that can significantly affect the result of the CFD simulations.

17.4 Conclusions and Future Perspectives

In this chapter we presented a diverse range of possible nonaffine shape morphing and parametriza-
tion techniques. In the general purpose group, we introduced FFD, RBF interpolation, and IDW
interpolation. These techniques are characterized by the fact that they can be applied to any
object or part of it, without any limitations. Then we presented more specific deformation meth-
ods which are necessary when we want to morph highly engineered artifacts while preserv-
ing some characteristics of these objects. We focused on naval engineering applications, where
deformations of a marine propeller and a sidewall of a planing motor yacht were presented.
 364 Chapter 17. Geometrical Parametrization and Morphing Techniques with Applications
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 17.17. A sectional view of a volumetric grid deformation carried out using RBFs in a
naval engineering problem. The blue lines denote the original volumetric mesh edges, while the red lines
indicate the deformed configuration.

These artifacts also have to be actually manufactured, so general purpose parametrizations can
easily generate objects which are very costly to produce or infeasible.
In the second part of the chapter we presented methodologies to guarantee that any modifica-
tion of the geometry is reflected in the computational mesh. The advantage of these techniques is
twofold: first, the shapes validated by simulations will also be available for manufacturing; sec-
ond, we can easily apply nonintrusive reduced order methods (refer to Chapter 9) to the database
of solutions since we preserve the same number of degrees of freedom.
Future studies will address the possibility of applying the FFD technique while satisfying
some constraints. For example, we might want to preserve the volume of the deformed object, or
we may want to impose the presence of a second object near the one being morphed. Addition-
ally, some machine learning techniques can be exploited to propagate the surface morphing to the
volumetric computational mesh; in this chapter we proposed such a deformation extension using
RBF, which shows good results but may decrease the quality of the deformed mesh. To over-
come this issue, a neural network can interpolate the deformation on grid nodes, adding some
constraints during the neural network training step to produce high-quality deformed meshes.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Chapter 18

Reduced Order Methods

for Hemodynamics
Applications

Zakia Zainib, Pierfrancesco Siena, Michele Girfoglio,

Martin W. Hess, Francesco Ballarin, Gianluigi Rozza

18.1 Introduction
Ischemic cardiovascular diseases, e.g., coronary artery disease (CAD), and diseases causing
blood regurgitation, such as mitral regurgitation (MR), are life-threatening events and there-
fore worth studying using computational fluid dynamics (CFD) tools. In CAD, plaque build-up
causes partial or complete blockage in blood perfusion to the heart muscle and risk of ischemia,
thrombosis, stroke, and heart attack. Likewise, severe MR leads to extra strain on the heart to
fulfill the blood supply demand and can result in arterial hypertension, atrial fibrillation, stroke,
and heart failure. Another cardiovascular problem worth exploring is the increased risk of aortic
thrombosis in the left ventricular assist devices (LVADs) used to provide mechanical circulatory
support during heart transplantation and destination therapy.
In studying these phenomena, one is interested in identifying the underlying causes, exploring
patency of surgical treatments and assisting devices, addressing the computational limitations,
and developing and implementing robust and efficient computational techniques. For instance,
coronary artery bypass graft (CABG) is a common surgical treatment for CAD and is needed by
the surgeons to identify the risk of restenosis in a noninvasive patient-specific manner. Thanks
to technological strides, the related techniques have advanced to patient-specific hemodynamics
studies; however, the challenge of computational efficiency and robustness remains. Similarly,
it is of interest in MR to identify the main features responsible for the blood flow conditions
caused by the nonsymmetric, wall-hugging jets that are responsible for the regurgitation, while
for LVADs, it is crucial to detect the optimal location and inclination of the anastomosis and to
tune the device flow rate. Corresponding computational modeling involves complex formulations
and thus reduced computational efficiency.
Full-order numerical techniques such as finite element methods (FEMs), finite volume meth-
ods (FVMs), and spectral element methods (SEMs), as is well known, provide computationally
stable and reliable solutions for complex problems such as the cardiovascular ones. Neverthe-
less, this accuracy depends on the degrees of freedom of the mathematical problem, which in turn
rely on the discretization size of the computational domain. Hemodynamics modeling in com-

365
 366 Chapter 18. Reduced Order Methods for Hemodynamics Applications

plex cardiovascular geometries, even if restricted to a particular and small part of the heart or the
vessel of interest, requires a sufficiently dense mesh for stable solutions. Therefore, such mod-
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

eling problems in the cardiovascular domain, especially with patient-specific anatomies, usually
comprise a large number of degrees of freedom and therefore have a high computational cost.
Furthermore, these problems are set in a many-query environment, meaning that cardiovascu-
lar studies cannot be beneficial unless repeated to tune some parameter values, which raises the
computational cost to an untenable level. This poses the challenge of developing and utilizing
reduced order modeling (ROM) techniques that keep the reliability of the full-order numerical
methods intact and increase the computational efficiency.
ROMs can be combined with other mathematical techniques to effectively address mathe-
matical limitations in cardiovascular hemodynamics problems. For instance, imposing accurate
and meaningful blood flow conditions at the boundaries of the cardiovascular domains is essen-
tial for realistic hemodynamics modeling. Multiscale modeling is used to face this challenge;
it uses lumped network and distributed network models to ensure the accuracy of the boundary
conditions using electrophysiology. The cost for these approaches is increased mathematical
complexity and lack of robustness. In other words, the physiological complexity of the biomed-
ical applications requires one to address several hemodynamics and mathematical challenges,
such as imposing accurate and meaningful boundary conditions and ever-changing hemodynam-
ics patterns.
This chapter illustrates the application of the ROM framework to the three cardiovascular
problems mentioned above: CABGs, the Coandă effect in regurgitant mitral valves, and LVADs.
These applications address several of the mentioned challenges regarding computational effi-
ciency and robustness.

18.2 CABGs
Coronary arteries have a diameter of only a few millimeters and are responsible for perfusing
the heart muscle with oxygenated blood. Atherosclerosis, or plaque build-up inside the arteries,
causes structural rigidity, and the resulting narrowing of the arteries reduces the blood supply.
Severe blockage leads to ischemia and can result in stroke, angina attack, or heart attack. In acute
stages, the disease requires invasive surgical treatments such as CABGs. CABGs are external
vessels anastomosed to the blocked coronary arteries and serve as alternative conduits for the
blood. Here we will illustrate the application of ROM to an optimal flow control problem (OFCP)
(Chapter 4) related to CABGs [638].

18.2.1 Problem Statement

For the following discussion, three-dimensional CABG geometries are considered as a bounded
and connected computational domain Ω with Lipschitz continuous boundary ∂Ω. For the imple-
mentation part, these domains are reconstructed as vessels with patient-specific anatomical struc-
ture from medical image data provided by the Sunnybrook Health Sciences Centre in Toronto,
Canada. The reconstruction process is overviewed in [638]. The boundary of the CABGs, ∂Ω,
is subdivided into three parts: the inlet boundary Γin , the walls Γw , and the outlets Γo . Owing
to the small diameter of coronary arteries, the viscosity of the blood can be considered a con-
stant and, hence, the blood is assumed to behave as a Newtonian fluid. Mathematically, this is
represented by the incompressible parametrized Navier–Stokes equations:

−η∆v (µ) + (v (µ) · ∇) v (µ) + ∇p (µ) = f (µ) in Ω, (18.1)

∇ · v (µ) = 0 in Ω. (18.2)
 18.2. CABGs 367

Next, different boundary conditions are chosen for ∂Ω. The boundary of the walls, i.e., Γw , is
assumed to be rigid and nonpermeable and, therefore, no-slip boundary conditions are considered
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

there. Furthermore, the parametrization is introduced through the inflow conditions; i.e., the
Reynolds number defining the blood flow velocity at the inlets Γin is the parameter at hand. The
inflow velocity is imposed as a Poiseuille profile owing to the pipelike structure of the vessels
and the Dirichlet conditions. Lastly, at the outlets Γo , the unknown control variable denoted
by u is imposed using the Neumann conditions. The control variable accounts for different
outflows corresponding to different inflows generated by the tuning the Reynolds number. These
conditions are mathematically written as

v (µ) = vin (µ) on Γin , (18.3)

v (µ) = 0 on Γw , (18.4)
−η (∇v (µ)) n + p (µ) n = u (µ) on Γo . (18.5)
The goal of an OFCP is to optimize (minimize/maximize) certain flow quantities while con-
trolling some unknown quantities that can affect the fluid flow behavior. Here, the quantity to be
optimized is set in a tracking-type manner, guided by the intended application at hand in patient-
specific data assimilation. This objective is motivated by the need to integrate patient-specific
physiological information in everyday health-care decisions and can be achieved in several ways,
one of which is to minimize the discrepancy between real-life physiological data and the output
of the above-defined mathematical model. Thus, the cost functional (also referred to as objective
functional, objective function, or cost function) J is defined as

1 α
Z Z
2
J (v, u; µ) = |v (µ) − vo | dΩ + |u (µ) |2 dΓo , (18.6)
2 Ω 2 Γo

where vo can be the patient-specific blood flow velocity read from patient-specific medical im-
ages, such as four-dimensional MRIs provided by the hospital or the clinic. In other words, J
in equation (18.6) is the distance between the blood flow velocity approximated using computa-
tional hemodynamics, i.e., v, and the real-life patient-specific blood flow velocity, i.e., vo . The
second term on the right-hand side of (18.6) denotes the energy of the control variable u, with
α > 0 as penalization parameter. The adjoint-based Lagrangian approach [468, 250], which is
commonly used in PDE-constrained OFCPs, is adopted here.

18.2.2 ROM
We use the Galerkin FEM approximation to discretize the full-order model (FOM) and proper
orthogonal decomposition (POD) [281, 352, 34] to construct the reduced order spaces. An alter-
native to POD is the greedy algorithm [281], which uses residual-based error estimators. In the
literature, both techniques have been applied to OFCPs [413, 414, 562, 150] and patient-specific
computational cardiovascular modeling [32, 33, 568]. Technical details can be found in Rozza
et al. [281, 472] and Ballarin et al. [34].

18.2.3 Numerical Results

The numerical tests reported here aim to match a certain desired blood flow velocity vo using J
defined in (18.6) for different parameter-induced inflow velocity scenarios. Such numerical tests
have used artificial target data for methodological verification, and, therefore, a blood vessel is
assumed to have the same target velocity in its entire volume. In real-life scenarios, this work can
be extended to four-dimensional data assimilation which can use blood velocity obtained from
 368 Chapter 18. Reduced Order Methods for Hemodynamics Applications

four-dimensional flow MRIs as the target velocity. In these cases, it is realistic to assume that
the velocity profile will vary from cross-section to cross-section. The target velocity vo in both
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

cases can be imposed in a parabolic shape pointwise along the center lines, which are defined to
be the lines between two sections of lumen such that their minimal distance from the boundary
of the vessel is maximal, using the expression
r2
 
vo = vconst 1 − 2 tc . (18.7)
R
Here, for a vessel, tc is the pointwise tangent along the corresponding center line in the axial
direction, R is the maximum radius of the vessel corresponding to points on the center line, and
r is the distance between mesh nodes and the nearest point on the center line.
To set the value of vo , it is observed that the maximum coronary flow velocity ranges from
1 m/s to 4 m/s [208, 524]. Therefore, one can choose any value for the target velocity that has
its maximum in this range. These numerical examples are run with the magnitude of vo equal
to 0.35 m/s or 350 mm/s  4 m/s. Moreover, the constant kinematic viscosity is assumed
to be η = 3.6 mm2 /s and the velocity at the inlets is imposed using the following parametric
expression, with the Reynolds number Re as the parameter:
r2
 
ηµ
vin (µ) = − 1 − 2 nin , µ := Re ∈ D. (18.8)
Rin Rin
Here, Rin is the maximum radius of an inlet, nin denotes the outward normal to the inlet, and r
is the distance between mesh nodes and the center of an inlet.
Furthermore, the average error Eδ for δ = s, z, u is defined to be the norm of the difference
between the FEM solution and the ROM solution for a fixed parameter value in the corresponding
continuous space, and the relative error (Erel ) is the average error relative to the truth solution.
Moreover, the absolute average error (ET ) is the sum of squares of the average errors, and the
absolute relative error is ETrel = (ksh k2 +kuh k2ET +kzh k2 )1/2 . Lastly, the difference be-
S(Ω) U (Γo ) Z(Ω)
tween the Galerkin FEM and the POD-Galerkin approximations of the objective functional is
calculated using the equation
EJ = |J (xh ; µ) − J (xN ; µ) |. (18.9)
These numerical examples are solved in RBniCS [38, 281] and multiphenics [6].
In the domain at hand, the graft connection is realized between the right internal mammary
artery (RIMA) and the stenosed left anterior descending artery (LAD). The parameter is assumed
3
to take values in D = [70, 80]. Moreover, vo ∈ L2 (Ωa ) and the parameter is fixed at µ = 80


for the reduced order problem. To solve the truth problem, Galerkin FEM is used with the
stable Taylor–Hood pair of spaces, i.e., P2-P1 for velocity and pressure and P2 for control. This
achieves 4 × 105 degrees of freedom, i.e., N = 4 × 105 . One such simulation requires 1213.3 s.
For the ROM problem, Nmax = 6 eigenvectors are used, which gets the reduced order
dimensions to N = 79. These reduced bases include 48 bases altogether for velocity and adjoint
velocity, 24 bases altogether for pressure and adjoint pressure, 6 bases for control, and 1 basis
for the lifting [501] for the nonhomogeneous parametrized Dirichlet inflow condition.
Figure 18.1 (right) shows that 99.99% of the energy of the full-order solution manifold is cap-
tured in six eigenvectors. Thus POD-Galerkin approximations retain the accuracy and reliability
of a FEM solution, while using only 79 bases as compared to the 4 × 105 degrees of freedom of
the latter (see Figure 18.2 (left) and (right)). Moreover, the magnitude of the Dirichlet boundary
control is reported in Figure 18.1 (left).
Figure 18.3 (top left) illustrates the average errors Eδ for δ = v, p, u, w, q. Ev decreases from
101 to approximately 10−5 , Ep decreases from 106 to 10−2 , and Eu decreases almost to 10−6
 18.2. CABGs 369
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 18.1. CABG: Boundary control magnitude mm2 /s2 (left) and eigenvalue reduction (right).

Figure 18.2. CABG: Simulation velocity v: comparison between FE approximation (left) and
ROM approximation (right).

as the number of eigenvectors used goes from one to six. Error reduction for adjoint variable
approximations follows a similar pattern as error reduction for state variables. Total error and
total relative error are reported in Figure 18.3 (top right). ET decreases from approximately
106 to 10−2 for n = 1, . . . , Nmax , and the maximum ET demonstrates the same behavior. A
reduction from 10−2 to 10−10 is achieved for both ETrel and the maximum ETrel in this case.
Figure 18.3 (bottom) shows the difference between J computed using FOM and ROM. It is
calculated using equation (18.9), and a reduction from approximately 106 to 10−2 is observed,
with similar behavior for the corresponding relative error when n goes from one to six.
 370 Chapter 18. Reduced Order Methods for Hemodynamics Applications
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 18.3. CABG: Average error between FE and POD-Galerkin approximation of δ =

v, p, u, w, q (top left); total error between Galerkin FE and POD-Galerkin approximations (top right);
and error between FE and POD-Galerkin reduction of J (bottom).

18.3 Coandă Effect in Mitral Valves

The Coandă effect is named after the Romanian engineer and physicist Henri Coandă (1886–
1972). Besides hemodynamics, it occurs in various technical applications where fast jet streams
are generated but is not commonly observed in nature.
This section discusses the physical principles underlying the Coandă effect and the mathe-
matical modeling required to simulate it.

18.3.1 Physical Principles

Figure 18.4 shows the Coandă effect in a computer simulation which solves the incompressible
Navier–Stokes equations. A concentrated jet flow enters from the right into a channel. The flow
attaches itself to the lower wall. Due to symmetry, the flow might just as well attach itself to
the upper wall. Figure 18.5 and Figure 18.6 show even more details by plotting the velocity as
a vector and not only the horizontal component. These simulations are used in [270], and they
occur, e.g., in the human heart; see Figure 18.7.11
In [586] the Coandă effect is described as a wall attachment of the flow. According to the
author [586, Section 22], “The reason for it can be understood as follows. In the case of an
unconfined jet, entrainment produces an inflow of fluid from the sides. In the presence of a wall,
the jet cannot draw fluid into itself in this way. Instead it is drawn down to the wall.”
11 Figure source: en.wikipedia.org/wiki/Mitral_valve, licensed under the Creative Commons Attribution-Share Alike

3.0 Unported license, https://creativecommons.org/licenses/by-sa/3.0/deed.en; no changes were made to

the figure.
 18.3. Coandă Effect in Mitral Valves 371
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 18.4. The horizontal velocity of a jet stream entering a channel.

Figure 18.5. The velocity in the channel.

Figure 18.6. The velocity in the channel, zoomed in to the wall attachment region.

Figure 18.7. The mitral valve between the left atrium and the left ventricle produces the Coandă effect.
 372 Chapter 18. Reduced Order Methods for Hemodynamics Applications

The book [586] uses the term entrainment, which is crucial to understanding the Coandă
effect. Entrainment means that the jet draws fluid into itself from the sides. It is due to a com-
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

bination of momentum conservation and energy dissipation; see [586]. Entrainment can occur
in laminar flows as well as turbulent flows. At the interface between turbulent and nonturbulent
flow, the turbulent flow spreads to or entrains the flow nearby. The rate of entrainment is much
higher for turbulent flows than for laminar flows. It is not only wall attachment but also the
attachment of two jets to each other which can be caused by the Coandă effect.
Moreover, a low-pressure area develops where the fluid that has been entrained used to be. In
the presence of a wall, there is no further fluid to flow into this area and cancel out the pressure
gradient. As a consequence, the jet itself is drawn to the wall. The low pressure area is sometimes
also included in the definition, e.g., according to the Merriam-Webster dictionary,12 the Coandă
effect is “the tendency of a jet of fluid emerging from an orifice to follow an adjacent flat or
curved surface and to entrain fluid from the surroundings so that a region of lower pressure
develops.”
A precise definition of the Coandă effect does not exist, which is why the two quotations are
descriptive in nature. The effect occurs due to the interplay of a jet stream, entrainment, and
the low-pressure area. When a jet stream is injected into a stationary fluid, entrainment occurs
naturally. From the interaction with the stationary fluid, the stream broadens and becomes slower;
if the Reynolds number is sufficiently high then it also becomes turbulent. In most cases there
will be more surrounding fluid at the sides, which will be sucked closer to the jet stream. In the
case of a wall nearby, the jet stream sucks itself toward the wall in a sense.
In hemodynamics, the Coandă effect can be observed in the blood flow close to the heart.
Immediately after pumping and opening of the mitral valve, a jet stream of blood is produced
which is fast enough to cause the Coandă effect. Computer simulation must thus be sensitive
enough to capture these effects, such that accurate simulations are provided.
A common misconception about the Coandă effect is that related to fluid flow in air. For
example, a spoon placed below a water tap will guide the water along its curvature, or a water
bottle which is skewed only slightly above 90 degrees might produce water flow along the bottle.
These occurrences do not relate to the Coandă effect but to surface tension and other physical
effects. In principle it would be very difficult for water to produce entrainment of the surrounding
air, because the air molecules are much more mobile.

18.3.2 Mathematical Modeling

Computational models such as the one used to generate Figure 18.4 are commonly used to gain
insight into the physical behavior of flows; see [621, 457, 454]. The Coandă effect can be ob-
served in models governed by the incompressible Navier–Stokes equations.
Let Ω denote the computational domain. Incompressible, viscous fluid motion in a domain
Ω over a time interval (0, T ) is governed by the incompressible Navier–Stokes equations:
∂u
+ u · ∇u = −∇p + ν∆u + f , (18.10)
∂t
∇ · u = 0, (18.11)
where u is the vector-valued velocity, p is the scalar-valued pressure, ν is the kinematic viscosity,
and f is a body forcing. Boundary and initial conditions are given as
u=d on ΓD × (0, T ), (18.12)
∇u · n = g on ΓN × (0, T ), (18.13)
u = u0 in Ω × 0, (18.14)
12 https://www.merriam-webster.com/dictionary/Coandaeffect.
 18.3. Coandă Effect in Mitral Valves 373

with d, g, and u0 given and ∂Ω = ΓD ∪ ΓN , ΓD ∩ ΓN = ∅. The Reynolds number Re,

which characterizes the flow [291], depends on ν, a characteristic velocity U , and a characteristic
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

length L:
UL
Re = . (18.15)
ν

In the simulations presented in [457, 454], the Coandă effect is computed as steady states,
i.e., solutions where ∂u ∂t vanishes. To observe the Coandă effect, a certain minimum Reynolds
number Re must be given; e.g., in the simulations of [454] the Reynolds number must be at least
35 to observe the Coandă effect. This shows that the Coandă effect can be produced either by fast
jet flows, i.e., by a large velocity U in (18.15), or by media with a low viscosity ν. If the process
of entrainment is to be studied, then time-dependent simulations are needed. Such simulations
allow various technical improvements, such as novel wind energy generators [214].
There is a large variety of numerical discretization methods to solve the incompressible
Navier–Stokes equations, such as FEM, FVM, SEM, and discontinuous Galerkin; see [96, 97,
468, 330]. The variational form of (18.10)–(18.11) is given by
 
∂u
,v + (v, u · ∇u)Ω = (∇ · v, p)Ω − ν (∇v, ∇u)Ω + (v, f )Ω , (18.16)
∂t Ω
(q, ∇ · u)Ω = 0, (18.17)

with (·, ·)Ω the L2 scalar product over Ω. The discrete velocity-pressure solution pair (u, p) is
then found by solving (18.16)–(18.17) for all velocity-pressure test functions pairs (v, q) with
zero boundary conditions.
In order to evaluate the flow patterns for changing Reynolds number, parametric model reduc-
tion can be used to quickly evaluate the model over a range of Reynolds numbers. For example,
when the change in Reynolds number is effected by a change in the viscosity ν, then the model
readily allows for an affine decomposition, and a ROM can be applied, as outlined in [270] or
Chapter 1 of this book.
As discussed in [457], the mitral valve of the human heart produces a jet flow which shows
the Coandă effect. Since the echocardiographic assessment of MR provides only limited data
about the flow profile, computer simulations are needed to assist medical doctors. A comparison
between experiments in vitro and simulations is given in [607].

18.3.3 ROM
The ROM approach followed in [457] uses an intrusive reduced basis method (RBM) with an
affine expansion in the Reynolds number and geometry. Full-order solutions are sampled over
a coarse grid of parameters, and the POD computes the ROM projection space. In the offline-
online procedure, the parameter-independent parts of the PDE operator are precomputed and
low-order matrices are formed online according to the affine form for fast evaluations. Besides
the reconstruction of the flow, bifurcation detection is performed, which determines the onset of
the Coandă effect.
The starting point of the investigation can be seen in Figure 1 of [457], where echocar-
diographic images of the mitral value are shown. However, it is insufficient to rely only on
the echocardiographic images, and simulation results provide additional insight, as discussed in
[457]. An intrusive ROM strategy is employed, where the discrete velocity space and the dis-
crete pressure space are projected onto a low-order space. In particular, the variational form
 374 Chapter 18. Reduced Order Methods for Hemodynamics Applications

(18.16)–(18.17) is reexpressed over the ROM space as

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

∂uN
 
+ v, uN · ∇uN = ∇ · v, pN − ν ∇v, ∇uN (18.18)




,v Ω Ω Ω
+ (v, f )Ω ,
∂t Ω
q, ∇ · uN (18.19)


Ω
=0

and is solved for all velocity-pressure test function pairs (v, q) in the low-order space. The
work of [457] employs the Piola transformation to keep the divergence-free condition (18.19)
for varying geometry.

18.4 LVAD
LVADs consist of a mechanical pump providing full or partial mechanical circulatory support
to the left ventricle of the heart for patients with critical cardiac failure. The device aspirates
oxygen-rich blood from the left ventricle to push it into the aorta through a graft (Figure 18.8);
see, e.g., [7, 540, 345]. This section briefly discusses a data-driven ROM based on the proper
orthogonal decomposition with interpolation (PODI) method used in order to investigate the
aortic blood flow patterns induced by the outflow cannula of a continuous LVAD in a patient-
specific framework. For further details about this work, the reader is referred to [228, 234].

Figure 18.8. Scheme for the LVAD.

18.4.1 High-Fidelity Problem

A patient, a 66-year-old man, was considered. The LVAD implanted is the Heartmate 3 Left
Ventricular Assist System [576]. The computational domain was reconstructed from computed
tomography (CT) images provided by San Camillo hospital in Rome (Italy) using the open-
source software package 3D Slicer (http://www.slicer.org).
The FOM consists of incompressible Navier–Stokes equations (18.10)–(18.11) discretized
with a finite volume technique. Boundary conditions are essential in order to obtain accurate
cardiovascular simulation results. The LVAD flow rate P F is imposed at the outflow cannula
section. Outflow boundary conditions were applied at each outlet of the model by using a three-
element Windkessel RCR model [615, 492]: the downstream pressure pk for each outlet k is
given by
dpp,k pp,k − pd,k
Ck + = Qk , (18.20)
dt Rd,k
pk − pp,k = Rp,k Qk , (18.21)
 18.4. LVAD 375

where Ck is a compliance, pp,k and pd,k are the proximal and the distal pressure, Rp,k and
Rd,k are the proximal and distal resistance, and Qk is the flow rate. The total resistance can be
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

evaluated as P
Ak
Rk = Rp,k + Rd,k = SVR k , (18.22)
Ak
where Ak is the cross-sectional area and SVR is the systemic vascular resistance. For each outlet
R
k, we assumed Rp,kk
= 0.056 [358]. The aortic compliance can be estimated as

SBP − DBP
C= , (18.23)
SV
with SBP and DBP the systolic and the diastolic blood pressure, respectively, and SV the
stroke volume. Then the compliance Ck related to each outlet k was evaluated as

Ak
Ck = C P . (18.24)
k Ak

We note that the values of SVR, SV , SAP , and DAP were measured by the right heart
catheterization (RHC) and echocardiography (ECHO) tests [228].

18.4.2 ROM
We have used a nonintrusive data-driven approach based on a PODI method, where the POD is
used to extract a reduced basis space from a collection of parametric full-order snapshots, and
an interpolation procedure based on radial basis functions (RBFs) is adopted to compute the
reduced coefficients. So the reduced approximation of a variable ϕ related to a certain value of
the parameter π can be obtained by the linear combination

k
X
ϕ(π) ≈ αj (π)φj , (18.25)
j=1

where k denotes the cardinality of a reduced basis, φj are the POD modes, and αj (π) are the
reduced coefficients.
Thanks to its nonintrusive nature, this procedure can be applied both for primal quanti-
ties, pressure p and velocity u, and for derived quantities such as wall shear stress (WSS). Re-
garding the technical implementation of the PODI method, we used the Python package called
EZyRB [158].

18.4.3 Numerical Results

We consider the LVAD flow rate P F as a parameter. To train the ROM, we consider an equis-
paced distribution with 11 samples in the typical clinical range [3, 5] l/min. Thus, the snapshots
are collected every ∆P F = 0.1 l/min.
In order to retain the 99.9% of the system’s energy, one mode for p, nine for WSS, and eight
for ux , uy and uz are selected. We show for two test values P F = 3.45 and P F = 4.35 the
relative error Eϕ in the L2 -norm computed as

||XF OM − XROM ||L2 (Ω)

EX = 100 %, (18.26)
||XF OM ||L2 (Ω)
 376 Chapter 18. Reduced Order Methods for Hemodynamics Applications

Table 18.1. L2 -norm relative errors for pressure p, WSS, and velocity components ux , uy , and
uz for P F = 3.45 l/min.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

p WSS ux uy uz
Eϕ 0.2% 4.1% 5% 7.8% 5.8%

Table 18.2. L2 -norm relative errors for pressure p, WSS, and velocity components ux , uy , and
uz for P F = 4.35 l/min.

p WSS ux uy uz
Eϕ 0.5% 7.2% 9.7% 13.5% 9.3%

Figure 18.9. LVAD: comparison of the FOM/ROM p (mmHg) at P F = 3.45 l/min.

where XF OM is the value of a particular field in the FOM model and XROM is the one that is
calculated using the ROM in Tables 18.1 and 18.2, respectively.
Figures 18.9, 18.10, and 18.11 display a qualitative comparison between FOM and ROM
for pressure, WSS, and velocity for P F = 3.45 l/min. The very similar spatial distribution
demonstrates that the ROM is able to provide a good reconstruction for all the variables.
Finally, we comment on the efficiency of our ROM approach. The CPU time of the FOM is
72000 s, while that of the ROM is 0.01 s. This corresponds to a speedup of about six orders of
magnitude.13

13 The authors want to thank the Medical Doctors of Sacco Hospital (Milano, Italy), San Camillo Hospital (Roma,

Italy), Sunnybrook Hospital (Toronto, Canada), and Houston Methodist Hospital (Houston, USA) for fruitful collabora-
tions and discussions.
 18.4. LVAD 377
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 18.10. LVAD: comparison of the FOM/ROM WSS (Pa) at P F = 3.45 l/min.

Figure 18.11. LVAD: comparison of the FOM/ROM u (m/s) at P F = 3.45 l/min.

 Chapter 19

Scientific Software
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Development and Packages

for Reduced Order Models
in Computational
Fluid Dynamics

Nicola Demo, Marco Tezzele, Giovanni Stabile, Gianluigi Rozza

19.1 Scientific Open-Source Software for Reduced Order

Models
We dedicate this section to the more technical aspects of the software for reduced order modeling
(ROM). Whereas in the previous chapters several methodologies were introduced, we consider
it important to present the developed packages to our readers, helping them choose the right
framework.
We present three packages, RBniCS, ITHACA-FV, and ITHACA-SEM, that perform the
reduced order method in a projection-based manner. These packages have been developed based
on three popular frameworks for solving PDEs: FEniCS for finite element, OpenFOAM for
finite volume, and Nektar++ for spectral element, making it easier to reduce models already
implemented in such frameworks. The other packages are not tightly connected to any particular
library, thanks to the data-driven nature of the implemented methods. In particular, EZyRB and
PyDMD provide implementation for the proper orthogonal decomposition with interpolation
(PODI) and dynamic mode decomposition (DMD) techniques, while PyGeM and ATHENA deal
with geometrical deformations and parameter space reduction, two numerical tools that have
synergy with ROM.
The next sections extend the description of all the packages mentioned above, focusing
mostly on technical details. We refer the reader to the other chapters for theoretical and method-
ological insights. Figure 19.1 provides the logos of the packages, while in Table 19.1 we sum-
marize the main information for each one, presenting a minimal description, the framework to
which it can be coupled, and finally the programming language and compatibility with the main
operating systems.

19.1.1 RBniCS
The RBniCS Project14 contains an implementation in FEniCS of several ROM techniques (and,
in particular, the certified reduced basis method and the proper orthogonal decomposition (POD)-
14 https://www.rbnicsproject.org.

379
 380 Chapter 19. Scientific Software Development and Packages for ROMs in CFD
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 19.1. The logos of the ROM packages.

Table 19.1. Summary table for all the packages.

Name Description HF framework Language OS

ROM framework for fi-
RBniCS FEniCS Python Linux
nite element
ROM framework for fi-
ITHACA-FV OpenFOAM C++ Linux
nite volume
ROM framework for
ITHACA-SEM Nektar++ C++ Linux
spectral element
ROM framework for dis-
ITHACA-DG HopeFOAM C++ Linux
continuous Galerkin
Nonintrusive ROM MacOS,
EZyRB Data-driven Python
framework Linux
DMD framework for dy- MacOS,
PyDMD Data-driven Python
namical system Linux
Parameter space reduc- MacOS,
ATHENA Data-driven Python
tion framework Linux
Geometrical deforma- MacOS,
PyGeM Data-driven Python
tion framework Linux
Parametrization frame-
MacOS,
BladeX work for propeller and Data-driven Python
Linux
rotor
 19.1. Scientific Open-Source Software for Reduced Order Models 381

Galerkin methods) for parametrized problems. It is ideally suited for an introductory course on
reduced basis methods and ROM, thanks to an object-oriented approach and an intuitive and
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

versatile Python interface. To this end, it has been employed in several doctoral courses on
reduced basis methods for computational mechanics.
RBniCS can also be used as a basis for more advanced projects to assess the capability of
ROMs in their existing FEniCS-based software, thanks to the availability of several reduced or-
der methods (such as reduced basis and POD) and algorithms (such as the successive constraint
method and the empirical interpolation method (EIM)) in the library. Several tutorials are pro-
vided.

19.1.2 ITHACA-FV
ITHACA-FV15 [549] (In real Time Highly Advanced Computational Applications for Finite Vol-
umes) is a C++ library based on the finite volume solver OpenFOAM [2]. It implements several
ROM techniques for parametrized problems. The library includes POD-Galerkin [550] methods
with hyperreduction functionalities [551], nonintrusive methods using PODI techniques [209],
and neural network approximations [639].
ITHACA-FV has been used to solve a variety of different problems, including heat trans-
fer problems [551], laminar fluid dynamics problems [548], turbulent fluid dynamics prob-
lems [284], multiphysics problems [220, 553], and geometrically parametrized problems [551].
ITHACA-FV can also be used as a basis for more advanced projects to assess the capability of
ROMs in their existing OpenFOAM-based software, thanks to the availability of several reduced
order methods and algorithms.
ITHACA-FV is designed for OpenFOAM 1812, 1906, 1912, 2006, 2012, 2106, and 2112,
but it can be easily adapted to other versions of OpenFOAM.

19.1.3 ITHACA-SEM
ITHACA-SEM (In real Time Highly Advanced Computational Applications for Spectral Ele-
ment Methods) is a C++ library based on the spectral element solver Nektar++.16 It implements
several ROM techniques for parametrized problems. ITHACA-SEM can also be used as a ba-
sis for more advanced projects to assess the capability of ROMs. ITHACA-SEM periodically
merges in the current master branch of Nektar++ and uses Eigen for matrix decompositions.
Nektar++ and the Eigen header files are part of the ITHACA-SEM package.
The open-source software ITHACA-SEM17 currently has the ability to generate POD-based
ROMs for two-dimensional incompressible Navier–Stokes equations with parametric variation
in geometry and/or kinematic viscosity. The parametric dependency on geometry parameters is
assumed to be affine, and the user can specify the affine form in a header file. After another
compilation of ITHACA-SEM, the affine form is then available. It can thus serve as a head
start for a developer seeking to work in ROMs with the SEM and also to a practitioner within
the boundaries mentioned. A few test cases are part of the Nektar++ unit test, and additional
examples are available.
ITHACA-SEM has been used in numerous publications; see [273, 276, 274, 272] for basic
usage, [270, 275] for localized ROMs, [454] for deflation methods in bifurcation problems, and
[277] for work on sparse polynomial interpolation.

15 Available at https://github.com/mathLab/ITHACA-FV.
16 www.nektar.info.
17 https://mathlab.sissa.it/ITHACA-SEM.
 382 Chapter 19. Scientific Software Development and Packages for ROMs in CFD

19.1.4 ITHACA-DG
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

ITHACA-DG18 is an implementation in HopeFOAM of several ROM techniques. ITHACA-DG

is designed for HopeFOAM 0.1,19 which is based on OpenFOAM 4.0.
Linear and nonlinear algebra operations which are not already implemented in OpenFOAM
are performed with the external library Eigen. The source code of Eigen 3.3.4 is included with
ITHACA-DG and is located in the src/thirdyparty/Eigen folder. For eigenvalue decomposition
it is also possible to rely on the Spectra-0.6.1 library, whose source code is provided in the
src/thirdyparty/spectra-0.6.1 folder.
ITHACA-DG has been tested on ubuntu 18.04 but can be easily compiled on any Linux
distribution with a compiled version of OpenFOAM 4.0.

19.1.5 EZyRB
EZyRB20 [158] is a Python package that performs a data-driven model order reduction for
parametrized problems using recent approaches. We refer the reader to Chapter 9 for a deeper
discussion of these methodologies. Here we summarize by saying that such techniques provide
a parametric model able to approximate the solution of a generic (potentially complex and non-
linear) problem in real time. The reduced model is totally built upon the numerical data obtained
by the original (to be reduced) model, without requiring any knowledge of the equations that
describe this model, resulting in a framework that is well suited for industrial contexts due to
its natural ability to integrate with commercial software. Figure 19.2 shows an approximation
obtained using EZyRB on an parametric automotive benchmark.

Figure 19.2. Example of EZyRB prediction on an automotive benchmark.

EZyRB offers several methods for dimensionality reduction, such as POD and autoencoders.
It also provides a wide range of regressors for manifold approximation, such as linear interpo-
lators, Gaussian processes, and artificial neural networks (ANNs), made accessible through an
18 https://mathlab.sissa.it/ITHACA-DG.
19 https://github.com/HopeFOAM/HopeFOAM.
20 Available at https://github.com/mathLab/EZyRB.
 19.1. Scientific Open-Source Software for Reduced Order Models 383

intuitive interface. In this way, the user can easily test different methods for the problem at hand
and make the best choice, also thanks to the already available error estimators. The package
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

functionalities are illustrated in different tutorials, which together with the online documentation
help users take the first step with this package. EZyRB has been employed in several contribu-
tions, such as parametric heat conduction in an isogeometric analysis setting [213], naval shape
optimization problems [157, 152], hydroacoustic analysis [209], and hemodynamics problems
[234].

19.1.6 PyDMD
PyDMD21 [159] is a Python package that uses DMD for a data-driven model simplification based
on spatiotemporal coherent structures.
DMD is a model reduction algorithm developed by Schmid [527] that we already carefully
treated in Chapter 9. DMD relies only on high-fidelity measurements such as experimental data
and numerical simulations. Its popularity is due to the equation-free nature of the algorithm,
which does not require any assumptions about the underlying system. See [353] for a compre-
hensive overview of the algorithm and its connections to the Koopman-operator analysis, along
with examples in computational fluid dynamics (CFD). In the last years many variants have
arisen, such as multiresolution DMD, compressed DMD, forward-backward DMD, and higher-
order DMD, to deal with noisy data, big datasets, or spurious data e.g.
In PyDMD we implemented the majority of the variants mentioned above with a user-friendly
interface. The package is actively maintained, providing detailed documentation and several
tutorials and examples to show the common use cases, like the one sketched in Figure 19.3.
PyDMD won first prize in the DSWeb 2019 contest Tutorial on Dynamical Systems Software
(Junior Faculty Category). It is also possible to easily extend the package with a new DMD
variant, following a developer tutorial.

Figure 19.3. Example of PyDMD application to a toy problem.

Regarding its usage for scientific publications, we cite its employment for industrial design
in CFD applications [573] and for speeding up numeric simulations of dynamical systems [572,
155] as some recent examples.
21 Available at https://github.com/mathLab/PyDMD.
 384 Chapter 19. Scientific Software Development and Packages for ROMs in CFD

19.1.7 ATHENA
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

ATHENA22 [496] is an open-source Python package for parameter space reduction. It imple-
ments advanced numerical analysis techniques such as active subspaces (ASs) [134], kernel-
based active subspaces (KASs) [493], and nonlinear level-set learning (NLL) [643]. Refer to
Chapter 16 for a deeper description of these methods. The package is intended as a tool for
regression and sensitivity analysis, and in general to enhance existing numerical simulations’
pipelines by tackling the curse of dimensionality. In Figure 19.4, we depict a summary of the
implemented techniques. The package provides a widespread application programming interface
(API) following the scikit-learn conventions.

KERNEL-BASED ACTIVE SUBSPACES ACTIVE SUBSPACES

Feature space
mapping
Active
subspace

NONLINEAR LEVEL-SET LEARNING

Regression

Projection

Active variable

Figure 19.4. Sketch of the techniques implemented in ATHENA for parameter space reduction.

The software has been successfully used for optimization tasks [161, 157], for local active
subspaces (LASs) [495], for multifidelity Gaussian process approximations [494, 497, 575], for
coupling with DMD [573], and to reduce the memory consumption of convolutional neural net-
works [394].

19.1.8 PyGeM
PyGeM23 [570] is a Python package using free-form deformation (FFD), radial basis functions
(RBFs), and inverse distance weighting (IDW) to parametrize and morph complex geometries.
These methodologies, as well as their applications in parametric systems, are discussed in detail
in Chapter 17. The package is organized to work with a large variety of different file formats,
dealing with discrete geometries and computational grids (e.g., STL or VTK files) and analyti-
cal geometries (CAD files). Moreover, we highlight that the capacity of operating over discrete
geometries can be in principle extended to any file format, since PyGeM only needs the coor-
dinates of the points describing the geometry. Such points are moved according to the chosen
deformation, without altering the topology of the input geometry. In this way, PyGeM can be
coupled with any file format, requiring only a way to extract the initial points and to save the de-
formed ones.
Thanks to its properties, this package is very useful for problems in the parametric domain, al-
lowing one not only to morph the studied object—e.g., industrial artifacts like the one illustrated
in Figure 19.5—but also to act directly on the computational grid, so that it is very powerful when
22 Available at https://github.com/mathLab/ATHENA.
23 Available at https://github.com/mathLab/PyGeM.
 19.2. Best Practices for Scientific Programming 385
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 19.5. Example of PyGeM deformation on an automotive test case.

Figure 19.6. Example of blade deformation using BladeX.

coupled to ROM. Documentation and tutorials are available online, helping the user to take the
first steps for using the package or to contribute to it.

19.1.9 BladeX
BladeX24 [210] is a Python package for geometrical parametrization and bottom-up construction
of propeller blades. It allows the user to generate and deform a blade based on the radial distri-
bution of its parameters, such as pitch, rake, and skew, and the sectional foils’ parameters, such
as chord and camber. The package is ideally suited for parametric simulations on a large number
of blade deformations. It provides an automated procedure for computer-aided design (CAD)
generation, reducing the time and effort required for modeling. A graphical example showing
some CAD files obtained with this package is shown in Figure 19.6. The main scope of BladeX
is to deal with propeller blades; however it is flexible enough to be applied to further applications
with analogous geometrical structures, such as aircraft wings, turbomachinery, and wind turbine
blades.

19.2 Best Practices for Scientific Programming

In this section, we move from numerical analysis to software engineering since we consider it
very important to introduce the reader to some best practices to keep in mind during software
24 Available at https://github.com/mathLab/BladeX.
 386 Chapter 19. Scientific Software Development and Packages for ROMs in CFD

development. The aim is to provide young researchers with an introductory perspective about
software engineering, helping them to structure and accelerate the programming phase in their
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

research. If interested, the reader can refer to [545, 203] for a deeper discussion of software
engineering.

19.2.1 Code Quality

Dealing with programs developed for research, readability—the ability of the source code to
be read and understood—of the code is even more important than for generic programming.
Typically, research software offers implementations of state-of-the-art methodologies, meaning
that the algorithmic core could change a lot during development. For this reason, researchers
need a simple and understandable interface, such that changing the code structure is an easy
and quick task. Beyond readability, another key point to reach this goal is code modularity,
indicating how independent and interchangeable the code components are. These aspects are
two of the indicators usually taken into consideration during software design; we avoid a deeper
focus on them, providing instead a few considerations about the reasons for their importance
during the implementation of state-of-the-art methods.
The first development step of a scientific code is primarily the validation of the algorithm at
hand, with a limited effort given to the design of the software. Of course, a dedicated analysis
results most of the time in a too-demanding procedure that does not fit the timing of academia
well, but a minimal effort in the coding style can greatly improve both the readability and the
modularity of the code. Taking care of these aspects, especially during the initial design, can
mitigate the presence of bugs during the entire life of the software and make future extensions
or changes easier. In this sense, the initial effort becomes a time investment, both for software
development and the research line related to it.
On the practical side, we list in this section the main features to take into account during the
development of scientific code, without a detailed discussion about code structure and style.

• Use functions: Rather than thinking about the source code as a (long) list of instructions,
try to semantically split it into minimal (and possibly hierarchical) pieces of code. For any
of these snippets, separate the expected input and output in order to create the functions.
The main benefit is the reusability of the code and the reduced number of duplicated lines,
but such splitting also helps readability, since several technical (and probably difficult to
understand) parts can be encapsulated in functions.

• Keep it simple: As already mentioned, scientific code should be easy to understand in or-
der to enable future methodological enhancements. Simplicity of the code is an extremely
nice feature, which we suggest preserving by avoiding overcomplicated structures, even
with minimal loss of efficiency. Production software of course needs a different approach,
but simplicity is much more important for collaborative and dissemination purposes—in
the first steps of scientific software development.

• Source code should be self-informative: Rather than adding a comment beside every
line of code, we suggest using meaningful names for variables, constants, functions, and
classes, such that reading the source code allows the user in principle to understand the
algorithm. Comments are of course very welcome in any snippet of code, but we suggest
using them to explain functions. Commenting on all the lines generally makes the source
more cryptic.

• Follow the standards: Especially in team development, the use of code standards helps
make the code uniform and, consequently, improves its readability. Several conventions
 19.2. Best Practices for Scientific Programming 387

are available for any language, and their use is widespread. This aspect becomes even
more important when you plan to release the source code to allow new developers to easily
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

contribute.

As a final remark, we add that in any case the fundamental point to keep in mind during de-
velopment is common sense, probably the most important of the aspects mentioned above. The
developer has to proceed in the coding always thinking about possible extensions and refactor-
izations, trying to select the design that maximizes the future use of the code.

19.2.2 Testing
Producing data to be published in scientific journals means that the reliability of the software
is of course another important feature. Things can become very frustrating when you discover
that your novel implementation returns awesome results for a specific problem but is not able to
deal with well-known benchmarks. It doesn’t matter if this issue is related to a coding bug or
a methodological error, its resolution can cost several hours (or even days): high-quality source
code definitely helps to detect the issue, but it still results in a demanding procedure. To overcome
these problems, we suggest planning a testing procedure starting at the first development stage.
Intuitively, the procedure is composed of a collection of additional snippets of code that check
that all the implemented functions return, for a specific input, the expected output. If these tests
have already been developed, it is easy to run them after any modifications of the main code to
be sure no errors have been introduced in the source code.

19.2.3 Documentation
Now that we have discussed the structure of the code and its testability, we focus in this last
subsection on the importance of having and maintaining informative documentation for the code.
This becomes a key feature when the source code is shared with other scientists and developers
to improve its understandability. Practically, documentation is a manual that illustrates all the
details of the source code, such as the expected input and output of any functions, as well as
a guide for code employment for final users. Despite its usefulness, it is quite common to see
projects that lack this resource. It must be said that of course preparing such a document is a time-
consuming procedure, but by using some frameworks we can at least shorten the process. We
suggest postponing the creation of documentation until clean source code has been implemented
with the corresponding tests. Documentation helps new developers and users to get oriented
with the source code. It explains the code at its current status and its possible applications.
Even if writing everything from scratch is an option, the most efficient choice for creating the
documentation is usually to extract it from the comments of the code, the docstring. With the
aim of illustrating how any function works, such comments are usually placed near the function
prototype: in this way, the documentation can be manipulated by simply editing the source
code, turning it into a less onerous activity. Several frameworks are available, depending on
the language used, for this operation, giving the developers the ability to make aesthetic changes.
Among others we cite Sphinx25 and Doxygen.26
As already mentioned, the documentation should not be limited to a summary of all the func-
tions implemented in the source code. To increase the software dissemination, and consequently
also the dissemination of the algorithm implemented in the code, the developers should pro-
vide different use cases with the most common applications, such that new users can easily and
quickly understand the typical usage.
25 https://www.sphinx-doc.org/.
26 https://www.doxygen.nl/.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Chapter 20

A Deep Learning
Approach to Improving
Reduced Order Models

Laura Meneghetti, Nirav Shah, Michele Girfoglio, Nicola Demo,

Marco Tezzele, Andrea Lario, Giovanni Stabile, Gianluigi Rozza

20.1 Introduction
This chapter is devoted to describing different approaches for reduced order modeling (ROM)
obtained by combining innovative deep learning techniques, artificial neural networks (ANNs),
and digital twins.
Deep learning is a thriving field with many practical applications and active research topics
[237]. It is widely used to tackle and solve difficult tasks, such as visual object recognition
[350, 130], speech recognition [243], natural language processing (NLP) [633], autonomous
vehicles [295, 306], digitalization [266], and regression problems [51, 547]. ANNs have achieved
impressive results in dealing with such problems thanks to their deep structures and their ability
to approximate functions.
Some of the aforementioned fields represent great challenges in academic research but also
in the industrial sector. In order to face the complexity of such topics, neural network (NN)
architectures are forced to go deep, leading to a high number of parameters to be calibrated.
In this case, developing a reduction technique based on methods widely used in ROM, such
as proper orthogonal decomposition (POD) and active subspaces (ASs), is a good solution for
overcoming the typical constraints of embedded systems [394, 142]. On the other hand, ROM is
applied to digital twins in order to obtain a better product by providing more accurate information
about the real counterpart.
A classical result in this context, the universal approximation theorem for neural networks
[143, 293], states that sufficiently wide (shallow) NNs can approximate any (continuous) func-
tion. This opened the possibility of applying ANNs to a wide range of fields and in particular to
ROM and inverse problems involving nonlinear partial differential equations (PDEs).
Therefore, in this chapter, we start by outlining in Sections 20.2 and in 20.3 the most im-
portant features of ANNs and of the usual NN architectures. This part is then followed by
Subsection 20.4.1, where a detailed description of the POD-neural network (NN) approach is
provided: the use of ANNs to approximate the functional relationship between input parame-
ters and the output solution field in ROM. Subsection 20.4.2 is then devoted to presenting the

389
 390 Chapter 20. A Deep Learning Approach to Improving Reduced Order Models

NN-shifted POD (NNsPOD) algorithm, a nonintrusive, data-driven machine learning method.

After this part, we present in Section 20.5 physics-informed neural networks (PINNs), a deep
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

learning framework for solving forward and inverse problems involving nonlinear PDEs. In Sec-
tion 20.6 we describe a technique to develop a reduced version of an ANN in order to overcome
the constraint of memory storage usually connected with embedded systems. Finally, we report
in Section 20.7 the importance of digital twins in modeling and simulations, with a focus on the
role of ROMs in this context.

20.2 ANNs
We start by briefly introducing the main notions about ANNs [237, 348, 92].
The structure of ANNs is inspired by that of the human brain, mimicking the way that bio-
logical neurons signal to one another. They are characterized by an input layer, a certain number
of hidden layers,27 and an output layer, as shown in Figure 20.1. Usually, each neuron in a layer
is connected to each neuron of the next layer, and the strength of this connection is described by
a parameter called the weight. It is also important to highlight that, as explained in Section 20.3
and depicted in Figure 20.1, not only do ANN inputs flow in one direction (from the input to the
output layer), but backward connections and loops may exist.

Input Hidden Hidden Output

layer layer 1 layer 2 layer

x1

.. ŷ1
x2

.. .. . ..
. . . ŷnout
xnin

Figure 20.1. Schematic structure of an ANN.

More technically speaking, let x ∈ Rnin be the input vector and L the total number of hidden
layers of the ANN. The output vector ŷ ∈ Rnout is obtained by applying an activation function
to the weighted sum of all the inputs arriving at the output layer. The activation function is used
to introduce nonlinearity in the network, and some common choices are the ReLu function, the
sigmoid, the logistic function, and the radial activation functions [348, 92].
In order to better understand how to get the general formula (20.2) for computing NN output,
we start by considering a simple ANN with a single output and one hidden layer. In this case the
final output is given by
nin
!
X
ŷ = σ wi xi + bi , (20.1)
i=1
27 A priori there is no right number of hidden layers to use: it depends on the field of application of your net and on

the problem at hand [583].

 20.2. ANNs 391

where σ is the activation function, w ∈ Rnin represents the weights of the net, and b the bias.28
Thus, if we consider L layers, the final output can be seen as a weighted sum of its inputs
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

followed by the activation function, where each input can be rewritten in the same way:
nL nL nL−1
! !!!
(L+1) (L) (L+1)
X X X (L)
(L−1)
ŷj = σ wji yi =σ wji σ wim ym
i=1 i=1 m=1
nL nL−1 nin
!!!!!!!
(L+1)
X X (L) X (1)
= ... = σ wji σ wim σ . . . σ wsk xk ,
i=1 m=1 k=1
(20.2)
where nl , l = 0, . . . , L, represents the number of neurons29 in layer l; y (l) ∈ Rnl the output of
(l)
layer l for l = 1, . . . , L; and W l = (wji )ji , j = 1, . . . , nl , i = 1, . . . , nl−1 the weight matrix
related to layer l. Note that the first number in any weight’s subscript matches the index of the
neuron in the next layer and the second number matches the index of the neuron in the previous
layer.
Two important concepts are connected to NNs:

• forward propagation and

• backward propagation or backpropagation.

With the term forward propagation [92] we refer to the process we have described before, where
starting from some inputs we compute the outputs using (20.2). Usually the weights and the
other parameters of the network are randomly initialized. Thus, obviously we could not expect
to obtain the right results at the end of the forward pass. In order to adjust these parameters to gain
better results, a loss function and the backpropagation algorithm [490, 92] have to be introduced
to take into account the error made in predicting the outputs using the net with respect to the true
outputs.
Let {xi , yi }ni=1 be a set of n observations and L = L(y, f (x)) a loss function, where the ANN
in consideration is summarized as a generic function ŷ = f (x; W ) of the inputs and weights. The
problem that needs to be solved is a minimization:
n1 Xn o
min L(yi , f (xi ; W )) , (20.3)
W n i=1

where L is chosen based on the problem to be solved [15, 237]. A common choice is the mean
squared error (MSE) function, defined by
n n
1X 1X i
L({yi }, {ŷi }) = Li = (y − ŷi )2 , (20.4)
n i=1 n i=1

where Li = Li (yi , ŷi ), yi represents the expected output, and ŷi is the prediction made by our
ANN.
Another possibility is the mean absolute error, which measures the mean absolute value of
the elementwise difference between inputs:
n n
1X 1X i
L({yi }, {ŷi })) = Li = y − ŷi . (20.5)
n i=1 n i=1
28 For simplicity the bias is set to zero in the following discussion.
29 Note that for l = 0 we are considering the input vector and thus n0 = nin .
 392 Chapter 20. A Deep Learning Approach to Improving Reduced Order Models

In order to solve the minimization problem (20.3), the gradient method is usually employed
[92, 490]. In this context, this method is also referred to as backpropagation because the gap
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

recorded at a certain data point through the loss function L is propagated backward in the network
to obtain the update formulas of the coefficients by computing the gradient of L. Computing
an analytical expression for the gradient is straightforward, but numerically evaluating such an
expression can be computationally expensive. Therefore, the backpropagation algorithm does so
using a simple and inexpensive procedure that we will now explain.
Given an observation (x, y), the backpropagation algorithm initially takes a step forward
along the entire network, propagating information along all the layers (feedforward step or for-
ward propagation). After this step, the loss function L is computed using the predicted output
ŷ and the expected output y. It is important to highlight that in this first step the weights are
randomly initialized and then adjusted during the backward pass.

Figure 20.2. The forward pass on the left calculates y as a function f (x1 , x2 ) using the input
variables x1 and x2 . The right side of the figure shows the backward pass. Receiving dL/dy, the gradient
of the loss function with respect to y from above, the gradients of x1 and x2 on the loss function can be
calculated by applying the chain rule, as shown in the figure.

In order to find the minimum that satisfies problem (20.3), the gradient of L needs to be
computed. Since the loss function depends on the output of the net ŷ = f (x; W ) and thus on the
architecture of the ANN itself represented by f , the chain rule needs to be exploited, as depicted
(l)
in Figure 20.2. Therefore, for a single weight wjk , with l = 1, . . . , L + 1, the gradient is given
by
(l) (l)
dL dL dyj dzj
(l)
= (l) (l) (l)
. (20.6)
dwjk dyj dzj dwjk

By using the explicit version of relation (20.2) written in components,

nl−1
!
(l) (l) (l−1) (l)
X
yj =σ wjk yk = σ(zj ), (20.7)
k=1

(20.6) becomes
dL dL (l) (l−1)
(l)
= (l)
σ 0 (zj )yk , (20.8)
dwjk dyj

where dL
(l) is usually defined as the local gradient δjl . The computation of the gradients then
dyj
 20.3. NN Topologies 393

allows us to optimize the model’s parameters by adjusting the network’s weights as follows
[490, 92]:
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

(m),t (m),t−1 dL
wki = wki −  (m) , (20.9)
dwki
where  is the learning rate and is appropriately chosen according to the problem at hand; t
represents the training epoch under consideration, i.e., a time interval during which all training
examples are employed to update the network parameters.
The process described up to now is used in the training phase, where we take part of the
complete data set (usually 80%) to fit the parameters of the model. The remaining part is used in
the testing phase to evaluate the final performance of the network.

20.3 NN Topologies
We provide in this section an overview of the usual NN topologies [348, 370, 92], i.e., the con-
nections between neurons of different layers that define different types of ANN architectures.

20.3.1 Feedforward Neural Network

The feedforward neural network (FNN), also called the multilayer perceptron, is the simplest
form of NN model, usually employed for function regression [197, 399, 498]. Its architecture is
similar to that of ANN and is depicted in Figure 20.3, with the difference that input data travels in
one direction only from left to right, passing through artificial neural nodes of the several layers
and exiting through the output nodes.

Input Hidden Hidden Output

layer layer 1 layer 2 layer

x1

.. ŷ1
x2

.. .. . ..
. . . ŷnout
xnin

Figure 20.3. Schematic structure of an FNN.

20.3.2 Recurrent Neural Network

A recurrent neural network (RNN) is a type of ANN characterized by a memory, where infor-
mation from previous states y (T −1) provides an outcome and an input to the next state y (T ) ,
with T being the state we are considering [182, 392]. As represented in Figure 20.4, in this
case, each hidden layer becomes an RNN layer, where its output depends on the prior elements
within the sequence because each of the hidden cells receives its output back with a fixed delay.
 394 Chapter 20. A Deep Learning Approach to Improving Reduced Order Models

Input Recurrent Recurrent Output

layer layer 1 layer 2 layer
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

x1

.. ŷ1
x2

.. .. . ..
. . . ŷnout
xnin

Figure 20.4. Schematic structure of an RNN.

convolutional convolutional output

block 1 block 3 layer

convolutional fully connected

input image
block 2 layer

Figure 20.5. Schematic structure of a CNN.

Hence, these algorithms are commonly used for ordinal or temporal problems, such as language
translation, NLP, and speech recognition.

20.3.3 Convolutional Neural Network

A convolutional neural network (CNN) is a deep learning algorithm used to solve complex prob-
lems, such as image recognition, speech recognition, and computer vision [15, 366]. As depicted
in Figure 20.5, a CNN is a combination of two basic building blocks: the convolutional blocks
and an FNN.
The convolutional block forms the essential component of feature extraction. It consists of
the following features:

• It has a convolutional layer followed by a nonlinear function, such as ReLU, sigmoid, or

tanh. As shown in Figure 20.6, the convolution step is computed by sliding a filter over
an image by s pixels (also called the stride) for every position, computing elementwise
 20.3. NN Topologies 395

input image
output feature maps
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

or input feature map

Figure 20.6. Illustration of a single convolutional layer.

multiplication (between the two matrices), and adding the multiplication outputs to get the
final integer, which forms a single element of the output matrix. In CNN terminology, the
output matrix is called a feature map and its size is controlled by three parameters that
have to be decided before the convolution step is performed: the depth D, i.e., the number
of filters that will also determine the number of output images at the end of the convolution
step; the stride S, i.e., the number of pixels by which the filter matrix is slid over the input
matrix; and the zero-padding P , i.e. adding zeros around the border of a matrix in order to
apply the filter to bordering elements of our input image matrix.
Therefore, given as input a matrix x with size W1 × H1 × D1 and D F × F filters ω, the
convolutional operation will output a W2 × H2 × D2 matrix, where
(W1 − F + 2P ) (H1 − F + 2P )
W2 = + 1, H2 = + 1, D2 = D. (20.10)
S S
In particular, the convolutional output will be obtained by summing up the components of
the input image weighted by the filter components:
D1 X
X F X
F
z(ij) = ω0 + ω(abc) x(i+a−1,j+b−1,c) ,
c=1 a=1 b=1
(20.11)
y(ij) = σ(z(ij) ),
where ω0 represents the bias.
• It has a pooling (or subsampling) layer. Its role is to downsample the output of a convo-
lution layer along both of the spatial dimensions of height and width. As can be seen in
Figure 20.7, if layer l is a pooling layer, the m(l−1) feature maps of the previous layer are
pooled individually by defining a fixed sliding window F × F and considering the average
or maximum within it. In other words, each unit in one of the pooled feature maps in layer
l represents the average or the maximum within the fixed window of the corresponding
feature map in layer (l − 1). Hence, given a feature map with dimensions W1 × H1 × D1
and a fixed filter F × F moving with stride s, the final dimensions of the pooled feature
map W2 × H2 × D2 are determined by
(W1 − F ) (H1 − F )
W2 = + 1, H2 = + 1, D2 = D1 . (20.12)
S S
As described before, after the feature extraction part, an FNN performs the task of classification
based on the input from the last convolutional block.
 396 Chapter 20. A Deep Learning Approach to Improving Reduced Order Models

feature maps feature maps

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

layer (l − 1) layer l

Figure 20.7. Illustration of a pooling layer.

20.3.4 Autoencoder
An autoencoder (AE) is used for classification, clustering, and feature compression [237, 288,
528, 367]. It mainly consists of three components: encoder, code, and decoder. Therefore, to
build an AE we need three things: an encoding method, a decoding method, and a loss function
to compare the output with the target.
More technically speaking, both the encoder and the decoder are fully connected FNNs com-
posed of a sequence of fully connected layers plus the bias followed by a nonlinear function.
As shown in Figure 20.8, first the input x passes through the encoder to produce the lower-
dimensional code z, i.e., the AE learns in this way a dimensionality reduction. Then, the decoder,
which is typically the mirror image of the encoder, maps the code to a reconstruction of the input
x̃. Note that the dimensionalities of the input and output need to be the same. To obtain an out-
put almost identical to the input, AEs are trained the same way as ANNs via backpropagation,

input output
encoder decoder

Figure 20.8. Schematic structure of an AE.

 20.4. NNs to Improve POD-Based Models 397

minimizing the reconstruction loss L(x̃, x) (e.g., cross-entropy loss or MSE) that quantifies the
quality of the reconstruction x̃.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

20.4 NNs to Improve POD-Based Models

20.4.1 The POD-ANN Approach
Conventional model order reduction methods have focused on solving a smaller system of equa-
tions to obtain the degrees of freedom of the reduced basis solution. Recently, the focus of model
order reduction approaches has shifted to deep learning methods, which extract the information
from data and approximate the functional relationship between input parameters and the output
solution field [279, 522] using an ANN. The application of ANNs for approximating functions
is primarily based on the universal approximator property of ANNs [143, 293]. An ANN is a
training network inspired by the massive, hierarchical NN of the human brain. The evolution of
ANNs and their similarity to NNs in the human brain is explained in the literature [267, 491].
Let Vh = span{φ1,h , . . . , φnh ,h } be the space of high-fidelity solutions and uh (Ξ) ∈ Vh be
the high-fidelity solution at a given parameter tuple Ξ from the d-dimensional parameter space
P, i.e., Ξ ∈ P ⊂ Rd . Let Vrb rb rb
h = span{φ1,h , . . . , φNh ,h } ⊂ Vh be the space of the reduced
basis solution constructed with POD and uh (Ξ) ∈ Vrb
rb
h be the reduced basis approximation of
uh (Ξ) ∈ Vh :
Nh
urb ûrb rb
X
h = i,h φi,h , Nh  nh . (20.13)
i=1

We consider here that urb rb

h (Ξ) is the projection of uh (Ξ) on Vh and ζrb (Ξ) is the vector of corre-
sponding degrees of freedom:
ζrb = {ûrb Nh
i,h }i=1 . (20.14)

In this context, we use FNN to learn the map f : Rd → RNh , which approximates the reduced
basis degrees of freedom at a given parameter tuple Ξ:

f (Ξ) = ζNN (Ξ) ≈ ζrb (Ξ), ζNN = {ûNN Nh

j,h }j=1 . (20.15)

In order to create a training set for FNN, samples from parameter space P are drawn and
split into two sets: the training set {Ξi }ni=1tr
and the validation set {Ξj }nj=1
val
. At each sample
Ξi , the high-fidelity solution field uh (Ξi ) ∈ Vh and its projection urb
h (Ξi ) ∈ V rb
h are computed.
Hence, a set of known input-output pairs {Ξi , ζrb (Ξi )} is generated for training. Similarly, a
known set of input-output pairs {Ξj , ζrb (Ξj )} is generated for validation. FNN is trained to
minimize the loss function L(ζrb (Ξi ), ζNN (Ξi )) over the training set using backpropagation (see
Section 20.1). Since backpropagation is an iterative approach, it is important to ensure that
the initial guess of weights and biases does not inadvertently affect the outcome at the end of
backpropagation. Also, it should be noted that a larger weight assigned to a particular synaptic
connection may compromise the robustness of the method. The accuracy of the network can be
tested by sampling the parameter space P and measuring the error, in a suitable norm, between
the high-fidelity solution field and the output computed by the POD-ANN method.
Training an ANNs is an expensive procedure due to the trial-and-error approach used to
adjust its hyperparameters. During training, the following steps occur:

• A high-fidelity solution is computed at each parameter tuple of the sampled parameter

space.
 398 Chapter 20. A Deep Learning Approach to Improving Reduced Order Models

• Data for training the ANN is prepared by projecting high-fidelity solutions onto the re-
duced basis space.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

• The ANN needs to be trained for each new configuration.

The higher offline cost, due to the training procedure, is disadvantageous for ANN-based meth-
ods as compared to the projection-based methods. One solution to this problem is to extract more
information from the high-fidelity solution. This can be done by modifying the loss function: loss
functions utilizing residuals can produce more accurate predictions [119]. Higher offline cost is
often associated with the complexity of the problem, which the ANN is required to learn. An
alternative to reduce this complexity is to assign an input-output map and use the ANN to de-
termine the parameters of this map. For example, fitting a linear regression line to approximate
the input-output map will require an ANN without any hidden layers. However, for real-world
applications, the underlying input-output relationship is complex, so such simple models may
have very limited applications.
Another important component of the training of the ANN is to avoid “overfitting.” Due to
overfitting, a trained ANN is accurate for predictions corresponding to the parameter tuples used
for training but fails to accurately predict output for parameters outside the training set. To avoid
overfitting, regularization techniques such as L2 regularization and early stopping criteria can be
used.
It is also pertinent to mention that the sampling strategy for generating training data for
an ANN should be chosen carefully. A good sampling strategy aims to take samples that are
distributed evenly over the parameter space. It needs to be noted that the sampling strategy used
for POD might not be the same as the sampling strategy used to create a training set. This is
mainly due to two reasons:

1. While POD aims to capture only dominant modes, an ANN tries to capture all patterns.

2. Often, the training set for an ANN is larger than the one used for POD.

During the online phase, the ANN only requires feedforward operation at a given parameter
tuple Ξ using (20.2). In the case of projection-based methods, the online time is higher due to the
time taken to assemble system equations if the online phase and offline phase are not completely
decoupled. Higher online time can also occur in coupled systems, where computing one solution
field requires solving more than one equation, or in the presence of too many geometric param-
eters [536]. The POD-ANN approach is a particular case of nonintrusive data-driven methods
(Chapter 9); i.e., system matrices corresponding to a discretized form of governing equations
are not needed. The nonintrusive nature of the method simplifies the interface between different
software libraries, such as the one used to compute high-fidelity solutions and the one used for an
ANN. It also needs to be noted that adding information about the governing equation certainly
helps to increase accuracy [119].
We now provide offline-online decomposition for the POD-ANN approach.

• Offline phase:

1. Construct the reduced basis space Vrb

h using POD.
2. Sample the parameter space P to construct subset {Ξi }ni=1 ⊂ P and split the set to
create a training set and a validation set.
3. At each parameter tuple, compute uh (Ξi ) ∈ Vh and its projection urb rb
h (Ξi ) ∈ Vh .
4. Train and validate the ANN against computed solutions {urb h (Ξi )}i=1 using back-
n

propagation and regularization to adjust the hyperparameters (weights and biases).

 20.4. NNs to Improve POD-Based Models 399

5. Repeat steps 2 to 4 for a different number of hidden layers, width of each hidden
layer, and number of sampled parameter tuples n to find the optimal configuration of
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

the ANN.

• Online phase:

1. For any given parameter tuple Ξ ∈ P, perform a forward pass using (20.2).

20.4.2 The NNsPOD Approach

NNs offer a different strategy to improve the effectiveness of the POD-based model. The ef-
ficiency of the latter stems directly from the number of basis functions needed to encode the
parametric structures of a model at reduced order. The reduction accuracy of a model has tra-
ditionally been quantified by the Kolmogorov n-width of the problem [419]. Problems that are
characterized by a fast Kolmogorov n-width decay are easily restricted by a low-rank linear
subspace approximation; on the other hand, a slow decay indicates that a linear approximation
assumption might be inappropriate and inaccurate for that problem.
Thus the reduction of these problems—e.g., problems with dominant advection—by linear
projection onto the POD space requires a large number R of basis functions to reach a high accu-
racy, nullifying the reduction in computational cost. Since models with dominant advection are
particularly recurrent in fluid dynamics, being associated with the conservation laws describing a
large multitude of phenomena (most notably Euler’s equations and Riemann’s problems, shallow
water equations, and multiphase models), a growing number of endeavors in recent years have
proposed alternative modifications to improve the performance of the approximation of such
manifolds. Among all the different proposed techniques, such as [444, 442, 367], we highlight
the shifted POD (sPOD) [484]. This technique aims to preprocess the snapshots by spatially
translating them, typically according to some preliminary knowledge about the problem, such
that the sequential linear approximation of the solution manifold results are more accurate. For-
mally, we can define the shifting operator Tb acting on the field u(x, t) as

ũ(x, t) = Tb u(x, t) = u(x − bt, t) , x, b ∈ Ω, t ∈ T, (20.16)

where ũ(x, t) is the shifted field and Ω and T are the spatial and temporal domains where the field
is defined. A visual example computed on a simple advection problem is depicted in Figure 20.9,
showing the benefits of the shifting paradigm in terms of the dimensions of the reduced space.
Machine learning extends this framework in order to generalize the method. The choice of
the translation to apply to all the snapshots is not always trivial and requires complete knowledge
of the problem at hand. A possible alternative is learning this shifting using only the collected
snapshots in an NN framework. This is the main idea of the neural network shifted proper
orthogonal decomposition (NNsPOD) [435, 436] algorithm, which is a fully nonintrusive, purely
data-driven machine learning method that seeks an optimal mapping of the various snapshots in
the low-rank linear subspace to a reference configuration via an automatic detection that does not
depend on the physical properties of the model. In this framework, the shifting operator becomes

ũ(x, t) = Tb u(x, t) = u(x − b(x, t), t), x, b ∈ Ω, t ∈ T, (20.17)

where the translation of the field is not limited by a fixed value linearly depending on the time as
in (20.16) but depends on spatial and temporal coordinates. The shifting quantity b : Ω × T → Ω
is approximated using an FNN. The learning process is built upon the set of snapshots—the
discrete field—defined as
{V 3 u(ti )}N Ns
i=1 = {u(V, ti )}i=1 ,
s
(20.18)
 400 Chapter 20. A Deep Learning Approach to Improving Reduced Order Models
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 20.9. Example of POD enhancement thanks to the shifting of the snapshots. A Gaussian
impulse advancing in time is a simple advection problem which shows the limitations of POD in these
contexts: the singular values obtained by the decomposition of the snapshots are not able to individuate
any reduced dimension (top). By shifting the snapshots, POD is able to detect the dimension needed to
linearly represent the solution manifold, as highlighted by the single nonzero singular value (bottom).

where ti are the Ns time steps where the corresponding snapshots have been collected and V
is the discretized spatial domain. The main idea is to transform these snapshots by minimizing
the Euclidean distance between them and a chosen reference configuration uref —e.g., the initial
condition, one of the snapshots. A graphical example showing a two-dimensional field in an
advection problem is presented in Figure 20.10: one snapshot translates toward the reference
configuration during the training of the network, demonstrating the ability to automatically detect
the optimal transformation. The loss function imposed during the learning process is therefore

Ns
1 X
J = kuref (x, ti ) − u(x − b(x, ti ), ti )k2 , x ∈ V. (20.19)
Ns i=1

Figure 20.10. Example of NNsPOD applied to an advection problem in a two-dimensional do-

main. The green lines represent the contour levels of the field, the blue ones refer to one of the snapshots of
the database, and the red ones refer to the same snapshot after the shift. The plots at incremental epochs
show that the network is able to self-learn the optimal transformation.
 20.5. A PINN for Solving Parametric PDEs 401

In a discrete context, it is not guaranteed that x − b(x, t) still belongs to the discrete space
V, requiring an additional step to compute the mean distance between the snapshots and the
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

reference configuration. To evaluate any shifted snapshots in the correct space, an additional NN
is employed to generate a continuous approximation of the reference configuration, such that it
is possible to evaluate it in the shifted coordinates. This methodology then relies on two different
networks:

• the ShiftNet computes the shifting quantity b(x, t) to obtain the new spatial coordinates of
the shifted snapshots;

• the InterpNet evaluates the reference configuration in the coordinates of the shifted snap-
shots to compute the distance.

The NNsPOD method results in a data-driven enhancement of the POD method thanks to the
combined employment of a complex neural architecture, overcoming several well-known issues
in the ROM community caused by advection.

20.5 A PINN for Solving Parametric PDEs

NNs are a powerful tool, rapidly and massively employed in many different fields, and of course
not only for enhancing some consolidated framework. In the field of PDEs, they represent a
growing approach, especially for resolving complex problems. Their ability to approximate non-
linear functions allows one to fight the limitations of linear methods for the numerical solution of
PDEs. In the last years, among the different approaches proposed in the literature, the physics-
informed neural network (PINN) [480] is a topic with growing impact. This technique consists
of a generic network whose training procedure is based on the physical equation(s) describing
the problem at hand. Let us assume, e.g., that we are looking for a numerical approximation of
the scalar field u : Ω → R defined on some space Ω ∈ Rd . The field is described by a generic
PDE, with some boundary conditions.
PINN creates a network g : Ω → R with d input neurons and a single output neuron,30
which aims to approximate the unknown field g(x) ≈ u(x) with x ∈ Ω. To achieve the goal, the
physical residual is defined as the error obtained by replacing the (unknown) field with g(x) in
the physical equation and evaluating it at several points of the domain. The natural derivability
of NNs allows one to compute the output derivative with respect to the input, a property which
is exploited in order to also treat the differential terms with this approach. In a similar way, an
additional residual is employed to measure the error on the boundary conditions; in this latter
case, of course, the sample points, where we evaluate the residual, belong to the boundary of
the domain. The sum of these residuals becomes the loss function we want to minimize during
the learning procedure: epoch after epoch, the PINN approximation converges to the unknown
field in order to respect the physical equations and the boundary conditions. A proper number of
neurons and a proper number of sample points are of course required to obtain the desired accu-
racy in the final approximation. NNs thus result in an innovative and potentially very impactful
nonlinear framework for PDE resolution. A deep discussion of theoretical details is beyond the
scope of this section; the interested reader can refer to [603, 90] for a complete and concise anal-
ysis of such method. To conclude our introduction, we highlight that, even if in the example we
present a simple case with a spatially dependent scalar field, the extension to time-dependent and
parametric problems with vector output is trivial and affects mainly the computational side.
Thanks to its generality, PINN is also an alternative tool for parametric models in many
contexts [604, 382, 254, 153]. Time and parameters can be seen as additional inputs of the
30 since we are dealing with an input domain with dimension d and a scalar field.
 402 Chapter 20. A Deep Learning Approach to Improving Reduced Order Models
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 20.11. Example of a PINN solution for a Poisson problem with parametric forcing term.
The two rows show the field of interest for two different parametric configurations, while in the columns the
analytical solution, the PINN approximation, and the absolute error are sketched (from left to right).

network, and their relation to the unknown field is learned by providing the equations describing
the temporal and parametric dependencies. A simple graphical example is shown in Figure 20.11,
comparing the PINN solution of a Poisson problem with a parametric forcing term against the
analytical solution.
In this case, the higher complexity due to the increased number of dimensions translates into
a higher computational cost to train the model. After the training, the network becomes a real-
time model able to approximate the output of interest for different time steps and parameters.
Even if it cannot be considered a reduced order method, it has an algorithmic structure similar
to the methods proposed in the previous chapters of this book: the learning step requires almost
all the computational cost, as typically happens for the offline phase. The network evaluation,
which means the field computation, is instead in real time, as in the large variety of ROMs.
The computational resources needed to train the NN make its application to parametric linear
problems quite inefficient compared to the consolidated techniques, but its nonlinear nature could
offer a valid approach to addressing more complex problems in the parametric context.

20.6 A Reduced Approach for CNNs

In the last decades, NNs have become a popular machine learning technique, increasingly ex-
ploited in several scientific and engineering fields, such as computer vision [350], NLP [633],
robotics [417], and speech recognition [243]. The continuous demand for application to more
complex tasks has forced the architecture to go deep. This has led to an increasing number of
parameters that need to be calibrated during the training phase, in terms of both the number of
layers (depth) or the number of neurons [338] and, on a practical side, the bigger structure to
manage. Therefore, the use of deep NNs, such as CNNs, in embedded systems [499] revealed
several computational issues since their limited hardware imposes memory constraints on the
architectures.
 20.6. A Reduced Approach for CNNs 403

In this section, we propose strategies for the development of a reduced version of an ANN
and in particular of a CNN [394]. We will then present the results obtained using this reduced
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

net, compared with the ones of the original network.

20.6.1 Reduction Strategies for CNNs

The application of deep NNs in several engineering fields and in particular in embedded systems
with particular space constraints led to the need for methods to compress and speed up the in-
ference in CNNs. We provide an overview of some of the existing techniques to enhance the
efficiency of CNNs regarding memory footprint and computation time [499]:

• Network pruning [263, 377]: The main goal of these approaches is to prune redundant
channels in the weight matrices of a trained net and hence accelerate and compress the
whole model. There exist different sparsity constraints to select which channels have to be
pruned: the sum of the absolute value of weights or the average percentage of zeros. De-
spite the fact that these methods achieve parameter sparsity in the model, they are usually
iterative procedures that require a fine-tuning of the parameters, leading to a computation-
ally expensive method for deep NNs and for some applications.

• Low-rank factorization [516, 426]: In order to estimate the informative parameters of

deep CNNs, the matrix/tensor decomposition is used. In this approach, a convolutional
layer is replaced with several smaller convolutional layers applied sequentially, achieving
a much lower total computational cost. However, a decent amount of retraining is needed
to achieve convergence when compared to the original model.

• Parameter quantization [140]: This technique consists of substituting the floating points
with integers inside the network. The created network requires less memory, and, on cer-
tain hardware, the calculations are faster. On the other hand, this can lead to a significant
loss in accuracy, because there is no guarantee that approximating the parameters by sac-
rificing precision for a compressed representation will not compromise the final results.

As described, most of the existing methods do not change the network architectures but only
delete and change model parameters directly. Thus, starting from the idea explored in [142], we
propose a general framework for constructing a reduced version of an (already trained) ANN,
and in particular of a CNN, based on techniques widely adopted in the model order reduction
community, such as POD, described in Section 1.1, and the AS method, presented in Chapter 16.
The reduced network, described in Figure 20.12 and in [394], is constructed by retaining a certain
number of layers of the original ANN and replacing the remaining ones with an input-output
mapping. We are, thus, splitting the net into two parts connected by the reduced method, which
helps reduce the typically large dimensions of the intermediate layers by keeping only the most
important information. In this way, we are performing a smart selection of the main parameters
of the network, allowing us to reduce the required resources and the computing time to infer the
model.

20.6.2 Reduced ANNs

We provide in this section a rigorous description of the proposed reduction framework for the
ANN [142, 394] summarized in Figure 20.12. Let AN N : Rn0 → RnL be the original network,
composed of L layers;31 D0 = {x0,j }N j Ntrain
j=1 a dataset with Ntrain input samples; and {y }j=1 the
train

31 This is the only assumption we are making regarding the original network.
 404 Chapter 20. A Deep Learning Approach to Improving Reduced Order Models
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 20.12. Graphical representation of the reduction method proposed for an ANN.

real outputs of the AN N . Given the index l, denoting the cutoff layer, the full net is split in two
different parts:
• the premodel, made of the first l layers, and
• the postmodel, formed by the last L − l layers.
It is important to specify that the cutoff layer l is chosen empirically as a trade-off between the
final accuracy and the compression ratio. This value plays a central role in the final outcome
since it is the parameter that controls how many layers of the original network are kept in the
reduced architecture and thus how much information of the original network we are discarding.
Describing the network as a composition of functions AN N ≡ fL ◦ fL−1 ◦ . . . ◦ f1 , the pre-
and postmodels can be formally described as

AN N lpre = fl ◦ fl−1 ◦ . . . ◦ f1 , AN N lpost = fL ◦ fL−1 ◦ . . . ◦ fl+1 , (20.20)

where fj : Rnj−1 → Rnj for j = 1, . . . , L are the functions representing the different layers
of the network—e.g., convolutional, fully connected, batch-normalization, ReLU, or pooling
layers. Therefore, we can rewrite the original model as a composition of the postmodel with the
premodel:
AN N (x0 ) = AN N lpost (AN N lpre (x0 )) (20.21)
for any 1 ≤ l < L and for any x0 ∈ D0 . The network is thus reduced by copying the premodel
from the original net and approximating the postmodel using the method we will describe.
The output of the premodel x(l) = AN N lpre (x0 ) usually lies in a high-dimensional space;
thus we aim to project x(l) onto a low-dimensional space of dimension r using the aforemen-
tioned reduction techniques:
• ASs: As described in Chapter 16 and in [142], we consider a function gl defined as the
composition of the loss function with the postmodel,

gl (x(l) ) = loss(AN N lpost (x(l) )), (20.22)

 20.6. A Reduced Approach for CNNs 405

in order to construct the AS and thus extract the most important directions. In this way, we
will determine the matrix of the project W1 , which will reduce the premodel output xl in
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

the following way:

zAS = WT1 x(l) ∈ Rr . (20.23)

• POD: As discussed in Section 1.1, the singular value decomposition (SVD) is used to
compute the projection matrix U and thus the reduced solution

zPOD = UT x(l) ∈ Rr . (20.24)

After we obtain the reduced solution z, the last step of the method deals with the construction
of the mapping between z and the final output of the network y. Two different techniques have
been used to find that map:
• The polynomial chaos expansion (PCE) proposed by Wiener in [616] states that a real-
valued random variable can be approximated by a finite sum of orthogonal polynomials
weighted by unknown deterministic coefficients cα [308]. Therefore, in our context the
final output of the network y ∈ RnL can be approximated as
P
X
ŷ ≈ cα φα (z), |α| = α1 + . . . + αr , (20.25)
|α|=0

where φα (z) are the multivariate polynomial functions chosen based on the probability
density function ρ associated with z. The number of unknown coefficients in this summa-
tion is given by P + 1 = (p+r)!
p!r! [224], where p is the degree of the polynomial we are
considering in the r-dimensional space. Based on the work of Askey and Wilson [21],
we can provide the orthogonal polynomials for different distributions. One of the possible
choices is the Gaussian distribution with the related Hermite polynomials.
Hence, to determine the PCE, we need to estimate the PCE coefficients cα for |α| =
0, . . . , P , which can be carried out in different ways [564]: following a projection method
based on the orthogonality of the polynomials or following a regression method, the latter
of which we will describe here.
In order to determine the coefficients cα , the following minimization problem has to be
solved:
2
Ntrain p
1 X j
X
j
min y − cα φα (z ) . (20.26)
cα Ntrain
j=1 |α|=0

In particular we need to consider the following matrix:

 φ (z1 ) φ1 (z1 ) ··· φP (z1 ) 
0
 φ0 (z2 ) φ1 (z2 ) ··· φP (z2 )
(20.27)

Φ= .. .. .. .. .

. . . .

φ0 (zNtrain ) φ1 (zNtrain ) · · · φP (zNtrain )
Then the solution of (20.26) is computed by a least-squares optimization,

c = (ΦT Φ)−1 ΦT y, (20.28)

where if the matrix ΦT Φ is ill conditioned, as may happen, the SVD method should be
employed.
 406 Chapter 20. A Deep Learning Approach to Improving Reduced Order Models

• With the FNN described in Section 20.3, the input matches the output of the reduction
layer z ∈ Rr , and thus by using (20.2) we obtain the fact that the final output ŷ of the
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

reduced net32 is determined by

n1 n1 r
!
(2) (1) (2) (1)
X X X
ŷj = wji yi = wji σ wim zm , j = 1, . . . , nout , (20.29)
i=1 i=1 m=1

where nout corresponds to the number of categories that compose the dataset being exam-
ined and σ is the Softplus function:
1
Softplus(x) = log(1 + exp(βx)). (20.30)
β

Once the reduced network has been initialized, we need to train it using the input-output pairs
(xj0 , yj ) for j = 1, . . . , Ntrain and with x(0),j ∈ D0 ∀ j. The technique chosen to perform the
training process is that of knowledge distillation [287], as is done in [142]. This is an indirect
approach where a small student model, in our case AN N red , is trained to mimic the behavior
of a large pretrained model, called the teacher model, i.e., our full AN N , and thus to gain a
comparable or even superior performance.
The output of the last layer in a deep NN y is referred to in this framework as the vector of
logits. Each network input is then characterized by the probability that it belongs to the ith class,
which is defined by the softmax function
exp(yi /T )
pi = Pnclass , (20.31)
j=0 exp(yj /T )

where T , the temperature factor, is introduced to control the importance of each target, as ex-
plained in [287]. Note that if T → ∞, all classes have the same probability, whereas if T → 0
the targets pi become one-hot labels.
The final loss in this method is made of two different losses:
• The distillation loss LD matches the logits between the teacher and the student model.
Following [238, 287], LD is represented by the Kullback–Leibler (KL) divergence loss
[343]:
LD (p(yt , T ), p(ys , T )) = LKL ((p(yt , T ), p(ys , T )), (20.32)
where yt and ys indicate the logits of the teacher and student networks, respectively,
whereas the KL loss LKL is defined by
X pj (yt,j , T )
LKL ((p(ys , T ), p(yt , T )) = T 2 pj (yt,j , T ) log . (20.33)
j
pj (ys,j , T )

• The student loss LS is defined as the cross-entropy loss between the ground truth label and
the logits of the student network [238]:
X
LS (y, p(ys , T )) = LCE (ŷ, p(ys , T )) = −ŷi log(pi (ys,i , T )), (20.34)
i

where ŷ is a ground truth vector, characterized by having only the component correspond-
ing to the ground truth label on the training sample set to one, while the others are set to
zero.
32 Note that in this case the number of hidden layers is set to one since, as discussed in Subsection 20.6.3, we notice

that one hidden layer is enough to gain a good level of accuracy (see, e.g., Section 20.1).
 20.6. A Reduced Approach for CNNs 407
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Figure 20.13. Graphical representation of the VGG-16 architecture.

Both losses, (20.32) and (20.34), are characterized by the use of the logits of the student model
but with different temperatures: T = τ > 1 in the distillation loss and T = 1 in the student loss.
Therefore, the loss is obtained as a weighted sum of the aforementioned losses:

L(x0 , W ) = λLD (p(yt , T = τ ), p(ys , T = τ )) + (1 − λ)LS (ŷ, p(ys , T = 1)), (20.35)

where λ is the regularization parameter, x0 ∈ D0 , and W are the parameters of the student model.

20.6.3 Numerical Results

In this section, we present the results obtained by applying the different reduced methods pro-
posed to a CNN [394]. In particular, as stated before, the problem to be solved is the running
of this net within an embedded system with particular memory constraints in order to perform
the image recognition task. The technique proposed is implemented using PyTorch [437] as the
development environment, while for the actual computation of the active subspaces we use the
open-source Python package ATHENA [496].
Among the several existing CNN architectures [11, 15], e.g., AlexNet, ResNet, Inception,
and VGGNet, the CNN chosen to test the procedure is VGG-16 [538]. As represented in Fig-
ure 20.13, the structure of VGG-16 is characterized by 13 convolutional blocks, made of a con-
volutional layer followed by a nonlinear layer; 5 max-pooling layers; and 3 fully-connected
layers. It is also useful to recall that there are 16 layers that have tunable parameters, i.e., the
convolutional and fully connected ones.
The original network VGG-16 and its reduced versions have been trained, for 60 and 10
epochs,33 respectively, on two datasets:
• The CIFAR-10 dataset [349] is a computer-vision dataset used for object recognition. It
consists of 60000 32 × 32 color images divided into 10 nonoverlapping classes: airplane,
automobile, bird, cat, deer, dog, frog, horse, ship, and truck.
• The Custom dataset is composed of 3448 32 × 32 color images organized in four classes:
three nonoverlapping classes and a mixed one, characterized by pictures with objects of
different categories presented at the same time.
Tables 20.2 and 20.3 show the levels of accuracy of VGG-16 at the end of the training phase,
77.98% for the CIFAR-10 and 95.65% for the custom dataset, which represent the values we
want to achieve or exceed with the reduced networks. In these tables are also presented the results
obtained with different versions of VGG-16, constructed using three cutoff layers34 l: 5, 6, and
7, as is done in [142]. The dimension of the reduced space r onto which we are projecting the
33 The number of epochs for the training phase was chosen as a trade-off between the final accuracy and the time

needed. For this reason, we kept the same value in the two different cases we are considering, i.e., for the two sets of
data, to have a fair comparison.
34 As explained in [142] and in the corresponding implementation, the indices 5, 6, and 7 correspond to the indices of

the convolutional layers in a list where only convolutional and linear layers are taken into consideration as possible cutoff
layers. Therefore, taking into account the whole net with all the different layers, these correspond to 11, 13, and 16.
 408 Chapter 20. A Deep Learning Approach to Improving Reduced Order Models

Table 20.1. Results obtained for the reduced net POD+FNN (7) trained on CIFAR10 with differ-
ent structures for the FNN.
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Hidden layers
1 2 3 4
Epoch 0 81.39% 67.92% 75.52% 81.57%
10 Epoch 10 87.89% 87.59% 87.46% 87.26%
Hidden neurons

Storage FNN (Mb) 0.0024 0.0028 0.0032 0.0036

Epoch 0 80.17% 80.05% 79.97% 78.28%
20 Epoch 10 87.45% 87.13% 87.42% 86.68%
Storage FNN (Mb) 0.0047 0.0063 0.0079 0.0095
Epoch 0 77.57% 80.36% 80.43% 76.26%
30 Epoch 10 86.92% 86.25% 86.30% 85.25%
Storage FNN (Mb) 0.0070 0.0106 0.0141 0.0177
Epoch 0 71.24% 70.38% 69.31% 68.15%
40 Epoch 10 85.04% 84.60% 84.18% 83.64%
Storage FNN (Mb) 0.0093 0.0156 0.0219 0.0281

high-dimensional features, the output of the premodel, is set to 50 both for AS and for POD in
analogy with the structural analysis considerations presented in [142]. We also remark that to
compute the AS we employed the frequent direction method [225] implemented inside ATHENA.
Table 20.1 then reports different results obtained by training the reduced net for 10 epochs
using different FNN architectures,35 i.e., a different number of hidden layers and also hidden
neurons, which are kept constant in each hidden layer of the net. The comparison is made in
terms of memory allocation for the FNN and accuracy of the corresponding reduced net at epoch
0, i.e., after its initialization, and at epoch 10. Note that increasing the depth of the network
does not lead to a remarkable gain in accuracy. Therefore, the structures chosen for testing the
reduction method using an FNN as input-output map are the following:

• CIFAR10: 50 input neurons, 10 output neurons, and 1 hidden layer with 20 hidden neu-
rons;

• custom dataset: 50 input neurons, 4 output neurons, and 1 hidden layer with 10 hidden
neurons.

Tables 20.2 and 20.3 contain the results obtained by initializing the reduced nets and training
them for 10 epochs on the corresponding datasets. The data in the tables compares different
reduced networks, i.e., AS+PCE, AS+FNN, POD+FNN, using three different cutoff layers, 5,
6, and 7, in terms of accuracy (before and after the final training), memory storage, and time
needed for the initialization and training processes. We also highlight that the results we present
are consistent between the two different sets of data, as can be seen in Tables 20.2 and 20.3.
Regarding the problem of memory constraints in an embedded system, note that the stor-
age required for the proposed nets is less than that of the original VGG-16 in both cases (see
Tables 20.2 and 20.3). The checkpoint file for the reduced versions of VGG-16 now requires
less than 10 Mb in all cases, whereas before it occupied 56.14 Mb. Hence, if we compute the

35 Each of the FNNs used was trained for 500 epochs during its initialization process and thus before the retraining

step of the whole net reported in the tables.

 20.6. A Reduced Approach for CNNs 409

Table 20.2. Results obtained with CIFAR10 dataset.

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Network Accuracy Storage (Mb) Time

VGG-16 77.98% 56.15 46 h
Epoch 0 Epoch 10 Pre-M AS/POD PCE/FNN Init Train
AS+PCE (5) 13.52% 82.01% 2.12 3.12 0.05 43 min 4.5 h
AS+FNN (5) 33.06% 80.43% 2.12 3.12 0.0047 5h 4.5 h
POD+FNN (5) 62.16% 80.24% 2.12 3.12 0.0047 79 min 5h
AS+PCE (6) 14.42% 84.69% 4.37 3.12 0.05 49 min 5.5 h
AS+FNN (6) 33.76% 82.13% 4.37 3.12 0.0047 5h 4.5 h
POD+FNN (6) 63.84% 83.93% 4.37 3.12 0.0047 83 min 5h
AS+PCE (7) 4.25% 85.60% 6.62 0.78 0.05 35 min 5.5 h
AS+FNN (7) 75.66% 86.03% 6.62 0.78 0.0047 1.5 h 5h
POD+FNN (7) 80.17% 87.45% 6.62 0.78 0.0047 12 min 5h

Table 20.3. Results obtained with a custom dataset.

Network Accuracy Storage (Mb) Time

VGG-16 95.65% 56.14 22 min
Epoch 0 Epoch 10 Pre-M AS/POD PCE/FNN Init Train
AS+PCE (5) 29.03% 95.21% 2.12 3.12 0.02 2 min 10 min
AS+FNN (5) 94.63% 94.92% 2.12 3.12 0.0021 12.5 min 12 min
POD+FNN (5) 96.52% 96.66% 2.12 3.12 0.0021 28 sec 11.5 min
AS+PCE (6) 29.75% 95.79% 4.37 3.12 0.02 2.5 min 10 min
AS+FNN (6) 94.92% 95.36% 4.37 3.12 0.0021 12.5 min 12.5 min
POD+FNN (6) 96.23% 96.37% 4.37 3.12 0.0021 33 sec 13 min
AS+PCE (7) 28.59% 94.05% 6.62 0.78 0.02 1.5 min 11 min
AS+FNN (7) 94.34% 94.63% 6.62 0.78 0.0021 4.5 min 13 min
POD+FNN (7) 96.37% 96.52% 6.62 0.78 0.0021 33 sec 14 min

compression ratio for each of these nets,

compress_size
compression_ratio = 1 − , (20.36)
uncompress_size

we obtain that the savings in memory is around 90% in all cases. It can also be highlighted
that using an FNN instead of the PCE brings a gain in saving memory space of two orders of
magnitude: 10−4 against 10−2 .
From the point of view of the final accuracy, it can be observed that increasing the cutoff layer
index l leads to nets with higher accuracy since we are retaining more information (i.e., features)
from the original VGG-16, but on the other hand, there is a smaller compression ratio. Hence, the
index l should be chosen carefully, taking into account the field of application and also making
a trade-off between the final accuracy and the reduction. Despite this, for all the reduced CNNs,
the final level of accuracy is similar to (in most cases also greater than) the original one.
Using the reduced net POD+FNN leads to another gain in terms of time, since, as shown in
Table 20.2, in this case, additional training with the whole dataset is not required by this net.
From the table, note that after the initialization, i.e., at epoch 0, we already have a performing
network and thus there is a savings of time of the order of 5 h. This is true and also consistent with
the custom dataset (see Table 20.3), where for the three choices of l we have that the POD+FNN
net has an accuracy greater than that of the original VGG-16 even after the initialization process.
 410 Chapter 20. A Deep Learning Approach to Improving Reduced Order Models

20.7 A Future Perspective on ROMs for Digital Twin

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

20.7.1 Introduction to Digital Twins

Digital twin [481] is one of the emerging technologies in the framework of Industry 4.0. The
digitalization of existing complex systems is one of the key ingredients for adding intelligence
layers to products, processes, factories, and facilities to make them smart. The main idea behind
the digital twin is to create a set of coupled models which can mimic the behavior of the physical
asset, to reliably and quickly estimate all the quantities of interest along with the asset’s entire
operative life, and to make them available for users through a proper interface.
The rising interest in digital twins in both industry and academia is due to its potential ap-
plications, among others fault detection, optimization, and predictive maintenance. Digital twins
can be used to create virtual prototypes which can be used to optimize the design of a product
under development. This helped overcome the build-and-test paradigm, reducing the need to
build physical prototypes, thus reducing costs and time-to-market. This type of digital twin can
also be used to test controllers and to simulate entire plants or environments.
Following [266], three main categories of deployed digital twins exist according to different
levels of integration between the physical asset and the digital one:

• The system integration digital twin can be used to simulate complex systems and test
settings on-premise. There is no direct data exchange between the virtual and the real
asset.

• The embedded digital twin receives data from the physical asset. It can be used for appli-
cation monitoring and virtual sensing.

• The companion digital twin exchanges data in a bidirectional way with its real counterpart.
The most advanced applications, which include dynamic optimization, optimal control,
and advanced diagnostics, belong to this category of digital twins.

Digital twin and physical assets, when bonded together, are usually referred to as cyber-physical
systems.
Hence, digital twins are opening new scenarios and possibilities by integrating virtual and
real assets, but at the same time, they propose new challenges. For example, there is a clear
need to find an effective way to exchange, collect, and analyze data coming from the real asset
and integrate it into the models which make up the digital twins. These challenges can be faced
by combining, in a unified framework, different techniques and technologies recently developed
in numerical methods, IoT, artificial intelligence, Big Data, and high-performance computing
(HPC). Some practical examples of applications of digital twin can be found in [266].

20.7.2 The Role of Modeling and Simulation in Digital Twin

One of the major consequences of the introduction of the digital twin technology is to advance
the importance of models and simulations in the industrial world; in fact, up to the end of the
nineteen-nineties, simulation was intended exclusively for virtual testing purposes and, more re-
cently, models became a support to the entire design phase in the model-based systems engineer-
ing (MBSE) framework. With the introduction of digital twin, models must be able to support
the entire life cycle of the asset, from design to disposal. In detail, the following challenges arise
from the modeling point of view in digital twins.
 20.7. A Future Perspective on ROMs for Digital Twin 411

Intrinsic Complexity
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Modern systems are complex, and their complexity does not reside in the number of components
but rather in the number of physical domains which are involved. For example, cars up to 1980
were almost exclusively mechanical products, while a modern vehicle includes electric compo-
nents, electronic boards, air conditioning, and software. Digital twin must take into account all
the diverse domains to get a complete model so digital twins are intrinsically multiphysical.

High Number of Parameters

Digital twins are a set of parameter-dependent models, and the global space of parameters can be
huge. Proper techniques for reducing the space of parameters can be used to reduce the number of
independent variables and to keep the complexity under control. Some results with applications
in naval engineering can be found in [575, 157, 404], while for an automotive application we
suggest [494].

Long-Time Consistency
Models implemented in digital twins must remain consistent with the corresponding physical as-
sets for their entire lifetimes, especially when the real system tends to degrade over time because
of wear, rusting, change of working conditions, and fatigue phenomena. To overcome this prob-
lem, models should be able to adapt their behavior by learning from the data generated by the
associated physical counterpart (see [321] for a probabilistic graphical model approach). Data
assimilation techniques are therefore mandatory in order to guarantee a tight consistency during
the life of the digital twin.

Near Real Time Simulation

Digital twins must include models which can perform simulation in near real time since the aim
is to continuously monitor the behavior of the physical counterpart or to provide information for
strategies of dynamic control.
Traditionally, digital twins were built by employing zero- and one-dimensional models gath-
ered in blocks that exchange information at given time intervals. This approach has the advantage
of being simple and so cheap from a computational point of view that it can be deployed on edge
devices with limited power. This paradigm, based on relatively simple equation-based models,
is not enough because of its inability to deal with data and its limited accuracy when complex
phenomena are involved. To overcome these limitations, digital twins which mix equation-based
models, ROMs, and data-driven models have been proposed.

20.7.3 ROMs for Digital Twin

Many applications related to digital twins require accurate predictions of physical phenomena
to get consistent results, but these predictions cannot be obtained by employing standard meth-
ods during the design phase because of their intrinsic high cost, despite the recent advancements
in computational power and the improvement in the efficiency of algorithms [265]. For exam-
ple, finite element methods (FEMs) and finite volume methods (FVMs) belong to this set of
computationally expensive methods, and they are commonly used to describe problems related
to structural mechanics and fluid dynamics, respectively. Modern methodologies allow one to
produce reduced order methods, which unlock the possibility of simulating complex two- and
three-dimensional phenomena in real time starting from a dataset of precomputed solutions. In
ROMs, one trades speed for a modest loss of accuracy; hence the adoption of ROMs is a natural
 412 Chapter 20. A Deep Learning Approach to Improving Reduced Order Models

choice when one has to create accurate digital twins to deal with complex dynamics. ROMs
consist of two distinct phases:
Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

• In the offline phase, high-fidelity simulations are performed for some given combinations
of input parameters, the dominant dynamics (or modes) are extracted, and the other infor-
mation needed for building the reduced order method is evaluated.

• In the online phase, the magnitudes of interest are constructed quickly by a linear super-
imposition of the modes multiplied by proper weights.

Two distinct ROM approaches exist: nonintrusive and intrusive. The latter relies on the
projection of the set of governing equations onto a reduced space individuated by the modes
computed during the offline phase. These methods are accurate, with a solid physical back-
ground, but they suffer from possible instabilities. On the other hand, the nonintrusive methods
use machine learning techniques to create mapping functions during the offline phase between
the weights which multiply the modes and the input parameters. Both of the approaches are
able to reconstruct complex phenomena in near real time, providing significant insight into the
phenomena.
ROMs in digital twin are relevant in two distinct phases. The most trivial application is to
apply digital twin to deployable digital twins to provide more accurate information about the real
counterpart. The second is related to the product development phase, when ROMs can be used to
obtain a better product by performing processes that are based on multiquery, e.g., optimization,
uncertainty quantification, and controls. In Figure 20.14 is an example of a model used for a
process of structural optimization, while Figure 20.15 shows the corresponding displacement
field for a given loading condition.

Figure 20.14. Example of a structural model in the naval engineering field. On the left is the
entire hull, while on the right is a sectional view.

Figure 20.15. A modern cruise ship example of displacement field for the hogging loading condition.
 20.7. A Future Perspective on ROMs for Digital Twin 413

20.7.4 Next Developments and Challenges

Downloaded 12/19/23 to 222.29.66.238 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Digital twin is a relatively new technology that started attracting the attention of the industrial
world in the last few years, and many challenges are still open. The first challenge is to devise
a robust and general methodological approach for building digital twins; the first step in this
direction is some recent works by Kapteyn et al. [320], who propose an approach for enabling
data-driven digital twins based on libraries of components based on ROMs. Other relevant as-
pects which are the subject of research are to create big models, coupling data and equation-based
models, which would help to overcome problems that are related to Big Data in applications of
practical interest such as lack of data and data veracity, and it would allow uncertainties to be
quantified.
Some practical aspects are also under investigation, in particular the ones regarding the de-
velopment of general tools which allow an effective way to exchange information among models
generated with different software.
Thanks to the integration of data, ROMs, and models, the concept of digital twin can be
further expanded, and the frontiers of digital twin are related to the creation of complex virtual
copies of more complex systems [416]. These extensions would help to implement effective
strategies for sustainability and a circular economy and would be beneficial for mobility, logis-
tics, the economy, and the environment. Smart factories are one of the main goals of Industry
4.0, and they consist of a digital twin that is strongly integrated with the corresponding physical
plant. This would allow us to constantly monitor and optimize production processes, instanta-
neous power consumption, logistics, and product quality. The next wave of digital twins will
include the augmented digital twin, which interacts not only with its real asset but also with its
surroundings and other digital twins. The augmented digital twin will be used to create smart
infrastructure (harbors, freeways, etc.) and entire cities or ecosystems, e.g., to optimize mobility,
improve efficiency, and monitor pollution. Finally, the most ambitious application of the digital
twin technology is the one related to human digital twin to enable full life cycle health manage-
ment of a human being; given the complexity of biological systems, ROMs will be one of the
key enablers of the application of this new type of digital twins, even though a lot of work still
has to be done in terms of pure modeling of biological processes whose end is to understand
fundamental phenomena [147, 611].
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Bibliography

[1] ITHACA-DG—In real Time Highly Advanced Computational Applications for Discontinuous
Galerkin-ROMs for HopeFOAM. https://github.com/mathLab/ITHACA-DG. (Cited on p. 241)
[2] OpenFOAM website. https://openfoam.com/. (Cited on p. 381)
[3] STAR CCM+ Users Manual. http://www.cd-adapco.com/products/star-ccm/
documentation. (Cited on p. 143)
[4] STEP-file, ISO 10303-21. https://www.loc.gov/preservation/digital/formats/fdd/
fdd000448.shtml. (Cited on p. 357)
[5] STL (STereoLithography) File Format Family. https://www.loc.gov/preservation/digital/
formats/fdd/fdd000505.shtml. (Cited on p. 356)
[6] multiphenics—easy prototyping of multiphysics problems in FEniCS. http://mathlab.sissa.
it/multiphenics, 2016. (Cited on pp. 318, 368)
[7] K. A ARONSON , M. S LAUGHTER , L. M ILLER , E. M C G EE , W. C OTTS , M. ACKER , M. J ES -
SUP, I. G REGORIC , P. L OYALKA , O. F RAZIER , V. J EEVANANDAM , A. A NDERSON , R. KORMOS ,
J. T EUTEBERG , W. L EVY, D. NAFTEL , R. B ITTMAN , F. PAGANI , D. H ATHAWAY, AND S. B OYCE,
HeartWare Ventricular Assist Device (HVAD) Bridge to Transplant ADVANCE Trial Investigators.
Use of an intrapericardial, continuous-flow, centrifugal pump in patients awaiting heart transplan-
tation, Circulation, 125 (2012), pp. 3191–3200, https://doi.org/10.1161/CIRCULATIONAHA.
111.058412. (Cited on p. 374)
[8] I. A BBOTT AND A. VON D OENHOFF, Theory of Wing Sections, Including a Summary of Airfoil
Data, Dover Books on Aeronautical Engineering Series, Dover Publications, 1959. (Cited on p. 355)
[9] R. A BGRALL , D. A MSALLEM , AND R. C RISOVAN, Robust model reduction by L1 -norm mini-
mization and approximation via dictionaries: Application to nonlinear hyperbolic problems, Ad-
vanced Modeling and Simulation in Engineering Sciences, 3 (2016), https://doi.org/10.1186/
s40323-015-0055-3. (Cited on p. 298)
[10] S. R. A HMED , G. R AMM , AND G. FALTIN, Some salient features of the time-averaged ground ve-
hicle wake, SAE Transactions (1984), pp. 473–503, https://doi.org/10.4271/840300. (Cited
on p. 361)
[11] A. A JIT, K. ACHARYA , AND A. S AMANTA, A review of convolutional neural networks, in 2020 In-
ternational Conference on Emerging Trends in Information Technology and Engineering (ic-ETITE),
IEEE, 2020, pp. 1–5, https://doi.org/10.1109/ic-ETITE47903.2020.049. (Cited on p. 407)
[12] S. A LI, Stabilized Reduced Basis Methods for the Approximation of Parametrized Viscous Flows,
PhD thesis, SISSA, 2018. (Cited on p. 128)
[13] S. A LI , F. BALLARIN , AND G. ROZZA, Stabilized reduced basis methods for parametrized steady
Stokes and Navier–Stokes equations, Computers & Mathematics with Applications, 80 (2020),
pp. 2399–2416, https://doi.org/10.1016/j.camwa.2020.03.019. (Cited on pp. 129, 131)

415
 416 Bibliography

[14] E. L. A LLGOWER AND K. G EORG, Introduction to Numerical Continuation Methods, Society

Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

for Industrial and Applied Mathematics, 2003, https://doi.org/10.1137/1.9780898719154.

(Cited on p. 104)

[15] L. A LZUBAIDI , J. Z HANG , A. J. H UMAIDI , A. A L -D UJAILI , Y. D UAN , O. A L -S HAMMA , J. S AN -

TAMARÍA , M. A. FADHEL , M. A L -A MIDIE , AND L. FARHAN , Review of deep learning: Concepts,
CNN architectures, challenges, applications, future directions, Journal of Big Data, 8 (2021), pp. 1–
74, https://doi.org/10.1186/s40537-021-00444-8. (Cited on pp. 391, 394, 407)

[16] A. A MBROSETTI AND A. M ALCHIODI, Nonlinear Analysis and Semilinear Elliptic Problems, no. v.
10 in Cambridge Studies in Advanced Mathematics, Cambridge University Press, 2007. (Cited on
p. 101)

[17] A. A MBROSETTI AND G. P RODI, A Primer of Nonlinear Analysis, Cambridge Studies in Advanced
Mathematics, Cambridge University Press, 1995. (Cited on p. 101)

[18] D. A MSALLEM , M. J. Z AHR , AND C. FARHAT, Nonlinear model order reduction based on lo-
cal reduced-order bases, International Journal for Numerical Methods in Engineering, 92 (2012),
pp. 891–916, https://doi.org/10.1002/nme.4371. (Cited on p. 298)

[19] P. F. A NTONIETTI , P. PACCIARINI , AND A. Q UARTERONI, A discontinuous Galerkin reduced basis

element method for elliptic problems, ESAIM: Mathematical Modelling and Numerical Analysis, 50
(2016), pp. 337–360, https://doi.org/10.1051/m2an/2015045. (Cited on p. 234)

[20] D. N. A RNOLD , F. B REZZI , B. C OCKBURN , AND D. M ARINI, Discontinuous Galerkin Methods

for Elliptic Problems, in Lecture Notes in Computational Science and Engineering, Springer, 2000,
pp. 89–101, https://doi.org/10.1007/978-3-642-59721-3_5. (Cited on p. 239)

[21] R. A SKEY AND J. A. W ILSON, Some basic hypergeometric orthogonal polynomials that generalize
Jacobi polynomials, Memoirs of the American Mathematical Society, 54 (1985), https://doi.
org/10.1090/memo/0319. (Cited on p. 405)

[22] M. A STORINO , F. C HOULY, AND M. A. F ERNÁNDEZ, Robin based semi-implicit coupling in fluid-
structure interaction: Stability analysis and numerics, SIAM Journal on Scientific Computing, 31
(2009), pp. 4041–4065, https://doi.org/10.1137/090749694. (Cited on p. 291)

[23] I. BABUŠKA, The finite element method with penalty, Mathematics of Computation, 27 (1973),
pp. 221–221, https://doi.org/10.1090/s0025-5718-1973-0351118-5. (Cited on pp. 154,
155)

[24] E. BADER , M. K ÄRCHER , M. G REPL , AND K. V EROY-G REPL, A certified reduced basis ap-
proach for parametrized linear-quadratic optimal control problems with control constraints, IFAC-
PapersOnLine, 48 (2015), pp. 719–720, https://doi.org/10.1016/j.ifacol.2015.05.167.
(Cited on p. 83)

[25] E. BADER , M. K ÄRCHER , M. A. G REPL , AND K. V EROY, Certified reduced basis methods for
parametrized distributed elliptic optimal control problems with control constraints, SIAM Journal on
Scientific Computing, 38 (2016), pp. A3921–A3946, https://doi.org/10.1137/16m1059898.
(Cited on p. 83)

[26] S. BADIA , F. N OBILE , AND C. V ERGARA, Fluid–structure partitioned procedures based on Robin
transmission conditions, Journal of Computational Physics, 227 (2008), pp. 7027–7051, https:
//doi.org/10.1016/j.jcp.2008.04.006. (Cited on p. 291)

[27] S. BADIA , A. Q UAINI , AND A. Q UARTERONI, Modular vs. non-modular preconditioners for
fluid–structure systems with large added-mass effect, Computer Methods in Applied Mechanics and
Engineering, 197 (2008), pp. 4216–4232, https://doi.org/10.1016/j.cma.2008.04.018.
(Cited on p. 286)
 Bibliography 417

[28] S. BADIA , A. Q UAINI , AND A. Q UARTERONI, Splitting methods based on algebraic factorization
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

for fluid-structure interaction, SIAM Journal on Scientific Computing, 30 (2008), pp. 1778–1805,
https://doi.org/10.1137/070680497. (Cited on p. 286)

[29] S. BADIA , A. Q UAINI , AND A. Q UARTERONI, Coupling Biot and Navier–Stokes equations for mod-
elling fluid–poroelastic media interaction, Journal of Computational Physics, 228 (2009), pp. 7986–
8014, https://doi.org/10.1016/j.jcp.2009.07.019. (Cited on p. 287)

[30] F. BALLARIN , T. C HACÓN R EBOLLO , E. D ELGADO Á VILA , M. G ÓMEZ M ÁRMOL , AND

G. ROZZA, Certified reduced basis VMS-Smagorinsky model for natural convection flow in a cav-
ity with variable height, Computers & Mathematics with Applications, 80 (2020), pp. 973–989,
https://doi.org/10.1016/j.camwa.2020.05.013. (Cited on pp. 69, 128)

[31] F. BALLARIN , A. D’A MARIO , S. P EROTTO , AND G. ROZZA, A POD-selective inverse distance
weighting method for fast parametrized shape morphing, International Journal for Numerical Meth-
ods in Engineering, 117 (2018), pp. 860–884, https://doi.org/10.1002/nme.5982. (Cited on
pp. 206, 351)

[32] F. BALLARIN , E. FAGGIANO , S. I PPOLITO , A. M ANZONI , A. Q UARTERONI , G. ROZZA , AND

R. S CROFANI, Fast simulations of patient-specific haemodynamics of coronary artery bypass grafts
based on a POD-Galerkin method and a vascular shape parametrization, Journal of Computational
Physics, 315 (2016), pp. 609–628, https://doi.org/10.1016/j.jcp.2016.03.065. (Cited on
pp. 283, 356, 367)
[33] F. BALLARIN , E. FAGGIANO , A. M ANZONI , A. Q UARTERONI , G. ROZZA , S. I PPOLITO , C. A N -
TONA , AND R. S CROFANI , Numerical modeling of hemodynamics scenarios of patient-specific coro-
nary artery bypass grafts, Biomechanics and Modeling in Mechanobiology, 16 (2017), pp. 1373–
1399, https://doi.org/10.1007/s10237-017-0893-7. (Cited on pp. 83, 367)

[34] F. BALLARIN , A. M ANZONI , A. Q UARTERONI , AND G. ROZZA, Supremizer stabilization of

POD-Galerkin approximation of parametrized steady incompressible Navier-Stokes equations, In-
ternational Journal for Numerical Methods in Engineering, 102 (2014), pp. 1136–1161, https:
//doi.org/10.1002/nme.4772. (Cited on pp. 15, 40, 107, 144, 159, 160, 189, 258, 367)
[35] F. BALLARIN , A. M ANZONI , G. ROZZA , AND S. S ALSA, Shape optimization by free-form defor-
mation: Existence results and numerical solution for Stokes flows, Journal of Scientific Computing,
60 (2014), pp. 537–563, https://doi.org/10.1007/s10915-013-9807-8. (Cited on p. 347)

[36] F. BALLARIN AND G. ROZZA, POD-Galerkin monolithic reduced order models for parametrized
fluid-structure interaction problems, International Journal for Numerical Methods in Fluids, 82
(2016), pp. 1010–1034, https://doi.org/10.1002/fld.4252. (Cited on pp. 286, 287, 299)
[37] F. BALLARIN , G. ROZZA , AND Y. M ADAY, Reduced-order semi-implicit schemes for fluid-structure
interaction problems, in Model Reduction of Parametrized Systems, Springer International Pub-
lishing, 2017, pp. 149–167, https://doi.org/10.1007/978-3-319-58786-8_10. (Cited on
pp. 286, 293)

[38] F. BALLARIN , A. S ARTORI , AND G. ROZZA, RBniCS—reduced order modelling in FEniCS. http:
//mathlab.sissa.it/rbnics, 2015. (Cited on pp. 26, 318, 368)

[39] J. BANKS , W. H ENSHAW, AND D. S CHWENDEMAN, An analysis of a new stable partitioned al-
gorithm for FSI problems. Part I: Incompressible flow and elastic solids, Journal of Computational
Physics, 269 (2014), pp. 108–137, https://doi.org/10.1016/j.jcp.2014.03.006. (Cited on
p. 292)

[40] M. BARRAULT, Y. M ADAY, N. C. N GUYEN , AND A. T. PATERA, An “empirical interpola-

tion” method: Application to efficient reduced-basis discretization of partial differential equations,
Comptes Rendus Mathematique, 339 (2004), pp. 667–672, https://doi.org/10.1016/j.crma.
2004.08.006. (Cited on pp. 55, 59, 63, 90, 108, 151, 316)
 418 Bibliography

[41] A. BARRETT AND G. R EDDIEN, On the reduced basis method, ZAMM—Journal of Applied Math-
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

ematics and Mechanics/Zeitschrift für Angewandte Mathematik und Mechanik, 75 (1995), pp. 543–
549, https://doi.org/10.1002/zamm.19950750709. (Cited on p. 3)

[42] J. W. BARRETT AND C. M. E LLIOTT, Finite element approximation of the Dirichlet problem using
the boundary penalty method, Numerische Mathematik, 49 (1986), pp. 343–366, https://doi.
org/10.1007/bf01389536. (Cited on pp. 154, 155)

[43] F. BASSI AND S. R EBAY, Accurate 2D Euler computations by means of a high order discontinu-
ous finite element method, in Fourteenth International Conference on Numerical Methods in Fluid
Dynamics, 453 (1995), pp. 234–240. (Cited on p. 234)

[44] F. BASSI AND S. R EBAY, A high-order accurate discontinuous finite element method for the numer-
ical solution of the compressible Navier–Stokes equations, Journal of Computational Physics, 131
(1997), pp. 267–279, https://doi.org/10.1006/jcph.1996.5572. (Cited on pp. 234, 236)

[45] S. BASTING , A. Q UAINI , S. Č ANI Ć , AND R. G LOWINSKI, Extended ALE method for
fluid–structure interaction problems with large structural displacements, Journal of Computational
Physics, 331 (2017), pp. 312–336, https://doi.org/10.1016/j.jcp.2016.11.043. (Cited on
p. 286)

[46] S. BASTING , A. Q UAINI , S. C ANIC , AND R. G LOWINSKI, On the implementation and benchmark-
ing of an extended ALE method for FSI problems, in Fluid-Structure Interaction, De Gruyter, 2017,
pp. 3–40, https://doi.org/10.1515/9783110494259-001. (Cited on p. 284)

[47] S. BASTING , A. Q UAINI , R. G LOWINSKI , AND S. C ANI Ć, Comparison of time discretization
schemes to simulate the motion of an inextensible beam, in Numerical Mathematics and Advanced
Applications - ENUMATH 2013, A. Abdulle, S. Deparis, D. Kressner, F. Nobile, and M. Picasso,
eds., vol. 103 of Lecture Notes in Computational Science and Engineering, Springer, 2015, pp. 175–
183. (Cited on p. 283)

[48] V. BATHELMANN , E. N OVAK , AND K. R ITTER, High dimensional polynomial interpolation on

sparse grids, Advances in Computational Mathematics, 4 (2000), pp. 273–288, https://doi.org/
10.1023/A:1018977404843. (Cited on p. 256)

[49] Y. BAZILEVS , V. C ALO , J. C OTTRELL , T. H UGHES , A. R EALI , AND G. S COVAZZI, Variational

multiscale residual-based turbulence modeling for large eddy simulation of incompressible flows,
Computer Methods in Applied Mechanics and Engineering, 197 (2007), pp. 173–201, https://
doi.org/10.1016/j.cma.2007.07.016. (Cited on p. 171)

[50] C. J. P. B ÉLISLE , H. E. ROMEIJN , AND R. L. S MITH, Hit-and-run algorithms for generating

multivariate distributions, Mathematics of Operations Research, 18 (1993), pp. 255–266, https:
//doi.org/10.1287/moor.18.2.255. (Cited on p. 329)

[51] Y. B ENGIO, Deep learning of representations for unsupervised and transfer learning, in Proceedings
of 2011 ICML Workshop on Unsupervised and Transfer Learning, vol. 27, JMLR Workshop and
Conference Proceedings, 2012, pp. 17–36. (Cited on p. 389)

[52] M. B ENÍTEZ AND A. B ERMÚDEZ, A second order characteristics finite element scheme for natural
convection problems, Journal of Computational and Applied Mathematics, 235 (2011), pp. 3270–
3284, https://doi.org/10.1016/j.cam.2011.01.007. (Cited on p. 69)

[53] P. B ENNER , S. G RIVET-TALOCIA , A. Q UARTERONI , G. ROZZA , W. H. A. S CHILDERS , AND

L. M. S ILVEIRA, eds., Model Order Reduction, De Gruyter, 2020. (Cited on pp. 1, 228)

[54] P. B ENNER , S. G UGERCIN , AND K. W ILLCOX, A survey of projection-based model reduction meth-
ods for parametric dynamical systems, SIAM Review, 57 (2015), pp. 483–531, https://doi.org/
10.1137/130932715. (Cited on p. 228)
 Bibliography 419

[55] P. B ENNER , V. M EHRMANN , AND D. C. S ORENSEN, Dimension Reduction of Large-Scale Systems,

Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

vol. 45 of Lecture Notes in Computational Science and Engineering, Springer, 2005, https://doi.
org/10.1007/3-540-27909-1. (Cited on p. 1)

[56] A. B ERLINET AND C. T HOMAS -AGNAN, Reproducing Kernel Hilbert Spaces in Probabil-
ity and Statistics, Springer Science & Business Media, 2004, https://doi.org/10.1007/
978-1-4419-9096-9. (Cited on p. 332)

[57] F. B ERNARD , A. I OLLO , AND S. R IFFAUD, Reduced-order model for the BGK equation based on
POD and optimal transport, Journal of Computational Physics, 373 (2018), pp. 545–570, https:
//doi.org/10.1016/j.jcp.2018.07.001. (Cited on p. 298)

[58] L. B ERTAGNA , A. Q UAINI , AND A. V ENEZIANI, Deconvolution-based nonlinear filtering for in-
compressible flows at moderately large Reynolds numbers, International Journal for Numerical Meth-
ods in Fluids, 81 (2015), pp. 463–488, https://doi.org/10.1002/fld.4192. (Cited on p. 171)

[59] W.-J. B EYN AND V. T HÜMMLER, Freezing solutions of equivariant evolution equations, SIAM
Journal on Applied Dynamical Systems, 3 (2004), pp. 85–116, https://doi.org/10.1137/
030600515. (Cited on p. 298)

[60] P. B INEV, A. C OHEN , W. DAHMEN , R. D E VORE , G. P ETROVA , AND P. W OJTASZCZYK, Con-

vergence rates for greedy algorithms in reduced basis methods, SIAM Journal on Mathematical
Analysis, 43 (2011), pp. 1457–1472, https://doi.org/10.1137/100795772. (Cited on pp. 5,
62)

[61] C. B ISHOP, Pattern Recognition and Machine Learning, Information Science and Statistics,
Springer, 1 ed., 2006. (Cited on p. 330)

[62] D. A. B ISTRIAN AND I. M. NAVON, Randomized dynamic mode decomposition for nonintrusive
reduced order modelling, International Journal for Numerical Methods in Engineering, 112 (2017),
pp. 3–25, https://doi.org/10.1002/nme.5499. (Cited on p. 209)

[63] K. B IZON AND G. C ONTINILLO, Reduced order modelling of chemical reactors with recycle by
means of POD-penalty method, Computers & Chemical Engineering, 39 (2012), pp. 22–32, https:
//doi.org/10.1016/j.compchemeng.2011.10.001. (Cited on p. 154)

[64] R. D. B LEVINS AND C. S. C OUGHRAN, Experimental investigation of vortex-induced vibration

in one and two dimensions with variable mass, damping, and Reynolds number, Journal of Fluids
Engineering, 131 (2009), p. 101202, https://doi.org/10.1115/1.3222904. (Cited on p. 188)

[65] D. B OFFI , F. B REZZI , AND M. F ORTIN, Mixed Finite Element Methods and Applications, vol. 44 of
Springer Series in Computational Mathematics, Springer-Verlag, 1 ed., 2013, https://doi.org/
10.1007/978-3-642-36519-5. (Cited on p. 318)

[66] J. B OLAND AND W. L AYTON, An analysis of the finite element method for natural convection
problems, Numerical Methods for Partial Differential Equations, 6 (1990), pp. 115–126, https:
//doi.org/10.1002/num.1690060202. (Cited on p. 69)

[67] J. B OLAND AND W. L AYTON, Error analysis for finite element methods for steady natural convec-
tion problems, Numerical Functional Analysis and Optimization, 11 (1990), pp. 449–483, https:
//doi.org/10.1080/01630569008816383. (Cited on p. 69)

[68] G. B ONCORAGLIO AND C. FARHAT, Active manifold and model reduction for multidisciplinary
analysis and optimization, in AIAA Scitech 2021 Forum, 2021, p. 1694, https://doi.org/10.
2514/6.2021-1694. (Cited on p. 338)

[69] J. B ORGGAARD , T. I LIESCU , AND J. P. ROOP, A bounded artificial viscosity large eddy simulation
model, SIAM Journal on Numerical Analysis, 47 (2009), pp. 622–645, https://doi.org/10.
1137/060656164. (Cited on p. 170)
 420 Bibliography

[70] N. B OTKIN AND V. T UROVA -B OTKINA, An algorithm for finding the Chebyshev center of a convex
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

polyhedron, Applied Mathematics and Optimization, 29 (1994), pp. 211–222, https://doi.org/

10.1007/BF01204183. (Cited on p. 329)

[71] J. B OUSSINESQ, Des eaux courantes. Memoires présentés par divers savants à l’Académie des
Sciences de l’Institut National de France, Tome XXIII, (1877). (Cited on p. 167)

[72] A. L. B OWERS , L. G. R EBHOLZ , A. TAKHIROV, AND C. T RENCHEA, Improved accuracy in reg-

ularization models of incompressible flow via adaptive nonlinear filtering, International Journal for
Numerical Methods in Fluids, 70 (2012), pp. 805–828, https://doi.org/10.1002/fld.2732.
(Cited on p. 170)

[73] H. B REZIS, Analyse fonctionnelle: Theorie et applications, Dunod, 2002. (Cited on p. 71)

[74] F. B REZZI, On the existence, uniqueness and approximation of saddle-point problems arising from
Lagrangian multipliers, Revue française d’automatique, informatique, recherche opérationnelle.
Analyse numérique, 8 (1974), pp. 129–151, https://doi.org/10.1051/m2an/197408R201291.
(Cited on pp. 8, 22, 64, 318)

[75] F. B REZZI AND J. P ITKÄRANTA, On the stabilization of finite element approximations of the Stokes
equations, in Efficient Solutions of Elliptic Systems, Vieweg+Teubner Verlag, 1984, pp. 11–19,
https://doi.org/10.1007/978-3-663-14169-3_2. (Cited on pp. 125, 128, 132)

[76] F. B REZZI , J. R APPAZ , AND P. A. R AVIART, Finite dimensional approximation of nonlinear prob-
lems, Numerische Mathematik, 36 (1980), pp. 1–25, https://doi.org/10.1007/BF01395985.
(Cited on pp. 64, 65, 72, 74, 102, 136, 137)

[77] R. A. B RIDGES , A. D. G RUBER , C. R. F ELDER , M. V ERMA , AND C. H OFF, Active manifolds:

A non-linear analogue to active subspaces, in Proceedings of the 36th International Conference on
Machine Learning, ICML 2019, Long Beach, California, USA, 9–15 June 2019, pp. 764–772. (Cited
on pp. 327, 337)

[78] A. N. B ROOKS AND T. J. H UGHES, Streamline upwind/Petrov-Galerkin formulations for convection

dominated flows with particular emphasis on the incompressible Navier-Stokes equations, Computer
Methods in Applied Mechanics and Engineering, 32 (1982), pp. 199–259, https://doi.org/10.
1016/0045-7825(82)90071-8. (Cited on pp. 125, 127, 130, 132)

[79] S. L. B RUNTON AND J. N. K UTZ, Data-Driven Science and Engineering: Machine Learning,
Dynamical Systems, and Control, Cambridge University Press, 2019. (Cited on p. 207)

[80] S. L. B RUNTON , J. L. P ROCTOR , AND J. N. K UTZ, Discovering governing equations from data
by sparse identification of nonlinear dynamical systems, Proceedings of the National Academy of
Sciences, 113 (2016), pp. 3932–3937, https://doi.org/10.1073/pnas.1517384113. (Cited
on p. 222)

[81] A. B UFFA , Y. M ADAY, A. T. PATERA , C. P RUD ’ HOMME , AND G. T URINICI, A priori conver-
gence of the greedy algorithm for the parametrized reduced basis method, ESAIM: Mathematical
Modelling and Numerical Analysis, 46 (2012), pp. 595–603, https://doi.org/10.1051/m2an/
2011056. (Cited on pp. 5, 15, 31, 32)

[82] M. D. B UHMANN, Radial Basis Functions: Theory and Implementations, vol. 12 of Cambridge
Monographs on Applied and Computational Mathematics, Cambridge University Press, 2003,
https://doi.org/10.1017/CBO9780511543241. (Cited on pp. 10, 206, 346, 348, 350)

[83] T. B UI -T HANH , M. DAMODARAN , AND K. W ILLCOX, Aerodynamic data reconstruction and in-
verse design using proper orthogonal decomposition, AIAA Journal, 42 (2004), pp. 1505–1516,
https://doi.org/10.2514/1.2159. (Cited on p. 205)
 Bibliography 421

[84] M. B UKA Č , S. Č ANI Ć , R. G LOWINSKI , J. TAMBAČA , AND A. Q UAINI, Fluid–structure interac-

Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

tion in blood flow capturing non-zero longitudinal structure displacement, Journal of Computational
Physics, 235 (2013), pp. 515–541, https://doi.org/10.1016/j.jcp.2012.08.033. (Cited on
p. 286)
[85] J. B URKARDT, M. G UNZBURGER , AND H. L EE, POD and CVT-based reduced-order modeling
of Navier–Stokes flows, Computer Methods in Applied Mechanics and Engineering, 196 (2006),
pp. 337–355. (Cited on pp. 15, 87, 92)
[86] N. C AGNIART, A Few Non Linear Approaches in Model Order Reduction, PhD thesis, École Doc-
torale de Sciences Mathématiques de Paris Centre, 2018. (Cited on p. 298)
[87] N. C AGNIART, R. C RISOVAN , Y. M ADAY, AND R. A BGRALL, Model Order Reduction for Hyper-
bolic Problems: A New Framework. hal:01583224, 2017. (Cited on p. 298)
[88] N. C AGNIART, Y. M ADAY, AND B. S TAMM, Model order reduction for problems with large con-
vection effects, in Contributions to Partial Differential Equations and Applications, W. Fitzgibbon,
Y. Kuznetsov, P. Neittaanmäki, J. Periaux, and O. Pironneau, eds., vol. 47 of Computational Methods
in Applied Sciences, Springer, Cham., 2019, pp. 131–150. (Cited on pp. 298, 299, 305)
[89] J. W. C AHN AND J. E. H ILLIARD, Free energy of a nonuniform system. I. Interfacial free en-
ergy, The Journal of Chemical Physics, 28 (1958), pp. 258–267, https://doi.org/10.1063/1.
1744102. (Cited on p. 280)
[90] S. C AI , Z. M AO , Z. WANG , M. Y IN , AND G. E. K ARNIADAKIS, Physics-informed neural networks
(PINNs) for fluid mechanics: A review, Acta Mechanica Sinica, 37 (2021), pp. 1727–1738, https:
//doi.org/10.1007/s10409-021-01148-1. (Cited on p. 401)
[91] A. C AIAZZO , T. I LIESCU , V. J OHN , AND S. S CHYSCHLOWA, A numerical investigation of veloc-
ity–pressure reduced order models for incompressible flows, Journal of Computational Physics, 259
(2014), pp. 598–616, https://doi.org/10.1016/j.jcp.2013.12.004. (Cited on pp. 144, 187)
[92] O. C ALIN, Deep Learning Architectures: A Mathematical Approach, Springer Series in the Data
Sciences, Springer International Publishing, 2020. (Cited on pp. 390, 391, 392, 393)
[93] G. C ALOZ AND J. R APPAZ, Numerical analysis for nonlinear and bifurcation problems, Handbook
of Numerical Analysis, 5 (1997), pp. 487–637. (Cited on pp. 65, 72, 102, 136)
[94] E. J. C ANDÈS , X. L I , Y. M A , AND J. W RIGHT, Robust principal component analysis?, Journal of
the ACM (JACM), 58 (2011), pp. 1–37, https://doi.org/10.1145/1970392.1970395. (Cited
on p. 205)
[95] C. C ANTWELL , D. M OXEY, A. C OMERFORD , A. B OLIS , G. ROCCO , G. M ENGALDO , D. D E
G RAZIA , S. YAKOVLEV, J.-E. L OMBARD , D. E KELSCHOT, B. J ORDI , H. X U , Y. M OHAMIED ,
C. E SKILSSON , B. N ELSON , P. VOS , C. B IOTTO , R. K IRBY, AND S. S HERWIN, Nektar++:
An open-source spectral/hp element framework, Computer Physics Communications, 192 (2015),
pp. 205–219, https://doi.org/10.1016/j.cpc.2015.02.008. (Cited on p. 229)
[96] C. C ANUTO , M. H USSAINI , A. Q UARTERONI , AND T. Z HANG, Spectral Methods Fundamentals
in Single Domains, Scientific Computation, Springer, 2006. (Cited on pp. 223, 373)
[97] C. C ANUTO , M. H USSAINI , A. Q UARTERONI , AND T. Z HANG, Spectral Methods Evolution to
Complex Geometries and Applications to Fluid Dynamics, Scientific Computation, Springer, 2007.
(Cited on pp. 223, 373)
[98] G. C ARERE , M. S TRAZZULLO , F. BALLARIN , G. ROZZA , AND R. S TEVENSON, A weighted POD-
reduction approach for parametrized PDE-constrained optimal control problems with random in-
puts and applications to environmental sciences, Computers & Mathematics with Applications, 102
(2021), pp. 261–276, https://doi.org/10.1016/j.camwa.2021.10.020. (Cited on pp. 95,
261)
 422 Bibliography

[99] K. C ARLBERG, Adaptive h-refinement for reduced-order models, International Journal for Numeri-
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

cal Methods in Engineering, 102 (2014), pp. 1192–1210, https://doi.org/10.1002/nme.4800.

(Cited on p. 298)

[100] K. C ARLBERG , C. FARHAT, J. C ORTIAL , AND D. A MSALLEM, The GNAT method for nonlinear
model reduction: Effective implementation and application to computational fluid dynamics and
turbulent flows, Journal of Computational Physics, 242 (2013), pp. 623–647, https://doi.org/
10.1016/j.jcp.2013.02.028. (Cited on p. 151)

[101] J. C ARLTON, Marine Propellers and Propulsion (Second Edition), Butterworth-Heinemann, Oxford,
2007, ch. 6—Propeller performance characteristics, pp. 88–135, https://doi.org/10.1016/
B978-075068150-6/50008-5. (Cited on pp. 346, 355)

[102] P. C AUSIN , J. F. G ERBEAU , AND F. N OBILE, Added-mass effect in the design of partitioned algo-
rithms for fluid–structure problems, Computer Methods in Applied Mechanics and Engineering, 194
(2005), pp. 4506–4527, https://doi.org/10.1016/j.cma.2004.12.005. (Cited on p. 286)

[103] F. C AVALLINI AND F. C RISCIANI, Quasi-geostrophic Theory of Oceans and Atmosphere: Topics in
the Dynamics and Thermodynamics of the Fluid Earth, vol. 45, Springer Science & Business Media,
2013, https://doi.org/10.1007/978-94-007-4691-6_2. (Cited on p. 89)

[104] A. C ESMELIOGLU , H. L EE , A. Q UAINI , K. WANG , AND S. Y. Y I, Optimization-based decoupling

algorithms for a fluid-poroelastic system, in Topics in Numerical Partial Differential Equations and
Scientific Computing, S. C. Brenner, ed., Springer, 2016, pp. 79–98. (Cited on p. 286)

[105] T. C HACÓN R EBOLLO, A term by term stabilization algorithm for finite element solution of incom-
pressible flow problems, Numerische Mathematik, 79 (1998), pp. 283–319, https://doi.org/10.
1007/s002110050341. (Cited on pp. 125, 132)

[106] T. C HACÓN R EBOLLO, An analysis technique for stabilized finite element solution of incompressible
flows, ESAIM: Mathematical Modelling and Numerical Analysis, 35 (2001), pp. 57–89, https:
//doi.org/10.1051/m2an:2001107. (Cited on p. 65)

[107] T. C HACÓN R EBOLLO , E. D ELGADO Á VILA , M. G ÓMEZ M ÁRMOL , F. BALLARIN , AND

G. ROZZA, On a certified Smagorinsky reduced basis turbulence model, SIAM Journal on Numer-
ical Analysis, 55 (2017), pp. 3047–3067, https://doi.org/10.1137/17m1118233. (Cited on
pp. 5, 60)

[108] T. C HACÓN R EBOLLO , M. G ÓMEZ M ÁRMOL , V. G IRAULT, AND I. S ÁNCHEZ M UÑOZ, A high
order term-by-term stabilization solver for incompressible flow problems, IMA Journal of Numeri-
cal Analysis, 33 (2012), pp. 974–1007, https://doi.org/10.1093/imanum/drs023. (Cited on
pp. 125, 132, 136)

[109] T. C HACÓN R EBOLLO , M. G ÓMEZ M ÁRMOL , F. H ECHT, S. RUBINO , AND I. S ÁNCHEZ M UÑOZ,
A high-order local projection stabilization method for natural convection problems, Journal of Sci-
entific Computing, 74 (2018), pp. 667–692, https://doi.org/10.1007/s10915-017-0469-9.
(Cited on pp. 69, 75)

[110] T. C HACÓN R EBOLLO , M. G ÓMEZ M ÁRMOL , AND S. RUBINO, Numerical analysis of a finite
element projection-based VMS turbulence model with wall laws, Computer Methods in Applied Me-
chanics and Engineering, 285 (2015), pp. 379–405, https://doi.org/10.1016/j.cma.2014.
11.023. (Cited on pp. 133, 135, 136)

[111] T. C HACÓN R EBOLLO AND R. L EWANDOWSKI, Mathematical and Numerical Founda-

tions of Turbulence Models and Applications, Springer, 2014, https://doi.org/10.1007/
978-1-4939-0455-6_3. (Cited on pp. 59, 60, 64, 66, 70)
 Bibliography 423

[112] S. C HANDRA, A three-dimensional insight into the complexity of flow convergence in mitral regur-
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

gitation: Adjunctive benefit of anatomic regurgitant orifice area, American Journal of Physiology.
Heart and Circulatory Physiology, 301 (2011), https://doi.org/10.1152/ajpheart.00275.
2011. (Cited on p. 311)

[113] B. C HANG , L. M ENG , E. H ABER , L. RUTHOTTO , D. B EGERT, AND E. H OLTHAM, Reversible

architectures for arbitrarily deep residual neural networks, in Proceedings of the AAAI Conference
on Artificial Intelligence, vol. 32, 2018. (Cited on p. 337)

[114] D. C HAPELLE , A. G ARIAH , P. M OIREAU , AND J. S AINTE -M ARIE, A Galerkin strategy with
proper orthogonal decomposition for parameter-dependent problems: Analysis, assessments and
applications to parameter estimation, ESAIM: Mathematical Modelling and Numerical Analysis,
47 (2013), pp. 1821–1843, https://doi.org/10.1051/m2an/2013090. (Cited on pp. 15, 87,
92)

[115] E. C HARALAMPIDIS , P. K EVREKIDIS , AND P. E. FARRELL, Computing stationary solutions of the

two-dimensional Gross–Pitaevskii equation with deflated continuation, Communications in Nonlin-
ear Science and Numerical Simulation, 54 (2018), pp. 482–499, https://doi.org/10.1016/j.
cnsns.2017.05.024. (Cited on p. 112)

[116] S. C HATURANTABUT AND D. C. S ORENSEN, Nonlinear model reduction via discrete empirical
interpolation, SIAM Journal on Scientific Computing, 32 (2010), pp. 2737–2764, https://doi.
org/10.1137/090766498. (Cited on pp. 56, 108)

[117] P. C HEN , A. Q UARTERONI , AND G. ROZZA, A weighted reduced basis method for elliptic partial
differential equations with random input data, SIAM Journal on Numerical Analysis, 51 (2013),
pp. 3163–3185, https://doi.org/10.1137/130905253. (Cited on pp. 251, 252, 253, 254)

[118] P. C HEN , A. Q UARTERONI , AND G. ROZZA, Multilevel and weighted reduced basis method for
stochastic optimal control problems constrained by Stokes equations, Numerische Mathematik, 133
(2016), pp. 67–102, https://doi.org/10.1007/s00211-015-0743-4. (Cited on p. 257)

[119] W. C HEN , Q. WANG , J. S. H ESTHAVEN , AND C. Z HANG, Physics-informed machine learning

for reduced-order modeling of nonlinear problems, Journal of Computational Physics, 446 (2021),
p. 110666, https://doi.org/10.1016/j.jcp.2021.110666. (Cited on pp. 11, 398)

[120] Y. C HEN , J. H ESTHAVEN , Y. M ADAY, AND J. RODRÍGUEZ, A monotonic evaluation of lower

bounds for inf-sup stability constants in the frame of reduced basis approximations, Comptes rendus
mathematique, 346 (2008), https://doi.org/10.1016/j.crma.2008.10.012. (Cited on p. 5)

[121] Y. C HEN , J. H ESTHAVEN , Y. M ADAY, AND J. RODRÍGUEZ, Improved successive constraint

method based a posteriori error estimate for reduced basis approximation of 2D Maxwell’s prob-
lem, ESAIM: Mathematical Modelling and Numerical Analysis, 43 (2009), pp. 1099–1116, https:
//doi.org/10.1051/m2an/2009037. (Cited on p. 5)

[122] W. C HERDRON , F. D URST, AND J. H. W HITELAW, Asymmetric flows and instabilities in symmetric
ducts with sudden expansions, Journal of Fluid Mechanics, 84 (1978), pp. 13–31, https://doi.
org/10.1017/s0022112078000026. (Cited on p. 311)

[123] F. C HINESTA , P. L ADEVEZE , AND E. C UETO, A short review on model order reduction based on
proper generalized decomposition, Archives of Computational Methods in Engineering, 18 (2011),
p. 395, https://doi.org/10.1007/s11831-011-9064-7. (Cited on pp. 43, 53)

[124] A. J. C HORIN, Numerical solution of the Navier-Stokes equations, Mathematics of Computation, 22

(1968), pp. 745–762, https://doi.org/10.1090/s0025-5718-1968-0242392-2. (Cited on
p. 241)
 424 Bibliography

[125] M. C IANFERRA , V. A RMENIO , AND S. I ANNIELLO, Hydroacoustic noise from different geometries,
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

International Journal of Heat Fluid Flow, 70 (2018), pp. 348–362, https://doi.org/10.1016/j.

ijheatfluidflow.2017.12.005. (Cited on p. 216)

[126] M. C IANFERRA , S. I ANNIELLO , AND V. A RMENIO, Assessment of methodologies for the so-
lution of the Ffowcs Williams and Hawkings equation using LES of incompressible single-phase
flow around a finite-size square cylinder, Journal of Sound and Vibration, 453 (2019), pp. 1–24,
https://doi.org/10.1016/j.jsv.2019.04.001. (Cited on p. 216)

[127] P. G. C IARLET, Mathematical Elasticity, Volume I: Three-Dimensional Elasticity, Elsevier Science,

1988, https://doi.org/10.1016/s0168-2024(08)70054-7. (Cited on p. 116)

[128] P. G. C IARLET, Mathematical Elasticity: Volume II: Theory of Plates, Studies in Mathemat-
ics and Its Applications, Elsevier Science, 1997, https://doi.org/10.1016/s0168-2024(97)
x8001-4. (Cited on p. 109)

[129] P. G. C IARLET, Linear and Nonlinear Functional Analysis with Applications, Other Titles in Ap-
plied Mathematics, Society for Industrial and Applied Mathematics, 2013. (Cited on pp. 99, 103,
104)

[130] D. C. C IRE ŞAN , U. M EIER , L. M. G AMBARDELLA , AND J. S CHMIDHUBER, Deep, big, sim-
ple neural nets for handwritten digit recognition, Neural Computation, 22 (2010), pp. 3207–3220,
https://doi.org/10.1162/neco_a_00052. (Cited on p. 389)

[131] B. C OCKBURN AND B. D ONG, An analysis of the minimal dissipation local discontinuous Galerkin
method for convection–diffusion problems, Journal of Scientific Computing, 32 (2007), pp. 233–262,
https://doi.org/10.1007/s10915-007-9130-3. (Cited on p. 234)

[132] R. C ODINA, Stabilization of incompressibility and convection through orthogonal sub-scales in finite
element methods, Computer Methods in Applied Mechanics and Engineering, 190 (2000), pp. 1579–
1599, https://doi.org/10.1016/s0045-7825(00)00254-1. (Cited on pp. 132, 133)

[133] A. C OHEN AND R. D E VORE, Kolmogorov widths under holomorphic mappings, IMA Journal of
Numerical Analysis, 36 (2015), pp. 1–12, https://doi.org/10.1093/imanum/dru066. (Cited
on pp. 32, 106, 297)

[134] P. G. C ONSTANTINE, Active Subspaces: Emerging Ideas for Dimension Reduction in Pa-
rameter Studies, vol. 2 of SIAM Spotlights, SIAM, 2015, https://doi.org/10.1137/1.
9781611973860. (Cited on pp. 9, 220, 327, 328, 330, 338, 384)

[135] P. G. C ONSTANTINE AND P. D IAZ, Global sensitivity metrics from active subspaces, Reliability
Engineering & System Safety, 162 (2017), pp. 1–13, https://doi.org/10.1016/j.ress.2017.
01.013. (Cited on pp. 340, 341)

[136] P. G. C ONSTANTINE , E. D OW, AND Q. WANG, Active subspace methods in theory and practice:
Applications to kriging surfaces, SIAM Journal on Scientific Computing, 36 (2014), pp. A1500–
A1524, https://doi.org/10.1137/130916138. (Cited on p. 327)

[137] M. C OUPLET, P. S AGAUT, AND C. BASDEVANT, Intermodal energy transfers in a proper orthog-
onal decomposition-Galerkin representation of a turbulent separated flow, Journal of Fluid Me-
chanics, 491 (2003), pp. 257–284, https://doi.org/10.1017/S0022112003005615. (Cited on
p. 164)

[138] R. C OURANT, K. F RIEDRICHS , AND H. L EWY, Über die partiellen Differenzengleichungen der
mathematischen Physik, Mathematische Annalen, 100 (1928), pp. 32–74, https://doi.org/10.
1007/bf01448839. (Cited on pp. 149, 187)
 Bibliography 425

[139] R. C OURANT, K. F RIEDRICHS , AND H. L EWY, On the partial difference equations of mathematical
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

physics, IBM Journal of Research and Development, 11 (1967), pp. 215–234, https://doi.org/
10.1147/rd.112.0215. (Cited on pp. 149, 187)

[140] M. C OURBARIAUX , I. H UBARA , D. S OUDRY, R. E L -YANIV, AND Y. B ENGIO, Binarized Neural

Networks: Training Deep Neural Networks with Weights and Activations Constrained to +1 or −1,
arXiv preprint arXiv:1602.02830 (2016). (Cited on p. 403)

[141] G. W. C OWLES , N. PAROLINI , AND M. L. S AWLEY, Numerical simulation using RANS-based tools
for America’s Cup design, Proceedings of the 16th Chesapeake Sailing Yacht Symposium (2003),
https://doi.org/10.5957/CSYS-2003-007. (Cited on p. 168)

[142] C. C UI , K. Z HANG , T. DAULBAEV, J. G USAK , I. O SELEDETS , AND Z. Z HANG, Active subspace of

neural networks: Structural analysis and universal attacks, SIAM Journal on Mathematics of Data
Science, 2 (2020), pp. 1096–1122, https://doi.org/10.1137/19M1296070. (Cited on pp. 389,
403, 404, 406, 407, 408)

[143] G. C YBENKO, Approximation by superpositions of a sigmoidal function, Mathematics of Control,

Signals and Systems, 2 (1989), pp. 303–314, https://doi.org/10.1007/BF02551274. (Cited
on pp. 389, 397)

[144] S. D E G ROOT AND P. M AZUR, Non-equilibrium Thermodynamics (1962), Dover Publications,

2013. (Cited on p. 280)

[145] J. C. DE LOS R EYES AND F. T RÖLTZSCH, Optimal control of the stationary Navier–Stokes equa-
tions with mixed control-state constraints, SIAM Journal on Control and Optimization, 46 (2007),
pp. 604–629, https://doi.org/10.1137/050646949. (Cited on p. 83)

[146] G. D E VAHL DAVIS, Natural convection of air in a square cavity: A bench mark numerical solution,
International Journal for Numerical Methods in Fluids, 3 (1983), pp. 249–264, https://doi.org/
10.1002/fld.1650030305. (Cited on p. 75)

[147] M. A. B. D EAKIN, Modelling Biological Systems, Birkhäuser Boston, 1990, pp. 2–16, https:
//doi.org/10.1007/978-1-4684-6784-0_1. (Cited on p. 413)

[148] L. D EDÈ, Optimal flow control for Navier–Stokes equations: Drag minimization, International Jour-
nal for Numerical Methods in Fluids, 55 (2007), pp. 347–366, https://doi.org/10.1002/fld.
1464. (Cited on p. 83)

[149] L. D EDÈ, Reduced basis method and a posteriori error estimation for parametrized linear-quadratic
optimal control problems, SIAM Journal on Scientific Computing, 32 (2010), pp. 997–1019, https:
//doi.org/10.1137/090760453. (Cited on pp. 83, 87)

[150] L. D EDÈ, Reduced basis method and error estimation for parametrized optimal control problems
with control constraints, Journal of Scientific Computing, 50 (2012), pp. 287–305, https://doi.
org/10.1007/s10915-011-9483-5. (Cited on p. 367)

[151] E. D ELGADO ÁVILA, Development of Reduced Numeric Models to Aero-thermic Flows in Buildings,
PhD thesis, University of Seville, 2018. (Cited on pp. 60, 69, 136)

[152] N. D EMO , G. O RTALI , G. G USTIN , G. ROZZA , AND G. L AVINI, An efficient computational

framework for naval shape design and optimization problems by means of data-driven reduced
order modeling techniques, Bollettino dell’Unione Matematica Italiana, 14 (2021), pp. 211–230,
https://doi.org/10.1007/s40574-020-00263-4. (Cited on pp. 212, 213, 214, 383)

[153] N. D EMO , M. S TRAZZULLO , AND G. ROZZA, An Extended Physics Informed Neural Network for
Preliminary Analysis of Parametric Optimal Control Problems, arXiv preprint arXiv:2110.13530
(2021). (Cited on pp. 95, 401)
 426 Bibliography

[154] N. D EMO , M. T EZZELE , G. G USTIN , G. L AVINI , AND G. ROZZA, Shape optimization by

Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

means of proper orthogonal decomposition and dynamic mode decomposition, in Technology and
Science for the Ships of the Future: Proceedings of NAV 2018: 19th International Conference
on Ship & Maritime Research, IOS Press, 2018, pp. 212–219, https://doi.org/10.3233/
978-1-61499-870-9-212. (Cited on pp. 11, 212, 213, 214, 347, 356)
[155] N. D EMO , M. T EZZELE , A. M OLA , AND G. ROZZA, An efficient shape parametrisation by free-
form deformation enhanced by active subspace for hull hydrodynamic ship design problems in open
source environment, in Proceedings of ISOPE 2018: The 28th International Ocean and Polar Engi-
neering Conference, vol. 3, 2018, pp. 565–572. (Cited on pp. 327, 347, 358, 360, 383)

[156] N. D EMO , M. T EZZELE , A. M OLA , AND G. ROZZA, A complete data-driven framework for the
efficient solution of parametric shape design and optimisation in naval engineering problems, in
Proceedings of MARINE 2019: VIII International Conference on Computational Methods in Marine
Engineering, R. Bensow and J. Ringsberg, eds., 2019, pp. 111–121. (Cited on p. 347)
[157] N. D EMO , M. T EZZELE , A. M OLA , AND G. ROZZA, Hull shape design optimization with param-
eter space and model reductions, and self-learning mesh morphing, Journal of Marine Science and
Engineering, 9 (2021), p. 185, https://doi.org/10.3390/jmse9020185. (Cited on pp. 11, 327,
340, 346, 363, 383, 384, 411)

[158] N. D EMO , M. T EZZELE , AND G. ROZZA, EZyRB: Easy reduced basis method, The Journal of Open
Source Software, 3 (2018), p. 661, https://doi.org/10.21105/joss.00661. (Cited on pp. 204,
375, 382)
[159] N. D EMO , M. T EZZELE , AND G. ROZZA, PyDMD: Python dynamic mode decomposition, The
Journal of Open Source Software, 3 (2018), p. 530, https://doi.org/10.21105/joss.00530.
(Cited on pp. 207, 383)

[160] N. D EMO , M. T EZZELE , AND G. ROZZA, A non-intrusive approach for reconstruction of POD
modal coefficients through active subspaces, Comptes rendus mécanique de l’Académie des Sci-
ences, DataBEST 2019 Special Issue, 347 (2019), pp. 873–881, https://doi.org/10.1016/j.
crme.2019.11.012. (Cited on pp. 206, 342, 347)
[161] N. D EMO , M. T EZZELE , AND G. ROZZA, A supervised learning approach involving active sub-
spaces for an efficient genetic algorithm in high-dimensional optimization problems, SIAM Journal
on Scientific Computing, 43 (2021), pp. B831–B853, https://doi.org/10.1137/20M1345219.
(Cited on pp. 11, 327, 340, 350, 384)

[162] S. D EPARIS, Reduced basis error bound computation of parameter-dependent Navier–Stokes equa-
tions by the natural norm approach, SIAM Journal on Numerical Analysis, 46 (2008), pp. 2039–
2067, https://doi.org/10.1137/060674181. (Cited on p. 131)
[163] S. D EPARIS , D. F ORTI , AND A. Q UARTERONI, A rescaled localized radial basis function interpola-
tion on non-Cartesian and nonconforming grids, SIAM Journal on Scientific Computing, 36 (2014),
pp. A2745–A2762, https://doi.org/10.1137/130947179. (Cited on p. 350)

[164] S. D EPARIS AND G. ROZZA, Reduced basis method for multi-parameter-dependent steady Navier-
Stokes equations: Applications to natural convection in a cavity, Journal of Computational Physics,
228 (2009), pp. 4359–4378, https://doi.org/10.1016/j.jcp.2009.03.008. (Cited on pp. 5,
131)
[165] A. D ERKEVORKIAN , P. AVERY, C. FARHAT, J. R ABINOVITCH , AND L. P ETERSON, Effects of
Structural Parameters on the FSI Simulation of Supersonic Parachute Deployments, in AIAA Avia-
tion 2019 Forum, Dallas, Texas, https://doi.org/10.2514/6.2019-3276. (Cited on p. 283)

[166] M. D EVILLE , P. F ISCHER , AND E. M UND, High-Order Methods for Incompressible Fluid Flow,
Cambridge University Press, 2002, https://doi.org/10.1017/CBO9780511546792. (Cited on
p. 223)
 Bibliography 427

[167] P. D IAZ , P. C ONSTANTINE , K. K ALMBACH , E. J ONES , AND S. PANKAVICH, A modified SEIR

Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

model for the spread of Ebola in Western Africa and metrics for resource allocation, Applied Math-
ematics and Computation, 324 (2018), pp. 141–155, https://doi.org/10.1016/j.amc.2017.
11.039. (Cited on pp. xxxiv, 339)

[168] M. D IEZ , E. F. C AMPANA , AND F. S TERN, Design-space dimensionality reduction in shape opti-
mization by Karhunen–Loève expansion, Computer Methods in Applied Mechanics and Engineering,
283 (2015), pp. 1525–1544, https://doi.org/10.1016/j.cma.2014.10.042. (Cited on p. 347)

[169] E. J. D OEDEL, Lecture Notes on Numerical Analysis of Nonlinear Equations, Springer Netherlands,
2007, pp. 1–49, https://doi.org/10.1007/978-1-4020-6356-5_1. (Cited on p. 104)

[170] E. J. D OEDEL , H. B. K ELLER , AND J. P. K ERNEVEZ, Numerical analysis and control of bifurcation
problems (i): Bifurcation in finite dimensions, International Journal of Bifurcation and Chaos, 01
(1991), pp. 493–520, https://doi.org/10.1142/S0218127491000397. (Cited on pp. 104, 316)

[171] V. D OLEJŠÍ AND M. F EISTAUER, Discontinuous Galerkin Method, Analysis and Applications to
Compressible Flow. Springer Series in Computational Mathematics, 48, Springer, 2015, https:
//doi.org/10.1007/978-3-319-19267-3. (Cited on p. 234)

[172] E. D ONADINI , M. S TRAZZULLO , M. T EZZELE , AND G. ROZZA, A Data-Driven Partitioned Ap-

proach for the Resolution of Time-Dependent Optimal Control Problems with Dynamic Mode De-
composition, arXiv preprint arXiv:2111.13906 (2021). (Cited on pp. 95, 210)

[173] J. D ONEA , S. G IULIANI , AND J. P. H ALLEUX, An arbitrary Lagrangian–Eulerian finite ele-

ment method for transient dynamic fluid–structure interactions, Computer Methods in Applied Me-
chanics and Engineering, 33 (1982), pp. 689–723, https://doi.org/10.1016/0045-7825(82)
90128-1. (Cited on pp. 284, 312)

[174] J. D OUGLAS AND T. D UPONT, Interior penalty procedures for elliptic and parabolic Galerkin
methods, in Computing Methods in Applied Sciences, Springer, 1976, pp. 207–216, https:
//doi.org/10.1007/BFb0120591. (Cited on p. 239)

[175] M. D ROHMANN , B. H AASDONK , AND M. O HLBERGER, Reduced basis approximation for non-
linear parametrized evolution equations based on empirical operator interpolation, SIAM Journal
on Scientific Computing, 34 (2012), pp. A937–A969, https://doi.org/10.1137/10081157X.
(Cited on p. 143)

[176] J. D UCHON, Splines minimizing rotation-invariant semi-norms in Sobolev spaces, in Constructive

Theory of Functions of Several Variables, Springer, 1977, pp. 85–100, https://doi.org/10.
1007/BFb0086566. (Cited on p. 348)

[177] D. DYLEWSKY, M. TAO , AND J. N. K UTZ, Dynamic mode decomposition for multiscale nonlinear
physics, Physical Review E, 99 (2019), p. 063311, https://doi.org/10.1103/PhysRevE.99.
063311. (Cited on p. 210)

[178] C. E CKART AND G. YOUNG, The approximation of one matrix by another of lower rank, Psychome-
trika, 1 (1936), pp. 211–218, https://doi.org/10.1007/BF02288367. (Cited on p. 4)

[179] J. E DMONDS, Matroids and the greedy algorithm, Mathematical Programming, 1 (1971), pp. 127–
136, https://doi.org/10.1007/BF01584082. (Cited on p. 131)

[180] J. L. E FTANG , D. J. K NEZEVIC , AND A. T. PATERA, An hp certified reduced basis method

for parametrized parabolic partial differential equations, Mathematical and Computer Modelling
of Dynamical Systems, 17 (2011), pp. 395–422, https://doi.org/10.1080/13873954.2011.
547670. (Cited on p. 5)
 428 Bibliography

[181] V. E HRLACHER , D. L OMBARDI , O. M ULA , AND F. X. V IALARD, Nonlinear Model Reduc-

Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

tion on Metric Spaces. Application to One-Dimensional Conservative PDEs in Wasserstein Spaces.

arXiv:1909.06626, 2019, https://doi.org/10.1051/m2an/2020013. (Cited on pp. 297, 298)

[182] J. L. E LMAN, Finding structure in time, Cognitive Science, 14 (1990), pp. 179–211, https://doi.
org/10.1207/s15516709cog1402_1. (Cited on p. 393)

[183] N. B. E RICHSON , S. L. B RUNTON , AND J. N. K UTZ, Compressed dynamic mode decomposi-

tion for background modeling, Journal of Real-Time Image Processing, 16 (2019), pp. 1479–1492,
https://doi.org/10.1007/s11554-016-0655-2. (Cited on pp. 209, 210)

[184] N. B. E RICHSON , L. M ATHELIN , J. N. K UTZ , AND S. L. B RUNTON, Randomized dynamic mode

decomposition, SIAM Journal on Applied Dynamical Systems, 18 (2019), pp. 1867–1891, https:
//doi.org/10.1137/18M1215013. (Cited on p. 209)

[185] D. E RRATE , M. J. E STEBAN , AND Y. M ADAY, Couplage fluid–structure. Un modele simplifié en

dimension 1., Comptes rendus de l’Académie des sciences. Série 1, Mathématique, 318 (1994),
pp. 275–281. (Cited on p. 283)

[186] L. C. E VANS, Partial Differential Equations, Graduate Studies in Mathematics, American Mathe-
matical Society, 2010, https://doi.org/10.1090/gsm/019. (Cited on p. 100)

[187] C. FARHAT AND V. K. L AKSHMINARAYAN, An ALE formulation of embedded boundary methods

for tracking boundary layers in turbulent fluid–structure interaction problems, Journal of Computa-
tional Physics, 263 (2014), pp. 53–70, https://doi.org/10.1016/j.jcp.2014.01.018. (Cited
on p. 283)

[188] C. FARHAT, K. G. VAN DER Z EE , AND P. G EUZAINE, Provably second-order time-accurate

loosely-coupled solution algorithms for transient nonlinear computational aeroelasticity, Computer
Methods in Applied Mechanics and Engineering, 195 (2006), pp. 1973–2001, https://doi.org/
10.1016/j.cma.2004.11.031. (Cited on p. 286)

[189] P. E. FARRELL , A. B IRKISSON , AND S. W. F UNKE, Deflation techniques for finding distinct solu-
tions of nonlinear partial differential equations, SIAM Journal on Scientific Computing, 37 (2015),
pp. A2026–A2045, https://doi.org/10.1137/140984798. (Cited on p. 114)

[190] G. E. FASSHAUER AND J. G. Z HANG, On choosing “optimal” shape parameters for RBF
approximation, Numerical Algorithms, 45 (2007), pp. 345–368, https://doi.org/10.1007/
s11075-007-9072-8. (Cited on p. 190)

[191] M. A. F ERNÁNDEZ, Incremental displacement-correction schemes for incompressible fluid–

structure interaction, Numerische Mathematik, 123 (2013), pp. 21–65, https://doi.org/10.
1007/s00211-012-0481-9. (Cited on p. 286)

[192] M. A. F ERNÁNDEZ , J. F. G ERBEAU , AND C. G RANDMONT, A projection semi-implicit scheme for

the coupling of an elastic structure with an incompressible fluid, International Journal for Numerical
Methods in Engineering, 69 (2007), pp. 794–821, https://doi.org/10.1002/nme.1792. (Cited
on p. 286)

[193] M. A. F ERNÁNDEZ , M. L ANDAJUELA , AND M. V IDRASCU, Fully decoupled time-marching

schemes for incompressible fluid/thin-walled structure interaction, Journal of Computational
Physics, 297 (2015), pp. 156–181, https://doi.org/10.1016/j.jcp.2015.05.009. (Cited on
p. 286)

[194] M. A. F ERNÁNDEZ , J. M ULLAERT, AND M. V IDRASCU, Explicit Robin–Neumann schemes for the
coupling of incompressible fluids with thin-walled structures, Computer Methods in Applied Me-
chanics and Engineering, 267 (2013), pp. 566–593, https://doi.org/10.1016/j.cma.2013.
09.020. (Cited on p. 286)
 Bibliography 429

[195] J. E. F FOWCS W ILLIAMS AND D. L. H AWKINGS, Sound generation by turbulence and surfaces
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

in arbitrary motion, Philosophical Transactions of the Royal Society A, 264 (1969), pp. 321–342,
https://doi.org/10.1098/rsta.1969.0031. (Cited on p. 214)

[196] L. F ICK , Y. M ADAY, A. T. PATERA , AND T. TADDEI, A stabilized POD model for turbulent flows
over a range of Reynolds numbers: Optimal parameter sampling and constrained projection, Journal
of Computational Physics, 371 (2018), pp. 214–243. (Cited on p. 164)

[197] T. L. F INE, Feedforward Neural Network Methodology, Springer Science & Business Media, 2006.
(Cited on p. 393)

[198] J. P. F INK AND W. C. R HEINBOLDT, On the discretization error of parametrized nonlinear equa-
tions, SIAM Journal on Numerical Analysis, 20 (1983), pp. 732–746, https://doi.org/10.
1137/0720049. (Cited on p. 3)

[199] J. P. F INK AND W. C. R HEINBOLDT, On the error behavior of the reduced basis technique
for nonlinear finite element approximations, ZAMM-Zeitschrift für Angewandte Mathematik und
Mechanik, 63 (1983), pp. 21–28, https://doi.org/10.1002/zamm.19830630105. (Cited on
p. 3)

[200] L. F ORMAGGIA , E. M IGLIO , A. M OLA , AND N. PAROLINI, Fluid–structure interaction problems

in free surface flows: Application to boat dynamics, International Journal for Numerical Methods in
Fluids, 56 (2008), pp. 965–978, https://doi.org/10.1002/fld.1583. (Cited on p. 168)

[201] D. F ORTI , M. B UKAC , A. Q UAINI , S. C ANIC , AND S. D EPARIS, A monolithic approach to fluid–
composite structure interaction, Journal of Scientific Computing, 72 (2017), pp. 396–421, https:
//doi.org/10.1007/s10915-017-0363-5. (Cited on p. 287)

[202] D. F ORTI AND G. ROZZA, Efficient geometrical parametrisation techniques of interfaces for
reduced-order modelling: Application to fluid–structure interaction coupling problems, International
Journal of Computational Fluid Dynamics, 28 (2014), pp. 158–169, https://doi.org/10.1080/
10618562.2014.932352. (Cited on pp. 350, 351)

[203] M. F OWLER, Refactoring: Improving the Design of Existing Code, Addison-Wesley Professional,
2018. (Cited on p. 386)

[204] L. P. F RANCA AND C. FARHAT, Bubble functions prompt unusual stabilized finite element methods,
Computer Methods in Applied Mechanics and Engineering, 123 (1995), pp. 299–308, https://
doi.org/10.1016/0045-7825(94)00721-x. (Cited on p. 132)

[205] L. P. F RANCA AND S. L. F REY, Stabilized finite element methods: II. The incompressible Navier-
Stokes equations, Computer Methods in Applied Mechanics and Engineering, 99 (1992), pp. 209–
233, https://doi.org/10.1016/0045-7825(92)90041-H. (Cited on p. 130)

[206] B. A. F RENO AND K. T. C ARLBERG, Machine-learning error models for approximate solutions to
parameterized systems of nonlinear equations, Computer Methods in Applied Mechanics and Engi-
neering, 348 (2019), pp. 250–296, https://doi.org/10.1016/j.cma.2019.01.024. (Cited on
p. 298)

[207] S. F U AND L. WANG, RANS modeling of high-speed aerodynamic flow transition with consideration
of stability theory, Progress in Aerospace Sciences, 58 (2013), pp. 36–59, https://doi.org/10.
1016/j.paerosci.2012.08.004. (Cited on p. 168)

[208] T. F UJIWARA , F. K AJIYA , S. K ANAZAWA , S. M ATSUOKA , Y. WADA , O. H IRAMATSU ,

M. K AGIYAMA , Y. O GASAWARA , K. T SUJIOKA , AND T. K ATSUMURA, Comparison of blood-
flow velocity waveforms in different coronary artery bypass grafts: Sequential saphenous vein grafts
and internal mammary artery grafts, Circulation, 78 (1988), pp. 1210–1217, https://doi.org/
10.1161/01.cir.78.5.1210. (Cited on p. 368)
 430 Bibliography

[209] M. G ADALLA , M. C IANFERRA , M. T EZZELE , G. S TABILE , A. M OLA , AND G. ROZZA, On the

Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

comparison of LES data-driven reduced order approaches for hydroacoustic analysis, Computers &
Fluids, 216 (2021), p. 104819, https://doi.org/10.1016/j.compfluid.2020.104819. (Cited
on pp. 214, 381, 383)

[210] M. G ADALLA , M. T EZZELE , A. M OLA , AND G. ROZZA, BladeX: Python blade morphing, The
Journal of Open Source Software, 4 (2019), p. 1203, https://doi.org/10.21105/joss.01203.
(Cited on pp. 355, 359, 385)

[211] G. G ALDI, An Introduction to the Mathematical Theory of the Navier-Stokes Equations: Steady-
State Problems, Springer Science & Business Media, 2011. (Cited on p. 257)

[212] C. G ARCÍA -A ZPEITIA AND D. P ELINOVSKY, Bifurcations of multi-vortex configurations in rotat-

ing Bose–Einstein condensates, Milan Journal of Mathematics, 85 (2017), pp. 331–367, https:
//doi.org/10.1007/s00032-017-0275-8. (Cited on p. 112)

[213] F. G AROTTA , N. D EMO , M. T EZZELE , M. C ARRATURO , A. R EALI , AND G. ROZZA, Reduced

order isogeometric analysis approach for PDEs in parametrized domains, in Quantification of
Uncertainty: Improving Efficiency and Technology: QUIET Selected Contributions, M. D’Elia,
M. Gunzburger, and G. Rozza, eds., vol. 137 of Lecture Notes in Computational Science and En-
gineering, Springer International Publishing, 2020, pp. 153–170, https://doi.org/10.1007/
978-3-030-48721-8_7. (Cited on pp. 206, 347, 383)

[214] A. G ARZOZI AND D. G REENBLATT, A pulsed Coandă-effect reciprocating wind energy generator,
Energy, 150 (2018), pp. 965–978, https://doi.org/10.1016/j.energy.2018.02.114. (Cited
on p. 373)

[215] R. G AUTIER , P. PANDITA , S. G HOSH , AND D. M AVRIS, A fully Bayesian gradient-

free supervised dimension reduction method using Gaussian processes, International Journal
for Uncertainty Quantification, 12 (2022), pp. 19–51, https://doi.org/10.1615/Int.J.
UncertaintyQuantification.2021035621. (Cited on p. 336)

[216] M. W. G EE , U. K ÜTTLER , AND W. A. WALL, Truly monolithic algebraic multigrid for fluid–
structure interaction, International Journal for Numerical Methods in Engineering, 85 (2011),
pp. 987–1016, https://doi.org/10.1002/nme.3001. (Cited on p. 287)

[217] R. G EELEN AND K. W ILLCOX, Localized non-intrusive reduced-order modeling in the operator
inference framework, Philosophical Transactions of the Royal Society A, 380 (2022), p. 20210206,
https://doi.org/10.1098/rsta.2021.0206. (Cited on p. 222)

[218] J. G ENOVESE , F. BALLARIN , G. ROZZA , AND C. C ANUTO, Weighted Reduced Order Methods
for Uncertainty Quantification in Computational Fluid Dynamics, in preparation (2022). (Cited on
p. 257)

[219] S. G EORGAKA , G. S TABILE , G. ROZZA , AND M. J. B LUCK, Parametric POD-Galerkin model or-
der reduction for unsteady-state heat transfer problems, Communications in Computational Physics,
27 (2020), pp. 1–32, https://doi.org/10.4208/cicp.oa-2018-0207. (Cited on p. 144)

[220] S. G EORGAKA , G. S TABILE , K. S TAR , G. ROZZA , AND M. J. B LUCK, A hybrid reduced order
method for modelling turbulent heat transfer problems, Computers & Fluids, 208 (2020), p. 104615,
https://doi.org/10.1016/j.compfluid.2020.104615. (Cited on pp. 205, 350, 381)

[221] J.-F. G ERBEAU AND D. L OMBARDI, Approximated Lax pairs for the reduced order integration of
nonlinear evolution equations, Journal of Computational Physics, 265 (2014), pp. 246–269, https:
//doi.org/10.1016/j.jcp.2014.01.047. (Cited on p. 298)

[222] J. F. G ERBEAU , D. L OMBARDI , AND E. S CHENONE, Reduced order model in cardiac electrophys-
iology with approximated Lax pairs, Advances in Computational Mathematics, 41 (2015), pp. 1103–
1130, https://doi.org/10.1007/s10444-014-9393-9. (Cited on p. 298)
 Bibliography 431

[223] A.-L. G ERNER AND K. V EROY, Certified reduced basis methods for parametrized saddle point
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

problems, SIAM Journal on Scientific Computing, 34 (2012), pp. A2812–A2836, https://doi.

org/10.1137/110854084. (Cited on pp. 51, 52, 83)

[224] R. G. G HANEM AND P. D. S PANOS, Stochastic Finite Elements: A Spectral Approach, Courier
Corporation, 2003. (Cited on p. 405)

[225] M. G HASHAMI , E. L IBERTY, J. M. P HILLIPS , AND D. P. W OODRUFF, Frequent directions: Simple

and deterministic matrix sketching, SIAM Journal on Computing, 45 (2016), pp. 1762–1792, https:
//doi.org/10.1137/15M1009718. (Cited on p. 408)

[226] S. F. G HOREISHI , S. F RIEDMAN , AND D. L. A LLAIRE, Adaptive dimensionality reduction for

fast sequential optimization with Gaussian processes, Journal of Mechanical Design, 141 (2019),
p. 071404, https://doi.org/10.1115/1.4043202. (Cited on p. 340)

[227] V. G IRAULT AND P. A. R AVIART, Finite Element Approximation of the Navier-Stokes Equations,
Springer, 1981. (Cited on p. 8)

[228] M. G IRFOGLIO , F. BALLARIN , G. I NFANTINO , F. N ICOLÓ , A. M ONTALTO , G. ROZZA , R. S CRO -

FANI , M. C OMISSO , AND F. M USUMECI , Non-intrusive PODI-ROM for Patient-Specific Aortic
Blood Flow in Presence of a LVAD Device. arXiv:2007.03527, 2021. (Cited on pp. 374, 375)

[229] M. G IRFOGLIO , A. Q UAINI , AND G. ROZZA, A finite volume approximation of the Navier-Stokes
equations with nonlinear filtering stabilization, Computers & Fluids, 187 (2019), pp. 27–45, https:
//doi.org/10.1016/j.compfluid.2019.05.001. (Cited on p. 171)

[230] M. G IRFOGLIO , A. Q UAINI , AND G. ROZZA, A POD-Galerkin reduced order model for a LES
filtering approach, Journal of Computational Physics, 436 (2021), p. 110260, https://doi.org/
10.1016/j.jcp.2021.110260. (Cited on pp. 171, 205)

[231] M. G IRFOGLIO , A. Q UAINI , AND G. ROZZA, Fluid-structure interaction simulations with a LES fil-
tering approach in solids4Foam, Communications in Applied and Industrial Mathematics, 12 (2021),
pp. 13–28, https://doi.org/doi:10.2478/caim-2021-0002. (Cited on p. 171)

[232] M. G IRFOGLIO , A. Q UAINI , AND G. ROZZA, A Hybrid Reduced Order Model for Nonlinear LES
Filtering. arXiv:2107.12933, 2021. (Cited on p. 171)

[233] M. G IRFOGLIO , A. Q UAINI , AND G. ROZZA, Pressure stabilization strategies for a LES filtering
reduced order model, Fluids, 6 (2021), https://doi.org/10.3390/fluids6090302. (Cited on
p. 171)

[234] M. G IRFOGLIO , L. S CANDURRA , F. BALLARIN , G. I NFANTINO , F. N ICOLO , A. M ONTALTO ,

G. ROZZA , R. S CROFANI , M. C OMISSO , AND F. M USUMECI, Non-intrusive data-driven ROM
framework for hemodynamics problems, Acta Mechanica Sinica, 37 (2021), pp. 1183–1191, https:
//doi.org/10.1007/s10409-021-01090-2. (Cited on pp. 374, 383)

[235] S. G LAS , A. M AYERHOFER , AND K. U RBAN, Two Ways to Treat Time in Reduced Ba-
sis Methods, Springer International Publishing, 2017, pp. 1–16, https://doi.org/10.1007/
978-3-319-58786-8_1. (Cited on pp. 83, 92)

[236] G. H. G OLUB AND C. F. VAN L OAN, Matrix Computations, third edition, Johns Hopkins University
Press, 1996. (Cited on p. 4)

[237] I. G OODFELLOW, Y. B ENGIO , AND A. C OURVILLE, Deep Learning, MIT Press, 2016, http:
//www.deeplearningbook.org. (Cited on pp. 11, 389, 390, 391, 396)

[238] J. G OU , B. Y U , S. J. M AYBANK , AND D. TAO, Knowledge distillation: A survey, Interna-

tional Journal of Computer Vision, 129 (2021), pp. 1789–1819, https://doi.org/10.1007/
s11263-021-01453-z. (Cited on p. 406)
 432 Bibliography

[239] GP Y, GPy: A Gaussian process framework in Python. http://github.com/SheffieldML/GPy,

Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

since 2012. (Cited on p. 220)

[240] W. R. G RAHAM , J. P ERAIRE , AND K. Y. TANG, Optimal control of vortex shedding us-
ing low-order models. Part I–Open-loop model development, International Journal for Numer-
ical Methods in Engineering, 44 (1999), pp. 945–972, https://doi.org/10.1002/(sici)
1097-0207(19990310)44:7<945::aid-nme537>3.0.co;2-f. (Cited on pp. 154, 155, 189)

[241] C. G RANDMONT, V. G UIMET, AND Y. M ADAY, Numerical analysis of some decoupling tech-
niques for the approximation of the unsteady fluid–structure interaction, Mathematical Models
and Methods in Applied Sciences, 11 (2001), pp. 1349–1377, https://doi.org/10.1142/
S0218202501001367. (Cited on p. 286)

[242] C. G RANDMONT AND Y. M ADAY, Existence de solutions d’un probléme de couplage fluide-
structure bidimensionnel instationnaire, Comptes Rendus de l’Académie des Sciences—Series I—
Mathematics, 326 (1998), pp. 525–530. (Cited on p. 283)

[243] A. G RAVES , S. F ERNÁNDEZ , F. G OMEZ , AND J. S CHMIDHUBER, Connectionist temporal classifi-

cation: Labelling unsegmented sequence data with recurrent neural networks, in Proceedings of the
23rd International Conference on Machine Learning, 2006, pp. 369–376, https://doi.org/10.
1145/1143844.1143891. (Cited on pp. 389, 402)

[244] C. G REIF AND K. U RBAN, Decay of the Kolmogorov N -width for wave problems, Applied Math-
ematics Letters, 96 (2019), pp. 216–222, https://doi.org/10.1016/j.aml.2019.05.013.
(Cited on p. 297)

[245] M. A. G REPL , Y. M ADAY, N. C. N GUYEN , AND A. T. PATERA, Efficient reduced-basis treatment

of nonaffine and nonlinear partial differential equations, ESAIM: Mathematical Modelling and Nu-
merical Analysis, 41 (2007), pp. 575–605, https://doi.org/10.1051/m2an:2007031. (Cited
on pp. 4, 59, 63, 316)

[246] M. A. G REPL AND A. T. PATERA, A posteriori error bounds for reduced-basis approximations of
parametrized parabolic partial differential equations, ESAIM: Mathematical Modelling and Numer-
ical Analysis, 39 (2005), pp. 157–181. (Cited on p. 53)

[247] B. G RIMSTAD, SPLINTER: a library for multivariate function approximation with splines. http:
//github.com/bgrimstad/splinter, 2015. (Cited on p. 182)

[248] A. G RUBER , M. G UNZBURGER , L. J U , Y. T ENG , AND Z. WANG, Nonlinear level set learning
for function approximation on sparse data with applications to parametric differential equations,
Numerical Mathematics: Theory, Methods and Applications, 14 (2021), pp. 839–861, https://
doi.org/10.4208/nmtma.OA-2021-0062. (Cited on p. 337)

[249] M. G UNZBURGER , N. J IANG , AND M. S CHNEIER, An ensemble-proper orthogonal decomposition

method for the nonstationary Navier–Stokes equations, SIAM Journal on Numerical Analysis, 55
(2017), pp. 286–304, https://doi.org/10.1137/16M1056444. (Cited on p. 284)

[250] M. D. G UNZBURGER, Perspectives in Flow Control and Optimization, vol. 5 of Advances in Design
and Control, SIAM, 2002, https://doi.org/10.1137/1.9780898718720. (Cited on p. 367)

[251] M. D. G UNZBURGER , J. S. P ETERSON , AND J. N. S HADID, Reduced-order modeling of time-

dependent PDEs with multiple parameters in the boundary data, Computer Methods in Applied
Mechanics and Engineering, 196 (2007), pp. 1030–1047, https://doi.org/10.1016/j.cma.
2006.08.004. (Cited on p. 154)

[252] M. G UO AND J. S. H ESTHAVEN, Reduced order modeling for nonlinear structural analysis us-
ing Gaussian process regression, Computer Methods in Applied Mechanics and Engineering, 341
(2018), pp. 807–826, https://doi.org/10.1016/j.cma.2018.07.017. (Cited on p. 206)
 Bibliography 433

[253] M. G UO AND J. S. H ESTHAVEN, Data-driven reduced order modeling for time-dependent problems,
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Computer Methods in Applied Mechanics and Engineering, 345 (2019), pp. 75–99, https://doi.
org/10.1016/j.cma.2018.10.029. (Cited on p. 187)

[254] M. G UO , A. M ANZONI , M. A MENDT, P. C ONTI , AND J. S. H ESTHAVEN, Multi-fidelity regression

using artificial neural networks: Efficient approximation of parameter-dependent output quantities,
Computer Methods in Applied Mechanics and Engineering, 389 (2022), p. 114378, https://doi.
org/10.1016/j.cma.2021.114378. (Cited on p. 401)

[255] M. E. G URTIN, An Introduction to Continuum Mechanics, Mathematics in Science and Engineering,

Elsevier Science, 1982. (Cited on pp. 116, 117)

[256] B. H AASDONK, Convergence rates of the POD–greedy method, ESAIM: Mathematical Modelling
and Numerical Analysis, 47 (2013), pp. 859–873, https://doi.org/10.1051/m2an/2012045.
(Cited on p. 5)

[257] B. H AASDONK, Reduced basis methods for parametrized PDEs—A tutorial introduction for station-
ary and instationary problems, in Model Reduction and Approximation: Theory and Algorithms,
P. Benner, A. Cohen, M. Ohlberger, and K. Willcox, eds., Computational Science and Engineering,
ch. 2, SIAM, 2017, pp. 65–136, https://doi.org/10.1137/1.9781611974829.ch2. (Cited on
p. 1)

[258] B. H AASDONK, Mor software, in Model Order Reduction: Volume 3, ch. 13, pp. 431–460, Applica-
tions, P. Benner, S. Grivet-Talocia, A. Quarteroni, G. Rozza, W. Schilders, and L. M. Silveira, eds.,
De Gruyter, 2020, https://doi.org/10.1515/9783110499001-013. (Cited on p. 3)

[259] B. H AASDONK AND M. O HLBERGER, Adaptive basis enrichment for the reduced basis method ap-
plied to finite volume schemes, in Proceedings of the 5th International Symposium on Finite Volumes
for Complex Applications, 2008, pp. 471–478. (Cited on p. 143)

[260] B. H AASDONK AND M. O HLBERGER, Reduced basis method for finite volume approximations of
parametrized linear evolution equations, ESAIM: Mathematical Modelling and Numerical Analysis,
42 (2008), pp. 277–302, https://doi.org/10.1051/m2an:2008001. (Cited on pp. 5, 143)

[261] B. H AASDONK AND M. O HLBERGER, Reduced Basis Method for Explicit Finite Volume Approx-
imations of Nonlinear Conservation Laws, American Mathematical Society, 2009, pp. 605–614,
https://doi.org/10.1090/psapm/067.2/2605256. (Cited on p. 143)

[262] E. H ABER AND L. RUTHOTTO, Stable architectures for deep neural networks, Inverse Problems, 34
(2017), p. 014004, https://doi.org/10.1088/1361-6420/aa9a90. (Cited on p. 337)

[263] S. H AN , H. M AO , AND W. J. DALLY, Deep compression: Compressing deep neural network with
pruning, trained quantization and Huffman coding, in 4th International Conference on Learning
Representations, 2016, arXiv:1510.00149. (Cited on p. 403)

[264] D. H ARTMAN AND L. K. M ESTHA, A deep learning framework for model reduction of dynam-
ical systems, in 2017 IEEE Conference on Control Technology and Applications (CCTA), 2017,
pp. 1917–1922, https://doi.org/10.1109/CCTA.2017.8062736. (Cited on p. 298)

[265] D. H ARTMANN , M. H ERZ , AND U. W EVER, Model order reduction a key technology for dig-
ital twins, in Reduced-Order Modeling (ROM) for Simulation and Optimization, W. Keiper,
A. Milde, and S. Volkwein, eds., Springer, 2018, pp. 167–179, https://doi.org/10.1007/
978-3-319-75319-5_8. (Cited on p. 411)

[266] D. H ARTMANN AND H. VAN DER AUWERAER, Digital twins, in Progress in Industrial Math-
ematics: Success Stories, M. Cruz, C. Parés, and P. Quintela, eds., Springer, 2021, pp. 3–17,
https://doi.org/10.1007/978-3-030-61844-5_1. (Cited on pp. 389, 410)
 434 Bibliography

[267] S. H AYKIN, Neural Networks and Learning Machines, vol. 10 of Neural Networks and Learning
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Machines, Prentice Hall, 2009. (Cited on p. 397)

[268] F. H ECHT, New development in FreeFem++, Journal of Numerical Mathematics, 20 (2012),

pp. 251–265, https://doi.org/10.1515/jnum-2012-0013. (Cited on pp. 66, 74, 138)

[269] T. H ERNANDEZ -B OUSSARD , P. M ACKLIN , E. J. G REENSPAN , A. L. G RYSHUK , E. S TAHLBERG ,

T. S YEDA -M AHMOOD , AND I. S HMULEVICH, Digital twins for predictive oncology will be a
paradigm shift for precision cancer care, Nature Medicine, 27 (2021), pp. 2065–2066, https:
//doi.org/10.1038/s41591-021-01558-5. (Cited on p. 12)

[270] M. H ESS , A. A LLA , A. Q UAINI , G. ROZZA , AND M. G UNZBURGER, A localized reduced-order

modeling approach for PDEs with bifurcating solutions, Computer Methods in Applied Mechanics
and Engineering, 351 (2019), pp. 379–403, https://doi.org/10.1016/j.cma.2019.03.050.
(Cited on pp. 229, 370, 373, 381)

[271] M. H ESS , S. G RUNDEL , AND P. B ENNER, Estimating the inf-sup constant in reduced basis methods
for time-harmonic Maxwell’s equations, IEEE Transactions on Microwave Theory and Techniques,
63 (2015), pp. 1–9, https://doi.org/10.1109/TMTT.2015.2473157. (Cited on pp. 5, 7)

[272] M. W. H ESS , A. L ARIO , G. M ENGALDO , AND G. ROZZA, Reduced Order Modeling for Spectral
Element Methods: Current Developments in Nektar++ and Further Perspectives. arXiv preprint
arXiv:2201.05404, 2022. (Cited on p. 381)

[273] M. W. H ESS , A. Q UAINI , AND G. ROZZA, A spectral element reduced basis method for Navier–
Stokes equations with geometric variations, in Spectral and High Order Methods for Partial Dif-
ferential Equations ICOSAHOM 2018, S. J. Sherwin, D. Moxey, J. Peiró, P. E. Vincent, and
C. Schwab, eds., Springer International Publishing, 2020, pp. 561–571, https://doi.org/10.
1007/978-3-030-39647-3_45. (Cited on pp. 229, 231, 381)

[274] M. W. H ESS , A. Q UAINI , AND G. ROZZA, Reduced basis model order reduction for Navier-Stokes
equations in domains with walls of varying curvature, International Journal of Computational Fluid
Dynamics, 34 (2020), pp. 119–126, https://doi.org/10.1080/10618562.2019.1645328.
(Cited on p. 381)

[275] M. W. H ESS , A. Q UAINI , AND G. ROZZA, A comparison of reduced-order modeling approaches us-
ing artificial neural networks for PDEs with bifurcating solutions, ETNA—Electronic Transactions
on Numerical Analysis, 56 (2022), pp. 52–65, https://doi.org/10.1553/etna_vol56s52.
(Cited on p. 381)

[276] M. W. H ESS AND G. ROZZA, A spectral element reduced basis method in parametric CFD, in
Numerical Mathematics and Advanced Applications—ENUMATH 2017, F. A. Radu, K. Kumar,
I. Berre, J. M. Nordbotten, and I. S. Pop, eds., vol. 126, Springer, 2019, pp. 693–701, https:
//doi.org/10.1007/978-3-319-96415-7_64. (Cited on pp. 231, 381)

[277] M. W. H ESS AND G. ROZZA, Model Reduction Using Sparse Polynomial Interpolation for the
Incompressible Navier-Stokes Equations. arXiv preprint arXiv:2201.03228, 2022. (Cited on p. 381)

[278] J. H ESTHAVEN , B. S TAMM , AND S. Z HANG, Efficient greedy algorithms for high-dimensional
parameter spaces with applications to empirical interpolation and reduced basis methods, ESAIM
Mathematical Modelling and Numerical Analysis, 48 (2014), https://doi.org/10.1051/m2an/
2013100. (Cited on p. 5)

[279] J. H ESTHAVEN AND S. U BBIALI, Non-intrusive reduced order modeling of nonlinear problems
using neural networks, Journal of Computational Physics, 363 (2018), pp. 55–78, https://doi.
org/10.1016/j.jcp.2018.02.037. (Cited on p. 397)
 Bibliography 435

[280] J. H ESTHAVEN AND T. WARBURTON, Nodal high-order methods on unstructured grids: I. Time-
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

domain solution of Maxwell’s equations, Journal of Computational Physics, 181 (2002), pp. 186–
221, https://doi.org/10.1006/jcph.2002.7118. (Cited on p. 235)

[281] J. S. H ESTHAVEN , G. ROZZA , AND B. S TAMM, Certified Reduced Basis Methods for Parametrized
Partial Differential Equations, Springer Briefs in Mathematics, 1 ed., Springer, 2015, https://
doi.org/10.1007/978-3-319-22470-1. (Cited on pp. 1, 5, 15, 83, 86, 87, 92, 102, 106, 108,
205, 251, 253, 254, 267, 327, 347, 367, 368)

[282] J. S. H ESTHAVEN AND T. WARBURTON, Nodal Discontinuous Galerkin Methods: Algorithms,

Analysis, and Applications, Springer Science & Business Media, 2007. (Cited on pp. 234, 241)

[283] R. M. H ICKS AND P. A. H ENNE, Wing design by numerical optimization, Journal of Aircraft, 15
(1978), pp. 407–412, https://doi.org/10.2514/3.58379. (Cited on p. 219)

[284] S. H IJAZI , S. A LI , G. S TABILE , F. BALLARIN , AND G. ROZZA, The effort of increasing Reynolds
number in projection-based reduced order methods: From laminar to turbulent flows, in FEF 2017
Selected Contributions, H. van Brummelen, A. Corsini, S. Perotto, and G. Rozza, eds., vol. 132 of
Lecture Notes in Computational Science and Engineering, Springer International Publishing, 2020,
pp. 245–264, https://doi.org/10.1007/978-3-030-30705-9_22. (Cited on pp. 175, 381)

[285] S. H IJAZI , G. S TABILE , A. M OLA , AND G. ROZZA, Non-intrusive polynomial chaos method ap-
plied to full-order and reduced problems in computational fluid dynamics: A comparison and per-
spectives, in QUIET Selected Contributions, M. D’Elia, M. Gunzburger, and G. Rozza, eds., vol. 137
of Lecture Notes in Computational Science and Engineering, Springer International Publishing,
2020, pp. 217–240, https://doi.org/10.1007/978-3-030-48721-8_10. (Cited on pp. 154,
251)

[286] S. H IJAZI , G. S TABILE , A. M OLA , AND G. ROZZA, Data-driven POD–Galerkin reduced order
model for turbulent flows, Journal of Computational Physics, 416 (2020), p. 109513, https://
doi.org/10.1016/j.jcp.2020.109513. (Cited on pp. 175, 205)

[287] G. H INTON , O. V INYALS , AND J. D EAN, Distilling the Knowledge in a Neural Network, in NIPS
Deep Learning and Representation Learning Workshop, 2015. (Cited on p. 406)

[288] G. E. H INTON AND R. Z EMEL, Autoencoders, minimum description length and Helmholtz free
energy, in Advances in Neural Information Processing Systems, vol. 6, J. Cowan, G. Tesauro, and
J. Alspector, eds., Morgan-Kaufmann, 1994, pp. 3–10. (Cited on p. 396)

[289] M. H INZE , M. K ÖSTER , AND S. T UREK, A space-time multigrid method for optimal flow control,
in Constrained Optimization and Optimal Control for Partial Differential Equations, G. Leugering, S.
Engell, A. Griewank, M. Hinze, R. Rannacher, V. Schulz, M. Ulbrich, and S. Ulbrich, ed., Springer,
2012, pp. 147–170, https://doi.org/10.1007/978-3-0348-0133-1_8. (Cited on p. 89)

[290] M. H INZE , R. P INNAU , M. U LBRICH , AND S. U LBRICH, Optimization with PDE constraints,
vol. 23 of Mathematical Modelling: Theory and Applications, Springer Netherlands, 2009, https:
//doi.org/10.1007/978-1-4020-8839-1. (Cited on pp. 84, 85, 89)

[291] P. H OLMES , J. L. L UMLEY, AND G. B ERKOOZ, Turbulence, Coherent Structures, Dynamical Sys-
tems and Symmetry, Cambridge Monographs on Mechanics, Cambridge University Press, 1996,
https://doi.org/10.1017/CBO9780511622700. (Cited on p. 373)

[292] M. H OLTZ, Sparse Grid Quadrature in High Dimensions with Applications in Finance and Insur-
ance, vol. 77, Springer, 2010. (Cited on p. 256)

[293] K. H ORNIK , M. S TINCHCOMBE , AND H. W HITE, Multilayer feedforward networks are uni-
versal approximators, Neural Networks, 2 (1989), pp. 359–366, https://doi.org/10.1016/
0893-6080(89)90020-8. (Cited on pp. 389, 397)
 436 Bibliography

[294] F. Q. H U , M. H USSAINI , AND P. R ASETARINERA, An analysis of the discontinuous Galerkin

Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

method for wave propagation problems, Journal of Computational Physics, 151 (1999), pp. 921–
946, https://doi.org/10.1006/jcph.1999.6227. (Cited on p. 234)

[295] Y. H UANG AND Y. C HEN, Survey of state-of-art autonomous driving technologies with deep learn-
ing, in 2020 IEEE 20th International Conference on Software Quality, Reliability and Security Com-
panion (QRS-C), 2020, pp. 221–228, https://doi.org/10.1109/QRS-C51114.2020.00045.
(Cited on p. 389)

[296] T. H UGHES , W. L IU , AND T. Z IMMERMANN, Lagrangian–Eulerian finite element formulation for

incompressible viscous flows, Computer Methods in Applied Mechanics and Engineering, 29 (1981),
pp. 329–349, https://doi.org/10.1016/0045-7825(81)90049-9. (Cited on p. 284)

[297] T. J. H UGHES , L. P. F RANCA , AND M. BALESTRA, A new finite element formulation for computa-
tional fluid dynamics: V. Circumventing the Babuška-Brezzi condition: A stable Petrov-Galerkin
formulation of the Stokes problem accommodating equal-order interpolations, Computer Meth-
ods in Applied Mechanics and Engineering, 59 (1986), pp. 85–99, https://doi.org/10.1016/
0045-7825(87)90093-4. (Cited on p. 128)

[298] T. J. H UGHES , L. P. F RANCA , AND G. M. H ULBERT, A new finite element formulation for
computational fluid dynamics: VIII. The Galerkin/least-squares method for advective-diffusive
equations, Computer Methods in Applied Mechanics and Engineering, 73 (1989), pp. 173–189,
https://doi.org/10.1016/0045-7825(89)90111-4. (Cited on pp. 130, 132)

[299] T. J. R. H UGHES , A. A. O BERAI , AND L. M AZZEI, Large eddy simulation of turbulent channel
flows by the variational multiscale method, Physics of Fluids, 13 (2001), pp. 1784–1799, https:
//doi.org/10.1063/1.1367868. (Cited on p. 132)

[300] J. H UNT, A. W RAY, AND P. M OIN, Eddies Stream and Convergence Zones in Turbulent Flows,
tech. Report CTR-S88, CTR report, 1988. (Cited on p. 170)

[301] D. H UYNH , G. ROZZA , S. S EN , AND A. PATERA, A successive constraint linear optimization

method for lower bounds of parametric coercivity and inf-sup stability constants, Comptes Rendus
Mathematique, 345 (2007), pp. 473–478, https://doi.org/10.1016/j.crma.2007.09.019.
(Cited on pp. 5, 48)

[302] L. I APICHINO , S. U LBRICH , AND S. VOLKWEIN, Multiobjective PDE-constrained optimization

using the reduced-basis method, Advances in Computational Mathematics, 43 (2017), pp. 945–972,
https://doi.org/10.1007/s10444-016-9512-x. (Cited on pp. 5, 83)

[303] A. I OLLO AND D. L OMBARDI, Advection modes by optimal mass transfer, Physical Review E, 89
(2014). (Cited on p. 298)

[304] R. I SSA, Solution of the implicitly discretised fluid flow equations by operator-splitting, Journal
of Computational Physics, 62 (1986), pp. 40–65, https://doi.org/10.1016/0021-9991(86)
90099-9. (Cited on p. 146)

[305] J. F. G ERBEAU AND M. V IDRASCU, A quasi-Newton algorithm based on a reduced model for
fluid–structure interaction problems in blood flows, ESAIM: Mathematical Modelling and Numerical
Analysis, 37 (2003), pp. 631–647, https://doi.org/10.1051/m2an:2003049. (Cited on p. 286)

[306] J. JANAI , F. G ÜNEY, A. B EHL , AND A. G EIGER, Computer vision for autonomous vehicles: Prob-
lems, datasets and state-of-the-art, Foundations and Trends in Computer Graphics and Vision, 12
(2020), pp. 1–308, https://doi.org/10.1561/0600000079. (Cited on p. 389)

[307] A. JANON , M. N ODET, AND C. P RIEUR, Certified reduced-basis solutions of viscous Burgers equa-
tion parametrized by initial and boundary values, ESAIM: Mathematical Modelling and Numerical
Analysis, 47 (2013), pp. 317–348, https://doi.org/10.1051/m2an/2012029. (Cited on p. 155)
 Bibliography 437

[308] C. JANYA -A NURAK, Framework for Analysis and Identification of Nonlinear Distributed Param-
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

eter Systems using Bayesian Uncertainty Quantification Based on Generalized Polynomial Chaos,
vol. 31, KIT Scientific Publishing, 2017, https://doi.org/10.5445/KSP/1000066940. (Cited
on p. 405)

[309] H. JASAK, Error Analysis and Estimation for the Finite Volume Method with Applications to Fluid
Flows, PhD thesis, Imperial College, University of London, 1996. (Cited on p. 146)

[310] W. J I , Z. R EN , Y. M ARZOUK , AND C. K. L AW, Quantifying kinetic uncertainty in turbulent com-

bustion simulations using active subspaces, Proceedings of the Combustion Institute, 37 (2019),
pp. 2175–2182, https://doi.org/10.1016/j.proci.2018.06.206. (Cited on p. 327)

[311] W. J I , J. WANG , O. Z AHM , Y. M. M ARZOUK , B. YANG , Z. R EN , AND C. K. L AW, Shared

low-dimensional subspaces for propagating kinetic uncertainty to multiple outputs, Combustion and
Flame, 190 (2018), pp. 146–157, https://doi.org/10.1016/j.combustflame.2017.11.021.
(Cited on p. 327)

[312] C. J OHNSON , U. N ÄVERT, AND J. P ITKÄRANTA, Finite element methods for linear hyperbolic
problems, Computer Methods in Applied Mechanics and Engineering, 45 (1984), pp. 285–312,
https://doi.org/10.1016/0045-7825(84)90158-0. (Cited on p. 127)

[313] H. J OHNSTON AND J.-G. L IU, Accurate, stable and efficient Navier–Stokes solvers based on explicit
treatment of the pressure term, Journal of Computational Physics, 199 (2004), pp. 221–259, https:
//doi.org/10.1016/j.jcp.2004.02.009. (Cited on pp. 152, 180)

[314] W. J ONES AND B. L AUNDER, The prediction of laminarization with a two-equation model of
turbulence, International Journal of Heat and Mass Transfer, 15 (1972), pp. 301–314, https:
//doi.org/10.1016/0017-9310(72)90076-2. (Cited on p. 168)

[315] M. R. J OVANOVI Ć , P. J. S CHMID , AND J. W. N ICHOLS, Sparsity-promoting dynamic mode de-

composition, Physics of Fluids, 26 (2014), p. 024103, https://doi.org/10.1063/1.4863670.
(Cited on p. 208)

[316] M. K AHLBACHER AND S. VOLKWEIN, Galerkin proper orthogonal decomposition methods for
parameter dependent elliptic systems, Discussiones Mathematicae, Differential Inclusions, Control
and Optimization, 27 (2007), pp. 95–117, http://eudml.org/doc/271156. (Cited on p. 107)

[317] I. K ALASHNIKOVA AND S. A RUNAJATESAN, A stable Galerkin reduced order model (ROM) for
compressible flow, in 10th World Congress for Computational Mechanics (WCCM), 2012, pp. 2012–
18407, https://doi.org/10.5151/meceng-wccm2012-18407. (Cited on pp. 242, 243)

[318] I. K ALASHNIKOVA AND M. BARONE, On the stability and convergence of a Galerkin reduced order
model (ROM) of compressible flow with solid wall and far-field boundary treatment, International
Journal for Numerical Methods in Engineering, 83 (2010), pp. 1345–1375, https://doi.org/10.
1002/nme.2867. (Cited on pp. 155, 243)

[319] I. K ALASHNIKOVA AND M. F. BARONE, Efficient non-linear proper orthogonal decomposition/

Galerkin reduced order models with stable penalty enforcement of boundary conditions, Interna-
tional Journal for Numerical Methods in Engineering, 90 (2012), pp. 1337–1362, https://doi.
org/10.1002/nme.3366. (Cited on pp. 154, 155)

[320] M. G. K APTEYN , D. J. K NEZEVIC , D. H UYNH , M. T RAN , AND K. E. W ILLCOX, Data-driven

physics-based digital twins via a library of component-based reduced-order models, International
Journal for Numerical Methods in Engineering, 123 (2022), pp. 2986–3003, https://doi.org/
10.1002/nme.6423. (Cited on p. 413)

[321] M. G. K APTEYN , J. V. P RETORIUS , AND K. E. W ILLCOX, A probabilistic graphical model founda-

tion for enabling predictive digital twins at scale, Nature Computational Science, 1 (2021), pp. 337–
347, https://doi.org/10.1038/s43588-021-00069-0. (Cited on pp. 12, 411)
 438 Bibliography

[322] E. N. K ARATZAS , F. BALLARIN , AND G. ROZZA, Projection-based reduced order models for a
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

cut finite element method in parametrized domains, Computers & Mathematics with Applications,
79 (2020), pp. 833–851, https://doi.org/10.1016/j.camwa.2019.08.003. (Cited on pp. 265,
280, 298)

[323] E. N. K ARATZAS , M. N ONINO , F. BALLARIN , AND G. ROZZA, A reduced order cut finite element
method for geometrically parameterized steady and unsteady Navier-Stokes problems, Computers &
Mathematics with Applications, 116 (2022), pp. 140–160, https://doi.org/10.1016/j.camwa.
2021.07.016. (Cited on p. 265)

[324] E. N. K ARATZAS AND G. ROZZA, A reduced order model for a stable embedded boundary
parametrized Cahn–Hilliard phase-field system based on cut finite elements, Journal of Scientific
Computing, 89 (2021), https://doi.org/10.1007/s10915-021-01623-8. (Cited on pp. 37,
265, 281)

[325] E. N. K ARATZAS , G. S TABILE , N. ATALLAH , G. S COVAZZI , AND G. ROZZA, A reduced order

approach for the embedded shifted boundary FEM and a heat exchange system on parametrized
geometries, in IUTAM Symposium on Model Order Reduction of Coupled Systems, Stuttgart, Ger-
many, J. Fehr and B. Haasdonk, eds., vol. 36 of IUTAM Bookseries, Springer, 2018, pp. 111–125,
https://doi.org/10.1007/978-3-030-21013-7_8. (Cited on pp. 37, 265, 266)

[326] E. N. K ARATZAS , G. S TABILE , L. N OUVEAU , G. S COVAZZI , AND G. ROZZA, A reduced basis

approach for PDEs on parametrized geometries based on the shifted boundary finite element method
and application to a Stokes flow, Computer Methods in Applied Mechanics and Engineering, 347
(2019), pp. 568–587, https://doi.org/10.1016/j.cma.2018.12.040. (Cited on pp. 265, 271)

[327] E. N. K ARATZAS , G. S TABILE , L. N OUVEAU , G. S COVAZZI , AND G. ROZZA, A reduced-

order shifted boundary method for parametrized incompressible Navier–Stokes equations, Computer
Methods in Applied Mechanics and Engineering, 370 (2020), pp. 113–273, https://doi.org/10.
1016/j.cma.2020.113273. (Cited on p. 265)

[328] M. K ÄRCHER AND M. A. G REPL, A certified reduced basis method for parametrized elliptic
optimal control problems, ESAIM: Control, Optimisation and Calculus of Variations, 20 (2014),
pp. 416–441, https://doi.org/10.1051/cocv/2013069. (Cited on pp. 83, 87, 262)

[329] M. K ÄRCHER , Z. T OKOUTSI , M. A. G REPL , AND K. V EROY, Certified reduced basis methods
for parametrized elliptic optimal control problems with distributed controls, Journal of Scientific
Computing, 75 (2018), pp. 276–307, https://doi.org/10.1007/s10915-017-0539-z. (Cited
on pp. 83, 87, 262)

[330] G. K ARNIADAKIS AND S. S HERWIN, Spectral/hp element methods for CFD, Numerical Mathe-
matics and Scientific Computation, Oxford University Press, 2005, https://doi.org/10.1093/
acprof:oso/9780198528692.001.0001. (Cited on pp. 223, 227, 373)

[331] G. K. K ARPOUZAS , E. M. PAPOUTSIS –K IACHAGIAS , T. S CHUMACHER , E. DE V ILLIERS , K. C.

G IANNAKOGLOU , AND C. OTHMER, Adjoint optimization for vehicle external aerodynamics, In-
ternational Journal of Automotive Engineering, 7 (2016), pp. 1–7, https://doi.org/10.20485/
jsaeijae.7.1_1. (Cited on p. 168)

[332] G. K ATSOULEAS , E. N. K ARATZAS , AND F. T RAVLOPANOS, Discrete empirical interpolation and

unfitted mesh FEMs: Application in PDE-constrained optimization, Optimization (2022), pp. 1–34,
https://doi.org/10.1080/02331934.2022.2032697. (Cited on p. 265)

[333] K. K EAN AND M. S CHNEIER, Error analysis of supremizer pressure recovery for POD based
reduced-order models of the time-dependent Navier–Stokes equations, SIAM Journal on Numeri-
cal Analysis, 58 (2020), pp. 2235–2264, https://doi.org/10.1137/19m128702x. (Cited on
p. 144)
 Bibliography 439

[334] H. B. K ELLER, Lectures on numerical methods in bifurcation problems, Applied Mathematics, 217
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

(1987). (Cited on p. 104)

[335] M. C. K ENNEDY AND A. O’H AGAN, Predicting the output from a complex computer code when
fast approximations are available, Biometrika, 87 (2000), pp. 1–13, https://doi.org/10.1093/
biomet/87.1.1. (Cited on p. 334)

[336] P. K ENNICOTT, IGES/PDES O RGANIZATION , AND U.S. P RODUCT DATA A SSOCIATION, Initial
Graphics Exchange Specification, IGES 5.3, U.S. Product Data Association, 1996. (Cited on p. 357)

[337] M. K HAMLICH , F. P ICHI , AND G. ROZZA, Model Order Reduction for Bifurcating Phenomena in
Fluid-Structure Interaction Problems, 2021, https://arxiv.org/abs/2110.06297. (Cited on
pp. 98, 312, 317)

[338] A. K HAN , A. S OHAIL , U. Z AHOORA , AND A. S. Q URESHI, A survey of the recent architectures
of deep convolutional neural networks, Artificial Intelligence Review, 53 (2020), pp. 5455–5516,
https://doi.org/10.1007/s10462-020-09825-6. (Cited on p. 402)

[339] S. K HAN AND E. G UNPINAR, Sampling CAD models via an extended teaching–learning-based
optimization technique, Computer-Aided Design, 100 (2018), pp. 52–67, https://doi.org/10.
1016/j.cad.2018.03.003. (Cited on p. 204)

[340] D. K HATAMSAZ AND D. L. A LLAIRE, Materials design using an active subspace batch Bayesian
optimization approach, in AIAA SCITECH 2022 Forum, 2022, p. 0075, https://doi.org/10.
2514/6.2022-0075. (Cited on p. 340)

[341] P. K HODABAKHSHI AND K. E. W ILLCOX, Non-intrusive data-driven model reduction for

differential–algebraic equations derived from lifting transformations, Computer Methods in Applied
Mechanics and Engineering, 389 (2021), p. 114296, https://doi.org/10.1016/j.cma.2021.
114296. (Cited on p. 222)

[342] H. K IELHÖFER, Bifurcation Theory: An Introduction with Applications to PDEs, Applied Mathe-
matical Sciences, Springer New York, 2006. (Cited on p. 101)

[343] T. K IM , J. O H , N. Y. K IM , S. C HO , AND S.-Y. Y UN, Comparing Kullback-Leibler divergence and

mean squared error loss in knowledge distillation, in Proceedings of the Thirtieth International Joint
Conference on Artificial Intelligence, IJCAI-21, Z.-H. Zhou, ed., International Joint Conferences on
Artificial Intelligence Organization, 2021, pp. 2628–2635, https://doi.org/10.24963/ijcai.
2021/362. (Cited on p. 406)

[344] D. P. K INGMA AND J. BA, Adam: A Method for Stochastic Optimization, arXiv preprint
arXiv:1412.6980 (2014). (Cited on p. 339)

[345] J. K IRKLIN , F. PAGANI , R. KORMOS , L. S TEVENSON , E. B LUME , S. M YERS , M. M ILLER ,

J. BALDWIN , J. YOUNG , AND D. NAFTEL, Eighth annual INTERMACS report: Special focus on
framing the impact of adverse events, The Journal of Heart and Lung Transplantation, 36 (2017),
pp. 1080–1086, https://doi.org/10.1016/j.healun.2017.07.005. (Cited on p. 374)

[346] D. J. K NEZEVIC AND A. T. PATERA, A certified reduced basis method for the Fokker–Planck equa-
tion of dilute polymeric fluids: FENE dumbbells in extensional flow, SIAM Journal on Scientific
Computing, 32 (2010), pp. 793–817, https://doi.org/10.1137/090759239. (Cited on p. 5)

[347] A. N. KOLMOGOROV, Equations of turbulent motion of an incompressible fluid, Izvestia Academy

of Sciences, USSR, 6 (1942), pp. 55–58. (Cited on p. 168)

[348] D. K RIESEL, A Brief Introduction to Neural Networks, 2007, http://www.dkriesel.com. (Cited

on pp. 11, 390, 393)
 440 Bibliography

[349] A. K RIZHEVSKY AND G. H INTON, Learning Multiple Layers of Features from Tiny Images, Mas-
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

ter’s thesis, Department of Computer Science, University of Toronto (2009). (Cited on p. 407)

[350] A. K RIZHEVSKY, I. S UTSKEVER , AND G. E. H INTON, Imagenet classification with deep convolu-
tional neural networks, Advances in Neural Information Processing Systems, 25 (2012), pp. 1097–
1105. (Cited on pp. 389, 402)

[351] K. K UNISCH AND S. VOLKWEIN, Galerkin proper orthogonal decomposition methods for a general
equation in fluid dynamics, SIAM Journal on Numerical Analysis, 40 (2002), pp. 492–515, https:
//doi.org/10.1137/s0036142900382612. (Cited on p. 284)

[352] K. K UNISCH AND S. VOLKWEIN, Proper orthogonal decomposition for optimality systems,
ESAIM: Mathematical Modelling and Numerical Analysis, 42 (2008), pp. 1–23, https://doi.
org/10.1051/m2an:2007054. (Cited on pp. 83, 367)

[353] J. N. K UTZ , S. L. B RUNTON , B. W. B RUNTON , AND J. L. P ROCTOR, Dynamic Mode Decompo-

sition: Data-Driven Modeling of Complex Systems, SIAM, 2016, https://doi.org/10.1137/1.
9781611974508. (Cited on pp. 207, 383)

[354] J. N. K UTZ , X. F U , AND S. L. B RUNTON, Multiresolution dynamic mode decomposition, SIAM

Journal on Applied Dynamical Systems, 15 (2016), pp. 713–735, https://doi.org/10.1137/
15M1023543. (Cited on pp. 209, 210)

[355] Y. K UZNETSOV, Elements of Applied Bifurcation Theory, Applied Mathematical Sciences, Springer
New York, 2004. (Cited on p. 101)

[356] R. R. L AM , O. Z AHM , Y. M. M ARZOUK , AND K. E. W ILLCOX, Multifidelity dimension reduction

via active subspaces, SIAM Journal on Scientific Computing, 42 (2020), pp. A929–A956, https:
//doi.org/10.1137/18M1214123. (Cited on p. 327)

[357] P. L ASAINT AND P. R AVIART, On a finite element method for solving the neutron transport equation,
Mathematical aspects of finite elements in partial differential equations, in C. de Boor, ed., New York,
1974, pp. 89–123. (Cited on p. 234)

[358] W. L ASKEY, H. PARKER , V. F ERRARI , W. K USSMAUL , AND A. N OORDERGRAAF, Estimation

of total systemic arterial compliance in humans, Journal of Applied Physiology (1985), 69 (1990),
pp. 112–119, https://doi.org/10.1152/jappl.1990.69.1.112. (Cited on p. 375)

[359] T. L ASSILA , A. M ANZONI , A. Q UARTERONI , AND G. ROZZA, A reduced computational and

geometrical framework for inverse problems in hemodynamics, International Journal for Numerical
Methods in Biomedical Engineering, 29 (2013), pp. 741–776, https://doi.org/10.1002/cnm.
2559. (Cited on p. 83)

[360] T. L ASSILA , A. M ANZONI , A. Q UARTERONI , AND G. ROZZA, Model order reduction in fluid dy-
namics: Challenges and perspectives, in Reduced Order Methods for Modeling and Computational
Reduction, A. Quarteroni and G. Rozza, eds., vol. 9 of MS&A series, Springer, 2014, pp. 235–273,
https://doi.org/10.1007/978-3-319-02090-7_9. (Cited on p. 7)

[361] B. L AUNDER AND D. S PALDING, The numerical computation of turbulent flows, Computer Meth-
ods in Applied Mechanics and Engineering, 3 (1974), pp. 269–289, https://doi.org/10.1016/
0045-7825(74)90029-2. (Cited on p. 168)

[362] W. L AYTON , L. R EBHOLZ , AND C. T RENCHEA, Modular nonlinear filter stabilization of methods
for higher Reynolds numbers flow, Journal of Mathematical Fluid Mechanics, 14 (2012), pp. 325–
354, https://doi.org/10.1007/s00021-011-0072-z. (Cited on pp. 170, 171)

[363] D. L AZZARO AND L. B. M ONTEFUSCO, Radial basis functions for the multivariate interpolation of
large scattered data sets, Journal of Computational and Applied Mathematics, 140 (2002), pp. 521–
536, https://doi.org/10.1016/s0377-0427(01)00485-x. (Cited on p. 172)
 Bibliography 441

[364] S. L E C LAINCHE AND J. M. V EGA, Higher order dynamic mode decomposition, SIAM Journal on
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Applied Dynamical Systems, 16 (2017), pp. 882–925, https://doi.org/10.1137/15M1054924.

(Cited on p. 209)

[365] P. L E TALLEC AND J. M OURO, Fluid structure interaction with large structural displacements,
Computer Methods in Applied Mechanics and Engineering, 190 (2001), pp. 3039–3067, https:
//doi.org/10.1016/S0045-7825(00)00381-9. (Cited on p. 286)

[366] Y. L E C UN AND Y. B ENGIO, Convolutional Networks for Images, Speech, and Time Series,
vol. 3361, MIT Press, 1995, p. 1995. (Cited on p. 394)

[367] K. L EE AND K. T. C ARLBERG, Model reduction of dynamical systems on nonlinear manifolds

using deep convolutional autoencoders, Journal of Computational Physics, 404 (2020), p. 108973,
https://doi.org/10.1016/j.jcp.2019.108973. (Cited on pp. 205, 298, 396, 399)

[368] M.-Y. L. L EE, Estimation of the error in the reduced basis method solution of differential algebraic
equation systems, SIAM Journal on Numerical Analysis, 28 (1991), pp. 512–528, https://doi.
org/10.1137/0728028. (Cited on p. 3)

[369] P. L E G RESLEY AND J. A LONSO, Investigation of Non-linear Projection for POD Based Reduced
Order Models for Aerodynamics, in 39th Aerospace Sciences Meeting and Exhibit, American Insti-
tute of Aeronautics and Astronautics, Reno, NV, U.S.A., Jan. 2001, https://doi.org/10.2514/
6.2001-926. (Cited on p. 144)

[370] S. L EIJNEN AND F. V. V EEN, The neural network zoo, in Multidisciplinary Digital Publishing In-
stitute Proceedings, vol. 47, 2020, p. 9, https://doi.org/10.3390/proceedings2020047009.
(Cited on p. 393)

[371] J. L ERAY, Sur le mouvement d’un fluide visqueux emplissant l’espace, Acta Mathematica, 63 (1934),
pp. 193–248, https://doi.org/10.1007/BF02547354. (Cited on p. 170)

[372] A. L EWIS , R. S MITH , AND B. W ILLIAMS, Gradient free active subspace construction using Morris
screening elementary effects, Computers & Mathematics with Applications, 72 (2016), pp. 1603–
1615, https://doi.org/10.1016/j.camwa.2016.07.022. (Cited on p. 337)

[373] T. L IEU , C. FARHAT, AND M. L ESOINNE, Reduced-order fluid/structure modeling of a complete

aircraft configuration, Computer Methods in Applied Mechanics and Engineering, 195 (2006),
pp. 5730–5742, https://doi.org/10.1016/j.cma.2005.08.026. (Cited on p. 283)

[374] J. L. L IONS, Optimal Control of System Governed by Partial Differential Equations, vol. 170 of
Grundlehren der mathematischen Wissenschaften, Springer-Verlag Berlin Heidelberg, 1971. (Cited
on p. 84)

[375] H. L IU, Optimal error estimates of the direct discontinuous Galerkin method for convection-diffusion
equations, Mathematics of Computation, 84 (2015), pp. 2263–2295, https://doi.org/10.1090/
S0025-5718-2015-02923-8. (Cited on p. 239)

[376] J.-G. L IU , J. L IU , AND R. L. P EGO, Stable and accurate pressure approximation for unsteady
incompressible viscous flow, Journal of Computational Physics, 229 (2010), pp. 3428–3453, https:
//doi.org/10.1016/j.jcp.2010.01.010. (Cited on pp. 152, 180)

[377] Z. L IU , M. S UN , T. Z HOU , G. H UANG , AND T. DARRELL, Rethinking the Value of Network Prun-
ing, in International Conference on Learning Representations, 2019. (Cited on p. 403)

[378] A. L OGG , K. M ARDAL , AND G. W ELLS, Automated Solution of Differential Equations by the Finite
Element Method, Springer-Verlag, Berlin, 2012. (Cited on p. 15)
 442 Bibliography

[379] M. L OMBARDI , N. PAROLINI , A. Q UARTERONI , AND G. ROZZA, Numerical simulation of sailing

Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

boats: Dynamics, FSI, and shape optimization, in Variational Analysis and Aerospace Engineering:
Mathematical Challenges for Aerospace Design, Springer, 2012, pp. 339–377, https://doi.org/
10.1007/978-1-4614-2435-2_15. (Cited on pp. 283, 347)

[380] S. L ORENZI , A. C AMMI , L. L UZZI , AND G. ROZZA, POD-Galerkin method for finite volume
approximation of Navier–Stokes and RANS equations, Computer Methods in Applied Mechanics
and Engineering, 311 (2016), pp. 151–179, https://doi.org/10.1016/j.cma.2016.08.006.
(Cited on pp. 143, 144, 150, 155, 181)

[381] L. L OVÁSZ AND S. V EMPALA, Hit-and-run from a corner, SIAM Journal on Computing, 35 (2006),
pp. 985–1005, https://doi.org/10.1137/S009753970544727X. (Cited on p. 329)

[382] L. L U , P. J IN , G. PANG , Z. Z HANG , AND G. E. K ARNIADAKIS, Learning nonlinear operators via

DeepONet based on the universal approximation theorem of operators, Nature Machine Intelligence,
3 (2021), pp. 218–229, https://doi.org/10.1038/s42256-021-00302-5. (Cited on p. 401)

[383] T. W. L UKACZYK , P. C ONSTANTINE , F. PALACIOS , AND J. J. A LONSO, Active subspaces for

shape optimization, in 10th AIAA multidisciplinary design optimization conference, 2014, p. 1171,
https://doi.org/10.2514/6.2014-1171. (Cited on p. 327)

[384] Y. M ADAY, Analysis of coupled models for fluid-structure interaction of internal flows, in Cardio-
vascular Mathematics: Modeling and simulation of the circulatory system, L. Formaggia, A. Quar-
teroni, and A. Veneziani, eds., Springer Milan, 2009, pp. 279–306, https://doi.org/10.1007/
978-88-470-1152-6_8. (Cited on p. 283)

[385] Y. M ADAY, N. C. N GUYEN , A. T. PATERA , AND G. S. H. PAU, A general multipurpose interpola-

tion procedure: The magic points, Communications on Pure & Applied Analysis, 8 (2009), p. 383,
https://doi.org/10.3934/cpaa.2009.8.383. (Cited on p. 108)

[386] Y. M ADAY, A. T. PATERA , AND G. T URINICI, Global a priori convergence theory for reduced-
basis approximations of single-parameter symmetric coercive elliptic partial differential equa-
tions, Comptes Rendus Mathematique, 335 (2002), pp. 289–294, https://doi.org/10.1016/
S1631-073X(02)02466-4. (Cited on p. 32)

[387] A. M ANZONI, An efficient computational framework for reduced basis approximation and a posteri-
ori error estimation of parametrized Navier-Stokes flows, ESAIM: Mathematical Modelling and Nu-
merical Analysis, 48 (2014), pp. 1199–1226, https://doi.org/10.1051/m2an/2014013. (Cited
on pp. 5, 64, 131)

[388] A. M ANZONI AND F. N EGRI, Heuristic strategies for the approximation of stability factors in
quadratically nonlinear parametrized PDEs, Advances in Computational Mathematics, 41 (2015),
pp. 1255–1288, https://doi.org/10.1007/s10444-015-9413-4. (Cited on p. 5)

[389] A. M ANZONI , A. Q UARTERONI , AND G. ROZZA, Computational reduction for parametrized

PDEs: Strategies and applications, Milan Journal of Mathematics, 80 (2012), pp. 283–309, https:
//doi.org/10.1007/s00032-012-0182-y. (Cited on p. 316)

[390] A. M ANZONI , A. Q UARTERONI , AND G. ROZZA, Model reduction techniques for fast blood flow
simulation in parametrized geometries, International Journal for Numerical Methods in Biomedical
Engineering, 28 (2012), pp. 604–625, https://doi.org/10.1002/cnm.1465. (Cited on pp. 11,
349)

[391] J. E. M ATSSON, An Introduction to ANSYS Fluent 2021, SDC Publications, 2021. (Cited on p. 143)

[392] L. R. M EDSKER AND L. JAIN, Recurrent neural networks, Design and Applications, 5 (2001),
pp. 64–67, https://doi.org/10.1201/9781420049176. (Cited on p. 393)
 Bibliography 443

[393] J. M. M ELENK, On n-widths for elliptic problems, Journal of Mathematical Analysis and Applica-
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

tions, 247 (2000), pp. 272–289, https://doi.org/10.1006/jmaa.2000.6862. (Cited on pp. 32,

106, 297)

[394] L. M ENEGHETTI , N. D EMO , AND G. ROZZA, A Dimensionality Reduction Approach for Convo-
lutional Neural Networks, 2021, https://arxiv.org/abs/2110.09163. (Cited on pp. 12, 343,
384, 389, 403, 407)

[395] F. R. M ENTER, Two-equation eddy-viscosity turbulence models for engineering applications, AIAA
Journal, 32 (1994), pp. 1598–1605, https://doi.org/10.2514/3.12149. (Cited on p. 168)

[396] C. A. M ICCHELLI, Interpolation of scattered data: Distance matrices and conditionally positive
definite functions, Constructive Approximation, 2 (1986), pp. 11–22, https://doi.org/10.1007/
bf01893414. (Cited on p. 172)

[397] S. M IDDELKAMP, P. K EVREKIDIS , D. F RANTZESKAKIS , R. C ARRETERO -G ONZÁLEZ , AND

P. S CHMELCHER, Bifurcations, stability, and dynamics of multiple matter-wave vortex states, Physi-
cal Review A, 82 (2010), p. 013646, https://doi.org/10.1103/PhysRevA.82.013646. (Cited
on p. 112)

[398] S. M IDDELKAMP, P. K EVREKIDIS , D. F RANTZESKAKIS , R. C ARRETERO -G ONZÁLEZ , AND

P. S CHMELCHER, Emergence and stability of vortex clusters in Bose-Einstein condensates: A bifur-
cation approach near the linear limit, Physica D: Nonlinear Phenomena, 240 (2011), pp. 1449–1459,
https://doi.org/10.1016/j.physd.2011.06.002. (Cited on p. 112)

[399] M. M INSKY AND S. A. PAPERT, Perceptrons: An Introduction to Computational Geometry, MIT

Press, 2017, https://doi.org/10.7551/mitpress/11301.001.0001. (Cited on p. 393)

[400] L. M IRSKY, Symmetric gauge functions and unitarily invariant norms, Quarterly Journal of Mathe-
matics, 11 (1960), pp. 50–59, https://doi.org/10.1093/qmath/11.1.50. (Cited on p. 4)

[401] A. M OELGAARD, PMM-Tests with a Model of a Frigate Class DDG-51, tech. Report 01, University
of Zurich, Department of Informatics, 2000. (Cited on p. 358)

[402] M. M OHEBUJJAMAN , L. G. R EBHOLZ , X. X IE , AND T. I LIESCU, Energy balance and mass con-
servation in reduced order models of fluid flows, Journal of Computational Physics, 346 (2017),
pp. 262–277, https://doi.org/10.1016/j.jcp.2017.06.019. (Cited on p. 187)

[403] A. M OLA , G. B ORDONARO , AND M. R. H AJJ, Low-frequency variations of force coefficients on

square cylinders with sharp and rounded corners, Journal of Structural Engineering, 135 (2009),
pp. 828–835, https://doi.org/10.1061/(ASCE)ST.1943-541X.0000034. (Cited on p. 187)

[404] A. M OLA , M. T EZZELE , M. G ADALLA , F. VALDENAZZI , D. G RASSI , R. PADOVAN , AND

G. ROZZA, Efficient reduction in shape parameter space dimension for ship propeller blade de-
sign, in Proceedings of MARINE 2019: VIII International Conference on Computational Methods
in Marine Engineering, R. Bensow and J. Ringsberg, eds., 2019, pp. 201–212, arXiv:1905.09815.
(Cited on pp. 327, 355, 356, 360, 411)

[405] H. M ONTAZERI AND B. B LOCKEN, CFD simulation of wind-induced pressure coefficients on build-
ings with and without balconies: Validation and sensitivity analysis, Building and Environment, 60
(2013), pp. 137–149, https://doi.org/10.1016/j.buildenv.2012.11.012. (Cited on p. 168)

[406] U. E. M ORELLI , P. BARRAL , P. Q UINTELA , G. ROZZA , AND G. S TABILE, A numerical approach

for heat flux estimation in thin slabs continuous casting molds using data assimilation, International
Journal for Numerical Methods in Engineering, 122 (2021), pp. 4541–4574, https://doi.org/
10.1002/nme.6713. (Cited on p. 251)
 444 Bibliography

[407] A. M ORRIS , C. A LLEN , AND T. R ENDALL, CFD-based optimization of aerofoils using radial ba-
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

sis functions for domain element parameterization and mesh deformation, International Journal for
Numerical Methods in Fluids, 58 (2008), pp. 827–860, https://doi.org/10.1002/fld.1769.
(Cited on p. 349)

[408] F. M OUKALLED , L. M ANGANI , AND M. DARWISH, The Finite Volume Method in Computational
Fluid Dynamics: An Advanced Introduction with OpenFOAM and Matlab, 1st ed., Springer, 2015,
https://doi.org/10.1007/978-3-319-16874-6. (Cited on pp. 143, 144, 146)

[409] M. M ROSEK , C. OTHMER , AND R. R ADESPIEL, Reduced-order modeling of vehicle aerodynamics

via proper orthogonal decomposition, SAE International Journal of Passenger Cars–Mechanical Sys-
tems, 12 (2019), pp. 225–236, https://doi.org/10.4271/06-12-03-0016. (Cited on p. 205)

[410] C. M UREA AND S. S Y, A fast method for solving fluid–structure interaction problems numerically,
International Journal for Numerical Methods in Fluids, 60 (2009), pp. 1149–1172, https://doi.
org/10.1002/fld.1931. (Cited on p. 303)

[411] N. J. NAIR AND M. BALAJEWICZ, Transported snapshot model order reduction approach for para-
metric, steady-state fluid flows containing parameter-dependent shocks, International Journal for Nu-
merical Methods in Engineering, 117 (2019), pp. 1234–1262, https://doi.org/10.1002/nme.
5998. (Cited on p. 298)

[412] J. N E ČAS, Les méthodes directes en théorie des équations elliptiques, Masson, 1967. (Cited on
p. 52)

[413] F. N EGRI , A. M ANZONI , AND G. ROZZA, Reduced basis approximation of parametrized optimal
flow control problems for the Stokes equations, Computers & Mathematics with Applications, 69
(2015), pp. 319–336, https://doi.org/10.1016/j.camwa.2014.12.010. (Cited on pp. 83, 87,
367)

[414] F. N EGRI , G. ROZZA , A. M ANZONI , AND A. Q UARTERONI, Reduced basis method for
parametrized elliptic optimal control problems, SIAM Journal on Scientific Computing, 35 (2013),
pp. A2316–A2340, https://doi.org/10.1137/120894737. (Cited on pp. 83, 87, 262, 367)

[415] N. C. N GUYEN , G. ROZZA , AND A. T. PATERA, Reduced basis approximation and a posteriori
error estimation for the time-dependent viscous Burgers’ equation, Calcolo, 46 (2009), pp. 157–185,
https://doi.org/10.1007/s10092-009-0005-x. (Cited on p. 292)

[416] S. A. N IEDERER , M. S. S ACKS , M. G IROLAMI , AND K. W ILLCOX, Scaling digital twins from the
artisanal to the industrial, Nature Computational Science, 1 (2021), pp. 313–320, https://doi.
org/10.1038/s43588-021-00072-5. (Cited on pp. 12, 413)

[417] K. N ODA , H. A RIE , Y. S UGA , AND T. O GATA, Multimodal integration learning of robot behavior
using deep neural networks, Robotics and Autonomous Systems, 62 (2014), pp. 721–736, https:
//doi.org/10.1016/j.robot.2014.03.003. (Cited on p. 402)

[418] M. N ONINO, On the Application of the Reduced Basis Method to Fluid-Structure Interaction Prob-
lems, PhD thesis, Scuola Internazionale Superiore di Studi Avanzati, Trieste, 2020. (Cited on p. 311)

[419] M. N ONINO , F. BALLARIN , G. ROZZA , AND Y. M ADAY, Overcoming Slowly Decaying Kol-
mogorov n-Width by Transport Maps: Application to Model Order Reduction of Fluid Dynamics
and Fluid–Structure Interaction Problems. arXiv:1911.06598, 2019. (Cited on p. 399)

[420] M. N ONINO , F. BALLARIN , G. ROZZA , AND Y. M ADAY, Projection Based Semi–Implicit Parti-
tioned Reduced Basis Method for Non Parametrized and Parametrized Fluid–Structure Interaction
Problems. arXiv:2201.03236, 2022. (Cited on p. 312)

[421] A. K. N OOR, Reduced basis technique for nonlinear analysis of structures, AIAA Journal, 18
(1980), pp. 455–462, https://doi.org/10.2514/3.50778. (Cited on p. 3)
 Bibliography 445

[422] A. K. N OOR, On making large nonlinear problems small, Computer Methods in Applied Me-
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

chanics and Engineering, 34 (1982), pp. 955–985, https://doi.org/10.1016/0045-7825(82)

90096-2. (Cited on p. 3)
[423] A. K. N OOR AND J. M. P ETERS, Recent advances in reduction methods for instability analy-
sis of structures, Computers & Structures, 16 (1983), pp. 67–80, https://doi.org/10.1016/
0045-7949(83)90148-7. (Cited on p. 3)
[424] A. K. N OOR , J. M. P ETERS , AND C. M. A NDERSEN, Reduced basis technique for collapse analysis
of shells, AIAA Journal, 19 (1981), pp. 393–397, https://doi.org/10.2514/3.7776. (Cited on
p. 3)
[425] N ORSOK, Norsok Standard: Design of Steel Structures, October 2004, https://www.standard.
no/en/sectors/energi-og-klima/petroleum/norsok-standard-categories/n-structu
ral/n-0041/. (Cited on pp. 115, 118)
[426] A. N OVIKOV, D. P ODOPRIKHIN , A. O SOKIN , AND D. V ETROV, Tensorizing neural networks,
in Proceedings of the 28th International Conference on Neural Information Processing Systems—
Volume 1, NIPS’15, MIT Press, 2015, pp. 442–450. (Cited on p. 403)
[427] M. O HLBERGER AND S. R AVE, Nonlinear reduced basis approximation of parameterized evolution
equations via the method of freezing, Comptes Rendus Mathematique, 351 (2013), pp. 901–906,
https://doi.org/10.1016/j.crma.2013.10.028. (Cited on p. 298)
[428] M. O HLBERGER AND S. R AVE, Reduced basis methods: Success, limitations and future challenges,
Proceedings of the Conference Algoritmy (2016), pp. 1–12. (Cited on p. 297)
[429] M. O LSHANSKII , G. L UBE , T. H EISTER , AND J. L ÖWE, Grad–div stabilization and subgrid pres-
sure models for the incompressible Navier–Stokes equations, Computer Methods in Applied Me-
chanics and Engineering, 198 (2009), pp. 3975–3988, https://doi.org/10.1016/j.cma.2009.
09.005. (Cited on p. 130)
[430] M. O LSHANSKII AND X. X IONG, A connection between filter stabilization and eddy viscosity
models, Numerical Methods for Partial Differential Equations, 29 (2013), pp. 2061–2080, https:
//doi.org/10.1002/num.21791. (Cited on p. 171)
[431] G. O RTALI , N. D EMO , AND G. ROZZA, A Gaussian process approach within a data-driven POD
framework for fluid dynamics engineering problems, Mathematics in Engineering, 4 (2022), pp. 1–
16, https://doi.org/10.3934/mine.2022021. (Cited on p. 206)
[432] C. OTHMER , T. W. L UKACZYK , P. C ONSTANTINE , AND J. J. A LONSO, On active subspaces in
car aerodynamics, in 17th AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference,
2016, p. 4294. (Cited on p. 327)
[433] P. PACCIARINI AND G. ROZZA, Stabilized reduced basis method for parametrized advec-
tion–diffusion PDEs, Computer Methods in Applied Mechanics and Engineering, 274 (2014), pp. 1–
18, https://doi.org/10.1016/j.cma.2014.02.005. (Cited on pp. 125, 127)
[434] M. PALACI -O LGUN, Gaussian Process Modeling and Supervised Dimensionality Reduction Algo-
rithms via Stiefel Manifold Learning, Master’s thesis, University of Toronto Institute of Aerospace
Studies (2018). (Cited on p. 329)
[435] D. PAPAPICCO, A Neural Network Framework for Reduced Order Modelling of Non-Linear Hyper-
bolic Equations in Computational Fluid Dynamics, Master’s thesis, Politecnico di Torino (2021),
https://webthesis.biblio.polito.it/17349/. (Cited on p. 399)
[436] D. PAPAPICCO , N. D EMO , M. G IRFOGLIO , G. S TABILE , AND G. ROZZA, The neural network
shifted-proper orthogonal decomposition: A machine learning approach for non-linear reduction
of hyperbolic equations, Computer Methods in Applied Mechanics and Engineering, 392 (2022),
p. 114687, https://doi.org/10.1016/j.cma.2022.114687. (Cited on pp. 11, 399)
 446 Bibliography

[437] A. PASZKE , S. G ROSS , F. M ASSA , A. L ERER , J. B RADBURY, G. C HANAN , T. K ILLEEN , Z. L IN ,

Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

N. G IMELSHEIN , L. A NTIGA , A. D ESMAISON , A. KOPF, E. YANG , Z. D E V ITO , M. R AISON ,

A. T EJANI , S. C HILAMKURTHY, B. S TEINER , L. FANG , J. BAI , AND S. C HINTALA, PyTorch:
An Imperative Style, High-Performance Deep Learning Library, in Advances in Neural Information
Processing Systems 32, H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and
R. Garnett, eds., Curran Associates, Inc., 2019, pp. 8024–8035. (Cited on p. 407)

[438] S. PATANKAR AND D. S PALDING, A calculation procedure for heat, mass and momentum transfer
in three-dimensional parabolic flows, International Journal of Heat and Mass Transfer, 15 (1972),
pp. 1787–1806, https://doi.org/10.1016/0017-9310(72)90054-3. (Cited on pp. 144, 146)

[439] A. PATERA, A spectral element method for fluid dynamics: Laminar flow in a channel expan-
sion, Journal of Computational Physics, 54 (1984), pp. 468–488, https://doi.org/10.1016/
0021-9991(84)90128-1. (Cited on p. 223)

[440] A. PATERA AND G. ROZZA, Reduced Basis Approximation and a Posteriori Error Estimation for
Parametrized Partial Differential Equation, MIT Pappalardo Monographs in Mechanical Engineer-
ing, Copyright MIT (2007–2010). (Cited on p. 106)

[441] W. PAZNER AND P.-O. P ERSSON, Stage-parallel fully implicit Runge–Kutta solvers for discon-
tinuous Galerkin fluid simulations, Journal of Computational Physics, 335 (2017), pp. 700–717,
https://doi.org/10.1016/j.jcp.2017.01.050. (Cited on p. 234)

[442] B. P EHERSTORFER, Model reduction for transport-dominated problems via online adaptive bases
and adaptive sampling, SIAM Journal on Scientific Computing, 42 (2020), pp. A2803–A2836,
https://doi.org/10.1137/19M1257275. (Cited on pp. 298, 399)

[443] B. P EHERSTORFER , K. W ILLCOX , AND M. G UNZBURGER, Survey of multifidelity methods in

uncertainty propagation, inference, and optimization, SIAM Review, 60 (2018), pp. 550–591,
https://doi.org/10.1137/16M1082469. (Cited on p. 334)

[444] Z. P ENG , M. WANG , AND F. L I, A Learning-Based Projection Method for Model Order Reduction
of Transport Problems, arXiv:2105.14633 [cs, math] (2021). (Cited on p. 399)

[445] P. P ERDIKARIS , M. R AISSI , A. DAMIANOU , N. D. L AWRENCE , AND G. E. K ARNIADAKIS,

Nonlinear information fusion algorithms for data-efficient multi-fidelity modelling, Proceedings of
the Royal Society A, 473 (2017), p. 20160751, https://doi.org/10.1098/rspa.2016.0751.
(Cited on p. 334)

[446] F. P ICHI, Reduced Order Models for Parametric Bifurcation Problems in Nonlinear PDEs, PhD
thesis, Scuola Internazionale Superiore di Studi Avanzati, Trieste. (Cited on pp. 25, 97, 311, 316)

[447] F. P ICHI , F. BALLARIN , G. ROZZA , AND J. S. H ESTHAVEN, An Artificial Neural Network Ap-
proach to Bifurcating Phenomena in Computational Fluid Dynamics (2021). arXiv:2109.10765.
(Cited on pp. 11, 97, 98, 119)

[448] F. P ICHI , F. BALLARIN , G. ROZZA , AND J. S. H ESTHAVEN, Artificial neural network for bifur-
cating phenomena modelled by nonlinear parametrized PDEs, PAMM, 20 (2021), p. e202000350,
https://doi.org/10.1002/pamm.202000350. (Cited on pp. 98, 206, 311)

[449] F. P ICHI , J. E FTANG , G. ROZZA , AND A. T. PATERA, Reduced Order Models for the Buckling of
Hyperelastic Beams. MIT-FVG “ROM2S" report, 2020. (Cited on pp. 97, 115)

[450] F. P ICHI , A. Q UAINI , AND G. ROZZA, A reduced order modeling technique to study bifurcating
phenomena: Application to the Gross–Pitaevskii equation, SIAM Journal on Scientific Computing,
42 (2020), pp. B1115–B1135, https://doi.org/10.1137/20M1313106. (Cited on pp. 97, 112,
316)
 Bibliography 447

[451] F. P ICHI AND G. ROZZA, Reduced basis approaches for parametrized bifurcation problems held
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

by non-linear Von Kármán equations, Journal of Scientific Computing, 81 (2019), pp. 112–135,
https://doi.org/10.1007/s10915-019-01003-3. (Cited on pp. 97, 104, 109)
[452] F. P ICHI , M. S TRAZZULLO , F. BALLARIN , AND G. ROZZA, Driving Bifurcating Parametrized
Nonlinear PDEs by Optimal Control Strategies: Application to Navier-Stokes Equations with Model
Order Reduction. arXiv:2010.13506, 2020. (Cited on pp. 25, 89, 95, 97, 98, 104, 119, 311)
[453] A. P INKUS, Ridge Functions, vol. 205, Cambridge University Press, 2015, https://doi.org/10.
1017/CBO9781316408124. (Cited on p. 328)
[454] M. P INTORE , F. P ICHI , M. H ESS , G. ROZZA , AND C. C ANUTO, Efficient computation of bi-
furcation diagrams with a deflated approach to reduced basis spectral element method, Advances
in Computational Mathematics, 47 (2020), https://doi.org/10.1007/s10444-020-09827-6.
(Cited on pp. 25, 98, 114, 119, 229, 372, 373, 381)
[455] S. P IPERNO AND C. FARHAT, Design of efficient partitioned procedures for the transient solution
of aeroelastic problems, Revue Européenne des Éléments Finis, 9 (2000), pp. 655–680, https:
//doi.org/10.1080/12506559.2000.10511480. (Cited on p. 286)
[456] L. P. P ITAEVSKII AND S. S TRINGARI, Bose-Einstein Condensation, International Series of Mono-
graphs on Physics, Oxford University Press, 2003, https://books.google.it/books?id=
rIobbOxC4j4C. (Cited on p. 112)
[457] G. P ITTON , A. Q UAINI , AND G. ROZZA, Computational reduction strategies for the detection of
steady bifurcations in incompressible fluid-dynamics: Applications to Coanda effect in cardiology,
Journal of Computational Physics, 344 (2017), pp. 534–557, https://doi.org/10.1016/j.jcp.
2017.05.010. (Cited on pp. 229, 311, 372, 373, 374)
[458] G. P ITTON AND G. ROZZA, On the application of reduced basis methods to bifurcation problems
in incompressible fluid dynamics, Journal of Scientific Computing, 73 (2017), pp. 157–177, https:
//doi.org/10.1007/s10915-017-0419-6. (Cited on p. 229)
[459] T. P ORSCHING, Estimation of the error in the reduced basis method solution of nonlinear
equations, Mathematics of Computation, 45 (1985), pp. 487–496, https://doi.org/10.1090/
S0025-5718-1985-0804937-0. (Cited on p. 3)
[460] T. A. P ORSCHING AND M. L. L EE, The reduced basis method for initial value problems, SIAM
Journal on Numerical Analysis, 24 (1987), pp. 1277–1287, https://doi.org/10.1137/0724083.
(Cited on p. 3)
[461] M. P OŠTA AND T. ROUBÍ ČEK, Optimal control of Navier–Stokes equations by Oseen approxima-
tion, Computers & Mathematics with Applications, 53 (2007), pp. 569–581, https://doi.org/
10.1016/j.camwa.2006.02.034. (Cited on p. 83)
[462] L. P RANDTL, Uber die ausgebildete Turbulenz, vol. 5, ZAMM, 1925. (Cited on p. 167)
[463] J. L. P ROCTOR , S. L. B RUNTON , AND J. N. K UTZ, Dynamic mode decomposition with control,
SIAM Journal on Applied Dynamical Systems, 15 (2016), pp. 142–161, https://doi.org/10.
1137/15M1013857. (Cited on pp. 209, 210)
[464] C. P RUD ’ HOMME , D. V. ROVAS , K. V EROY, L. M ACHIELS , Y. M ADAY, A. T. PATERA , AND
G. T URINICI, Reliable real-time solution of parametrized partial differential equations: Reduced-
basis output bound methods, Journal of Fluids Engineering, 124 (2002), pp. 70–80, https://doi.
org/10.1115/1.1448332. (Cited on p. 83)
[465] C. P RUD ’ HOMME , D. V. ROVAS , K. V EROY, AND A. T. PATERA, A mathematical and com-
putational framework for reliable real-time solution of parametrized partial differential equa-
tions, ESAIM: Mathematical Modelling and Numerical Analysis, 36 (2002), pp. 747–771, https:
//doi.org/10.1051/m2an:2002035. (Cited on p. 127)
 448 Bibliography

[466] A. Q UAINI, Algorithms for Fluid–Structure Interaction Problems Arising in Hemodynamics, PhD
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

thesis, EPFL, Lausanne, 2009. (Cited on p. 283)

[467] A. Q UAINI , R. G LOWINSKI , AND S. C ANIC, Symmetry breaking and preliminary results about
a Hopf bifurcation for incompressible viscous flow in an expansion channel, International Journal
of Computational Fluid Dynamics, 30 (2016), pp. 7–19, https://doi.org/10.1080/10618562.
2016.1144877. (Cited on p. 311)

[468] A. Q UARTERONI, Numerical Models for Differential Problems, MS&A Series Vol. 16, Springer
International, 2017, https://doi.org/10.1007/978-3-319-49316-9. (Cited on pp. 8, 103,
367, 373)

[469] A. Q UARTERONI AND L. F ORMAGGIA, Mathematical modelling and numerical simulation of the
cardiovascular system, Handbook of Numerical Analysis, 12 (2004), pp. 3–127. (Cited on pp. 283,
300)

[470] A. Q UARTERONI , A. M ANZONI , AND F. N EGRI, Reduced Basis Methods for Partial Dif-
ferential Equations: An Introduction, vol. 92, Springer, 2015, https://doi.org/10.1007/
978-3-319-15431-2. (Cited on pp. 62, 103, 106, 108, 128, 205)

[471] A. Q UARTERONI , A. M ANZONI , AND G. ROZZA, Certified reduced basis approximation for
parametrized partial differential equations and applications, Journal of Mathematics in Industry,
1 (2011), https://doi.org/10.1186/2190-5983-1-3. (Cited on pp. 62, 125, 127, 129, 130)

[472] A. Q UARTERONI AND G. ROZZA, Numerical solution of parametrized Navier–Stokes equations by

reduced basis methods, Numerical Methods for Partial Differential Equations, 23 (2007), pp. 923–
948, https://doi.org/10.1002/num.20249. (Cited on p. 367)

[473] A. Q UARTERONI , G. ROZZA , L. D EDÈ , AND A. Q UAINI, Numerical approximation of a control

problem for advection-diffusion processes, in Ceragioli F., Dontchev A., Futura H., Marti K., Pandolfi
L. (eds.), System Modeling and Optimization. International Federation for Information Processing,
CSMO Conference on System Modeling and Optimization, Springer, 2005, vol. 199, pp. 261–273.
(Cited on p. 83)

[474] A. Q UARTERONI , G. ROZZA , AND A. M ANZONI, Certified reduced basis approximation for
parametrized partial differential equations and applications, Journal of Mathematics in Industry,
1 (2011), https://doi.org/10.1186/2190-5983-1-3. (Cited on p. 6)

[475] A. Q UARTERONI , G. ROZZA , AND A. Q UAINI, Reduced basis methods for optimal control of
advection-diffusion problems, in Advances in Numerical Mathematics, no. CMCS-CONF-2006-003,
RAS and University of Houston, 2007, pp. 193–216. (Cited on p. 83)

[476] A. Q UARTERONI , M. T UVERI , AND A. V ENEZIANI, Computational vascular fluid dynamics:

problems, models and methods, Computing and Visualization in Science, 2 (2000), pp. 163–197,
https://doi.org/10.1007/s007910050039. (Cited on pp. 283, 300)

[477] A. Q UARTERONI AND A. VALLI, Numerical Approximation of Partial Differential Equations,

vol. 23, Springer Berlin Heidelberg, 1994, https://doi.org/10.1007/978-3-540-85268-1.
(Cited on pp. 15, 16, 21, 33, 86, 103, 125, 127, 129, 132, 259)

[478] J. R ABINOVITCH , D. Z. H UANG , R. B ORKER , P. AVERY, C. FARHAT, A. D ERKEVORKIAN , AND

L. P ETERSON, Towards a Validated FSI Computational Framework for Supersonic Parachute De-
ployments, in AIAA Aviation 2019 Forum. (Cited on p. 283)

[479] A. R AHIMI AND B. R ECHT, Random features for large-scale kernel machines, in Proceedings of
the 20th International Conference on Neural Information Processing Systems, NIPS’07, Curran As-
sociates Inc., 2008, pp. 1177–1184. (Cited on p. 332)
 Bibliography 449

[480] M. R AISSI , P. P ERDIKARIS , AND G. E. K ARNIADAKIS, Physics-informed neural networks: A deep

Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

learning framework for solving forward and inverse problems involving nonlinear partial differential
equations, Journal of Computational Physics, 378 (2019), pp. 686–707, https://doi.org/10.
1016/j.jcp.2018.10.045. (Cited on pp. 11, 401)

[481] A. R ASHEED , O. S AN , AND T. K VAMSDAL, Digital twin: Values, challenges and enablers from
a modeling perspective, IEEE Access, 8 (2020), pp. 21980–22012, https://doi.org/10.1109/
ACCESS.2020.2970143. (Cited on pp. 12, 410)

[482] S. S. R AVINDRAN, A reduced-order approach for optimal control of fluids using proper or-
thogonal decomposition, International Journal for Numerical Methods in Fluids, 34 (2000),
pp. 425–448, https://doi.org/10.1002/1097-0363(20001115)34:5<425::AID-FLD67>3.
0.CO;2-W. (Cited on p. 284)

[483] W. R EED AND T. H ILL, Triangular mesh methods for the neutron transport equation, Journal of
Scientific Computing, (1973). (Cited on p. 234)

[484] J. R EISS , P. S CHULZE , J. S ESTERHENN , AND V. M EHRMANN, The shifted proper orthogonal de-
composition: A mode decomposition for multiple transport phenomena, SIAM Journal on Scientific
Computing, 40 (2018), pp. A1322–A1344, https://doi.org/10.1137/17M1140571. (Cited on
pp. 11, 399)

[485] O. R EYNOLDS, IV. On the dynamical theory of incompressible viscous fluids and the determination
of the criterion, Philosophical Transactions of the Royal Society of London (A), 186 (1895), pp. 123–
164, https://doi.org/10.1098/rsta.1895.0004. (Cited on p. 165)

[486] C. M. R HIE AND W. L. C HOW, Numerical study of the turbulent flow past an airfoil with trailing
edge separation, AIAA Journal, 21 (1983), pp. 1525–1532, https://doi.org/10.2514/3.8284.
(Cited on p. 145)

[487] T. R ICHTER, Fluid–Structure Interactions. Model, Analysis and Finite Element, vol. 118 of Lecture
Notes in Computational Science and Engineering, Springer International, 2017. (Cited on pp. 284,
285, 286, 287)

[488] D. F. ROGERS, Chapter 7—Rational B-spline surfaces, in An Introduction to NURBS, D. F. Rogers,

ed., The Morgan Kaufmann Series in Computer Graphics, Morgan Kaufmann, 2001, pp. 209–262,
https://doi.org/10.1016/B978-155860669-2/50007-5. (Cited on pp. 357, 358)

[489] D. F. ROGERS AND J. A. A DAMS, Mathematical Elements for Computer Graphics (2nd ed.),
McGraw-Hill, Inc., USA, 1989. (Cited on p. 357)

[490] R. ROJAS, The backpropagation algorithm, in Neural Networks, Springer, 1996, pp. 149–182,
https://doi.org/10.1007/978-3-642-61068-4_7. (Cited on pp. 391, 392, 393)

[491] R. ROJAS, Neural Networks: A Systematic Introduction, Springer-Verlag, 1996. (Cited on p. 397)

[492] R. ROMAROWSKI , A. L EFIEUX , A. V ENEZIANI , S. M ORGANTI , AND F. AURICCHIO, Patient-

specific CFD modeling in the thoracic aorta with PC-MRI based boundary conditions: A least-
square 3-element Windkessel approach: Patient-specific tuning of 3-element Windkessel in the tho-
racic aorta, International Journal for Numerical Methods in Biomedical Engineering, 34 (2018),
p. e3134, https://doi.org/10.1002/cnm.3134. (Cited on p. 374)

[493] F. ROMOR , M. T EZZELE , A. L ARIO , AND G. ROZZA, Kernel-based Active Subspaces with
Application to CFD Parametric Problems Using Discontinuous Galerkin Method, arXiv preprint
arXiv:2008.12083 (2020). (Cited on pp. 10, 327, 329, 333, 339, 384)

[494] F. ROMOR , M. T EZZELE , M. M ROSEK , C. OTHMER , AND G. ROZZA, Multi-fidelity Data Fusion
through Parameter Space Reduction with Applications to Automotive Engineering, arXiv preprint
arXiv:2110.14396 (2021). (Cited on pp. 9, 206, 327, 334, 336, 384, 411)
 450 Bibliography

[495] F. ROMOR , M. T EZZELE , AND G. ROZZA, A Local Approach to Parameter Space Reduction for
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Regression and Classification Tasks, arXiv preprint arXiv:2107.10867 (2021). (Cited on pp. 10, 327,
333, 384)

[496] F. ROMOR , M. T EZZELE , AND G. ROZZA, ATHENA: Advanced techniques for high dimensional
parameter spaces to enhance numerical analysis, Software Impacts, 10 (2021), p. 100133, https:
//doi.org/10.1016/j.simpa.2021.100133. (Cited on pp. 9, 327, 330, 337, 384, 407)

[497] F. ROMOR , M. T EZZELE , AND G. ROZZA, Multi-fidelity data fusion for the approximation of scalar
functions with low intrinsic dimensionality through active subspaces, in Proceedings in Applied
Mathematics & Mechanics, vol. 20, Wiley Online Library, 2021, https://doi.org/10.1002/
pamm.202000349. (Cited on pp. 334, 384)

[498] F. ROSENBLATT, The perceptron: A probabilistic model for information storage and organization
in the brain, Psychological Review, 65 (1958), p. 386, https://doi.org/10.1037/h0042519.
(Cited on p. 393)

[499] W. ROTH , G. S CHINDLER , M. Z ÖHRER , L. P FEIFENBERGER , R. P EHARZ , S. T SCHIATSCHEK ,

H. F RÖNING , F. P ERNKOPF, AND Z. G HAHRAMANI, Resource-Efficient Neural Networks for Em-
bedded Systems, arXiv preprint arXiv:2001.03048 (2020). (Cited on pp. 402, 403)

[500] C. W. ROWLEY, I. G. K EVREKIDIS , J. E. M ARSDEN , AND K. L UST, Reduction and reconstruction

for self-similar dynamical systems, Nonlinearity, 16 (2003), pp. 1257–1275, https://doi.org/
10.1088/0951-7715/16/4/304. (Cited on p. 298)

[501] G. ROZZA, On optimization, control and shape design of an arterial bypass, International Journal for
Numerical Methods in Fluids, 47 (2005), pp. 1411–1419, https://doi.org/10.1002/fld.888.
(Cited on p. 368)

[502] G. ROZZA, Reduced basis methods for Stokes equations in domains with non-affine parameter de-
pendence, Computing and Visualization in Science, 12 (2009), pp. 23–35, https://doi.org/10.
1007/s00791-006-0044-7. (Cited on p. 4)

[503] G. ROZZA , M. H ESS , G. S TABILE , M. T EZZELE , AND F. BALLARIN, Basic ideas and tools
for projection-based model reduction of parametric partial differential equations, in Model Or-
der Reduction, P. Benner, S. Grivet-Talocia, A. Quarteroni, G. Rozza, W. H. A. Schilders, and
L. M. Silveira, eds., vol. 2, De Gruyter, 2020, ch. 1, pp. 1–47, https://doi.org/10.1515/
9783110671490-001. (Cited on pp. 9, 10, 327, 347)

[504] G. ROZZA , D. H UYNH , N. N GUYEN , AND A. PATERA, Real-time reliable simulation of heat
transfer phenomena, in ASME—American Society of Mechanical Engineers—Heat Transfer Sum-
mer Conference Proceedings, San Francisco, 2009, https://doi.org/10.1115/HT2009-88212.
(Cited on p. 253)

[505] G. ROZZA , D. H UYNH , AND A. T. PATERA, Reduced basis approximation and a posteriori er-
ror estimation for affinely parametrized elliptic coercive partial differential equations, Archives of
Computational Methods in Engineering, 15 (2008), pp. 229–275, https://doi.org/10.1007/
s11831-008-9019-9. (Cited on pp. 3, 5, 83, 91, 94, 127, 251, 253, 258, 268)

[506] G. ROZZA , D. B. P. H UYNH , AND A. M ANZONI, Reduced basis approximation and a poste-
riori error estimation for Stokes flows in parametrized geometries: Roles of the inf-sup stabil-
ity constants, Numerische Mathematik, 125 (2013), pp. 115–152, https://doi.org/10.1007/
s00211-013-0534-8. (Cited on p. 83)

[507] G. ROZZA , A. KOSHAKJI , AND A. Q UARTERONI, Free form deformation techniques applied to
3D shape optimization problems, Communications in Applied and Industrial Mathematics, 4 (2013),
https://doi.org/10.1685/journal.caim.452. (Cited on p. 347)
 Bibliography 451

[508] G. ROZZA , M. H. M ALIK , N. D EMO , M. T EZZELE , M. G IRFOGLIO , G. S TABILE , AND A. M OLA,

Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Advances in reduced order methods for parametric industrial problems in computational fluid dy-
namics, in ECCOMAS ECFD 7—Proceedings of 6th European Conference on Computational Me-
chanics (ECCM 6) and 7th European Conference on Computational Fluid Dynamics (ECFD 7),
R. Owen, R. de Borst, J. Reese, and P. Chris, eds., Glasgow, UK, 2018, pp. 59–76. (Cited on pp. 11,
37, 203, 327, 347)

[509] G. ROZZA , A. M ANZONI , AND F. N EGRI, Reduction strategies for PDE-constrained optimization
problems in haemodynamics, in Proceedings of the 6th European Congress on Computational Meth-
ods in Applied Sciences and Engineering, Vienna Technical University, 2012, pp. 1748–1769. (Cited
on p. 347)

[510] G. ROZZA , N. N GUYEN , A. PATERA , AND S. D EPARIS, Reduced Basis Methods and a Pos-
teriori Error Estimators for Heat Transfer Problems, in ASME—American Society of Mechan-
ical Engineers—Heat Transfer Summer Conference Proceedings, San Francisco, 2009, https:
//doi.org/10.1115/HT2009-88211. (Cited on p. 253)

[511] G. ROZZA AND K. V EROY, On the stability of the reduced basis method for Stokes equations in
parametrized domains, Computer Methods in Applied Mechanics and Engineering, 196 (2007),
pp. 1244–1260, https://doi.org/10.1016/j.cma.2006.09.005. (Cited on pp. 39, 258, 272)

[512] S. RUBINO, Numerical Modeling of Turbulence by Richardson Number-Based and VMS Models,
PhD thesis, University of Seville, 2014. (Cited on pp. 61, 69, 133)

[513] S. RUBINO, Numerical analysis of a projection-based stabilized POD-ROM for incompressible

flows, SIAM Journal on Numerical Analysis, 58 (2020), pp. 2019–2058, https://doi.org/10.
1137/19m1276686. (Cited on p. 144)

[514] P. G. S AFFMAN, A model for inhomogeneous turbulent flow, Proceedings of the Royal Society of
London. A. Mathematical and Physical Sciences, 317 (1970), pp. 417–433, https://doi.org/10.
1098/rspa.1970.0125. (Cited on p. 168)

[515] P. S AGAUT, Large Eddy Simulation for Incompressible Flows: An Introduction., Springer-Verlag,
2006, https://doi.org/10.1007/b137536. (Cited on pp. 59, 60)

[516] T. N. S AINATH , B. K INGSBURY, V. S INDHWANI , E. A RISOY, AND B. R AMABHADRAN, Low-

rank matrix factorization for deep neural network training with high-dimensional output targets,
in 2013 IEEE International Conference on Acoustics, Speech and Signal Processing, IEEE, 2013,
pp. 6655–6659, https://doi.org/10.1109/ICASSP.2013.6638949. (Cited on p. 403)

[517] F. S ALMOIRAGHI , F. BALLARIN , G. C ORSI , A. M OLA , M. T EZZELE , AND G. ROZZA, Advances

in geometrical parametrization and reduced order models and methods for computational fluid dy-
namics problems in applied sciences and engineering: Overview and perspectives, ECCOMAS
Congress 2016—Proceedings of the 7th European Congress on Computational Methods in Applied
Sciences and Engineering, 1 (2016), pp. 1013–1031, https://doi.org/10.7712/100016.1867.
8680. (Cited on pp. 11, 203, 327, 347)

[518] F. S ALMOIRAGHI , F. BALLARIN , L. H ELTAI , AND G. ROZZA, Isogeometric analysis-based re-

duced order modelling for incompressible linear viscous flows in parametrized shapes, Advanced
Modeling and Simulation in Engineering Sciences, 3 (2016), p. 21, https://doi.org/10.1186/
s40323-016-0076-6. (Cited on p. 347)

[519] F. S ALMOIRAGHI , A. S CARDIGLI , H. T ELIB , AND G. ROZZA, Free-form deformation, mesh mor-
phing and reduced-order methods: Enablers for efficient aerodynamic shape optimisation, Interna-
tional Journal of Computational Fluid Dynamics, 32 (2018), pp. 233–247, https://doi.org/10.
1080/10618562.2018.1514115. (Cited on pp. 173, 205, 206, 347)
 452 Bibliography

[520] S. S ALSA, Partial Differential Equations in Action: From Modelling to Theory, UNITEXT, Springer
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

International, 2016, https://doi.org/10.1007/978-3-319-15093-2. (Cited on p. 100)

[521] J. S AMAREH, Aerodynamic shape optimization based on free-form deformation, in 10th

AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference, 2004, p. 4630, https:
//doi.org/10.2514/6.2004-4630. (Cited on p. 347)

[522] O. S AN , R. M AULIK , AND M. A HMED, An artificial neural network framework for reduced order
modeling of transient flows, Communications in Nonlinear Science and Numerical Simulation, 77
(2019), pp. 271–287, https://doi.org/10.1016/j.cnsns.2019.04.025. (Cited on p. 397)

[523] D. T. S ANDWELL, Biharmonic spline interpolation of GEOS-3 and SEASAT altimeter

data, Geophysical Research Letters, 14 (1987), pp. 139–142, https://doi.org/10.1029/
GL014i002p00139. (Cited on p. 348)

[524] M. S ANKARANARAYANAN , L. C HUA , D. N. G HISTA , AND Y. S. TAN, Computational model of

blood flow in the aorto-coronary bypass graft, BioMedical Engineering OnLine, 4 (2005), p. 14,
https://doi.org/10.1186/1475-925X-4-14. (Cited on p. 368)

[525] M. S CHÄFER , S. T UREK , F. D URST, E. K RAUSE , AND R. R ANNACHER, Benchmark Computations

of Laminar Flow Around a Cylinder, Vieweg+Teubner Verlag, 1996, pp. 547–566, https://doi.
org/10.1007/978-3-322-89849-4_39. (Cited on p. 157)

[526] M. S CHÄFER , S. T UREK , F. D URST, E. K RAUSE , AND R. R ANNACHER, Benchmark Computations

of Laminar Flow Around a Cylinder, vol. 48, Vieweg+Teubner Verlag, 1996, pp. 547–566, https:
//doi.org/10.1007/978-3-322-89849-4_39. (Cited on p. 187)

[527] P. J. S CHMID, Dynamic mode decomposition of numerical and experimental data, Journal of Fluid
Mechanics, 656 (2010), pp. 5–28, https://doi.org/10.1017/S0022112010001217. (Cited on
pp. 207, 208, 383)

[528] J. S CHMIDHUBER, Deep learning in neural networks: An overview, Neural networks, 61 (2015),
pp. 85–117, https://doi.org/10.1016/j.neunet.2014.09.003. (Cited on p. 396)

[529] E. S CHMIDT, Zur theorie der linearen und nichtlinearen integralgleichungen, Mathematische An-
nalen, 63 (1907), pp. 433–476, https://doi.org/10.1007/BF01449770. (Cited on p. 4)

[530] I. J. S CHOENBERG, Spline functions and the problem of graduation, Proceedings of the National
Academy of Sciences, 52 (1964), pp. 947–950, https://doi.org/10.1073/pnas.52.4.947.
(Cited on p. 357)

[531] B. S CHÖLKOPF, A. S MOLA , AND K.-R. M ÜLLER, Kernel principal component analysis, in
International Conference on Artificial Neural Networks, Springer, 1997, pp. 583–588, https:
//doi.org/10.1007/BFb0020217. (Cited on p. 205)

[532] R. S CHREIBER AND H. K ELLER, Driven cavity flows by efficient numerical techniques, Journal of
Computational Physics, 49 (1983), pp. 310–333, https://doi.org/10.1016/0021-9991(83)
90129-8. (Cited on p. 157)

[533] T. S EDERBERG AND S. PARRY, Free-form deformation of solid geometric models, in Proceedings
of SIGGRAPH—Special Interest Group on GRAPHics and Interactive Techniques, vol. 20, SIG-
GRAPH, 1986, pp. 151–159, https://doi.org/10.1145/15886.15903. (Cited on pp. 10, 212,
346, 347)

[534] P. S ESHADRI , S. S HAHPAR , P. C ONSTANTINE , G. PARKS , AND M. A DAMS, Turbomachinery

active subspace performance maps, Journal of Turbomachinery, 140 (2018), p. 041003, https:
//doi.org/10.1115/1.4038839. (Cited on p. 327)
 Bibliography 453

[535] R. S EYDEL, Practical Bifurcation and Stability Analysis, Interdisciplinary Applied Mathematics,
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Springer, 2009, https://doi.org/10.1007/978-1-4419-1740-9. (Cited on p. 101)

[536] N. V. S HAH , M. G IRFOGLIO , P. Q UINTELA , G. ROZZA , A. L ENGOMIN , F. BALLARIN , AND

P. BARRAL, Finite Element Based Model Order Reduction for Parametrized One-Way Coupled
Steady State Linear Thermomechanical Problems, arXiv:2111.08534 (2021). (Cited on p. 398)

[537] D. S HEPARD, A two-dimensional interpolation function for irregularly-spaced data, in

Proceedings—1968 ACM National Conference, ACM, 1968, pp. 517–524, https://doi.org/10.
1145/800186.810616. (Cited on pp. 10, 346, 350)

[538] K. S IMONYAN AND A. Z ISSERMAN, Very Deep Convolutional Networks for Large-Scale Image
Recognition, in 3rd International Conference on Learning Representations, 2015. (Cited on p. 407)

[539] S. S IRISUP AND G. K ARNIADAKIS, Stability and accuracy of periodic flow solutions obtained
by a POD-penalty method, Physica D: Nonlinear Phenomena, 202 (2005), pp. 218–237, https:
//doi.org/10.1016/j.physd.2005.02.006. (Cited on pp. 154, 155, 189)

[540] M. S LAUGHTER , J. ROGERS , C. M ILANO , S. RUSSELL , J. C ONTE , D. F ELDMAN , B. S UN ,

A. TATOOLES , R. D ELGADO , J. L ONG , T. W OZNIAK , W. G HUMMAN , D. FARRAR , O. F RA -
ZIER , AND A. E HSAN , Advanced heart failure treated with continuous-flow left ventricular assist
device, The New England Journal of Medicine, 361 (2009), pp. 2241–2251, https://doi.org/
10.1056/NEJMoa0909938. (Cited on p. 374)

[541] J. S MAGORINSKY, General circulation experiments with the primitive equations. I. The basic
experiment, Monthly Weather Review, 91 (1963), pp. 99–164, https://doi.org/10.1175/
1520-0493(1963)091<0099:GCEWTP>2.3.CO;2. (Cited on pp. 59, 60, 69, 169)

[542] R. L. S MITH, The hit-and-run sampler: A globally reaching Markov chain sampler for generat-
ing arbitrary multivariate distributions, in Proceedings Winter Simulation Conference, IEEE, 1996,
pp. 260–264, https://doi.org/10.1109/WSC.1996.873287. (Cited on p. 329)

[543] I. J. S OBEY AND P. G. D RAZIN, Bifurcations of two-dimensional channel flows, Journal of Fluid
Mechanics, 171 (1986), pp. 263–287, https://doi.org/10.1017/S0022112086001441. (Cited
on p. 311)

[544] I. M. S OBOL’, On the distribution of points in a cube and the approximate evaluation of integrals,
Zhurnal Vychislitel’noi Matematiki i Matematicheskoi Fiziki, 7 (1967), pp. 784–802, https://
doi.org/10.1016/0041-5553(67)90144-9. (Cited on p. 204)

[545] I. S OMMERVILLE, Software Engineering, (7th ed.), Pearson Education India, 2004. (Cited on p. 386)

[546] P. S PALART AND S. A LLMARAS, A One-Equation Turbulence Model for Aerodynamic Flows, in
30th Aerospace Sciences Meeting and Exhibit, American Institute of Aeronautics and Astronautics,
January 1992, https://doi.org/10.2514/6.1992-439. (Cited on pp. 167, 219)

[547] D. F. S PECHT, A general regression neural network, IEEE Transactions on Neural Networks, 2
(1991), pp. 568–576, https://doi.org/10.1109/72.97934. (Cited on p. 389)

[548] G. S TABILE , S. H IJAZI , A. M OLA , S. L ORENZI , AND G. ROZZA, POD-Galerkin reduced order
methods for CFD using finite volume discretisation: Vortex shedding around a circular cylinder,
Communications in Applied and Industrial Mathematics, 8 (2017), pp. 210–236, https://doi.
org/10.1515/caim-2017-0011. (Cited on pp. 143, 144, 151, 155, 187, 381)

[549] G. S TABILE AND G. ROZZA, ITHACA-FV—In real Time Highly Advanced Computational Appli-
cations for Finite Volumes. http://www.mathlab.sissa.it/ithaca-fv. (Cited on pp. 181,
381)
 454 Bibliography

[550] G. S TABILE AND G. ROZZA, Finite volume POD-Galerkin stabilised reduced order methods for the
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

parametrised incompressible Navier–Stokes equations, Computers & Fluids, 173 (2018), pp. 273–
284, https://doi.org/10.1016/j.compfluid.2018.01.035. (Cited on pp. 39, 143, 144, 219,
381)
[551] G. S TABILE , M. Z ANCANARO , AND G. ROZZA, Efficient geometrical parametrization for finite-
volume-based reduced order methods, International Journal for Numerical Methods in Engineering,
121 (2020), pp. 2655–2682, https://doi.org/10.1002/nme.6324. (Cited on pp. 144, 381)

[552] K. S TAR , G. S TABILE , S. G EORGAKA , F. B ELLONI , G. ROZZA , AND J. D EGROOTE, POD-

Galerkin reduced order model of the Boussinesq approximation for buoyancy-driven enclosed flows,
in Building Theory and Applications: Proceedings of M&C 2019, American Nuclear Society (ANS),
Illinois, 2019, pp. 2452–2461. (Cited on p. 144)
[553] K. S TAR , G. S TABILE , G. ROZZA , AND J. D EGROOTE, A POD-Galerkin reduced order model of
a turbulent convective buoyant flow of sodium over a backward-facing step, Applied Mathematical
Modelling, 89 (2021), pp. 486–503, https://doi.org/10.1016/j.apm.2020.07.029. (Cited
on p. 381)

[554] S. K. S TAR , G. S TABILE , F. B ELLONI , G. ROZZA , AND J. D EGROOTE, Extension and comparison
of techniques to enforce boundary conditions in finite volume POD-Galerkin reduced order models
for fluid dynamic problems, arXiv:1912.00825 (2019). (Cited on p. 156)
[555] M. S TEIN, Large sample properties of simulations using Latin hypercube sampling, Technomet-
rics, 29 (1987), pp. 143–151, https://doi.org/10.1080/00401706.1987.10488205. (Cited
on p. 204)

[556] G. S TRANG, Linear Algebra and Its Applications, Academic Press, New York, 3rd ed., 1988. (Cited
on p. 4)

[557] M. S TRAZZULLO , F. BALLARIN , R. M OSETTI , AND G. ROZZA, Model reduction for parametrized
optimal control problems in environmental marine sciences and engineering, SIAM Journal on
Scientific Computing, 40 (2018), pp. B1055–B1079, https://doi.org/10.1137/17M1150591.
(Cited on pp. 83, 261, 262)

[558] M. S TRAZZULLO , F. BALLARIN , AND G. ROZZA, POD-Galerkin model order reduction for
parametrized time dependent linear quadratic optimal control problems in saddle point formulation,
Journal of Scientific Computing, 83 (2020), https://doi.org/10.1007/s10915-020-01232-x.
(Cited on pp. 84, 92, 205)

[559] M. S TRAZZULLO , F. BALLARIN , AND G. ROZZA, A Certified Reduced Basis Method for Linear
Parametrized Parabolic Optimal Control Problems in Space-Time Formulation, arXiv:2103.00460.
(Cited on p. 91)

[560] M. S TRAZZULLO , F. BALLARIN , AND G. ROZZA, POD-Galerkin model order reduction for
parametrized nonlinear time dependent optimal flow control: An application to shallow water equa-
tions, Journal of Numerical Mathematics, 30 (2022), pp. 63–84, https://doi.org/10.1515/
jnma-2020-0098. (Cited on pp. 89, 91, 95)

[561] M. S TRAZZULLO , M. G IRFOGLIO , F. BALLARIN , T. I LIESCU , AND G. ROZZA, Consistency of

the full and reduced order models for evolve-filter-relax regularization of convection-dominated,
marginally-resolved flows, International Journal for Numerical Methods in Engineering, 123 (2022),
pp. 3148–3178, https://doi.org/10.1002/nme.6942. (Cited on p. 171)

[562] M. S TRAZZULLO , Z. Z AINIB , F. BALLARIN , AND G. ROZZA, Reduced order methods for
parametrized non-linear and time dependent optimal flow control problems, towards applications
in biomedical and environmental sciences, in Numerical Mathematics and Advanced Applications
ENUMATH 2019, F. J. Vermolen and C. Vuik, eds., Springer International, 2020, pp. 841–850,
https://doi.org/10.1007/978-3-030-55874-1_83. (Cited on pp. 83, 367)
 Bibliography 455

[563] V. S TROUHAL, Ueber eine besondere art der tonerregung, Annalen der Physik und Chemie, 241
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

(1878), pp. 216–251, https://doi.org/10.1002/andp.18782411005. (Cited on p. 188)

[564] B. S UDRET, Global sensitivity analysis using polynomial chaos expansions, Reliability Engineering
& System Safety, 93 (2008), pp. 964–979, https://doi.org/10.1016/j.ress.2007.04.002.
(Cited on p. 405)

[565] T. S ULLIVAN, Introduction to Uncertainty Quantification, Springer, 2015, https://doi.org/10.

1007/978-3-319-23395-6. (Cited on p. 251)

[566] S. S Y AND C. M UREA, Algorithm for solving fluid–structure interaction problem on a global moving
mesh, Coupled Systems Mechanics, 1 (2012), pp. 99–113, https://doi.org/10.12989/csm.
2012.1.1.099. (Cited on p. 303)
[567] T. TADDEI, A registration method for model order reduction: Data compression and geometry re-
duction, SIAM Journal on Scientific Computing, 42 (2020), pp. A997–A1027, https://doi.org/
10.1137/19M1271270. (Cited on p. 298)

[568] M. T EZZELE , F. BALLARIN , AND G. ROZZA, Combined parameter and model reduction of car-
diovascular problems by means of active subspaces and POD-Galerkin methods, in Mathematical
and Numerical Modeling of the Cardiovascular System and Applications, D. Boffi, L. F. Pavarino,
G. Rozza, S. Scacchi, and C. Vergara, eds., vol. 16 of SEMA-SIMAI Series, Springer International,
2018, pp. 185–207, https://doi.org/10.1007/978-3-319-96649-6_8. (Cited on pp. 11, 342,
350, 356, 367)
[569] M. T EZZELE , N. D EMO , M. G ADALLA , A. M OLA , AND G. ROZZA, Model order reduction
by means of active subspaces and dynamic mode decomposition for parametric hull shape de-
sign hydrodynamics, in Technology and Science for the Ships of the Future: Proceedings of NAV
2018: 19th International Conference on Ship & Maritime Research, IOS Press, 2018, pp. 569–576,
https://doi.org/10.3233/978-1-61499-870-9-569. (Cited on pp. 37, 347)

[570] M. T EZZELE , N. D EMO , A. M OLA , AND G. ROZZA, PyGeM: Python Geometrical Morphing, Soft-
ware Impacts, 7 (2021), p. 100047, https://doi.org/10.1016/j.simpa.2020.100047. (Cited
on pp. 10, 347, 359, 384)

[571] M. T EZZELE , N. D EMO , A. M OLA , AND G. ROZZA, An integrated data-driven computational

pipeline with model order reduction for industrial and applied mathematics, in Novel Mathemat-
ics Inspired by Industrial Challenges, M. Günther and W. Schilders, eds., vol. 38 of Mathematics
in Industry, Springer International, 2022, https://doi.org/10.1007/978-3-030-96173-2_7.
(Cited on pp. 11, 203, 205, 206, 327, 347)

[572] M. T EZZELE , N. D EMO , AND G. ROZZA, Shape optimization through proper orthogonal decom-
position with interpolation and dynamic mode decomposition enhanced by active subspaces, in Pro-
ceedings of MARINE 2019: VIII International Conference on Computational Methods in Marine
Engineering, R. Bensow and J. Ringsberg, eds., 2019, pp. 122–133. (Cited on pp. 327, 347, 383)

[573] M. T EZZELE , N. D EMO , G. S TABILE , A. M OLA , AND G. ROZZA, Enhancing CFD predictions
in shape design problems by model and parameter space reduction, Advanced Modeling and Sim-
ulation in Engineering Sciences, 7 (2020), https://doi.org/10.1186/s40323-020-00177-y.
(Cited on pp. 11, 217, 219, 342, 350, 383, 384)
[574] M. T EZZELE , F. S ALMOIRAGHI , A. M OLA , AND G. ROZZA, Dimension reduction in hetero-
geneous parametric spaces with application to naval engineering shape design problems, Ad-
vanced Modeling and Simulation in Engineering Sciences, 5 (2018), https://doi.org/10.1186/
s40323-018-0118-3. (Cited on pp. 327, 346, 347, 356, 360)

[575] M. T EZZELE , M. S IDARI , M. S ICCHIERO , AND G. ROZZA, A multi-fidelity approach coupling

parameter space reduction and non-intrusive POD with application to structural optimization of
passenger ship hulls, submitted (2021). (Cited on pp. 206, 327, 342, 384, 411)
 456 Bibliography

[576] T HORATEC C ORPORATION, HeartMate 3 Left Ventricular Assist System: Instructions for Use. 2017.
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

(Cited on p. 374)

[577] S. P. T IMOSHENKO AND J. M. G ERE, Theory of Elastic Stability, Dover Civil and Mechanical
Engineering, Dover Publications, 2009. (Cited on p. 116)

[578] L. T OBISKA AND R. V ERFÜRTH, Analysis of a streamline diffusion finite element method for the
Stokes and Navier–Stokes equations, SIAM Journal on Numerical Analysis, 33 (1996), pp. 107–127,
https://doi.org/10.1137/0733007. (Cited on p. 130)

[579] D. T ORLO, Model Reduction for Advection Dominated Hyperbolic Problems in an ALE Framework:
Offline and Online Phases, arXiv:2003.13735 (2020). (Cited on p. 305)

[580] D. T ORLO , F. BALLARIN , AND G. ROZZA, Stabilized weighted reduced basis methods for
parametrized advection dominated problems with random inputs, SIAM/ASA Journal on Uncer-
tainty Quantification, 6 (2018), pp. 1475–1502, https://doi.org/10.1137/17M1163517. (Cited
on pp. 254, 259)

[581] E. F. T ORO, Riemann Solvers and Numerical Methods for Fluid Dynamics: A Practical Introduction,
Springer Science & Business Media, 2013. (Cited on p. 239)

[582] E. F. T ORO , M. S PRUCE , AND W. S PEARES, Restoration of the contact surface in the HLL-Riemann
solver, Shock Waves, 4 (1994), pp. 25–34, https://doi.org/10.1007/BF01414629. (Cited on
p. 239)

[583] S. T RENN, Multilayer perceptrons: Approximation order and necessary number of hidden units,
IEEE Transactions on Neural Networks, 19 (2008), pp. 836–844, https://doi.org/10.1109/
TNN.2007.912306. (Cited on p. 390)

[584] R. T RIPATHY, I. B ILIONIS , AND M. G ONZALEZ, Gaussian processes with built-in dimensional-
ity reduction: Applications to high-dimensional uncertainty propagation, Journal of Computational
Physics, 321 (2016), pp. 191–223, https://doi.org/10.1016/j.jcp.2016.05.039. (Cited on
p. 336)

[585] R. K. T RIPATHY AND I. B ILIONIS, Deep UQ: Learning deep neural network surrogate models for
high dimensional uncertainty quantification, Journal of Computational Physics, 375 (2018), pp. 565–
588, https://doi.org/10.1016/j.jcp.2018.08.036. (Cited on p. 332)

[586] D. J. T RITTON, Physical Fluid Dynamics, The Modern University in Physics Series, Springer
Netherlands, 1977, https://doi.org/10.1007/978-94-009-9992-3. (Cited on pp. 119, 311,
370, 372)

[587] F. T RÖLTZSCH, Optimal control of partial differential equations, Graduate Studies in Mathematics,
112 (2010), https://doi.org/10.1090/gsm/112. (Cited on p. 262)

[588] J. T U , C. ROWLEY, D. L UCHTENBURG , S. B RUNTON , AND N. K UTZ, On dynamic mode de-

composition: Theory and applications, Journal of Computational Dynamics, 1 (2014), pp. 391–421,
https://doi.org/10.3934/jcd.2014.1.391. (Cited on p. 208)

[589] M. U DELL AND A. T OWNSEND, Why are big data matrices approximately low rank?, SIAM
Journal on Mathematics of Data Science, 1 (2019), pp. 144–160, https://doi.org/10.1137/
18M1183480. (Cited on p. 205)

[590] K. U RBAN AND A. T. PATERA, A new error bound for reduced basis approximation of parabolic
partial differential equations, Comptes Rendus Mathematique, 350 (2012), pp. 203–207, https:
//doi.org/10.1016/j.crma.2012.01.026. (Cited on pp. 83, 92)

[591] S. VALLAGHÉ , M. F OUQUEMBERGH , A. H YARIC , AND C. P RUD ’ HOMME, A Successive Con-

straint Method with Minimal Offline Constraints for Lower Bounds of Parametric Coercivity Con-
stant (2011). (Cited on p. 5)
 Bibliography 457

[592] L. V ENTURI , F. BALLARIN , AND G. ROZZA, A weighted POD method for elliptic PDEs with
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

random inputs, Journal of Scientific Computing, 81 (2019), pp. 136–153, https://doi.org/10.

1007/s10915-018-0830-7. (Cited on pp. 251, 252, 253, 254)

[593] L. V ENTURI , D. T ORLO , F. BALLARIN , AND G. ROZZA, Weighted Reduced Order Methods
for Parametrized Partial Differential Equations with Random Inputs, Springer International, 2019,
pp. 27–40, https://doi.org/10.1007/978-3-030-04870-9_2. (Cited on p. 254)

[594] K. V EROY AND A. T. PATERA, Certified real-time solution of the parametrized steady incompress-
ible Navier–Stokes equations: Rigorous reduced-basis a posteriori error bounds, International Jour-
nal for Numerical Methods in Fluids, 47 (2005), pp. 773–788, https://doi.org/10.1002/fld.
867. (Cited on p. 5)

[595] K. V EROY, C. P RUD ’ HOMME , D. ROVAS , AND A. T. PATERA, A Posteriori Error Bounds for
Reduced-Basis Approximation of Parametrized Noncoercive and Nonlinear Elliptic Partial Dif-
ferential Equations, in 16th AIAA Computational Fluid Dynamics Conference, Orlando, Florida,
American Institute of Aeronautics and Astronautics, June 2003, https://doi.org/10.2514/6.
2003-3847. (Cited on pp. 4, 7)

[596] H. K. V ERSTEEG AND W. M ALALASEKERA, An Introduction to Computational Fluid Dynamics.

The Finite Volume Method, Longman Group, 1995. (Cited on pp. 143, 148)

[597] M. VOHRA , A. A LEXANDERIAN , H. G UY, AND S. M AHADEVAN, Active subspace-based dimen-

sion reduction for chemical kinetics applications with epistemic uncertainty, Combustion and Flame,
204 (2019), pp. 152–161, https://doi.org/10.1016/j.combustflame.2019.03.006. (Cited
on p. 327)

[598] T. VON K ÁRMÁN, Festigkeitsprobleme im Maschinenbau, Encyclopädie der Mathematischen Wis-

senschaften, 4 (1910). (Cited on p. 109)

[599] A. V REMAN, An eddy-viscosity subgrid-scale model for turbulent shear flow: Algebraic theory
and applications, Physics of Fluids, 16 (2004), pp. 3670–3681, https://doi.org/10.1063/1.
1785131. (Cited on p. 170)

[600] S. WALTON , O. H ASSAN , AND K. M ORGAN, Reduced order modelling for unsteady fluid flow using
proper orthogonal decomposition and radial basis functions, Applied Mathematical Modelling, 37
(2013), pp. 8930–8945, https://doi.org/10.1016/j.apm.2013.04.025. (Cited on p. 173)

[601] D. WAN , B. PATNAIK , AND G. W EI, A new benchmark quality solution for the buoyancy-driven
cavity by discrete singular convolution, Numerical Heat Transfer, Part B: Fundamentals, 40 (2001),
pp. 199–228, https://doi.org/10.1080/104077901752379620. (Cited on p. 75)

[602] Q. WANG , J. S. H ESTHAVEN , AND D. R AY, Non-intrusive reduced order modeling of unsteady
flows using artificial neural networks with application to a combustion problem, Journal of Com-
putational Physics, 384 (2019), pp. 289–307, https://doi.org/10.1016/j.jcp.2019.01.031.
(Cited on p. 206)

[603] S. WANG , Y. T ENG , AND P. P ERDIKARIS, Understanding and mitigating gradient flow pathologies
in physics-informed neural networks, SIAM Journal on Scientific Computing, 43 (2021), pp. A3055–
A3081, https://doi.org/10.1137/20M1318043. (Cited on p. 401)

[604] S. WANG , H. WANG , AND P. P ERDIKARIS, Learning the solution operator of parametric partial
differential equations with physics-informed DeepOnets, Science Advances, 7 (2021), https://
doi.org/10.1126/sciadv.abi8605. (Cited on p. 401)

[605] W. Q. WANG AND Y. YAN, Strongly coupling of partitioned fluid–solid interaction solvers using
reduced-order models, Applied Mathematical Modelling, 34 (2010), pp. 3817–3830, https://doi.
org/10.1016/j.apm.2010.03.022. (Cited on p. 286)
 458 Bibliography

[606] Y. WANG , A. Q UAINI , AND S. C ANI Ć, A higher-order discontinuous Galerkin/arbitrary La-
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

grangian Eulerian partitioned approach to solving fluid–structure interaction problems with incom-
pressible, viscous fluids and elastic structures, Journal of Scientific Computing, 76 (2018), pp. 481–
520, https://doi.org/10.1007/s10915-017-0629-y. (Cited on p. 286)

[607] Y. WANG , A. Q UAINI , S. C ANI Ć , M. V UKICEVIC , AND S. H. L ITTLE, 3D experimental and

computational analysis of eccentric mitral regurgitant jets in a mock imaging heart chamber, Car-
diovascular Engineering and Technology, 8 (2017), pp. 419–438, https://doi.org/10.1007/
s13239-017-0316-z. (Cited on pp. 283, 373)

[608] Y. WANG , B. Y U , Z. C AO , W. Z OU , AND G. Y U, A comparative study of POD interpolation and

POD projection methods for fast and accurate prediction of heat transfer problems, International
Journal of Heat and Mass Transfer, 55 (2012), pp. 4827–4836, https://doi.org/10.1016/j.
ijheatmasstransfer.2012.04.053. (Cited on p. 173)

[609] K. M. WASHABAUGH , M. J. Z AHR , AND C. FARHAT, On the Use of Discrete Nonlinear Reduced-
Order Models for the Prediction of Steady-State Flows Past Parametrically Deformed Complex Ge-
ometries, in 54th AIAA Aerospace Sciences Meeting, American Institute of Aeronautics and Astro-
nautics, January 2016, https://doi.org/10.2514/6.2016-1814. (Cited on p. 144)

[610] G. W. WASILKOWSKI, Explicit cost bounds of algorithms for multivariate tensor product problems,
Journal of Complexity, 1 (1995), pp. 1–56, https://doi.org/10.1006/jcom.1995.1001. (Cited
on p. 256)

[611] S. W EI, Is human digital twin possible?, Computer Methods and Programs in Biomedicine Update,
1 (2021), p. 100014, https://doi.org/10.1016/j.cmpbup.2021.100014. (Cited on p. 413)

[612] H. G. W ELLER , G. TABOR , H. JASAK , AND C. F UREBY, A tensorial approach to computational

continuum mechanics using object-oriented techniques, Computers in Physics, 12 (1998), pp. 620–
631, https://doi.org/10.1063/1.168744. (Cited on pp. 143, 181, 213, 214, 219, 363)

[613] G. W ELPER, Interpolation of functions with parameter dependent jumps by transformed snapshots,
SIAM Journal on Scientific Computing, 39 (2017), pp. A1225–A1250, https://doi.org/10.
1137/16M1059904. (Cited on p. 298)

[614] G. W ELPER, Transformed snapshot interpolation with high resolution transforms, SIAM Journal on
Scientific Computing, 42 (2020), pp. A2037–A2061, https://doi.org/10.1137/19M126356X.
(Cited on p. 298)

[615] N. W ESTERHOF, J.-W. L ANKHAAR , AND B. W ESTERHOF, The arterial Windkessel, Medical
& Biological Engineering & Computing, 47 (2008), pp. 131–41, https://doi.org/10.1007/
s11517-008-0359-2. (Cited on p. 374)

[616] N. W IENER, The homogeneous chaos, American Journal of Mathematics, 60 (1938), pp. 897–936,
https://doi.org/10.2307/2371268. (Cited on p. 405)

[617] D. C. W ILCOX, Numerical Study of Separated Turbulent Flows, in 7th Fluid and PlasmaDynamics
Conference, American Institute of Aeronautics and Astronautics, June 1974, https://doi.org/
10.2514/6.1974-584. (Cited on p. 168)

[618] D. C. W ILCOX, Reassessment of the scale-determining equation for advanced turbulence mod-
els, AIAA Journal, 26 (1988), pp. 1299–1310, https://doi.org/10.2514/3.10041. (Cited on
p. 168)

[619] K. W ILLCOX, Unsteady flow sensing and estimation via the gappy proper orthogonal decomposi-
tion, Computers & Fluids, 35 (2006), pp. 208–226, https://doi.org/10.1016/j.compfluid.
2004.11.006. (Cited on p. 205)
 Bibliography 459

[620] K. W ILLCOX AND J. P ERAIRE, Balanced model reduction via the proper orthogonal decomposition,
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

AIAA Journal, 40 (2002), pp. 2323–2330, https://doi.org/10.2514/2.1570. (Cited on p. 284)

[621] R. W ILLE AND H. F ERNHOLZ, Report on the first European Mechanics Colloquium, on the
Coanda effect, Journal of Fluid Mechanics, 23 (1965), pp. 801–819, https://doi.org/10.1017/
S0022112065001702. (Cited on p. 372)

[622] C. K. W ILLIAMS AND C. E. R ASMUSSEN, Gaussian Processes for Machine Learning, vol. 2, MIT
Press, 2006. (Cited on pp. 206, 338)

[623] J. W ITTEVEEN AND H. B IJL, Explicit Mesh Deformation Using Inverse Distance Weighting In-
terpolation, in 19th AIAA Computational Fluid Dynamics, AIAA, 2009, https://doi.org/10.
2514/6.2009-3996. (Cited on pp. 206, 351)

[624] K. W U , T. O’L EARY-ROSEBERRY, P. C HEN , AND O. G HATTAS, Derivative-Informed Projected

Neural Network for Large-Scale Bayesian Optimal Experimental Design, arXiv:2201.07925. (Cited
on p. 342)

[625] Y. W U AND X. C. C AI, A fully implicit domain decomposition based ALE framework for three-
dimensional fluid–structure interaction with application in blood flow computation, Journal of Com-
putational Physics, 258 (2014), pp. 524–537, https://doi.org/10.1016/j.jcp.2013.10.046.
(Cited on p. 286)

[626] N. W YCOFF , M. B INOIS , AND S. M. W ILD, Sequential learning of active subspaces, Journal
of Computational and Graphical Statistics, 30 (2021), pp. 1–14, https://doi.org/10.1080/
10618600.2021.1874962. (Cited on p. 336)

[627] D. X IAO , F. FANG , A. B UCHAN , C. PAIN , I. NAVON , J. D U , AND G. H U, Non linear model re-
duction for the Navier–Stokes equations using residual DEIM method, Journal of Computational
Physics, 263 (2014), pp. 1–18, https://doi.org/10.1016/j.jcp.2014.01.011. (Cited on
p. 151)

[628] X. X IE , M. M OHEBUJJAMAN , L. G. R EBHOLZ , AND T. I LIESCU, Data-driven filtered reduced

order modeling of fluid flows, SIAM Journal on Scientific Computing, 40 (2018), pp. B834–B857,
https://doi.org/10.1137/17m1145136. (Cited on p. 187)

[629] L. X U , Y. TANG , X. X U , Y. F ENG , AND Y. G UO, A high order discontinuous Galerkin method
based RANS turbulence framework for OpenFOAM, in Proceedings of the 2017 2nd International
Conference on Communication and Information Systems, ICCIS 2017, New York, 2017, Association
for Computing Machinery, pp. 404–408, https://doi.org/10.1145/3158233.3159368. (Cited
on p. 241)

[630] J. YAN AND C.-W. S HU, Local discontinuous Galerkin methods for partial differential equations
with higher order derivatives, Journal of Scientific Computing, 17 (2002), pp. 27–47, https://
doi.org/10.1023/A:1015132126817. (Cited on p. 236)

[631] M. YANO, A space-time Petrov–Galerkin certified reduced basis method: Application to the Boussi-
nesq equations, SIAM Journal on Scientific Computing, 36 (2014), pp. A232–A266, https:
//doi.org/10.1137/120903300. (Cited on pp. 5, 83, 92)

[632] M. YANO, Discontinuous Galerkin reduced basis empirical quadrature procedure for model re-
duction of parametrized nonlinear conservation laws, Advances in Computational Mathematics, 45
(2019), pp. 2287–2320, https://doi.org/10.1007/s10444-019-09710-z. (Cited on p. 234)

[633] T. YOUNG , D. H AZARIKA , S. P ORIA , AND E. C AMBRIA, Recent trends in deep learning based
natural language processing, IEEE Computational Intelligence Magazine, 13 (2018), pp. 55–75.
(Cited on pp. 389, 402)
 460 Bibliography

[634] Z. Y U, A DLM/FD method for fluid/flexible-body interactions, Journal of Computational Physics,

Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

207 (2005), pp. 1–27, https://doi.org/10.1016/j.jcp.2004.12.026. (Cited on p. 315)

[635] O. Z AHM , P. G. C ONSTANTINE , C. P RIEUR , AND Y. M. M ARZOUK, Gradient-based dimension

reduction of multivariate vector-valued functions, SIAM Journal on Scientific Computing, 42 (2020),
pp. A534–A558, https://doi.org/10.1137/18M1221837. (Cited on p. 327)

[636] O. Z AHM , T. C UI , K. L AW, A. S PANTINI , AND Y. M ARZOUK, Certified dimension reduction in

nonlinear Bayesian inverse problems, arXiv:1807.03712 (2018), (Cited on p. 341)

[637] M. J. Z AHR AND C. FARHAT, Progressive construction of a parametric reduced-order model for
PDE-constrained optimization, International Journal for Numerical Methods in Engineering, 102
(2014), pp. 1111–1135, https://doi.org/10.1002/nme.4770. (Cited on p. 144)

[638] Z. Z AINIB , F. BALLARIN , S. F REMES , P. T RIVERIO , L. J IME ŃEZ -J UAN , AND G. ROZZA, Re-
duced order methods for parametric optimal flow control in coronary bypass grafts, towards patient–
specific data assimilation, International Journal for Numerical Methods in Biomedical Engineering,
37 (2021), p. e3367, https://doi.org/10.1002/cnm.3367. (Cited on pp. 83, 366)

[639] M. Z ANCANARO , M. M ROSEK , G. S TABILE , C. OTHMER , AND G. ROZZA, Hybrid neural network
reduced order modelling for turbulent flows with geometric parameters, Fluids, 6 (2021), p. 296,
https://doi.org/10.3390/fluids6080296. (Cited on pp. 361, 381)

[640] M. M. Z DRAVKOVICH, Flow around Circular Cylinders: Volume 1: Fundamentals, Cambridge

University Press, 1997. (Cited on p. 187)

[641] M. M. Z DRAVKOVICH, Flow around Circular Cylinders: Volume 2: Applications, Oxford Univer-
sity Press, 2003. (Cited on p. 187)

[642] E. Z EIDLER, Nonlinear Functional Analysis and its Applications, vol. I: Fixed-Point Theorems,
Springer-Verlag New York, 1985, https://doi.org/10.1007/BF00047050. (Cited on pp. 99,
104)

[643] G. Z HANG , J. Z HANG , AND J. H INKLE, Learning nonlinear level sets for dimensionality reduc-
tion in function approximation, in Advances in Neural Information Processing Systems Conference,
2019, pp. 13199–13208. (Cited on pp. 327, 330, 337, 384)

[644] R. Z IMMERMANN AND S. G ÖRTZ, Non-linear reduced order models for steady aerodynamics, Pro-
cedia Computer Science, 1 (2010), pp. 165–174, https://doi.org/10.1016/j.procs.2010.
04.019. (Cited on p. 144)

[645] X. Z OU , M. C ONTI , P. D ÍEZ , AND F. AURICCHIO, A nonintrusive proper generalized decompo-

sition scheme with application in biomechanics, International Journal for Numerical Methods in
Engineering, 113 (2018), pp. 230–251, https://doi.org/10.1002/nme.5610. (Cited on p. 222)
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Index

Active manifolds, 337 Discontinuous Galerkin linear elasticity, 318

Active subspaces, 9, 328 method, 234 Navier–Stokes, 24, 120
kernel-based, 10, 329 Distribution nonlinear hyperelasticity, 322
local, 10, 333 Beta, 257 Poisson, 16, 265
multifidelity, 334 uniform, 260 quasigeostrophic, 89
Affine parameter dependency, 3 Domain parametrization, 361 steady Stokes, 21, 127
Aggregated space, 87 Dynamic mode decomposition, von Kármán, 109
Algorithm 207 Error estimation, 44
Genetic, 340 compressed, 210 a posteriori, 31
PISO, 150 multiresolution, 210 effectivity, 45
reduced order SIMPLE, 153 parametric, 217 elliptic problem, 44
SIMPLE, 149 with control, 209 parabolic problem, 52
Artificial neural networks, 390 Stokes equation, 51
reduced approach, 403 Eddy viscosity EZyRB, 375, 382
topologies, 393 interpolation, 189
ASGA, 340 FEniCS, 15
models, 167
ATHENA, 384 Finite elements, 15
Embedded boundary method,
Finite volume, 39
282
Bifurcation, 97, 311 Finite volume–based ROMs,
CutFEM, 276
branchwise reconstruction, 150
shifted boundary, 265
104, 315 Finite–volume method, 143
Empirical interpolation method,
diagrams, 102, 318 Fluid-structure interaction
55, 63, 90, 108, 134
phenomena, 101 problems, 98, 283
Equation
points, 101 Free-form deformation, 11, 347
advection-diffusion, 126,
secondary, 101, 111 FreeFem++, 66, 74, 138
259, 262
Biharmonic operator, 109
BladeX, 385 compressible Navier–Stokes, Gaussian process regression,
Brezzi–Rappaz–Raviart theory, 242 206
64, 72, 99, 131, 135 Gross–Pitaevskii, 112 Geometrical parametrization,
heat, 35, 256 35, 77, 265, 345
Coandă effect, 119, 311, 317, hyperelastic, 116 Gram–Schmidt procedure, 32
370 incompressible Greedy algorithm, 4, 62, 133
Convolutional neural network, Navier–Stokes, 129, 144, actual, 31
394, 407 241, 366, 372, 374 prototype, 31
Coronary artery bypass grafts, incompressible weighted, 253
366 Navier–Stokes with
Cost functional, 84 Poisson pressure Hemodynamics, 365
equation, 152 Hybrid ROMs, 173
Data assimilation, 90 incompressible Hydroacoustics analysis, 214
Data-driven ROMs, 203 Reynolds-averaged
Deep learning, 389 Navier–Stokes (RANS), Inf-sup condition, 6, 86, 100,
Digital twin, 12, 410 165 129, 130, 136

461
 462 Index

Inverse distance weighting, 350 Problem Sensitivity analysis, 340

Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Inverse problems, 341 Cahn–Hilliard, 280 Shape optimization, 212

ITHACA-DG, 241, 382 cavity flow in a duct, 244 Shape parametrization
ITHACA-FV, 381 coercive, 6, 16 CAD surfaces, 357
ITHACA-SEM, 381 convection dominated, 259 discrete surfaces, 356
elliptic, 27 general purpose, 347
Lagrangian functional, 85 flow around a cylinder, 187, object specific, 351
Left ventricular assist device, 244 Software list, 379
374 flow past a backward step, Space-time, 83
Lift and drag forces, 156 181 Spectral element method, 223
Lifting Method, 154 fluid-structure interaction, Stabilization method
312 Brezzi–Pitkäranta, 129
Method parabolic, 41 LPS, 132
Chorin–Temam, 241 steady Stokes, 257, 272 SUPG, 127, 130
continuation, 104 sudden expansion channel, Stochastic
Galerkin projection, 17 317 parameters, 252
Lagrange multipliers, 315 Programming PDE, 252
Newton–Kantorovich, 103 code style, 386 Stokes equations, parametrized,
POD-greedy, 42 documentation, 387 steady, 271
successive constraint, 48 testing, 387 Streamline upwind
Min-θ approach, 48 Programming best practices, Petrov–Galerkin, 259
Model 385 Supremizer enrichment, 39, 62,
Saint Venant-Kirchoff, 322 Proper orthogonal 72
Smagorinsky, 60 decomposition, 4, 29, 87,
Multifidelity regression, 334 165, 240, 268, 367 Theorem
adapted to flows, 272 a priori estimate, 21
Nitsche boundary weak neural network shifted, 399 Brezzi, 85
formulation, 266 weighted, 254 Lax–Milgram, 16
Nonaffine problems, 53 with artificial neural network, Theory
Nonlinear level-set learning, 98, 397 Banach–Nečas–Babuška, 52
337 with interpolation, 204, 374 Thermal-heat exchange model,
PyDMD, 383 parametrized, 267
Offline phase, 50, 86, 151, 188 PyGeM, 384 Trace back, 92
Offline-online decomposition, Turbulence modeling, 165
2, 239 Radial basis functions, 11, 165, large eddy simulations, 165
Online phase, 50, 86, 151, 189 348 Reynolds-averaged
OpenFOAM, 146, 165 RBniCS, 26, 367, 379 Navier–Stokes (RANS),
Optimal control, 83, 98, 261, Reduced forces, 156 165
366 Reduced order models, 15 the closure problem, 166
linear constraint, 84 long time extrapolation, 200 Turbulent ROM, 171
nonlinear constraint, 88 Reduced residual, 44
time dependent constraint, 91 Regression, 338 Uncertainty quantification, 251
Optimality system, 85 Response surface design, 338 Uniqueness, 99
Optimization, 340
shape, 212 Saddle point, 22, 83 Variational multiscale
Sampling strategy, 255 Boussinesq, 69
Parameter space reduction, 219, distributed Monte Carlo, 255 Smagorinsky, 131
327 Monte Carlo, 255
Parametrization techniques, 347 Smolyak quadrature, 255 Weighted ROMs
Partial differential equations, 15 univariate quadrature, 255 greedy algorithm, 253
Physics-informed neural Segregated pressure-based proper orthogonal
network, 401 solvers, 146 decomposition, 254
 GIANLUIGI ROZZA • GIOVANNI STABILE • FRANCESCO BALLARIN
CS&E

Applications in Computational Fluid Dynamics

Advanced Reduced Order Methods and
Downloaded 12/19/23 to 103.172.41.151 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Reduced order modeling is an important, growing field in computational science

and engineering, and this is the first book to address the subject in relation to
computational fluid dynamics. It focuses on complex parametrization of shapes for
Advanced
Reduced Order
their optimization and includes recent developments in advanced topics such as
turbulence, stability of flows, inverse problems, optimization, and flow control, as
well as applications.

This book will be of interest to researchers and graduate students in the field of

Methods and
reduced order modeling.

Gianluigi Rozza is a full professor of Numerical Analysis and Scientific Computing

at SISSA MathLab – International School for Advanced Studies, in Trieste, Italy.
His research interests include numerical analysis, numerical simulation, scientific

Applications in
computing, and reduced order modeling and methods with a special focus on
viscous flows and complex geometrical parametrizations.

Giovanni Stabile is an assistant professor at SISSA MathLab – International

School for Advanced Studies, in Trieste, Italy. His current research interests include

Computational
several aspects concerning the numerical approximation of partial differential
equations with a special emphasis on model order reduction techniques for
computational fluid dynamic problems. He has worked with industrial partners on a
variety of projects aimed at exporting advanced numerical methods to the industrial
world.

Francesco Ballarin is an assistant professor in Numerical Analysis in the

Department of Mathematics and Physics at Catholic University of the Sacred Heart,
Brescia, Italy. His research focuses on reduced order models for parametrized
problems in computational fluid dynamics and applications in engineering and
Fluid Dynamics
life sciences, such as those arising from environmental problems or patient-specific
mathematical modeling of the cardiovascular system. He also develops open source
software libraries.

FRANCESCO BALLARIN
GIOVANNI STABILE
Society for Industrial and Applied Mathematics GIANLUIGI ROZZA
3600 Market Street, 6th Floor
Philadelphia, PA 19104-2688 USA
+1-215-382-9800
[email protected] • www.siam.org

CS27
ISBN: 978-1-61197-724-0
90000

CS27
Computational Science and Engineering
9 781611 977240

CS27_ROZZA_A_B_COMBO_1_V2_416PPI.indd 1 10/13/2022 3:18:46 PM