Data-Driven Fluid Mechanics

Data-driven methods have become an essential part of the methodological portfolio of

fluid dynamicists, motivating students and practitioners to gather practical knowledge
from a diverse range of disciplines. These fields include computer science, statistics,
optimization, signal processing, pattern recognition, nonlinear dynamics, and control.
Fluid mechanics is historically a big data field and offers a fertile ground for develop-
ing and applying data-driven methods, while also providing valuable shortcuts, con-
straints, and interpretations based on its powerful connections to basic physics. Thus,
hybrid approaches that leverage both methods based on data as well as fundamental
principles are the focus of active and exciting research.
Originating from a one-week lecture series course by the von Karman Institute for
Fluid Dynamics, this book presents an overview and a pedagogical treatment of some
of the data-driven and machine learning tools that are leading research advancements
in model-order reduction, system identification, flow control, and data-driven turbu-
lence closures.

Miguel A. Mendez is Assistant Professor at the von Karman Institute for Fluid Dynam-
ics, Belgium. He has extensively used data-driven methods for post-processing numer-
ical and experimental data in fluid dynamics. He developed a novel multi-resolution
extension of POD which has been extensively used in various flow configurations of
industrial interest. His current interests include data-driven modeling and reinforce-
ment learning.

Andrea Ianiro is Associate Professor at Universidad Carlos III de Madrid, Spain. He is

a well-known expert in the field of experimental thermo-fluids. He has pioneered the
use of data-driven modal analysis in heat transfer studies for impinging jets and wall-
bounded flows with heat transfer. He extensively applies these techniques in combina-
tion with advanced measurement techniques such as 3D PIV and IR thermography.

Bernd R. Noack is National Talent Professor at the Harbin Institute of Technology,

China. He has pioneered the automated learning of control laws and reduced-order
models for real-world experiments as well as nonlinear model-based control from
first principles. He is Fellow of the American Physical Society and Mendeley/Web-
of-Science Highly Cited Researcher with about 300 publications including 5 books,
2 US patents and over 120 journal publications.

Steven L. Brunton is Professor at the University of Washington, USA. He has pioneered

the use of machine learning to fluid mechanics in areas ranging from system identi-
fication to flow control. He has an international reputation for his excellent teaching
and communication skills, which have contributed to the dissemination of his research
through textbooks and online lectures.

bli h d li b C b id i i
bli h d li b C b id i i
 Data-Driven Fluid Mechanics
Combining First Principles and Machine Learning
Based on a von Karman Institute Lecture Series

Edited by

MIGUEL A. MENDEZ
von Karman Institute for Fluid Dynamics, Belgium

ANDREA IANIRO
Universidad Carlos III de Madrid

BERND R. NOACK
Harbin Institute of Technology, China

STEVEN L. BRUNTON
University of Washington

bli h d li b C b id i i
 Shaftesbury Road, Cambridge CB2 8EA, United Kingdom
One Liberty Plaza, 20th Floor, New York, NY 10006, USA
477 Williamstown Road, Port Melbourne, VIC 3207, Australia
314–321, 3rd Floor, Plot 3, Splendor Forum, Jasola District Centre, New Delhi – 110025, India
103 Penang Road, #05–06/07, Visioncrest Commercial, Singapore 238467

Cambridge University Press is part of Cambridge University Press & Assessment,

a department of the University of Cambridge.
We share the University’s mission to contribute to society through the pursuit of
education, learning and research at the highest international levels of excellence.

www.cambridge.org
Information on this title: www.cambridge.org/9781108842143
DOI: 10.1017/9781108896214
© Cambridge University Press 2023
This publication is in copyright. Subject to statutory exception and to the provisions
of relevant collective licensing agreements, no reproduction of any part may take
place without the written permission of Cambridge University Press & Assessment.
First published 2023
Printed in the United Kingdom by TJ Books Limited, Padstow Cornwall
A catalogue record for this publication is available from the British Library.
ISBN 978-1-108-84214-3 Hardback
Cambridge University Press & Assessment has no responsibility for the persistence
or accuracy of URLs for external or third-party internet websites referred to in this
publication and does not guarantee that any content on such websites is, or will
remain, accurate or appropriate.

bli h d li b C b id i i
 Contents

List of Contributors page x

Preface xiii
A Note on the Notation xix

Part I Motivation 1

1 Analysis, Modeling, and Control of the Cylinder Wake 3

B. R. Noack, A. Ehlert, C. N. Nayeri, and M. Morzynski
1.1 Introduction 3
1.2 The Cylinder Wake 4
1.3 Features 8
1.4 Low-Dimensional Representations 11
1.5 Dynamic Models 14
1.6 Control 16
1.7 Conclusions 19

2 Coherent Structures in Turbulence: A Data Science Perspective 20

J. Jiménez
2.1 Introduction 20
2.2 Coherent Structures 24
2.3 Discussion 33

3 Machine Learning in Fluids: Pairing Methods with Problems 34

S. Brunton
3.1 Overview 34
3.2 Machine Learning Basics 39
3.3 Flow Feature Extraction 46
3.4 Modeling Flow Dynamics 50
3.5 Control and Optimization 55
3.6 Challenges for Machine Learning in Fluid Dynamics 56

bli h d li b C b id i i
 vi Contents

Part II Methods from Signal Processing 59

4 Continuous and Discrete LTI Systems 61

M. A. Mendez
4.1 On Signals, Systems, Data, and Modeling of Fluid Flows 61
4.2 A Note about Notation and Style 64
4.3 Signal and Orthogonal Bases 65
4.4 Convolutions and Eigenfunctions 69
4.5 Causal and Stable Systems 71
4.6 Linear Time-Invariant Systems in Their Eigenspace 72
4.7 Application I: Harmonic Analysis and Filters 78
4.8 Application II: Time-Series Analysis and Forecasting 84
4.9 What’s Next? 87

5 Time–Frequency Analysis and Wavelets 89

S. Discetti
5.1 Introduction 89
5.2 Windowed Fourier Transform 90
5.3 Wavelet Transform 95
5.4 Tutorials and Examples 107
5.5 Conclusions 113

Part III Data-Driven Decompositions 115

6 The Proper Orthogonal Decomposition 117

S. Dawson
6.1 Introduction 117
6.2 The Singular Value Decomposition 118
6.3 The POD of Discrete Data 122
6.4 The POD of Continuous Systems 124
6.5 The POD for Projection-Based Modeling 126
6.6 Extensions of the POD 127
6.7 Examples 127

7 The Dynamic Mode Decomposition: From Koopman Theory

to Applications 133
P. J. Schmid
7.1 Introduction 133
7.2 The Koopman Idea 135
7.3 The Dynamic Mode Decomposition 138
7.4 Sparsity Promotion for Finding Amplitudes 141
7.5 Applications 144
7.6 Extensions and Generalizations 151
7.7 Summary and Conclusions 152

bli h d li b C b id i i
 Contents vii

8 Generalized and Multiscale Modal Analysis 153

M. A. Mendez
8.1 The Main Theme 153
8.2 The General Architecture 156
8.3 Common Decompositions 163
8.4 The Multiscale POD (mPOD) 171
8.5 Tutorial Test Cases 178
8.6 What’s Next? 181

9 Good Practice and Applications of Data-Driven Modal Analysis 182

A. Ianiro
9.1 Introduction 182
9.2 Data Set Size and Richness 185
9.3 Extracting Phase Information 189
9.4 Extended POD 193

Part IV Dynamical Systems 201

10 Linear Dynamical Systems and Control 203

S. Dawson
10.1 Linear Systems 203
10.2 PID Feedback Control 206
10.3 The Root Locus Plot 210
10.4 Controllability and Observability 211
10.5 Full-State Feedback 212
10.6 Additional Control Techniques and Considerations 214

11 Nonlinear Dynamical Systems 217

S. Brunton
11.1 Introduction 217
11.2 Dynamical Systems 218
11.3 Goals and Challenges in Modern Dynamical Systems 223
11.4 Koopman Theory 224

12 Methods for System Identification 232

S. Brunton
12.1 Balanced Model Reduction 233
12.2 System Identification with Hankel Matrices 244
12.3 Sparse Identification of Nonlinear Dynamics (SINDy) 250

13 Modern Tools for the Stability Analysis of Fluid Flows 255

P. J. Schmid
13.1 Introduction 255
13.2 Definitions of Stability 259

bli h d li b C b id i i
 viii Contents

13.3 General Formulation 262

13.4 Reduced Formulation for Special Cases 268
13.5 Practical Issues and Implementation Details 273
13.6 Tonal Noise Analysis: Quantifying a Complex Feedback Mechanism 280
13.7 Further Extensions 283
13.8 Summary and Conclusions 284

Part V Applications 285

14 Machine Learning for Reduced-Order Modeling 287

B. R. Noack, D. Fernex, and R. Semaan
14.1 Introduction 287
14.2 POD Galerkin Model 288
14.3 Cluster-Based Reduced-Order Modeling 294
14.4 General Principles 296
14.5 Tutorial: xROM 298

15 Advancing Reacting Flow Simulations with Data-Driven Models 304

K. Zdybal, G. D’Alessio, G. Aversano, M. R. Malik, A. Coussement,
J. C. Sutherland, and A. Parente
15.1 Introduction 304
15.2 Combustion Data Sets 306
15.3 Feature Extraction Using Dimensionality Reduction Techniques 309
15.4 Transport of Principal Components 320
15.5 Chemistry Acceleration via Adaptive-Chemistry 325
15.6 Available Software 328
15.7 Summary 328

16 Reduced-Order Modeling for Aerodynamic Applications and

Multidisciplinary Design Optimization 330
S. Görtz, P. Bekemeyer, M. Abu-Zurayk, T. Franz, and M. Ripepi
16.1 Introduction and Motivation 330
16.2 Reduced-Order Modeling 332
16.3 Implementation 334
16.4 Applications 335
16.5 Conclusions 349

17 Machine Learning for Turbulence Control 350

B. R. Noack, G. Y. Cornejo Maceda, and F. Lusseyran
17.1 Introduction 350
17.2 Goals, Tools, and Principles 351
17.3 Model-Based Control 354
17.4 Model-Free Machine Learning Control 359
17.5 MLC Tutorial 362

bli h d li b C b id i i
 Contents ix

18 Deep Reinforcement Learning Applied to Active Flow Control 368

J. Rabault and A. Kuhnle
18.1 The Promise of Deep Reinforcement Learning in Fluid Mechanics 368
18.2 A Brief Introduction to Deep Reinforcement Learning 370
18.3 Recent Applications of Deep Reinforcement Learning to Fluid
Mechanics Problems 378
18.4 A Guide to Deploying DRL in Fluid Mechanics 383
18.5 Perspectives and Remaining Questions 389

Part VI Perspectives 391

19 The Computer as Scientist 393

J. Jiménez
19.1 Introduction 393
19.2 2-D Turbulence: A Case Study 395
19.3 Vortices and Dipoles 403
19.4 Discussion and Conclusions 404

References 409

bli h d li b C b id i i
 Contributors

Mohammad Abu-Zurayk
Institute of Aerodynamics and Flow Technology, German Aerospace Center (DLR),
Germany

Gianmarco Aversano
Université Libre de Bruxelles, Belgium

Philipp Bekemeyer
Institute of Aerodynamics and Flow Technology, German Aerospace Center (DLR),
Germany

Steven L. Brunton
University of Washington, USA

Guy Yoslan Cornejo Maceda

CNRS, LISN, Université Paris Saclay, France

Axel Coussement
Université Libre de Bruxelles, Belgium

Giuseppe D’Alessio
Université Libre de Bruxelles, Belgium
Politecnico di Milano, Italy

Scott Dawson
Illinois Institute of Technology, USA

Stefano Discetti
Universidad Carlos III de Madrid, Spain

Arthur Ehlert
Technische Universität, Berlin, Germany
Industrial Analytics, Berlin, Germany

bli h d li b C b id i i
 List of Contributors xi

Daniel Fernex
École Polytechnique Fédérale de Lausanne, Switzerland

Thomas Franz
Institute of Aerodynamics and Flow Technology, German Aerospace Center (DLR),
Germany

Stefan Görtz
Institute of Aerodynamics and Flow Technology, German Aerospace Center (DLR),
Germany

Andrea Ianiro
Universidad Carlos III de Madrid, Spain

Javier Jiménez
Universidad Politécnica Madrid, Spain

Alex Kuhnle
University of Cambridge, United Kingdom

François Lusseyran
CNRS, LISN, Université Paris Saclay, France

Mohammad R. Malik
Université Libre de Bruxelles, Belgium

Miguel A. Mendez
von Karman Institute for Fluid Dynamics, Belgium

Marek Morzyński
Poznań University of Technology, Poland

Christian N. Nayeri,
Technische Universität Berlin, Germany

Bernd R. Noack,
Harbin Institute of Technology, China
Technische Universität Berlin, Germany

Alessandro Parente
Université Libre de Bruxelles, Belgium

Jean Rabault
Norwegian Meteorological Institute, Norway
University of Oslo, Norway

bli h d li b C b id i i
 xii List of Contributors

Matteo Ripepi
Institute of Aerodynamics and Flow Technology, German Aerospace Center (DLR),
Germany

Peter J. Schmid
Imperial College London, United Kingdom

Richard Semaan
Technische Universität Braunschweig, Germany

James C. Sutherland
University of Utah, USA

Kamila Zdybal
Université Libre de Bruxelles, Belgium

bli h d li b C b id i i
 Preface

This book is for scientists and engineers interested in data-driven and machine learning
methods for fluid mechanics. Big data and machine learning are driving profound
technological progress across nearly every industry, and they are rapidly shaping
research into fluid mechanics. This revolution is driven by the ever-increasing amount
of high-quality data, provided by rapidly improving experimental and numerical
capabilities. Machine learning extracts knowledge from data without the need for first
principles and introduces a new paradigm: use data to discover, rather than validate,
new hypotheses and models. This revolution brings challenges and opportunities.
Data-driven methods are an essential part of the methodological portfolio of fluid
dynamicists, motivating students and practitioners to gather practical knowledge
from a diverse range of disciplines. These fields include computer science, statistics,
optimization, signal processing, pattern recognition, nonlinear dynamics, and control.
Fluid mechanics is historically a big data field and offers a fertile ground to develop
and apply data-driven methods, while also providing valuable shortcuts, constraints,
and interpretations based on its powerful connections to first principles physics.
Thus, hybrid approaches that leverage both data-driven methods and first principles
approaches are the focus of active and exciting research. This book presents an
overview and a pedagogical treatment of some of the data-driven and machine
learning tools that are leading research advancements in model-order reduction,
system identification, flow control, and data-driven turbulence closures.

About the Book and the VKI Lecture Series

This book originated from a one-week course at the von Karman Institute for
Fluid Dynamics, Belgium (VKI; www.vki.ac.be/). The course was hosted by the
Université Libre de Bruxelles, Belgium, from February 24 to 28, 2020, in the classic
VKI lecture series format. These are one-week courses on specialized topics, selected
by the VKI faculty and typically organized 8–12 times per year. These courses have
gained worldwide recognition and are among the most influential and distinguished
European teaching forums, where pioneers in fluid mechanics have been training
young talents for many decades.
The lecture series was co-organized by Miguel A. Mendez from the von Karman
Institute, Alessandro Parente from the Université Libre de Bruxelles, Andrea Ianiro

h d i 0 0 9 8 088962 4 00 bli h d li b C b id i i
 xiv Preface

from the Universidad Carlos III de Madrid (Spain), Bernd R. Noack from the Harbin
Institute of Technology, Shenzhen (China), and TU Berlin (Germany), and Steven L.
Brunton from the University of Washington (USA).

Online Material

The book is supported by supplementary material, including codes, experimental

and numerical data, exercises, and the video lectures recorded from the course. All
material is hosted on the course website:
www.datadrivenfluidmechanics.com/
The supplementary material covers more exercises, tutorials, and practicalities than
could be included in this book while preserving its conciseness. Readers interested in
gaining a working knowledge on the subject are encouraged and expected to download
this material, study it along with the book, and test it on their own data.

The Audience

The book is intended for anyone interested in the use of data-driven methods for fluid
mechanics. We believe that the book provides a unique balance between introductory
material, practical hands-on tutorials, and state-of-the-art research. While keeping
the approach pedagogical, the reader is exposed to topics at the frontiers of fluid
mechanics research. Therefore, the book could be used to complement or support
classes on data-driven science, applied mathematics, scientific computing, and fluid
mechanics, as well as to serve as a reference for engineers and scientists working
in these fields. Basic knowledge of data processing, numerical methods, and fluid
mechanics is assumed.

The Book’s Road Map

Like the course from which it originates, this book results from the contribution of
many authors. The use of machine learning methods in fluid mechanics is in its early
days, and a large team of lecturers allowed the course attendees to learn from the
expertise and perspectives of leading scientists in different fields.
Here we provide a road map of the book to guide the reader through its structure
and link all the chapters into a coherent narrative. The book chapters can be clustered
into six interconnected parts, slightly adapted from the VKI lecture series.
Part I: Motivation. This part includes the first three chapters, which introduce the
motivation for data-driven techniques from three perspectives.
Chapter 1, by B. R. Noack and coauthors, opens with a tour de force on machine
learning tools for dimensionality reduction and flow control. These techniques are

h d i 0 0 9 8 088962 4 00 bli h d li b C b id i i
 Preface xv

introduced to analyze, model, and control the well-known cylinder wake problem,
building confidence and intuition about the challenges and opportunities for machine
learning in fluid mechanics. Chapter 2, by J. Jiménez, takes a step back and gives
both a historical and a data science perspective. Most of the dimensionality reduction
techniques presented in this book have been developed to identify patterns in the data,
known as coherent structures in turbulent flows. But what are coherent structures?
This question is addressed by discussing the relationship between data analysis and
conceptual modeling and the extent to which artificial intelligence can contribute to
these two aspects of the scientific method. Chapter 3, by S. L. Brunton, gives an
overview of how machine learning tools are entering fluid mechanics. This chapter
provides a short introduction to machine learning, its categories (e.g., supervised ver-
sus unsupervised learning), its subfields (regression and classification, dimensionality
reduction and clustering), and the problems in fluid mechanics that can be addressed
by these methods (e.g., feature extraction, turbulence modeling, and flow control).
This chapter contains a broad literature review, highlights the key challenges of the
field, and gives perspectives for the future.
Part II: Methods from Signal Processing. This part brings the reader back
to classic tools from signal processing, usually covered in curricula crossed by
experimental fluid dynamicists, although with a large variety of depth. This part of
the book is motivated by two reasons. First, tools from signal processing are, and
will likely remain, the first “off-the shelf” solutions for many practical problems.
Examples include filtering, time-frequency analysis, and data compression using filter
banks or wavelets, or the use of linear system identification and time series analysis via
autoregressive methods. The second reason – and this is a central theme of the book –
is that much can be gained by combining machine learning tools with methods from
classic signal processing, as later discussed in Chapter 8. Therefore, Chapter 4, by M.
A. Mendez, reviews the theory of linear time-invariant (LTI) systems along with their
properties and the fundamental transforms used in their analysis: the Laplace, Fourier,
and Z transforms. This chapter draws several parallels with more advanced techniques.
For example, the use of the Laplace transform to reduce ordinary differential equations
(ODEs) to algebraic equations parallels the use of Galerkin methods to reduce the
Navier–Stokes equation to a system of ODEs. Similarly, there is a link between
the classical Z-transform and the modern dynamic mode decomposition (DMD).
Chapter 5, by S. Discetti, complements Chapter 4 by focusing on time–frequency
analysis. The fundamental Gabor and continuous/discrete wavelet transforms are
introduced along with the related Heisenberg uncertainty principle and multiresolution
analysis. The methods are illustrated on a time series obtained from hot-wire
anemometry in a turbulent boundary layer and from flow fields obtained via numerical
simulations.
Part III: Data-Driven Decompositions. This part of the book consists of four
chapters dedicated to a cornerstone (and rapidly growing subfield) of fluid mechanics:
modal analysis. This part is mostly concerned with methods for linear dimensionality
reduction, originally introduced to identify, and “objectively” define, coherent struc-
tures in turbulent flows.

h d i 0 0 9 8 088962 4 00 bli h d li b C b id i i
 xvi Preface

Chapter 6, by S. Dawson, is dedicated to the proper orthogonal decomposition

(POD), the first and most popular tool introduced in the fluid mechanics community
in the 1970s. This chapter reviews the link between POD with the singular value
decomposition (SVD), its essential properties (e.g., optimality, relation to eigenvalue
decomposition, and generalization to weighted inner products), its practical computa-
tion on discrete data sets, and its extension to continuous systems. This chapter closes
with illustrative exercises that guide the reader to practical computation. Chapter 7,
by P. Schmid, is dedicated to the dynamic mode decomposition (DMD), a powerful
alternative to POD introduced by P. Schmid a decade ago. This chapter reviews the
derivation of DMD and its roots in dynamical systems and Koopman operator theory.
The main DMD algorithm is presented along with its “sparsity promoting” variant,
and this chapter is enriched by three applications to experimental and numerical data,
as well as a brief outlook at new extensions and generalizations.
Chapter 8, by M. A. Mendez, presents a generalized framework for deriving,
computing, and interpreting any linear decomposition. Modal decompositions are
analyzed in terms of matrix factorization and viewed as a special case of 2D discrete
transforms. This framework is used to combine multiresolution analysis via filter
banks with the classic POD, and derive the multiscale POD (mPOD). The mPOD
is a recent decomposition that generalized the energy-based (POD-like) and the
frequency-based (DMD-like) formalism. This chapter includes several exercises and
tutorials, allowing the reader to test these decompositions on experimental data.
Finally, this part on modal analysis closes with Chapter 9, by A. Ianiro, with an
overview of good practices and applications of modal analysis. This chapter addresses
essential questions on the statistical convergence of POD, the impact of random noise,
and the possibility to extract phase information about the modes even if the data is not
time-resolved. Moreover, this chapter presents interesting applications of the extended
POD – in which decompositions of different data sets are correlated – to experimental
and numerical data.
Part IV: Dynamical Systems. This part of the book consists of four chapters
dedicated to various aspects of dynamical systems. Chapter 10, by S. Dawson, gives a
brief overview of linear dynamical systems and linear control. This is one of the most
developed disciplines in engineering, with applications across robotics, automation,
aeronautics, and mechanical systems in general. Linear techniques provide a standard
approach for closed-loop control and have been successfully used in fluid flows. This
chapter illustrates the main concepts (state-space representation, controllability and
observability, and optimal control) and tools (root locus, pole placement, proportional,
integral and derivative (PID) controllers) focusing on a specific example from fluid
mechanics, namely the stabilization of a wake flow. An overview of additional control
techniques and a brief literature review for flow control are also provided.
Chapter 11, by S. L. Brunton, provides an overview of nonlinear dynamical
systems. This chapter introduces fundamental concepts such as flow maps, attracting
sets, and bifurcations, and gives a modern perspective on the field, with its current
goals and open challenges. These include recent advances in the operator-theoretic
views that seek to identify a linear representation of nonlinear systems and identify

h d i 0 0 9 8 088962 4 00 bli h d li b C b id i i
 Preface xvii

dynamical systems from data, further discussed in Chapter 12. Chapter 12, also by
S. L. Brunton, builds on Chapter 11 and Part III of the book to introduce several
advanced topics in model reduction and system identification. This chapter opens with
a review of balanced model reduction goals for linear systems and builds the required
mathematical background and the fundamentals of balanced POD (BPOD). Linear and
nonlinear identification tools are introduced. Among the linear identification tools,
this chapter presents the eigensystem realization algorithm (ERA) and the observer
Kalman filter identification (OKID). Among the nonlinear identification tools, the
chapter presents the sparse identification of nonlinear dynamics (SINDy) algorithm,
which leverages the LASSO regression from statistics to identify nonlinear systems
from data.
This part closes with Chapter 13, by P. Schmid, providing a modern account
of stability analysis of fluid flows. This chapter begins with a brief review of the
classic definition of stability (e.g., Lyapunov, asymptotic, and exponential stability)
and moves toward a modern formulation of stability as an optimization problem:
unstable modes are those along with which the growth of disturbances is maximized.
This chapter introduces a powerful, adjoint-based, iterative method to solve such an
optimization and shows how to recover common stability and receptivity results from
the general framework. Finally, an illustrative application to the problem of tonal noise
is given.
Part V: Applications. This part of the book is dedicated to the application of data-
driven and machine learning methods to fluid mechanics.
Chapter 14, by B. R. Noack and coworkers, is dedicated to reduced-order modeling.
This chapter gives an overview of the classic POD Galerkin approach, reviewing the
main challenges in closure and stabilization as well as classic applications. It then
moves to emerging cluster-based Markov models and their possible generalization. A
detailed tutorial is also provided to offer the reader hands-on experience with reduced-
order modeling.
Chapter 15, by A. Parente and coworkers, focuses on the use of data-driven
models for studying reacting flows. The numerical simulation of these flows is
extremely challenging because of the vast range of scales involved. This chapter
gives a broad overview of how machine learning techniques can help reduce the
computational burden. The key challenges of high dimensionality are discussed along
with an overview of dimensionality reduction methods, ranging from classic principal
component analysis (PCA) to local PCA, nonnegative matrix factorization (NMF), and
artificial neural network (ANN) autoencoders. The application of these tools to reduce
dimensionality in the modeling of transport and chemical reactions is illustrated in a
challenging test case.
Chapter 16, by S. Görtz and coworkers, is dedicated to the application of reduced-
order modeling for multidisciplinary analysis and design optimization in aerodynam-
ics. The design of an aircraft involves thousands of extremely expensive numerical
simulations. This chapter shows how linear and nonlinear dimensionality reduction
tools can help speed up the process. POD, cluster POD, and Isomaps, combined with

h d i 0 0 9 8 088962 4 00 bli h d li b C b id i i
 xviii Preface

nonlinear regression, are discussed and demonstrated in industrially relevant cases

such as the aero/structure optimization of an entire aircraft.
Chapter 17, by B. R. Noack and coworkers, presents how machine learning
revolutionizes flow control. This chapter gives first an overview of flow control,
its purposes, goals, tools, and strategies. Then, two paradigms for flow control are
introduced and compared. On the one hand, there are model-based approaches, rooted
in first principles and our ability to derive models that predict how a system responds
to inputs. On the other hand, there are model-free approaches rooted in powerful
optimization strategies that can “learn” the best control laws from data, by simply
interacting with the system. Cluster-based control and linear genetic programming are
illustrated, and the chapter closes with a tutorial on an illustrative nonlinear benchmark
problem.
Chapter 18, by J. Rabault and A. Kuhnle complements Chapter 17 with an
overview of deep reinforcement learning (DRL) for active flow control. Reinforcement
learning is one of the three paradigms of machine learning. Contrary to the other two
(supervised and unsupervised learning), a reinforcement learning algorithm starts with
no data and learns through experience, that is, via trial and error. This framework
is meant to tackle decision-making processes, such as teaching a computer to play
chess or drive a car or, as the authors show, to control a fluid flow. This chapter
introduces the main approaches of reinforcement learning (e.g., Q-learning vs. policy
gradient methods), the current research directions, and the recent applications to fluid
mechanics problems. Guidelines to practically deploy DRL are given, along with a
perspective for the future of the field.
Part VI: Perspectives. The book closes with Chapter 19, by J. Jiménez, with a
fascinating perspective and important questions for the field. Combined with Chapter
2, this chapter explores how much the progressive synergy between machine learning
and fluid dynamics, fostered by ever-increasing computational capabilities, could
promote the “automation” of science and ultimately turn machines into colleagues.
This, as masterfully illustrated with a simple case study, ultimately depends on
whether “blind” randomized trials can be integrated in the process of formulating
hypothesis, eventually giving computers the ability to ask questions rather than just
providing answers.

h d i 0 0 9 8 088962 4 00 bli h d li b C b id i i
 A Note on the Notation

The reader will quickly realize that different chapters have (slightly) different notation.
Among these, the same symbol is sometimes used for different purposes, and different
symbols are sometimes used for the same quantities. This choice is deliberate and
motivated by the wide range of intersected disciplines, each of which has well-
established notations. For example, the symbol u usually denotes the actuation in
control theory and the velocity field in fluid mechanics. In reinforcement learning,
the actuation is denoted by at and called the “action” while the sensor measurement
is denoted by st and called the “state” (it is usually denoted by y in control theory).
Resolving these ambiguities would make it difficult for readers to link the material in
this book with the literature of each field.
Moreover, each chapter represents the starting point toward more advanced and
specialized literature, in which a standard notation has not yet been settled. Keeping
the notation as close as possible to the cited literature helps the reader make essential
connections. We hope that the reader will approach each chapter with the required
flexibility, and we welcome comments, corrections, and suggestions to benefit students
for the next reprint.

h d i 0 0 9 8 088962 4 002 bli h d li b C b id i i

h d i 0 0 9 8 088962 4 002 bli h d li b C b id i i
 Part I
Motivation

h d i 0 0 9 8 088962 4 003 bli h d li b C b id i i

h d i 0 0 9 8 088962 4 003 bli h d li b C b id i i
 1 Analysis, Modeling, and Control
of the Cylinder Wake
B. R. Noack, A. Ehlert, C. N. Nayeri, and M. Morzynski

We give a tour de force through select powerful methods of machine learning for
data analysis, dynamic modeling, model-based control, and model-free control. Focus
is placed on a few Swiss army knife methods that have proven capable of solving
a large variety of flow problems. Examples are proximity maps, manifold learning,
proper orthogonal decomposition, clustering, dynamic modeling, and control theory
methods as contrasted with machine learning control (MLC). In Chapters 14 and 17
of this book, the mentioned machine learning approaches are detailed for reduced-
order modeling and for turbulence control. All methods are applied to a classical,
innocent looking benchmark: the oscillatory two-dimensional incompressible wake
behind a circular cylinder at Re = 100 without and with actuation. This example has
the advantage of being visually accessible to interpretation and foreshadows already
key challenges and opportunities with machine learning.

1.1 Introduction

Machine learning, and more generally, artificial intelligence, is increasingly trans-

forming fluid mechanics (Brunton et al. 2020). This change is based on several trends.
First, the efforts from first principles to new theoretical insights have diminishing
returns after hundreds of years of theoretical research. Second, multiphysics mul-
tiscale problems of engineering interest significantly increase in complexity. Third,
fluid mechanics creates increasing amounts of high-quality data from experiments,
for example, participle image velocimetry, to simulations. Finally, the methods of
computer science become increasingly powerful with increasing performance of
computers and with continual discoveries of new algorithms and their improvements.
In this chapter, we focus on machine learning methods for three classical fields
of fluid mechanics: analysis of snapshot data, dynamic reduced-order modeling, and
the control of a given configuration. Following the literature, we take the transient
oscillatory cylinder wake as the most simple, yet nontrivial benchmark. The flow
physics is phenomenologically easy to visualize and can be conceptualized as a
nonlinear oscillator. Yet, an accurate description poses already challenges. Section
1.2 describes the employed configuration and data.

h d i 0 0 9 8 088962 4 004 bli h d li b C b id i i

 4 Data-Driven Fluid Mechanics Chapter 1

Analysis may start with the search of a few features helping to categorize or explain
fluid mechanics. In the case of the cylinder wake, the amplitude and phase are two
such features helping to parameterize drag, lift, and even the flow with good accuracy.
Section 1.3 presents two tools for this purpose. First, the proximity map cartographs
all snapshot data in a two-dimensional plane with classical multidimensional scaling
(CMDS) as the prominent approach. Second, an automated manifold extraction, local
linear embedding (LLE), is presented. Both methods allow us to extract the amplitude
and the phase of vortex shedding. CMDS can be applied to arbitrary even turbulent
data. LLE comes with an estimate of the embedding dimension, if the dynamics is
simple enough.
The analysis may continue with the search of a low-dimensional representation of
the flow data – as described in Section 1.4. Two different approaches are presented.
First, the flow data is represented by a data-driven Galerkin expansion with proper
orthogonal decomposition (POD). POD minimizes the averaged error of the expansion
residual. Second, the flow data may also be represented by a small number of
representative state, called centroids. K-means++ clustering achieves this goal by
minimizing the averaged representation error between the snapshots and their closest
centroids.
The dynamics may be understood by reduced-order models (ROM) building on
such low-dimensional representations. The spectrum of ROM has a bewildering
richness with a myriad of enabling auxiliary methods. Hence, an overview is
postponed until Chapter 14. We focus on simple dynamical models of the cylinder
wake, illustrating the analytical insights that may be gained (see Section 1.5).
The stabilization of the wake is discussed in Section 1.6. This discussion starts with
a classical approach employing ROM for the derivation of the control law. A powerful
model-free alternative is MLC, which learns the control laws in hundreds or thousands
of test runs.
Finally, Section 1.7 summarizes some good practices of analysis, modeling, and
control. The chapter cannot be an exhaustive state-of-the-art compendium of machine
learning approaches. Instead, the presented methods can be seen as the Swiss army
knife of machine learning. They are simple yet powerful and can already be applied
in experimental projects with no or limited availability to first principle equations.

1.2 The Cylinder Wake: A Classical Benchmark

This section describes a classical, innocent looking benchmark of modeling and

control, the two-dimensional oscillatory flow behind a circular cylinder. The first wake
models were proposed over 100 years ago (von Kármán 1911), giving the von Kármán
vortex street its name. The configuration and direct numerical simulation are described
in Section 1.2.1. The transient flow behavior is outlined in Section 1.2.2. A sketch of
the dynamics is previewed in Section 1.2.3.

h d i 0 0 9 8 088962 4 004 bli h d li b C b id i i

 Noack et al: Analysis, Modeling, and Control of the Cylinder Wake 5

1.2.1 Configuration and Direct Numerical Simulation

The two-dimensional viscous, incompressible wake behind a circular cylinder is
computed. This flow is characterized by the Reynolds number Re = UD/ν where
D represents the cylinder diameter, U the oncoming velocity, and ν the kinematic
viscosity of the fluid. The reference Reynolds number is set to Re = 100, which is
significantly above the onset of vortex shedding at Re = 47 (Jackson 1987, Zebib
1987) and also far below the onset of three-dimensional instabilities around Re = 160
(Zhang et al. 1995, Barkley & Henderson 1996).
In the following, all quantities are assumed to be normalized with the cylinder
diameter D, the oncoming velocity U, and the density of the fluid ρ. The two-
dimensional cylinder wake is described by a Cartesian coordinate system (x, y)
with the origin in the cylinder center, the x-axis pointing in streamwise, and the
y-axis in the transverse direction. The incompressibility condition and Navier–Stokes
equations read
∇ · u = 0, (1.1a)
1
∂t u + u · ∇u = −∇p + 4u, (1.1b)
Re
where p represents the pressure, “∂t ” partial differentiation with respect to time, “∇”
the Nabla operator, “4” the Laplace operator, and “·” an inner product or contraction
in tensor algebra.
The rectangular computational domain ΩDNS has a length and width of 50 and 20
diameters, respectively. The cylinder center has a distance of 10 diameter to the front
and lateral sides. Summarizing,
ΩDNS = (x, y) ∈ R 2 : x 2 + y 2 ≥ 1/4 ∩ −10 ≤ x ≤ 40 ∩ |y | ≤ 10 .


On the cylinder, the no-slip condition u = 0 is enforced. At the front x = −10 and
lateral sides of the domain y = ±10 , a uniform oncoming flow u∞ = (1, 0) is assumed.
A vanishing stress condition is employed at the outflow boundary x = 40.
Simulations are performed with a finite-element method on an unstructured grid
with implicit time integration. This solver is third-order accurate in time and second-
order accurate in space. Details about the Navier–Stokes and stability solvers are
described in Morzyński et al. (1999) and Noack et al. (2003). The triangular mesh
consists of 59 112 elements.
The employed initial conditions are based on the unstable steady solution u s and
a small disturbance with the most unstable eigenmode f1 . The steady solution is
computed with a Newton gradient solver. The eigenmode computation is described
in our earlier work (Noack et al. 2003). The disturbance is the real part of the product
of the eigenmode and unit phase factor ejφ . Here, “j” denotes the imaginary unit and φ
the phase. The amplitude ε is chosen to create a perturbation with a fluctuation energy
of 10−4 . The resulting initial condition reads
u(x, t = 0) = u s (x) + εR f1 (x) ejφ . (1.2)


Sixteen initial conditions are considered. These correspond to equidistantly sam-

pled phases φ ∈ [22.5◦, 45◦, . . . , 337.5◦, 360◦ ]. Integration is performed from t = 0

h d i 0 0 9 8 088962 4 004 bli h d li b C b id i i

 6 Data-Driven Fluid Mechanics Chapter 1

to t = 200, capturing the complete transient and post-transient state. The time step is
∆t = 0.1, corresponding to roughly one 50th of the period.

1.2.2 Unforced Transients

In this section, the transients from the steady solution to periodic vortex shedding are
investigated. The flow is analyzed in the observation domain

Ω := {(x, y) ∈ ΩDNS : 5 ≤ x ≤ 15 ∧ 5 ≤ y ≤ 5} . (1.3)

This domain is about twice as long as the vortex bubble of the steady solution. The
streamwise extent is large enough to resolve over one wavelength of the initial vortex
shedding as characterized by the stability eigenmode. A larger domain is not desirable,
because a small increase in wavenumber during the transient will give rise to large
phase differences in the outflow region, complicating the comparison between flow
states. The domain is consistent with earlier investigations by the authors (Gerhard
et al. 2003, Noack et al. 2003) and similar to the domains of other studies (Deane
et al. 1991).
The analysis is based on the inner product of the Hilbert space of square-integrable
functions over the observation domain Ω. This inner product between two velocity
fields v and w is defined by
∫
(v, w)Ω = d x v · w, (1.4)
Ω

where “·” denotes the Euclidean inner product. The corresponding norm of the
velocity field v reads
p
kvkΩ = (v, v)Ω . (1.5)

The flow u is decomposed into a slowly varying base flow u D and an oscillatory
fluctuation u 0,
u = u D + u 0. (1.6)

The short-term averaged flow is approximated as the projection of the flow on the one-
dimensional affine space containing the steady solution u s and the post-transient mean
flow u0 . The superscript “D” comes from the term distorted mean flow of mean-field
theory (Stuart 1958). In other words,

u D (x, t) = u s (x) + a∆ (t) u∆ (x), (1.7)

with the shift-mode u∆ = (u0 − u s ) /ku0 − u s kΩ and amplitude a∆ = (u − u s , u∆ )Ω .

This definition approximates a short-term averaged flow and generalizes the notion in
the stability literature where the steady solution is identified with the base flow.
The shift-mode amplitude a∆ characterizes the mean-flow distortion while the
fluctuation energy
K := ku 0 k 2Ω /2 (1.8)

h d i 0 0 9 8 088962 4 004 bli h d li b C b id i i

 Noack et al: Analysis, Modeling, and Control of the Cylinder Wake 7

K/Kmax 0.5

0
0 25 50 75 100 125 150 175 200
t

Figure 1.1 Evolution of the turbulent kinetic energy K with time t associated with an initial
condition for φ = 22.5◦ . The values are normalized with the maximum value Kmax . Red points
indicate normalized fluctuation levels of 0%, 10%, 50%, and 100%. For details, see Ehlert
et al. (2020).

parameterizes the fluctuation level. We also refer to K as turbulent kinetic energy

(TKE) following the mathematical definition of statistical fluid mechanics, realizing
that the flow is laminar, not turbulent.
Figure 1.1 displays the TKE evolution with time. The maximum TKE value Kmax is
used for normalization. Three dynamic phases can be distinguished. Within the first 30
convective time units D/U, the flow exhibits linear dynamics or exponential growth in
the neighborhood of the steady solution. This exponential growth can clearly be seen
in a logarithmic plot (Noack et al. 2003). In the second, nonlinear transient phase for
50 < t < 100 the flow transitions from the steady solution to the limit cycle with
decreasing growth rate. In the post-transient phase for t > 150, a periodic vortex
shedding or, equivalently, limit-cycle dynamics is observed. The figure marks four
times for TKE levels near 0%, 10% , 50%, and 100%, corresponding to the linear
dynamics phase, the beginning and middle of the nonlinear transient phase and the
limit cycle.
In Figure 1.2, the vorticity for the four selected time instants is shown. Positive
(negative) values of vorticity are shown in red (blue) bounded by solid (dashed) lines.
The three dynamic phases can be distinguished based on the closeness of vortex
shedding to the cylinder and on the formation of pronounced individual vortices.
This discussion provides a basis for the time interval [tmin, tmax ] for snapshot
selection. A lower bound tmin = 40 is chosen. This bound guarantees a TKE below
0.01% or, equivalently 10−4 of the asymptotic maximum value. The upper bound
tmax = 110 includes few periods on the limit cycle.

1.2.3 Wake Dynamics

Figure 1.3 foreshadows a state-space picture of the transient dynamics. Near the steady
solution (bottom of the paraboloid), the flow spirals outward on a plane spanned
by the real and imaginary part of the unstable stability mode (bottom right). With
increasing fluctuation amplitude, the Reynolds stress deforms the short-term averaged
flow in the direction of the shift mode depicted middle left. This deformation leads

h d i 0 0 9 8 088962 4 004 bli h d li b C b id i i

 8 Data-Driven Fluid Mechanics Chapter 1

t = 40

t = 60.6

t = 71.5

t = 158.2

Figure 1.2 Vorticity snapshots corresponding to 0%, 10%, 50%, and 100% fluctuation level
for the simulation displayed in Figure 1.1. The flow is visualized by iso-contours of vorticity
with positive (negative) values marked by solid (dashed) lines and red (blue) background. The
iso-contour levels and color scales are the same for all snapshots. For details, see Ehlert et al.
(2020).

to a shorting of the vortex bubble and an upward motion of the fluctuation energy.
The state moves outward and upward on a paraboloid until it converges against a
limit cycle. The center of this limit cycle is characterized by the mean flow while the
fluctuations are well approximated by the first two POD modes. In Chapters 14 and
17, this dynamics will be distilled from the data, dynamically modeled, and reversed
by stabilizing control.

1.3 Cartographing the Data with Features

In this section, feature extraction is discussed. For the sake of concreteness, fea-
tures are considered for an ensemble of M velocity snapshots u m (x) = u(x, t m ),
m = 1, . . . , M. For an oscillatory flow, amplitude and frequencies are important
features that can completely or, in case of slow drifts, partially characterize the state.
For a turbulent flow, features are far more challenging to design. In case of skin
friction reduction of a turbulent boundary layer, features might be the position and
amplitude of sweeps and ejections. In the following, two feature extraction methods
are presented. First (Section 1.3.1), proximity maps via classical multidimensional

h d i 0 0 9 8 088962 4 004 bli h d li b C b id i i

 Noack et al: Analysis, Modeling, and Control of the Cylinder Wake 9

Figure 1.3 Principal sketch of the wake dynamics. The left side displays the mean flow (top),
shift-mode (middle) and steady solution (bottom). The right side illustrates interpolated vortex
streets on the mean-field paraboloid (middle column). The short-term averaged flows are
depicted also as the streamline plots. Adapted from Morzyński et al. (2007).

scaling (CMDS) can be employed for any data. Second (Section 1.3.2), manifold
extraction with local linear embedding (LLE) is particularly powerful for low-
dimensional dynamics.

1.3.1 Proximity Maps with Multidimensional Scaling

The goal of a proximity map is to cartograph high-dimensional snapshots in a visually
accessible, often two-dimensional, feature space such
 T that neighborhood relations are
preserved as much as possible. Let γ m = γ1m, γ2m ∈ R 2 (the superscript “T” denotes
the transpose) with m = 1, . . . , M the two-dimensional feature vectors corresponding
to the snapshots u m, m = 1, . . . , M. In CMDS (Cox & Cox 2000), these features
minimize the accumulative error of the distances between the snapshots
M Õ
M
[ku m − u n k − kγ m − γ n k]2 .
Õ
E= (1.9)
m=1 n=1

The translational degree of freedom is removed by requesting centered features,

m=1 γ = 0. The rotational degree of freedom is fixed by requiring the first feature
ÍM m

coordinate to have maximum variation. In general, for an N-dimensional feature

space, the sum of first I variances is maximized for all I ∈ {1, . . . , N }. For the
invariance of the error under mirroring, however, there is no cure, as with the sign
indeterminacy of POD modes and amplitudes. In fact, the resulting proximity map
yields the first two POD amplitudes a1 , a2 . The resulting metric may be tailored
to specific applications, for example, identifying regions with similar cost functions
(Kaiser et al. 2017b). Since proximity maps are based on preserving neighborhood
information, it is strongly related to LLE.
In this formulation, the features γ1 and γ2 coincide exactly with the POD mode
amplitudes a1 and a2 discussed in Section 1.4.1. Hence, the proximity map is identical
with the phase portrait of these POD mode coefficients. Numerous generalizations of

h d i 0 0 9 8 088962 4 004 bli h d li b C b id i i

 10 Data-Driven Fluid Mechanics Chapter 1

γ2
−1

−2
−2 −1 0 1 2
γ1

Figure 1.4 LLE of 16 cylinder wake transients. The figures displays the first two embedding
coordinates γ = [γ1, γ2 ]T resulting from K = 15 nearest neighbors. For details, see Ehlert et al.
(2020).

proximity maps have been proposed. In case of control, the error may include also the
cost function to bring similarly performing states closer together (Kaiser et al. 2017).

1.3.2 Identifying the Manifold with Local Linear Embedding

LLE (Roweis & Lawrence 2000) targets to approximate M data points of a typically
high-dimensional space by a low-dimensional manifold. In particular, neighboring
points in the original data space remain neighbors in the low-dimensional embedding
space.
The result is an optimal mapping from the snapshots u m to N-dimensional features
γ ∈ R N . The neighborhood relation is preserved as follows. Let i1m, . . . , iK
m m be the

indices of the K closest snapshots to the mth one. Let

K
Õ m
um ≈ wmk u ik
k=1

be the best approximation of the mth snapshot by its neighbors with optimized
nonnegative weights wmk adding up to unity. These constraints on the weights enforce
a local interpolation. Then, also the feature vector can be approximated by the same
m
expansion, γ m ≈ k=1 wmk γ ik . Here, K is a design parameter. It must be larger than
ÍK
the dimension of the manifold yet sufficiently small for the assumed locally linear
behavior of the manifold. N is increased until convergence of the error is reached.
Then, N denotes the dimension of the manifold. For the details, we refer to the original
literature (Roweis & Lawrence 2000).
Figure 1.4 displays the LLE feature coordinates of the wake snapshot data.
The origin corresponds to the steady solution u s , while the outer circle represents

h d i 0 0 9 8 088962 4 004 bli h d li b C b id i i

 Noack et al: Analysis, Modeling, and Control of the Cylinder Wake 11

post-transient vortex shedding states. An analysis of the polar representation γ1 +jγ2 =

r exp(jθ) reveals that the distance r correlates well with the fluctuation energy, while
θ can be considered the vortex shedding phase (Ehlert et al. 2020).

1.3.3 Other Features

Evidently, many other features can be constructed. An obvious design parameter of
CMDS and LLE is the chosen distance between two snapshots, for example, the
domain and the considered state variables.
Here, we mention one feature vector which is of large relevance to experiments:
time-delay coordinates from sensor signals (Takens 1981). While reduced-order
representations of fluid flows reduce the dimension of the state, time-delay coordinates
increase the dimension of the measured signal to a level where the trajectories do
not cross each other (no false neighbors) and a dynamical system can be identified.
Loiseau et al. (2018) discuss the construction of velocity field manifold for the
transient cylinder wake from the lift coefficient.

1.4 Low-Dimensional Representations

Low-dimensional state representations are key enablers for understanding, state

estimation, dynamic modeling, and control. The starting point of this section is the
ensemble of M flow snapshots u m (x), m = 1, . . . , M (Section 1.2). A low-dimensional
data-driven representation is synonymous with an autoencoder in machine learning.
An autoencoder targets a low-dimensional parameterization of the snapshot data, say
in R N . More precisely, an autoencoder comprises an encoder G from the high- or
infinite-dimensional state-space to a low-dimensional feature space, for example,

u m 7→ a m := G (u m ) ∈ R N , m = 1, . . . , M (1.10)

and a decoder or state estimator H, for example,

a m 7→ û m := H (a m ) , m = 1, . . . , M (1.11)

for the reconstruction of the state. Ideally, the autoencoder identifies the best possible
pair of encoder G and decoder H that minimizes the in-sample error of the estima-
tor/decoder
M
1 Õ m
Ein := k û − u m k 2Ω . (1.12)
M m=1

1.4.1 Optimal Subspaces with Proper Orthogonal Decomposition

POD can be considered as an optimal linear autoencoder onto an N-dimensional affine
subspace. Let u0 be the average of the snapshot ensemble, ui , i = 1, . . . , N be the N

h d i 0 0 9 8 088962 4 004 bli h d li b C b id i i

 12 Data-Driven Fluid Mechanics Chapter 1

POD modes, and ai be the corresponding mode coefficients. Then the encoder G of a
velocity field u to the mode amplitudes a = [a1, . . . , a N ]T is defined by
ai := (u − u0, ui )Ω , i = 1, . . . , N, (1.13)
while the decoder H reads
M
Õ
û(x) = u0 (x) + ai ui (x). (1.14)
i=1

POD modes are an orthonormal set and sorted by energy content. The optimality
condition (e.g., Holmes et al. 2012) implies a minimal in-sample representation error
from (1.12). We cannot find another Nth order Galerkin expansion (more precisely, a
different N-dimensional subspace) yielding a smaller error.
For the transient wake data, the most energetic POD modes ui and the amplitude
evolution ai are displayed in Fig. 3 and 4 of Noack et al. (2016). The first two modes
correspond to von Kármán vortex shedding. The third mode resolves the change of
the mean field. The following modes mix different frequencies and wavelengths.

1.4.2 Coarse-Graining the Data into Bins with Clustering

The key idea of clustering is representing the snapshots by a small number, say K, of
centroids ck with k = 1, . . . , K. Every snapshot u m can be associated with its closest
centroid ck . Thus, the encoder G maps the velocity field u to k ∈ {1, . . . , K }, the index
of the closest centroid. In other words, the encoder creates “bins” of similar snapshots.
The decoder H approximates the velocity field by the closest centroid û = ck . The
k-means algorithm aims to minimize the in-sample representation error (Arthur &
Vassilvitskii 2007). For generic data, the centroids can be expected to be unique
modulo the index numbering. Clustering with K bins cannot yield a lower in-sample
error than a POD representation with K modes ui , i = 0, . . . , K − 1. Both clustering
and POD span a K − 1-dimensional subspace, but a POD expansion can interpolate
states while the centroids are fixed representations of the corresponding snapshot bin.
Figure 1.5 displays the results for 10 centroids visualized with LLE features. The
centroids attempt to fill the circular region. One centroid is the steady solution; seven
centroids resolve the limit cycle; and a phase opposite pair of centroids represent low-
amplitude oscillations.

1.4.3 Comparison and Discussion

POD and clustering are quite different autoencoders. While POD is based on a super-
position of modes, centroids are representative states that cannot be superimposed.
LLE can be generalized to an autoencoder. The feature vector γ is obtained in
the encoder step. A decoder can easily be constructed, for instance, with K-nearest
neighbor interpolation (Ehlert et al. 2020).
Figure 1.6 displays the reconstruction error of the three methods for the simulation
data. All three methods have the largest reconstructing error in the transient phase

h d i 0 0 9 8 088962 4 004 bli h d li b C b id i i

 Noack et al: Analysis, Modeling, and Control of the Cylinder Wake 13

γ2
−1

−2
−2 −1 0 1 2
γ1

Figure 1.5 Cluster centroids localized in the LLE-based feature space. One centroid represents
the steady-state solution; two resolve the opposite transient phases; and the remaining eight
centroids are close to the limit cycle. For details, see Ehlert et al. (2020).

10 clusters 2 LLE coordinates 10 POD modes

102
Eout

10−1

10−4
40 60 80 100 120 140 160
t

Figure 1.6 Out-of-sample error Eout for a new simulation trajectory at Re = 100. The solid
line corresponds to LLE representations. The red dash-dotted curve and blue dashed curve
refer to approximations with 10 centroids and 10 POD modes, respectively. For details, see
Ehlert et al. (2020).

between t = 60 and t = 80. LLE significantly outperforms both POD and clustering
by up to three orders of magnitudes, highlighting the two-dimensional manifold of
the Navier–Stokes dynamics and a niche application of LLE. As expected, clustering
performs worst lacking any intrinsic interpolation. The low error of the LLE-based
autoencoder demonstrates that the dynamics is effectively two-dimensional. Yet, about
50 POD modes are necessary for a similar resolution as corroborated by Loiseau et al.
(2018) for a similar manifold approximation. Evidently, POD-based representations
are not efficient for slowly changing oscillatory coherent structures.
Data-driven Galerkin expansions have been optimized for a myriad of purposes.
Dynamic mode decomposition (DMD) (Rowley et al. 2009, Schmid 2010) can extract

h d i 0 0 9 8 088962 4 004 bli h d li b C b id i i

 14 Data-Driven Fluid Mechanics Chapter 1

stability modes from initial transients and Fourier modes from converged post-
transient data. However, the performance for transient wakes is disappointing while
a recursive DMD can keep some advantages of POD and DMD (Noack et al. 2016).
Flexible state-dependent modes may significantly improve the accuracy of a low-order
representation (Siegel et al. 2008, Tadmor et al. 2011, Babaee & Sapsis 2016).

1.5 Dynamic Models: From Proper Orthogonal

Decomposition to Manifolds

In this section, a path to a least-order model for the transient cylinder wake is pursued.
The starting point is the POD Galerkin method (Section 1.5.1). Then (Section
1.5.2), mean-field theory is employed to significantly simplify the dynamics. The
simplification culminates in a manifold model with the Landau equation as dynamics
(Section 1.5.3).

1.5.1 Proper Orthogonal Decomposition Galerkin Method

The traditional Galerkin method (Fletcher 1984) derives the dynamics of the Galerkin
approximation (1.14) with orthonormal modes from the Navier–Stokes equations
(1.1). Under weak conditions, a constant-linear-quadratic system of ordinary differ-
ential equations is obtained
dai Õ Õ
= ci + li j a j + qi jk a j ak . (1.15)
dt j=1,..., N j,k=1,..., N

These conditions may include the incompressibility of the flow, a stationary domain,
stationary Dirichlet, periodic, or von Neumann boundary conditions, and smoothness
of the flow. The coefficients ci , li j , and qi jk are functions of the modes and of the
Reynolds number. The coefficients may also be identified from numerical solutions or
experimental data (Galletti et al. 2004, Cordier et al. 2013), for instance, if the flow
domain is too small or if the turbulent fluctuations are not resolved in (1.14). We refer
to exquisite textbooks (Fletcher 1984, Holmes et al. 2012) for details. Schlegel and
Noack (2015) have derived necessary and sufficient conditions for bounded solutions,
which can be considered a requirement for a physical sound model.

1.5.2 Mean-Field Model

The Galerkin system (1.15) can be significantly simplified exploiting the manifold
dynamics depicted in Figure 1.3 and derived in Section 1.5.1. The flow is dominated
by a zeroth and first harmonics component (see (1.6)).
The zeroth component is on the line from the steady solution u s to the mean flow u.
This line approximates the one-period averaged flow u D and is parameterized by the
shift mode u3 and its amplitude a3 (see (1.7)). The shift mode is effectively a backflow
in the wake. The streamlines look like a fly and are depicted in Figure 1.3 (left middle
subfigure).

h d i 0 0 9 8 088962 4 004 bli h d li b C b id i i

 Noack et al: Analysis, Modeling, and Control of the Cylinder Wake 15

The first harmonics represents von Kármán vortex shedding, which may be resolved
by a cosine u1 and sine u2 mode, ignoring the shape deformation for a moment:
u 0(x, t) = a1 (t) u1 (x) + a2 (t) u2 (x). (1.16)
The modes may be inferred from Figure 1.3 in the rightmost column. In the following,
u1 , u2 , and u3 are assumed to be orthonormalized. The first three POD modes of the
transient yield are already a good approximation of these modes.
Following mean-field arguments (Noack et al. 2003), the Galerkin system (1.15)
simplifies to a self-amplified, amplitude limited oscillator,
da1 /dt = σa1 − ωa2, σ = σ1 − βa3, (1.17a)
da2 /dt = σa2 + ωa1, ω = ω1 + γa3, (1.17b)
 
da3 /dt = σ3 a3 + α a1 + a22 .
2
(1.17c)
The oscillator has three parameters σ1 , ω1 , σ3 for linear dynamics and three
parameters for α, β, γ for the weakly nonlinear effects of Reynolds stress. Intriguingly,
sparse identification of nonlinear dynamics (SINDy) extracts precisely this sparse
dynamical system from transient simulation data (Brunton, Proctor & Kutz 2016a).
a

a
a

Figure 1.7 Transient dynamics of the cylinder wake from the DNS (solid line) and the
mean-field Galerkin system (1.17) (dashed line). Here, a∆ = a3 . Phase portrait from Tadmor
and Noack (2004).

Figure 1.7 shows that the mean-field Galerkin system (1.17) and the direct numeri-
cal simulation agree well. A detailed analysis (Tadmor & Noack 2004) quantitatively
corroborates this agreement for the manifold and the temporal evolution.

1.5.3 Manifold Model

The mean-field Galerkin system can be further simplified exploiting the slaving of the
shift mode to the fluctuation level. In fact, the calibrated σ3 ≈ −6σ1 shows a much
quicker convergence to manifold than the growth of the fluctuation. Hence, (1.17c) is
well approximated by the algebraic equation
α3  2 
a3 = a1 + a22 , (1.18)
|σ3 |

h d i 0 0 9 8 088962 4 004 bli h d li b C b id i i

 16 Data-Driven Fluid Mechanics Chapter 1

explaining the mean-field parabola shape in the figure. Equation (1.18) in (1.17a),
(1.17b) and the introduction to polar coordinates a1 = r cos θ, a2 = r sin θ lead to the
famous Landau equation (Landau & Lifshitz 1987) for a supercritical instability,

dr/dt = σ1 r − β 0r 3, dθ/dt = ω1 + γ 0r 2 . (1.19)

These equations show an exponential growth by linear instability and a cubic damping
by Reynolds stress and mean-field deformation. The frequency changes as well. The
nonlinearity parameters β 0 and γ 0 can easily be derived from (1.17). Intriguingly, this
equation is found to remain accurate even far from the onset of vortex shedding.
The resulting Landau model does not include higher harmonics. Even worse, we
have assumed fixed modes u1 , u2 . Yet, the prominent stability mode near the steady
solution is distinctly different from the POD modes characterizing the limit cycle. a3 -
dependent modes can cure this shortcoming (Morzyński et al. 2006), but the model is
still blind to higher harmonics. A more accurate model is based on the LLE feature
coordinates γ1 = r cos θ and γ2 = r sin θ and an identified Landau equation for the
dynamics of r and θ. The LLE autoencoder incorporates the mode deformation and
higher harmonics (Ehlert et al. 2020). Loiseau et al. (2018) has derived such a model
based on similar premises. The mean-field Galerkin model can be generalized for two
(and more) frequencies (Luchtenburg et al. 2009).

1.6 Control: Model-Based and Model-Free Approaches

This section previews two flow control approaches. Section 1.6.1 follows the classical
paradigm of model-based control design, while Section 1.6.2 outlines a model-free
machine-learned control optimization.

1.6.1 Model-Based Control

As a control benchmark (Hinze & Kunisch 2000), we aim to stabilize the cylinder
wake with a transversal volume force in the near wake (see Figure 1.8). This
admittedly academic volume force actuator is surprisingly often used in the compu-
tational flow control literature and significantly simplifies the discussion. As added
complexity, experimental conditions are emulated by the hot-wire measurement of the
streamwise velocity component s. The goal is stabilizing sensor-based control law
(see, again, Figure 1.8).
The volume force can be shown to lead to an extra term gb in the mean-field system
(Gerhard et al. 2003),

da1 /dt = σa1 − ωa2 + gb, σ = σ1 − βa3, (1.20a)

da2 /dt = σa2 + ωa1, ω = ω1 + γa3, (1.20b)
 
a3 = α 0 a12 + a22 , α = α/|σ3 |.
0
(1.20c)

h d i 0 0 9 8 088962 4 004 bli h d li b C b id i i

 Noack et al: Analysis, Modeling, and Control of the Cylinder Wake 17

Volume force Hot-wire

y
−2
−2 0 2 4 6 8
x

Figure 1.8 Cylinder wake configuration with volume force actuation and hot-wire sensor.

In this differential algebraic system, a3 is slaved to the fluctuation energy (1.18). Here,
b is the actuation command, for example, the induced acceleration in the circular
region, while the positive gain g quantifies the effect on the dynamics and depends,
for instance, on the size and location of the volume force support. Without loss of
generality, the actuation term only effects a1 by suitable rotation of the modes u1 , u2 .
The forced growth rate of the fluctuation energy K = r 2 /2 = (a12 + a22 )/2 reads
dK da1 da2
= a1 + a2 = σr 2 + g b a1 . (1.21)
dt dt dt
The fluctuation energy can be mitigated with negative actuation power g b a1 , that is,
b has to have the opposite sign of a1 . For simplicity, a linear control law is assumed,

b = −ka1, k > 0. (1.22)

The control gain k shall ensure a forced exponential decay rate σc < 0 of the
amplitude r. This implies with (1.22) and (1.21),
dK
= σc r 2 = σr 2 + g k a12 . (1.23)
dt

Averaging over one period and exploiting a12 = r 2 /2 yields the control gain k and thus
the control law
σc − σ
b=2 a1 . (1.24)
g
The implications of this law are plausible: the higher the unforced growth rate and
the higher the desired damping, the larger the volume force amplitude must become.
Contrary, the larger the gain g of actuation on the dynamics, the smaller the volume
force needs to be. It should be born in mind that σ is r dependent.
For the sensor-based control, the state a = [a1, a2 ]T is estimated from the sensor
signal with a dynamic observer. The resulting stabilization effect is shown in Figure
1.9. While the model allows complete stabilization, the fluctuation energy of the
DNS has been reduced by 30% in the observation region x < 6. The model is only
partially accurate as it ignores shedding mode changes due to actuation and convection

h d i 0 0 9 8 088962 4 004 bli h d li b C b id i i

 18 Data-Driven Fluid Mechanics Chapter 1

(a) (b)

Figure 1.9 Model-based cylinder wake stabilization. (a) unforced and (b) controlled flow.
Here, the vorticity field is shown: red and blue mark negative and positive vorticity, while
green corresponds to potential flow.

effects. The maximum reduction of the fluctuation level is 60%, leading to a complete
stabilization of the near wake and a residual fluctuation in the far wake.

1.6.2 Model-Free Machine Learning Control

In Section 1.6.1, the inaccuracy of the model has led to a reduced control performance.
In model-free control, such model-based errors are avoided. Instead, a regression
problem is solved: find a control law b = K(s) that minimizes the cost function
J, for example, the fluctuation energy. Any regression solver requires repeated tests
of control laws in the full plant for the optimization problem. A linear ansatz for
the control law allows for a gradient-based approach, for example, the downhill
simplex method, for parameter optimization. If no structure of the control law shall
be assumed, for example, in case of strongly nonlinear dynamics, MLC has proven
to be very powerful. MLC is based on genetic programming as regression solver and
has optimized the nonlinear control law in dozens of experiments and simulations
(Noack 2019). The observed testing requires hundreds to one thousand control laws
from simple single-input single-output (SISO) to complex nonlinear multiple-input
multiple-output (MIMO) control.

(a) (b)

Figure 1.10 Machine learning control of the fluidic pinball with 3 cylinder rotations
responding to 15 undisplayed downstream sensors. Contour levels of vorticity for the unforced
flow (a) and controlled flow (b) by the control law. Solid lines and dashed lines represent
respectively positive and negative vorticity. For details, see Cornejo Maceda et al. (2019).

Figure 1.10 provides a synopsis for MLC applied to the fluidic pinball configuration
(Cornejo Maceda et al., (2019), (2021)). Forty-two percent reduction of the effective
drag power is achieved accounting for actuation energy. The enabling MIMO feedback

h d i 0 0 9 8 088962 4 004 bli h d li b C b id i i

 Noack et al: Analysis, Modeling, and Control of the Cylinder Wake 19

control has 15 downstream sensors (not displayed) and commands the rotation of the
three cylinders. The mechanism is a combination of open-loop Coanda forcing and
closed-loop phasor control (as in the cylinder wake example).

1.7 Conclusions

This chapter provides examples of machine learning applications in fluid mechanics.

The examples are from analysis, modeling, and control of the oscillatory cylinder
wake. All applications are based on regression problems, that is, finding a function
that minimizes a cost. Feature extraction leads to a mapping from the flow to
few coordinates optimizing some neighborhood/distance criteria. An autoencoder
comprises an encoder and decoder for a low-dimensional flow representation that
minimizes the representation error for the snapshot data. Dynamic modeling relies
on a mapping from the state to the change of state. State estimation is a function from
the measured sensor signals to the flow field. Control design leads to a function from
the state or sensor signals to the actuation command minimizing a given cost. Closures
can be seen as control terms that minimize the tracking error between computed and
observed states. More examples can be added (Brunton & Noack 2015).
Machine learning provides powerful tools for regression solvers based on existing
data (“curve fitting”) or based on in situ optimization (“variational problems”). The
discussed data-based regressions – classical multidimensional scaling, local linear
embedding, proper orthogonal decomposition, clustering – can be used for a wide
range of problems. Genetic programming is a powerful tool for solving variational
problem, control design or closures, without required data but with in situ testing.
The oscillatory cylinder wake and its stabilization looks like an innocently simple
benchmark. Yet, it shows already that machine learning methods need to be carefully
chosen. For instance, local linear embedding does a remarkable job of compressing
the data to a two-dimensional manifold. In contrast, 50 POD modes are required
for a similar representation error making understanding, state estimation, dynamic
modeling, and model-based control next to impossible. We will elaborate on machine
learning methods and more advanced applications for reduced-order modeling (Chap-
ter 14) and for control (Chapter 17). We highly recommend three introductory books
for machine learning: Abu-Mostafa et al. (2012) as a deep introduction to how one
can learn from data, Wahde (2008) for optimization solvers, and Burkov (2019) for an
inspiring overview of machine learning methods.

h d i 0 0 9 8 088962 4 004 bli h d li b C b id i i

 2 Coherent Structures in Turbulence:
A Data Science Perspective
J. Jiménez∗

This chapter and Chapter 19 explore how far the scientific discovery process can
be automated. After discussing the scientific method and its possible relation with
artificial intelligence techniques, this first chapter deals with how to extract correla-
tions and models from data, using examples of how computers have made possible
the identification of structures in shear flows. It is concluded that, besides the data-
processing techniques required for the extraction of useful correlations, the most
important part of the discovery process remains the generation of rules and models.

2.1 Introduction

This chapter is intended to be used together with Chapter 19 of this book. Both deal
with the problem of how much the process of scientific discovery can be automated
and, in this sense, address the core question of the use of artificial intelligence (AI)
in fluid mechanics, or in science in general. The two chapters are, however, very
different. The present one discusses techniques and results that have been current in
fluid mechanics for the past few decades. Specifically in turbulence research. They
owe a lot to the growth of computers, and have changed the field to an extent that
would have been hard to predict 40 years ago. But, during this period, computers were
basically used as research tools, and the way research was done was not very different
from the way it had always been.
Chapter 19 deals with one way in which the role of computers in research could
develop in the near future, and will argue that the scientific method is about to
undergo changes by which computers may evolve from being used as research tools to
being considered minor research “participants.” Since that chapter discusses nascent
technology, closer to AI than what has been used up to now in fluid mechanics, the
scientific results in it will necessarily be more limited than those in the present chapter.
But the methods will be more interesting for being new. In this sense, the present
chapter deals with the past, and Chapter 19 deals with a possible future.
We begin by specifying what science and the scientific method means for us in
these two chapters. There are two meanings to the word “science.” The first one is a
* This work was supported by the European Research Council under the Coturb grant
ERC-2014.AdG-669505.

h d i 0 0 9 8 088962 4 00 bli h d li b C b id i i
 Jiménez: Coherent Structures in Turbulence 21

systematic method for discovering how things work, and the second is the resulting
body of knowledge about how they work. Depending on who you ask, one aspect
is considered to be more important than the other. Empiricists tend to focus on the
scientific method (Poincaré 1920), and practitioners tend to be more interested in the
results (Kuhn 1970). These two chapters deal mostly with methods while illustrating
them with examples from real applications to fluid mechanics.
We take the view that the basic goal of the scientific method is to search for rules
that can be used to make predictions. But the second important goal is to make those
rules “beautiful” for the researcher (Jho 2018). This is not an arbitrary requirement
because, at least for now, science depends on the work of scientists, and people work
better when they like what they are working with. The logarithmic function can be
defined by a computer algorithm to evaluate it, but most mathematicians would be
more likely to use logarithms if they are defined as
d log(t)/dt = 1/t.
The well-known conundrum about why mathematics is useful to the physical sciences,
even when it is primarily developed by mathematicians for its aesthetic value (Wigner
1960), may have more to do with the motivation of mathematicians and physicists
than with mathematics.
The classical point of view is that a prerequisite for science is the determination
of causes (Does A cause B?), preferably complemented by mechanisms (How does
A cause B?). Causality without mechanisms is unlikely to lead to quantitative
predictions, and mechanisms without causality have a high probability of being wrong.
Both causes and mechanisms have to be testable, preferably on data sets different from
the ones used to train them. The ancient Greeks knew about the alternation of night
and day, and imagined causes and mechanisms involving gods and chariots. They
explained the facts known to them, but the Greeks did not have data from other planets,
or even from other latitudes, and we know today that their mechanisms are hard to
generalize. One possible characterization of these two chapters is that they explore
how to determine causality in physical systems, and whether recent developments
provide us with new tools for doing so.
There are several distinctions that need to be made before we begin our argument.
We have mentioned the importance of causes, but research is not always geared toward
them. The classical description of the scientific method is summarized in Figure 2.1,
which emphasized its iterative character. Causality is encapsulated in the modeling
and testing steps, S2 and S3, and the emphasis is often put on these two steps rather
than on the data-gathering one, S1. In fact, the loop in Figure 2.1 often starts with step
S2, and only later are hypotheses tested against observations, as in S2–S1–S3. The
implied relation is not always causal. Even in data-driven research, where observations
precede hypotheses, the argument is often that “if A precedes B,” A is likely to be the
cause of B, although it is generally understood that correlation and causation are not
equivalent. A classical example is the observation of night and day mentioned earlier.
The correlation between earlier days and later nights is perfect, but it does not imply
causation (Mackie 1974).

h d i 0 0 9 8 088962 4 00 bli h d li b C b id i i
 22 Data-Driven Fluid Mechanics Chapter 2

Figure 2.1 The scientific method.

Moreover, the concept of causality is not without problems, and it has been argued
that it is indistinguishable from initial conditions in systems described by differential
equations, such as fluid mechanics (Russell 1912). While this is true, and we could
formulate science as a quest to classify initial conditions in terms of their outcome, the
result may not be very informative, especially in turbulence and other chaotic systems,
which lose memory of their previous evolution after a while. We will restrict ourselves
here to a prescribed-time definition of cause, of the type of: “The falling tree causes
the crashing noise,” even if we know that the fall of the tree has its own reasons for
happening, and that intermediate causes depend on the time horizon that we impose
on them. Such investigations require something beyond correlations, and imply an
active intervention of the observer in the generation of data. This is particularly true
when we are interested in the possibility of controlling our system, instead of simply
describing it. Consider the difference between formulating sub-grid models for large-
eddy simulations, and devising strategies to decrease drag in a pipe.
Simplifying a lot, the difference between the two points of view, which mostly
affects how and when step S1 in Figure 2.1 is undertaken, can be summarized as
follows:

• Observational science. Searches for correlations between different observations,

and generalizes them into models. It can be described as the science of prediction.
• Interventional science. Relies on the results of experiments in which some condition
is changed by the observer, generally in the hope of uncovering “causal” (i.e.,
if this, then that), or “counterfactual” relations (if not this, then not that).
Interventional science is the science of control.

The present chapter is oriented toward predictions, and therefore toward establish-
ing correlations among observations, without being especially interested in causality.
Chapter 19 is oriented toward causality and control, and looks very different from the
present one. In the aforementioned example, Chapter 19 will want to know about the
falling tree because it might be interested in doing something to prevent it from falling.
The present chapter will center (although not completely) on establishing correlations
between tree health and noise levels.

h d i 0 0 9 8 088962 4 00 bli h d li b C b id i i
 Jiménez: Coherent Structures in Turbulence 23

There is a second distinction that, although related to the previous one, is indepen-
dent from it. It has do with how AI, or intelligence in general, is supposed to work
(Nilsson 1998), and is traditionally divided into symbolic (Newell & Simon 1976) and
sub-symbolic (Brooks 1990). Symbolic AI is the classical kind, which manipulates
symbols representing real-world variables according to rules set by the programmer,
typically embodied into an “expert system.” For example, assuming that we agree
that “vortices are localized pressure minima” (or any favorite personal definition),
symbolic AI encapsulates this knowledge into a set of rules to identify and isolate
the vortices. On the contrary, sub-symbolic AI is not interested in rules, but in
algorithms to do things. For example, given enough snapshots of a flow in which
vortices have been identified (maybe manually by the researcher), we can train a
neural network to distinguish them from vortex sheets or from other flow features.
The interesting result of sub-symbolic AI is not the rule, but the algorithm, and it
does not imply that a rule exists. After training our system, we may not know (or
care) what a vortex is, or what distinguishes it from a vorticity sheet, but we may
have a faster way of distinguishing one from the other than would have been possible
using only preordained physics-based rules. Copernicus and the Greeks had symbolic
representations of the day-night cycle, although with very different ideas of the rules
involved. Most other living beings, which can distinguish night from day and usually
predict quite accurately dawn and dusk, are (probably) sub-symbolic. The classical
scientific method, including causality, is firmly symbolic: we are not only interested
in the result, but also in the rule. But there is a small but growing body of scientists and
engineers who feel that data and a properly trained algorithm are all the information
required about Nature, and that no further rule is necessary (Coveney et al. 2016, Succi
& Coveney 2018). Observational science is sub-symbolic at heart, although we will
argue that this point of view is incomplete even for it. Interventional science at least
aims for symbolism.
In fluid mechanics, the availability of enough data to even consider sub-symbolic
science is a recent development, made possible by the numerical simulations of the
1990s, which gave us for the first time the feeling that “we knew everything,” and
that any question that could be posed to a computer would eventually be answered
(Brenner et al. 2019, Jiménez 2020a). The present chapter is partly a review of this
work. Of course, even without considering practical problems of cost, this did not
mean that all questions were answered, because they first had to be put to the computer
by a researcher. But this discussion has to be postponed to Chapter 19.
Some readers could be excused for wondering whether these two chapters,
and especially the present one, deal with AI at all. They may be right. Artificial
intelligence, as the term is mostly used at the moment, is a collection of techniques
for the analysis of large quantities of data. In terms of the scientific method, this
corresponds to the generation of empirical knowledge, usually from observations, but
this is only a small part of the overall method (distributed among steps S1 and S2 in
Figure 2.1). What interests us here is the full process of data exploitation, from their
generation to their final incorporation into hypotheses. It will become clear that AI is
only occasionally useful in the process, even if computers have become indispensable
at most stages of it.

h d i 0 0 9 8 088962 4 00 bli h d li b C b id i i
 24 Data-Driven Fluid Mechanics Chapter 2

The rest of this chapter is structured as follows. Section 2.2 discusses the concept
of coherent structure, including, in Section 2.2.1, examples from free-shear flows and,
in Section 2.2.2 and Section 2.2.3, examples from wall-bounded flows. Section 2.3
summarizes the results and discusses the relation between data analysis and conceptual
modeling.

2.2 Coherent Structures

Although turbulence is often treated as a random process in which questions are posed
in terms of statistics, a competing point of view that describes it in terms of structures
and eddies has had its followers from the beginning. Thus, while Reynolds (1894)
centered on the statistics of the fluctuations, Richardson (1920) described them in
terms of “little and big whorls.” And while Kolmogorov (1941) framed his cascade
theory as a statistical relation between fluctuation intensity and scale, Obukhov (1941)
interpreted it in terms of interactions among eddies. There is probably a reason for
the different emphases. Reynolds was an engineer and Kolmogorov a mathematician.
Both could disregard individual fluctuations, which they saw from the outside.
Richardson and Obukhov were primarily atmospheric scientists, and atmospheric
turbulent structures cannot be ignored, because they are large enough for us to live
within them.
These notes take the view that randomness is an admission of ignorance that should
be avoided whenever possible (Voltaire 1764), and that turbulence is a deterministic
dynamical system satisfying the Navier–Stokes equations. Specifically, we will be
interested in whether the description of the flow can be simplified by decomposing
it into “coherent” structures that can be extracted by observation or predicted from
theoretical considerations, and whether the dynamics of these structures can be used
to model the evolution of the flow, or even to guide us into ways of controlling it.
Up to fairly recently, the statistical point of view was all but inevitable, because it is
difficult to extract structural information from single-point measurements. It was not
until shear-flow visualizations in the 1970s revealed structures whose lives were long
enough to visually influence scalar tracers (Kline et al. 1967, Corino & Brodkey 1969,
Brown & Roshko 1974) that the structural point of view began to gain acceptance. And
it was only after probe rakes, numerical simulations, and PIV experiments routinely
provided multidimensional flow fields that that point of view became fully established.
In the same way, the more recent introduction of time-resolved three-dimensional flow
information, mostly from simulations (Perlman et al. 2007, Lozano-Durán & Jiménez
2014, Cardesa et al. 2017) has opened a window into the dynamics of those structures,
not only allowing us to conceive structures as predictable, but forcing us to try to
predict them.
However, any attempt to simplify complexity should be treated with caution,
because it implies neglecting something that may be important. There are several
ways of approaching these simplifications, some of which are described in detail in
other parts of this course. Many are based on projecting the equations of motion onto

h d i 0 0 9 8 088962 4 00 bli h d li b C b id i i
 Jiménez: Coherent Structures in Turbulence 25

Figure 2.2 The evolution of the coherent structures is expected to depend mostly on a few
other structures, plus perhaps nonessential residues.

a smaller set of variables, typically a linear subspace or a few Fourier modes, but this
approach becomes less justifiable as the Reynolds number increases, or as the system
becomes more extended. If we consider a pipe at even moderate Reynolds number,
whose length is several hundred times its diameter, we can expect to find several
thousands eddies of any definition, with widely varying scales, far enough from each
other to be essentially independent. Projection methods become less useful in those
cases because they do not respect locality, and treat together unrelated quantities.
Moreover, the Navier–Stokes equations are partial differential equations in physical
space (except for the pressure in incompressible flows), but not in Fourier space, and
our approach will be to treat the evolution of the flow as largely local, and to look for
solutions that evolve relatively independently from other similar solutions far away.
This view of turbulence is based on the “hope” that at least part of its dynamics can
be described in terms of a relatively small number of elementary objects that, at least
for some time, depend predominantly on a “few” similar objects and on themselves,
with only minor contributions from an ‘unstructured’ background. We will refer to
them as “coherent structures.” The underlying model is sketched in Figure 2.2, but it
is important to understand that this simple dependence, and therefore the existence of
coherent structures, is mostly a hope that requires testing at every step of the process.
“Self” dependence suggests properties that the structures should possess to be
relevant and practical, such as that they should be strong enough with respect to
their surroundings to have some dynamics of their own, and that they should be
observable, predictable, or computable. These requirements generally imply that
coherent structures are either “engines” that extract energy from some relevant forcing,
sinks that dissipate it, or “repositories” that hold energy long enough to be important
for the general budget of the flow.
The emphasis on structure does not exclude statistics. The descriptions of Reynolds
(1894) and Kolmogorov (1941) have been more useful in applications than most
structural approaches, and it cannot be forgotten that turbulence is as important
for engineering as it is for science. Moreover, statistics are required even when
analysing structure. Turbulent flows are chaotic and high-dimensional (Tennekes &
Lumley 1972, Keefe et al. 1992) and, even if dissipation probably restricts them to a
finite-dimensional attractor in phase space, they explore that space widely. In essence,
anything that is not strictly inconsistent with the equations of motion is bound to
occasionally happen in a turbulent flow, and it is important to approach structure
identification statistically, and to make sure that any observed phenomenon, even
if intellectually appealing, occurs often enough to influence the overall dynamics.

h d i 0 0 9 8 088962 4 00 bli h d li b C b id i i
 26 Data-Driven Fluid Mechanics PETER FREYMUTH Chapter 2

(a)

11 = 7 . 6 cm, 77" = 3 m/ser, 4!j = 0.1 em.

28. Deperideiicc of the sinoke pictm-c on the Stronhal number.
FIGCJRE
3. Shadowgraphs of mixing Iaycr t,alteti at random times. Liiies show rrlrthod of
FIGURE
determining 8tlz.Channel w i d t h at lcft side of picture is 6 mi.
BROWN AND EOSHKO

(b)

0.011s s

(c)

d
d

Figure 2.3 (a) Shadowgraph of a turbulent-free mixing layer, from Brown and Roshko (1974).
Flow is from left to right. (b) Transition in a laminar mixing layer, from Freymuth (1966).
(c) Fluctuation amplitude in a forced turbulent mixing layer. Lines are linear stability theory,
and symbols are laboratory measurements. From Gaster et al. (1985). All reproduced with
permission.
Plate 7 Journal of Flicid JIechanics, 1'01. 2 5 , part 4

On the other hand, rare events can be important if they can be exploited for control
purposes (Kawahara 2005), or if they are harmful enough to justify investing
in avoiding them, for example, tornadoes. Turbulence, as an example of natural
macroscopic chaos, may be one of the first systems in which we may hope for the
engineering implementation of a “Maxwell daemon,” in the sense of extracting useful
work from apparently random fluctuations. In a way, soaring albatrosses, or glider
pilots riding thermals, have been doing this for a long time.

2.2.1 Free-Shear Flows

Probably the most influential early visualizations of coherent structures in a fully
turbulent flow were the free-shear layers in Brown and Roshko (1974) (Figure 2.3(a)).
They were not the earliest; streaks and ejections had been observed before in
wall-bounded flows by Kline et al. (1967) and by Corino and Brodkey (1969), but
the mechanism for the structures in Brown and Roshko (1974) was immediately
obvious, because similar structures were known in transitional flows at lower
Reynolds numbers, and had been theoretically analyzed in terms of linear stability

h d i 0 0 9 8 088962 4 00 bli h d li b C b id i i
 Jiménez: Coherent Structures in Turbulence 27

Analysis of film of a turbulent mixing layer 331

Figure 2.4 Automatic tracking of structures in a free-shear layer, from Hernán and Jiménez
(1982), reproduced with permission.
FIGURE7. A part of the final (2,t)-diagram showing several amalgamations; note the tripling
on the left of the picture.

committed during the segmentation step. There, some segments of the layer are left
theory by Freymuth (1966)
unclassified (Figure
on account 2.3(b)).
of the difficulty The their
in recognizing mean velocity
shapes. These segments profile of free-shear
leave gaps in the sequence of eddies representing one frame, and it happens occasion-
flows, such as shearally layers, wakes and jets, are modally unstable, and the associated
that a history can be continued by assuming the existence of an eddy
appropriate size in such a gap. These occasions are especially easy to recognize when
of the

Kelvin–Helmholtz be instability
there is proper
are eddies a t the wellpositions
understood. Inbefore
in the frames just fact,and even
after the if
interpolated, and when moreover the size and position of the gap are such as to
onethe
to stability analysis
is linear, while turbulence is not,
permit the interpolation. The it was
algorithm soon
used confirmed
in our analysis that
inspects sequences
consecutive frames, allowing the interpolation of up to two consecutive gaps. A part
of the
four linearized results

match the large-scale dynamics of forced shear layers well (Gaster et al. 1985),
of the (x, t)-diagram drawn in this way is shown in figure 7.

within well-understood limits (Figure 2.3(c)). We noted in the introduction that it is

important to have mechanisms to supplement and support raw observations. This was
the case for the structures of free-shear flows, and it was probably the main reason
for their early scientific acceptance. The instability of shear layers is used industrially
today to enhance mixing and to reduce noise.
Of interest for the present course is that these experiments were some of the first
computer-analyzed temporal series of images in any experimental flow. A film of
the experiments in Bernal (1981) was processed by Hernán and Jiménez (1982) to
catalogue the properties of the structures and of their interactions (Figure 2.4). It
may add historical perspective to the present course that that paper, which included
automatic structure identification, classification and temporal analysis, refers in
several places to “syntactic programming,” “formal grammars,” and similar techniques
of current AI flavor. The paper did not even claim to have invented those techniques,
but cited existing textbooks on image processing.

2.2.2 Wall-Bounded Flows

Structures in wall-bounded flows were observed even earlier than in free-shear layers
(Kline et al. 1967, Corino & Brodkey 1969), but the lack of a theoretical model
comparable to the linear instability of the latter leads to continued disagreements about
the meaning of the wall-bounded structures. It had been found early that the mean
velocity profile of a turbulent channel is modally stable (Reynolds & Tiederman 1967),
and it was not until it was understood that modally stable perturbations can transiently

h d i 0 0 9 8 088962 4 00 bli h d li b C b id i i
 28 Data-Driven Fluid Mechanics Chapter 2

Figure 2.5 Near-wall structure of a turbulent channel. The flow is from left to right and the
figure only includes the region below y + ≈ 60, looking toward the wall. The gray background
is the shear at the wall and represents the streamwise-velocity streaks, and the colored shorter
objects are vortices with either positive or negative streamwise vorticity. The axes labels are in
viscous units. From Jiménez (2002), reproduced with permission of the Licensor through
PLSclear.

(a) (b)

Figure 2.6 Near-wall structure of a turbulent channel. The flow is from left to right, and the
figure represents the region below y + ≈ 80, looking toward the wall. The gray objects are the
streamwise-velocity streak, and the colored shorter objects are the vortices responsible for the
sweeps and ejections, with either positive or negative streamwise vorticity. The axes labels are
in viscous units. (a) Minimal channel at Reτ = 180. (b) A weakly oscillatory wave in a
minimal autonomous channel, from Jiménez and Simens (2001), reproduced with permission.

grow by considerable amounts (Butler & Farrell 1993) that the question of wall-
bounded turbulent structures could be treated with some rigor.
Visualizations show that the region near the wall of boundary layers, pipes, and
channels is dominated by long “streaks” (Kline et al. 1967), eventually shown to
be jets of high or low streamwise velocity, sprinkled with shorter fluid eruptions, or
“bursts” (Kim et al. 1971). It was initially hypothesized that bursts were due to the
intermittent breakup of the near-wall streaks, but even the original authors acknowl-
edged that their visualizations could be consistent with more permanent objects being
advected past the observation window. The term “burst” eventually became associated
with the “sweeps” and “ejections” observed by stationary velocity probes, which
respectively move fluid toward and away from the wall (Lu & Willmarth 1973). After
the careful visual analysis by Robinson (1991) of a temporally resolved film of one of
the first numerical simulations of a turbulent boundary layer (Spalart 1988), it became
clear that, at least in the viscous layer near the wall, the sweeps and ejections known
from single-point measurements reflected the passing of shorter quasi-streamwise
vortices, intermittent in space but not necessarily in time (see Figure 2.5). It was also
soon understood that ejections and sweeps create the streaks by deforming the mean
velocity profile, but the origin and dynamics of the bursts remained unclear.

h d i 0 0 9 8 088962 4 00 bli h d li b C b id i i
 Jiménez: Coherent Structures in Turbulence 29

The next step was taken by Jiménez and Moin (1991), who shrank the dimensions
of a channel simulation to the minimum value that could accommodate a single
sweep–ejection pair and a short streak segment, thus allowing the study of their
interactions “in isolation.” These simulations showed that wall turbulence is
largely independent of the chaotic interaction among structures in the flow, thus
satisfying one of the conditions for coherence in Figure 2.2, and that this minimal
unit was unambiguously intermittent in time. These and other authors eventually
converged to the description of a “cogeneration” cycle in which the vortices
sustain the streak, and an undetermined instability of the streak creates the vortices
(Jiménez & Moin 1991, Hamilton et al. 1995, Jiménez & Pinelli 1999, Schoppa &
Hussain 2002).
A parallel, and largely independent, development was led by Nagata (1990), who
studied wall turbulence from the point of view of dynamical systems, and was able
to compute fully nonlinear steady-state solution of the Navier–Stokes equations in
Couette flow. These solutions are typically unstable, and therefore unlikely to be
found in a real flow, but the idea was that the system evolves relatively slowly in
their phase-space neighborhood, or in that of related oscillatory solutions, and that the
relatively large fraction of time spent near them would “anchor” the flow statistics.
In fact, these simple solutions look extraordinarily similar to the bursting solutions in
minimal channels (Figure 2.6), and predict fairly well the amplitude and dimensions of
the observations in the viscous layer of real flows (Jiménez et al. 2005). They remain
to this day one of the main reasons to believe that coherent structures are important
components in the dynamics of wall-bounded turbulence. A review of their general
significance is Kawahara et al. (2012).
The reason why Figures 2.5 and 2.6 can be compared visually is that they are
restricted to the viscous layer of the flow, where the internal Reynolds number of the
structures, ∆+y = uτ ∆y /ν, is low, where uτ is the friction velocity, ν is the kinematic
viscosity, and ∆y is the height of the structure. The resulting structures are relatively
smooth and can be recognized as “objects.” As the Reynolds number of the flow
increases and some of the structures become larger, their internal Reynolds number
grows, and their shape can best be described as a fractal (see Figure 2.7(a)). Such
objects should be treated statistically, and their study had to wait until the Reynolds
number of the simulations grew and methods of analysis for the large data sets
involved were developed.
The first such studies examined the structure of vorticity in channels. Individual
intense vortices had been studied in isotropic turbulence (Vincent & Meneguzzi 1991,
Jiménez et al. 1993), where they are associated with dissipation. They were known
to be predominantly concentrated in low-velocity regions in channels (Tanahashi
et al. 2004), but del Álamo et al. (2006) showed that the relevant objects in wall-
bounded turbulence were not individual vortices but vortex tangles, or “clusters”
(Figure 2.7(b)). They are approximately collocated with the ejections, which explains
their association with low-velocity regions, and form a self-similar family of clusters
(not of individual vortices) across the logarithmic layer. The minimal-channel tech-
nique was extended by Flores and Jiménez (2010) to the logarithmic layer, where

h d i 0 0 9 8 088962 4 00 bli h d li b C b id i i
 30 Data-Driven Fluid Mechanics Chapter 2

(a)

(b)

Figure 2.7 (a) Reynolds-stress structures in the outer layer of a turbulent channel, Reτ = 2000
(Hoyas & Jiménez 2006). The flow is from left to right. The gray objects are the ejections, and
approximately correspond to low-velocity streaks. The yellow ones are sweeps. Image courtesy
of A. Lozano-Durán. (b) Temporal evolution of a vortex cluster along its lifetime. Channel at
Reτ = 4200. The maximum height of the cluster is uτ ∆y /ν ≈ 300 at which time the cluster is
attached to the wall. From Lozano-Durán and Jiménez (2014), reproduced with permission.

they used it to show that the vortex clusters also burst intermittently, and that this
process is associated with the deformation of the large-scale streaks found in that
region. Probably related to these clusters are the self-similar “hairpin packets” that
have been intensively studied by experimentalists (Adrian 2007), although there is
little information about their temporal behavior.
The next step was to analyze whether a similar organization exists for sweeps and
ejections, which had been studied for a long time as being responsible for carrying
the tangential Reynolds stress. This was done by Lozano-Durán et al. (2012), using
three-dimensional data from simulations. The structures were defined by thresholding
the Reynolds stress, and proved to also form self-similar families of side-by-side
pairs of a sweep and an ejection, which are the logarithmic-layer equivalent of the
quasi-streamwise vortices of the buffer layer. A typical flow field in which these
Reynolds-stress structures have been isolated by thresholding their intensity is given
in Figure 2.7(a), confirming that they are no longer smooth objects, and that it is
difficult to analyze them individually rather than statistically. The study of large-scale
structures of either velocity or Reynolds stress has since been extended to boundary
layers (Sillero 2014), homogeneous shear flows (Dong et al. 2017), and pipes (Lee &
Sung 2013), with broadly similar results. The elementary unit is a sweep–ejection pair
with a vortex cluster in between, located in the lateral boundary between a high- and a
low-velocity streak. We will use this structure interchangeably with the term “burst.”

h d i 0 0 9 8 088962 4 00 bli h d li b C b id i i
 Jiménez: Coherent Structures in Turbulence 31

(a) (b) (c)

Figure 2.8 Effect of the inhomogeneity of the mean velocity profile on the flow field
conditioned to sweep–ejection pairs. The central opaque S-shaped object is an isosurface of
the magnitude of the conditional perturbation vorticity, at 25% of its maximum. The two
translucent objects are isosurfaces of the conditional perturbation streamwise velocity,
u+ = ±0.6. The axes are scaled with the average of the diagonal sizes of the sweep and the
ejection. (a) Homogeneous shear turbulence. (b) Detached eddy in a channel, Reτ = 950.
(c) As in (b), for an attached eddy. Adapted from Dong et al. (2017), reproduced with
permission.

As temporally resolved data sets became available, individual bursts were tracked
by Lozano-Durán and Jiménez (2014), finally answering the question of whether
ejections are permanent passing objects or transient ones. They have a definite
lifetime, shown in Figure 2.7(b) for a vortex cluster, but this lifetime is longer than
the passing time across a stationary probe, so that the signal that the latter sees is due
to advection. Individual bursts are born and disappear at an approximately constant
distance from the wall but, as they grow toward the middle of their lifetime, some
of them become large enough for their root to become attached to the wall. On
the general principle that the largest eddies carry the Reynolds stresses (Tennekes
& Lumley 1972), these attached eddies are responsible for most of the momentum
transfer in the flow (Townsend 1976), where the definition of “large” is linked to the
Corrsin (1958) scale, which separates the eddies that feel the effect of the shear from
the smaller ones that do not (Jiménez 2018a). The temporal behavior of the large
eddies is also self-similar. The lifetime of an eddy whose maximum height is ∆y is
approximately an eddy turnover, uτ T/∆y ≈ 1.
The confirmation that bursts are transient reopened the question of their origin.
There is persuasive evidence that they are essentially linearizable Orr (1907b)
solutions, which transiently extract energy from the mean flow as they are tilted
forward by the shear (Jiménez 2013, Encinar & Jiménez 2020). In retrospect, this
was to be expected, because the only energy source in the flow is the mean shear,
but it leaves open the question of how the bursts are seeded. Orr bursts are temporary
structures of the wall-normal velocity; they are born weak, leaning backward with
respect to the shear; they intensify as the shear tilts them toward the normal to the
wall, and eventually disappear when the forward tilt becomes too pronounced. In
the process, some of their energy is transferred to the streaks and remains in the
flow. This is a plausible mechanism for turbulent-energy production, but there is
no obvious process to restart the eddies as backward-leaning structures. There is

h d i 0 0 9 8 088962 4 00 bli h d li b C b id i i
 32 Data-Driven Fluid Mechanics Chapter 2

overwhelming “circumstantial” evidence that the picture sketched earlier corresponds

to the kinematics of the flow, but using the same argument as in free-shear flows, its
definitive acceptance will only come when a complete model is available. Note that the
model presented here refers only to the shear, and does not reserve any special role for
the wall. In fact, the structures studied by Dong et al. (2017) in a homogeneous shear
flow, which has no walls, behave very similarly to those in channels, and there is a
smooth transition from the symmetric sweep–ejection pairs of the homogeneous shear
to the very asymmetric wall-attached ones (Figure 2.8). An extended recent review of
these and other aspects of the structures of wall-bounded turbulent flows is Jiménez
(2018a).

2.2.3 Conceptual Experiments

Although most of the results mentioned earlier exploit correlations among data
that had not been compiled in response to specific questions, settling controversies
often requires direct experimentation of the kind discussed in the introduction as
“interventional science.” Conceptual, “hypothesis-driven,” experiments have been a
fixture of physics for a long time. In simple dynamical situations, such as point
masses or moving clocks, the outcome of such exercises can often be used to settle
the soundness of a particular model with little more than pencil, paper, and a good
intellect, but in complex cases, such as turbulence, the guessing almost always has to
be substituted by actual experimentation on the modified system.
The first experiment of this type in structural turbulence was probably Cimbala
et al. (1988), and predates simulations. There was some controversy at the time
about whether the structures in turbulent wakes were remnants of vortex shedding
by the obstacle that creates the wake, or local instabilities of the wake profile. Their
experiment showed that the latter was the case by artificially damping the structures
present in the wake at some distance downstream, and observing how they reappear
afterward.
After computers started to be powerful enough to simulate turbulence, concep-
tual experiments became mostly computational, because simulations permit many
“unphysical” manipulations that are not easily implemented otherwise. In wall-
bounded flows, we have already mentioned the minimal experiments by Jiménez and
Moin (1991). Later, to settle a controversy about whether turbulence near the wall
was autonomous or forced by outer motions, Jiménez and Pinelli (1999) performed
simulations in which the flow equations were artificially damped away from the wall,
and complemented these experiments by selective modifications of the equations
of motion and of the boundary conditions, intended to falsify different versions of
the regeneration cycle. Mizuno and Jiménez (2013) did the opposite, damping only
the near-wall region. Both experiments showed that the undamped part of the flow
survives by itself, and that wall turbulence at a given distance from the wall can
be studied relatively independently of other wall distances. The homogeneous-shear
simulations by Dong et al. (2017), mentioned earlier, were intended to complement
these experiments, removing both the near-wall region and the outer flow.

h d i 0 0 9 8 088962 4 00 bli h d li b C b id i i
 Jiménez: Coherent Structures in Turbulence 33

2.3 Discussion

The previous sections have reviewed some of what is known about structures in
shear flows. Except for the earliest visualizations of free-shear layers, all the data
mentioned here have been “big” in relation to the dates in which they were generated
and analyzed: from tens of Mbytes for the first channel simulations in the 1980s, to
hundreds of Tbytes for the current time-resolved data sets. The field has been driven
as much by storage capacity as by computational power or experimental methods, and
the analysis of the data has always involved new techniques that were not available,
or required, before the new data came online. It would appear that the description of
these methods should have been the main thrust of a set of notes on the application
of AI to fluid mechanics, but a rereading of the previous sections shows that this has
not been the case. Most of them are dedicated to the discussion of physical models.
We have noted several times that the quest for structures is driven by the hope that
they simplify the description and control of the flow, and that their acceptance by the
community depends on the acceptance of these simplifications.
Coherent structures are not an obvious fit for fluid mechanics, which is best
described by continuous vector fields. Structures are patterns that we hope are useful
for us, but we should be weary of the evidence that humans are notoriously prone to
finding spurious patterns. Consider constellations in the night sky, or divination from
tea leaves, and note that even those spurious patterns keep being considered relevant
because there is a community that finds them useful.
The Kelvin–Helmholtz rollers of free-shear layers were considered useful from the
start because their mechanism is clear and offers a distinct avenue for the description
and control of the flow. The bursts and streaks of wall-bounded flows are still
considered less definitive because the description of their regeneration cycle remains
incomplete.
If any conclusion can be drawn from the present discussion is that the scientific
cycle in Figure 2.1 should be considered as a whole, that data analysis should include
a parallel development of physical models, and that no single isolated step should be
taken as final. Apparently convincing flow descriptions in terms of a reduced set of
structures or modes should be taken with suspicion unless they are complemented by
plausible mechanisms. In the nomenclature introduced in Section 2.1, sub-symbolic
AI only becomes scientifically conclusive once it becomes symbolic.
Appealing mechanisms should also be carefully examined for evidence that they are
really important for the flow. No single description is likely to span the full dynamics
of turbulence, and we are probably bound to always “make do” with partial results.
But it is important to consider the phenomenon as a whole, and to recognize which
part of it is described by our models. One of the reasons to appreciate the bursting
model in Section 2.2.2 is that it accounts for 60% of the skin friction using only 10%
of the flow volume (Jiménez 2018a). This is appealing, but it is fair to wonder what
happens to the other 40%. One of the most dangerous sentences in scientific research
is: “It is a start.”

h d i 0 0 9 8 088962 4 00 bli h d li b C b id i i
 3 Machine Learning in Fluids:
Pairing Methods with Problems
S. Brunton∗

The modeling, optimization, and control of fluid flows remain a grand challenge
of the modern era, with potentially transformative scientific, technological, and
industrial impact. An improved understanding of complex flow physics may enable
drag reduction, lift increase, mixing enhancement, and noise reduction in domains
as diverse as transportation, energy, security, and medicine. Fluid dynamics is
challenged by strong nonlinearity, high-dimensionality, and multiscale physics; both
modeling and control may be thought of as non-convex optimization tasks. Recent
advances in machine learning and data-driven optimization are revolutionizing how
we approach these traditionally intractable problems. Indeed, machine learning may
be considered as a growing set of techniques in data-driven optimization and applied
regression that are tailored for high-dimensional and nonlinear problems. Thus, the
modeling, optimization, and control of complex fluid systems via machine learning
are complementing mature numerical and experimental techniques that are generating
increasingly large volumes of data.

3.1 Overview

The large range of spatial and temporal scales in a complex fluid flow requires exceed-
ingly high-dimensional measurements and computational discretization to resolve all
relevant features, resulting in vast data sets and time-intensive computations. There
is no indication that the increasing volumes of data generated and analyzed in fluid
mechanics will slow any time soon, and we are decades away from fully resolving the
most complex engineering flows (aircraft, submarines, etc.) fast enough for iterative
optimization and in-time control. Despite the large number of degrees of freedom
required to describe fluid systems, there are often dominant patterns that emerge,
characterized by energetically or dynamically important coherent structures (see
Figure 3.1 and Chapter 2). These coherent structures provide a compact representation
of these flows, and capturing their evolution dynamics has been the focus of intense
research in reduced-order modeling for decades, as described in Chapter 1. Advances
in machine learning has promised a renaissance in the analysis and understanding
of such complex data, extracting patterns in high-dimensional, multimodal data
that are beyond the ability of humans to grasp. As more complete measurements

h d i 0 0 9 8 088962 4 006 bli h d li b C b id i i

 Brunton: Machine Learning in Fluids 35

Figure 3.1 Fluid dynamics are often characterized by complex, multiscale dynamics in space
and time. For example, by increasing the Reynolds number, the flow past a circular cylinder
becomes unsteady and turbulent. Figure adapted from Feynman et al. (2013), reprinted by
permission of Basic Books, an imprint of Hachette Book Group, Inc.

become available for more complicated fluid systems, the data associated with these
investigations will only become increasingly large and unwieldy. Data-driven methods
are central to managing and analyzing this data, and generally, we consider data
methods to include techniques that apply equally well to data from numerical or
experimental systems. In fact, data-driven methods in machine learning provide a
common framework for both experimental and numerical data.
In this chapter, we will provide an overview of the established and emerging
applications of machine learning in fluid dynamics, emphasizing the potentially
transformative impact in modeling, optimization, and control; for more details and
references, see Brenner et al. (2019) and Brunton et al. (2020). We take the perspective
that machine learning is a collection of applied optimization techniques that build
models based on available data, which is abundant in fluid systems. Because of the
data-intensive nature of fluid dynamics, machine learning complements efforts in both
experiments and simulations, which generate incredible amounts of data. In machine
learning, it is just as important to understand when methods fail as when they succeed,
and we will attempt to convey the limitations of current methods, exploring what is
easy and difficult in machine learning. We will also emphasize the need to incorporate
physics into machine learning to promote interpretable and generalizable models.
Finally, we cannot overstate the importance of cross-validation to prevent overfitting.
Fluid flow fields may be naturally viewed as images, as in Figure 3.2. Thus, there is
considerable low-hanging fruit in applying standard techniques from machine learning
for image analysis to flow fields, for example for occlusion inference (Scherl et al.
2020) and super-resolution (Fukami et al. 2018, Erichson et al. 2020). However, fluid
dynamics departs from many of the typical application domains of machine learning,
such as image recognition, in that it is fundamentally governed by physics. Many of the

h d i 0 0 9 8 088962 4 006 bli h d li b C b id i i

 36 Data-Driven Fluid Mechanics Chapter 3

Figure 3.2 Illustration of how simple it is to add texture to an image using modern neural
networks. Is it possible to add the texture of turbulence to a laminar flow?

goals of fluid dynamics involve actively or passively manipulating these dynamics for
an engineering objective through intervention, which inherently changes the nature
of the data collected; see Chapters 17 through 19. Thus, the dynamic and physical
nature of fluids provides both a challenge and an opportunity for machine learning,
potentially providing deeper theoretical insights into algorithms by applying them to
systems with known physics.
Although fluid data is vast in some dimensions (i.e., fully resolved velocity fields
may easily contain millions or billions of degrees of freedom), it may be relatively
sparse in others (e.g., it may be expensive to sample many geometries and perform
parametric studies). This heterogeneity of the data can limit the available techniques
in machine learning, and it is important to carefully consider whether or not the model
will be used for interpolation within a parameter regime or for extrapolation, which
is generally much more challenging. In addition, many fluid systems are naturally
nonstationary, and even for stationary flows it may be prohibitively expensive to obtain
statistically converged results when behavior is broadband; see Chapter 9. Typically,
engineering goals that involve a working fluid are generally not simple classification
tasks, but often involve more subtle multi-objective optimization, for example to
maximize lift while minimizing drag. Finally, because fluid dynamic systems are
central to transportation, health, and defense systems, it is essential that machine
learning solutions are interpretable, explainable, and generalizable, and it is often
necessary to provide guarantees on performance.
Fluid dynamics challenges often amount to large, high-dimensional, non-convex
optimization problems for which machine learning techniques are becoming
increasingly well adapted. Although the opportunities of machine learning in fluid
mechanics are nearly limitless, we roughly categorize efforts into (1) modeling, (2)
closed-loop control, and (3) engineering optimization. These categories necessarily
share some overlap. The traditional wisdom in fluid dynamics is that modeling is
essential to support the downstream goals of optimization and control. However,
effective optimization and control may bypass modeling. In fact, by manipulating

h d i 0 0 9 8 088962 4 006 bli h d li b C b id i i

 Brunton: Machine Learning in Fluids 37

the flow and exciting transients, we may indeed generate valuable data to enrich our
modeling efforts.

Flow feature extraction

– There is a tremendous opportunity to improve efforts in dimensionality reduction to
identify better coordinates for compact descriptions. This topic is related to coherent
structures from Chapter 2 and will be discussed in Section 3.3.

Modeling flow dynamics

– Machine learning is providing new computational techniques to develop models
from data, which may improve our understanding of key dynamic mechanisms for
efficient and accurate prediction, estimation, and control. This will be discussed in
Section 3.4 and Chapter 14.

Control and optimization

– Flow control is a long-held goal of fluid dynamics. Sensor-based feedback control
of a fluid flow may be viewed as a non-convex optimization based on measurement
data. Machine learning is poised to transform the field of turbulence control, with
potential impact across the industrial and technological landscape. It may also be used
for design optimization in fluids, for example to optimize geometry and parameters
for multi-objective performance goals. This will be discussed in Section 3.5 and
Chapter 17.

3.1.1 Kinematic and Dynamics Modeling of Fluid Flows

Modeling fluid flows has been the focus of efforts in mathematical physics and applied
engineering for centuries. Flow modeling is particularly useful for understanding
essential flow physics, design and optimization, and active control. Fluid dynamics
is a notoriously challenging field of classical physics, especially the characterization
of turbulence, which is nonlinear, high-dimensional, and multiscale. Although the
Navier–Stokes equations provide a precise mathematical model, it is difficult to use
this representation for design, optimization, and control. In fact, much of the observed
richness observed in fluids is often obscured by the simplicity of the governing
equations, as elegantly observed by Feynman:
The test of science is its ability to predict. Had you never visited the earth, could you predict
the thunderstorms, the volcanos, the ocean waves, the auroras, and the colorful sunset?

Instead of analyzing the governing equations directly, we typically simulate high-

dimensional discretized equations or perform laboratory experiments to explore the
phenomenology associated with a specific relevant configuration. However, simula-
tions and experiments are expensive, both in financial and human resources, and they
are often not suitable for iterative optimization or real-time control. Thus, considerable
effort has gone into obtaining accurate and efficient reduced-order models that capture
the essential flow mechanisms at a fraction of the cost of a fully resolved flow solver.

h d i 0 0 9 8 088962 4 006 bli h d li b C b id i i

 38 Data-Driven Fluid Mechanics Chapter 3

This section will distinguish between kinematic and dynamic flow modeling, which
will be important for later sections.
There are several important aspects of flow modeling that we will investigate here.
Modeling may be roughly categorized into two complementary efforts: dimensionality
reduction and reduced-order modeling. Dimensionality reduction involves extracting
dominant patterns that may be used as reduced coordinates where the fluid is
compactly and efficiently described. Reduced-order modeling describes the evolution
of the flow in time as a parameterized dynamical system, although it may also
involve developing a statistical map from parameters to averaged quantities, such as
drag. Models describe how the flow varies, either in time, with a parameter (e.g.,
for optimization), or with actuation (e.g., for control). Identifying and extracting
relevant flow features is essential to describe and model the flow. In fact, much
of the progress in mathematical physics has revolved around identifying coordinate
transformations that simplify dynamics and capture essential physics, such as the
Fourier transform and modern data-driven modal decompositions such as the proper
orthogonal decomposition (POD); see Chapter 6.
Reduced-order modeling encompasses model reduction, which begins with the
governing equations, and system identification, which begins with data. Model
reduction, such as Galerkin projection of the Navier–Stokes equations onto an
orthogonal basis of POD modes, benefits from a close connection to the governing
equations; however, it is intrusive, requiring human expertise to develop models
from a working simulation. System identification provides a flexible and data-driven
alternative, although often resulting in black box models that lack a deep connection
to the physics; see Chapters 7 and 8 for more details on system identification. Machine
learning constitutes a rapidly growing body of algorithms that may be used for
advanced system identification. A central goal of modeling is to balance efficiency
and accuracy, which are often dueling objectives. When modeling physical systems,
interpretability and generalizability are also critical considerations. Other unique
aspects of data-driven modeling of physical systems include partial prior knowledge
of the governing equations, constraints, and symmetries. Finally, the volume, variety,
and fidelity of data will impact the viability of various machine learning methods.
We will separate the discussion into the use of machine learning for (1) flow feature
extraction to model flow kinematics, and (2) modeling dynamics. Of course, there is
some overlap in these topics. For example, the dynamic mode decomposition (DMD)
and related Koopman operator approaches are simultaneously concerned with obtain-
ing effective coordinates to represent coherent structures (kinematics) and a model
for how these coherent structures evolve (dynamics). Similarly, in the discussion of
closure models for turbulence, there are several approaches, ranging from using super-
resolution to fill in low-resolution flow images (kinematics), to directly modeling the
closure terms as a dynamical system (dynamics). Mathematically, the discussion of
kinematics amounts to obtaining a change of coordinates from some variables or
measurements x to a new set of variables a through a possibly nonlinear function ϕ:

a = ϕ(x). (3.1)

h d i 0 0 9 8 088962 4 006 bli h d li b C b id i i

 Brunton: Machine Learning in Fluids 39

...

Data Linear subspace (POD) Nonlinear manifold

Figure 3.3 Illustration of linear versus nonlinear embeddings of fluid flow data for flow past a
circular cylinder. It is possible to describe the dominant evolution in a three-dimensional linear
subspace obtained via POD (Noack et al. 2003). However, in this subspace, the flow actually
evolves on a two-dimensional submanifold, given by a parabolic inertial manifold. In general,
linear subspaces are simpler to obtain, although nonlinear embeddings may yield enhanced
reduction.

The original variables x may be high-resolution flow fields from a simulation or

particle image velocimetry (PIV) experiment, and a may represent the amplitudes
of a modal expansion of the flow in terms of dominant coherent structures (Taira
et al. 2017), resulting in dimensionality reduction. In contrast, the variables x may
contain under-resolved measurements and a may be the desired high-resolution flow
field, which amounts to a super-resolution problem. Similarly, the discussion on
dynamics amounts to obtaining a differential equation model of the form
d
a = f(a, u, t; β), (3.2)
dt
in terms of these variables a, some control inputs u, time t, and a set of parameters β.

3.2 Machine Learning Basics

Machine learning is one of the great technological developments of our generation,

and it is undoubtedly changing the art of what is possible across the scientific and
industrial landscapes, especially in the field of fluid mechanics (Duraisamy et al.
2019, Brenner et al. 2019, Brunton, Noack & Koumoutsakos 2020, Brunton, Hemati
& Taira 2020). Simply put, machine learning refers to the ability of a computer
system to learn and improve from experiential data, rather than being explicitly
or deterministically programmed. The rise of machine learning is driven by the
confluence of (1) vast and increasing volumes of data; (2) tremendous advances
in computational hardware and orders of magnitude less expensive data storage
and transfer; (3) advanced optimization algorithms in applied math and statistics;
(4) increasingly expressive architectures (e.g., neural networks); and (5) significant
investment by industry, leading to open source development tools (see Figure 3.4). For
example, the ImageNet database of labeled images (Deng et al. 2009) and increasingly
powerful GPU technology were critical in the development and demonstration of

h d i 0 0 9 8 088962 4 006 bli h d li b C b id i i

 40 Data-Driven Fluid Mechanics Chapter 3

Figure 3.4 The success of machine learning in the past decade can largely be attributed to (1)
large-scale labeled training data, such as image databases; (2) advanced computational
hardware, such as GPUs; (3) improved neural network architectures and optimization
algorithms to train them; and (4) industry investment, leading to open source tools. GPU image
used courtesy of NVIDIA, www.nvidia.com. PyTorch, the PyTorch logo, and any related
marks are trademarks of Facebook, Inc.

modern deep learning architectures. There is tremendous potential for data science
and machine learning to revolutionize nearly every aspect of our modern industrial
world, and few fields stand to benefit as clearly or as significantly as the field
of fluid mechanics. Indeed, fluid dynamics is one of the original big data fields,
and many high-performance computing architectures, experimental measurement
techniques, and advanced data processing and visualization algorithms were driven
by the decades of research on fluid systems. However, it is important to recognize
that machine learning will not replace experimental and numerical efforts in fluid
mechanics, but will instead complement them (see Figure 3.5), much as the rise of
computational fluid dynamics (CFD) complemented experimental efforts (Brunton,
Hemati & Taira 2020).
Roughly speaking, there are five major stages in machine learning: (1) asking a
question, defining a hypothesis to be tested, or identifying a function to be modeled;
(2) preparing a data pipeline, either through collecting and curating the training
data (e.g., for supervised and unsupervised learning), or setting up an interactive
environment (e.g., for reinforcement learning); (3) identifying the model architecture
and parameterization; (4) defining a loss function to be minimized; and (5) choosing
an optimization strategy to optimize the parameters of the model to fit the data.
Human intelligence is critical in each of these stages. Although considerable attention
is typically given to the learning architecture, it is often the data collection and
optimization stages that require the most time and resources. It is also important to
note that known physics (e.g., invariances, symmetries, conservation laws, constraints,
etc.) may be incorporated in each of these stages (Battaglia et al. 2018, Loiseau

h d i 0 0 9 8 088962 4 006 bli h d li b C b id i i

 Brunton: Machine Learning in Fluids 41

Tasks aided by
machine learning

Flow
data

Figure 3.5 Schematic overview of the use of machine learning in fluid dynamics.

et al. 2018, Lusch et al. 2018, Champion et al. 2019, Cranmer et al. 2019, Zheng
et al. 2019, Greydanus et al. 2019, Noé et al. 2019, Champion et al. 2020, Zheng
et al. 2019, Cranmer et al. 2020, Finzi et al. 2020, Raissi et al. 2020, Zhong &
Leonard 2020). For example, rotational invariance is often incorporated by enriching
the training data with rotated copies, and translational invariance is often captured
using convolutional neural network (CNN) architectures. Additional physics and
prior knowledge may also be incorporated as extra loss functions or constraints in
the optimization problem (Loiseau et al. 2018, Lusch et al. 2018, Champion et al.
2019, Zheng et al. 2019, Champion et al. 2020). There is also considerable work in
developing physics-informed neural network architectures (Lu et al. 2019, Raissi et al.,
2019, 2020), networks that capture Hamiltonian or Lagrangian dynamics (Cranmer
et al. 2019, 2020, Greydanus et al. 2019, Finzi et al. 2020, Zhong & Leonard
2020), and that uncover coordinate transformations where nonlinear systems become
approximately linear (Takeishi et al. 2017, Yeung et al. 2017, Lusch et al. 2018, Mardt
et al. 2018, Wehmeyer & Noé 2018, Otto & Rowley 2019).

3.2.1 Machine Learning Categorizations

Machine learning is a growing set of optimization and regression techniques to
build models from data. There are a number of important dichotomies with which
we may organize the variety of machine learning techniques. Here, we will group
these into supervised, unsupervised, and semi-supervised techniques, based on the
extent to which the training data is labeled. Reinforcement learning is another major
branch of machine learning research, which will be described in detail in Chapter 18.
A rough organization is shown in Figure 3.6. In many of these approaches, a
function mapping inputs to outputs must be learned. There are many approaches,
including linear and nonlinear regression, genetic programming (i.e., symbolic regres-

h d i 0 0 9 8 088962 4 006 bli h d li b C b id i i

 42 Data-Driven Fluid Mechanics Chapter 3

Clustering
Discrete
e.g., K-means,
Spectral
clustering
No Discrete or
Unsupervised Continuous?
Embedding
Continuous e.g., POD/PCA
Autoencoder
Diffusion maps

Model Generative
Training Models (GAN)
Data Partial (self-supervised)
Complex Labeled? Semi- Model or
System (truth) supervised Modify?
Modify Reinforcement
Learning

Discrete Classification
e.g, SVM,
Neural network
Yes Model or Model Discrete or Decision tree, …
Supervised Continuous?
Modify? Regression
Continuous e.g., Linear
Generalized linear
Gaussian process

Optimization
Modify
and
Control
Feedback signal to modify system or control parameters

Figure 3.6 Schematic overview of various machine learning techniques.

sion) (Holland 1975, Koza 1992), and neural networks (Goodfellow et al. 2016). In
addition, it is increasingly common to add regularization terms to the optimization
problem to promote certain beneficial problems in the resulting model. For example,
promoting sparsity often results in models that are more interpretable (i.e., fewer
parameters to interpret), and less prone to overfitting (i.e., fewer parameters to
overfit) (Hastie et al. 2009, Brunton, Proctor & Kutz 2016a, Brunton & Kutz 2019).
Including priors and regularizers to embed known physics is a particularly exciting
area of development.

Supervised Learning
Supervised learning assumes the availability of labeled training data, and it is
concerned with learning a function mapping the training data to the label. Many of
the most powerful approaches in machine learning are supervised, as there are many
data sets where we have empirical knowledge of the labels, but where it is challenging
to develop a deterministic algorithm to describe the function between the data and the
labels. Supervised learning solves this fuzzy problem by leveraging data rather than
hand-coded algorithms.
If the labels are discrete, such as categories to classify an image (e.g., dog vs.
cat), then the supervised learning task is a classification. If the labels are continuous,
such as the value of the lift coefficient for a particular airfoil shape, then the task is a
regression. The labels may also correspond to an objective that should be minimized
or maximized, resulting in an optimization or control problem.
Before the rise of deep learning, support vector machines (SVMs) (Schölkopf
& Smola 2002) and random forests (Breiman 2001) dominated classification tasks,
and were industry standard methods. Continuous regression methods, such as linear
regression, logistic regression, and Gaussian process regression are still widespread.

h d i 0 0 9 8 088962 4 006 bli h d li b C b id i i

 Brunton: Machine Learning in Fluids 43

Figure 3.7 Generative art based on Arcimbaldo’s La Primavera. Reproduced from

https://commons.wikimedia.org/wiki/File:Deepdream-arcimboldo.gif.

Unsupervised Learning
Unsupervised learning, also known as data mining or pattern extraction, determines
the underlying structure of a data set without labels. Again, if the data is to be grouped
into distinct categories, then the task is clustering, while if the data has a continuous
distribution, the task is an embedding.
There are several popular approaches to clustering. The most common and simple
clustering algorithm is the k-means algorithm, which partitions data instances into k
clusters; an observation will belong to the cluster with the nearest centroid, resulting
in a partition of the data space into Voronoi cells. Spectral clustering (Ng et al. 2002)
is also quite common. Likewise, there are many approaches to learn embeddings. The
singular value decomposition (SVD) and related principal component analysis (PCA)
and proper orthogonal decomposition (POD) are perhaps the most widely used linear
embedding techniques, resulting in an orthogonal low-dimensional basis that captures
most of the variance in a data set. These embeddings have recently been generalized
in the neural network autoencoder, discussed in Section 3.2.2.

Semi-Supervised and Reinforcement Learning

Semi-supervised learning, where only partially labeled data is available, is becoming
increasingly popular. In many situations, it is prohibitively expensive to generate a
labeled data set that is large enough to train a modern deep learning architecture.
Similarly, in many cases, full state information is not available, for example in control
scenarios.
Generative adversarial networks (GANs) (Goodfellow et al. 2014) are an important
and growing avenue of semi-supervised learning. GANs result in a generative model
that creates new data to mimic the distribution of the training data. A GAN consists of
two networks, the generative network, and a critic network that decide if the generated
data is fictitious or not. These two networks are trained in an adversarial fashion so
that with even a small amount of labeled training data, this architecture is able to
begin training itself. Many of the most impressive recent trends in neural networks,
such as deep fakes, are based on GANs. Figure 3.7 shows generative art based on
Arcimbaldo’s La Primavera. Interestingly, the original art may be considered as a
predecessor of the current deep dream trend in generative art.
Reinforcement learning (RL) (Sutton & Barto 2018) is another important class
of machine learning algorithms in which an actor is learning to interact with an
environment. RL will be the subject of Chapter 18. Often, the reward information

h d i 0 0 9 8 088962 4 006 bli h d li b C b id i i

 44 Data-Driven Fluid Mechanics Chapter 3

about whether or not a given action was beneficial will not occur regularly; for
example, in the game of chess, it is not always obvious exactly which moves led
to a player winning or losing. Thus, only partial information is available to guide the
control strategy, making it closely related to semi-supervised learning, although it is
typically considered a distinct branch of machine learning. Reinforcement learning
has been widely used to learn game strategies, for example in Alpha Go (Mnih
et al. 2015). Increasingly, these approaches are being applied to scientific and robotic
applications (Mnih et al. 2015, Silver et al. 2016).

3.2.2 Neural Networks

Neural networks are particularly powerful and expressive architectures (Goodfellow
et al. 2016). Many of the most impressive advances in machine learning over the
past decade have involved deep neural network architectures that are trained on
powerful graphics processing units. Here, we will review some of the most prominent
architectures, although new developments are taking place every day in the field.
Before beginning, it is worth noting that neural networks have a long and rich
history spanning decades (Hopfield 1982, Hinton & Sejnowski 1986, Gardner 1988),
beginning with the perceptron (Rosenblatt 1958). However, early neural networks
did not reach their full potential, lacking sufficiently large data sets, computational
hardware, and optimization algorithms. There are several reasons for the recent
success of neural networks. First, there are vast and increasing volumes of data in
nearly every field, providing the data necessary to train increasingly large networks.
Second, dramatic improvements to computational hardware make it possible to train
these large multi-layer networks. Third, there have been considerable advances in
non-convex optimization, with back-propagation being a particularly critical devel-
opment (Rumelhart et al. 1986). The universal approximation theorem has also been
important to establishing the fact that a sufficiently deep neural network is capable
of arbitrary function approximation (Hornik et al. 1989). The modern era of deep
learning (Goodfellow et al. 2016) began with the incredible success of neural networks
on the ImageNet image classification task (Krizhevsky et al. 2012), beating out all
other algorithms. Since then, neural networks have dominated nearly every task in
image processing and natural language processing.
Artificial neural networks (ANNs) seek to mimic the learning capabilities of
animals, and they generally consist of a set of neurons connected in a directed graph to
process input data and produce an output. Modern deep learning refers to ANNs that
have many distinct layers of neurons, much as the human visual system has several
distinct processing layers; each layer learns a different level of abstraction. Each
individual neuron performs an operation on the input data, known as an activation
function. Traditionally, sigmoidal activation functions were used to mimic the firing
thresholds of biological neurons, although now there are many activation functions,
with the rectified linear unit (ReLU) being a popular choice. Because of the layered
structure of many ANNs, mathematically, these may be described by a composition of
functions

h d i 0 0 9 8 088962 4 006 bli h d li b C b id i i

 Brunton: Machine Learning in Fluids 45

(a) Feedforward network (b) Autoencoder network

(c) Convolutional network (d) Residual network

(e) Recurrent network

Figure 3.8 (a) Simple feedforward neural network. (b) Deep autoencoder network. (c)
Convolutional neural network. Reproduced from
https://commons.wikimedia.org/wiki/File:
3_filters_in_a_Convolutional_Neural_Network.gif. (d) Deep residual network. (e)
Recurrent neural network. Reproduced from
https://en.wikipedia.org/wiki/File:Recurrent_neural_network_unfold.svg.

y = f1 ( f2 (· · · ( f n (x; θ n ); · · · ); θ2 ); θ1 ), (3.3)

where θ j describe the network weights for each layer (i.e., the weights of the graph)
that must be learned.
There are several challenges in training these large networks, as there may be
millions or billions of parameters in a modern ANN. Backpropagation essentially uses
the chain rule for the composition of functions to compute the gradient of the objective
function with respect to the weights. These gradient computations make it possible to
optimize parameters with gradient descent; stochastic gradient descent algorithms,
such as ADAM, are typically used because of the large scale of the optimization
problem (Kingma & Ba 2014). Other challenges include vanishing gradients and the
need for good network initializations. These considerations all depend on having an
effective network architecture and loss function, which are meta-challenges. More
details on network training can be found in Goodfellow et al. (2016) and Brunton and
Kutz (2019).
We will now review a number of commonly used architectures. The simple feed-
forward architecture, shown in Figure 3.8(a), is useful for approximating input–output
functions. Figure 3.8(b) shows a deep autoencoder network. Autoencoders generalize
the principal component analysis (PCA) to nonlinear coordinate embeddings; the
connection between PCA and linear autoencoders was first established by Baldi

h d i 0 0 9 8 088962 4 006 bli h d li b C b id i i

 46 Data-Driven Fluid Mechanics Chapter 3

Underfit

Training error

Test error
Parsimonious
0 Overfit

Model complexity

Figure 3.9 Schematic illustrating the importance of cross-validation to prevent overfitting.

and Hornik (1989), and the first use of deep nonlinear autoencoders in fluids was
in Milano and Koumoutsakos (2002). Recently, deep residual networks (Szegedy
et al. 2017), shown in Figure 3.8(d), have enabled better optimization performance
by including jump connections in exceedingly deep networks. Figure 3.8(e) depicts a
recurrent neural network (RNN), which, unlike feedforward networks, has feedback
connections from later layers back to earlier layers. The long-short term memory
(LSTM) network (Hochreiter & Schmidhuber 1997) is a particular variant of an RNN
that is useful for dynamical systems and other time-series data, such as audio signals.
Finally, Figure 3.8(c) depicts a CNN, which is a useful architecture for processing data
with spatial correlations, such as images.

3.2.3 Cross-Validation
When discussing machine learning, it is critical to introduce the notion of cross-
validation. Cross-validation plays the same role in machine learning that grid refine-
ment studies play in CFD. Because many modern machine learning architectures have
millions of degrees of freedom, it is easy to overfit the model to the training data.
Thus, it is essential to split the data into training and testing data; the model is trained
using the training data and evaluated on the testing data. Depending on the volume of
available data, it is common to randomly split the data into training and testing data,
and then repeat this process many times to get statistically averaged results. Figure 3.9
shows the idea behind cross-validation, where a model may overfit the training data
unless it is cross-validated against test data.

3.3 Flow Feature Extraction

Pattern extraction is a core strength of machine learning from which fluid dynamics
stands to benefit directly. A tremendous amount of machinery has been developed for
image and audio processing, and it is fortunate that fluid flow measurements, both
spatial and temporal, may readily capitalize on these techniques. Here, we begin by
discussing linear and nonlinear dimensionality reduction techniques, also known as

h d i 0 0 9 8 088962 4 006 bli h d li b C b id i i

 Brunton: Machine Learning in Fluids 47

(a) (b)
U V ϕ(x) ψ(x)

a a

x x x x

Figure 3.10 Illustration of shallow, linear autoencoder (a), versus deep nonlinear autoencoder
(b). In the linear autoencoder, the node activation functions are linear, so that the encoder
matrix U and decoder matrix V are chosen to minimize the loss function kx − VUxk. In the
nonlinear autoencoder, the node activation functions may be nonlinear, and the encoder ϕ and
decoder ψ are chosen to minimize the loss function kx − ψϕ(x)k. In each case, the latent
variable a describes the low-dimensional subspace or manifold on which the data evolves.
Figure modified from Brunton, Noack and Koumoutsakos (2020).

embeddings of the high-dimensional data into either a linear subspace or a nonlinear

manifold, as depicted in Figure 3.3. These techniques are related to the classical POD
approach (see Chapter 6), but may be extended for nonlinear mappings with machine
learning. Next, we discuss the use of clustering and classification, a well-developed
field of machine learning, for decomposing flow behaviors.

3.3.1 Dimensionality Reduction

It is common in fluid dynamics to define an orthogonal linear coordinate transforma-
tion from physical coordinates into a modal basis. For example, the Fourier transform
is one such change of basis that has advantages for computational speed and accuracy.
However, the Fourier transform is rather restrictive on allowable boundary conditions
and geometries. The proper orthogonal decomposition (Lumley 1970) provides a
generalization of the Fourier transform, resulting in an orthogonal basis for complex
geometries based on empirical measurements. Sirovich (1987) developed the snapshot
POD, reducing the computation to a singular value decomposition of a matrix of flow
field snapshots. In the same year, Sirovich applied the same POD algorithm to extract a
low-dimensional feature space for human faces, which originated much of the modern
computer vision literature (Sirovich & Kirby 1987).
POD is closely related to principal component analysis (PCA), which is a cor-
nerstone method of applied statistics that is used to mine dominant correlations in
high-dimensional data. POD/PCA is also known under other names, including the
Hotelling transformation, Karhunen–Loève decomposition, and empirical orthogonal
functions. It is possible to build a simple neural network, called an autoencoder, to
compress high-dimensional data for a compact representation, shown in Figure 3.10.
This network comprises an encoder network that embeds high-dimensional data into a
latent space, followed by a decoder network that lifts from the latent space back to the

h d i 0 0 9 8 088962 4 006 bli h d li b C b id i i

 48 Data-Driven Fluid Mechanics Chapter 3

ambient high-dimensional space. When the neural network activation functions are
linear and the encoder and decoder are transposes of one another, the autoencoder can
be shown to be closely related to the standard POD/PCA decomposition. However,
the neural network autoencoder is more general, and with nonlinear activation units
for the nodes, it is possible to represent nonlinear analogues of POD/PCA, potentially
providing a better and more compact coordinate system. Milano and Koumoutsakos
(2002) were the first to apply the nonlinear autoencoder for dimensionality reduction
in fluid systems, building nonlinear embeddings to represent near wall turbulence.
Because of the universal approximation theorem (Hornik et al. 1989), which
states that a sufficiently large neural network can represent an arbitrarily complex
input–output function, deep neural networks will continue be leveraged to obtain
more effective nonlinear coordinates for increasingly complex flows. However, it is
important to note that deep learning requires extremely large volumes of training
data, and the resulting models are typically only good for interpolation and cannot be
trusted for extrapolation beyond the training data. In many modern machine learning
applications, such as image classification, the training data is becoming so vast that it
may be expected that most future classification tasks will fall under interpolation of
the training data. For example, the ImageNet data set in 2012 (Krizhevsky et al. 2012)
contained over 15 million labeled images, which sparked the current movement in
deep learning (LeCun et al. 2015, Goodfellow et al. 2016). We are still far away from
this paradigm in fluid mechanics, with neural network solutions providing targeted
and focused models that do not readily generalize. However, it may be possible in the
coming years and decades to curate large and complete enough fluid databases that
such deep interpolation may be more universally useful.

3.3.2 Clustering and Classification

Clustering and classification are core machine learning techniques, with dozens of
mature algorithms to choose from, depending on various factors, such as the size of the
data and the number of desired categories. Clustering is an unsupervised data mining
technique that identifies similarity-based groupings in the data. The most common and
simple clustering algorithm is the k-means algorithm, which partitions data instances
into k clusters; an observation will belong to the cluster with the nearest centroid,
resulting in a partition of the data space into Voronoi cells. Kaiser et al. (2014) used k-
means clustering to dramatically reduce the effective dimension of a high-dimensional
fluid mixing layer, resulting in an accurate and efficient data-driven model. This
cluster-based representation scaled with the number of clusters, not the ambient
measurement dimension of the fluid, enabling tractable Markov transition models to
describe the evolution of the flow from one state to another. Their modeling approach
is also interpretable, in the sense that each cluster centroid may be associated with a
physical flow field. In Amsallem et al. (2012) the k-means clustering approach was
used to partition phase space into distinct regions where separate local reduced-order
bases were constructed, resulting in improved stability and robustness to parameter
variations.

h d i 0 0 9 8 088962 4 006 bli h d li b C b id i i

 Brunton: Machine Learning in Fluids 49

Classification is a related supervised learning approach where the algorithm learns

to group data into several labeled categories. Colvert et al. (2018) developed a neural
network classifier to characterize bluff-body wake topology, such as 2S, 2P+2S,
2P+4S, and so on, in the wake of a pitching airfoil from local measurements of
the vorticity. Wake detection is believed to be a biologically relevant behavior, for
example to detect or avoid certain species of fish. Related work investigated the
classification performance with different types of sensors (Alsalman et al. 2018) and
other classifiers, such as the k-nearest neighbors (KNN) algorithm (Wang & Hemati
2017). Neural network classifiers may also be combined with dynamical systems
models, for example to detect flow disturbances and estimate their parameters (Hou
et al. 2019). Community detection (Meena et al. 2018) and other network-based
analyses (Nair & Taira 2015) have been applied with great success to model and
understand flows. Sparse representation for classification (Wright et al. 2009) has also
been applied to classify wakes by Bright et al. (2013) in one of the earliest examples
of classification in fluid dynamics. Extensions to DMD (Kramer et al. 2015) and
sparse sensor optimization (Brunton, Brunton, Proctor & Kutz 2016a) have since been
developed.

3.3.3 Randomized Linear Algebra and Sparse Sampling

Despite the high ambient dimension of fluid measurements and simulations, there
are often low-dimensional coherent structures that adequately describe the flow.
Randomized linear algebra (Mahoney 2011, Halko, Martinsson & Tropp 2011, Liberty
et al. 2007) is a growing field that leverages this low-dimensional structure to
accelerate computations, such as POD based on the SVD (Sarlos 2006, Rokhlin et al.
2009, Halko, Martinsson, Shkolnisky & Tygert 2011). Randomized algorithms rely on
the fact that if a large data matrix has structure, then this structure is often preserved
after multiplying with a smaller random matrix to project onto a random lower-
dimensional subspace where computations are more efficient. Randomized DMD
has also been developed recently for fluid flows (Bistrian & Navon 2016, Erichson
et al. 2019). In addition to facilitating fast numerical linear algebra, low-dimensional
structure also facilitates sparse and efficient measurements (Manohar et al. 2018). If
there are dominant patterns in the flow that are known, it is unnecessary to measure
the flow in every location. Instead, it is possible to measure at a few key locations
and infer the amplitude of these patterns, and therefore estimate the flow. Sparse
measurements are desirable for control applications, where rapid computations and
decisions reduce latency and improve robust performance. Classification and detection
of fluid events based on measurements are a crucial capability for intelligent flow
control systems, and it is desirable to perform such categorical tasks with limited or
incomplete information (Bright et al. 2013, Kramer et al. 2015, Brunton, Brunton,
Proctor & Kutz 2016a, Callaham et al. 2019).

h d i 0 0 9 8 088962 4 006 bli h d li b C b id i i

 50 Data-Driven Fluid Mechanics Chapter 3

3.3.4 Super Resolution and Experimental Flow Processing

Billions of dollars of industrial machine learning research has focused on image
science, providing robust techniques that may be directly leveraged for flow field
analysis, as flow fields closely resemble images. Super resolution is a particularly
intriguing field of research, where it is possible to statistically infer a high-resolution
image from low-resolution measurements, given sufficiently rich high-resolution
training data. At first, super-resolution sounds too good to be true, but in reality
it is simply a statistical inference based on past experience. If one is filling in an
image of a mountain landscape, and there is a low-resolution patch of green, it is
much more likely that this region is a tree or a bush, and not a green elephant;
thus, super resolution seeks to find the microstructure that is most consistent with
the neighboring low-resolution pixels and a library of high-resolution examples.
Several approaches have been developed for super resolution, for example based on a
library of examples (Freeman et al. 2002), neighbor embedding (Chang et al. 2004),
sparse representation in a library (Yang et al. 2010), and most recently based on
CNNs (Dong et al. 2014). Both fluid simulations and experiments stand to benefit
from super-resolution techniques, as improved resolution would improve our ability
to resolve relevant flow structures and dynamic mechanisms. For example, particle
image velocimetry (PIV) (Adrian 1991, Willert & Gharib 1991) would directly benefit
from improved resolution, potentially improving hardware around the world with what
essentially amounts to a software update. Large eddy simulations (LES) (Germano
et al. 1991, Meneveau & Katz 2000) suffer from similar resolution limitations, and the
ability to infer sub-grid-scale models with advanced machine learning techniques is
an area of active research. Fukami et al. (2018) recently developed a super-resolution
algorithm based on a CNN and demonstrated that it is possible to reconstruct turbulent
flows from coarse data, preserving the energy spectrum.
Machine learning may also be used for flow cleansing and imaging in experiments.
There are several learning approaches that are used for speeding up PIV (Knaak
et al. 1997) and for particle tracking (Labonté 1999), with impressive demonstrations
for three-dimensional Lagrangian particle tracking (Ouellette et al. 2006). Deep
CNNs are now being used to directly construct velocity fields from PIV image pairs,
bypassing traditional cross-correlation techniques (Lee et al. 2017). Robust principal
component analysis (RPCA) (Wright et al. 2009) has recently been applied to fluid
flow filtering (Scherl et al. 2020), making it possible to remove outliers and spurious
measurements.

3.4 Modeling Flow Dynamics

3.4.1 Linear Models: Dynamic Mode Decomposition and Koopman Analysis

Many of the classical techniques in system identification may be considered as
machine learning, as they are data-driven models that hopefully generalize beyond
the training data. A majority of classical system identification methods result in

h d i 0 0 9 8 088962 4 006 bli h d li b C b id i i

 Brunton: Machine Learning in Fluids 51

linear models (Juang & Pappa 1985, Juang 1994, Ljung 1999), which are simple
to analyze and use for control design (Dullerud & Paganini 2000, Skogestad &
Postlethwaite 2005). Despite the powerful tools for linear model reduction and control,
the assumption of linearity is often overly restrictive for real-world flows. Turbulent
fluctuations are inherently nonlinear, and often our goal is not to stabilize an unstable
laminar solution, but instead to modify the nature of the turbulent statistics. However,
linear modeling and control can be quite useful for stabilizing unstable fixed points,
for example to maintain a laminar boundary layer profile or suppress oscillations in
an open cavity (Kim & Bewley 2007, Brunton & Noack 2015). Refer to Chapter 12
for an overview of system identification and to Chapter 10 for an overview of linear
dynamical systems.
The DMD (Schmid 2010, Tu, Rowley, Luchtenburg, Brunton & Kutz 2014, Kutz,
Brunton, Brunton & Proctor 2016) is a recent technique introduced by Schmid
(2010) in the fluid dynamics community to extract spatiotemporal coherent structures
from high-dimensional time-series data, resulting in a low-dimensional linear model
for the evolution of these dominant coherent structures. DMD is an entirely data-
driven regression technique, and is equally valid for time-resolved experimental and
numerical data. Shortly after being introduced, it was shown that DMD is closely
related to the Koopman operator (Rowley et al. 2009, Mezić 2013), which is an
infinite-dimensional linear operator that describes how all measurement functions of
the fluid state will evolve in time. There have also been connections between the linear
resolvent analysis and Koopman (Sharma et al. 2016). Because the original DMD
algorithm is based on linear measurements of the flow field (i.e., direct measurements
of the fluid velocity or vorticity field), the resulting models are generally not able to
capture nonlinear transients, but are well suited to capture periodic phenomena. See
Chapter 7 for more details on DMD and Koopman operator theory.
To improve the performance of DMD, researchers are using machine learning to
uncover nonlinear coordinate systems where the dynamics appear linear (Brunton &
Kutz 2019). The extended DMD (eDMD) (Williams, Rowley & Kevrekidis 2015)
and variational approach of conformation dynamics (VAC)) (Noé & Nuske 2013,
Nüske et al. 2016) enrich DMD models with nonlinear measurements, leveraging
kernel methods (Williams, Rowley & Kevrekidis 2015) and dictionary learning
approaches (Li et al. 2017) from machine learning. Although the resulting models
are simple and linear, obtaining nonlinear coordinate systems may be arbitrarily
complex, as these special nonlinear measurement functions may not have simple, or
even closed-form solutions. Fortunately, this type of arbitrary function representation
problem is ideally suited for neural networks, especially the emerging deep learning
approaches with large multilayer networks. Deep learning is also being used to
identify these nonlinear coordinate systems, related to eigenfunctions of the Koopman
operator (Takeishi et al. 2017, Yeung et al. 2017, Lusch et al. 2018, Mardt et al. 2018,
Wehmeyer & Noé 2018, Otto & Rowley 2019). The VAMPnet architecture (Mardt
et al. 2018, Wehmeyer & Noé 2018) uses a time-lagged auto-encoder and a custom
variational score to identify Koopman coordinates on an impressive protein folding
example. The field of fluid dynamics may benefit from leveraging these techniques

h d i 0 0 9 8 088962 4 006 bli h d li b C b id i i

 52 Data-Driven Fluid Mechanics Chapter 3

from neighboring fields, such as molecular dynamics, which have similar modeling
issues, including stochasticity, coarse-grained dynamics, and massive separation of
timescales.

3.4.2 Modeling Dynamics with Neural Networks

Neural networks are among the most powerful and expressive machine learning
architectures, and they are well suited to represent the evolution of complex dynamical
systems. Even for simple continuous-time dynamics, the discrete-time flow map
that advances the system state forward by some amount of time may be arbitrarily
complex, and generally the map becomes more complex for longer advances in
time. Fortunately, deep learning is ideal for this type of arbitrary function approx-
imation. The history of neural networks in dynamics goes back decades, with
early examples modeling ordinary and partial differential equations (Gonzalez-Garcia
et al. 1998, Lagaris et al. 1998). More recently, neural networks are used for nonlinear
system identification to model fluids (Glaz et al. 2010, Zhang et al. 2012, Semeraro
et al. 2017). LSTM networks (Hochreiter & Schmidhuber 1997) are a type of
RNNs (Elman 1990) that are increasingly being used to model dynamical systems and
extreme events (Wan et al. 2018, Vlachas et al. 2018). More generally, deep learning
is providing a new approach to model systems in physics (Raissi & Karniadakis 2018,
Cranmer et al. 2019, 2020, Noé et al. 2019, Raissi et al. 2019, Greydanus et al. 2019,
Finzi et al. 2020, Zhong & Leonard 2020).
Although deep learning is a promising technique to model dynamical systems in
fluid mechanics, there are several challenges that must be addressed. Deep learning
relies on tremendous volumes of training data that are sufficiently rich so as to
represent any example that will likely be seen in the future. Further, deep learning
is inherently interpolative, simply providing a model based on the training examples.
In many computer vision and speech recognition examples, the training data sets are
so vast that nearly all future tasks may be viewed as an interpolation on the training
data, although this scale of training has not been achieved to date in fluid mechanics.
Therefore, neural network models should be used with extreme care for extrapolation
tasks. Neural networks are also prone to overfitting, and the resulting models are often
opaque and uninterpretable. Finally, the structure and scale of neural networks make it
difficult to incorporate partially known physics, such as symmetries, constraints, and
conserved quantities, although this is an area of active research.

3.4.3 Parsimonious Nonlinear Models

Parsimony is a common theme in the natural sciences, where the simplest model
describing observed data is often presumed to be the correct model. Moreover,
simplicity is a powerful technique to promote generalizability in models. Although
neural networks are a powerful branch of machine learning, it is important not
to forget that there are many other techniques that identify parsimonious models,

h d i 0 0 9 8 088962 4 006 bli h d li b C b id i i

 Brunton: Machine Learning in Fluids 53

balancing predictive accuracy with model complexity. These minimalistic models

tend to prevent overfitting and promote both interpretability and generalizability. One
of the seminal approaches to parsimonious modeling used genetic programming to
discover conservation laws and parsimonious governing equations (Bongard & Lipson
2007, Schmidt & Lipson 2009). Sparse regression has also been used to identify
ordinary (Brunton, Proctor & Kutz 2016a) and partial differential equations (Rudy
et al. 2017, Schaeffer 2017).
Sparse model identification has been successfully adapted to obtain reduced-order
models of several fluid systems, including fluid flow past a cylinder and the shear-
driven cavity flow (Brunton, Proctor & Kutz 2016a, Loiseau & Brunton 2018).
Because the SINDy algorithm is based on an iteratively thresholded least squares
regression, it is straightforward to incorporate known constraints via the Karush–
Kuhn–Tucker (KTT) equations, as used by Loiseau and Brunton to enforce energy
conserving constraints on the quadratic nonlinearities (Loiseau & Brunton 2018).
Later, it was shown that SINDy models can be obtained from sparse, physical
measurements of a system, such as time-series measurements of lift and drag
forces (Loiseau et al. 2018), producing models that may be more robust to mode
deformation, moving geometry, and varying operating conditions, compared with
Galerkin projection models. More recently, there have been several applications of
this approach in fluid dynamics for reduced-order modeling (Murata et al. 2019),
for turbulence closure modeling (Beetham & Capecelatro 2020, Schmelzer
et al. 2020), and in the related fields of electroconvection and plasma physics (Guan
et al. 2020, Kaptanoglu et al. 2020).

3.4.4 Closure Models with Machine Learning

One of the most promising applications of machine learning in fluid dynamics is to
develop advanced turbulence closure models (Duraisamy et al. 2019). The extreme
separation of scales, in both space and time, for the turbulent flows over engineering
relevant configurations (e.g., aircraft wings, submarines, etc.), makes it exceedingly
costly to resolve all scales in simulation, as shown in Figure 3.11. Even with
Moore’s law, we are decades away from resolving all scales in complex turbulent
flows. Because large, energetic coherent structures are often the most relevant for
design, it is common to truncate small scales and model their effect on the large
scales with a turbulence closure model, shown in Figure 3.12. There are several
common approaches, including Reynolds averaged Navier–Stokes (RANS) and LES
models, which make it possible to apply simulations to approximate the flow over
engineering relevant configurations. However, the underlying physics and accuracy of
these turbulence closures is often dubious, and models often require careful tuning to
match fully resolved simulations or experiments.
There are several recent efforts to incorporate data-driven methods from machine
learning to improve turbulence modeling, as nicely summarized in Kutz (2017) and
Duraisamy et al. (2019). In one prominent line of research, researchers have used
various machine learning techniques to identify and model discrepancies in the
Reynolds stress tensor between RANS model and high-fidelity simulations (Ling &

h d i 0 0 9 8 088962 4 006 bli h d li b C b id i i

 54 Data-Driven Fluid Mechanics Chapter 3

Figure 3.11 Turbulence is characterized by a large range of spatial and temporal scales, as
illustrated in this famous turbulence energy cascade. Resolving all scales for industrially
relevant flows has been notoriously challenging. See Pope (2000) for details.

Figure 3.12 The turbulence closure problem may be formulated as modeling the effect of high
frequency structures on the energy containing low frequency structures.

Templeton 2015, Ling, Kurzawski & Templeton 2016, Parish & Duraisamy 2016,
Xiao et al. 2016, Singh et al. 2017, Wang, Wu & Xiao 2017). Ling and Templeton
(2015) compare support vector machines, Adaboost decision trees, and random forests
to classify and predict regions of high uncertainty in the Reynolds stress tensor. Wang,
Wu and Xiao (2017) went on to use random forests to build a supervised model for the
discrepancy in the Reynolds stress tensor. Xiao et al. (2016) leveraged sparse online
velocity measurements in a Bayesian framework to infer these discrepancies. In a
related line of work, Parish and Duraisamy (2016) develop the field inversion and
machine learning modeling framework that builds corrective models based on inverse
modeling. This framework was later used by Singh et al. (2017) to develop a neural
network enhanced correction to the Spalart–Allmaras RANS model, with excellent

h d i 0 0 9 8 088962 4 006 bli h d li b C b id i i

 Brunton: Machine Learning in Fluids 55

Figure 3.13 Comparison of standard neural network architecture (a) with modified neural
network architecture for identifying Galilean invariant Reynold stress models (b), reproduced
with permission from Ling, Kurzawski and Templeton (2016).

performance. One particular interesting application of machine learning for turbulence

modeling was developed by Ling, Kurzawski and Templeton (2016), which involved
the first deep neural network architecture (i.e., with 8–10 hidden layers) trained to
model the anisotropic Reynolds stress tensor, as shown in Figure 3.13. What was so
interesting about this work is that they developed a novel architecture that incorporates
a multiplicative layer to embed Galilean invariance into the tensor predictions. This
provides an innovative and simple approach to embed known physical symmetries
and invariances into the learning architecture (Ling, Jones & Templeton 2016). For
LES closures, Maulik et al. (2019) employ artificial neural networks to predict the
turbulence source term from coarsely resolved quantities.

3.5 Control and Optimization

Fluid flow control is one of the grand challenge problems of the modern era, with
nearly limitless potential to enable advanced technologies. Indeed, working fluids are
central to many trillion dollar industries (transportation, health, energy, defense), and
even modest improvements to flow control could have transformative impact through
lift increase, drag reduction, mixing enhancement, and noise reduction. However, flow
control is generally a highly non-convex and high-dimensional optimization problem,
which has remained challenging despite many concerted efforts (Brunton & Noack
2015). Fortunately, machine learning may be considered as a growing body of data-
driven optimization procedures that are well suited to highly nonlinear and non-convex
problems. Thus, there are renewed efforts to solve traditionally intractable flow control
problems with emerging techniques in machine learning.
There is a considerable effort applying reinforcement learning (Sutton & Barto
2018) to fluid flow control problems (Beintema et al. 2020, Rabault & Kuhnle 2020),
as will be discussed more in Chapter 18. In particular, controlling the motion of

h d i 0 0 9 8 088962 4 006 bli h d li b C b id i i

 56 Data-Driven Fluid Mechanics Chapter 3

fish (Gazzola et al. 2014, Gazzola et al. 2016, Novati et al. 2017, Verma et al. 2018)
and of robotic gliders (Reddy et al. 2018) has experienced great strides with more
powerful reinforcement learning architectures. These techniques are also being used to
optimize the flight of uninhabited aerial vehicles (Kim et al. 2004, Tedrake et al. 2009)
and for path planning applications (Colabrese et al. 2017). It is believed that advances
will continue, as reinforcement learning with deep neural networks is an active area
of machine learning research, with considerable progress, for example in Alpha
Go (Mnih et al. 2015).
Genetic algorithms (Holland 1975) and genetic programming (Koza 1992) have
also been widely applied to fluid flow control (Dracopoulos 1997, Fleming &
Purshouse 2002, Duriez et al. 2017, Noack 2019). Neural networks have also been
used for flow control, with decades of rich history (Phan et al. 1995, Lee et al. 1997).
Recently, deep learning has been used to improve model predictive control (MPC)
efforts (Kaiser et al. 2018, Bieker et al. 2019) with impressive performance gains.
Finally, machine learning approaches have also been used for aerodynamic and
hydrodynamic shape and motion optimization (Hansen et al. 2009, Strom et al. 2017).
In another important vein of research, machine learning and sparse optimiza-
tion (Brunton & Kutz 2019) are being leveraged for sensor and actuator place-
ment (Manohar et al. 2018). Sensors and actuators are the workhorses of active
flow control. Developments in sensing and actuation hardware will continue to
drive advances in flow control (Cattafesta III & Sheplak 2011), including smaller,
higher-bandwidth, cheaper, and more reliable devices that may be integrated directly
into existing hardware, such as wings or flight decks. Many competing factors
impact control design, and a chief consideration is the latency in making a control
decision, with larger latency imposing limitations on robust performance (Skogestad
& Postlethwaite 2005). Thus, as flow speeds increase and flow structures become
more complex, it is increasingly important to make fast control decisions based on
efficient low-order models, with sensors and actuators placed strategically to gather
information and exploit flow sensitivities. Optimal sensor and actuator placement
are one of the most challenging unsolved problems in flow control (Giannetti &
Luchini 2007, Chen & Rowley 2011). Nearly every downstream control decision is
affected by these sensor or actuator locations, although determining optimal locations
amounts to an intractable brute force search among the combinatorial possibilities.
Therefore, the placement of sensors and actuators is typically chosen according to
heuristics and intuition. Figure 3.14 shows schematically how sensors and actuators
are used for flow control. Recent work by Manohar et al. (2018) has demonstrated
efficient, near-optimal sensor placement based on sparse optimization. This is a
promising area of development for flow control.

3.6 Challenges for Machine Learning in Fluid Dynamics

Applying machine learning methods to model dynamical systems from physics, such
as fluid dynamics, poses a number of unique challenges and opportunities. In physical

h d i 0 0 9 8 088962 4 006 bli h d li b C b id i i

 Brunton: Machine Learning in Fluids 57

Actuators: Sensors:

Fluid dynamic system

Offline learning

Build
dictionary
basis
Feedback control
(with sparse sensors & actuators)

Learn sparse Construct low-dim.,

sensors & actuators dynamic model

Figure 3.14 Schematic illustrating optimal sensor and actuator design for closed-loop
feedback fluid flow control. Learning sparse sensor and actuator locations is now possible
using convex optimization, and coherent structures in the fluid facilitate sparse sensing and
sparse actuation. The offline learning represents a slow optimization, enabling online
feedback (red).

systems, interpretability and generalizability are essential, as obtaining a simple

model that explains a range of phenomena is a cornerstone of physics. In fact, a well-
crafted model will yield hypotheses for new phenomena that have not been observed
before. This principle is clearly exhibited in the parsimonious formulation of classical
mechanics in Newton’s second law, and later by the Hamiltonian and Lagrangian
formulations, which provide an extremely general model framework (Abraham
et al. 1978, Marsden & Ratiu 1999).
Unsteady fluid mechanical systems may be viewed as high-dimensional dynamical
systems, with the associated challenges of varying parameters, multiscale dynamics,
noise and disturbances, latent and hidden variables, and transients, all of which
require care when using machine learning. In machine learning for dynamics, there
are often two dueling tasks of discovering unknown physics and improving models
by incorporating partially known physics. In many learning architectures, such as
neural networks, it is unclear how to incorporate partially known physics, which often
come in the form of symmetries, constraints, and conservation laws. For example, in
fluid mechanics, energy-preserving constraints are known to improve the long-time
stability and performance of nonlinear models, while standard Galerkin projection
methods often suffer from stability issues (Carlberg et al. 2017). However, a major
drawback of regression-based methods is the possible loss of existing symmetries
in the governing equations, which may otherwise be included in the physics-based
Galerkin projection methods described previously (Noack et al. 2011, Balajewicz
et al. 2013, Carlberg et al. 2015, Schlegel & Noack 2015). A notable exception is the
physics-constrained multilevel quadratic regression used to identify models in climate
and turbulence (Majda & Harlim 2012).

h d i 0 0 9 8 088962 4 006 bli h d li b C b id i i

 58 Data-Driven Fluid Mechanics Chapter 3

It is believed that the widespread application and development of machine learning

techniques to physical systems, such as fluids, will benefit both fields. Indeed, one of
the remarkable features of biological learning systems is their ability to model and
interact with a physical world, learning representations of the rules that are useful
for complex multi-objective optimizations. Pushing the limits of machine learning
by applying them to challenging physical systems may well yield transformative
breakthroughs in how we structure and train models. However, machine learning is
unlikely to displace established experimental and numerical methods, and will rather
complement these approaches, much as CFD has complemented earlier experimental
methods.
It is important to be clear about where not to employ machine learning. For
example, it is highly unlikely that machine learning will be more efficient than the
fast Fourier transform (FFT) in most applications. Similarly, when optimal or near-
optimal solutions are known, it may be best to use these solutions. However, it is rare
to have such optimal solutions in the real-world, and so machine learning will likely be
a useful alternative in these cases, where heuristics and simplified models are currently
employed.

h d i 0 0 9 8 088962 4 006 bli h d li b C b id i i

 Part II
Methods from Signal
Processing

h d i 0 0 9 8 088962 4 00 bli h d li b C b id i i
h d i 0 0 9 8 088962 4 00 bli h d li b C b id i i
 4 Continuous and Discrete Linear
Time-Invariant Systems
M. A. Mendez

This chapter reviews the fundamentals of continuous and discrete linear time-invariant
(LTI) systems with single-input single-output (SISO). We start from the general
notions of signals and systems, the signal representation problem, and the related
orthogonal bases in discrete and continuous forms. We then move to the key properties
of LTI systems and discuss their eigenfunctions, the input–output relations in the
time and frequency domains, the conformal mapping linking the continuous and the
discrete formulations, and the modeling via differential and difference equations.
Finally, we close with two important applications: (linear) models for time-series
analysis and forecasting and (linear) digital filters for multi-resolution analysis. This
chapter contains seven exercises, the solution of which is provided in the book’s
webpage.1

4.1 On Signals, Systems, Data, and Modeling of Fluid Flows

A signal is any function that conveys information about a specific variable (e.g., veloc-
ity or pressure) of interest in our analysis. A signal can be a function of one or multiple
variables; it can be continuous or discrete; and can have infinite or finite duration or
extension (that is be non-null only within a range of its domain). Signals are produced
by systems, usually in response to other signals or due to the interaction between
different interconnected subsystems. In the most general setting, a system is any entity
that is capable of manipulating an input signal and producing an output signal.
At such a level of abstraction, a vast range of problems in applied science fall
within the framework of this chapter. For a fluid dynamicist, the flow past an airfoil
is a system in which the inputs are the flow parameters (e.g., free-stream velocity and
turbulence) and control parameters (e.g., the angle of attack), and the outputs are the
drag and lift components of the aerodynamic force exchanged with the flow.
Any measurement chain is a system in which the input is the quantity to be
measured, and the output is the quantity that is measured. A hot-wire anemometer,
for example, is a complex system that measures the velocity of flow by measuring
the heat loss by a wire that is heated by an electrical current. Any signal processing
1 www.datadrivenfluidmechanics.com/download/book/chapter4.zip

h d i 0 0 9 8 088962 4 008 bli h d li b C b id i i

 62 Data-Driven Fluid Mechanics Chapter 4

technique for denoising, smoothing, and filtering can be seen as a system that takes in
input the raw signal and outputs its processed version (e.g., with enhanced details or
reduced noise).
Regardless of the number of subsystems composing a system, and whether the
system is a physical system, a digital replica of it, or an algorithm in a computer
program, the relations between input and output are governed by a mathematical
model. The derivation of such models is instrumental for applications encompassing
simulation, prediction, or control. Models can be phrased with various degrees of
sophistication, depending on the purposes for which these are developed. They usually
take the form of partial differential equations (PDEs), ordinary differential equations
(ODEs) or difference equations (DEs), or simple algebraic relation. Different models
might have different ranges of validity (hence different degrees of generalization) and
might involve different levels of complexity in their validation.
Models can be derived from two different routes, hinging on data and experimen-
tation in various ways. The first route is that of fundamental principles, based on the
division of a system into subsystems for which empirical observations have allowed
to derive well established and validated relations. In the example of the hot-wire
anemometer, the subsystems operate according to Newton’s cooling law for forced
convection, the resistive heating governed by Joule’s law, the thermoelectric laws
relating the wire resistance to its temperature, and the Ohm’s and Kirchhoff’s laws
governing the electric circuits that are designed to indirectly measure the heat losses.
Arguably, none of these laws were formulated with a hot-wire anemometer in mind.
Yet, their range of validity is wide enough to accommodate also such an application:
these laws generalize well.
In the example of the flow past an airfoil, the system is governed by Navier–Stokes
equations, which incorporates other laws such as constitutive relations for the shear
stresses (e.g., Newton’s law for a Newtonian fluid) and the heat fluxes (e.g., Fourier’s
Law for conduction). These closure relations led to the notion of fluid properties
such as dynamic viscosity and thermal conductivity and were also derived in simple
experiments that did not target any specific application.
Because of their remarkable level of generalization, we tend to see these laws
as simple mathematical representations of the “laws of nature.” Whether nature is
susceptible to mathematical treatment is a question with deep philosophical aspects.
As engineers, we accept the pragmatic view of relying on models and laws if these are
validated and useful. This is the foundation of our scientific and technical background.
The second route is that of system identification or data-driven approach, based on
the inference of a suitable model from a (usually large) set of input–output data. The
model might be constrained to a certain parametric form (e.g., with a given order in
the differential equations) or can be completely inferred from data. In the first case,
we face a regression problem of identifying the parameters such that the model fits
the data. In the second case, an algorithm proposes possible models (e.g., in Genetic
Programming, see Chapter 14) or uses such a complex parametrization that an analytic
form is not particularly interesting (e.g., in Artificial Neural Networks, see Chapter 3).
This paradigm is certainly not new. An excellent example of system identification
is the method proposed by the Swedish physicist Angström to measure thermal

h d i 0 0 9 8 088962 4 008 bli h d li b C b id i i

 Mendez: Continuous and Discrete LTI Systems 63

conductivity of a material in 1861 (Sundqvist 1991). This method consists in using

a long metal rod in which a harmonic heat wave is produced by periodically varying
the temperature at one end. The spatial attenuation of the heat wave is modeled by
the unsteady one-dimensional heat equation, leaving the thermal conductivity as an
unknown. A simple formula can be derived for the harmonic response2 of the system,
and the thermal conductivity is then identified by fitting the model to data.
Although the data-driven paradigm has a long history (Ljung 2008), its capabilities
have been significantly augmented and empowered by our increasing ability to
generate enormous amounts of data and by the powerful tools popularized by the
ongoing machine learning revolution. Some of these are described in Chapters 1 and
3. The big challenge (and the big opportunity) in big data is to combine physical
modeling with the data-driven modeling. It is the author’s opinion that the combination
of machine learning tools and the general framework of signals and systems can
accommodate the formulation of many problems of applied science and, with the
required caution, problems in fluid mechanics.
Caution is certainly required as the Navier–Stokes equations governing fluid
flows feature the entire spectra of complexities and challenges that a system analyst
could think of. In most configurations of interest, fluid flows are nonlinear: a linear
combination of input does not produce a predictable linear combination of outputs. It
is thus not possible to predict the response of the system from a dictionary of known
outputs. Fluid flows are often high dimensional and involve many scales: the amount
of information required to identify the state of a flow system is tremendous. Fluid
flows are often chaotic systems and are thus unpredictable: an infinitesimal change
in initial or boundary conditions quickly yields very different instantaneous states of
the system. Decades of fluid mechanics research aimed at models that are of reduced
order, which include engineering closure laws (e.g., turbulence modeling) or treat the
flows in terms of statistical quantities that are usually predictable.
While the Navier–Stokes equations are on the top of the ladder of complexity
in the signal-system framework, this chapter treats systems that are at the bottom
of the ladder: systems that are linear, single dimensional, deterministic, and time
invariant. These are a subclass of linear time (or Translation) invariant systems
(LTI). The interest in such a review is justified by three reasons. The first is that
LTI systems are still often encountered in practice, both in physical systems3 and
in most signal processing operations. The second is that many nonlinear systems can
be treated reasonably well as linear systems if the range of operations is close enough
to a fixed state; this is what led to the remarkable success of many linear control
methods described in Chapter 10. The third reason is that this chapter provides the
background material to understand time-frequency analysis in Chapter 5 and the multi-
resolution analysis underpinning the data decompositions in Chapter 8, the state-space
models and the linear control theory in Chapter 10, and the system identification tools
in Chapter 12.

2 The harmonic response of a system is described in Section 4.7.

3 See the previous example of Angström’s system.

h d i 0 0 9 8 088962 4 008 bli h d li b C b id i i

 64 Data-Driven Fluid Mechanics Chapter 4

Most of the material presented in this chapter can be found in classic textbooks
on signal and systems (Ljung & Glad 1994, Oppenheim et al. 1996, Hsu 2013),
signal processing (Williamson 1999, Hayes 2011, Ingle & Proakis 2011), orthogonal
transforms (Wang 2009), system identification (Ljung 1999, Oppenheim 2015), or
control theory (Sigurd Skogestad 2005, Ogata 2009). We begin this chapter by
introducing the relevant notation.

4.2 A Note about Notation and Style

In their most general form, continuous and discrete LTI systems admit multiple inputs
and respond with multiple outputs (MIMO systems) or have a single input and respond
with a single output (SISO systems)4 . This classification and the relevant notation are
further illustrated in the block diagram in Figure 4.1.
In a continuous SISO system, inputs and outputs are denoted, respectively, as con-
tinuous functions u(t), y(t) ∈ R with t ∈ R. Following the signal processing literature,
for discrete systems, these are denoted using an index notation, as u[k], y[k] ∈ R with
k ∈ Z. In this chapter, discrete and continuous signals are assumed to be linked by a
sampling process, hence the time domain is discretized as t → tk = k∆t with an index
k ∈ Z, sampling period ∆t = 1/ fs , and (constant) sampling frequency fs . Therefore,
the notation u(tk ) is equivalent to the index notation u[k], but the second makes no
link to the sampling process nor the time axis.
In a MIMO system, both the inputs and the outputs are vectors u(t), u[k] ∈ Rn I and
y(t), y[k] ∈ RnO , with nI and nO the number of inputs and outputs. MIMO systems are
better treated in state-space representation presented in Chapter 10, hence this chapter
only focuses on SISO systems.

Figure 4.1 Block diagram representation of systems, classified as single-input single-output

(SISO) or multiple-input multiple-output (MIMO) and continuous or discrete.

Signals and systems can contain a deterministic and a stochastic part. While the
focus is mostly on the deterministic part, the treatment of stochastic signals is briefly
recalled in Section 4.8.

4 Or can be a combination of the two as in MISO/SIMO systems.

h d i 0 0 9 8 088962 4 008 bli h d li b C b id i i

 Mendez: Continuous and Discrete LTI Systems 65

4.3 Signal and Orthogonal Bases

For reasons that will become clear in Section 4.4, it is convenient to represent a signal
in a way that allows decoupling the contribution of every time instance. In other
words, we seek to define a signal with respect to a very localized basis that allows for
sampling the signal at a given time. With such a (unitary) basis, the sampling process
can be done by direct comparison: for example, we say that a continuous signal u(t)
has u(3) = 2 because at time t = 3 this signal equals two times the element of the basis
that is unitary (in a sense to be defined) at t = 3 and zero elsewhere. Mathematically,
this “comparison” process is a correlation, the signal processing equivalent of the
inner product. This notion is more easily introduced for discrete signals, considered
in Section 4.3.2. Continuous signals are treated in 4.3.1.

4.3.1 Discrete Signals

Consider a discrete signal of finite duration u[k] = [1, 3, 0, 4, 5]T , that is,5 k ∈
[0, . . . , 4], which can be arranged as a column vector u ∈ R5×1 . In many applications,
a finite duration signal is assumed to be a special case of an infinite duration signal,
which is zero outside the available points. This is a common practice referred to as
zero padding that will be further discussed in Chapter 8. This signal can be written as
a linear combination of shifted unitary impulses:
1 0 0 0 0
0 1 0 0 0
         

u = 1 0 +3 0 +0 1 +4 0 +5 0 . (4.1)

         
       
0 0 0 1 0
         
0 0 0 0 1
  δ[l−0]   δ[l−1]   δ[l−2]   δ[l−3]   δ[l−4]
The elementary basis to describe such a signal is thus the set:
( l=∞
0 if l , k Õ
δk = δ[l − k] = and δ[l − k] = 1 . (4.2)
1 if l = k l=−∞

Note that two notations are introduced. δk is a vector of the same size of u, which
is zero except at l = k, where it is equal to one; δ[l − k] is a sequence of numbers
collecting the same information. In this sequence, k is the index spanning the position
of the impulse, while the index l spans the time (shift) domain. Infinite duration signals
are vectors of infinite length. For two such signals a[k], b[k] or vectors a, b, the inner
product for the “comparison procedure” is
t −1
nÕ
ha, bi = b † a = a[k] b[k] , (4.3)
k=0

5 Note that we will here use a “Python-like” indexing, that is, starting from k = 0 rather than k = 1.

h d i 0 0 9 8 088962 4 008 bli h d li b C b id i i

 66 Data-Driven Fluid Mechanics Chapter 4

T
where b† = b is the Hermitian transpose, with the superscript T denoting transposition
and the overline denoting conjugation. With such a basis, the value of the signal at a
specific location u[k] = uk can be written as
∞
Õ
uk = u[k] = hu, δk il or u[k] = u[l]δ[k − l] . (4.4)
l=−∞

The operation on the left is a correlation: for a given location of the impulse k, the
inner product is performed over the index l spanning the entire length of the signal and
the result is a scalar – the signal’s value at the index k. The operation on the right is a
discrete convolution and the result is a signal: the entire set of shifts will be spanned.
We shall return to the algebra of this operation in Chapter 8.
Note the flipping of the indices δ[l − k] in (4.2) to δ[k − l] in (4.4). In the first case,
the location of the impulse k is fixed and l spans the vector entries; in the second case,
within the summation, the time domain k is fixed and l spans the possible locations of
the delta functions6 .
The inner product in a vector space is the fundamental operation that allows for a
rigorous definition of intuitive geometrical notions such as the length of a vector and
the angle between two vectors. The length (l 2 norm) of a vector ||a|| and the cosine of
the angle β between two vectors a and b of equal size are defined, respectively, as
p p ha, bi
||a|| = ha, bi = b† a and cos(β) = . (4.5)
||a||||b||
In signal processing, the first quantity is the root of the signal’s energy7 , defined as
E{y} = ||y|| 2 while the second is the normalized correlation between two signals.
Signals with finite energy are square-
summable. If cos β = 0, two vectors are
orthogonal and two signals are uncor-
related; if cos β = 1, two vectors are
aligned and two signals are perfectly
correlated.
Notice that the projection of a vector
a onto a vector b (see Figure 4.2) is
ha, bi Figure 4.2 Projections and inner products.
ab = ||a|| cos(β) = . (4.6)
||b||
Hence if b is a basis vector of unitary length, the inner product equals the projection.
When this occurs, as in most of the cases presented in what follows, the notion of
inner product and projection are used interchangeably.
The basis of shifted impulses has a very special property: it is orthonormal. This
means that the inner product of two basis elements (in this case the shifted delta
functions) gives zero unless the same basis element is considered in which case we
6 This distinction is irrelevant for a symmetric function such as δ, but it is essential in the general case: if
δ[k − l] is replaced by δ[l + k] in (4.4), the operation is called cross-correlation.
7 Note that the notion of energy is used in signal processing for the square of a signal independently of
whether this is actually linked to physical energy.

h d i 0 0 9 8 088962 4 008 bli h d li b C b id i i

 Mendez: Continuous and Discrete LTI Systems 67

recover its unitary norm (energy). We return to this property in Chapter 8. Before
moving to continuous signals, it is worth introducing another important signal that is
linked to delta functions, namely the Heaviside step function. This is defined as
( k
1 if k ≥ l Õ
uS [k − l] = that is −→ uS [k − l] = δ[r − l]. (4.7)
0 if l < k r=−∞

The difference between two shifted step functions generates a box function u Ba,b ,
which is unitary in the range k = [a, b − 1] and is zero outside. The difference between
two step functions shifted by a single step is a delta function, that is, δ[k − l] =
uS [k − l] − uS [k − l − 1]. In the discrete setting, this is equivalent to a differentiation.
Hence, the delta function is the derivative of the step function and the summation
in (4.7) shows that the step function is an integral of the delta functions. The reader
should close this subsection with an exercise on some distinctive features of discrete
signals.

Exercise 1: Continuous versus Discrete

Consider the continuous signal y(t) = e−0.2t cos(2πt) where t is in seconds.

Assuming that the signal is sampled at f = 100Hz, and that nt = 320 points
are collected from t = 0, what is the resulting discrete signal? Plot this signal
and comment on its link with the continuous one.

4.3.2 Continuous Signals

The extension of the previously introduced notion to continuous signals brings several
complications, a detailed resolution of which is out of the scope of this chapter.
Interested readers are referred to Kaiser (2010) for a gentle introduction. The main
difficulty arises from the need to define an inner product space which can generalize
(4.3) for functions while allowing for a basis of impulses, that is, functions that are
zero at all but one point. This generalization is provided by the Hilbert space within
which the inner product of two complex-valued functions a(t), b(t) ∈ C is
∫ ∞
ha, bi = a(t) b(t) dt. (4.8)
−∞
The notion of energy (norm), correlation, and projection in (4.5)–(4.6) extends
to continuous signals using the inner product8 in (4.8). The generalization of (4.4)
requires the definition of a continuous delta function δ(t) with the same properties
as δ[k]: it is nonzero only at a given time and it is absolutely integrable. This can
be constructed as the limit of a suitably chosen function having unit area over an
infinitesimal time interval. An example is a normalized Gaussian G(t, µ, σ), which has
unit integral regardless of its standard deviation σ:
1
∫ ∞ ∫ ∞
−t 2
G(t, µ = 0, σ)dt = √ e 2σ 2 dt = 1 . (4.9)
−∞ 2πσ 2 −∞
8 Note that the upper case is used for the norms of a continuous function, that is, | |a(t) | |2 =
p
ha, ai,
with the inner product in (4.8), is the L 2 norm of the function a(t).

h d i 0 0 9 8 088962 4 008 bli h d li b C b id i i

 68 Data-Driven Fluid Mechanics Chapter 4

This function gets narrower and taller, as σ → 0, to the point at which it becomes
infinite at t = 0 and null everywhere else while still having unit area. This is the
definition of a continuous Dirac delta function, which in its shifted form is
( ∫ +∞
0 if t , τ,
δ(t − τ) = and δ(t − τ)dt = 1. (4.10)
∞ if t = τ −∞

This function is not an ordinary one, as its integration poses several technical
difficulties. Without entering into details of measure and distribution theory (Richards
& Youn 1990), we shall accept this as a generalized function that serves well our
purpose of sampling a continuous signal. From the definition in (4.10), it is easy to
derive the sifting (sampling property):
∫ ∞ ∫ ∞
y(t) = hy(τ), δ(t − τ)i = y(τ)δ(t − τ)dτ = y(t)δ(t − τ)dτ
−∞ −∞
(4.11)
∫ ∞
= y(t) δ(t − τ)dτ = y(t).
−∞

The equivalence y(τ) = y(t) in the integral results from the product y(τ)δ(t − τ) being
null everywhere but at t; then, in the last step it is possible to move y(t) outside the
integral as this is independent from the integration domain τ.
As for the discrete case, it is interesting to introduce the unitary step function and
its link with the delta function as
(
1 if t > t0,
∫ t
uS (t − t0 ) = that is, −→ uS (t − t0 ) = δ(τ − t0 )dτ. (4.12)
0 if t < t0 −∞

Notice that the step function is not defined at t = t0 . This definition is the continuous
analogue of (4.7). To show that the delta function is the derivative of the step function,
we must introduce the notion of generalized derivative. For a continuous signal
u(t), denoting as u 0 and u(n) its first and nth derivative, integration by part using an
appropriate test function ξ(t) gives
∫ ∞ ∫ ∞
ξ(t)u(n) (t)dt = (−1)n ξ (n) u(t)dt, (4.13)
−∞ −∞

where the test function ξ(t) is assumed to be continuous and differentiable at least n
times and is such that ξ(t) → 0 for t → ±∞. The first derivative of u S is
∫ ∞ ∫ ∞ ∫ ∞
0
ξ(t) u 0S (t − t0 ) dt = − ξ 0(t)u S (t − t0 )dt = − ξ (t)
−∞ −∞ t
∫ ∞ 0 (4.14)
= ξ(t0 ) − ξ(∞) = ξ(t0 ) = ξ(t) δ(t − t0 ) dt.
−∞

The first equality results from direct application of (4.13); the second from u S (t−t0 )
being zero in t < t0 . After integration, the sifting property of the delta function (4.11)
is used, and the result is obtained by equivalence of last step with the first, holding on
the fact that ξ(t) is arbitrary.

h d i 0 0 9 8 088962 4 008 bli h d li b C b id i i

 Mendez: Continuous and Discrete LTI Systems 69

4.4 Convolutions and Eigenfunctions

In a linear system, the input–output relation satisfies homogeneity and superposition

conditions. For continuous (Sc {u(t)} = y(t)) and discrete (Sd {u[k]} = y[k]) systems,
these are:

Homegeneity:
Sc {a u(t)} = aSc { u(t)} = ay(t) for all a ∈ C,
(4.15)
Sd {a u[k]} = aSd { u[k]} = ay[k] for all a ∈ C.

Superposition:
(∫ )
∞ ∫ ∞ ∫ ∞
Sc u(t) dt = Sd {u(t)}dt = y(t)dt,
−∞ −∞ −∞
( ) (4.16)
N
Õ N
Õ N
Õ
Sd un [k] = Sd {un [k]} = yn [k],
n=1 n=1 n=1

where we considered a finite summation of N inputs for the discrete case and an
infinite summation of infinitesimally close inputs for the continuous one. Combining
these properties, we see that a linear combination of inputs results in the same linear
combination of outputs:
(∫ )
∞ ∫ ∞ ∫ ∞
Sc a(τ)u(t, τ) dτ = a(τ)Sd {u(t, τ)}dτ = a(τ)y(t, τ)dτ,
−∞ −∞ −∞
( ) (4.17)
N
Õ N
Õ N
Õ
Sd an un [k] = an Sd {un [k]} = an yn [k].
n=1 n=1 n=1

A system is time-invariant (or translation-invariant, if time is replaced by space) if

the response to the input does not change over time (or space), that is,

if Sc {u(t)} = y(t) then Sc {u(t − t0 )} = y(t − t0 ) for all t0 ∈ R,

(4.18)
if Sd {u[k]} = y[k] then Sd {u[k − k0 ]} = y[k − k0 ] for all k0 ∈ Z.

A system is linear time/translation invariant (LTI) if it is both linear and

time/translation invariant. These two properties, combined with the signal
representations in (4.10)–(4.11), make the analysis of LTI systems particularly
simple: the response to any input can be fully characterized from the response to a
single impulse.
Defining as h(t) = Sc {δ(t)} the impulse response of a continuous system, the
response to any input is

h d i 0 0 9 8 088962 4 008 bli h d li b C b id i i

 70 Data-Driven Fluid Mechanics Chapter 4

(∫ )
∞ ∫ ∞
y(t) = Sc {u(t)} = Sc u(t)δ(t − τ)dτ = u(t)Sc {δ(t − τ)}dτ
−∞ −∞
(4.19)
∫ ∞ ∫ ∞
→ y(t) = u(τ) h(t − τ)dτ = h(τ) u(t − τ)dτ .
−∞ −∞

The last two integrals are equivalent forms of the convolution integral, hinging on
its commutative property. Similarly, defining h[k] = Sd {δ[k]} the impulse responses
of a discrete system, the response to any input is
( ∞ ) ∞
Õ Õ
y[k] = Sd {u[k]} = Sd u[k]δ[k − l] = u[k]Sd {δ[k − l]}
l=−∞ l=−∞
∞ ∞
(4.20)
Õ Õ
→ y[k] = u[l] h[k − l] = h[l] u[k − l] ,
l=−∞ l=−∞

having introduced the discrete convolution and its commutative property.

Note that the system response obtained via (4.19) or (4.20) is independent from the
initial state of a system. Such a response is thus a particular solution, that is, after the
transitory from the initial condition vanishes.
A special case is produced when the input is an exponential of the form u(t) = est
for continuous systems and u[k] = z k for the discrete ones, with s, z ∈ C. In what
follows, we write s in a Cartesian form as s = σ + jω and z in a polar form as
z = ρej θ ; the convenience in this is evident in Section 4.7. For the moment, note that
these two variables are the continuous and discrete complex frequencies.
In continuous systems, from (4.19), the output is
!
∫ ∞ ∫ ∞
y(t) = Sc {est } = h(τ) e(t−τ)s dτ = h(τ) e−τs dτ est = λc (s)est . (4.21)
−∞ −∞

In discrete systems, from in (4.20), the output is

∞ ∞
!
Õ Õ
y[k] = Sd {z k } = h[l] z k−l = h[l] z −l z k = λd (z)z k . (4.22)
l=−∞ l=−∞

In both cases, this result shows that LTI system responds to a complex exponential
with the same input multiplied by a complex number (λd or λc ). This number solely
depends on the complex frequencies z ∈ C and s ∈ C. Therefore, these special input
functions are eigenfunctions of the LTI operators and their complex eigenvalues are
∫ ∞
λc (s) = H(s) = h(τ) e−τs dτ, (4.23a)
−∞

∞
Õ
λd (z) = H(z) = h[l] z −l . (4.23b)
l=−∞

h d i 0 0 9 8 088962 4 008 bli h d li b C b id i i

 Mendez: Continuous and Discrete LTI Systems 71

These are, respectively, the Laplace transform and the Z-transform of the impulse
response. These are the transfer functions of the LTI systems and link input and output
in the complex frequency domain. For time varying or nonlinear systems, the notion
of transfer function is not useful. Finally, note that in discrete systems, the transfer
function is a continuous function of z.

Exercise 2: Convolution with Impulse Responses

To identify a SISO system, a step function test is carried out. A regression

analysis reveals that the response of the system to a step u S is well described
by the function
1 1 1
 
y S = − e−t cos(2 t) + sin(2 t) . (4.24)
5 5 2
Compute the response of the system to an exponential input u(t) = e−t in the
continuous and the discrete domains, assuming that the system can be sampled
at fs = 10 Hz or at fs = 3 Hz for T = 5s.

4.5 Causal and Stable Systems

In most applications of interest, signals are assumed to be null at time t < 0. This is
important for the impulse response of systems that are causal in which the impulse
response is h(t) = 0 for t < 0 and h[k] = 0 for k < 0. This means that no output
can be produced before the input, and hence the system is not anticipatory. Models
of physical systems and online data processing must be causal. On the other hand,
many data processing schemes operating off-line are not causal (e.g., zero-phase filters
described in Section 4.8).
As anticipated in the previous exercise, the convolution integral and summations in
(4.19)–(4.20) for causal systems become
∫ t ∫ t
y(t) = Sc {u(t)} = h(τ) u(t − τ)dτ = u(τ) h(t − τ)dτ, (4.25a)
0 0

k
Õ k
Õ
y[k] = Sd {u[k]} = h[l] u[k − l] = u[l] h[k − l] . (4.25b)
l=0 l=0

The upper limit is replaced by t or k since signals and impulse responses are null
for τ > t or k > l, and the lower one is replaced by 0 since both are null for t < 0 or
l < 0. The continuous and discrete transfer functions in (4.23) become
∫ ∞ Õ∞
λc (s) = H(s) = h(τ) e−τs dτ and λd (z) = H(z) = h[n] z−n . (4.26)
0− n=0

In the continuous case, t = 0 must be included in the integration; hence the lower
bound is tuned to accommodate for any peculiarity occurring at t = 0 (notably an

h d i 0 0 9 8 088962 4 008 bli h d li b C b id i i

 72 Data-Driven Fluid Mechanics Chapter 4

impulse). Nevertheless, to avoid the extra notational burden, the minus subscript in
the lower bound is dropped in what follows.
Finally, another important class of interest is that of stable systems. The stability
analysis of complex systems is a broad topic (see Chapters 10 and 13). Here, we limit
the focus to bounded-input/bounded-output (BIBO) stability. A system is BIBO stable
if its response to any bounded input is a bounded output. This requires that the impulse
response of continuous and discrete signals satisfies
∫ ∞ Õ∞
|h(t)|dt < ∞ and |h[k]| < ∞ . (4.27)
0 k=0

A different notion is that of asymptotic stability, which is related to the internal

stability of a system. A system is asymptotically stable if every initial state, in the
absence of inputs, produces a bounded response that converges to zero. In an LTI
system, asymptotic stability implies BIBO stability, but the reverse is not true.

4.6 Linear Time-Invariant Systems in Their Eigenspace

In Section 4.5, we have seen that complex exponentials are eigenfunctions of LTI
systems. Great insights on a system’s behavior can be obtained by projecting their
input–output relation onto the system’s eigenfunctions. The projection of signals into
complex exponentials leads to the Laplace transform in the continuous domain and
the Z transform in the discrete domain. This section is divided into four subsections.
We start with some definitions.

4.6.1 Laplace and Z Transforms

Laplace Transforms.
Given a continuous signal and causal signal u(t), the Laplace transforms are
∫ ∞ ∫ ∞
U(s) = L{u(t)} = u(t)e−st dt = u(t)e−st dt . (4.28)
−∞ 0

The first integral is the bilateral transform; the second is the unilateral transform.
As we here focus on causal signals (i.e., u(t < 0) = 0), these are identical. Neverthe-
less, these are different tools required for different purposes: the first is suitable for
infinite duration signals for which it can be linked to the Fourier Transform (Section
4.7); the second is developed for solving initial value problems, as it naturally handles
initial conditions.
These integrals converge, and hence the Laplace transforms exist, if u(t)e−st → 0
for t → ±∞ (only t → +∞ for the unilateral). This requires that the signal is of
exponential order, that is, grows more slowly than a multiple of some exponential:
|u(t)| ≤ Meαt . If this is the case, the range of values R{s} > α is the region of
convergence (ROC) of the transform.

h d i 0 0 9 8 088962 4 008 bli h d li b C b id i i

 Mendez: Continuous and Discrete LTI Systems 73

The reader is referred to Beerends et al. (2003), Wang (2009), Hsu (2013) for a
review of all the properties of the Laplace transform; we here focus on the key oper-
ations enabled by this powerful tool and we omit formulation of the inverse Laplace
transform, as it requires notions of complex variables theory that are out of the scope of
this chapter. The Python script E X3. PY for solving Exercise 3 provides the commands
to compute both the transform and its inverse using the Python library S YM P Y9 .
The key property of interest in this chapter is that of time derivation, which can be
easily demonstrated using integration by parts. The bilateral (L b ) and unilateral (L u )
transforms of a time derivative are
L b {u 0(t)} = sU(s) and L u {u 0(t)} = sU(s) − u(0) . (4.29)
That is differentiation in the time domain corresponds to multiplication by s in the
frequency domain; similarly, one can show that integration in the time domain corre-
sponds to division by s. Notice that no distinction between L b and L a is needed for a
system initially at rest and the bilateral transform cannot handle initial conditions.
Finally, compare the inner product in (4.28) with (4.8), taking a(t) = u(t). Note
that the Laplace transform is a projection of the signal u(t) onto an exponential basis
b L (t, s) = e−σ+jω . It is left as an exercise to show that the Laplace basis is not
orthogonal, unless σ = 0. This is the basis of the continuous Fourier transform.

Z Transform.
Given a discrete and causal signal u[k], the Z transforms are
+∞
Õ ∞
Õ
U(z) = Z{u[k]} = u[k]z−k = u[k]z −k . (4.30)
k=−∞ k=0

The same distinction on bilateral or unilateral transforms, as well as their

equivalence for causal signals, applies to the discrete transform. Observe that the Z
transform is a continuous function of z. These summations converge, and hence the
transforms exist, if |u[k]z−k | < 1. The domain within which this occurs is the ROC of
the transforms, and this is typically within a domain such that |z| > α. As before, we
avoid a review of all the properties of this transform, and the formulation of its inverse
(see Wang 2009, Hayes 2011).
The Z transform equivalent of the time differentiation and integration properties of
the Laplace transform are the time-shifting properties
Z{u[k − 1]} = z−1 U(z) and Z{u[k + 1]} = z U(z) . (4.31)
Finally, compare (4.30) with (4.3) taking a[k] = u[k] to see that the Z transform is
a projection of the signal onto a basis of powers b Z [k] = ρ−k ejθk . This basis is not
orthogonal unless ρ = 1. This is the basis of the discrete Fourier transform.
The Z transform is not suited for compression purposes, as the dimensionality of
the problem in the transformed domain is increased: the original signal has a (finite)
dimension tk , but the same information in the frequency domain is mapped onto a
continuous complex plane z.
9 see www.sympy.org/en/index.html.

h d i 0 0 9 8 088962 4 008 bli h d li b C b id i i

 74 Data-Driven Fluid Mechanics Chapter 4

4.6.2 Discrete and Continuous Frequencies

If a discrete signal is obtained by sampling a continuous one, the link between these
two transforms reveals the important impact of the sampling process on the frequency
domain. The discretization tk = k∆t creates a point-wise equivalence such that any
signal (causal or not) can be equivalently written as
∞
Õ ∞
Õ
u[k] = u[l]δ[k − l] ←→ u(tl ) = u(tl )δ(tl − l∆t) . (4.32)
l=−∞ l=−∞

With e−s tk = e−s k∆t = z k , the Laplace transform of this discrete signal is
∫ ∞ Õ ∞
L{u[k]} = u[k]δ(t − l∆t)e−st dt
−∞ l=−∞
∫ ∞ Õ ∞
= u[k]δ(t − l∆t)z−k dt (4.33)
−∞ l=−∞
∞
Õ ∫ ∞
= u[k]z −k
δ(t − l∆t)dt = Z{u[k]}.
l=−∞ −∞

The equivalence of these transforms relies on the change of variables z = es∆t .

This maps s onto the z while preserving angles in the two domains: such mapping is
a conformal mapping. Introducing s = σ + jω and z = ρejθ shows that ρ = eσ∆t and
θ = ω∆t.
Figure 4.3 shows several important features of this mapping. Observe that the
imaginary axis ω in the continuous domain is mapped onto an angular coordinate
θ ∈ [−π, π] or θ ∈ [0, 2π]. This offers yet another way of introducing the notion
of aliasing and the Nyquist–Shannon sampling theorem (encountered in Exercise 1),
arising from the fact that the frequency domain of a digital signal is periodic. Another
important observation is that the left side of the complex domain s = σ < 0 is mapped
inside the unit circle ρ = 1, while the axis s = σ = 0 is mapped on the unit circle.

Figure 4.3 The conformal mapping linking the Laplace and the Z transforms.

h d i 0 0 9 8 088962 4 008 bli h d li b C b id i i

 Mendez: Continuous and Discrete LTI Systems 75

4.6.3 The Convolution Theorem

While the convolution integrals and summations in (4.19) and (4.20) provide the
input–output relation in the time domain using the impulse response, further insights
can be obtained by analyzing this relation in the frequency domain. We here give,
without proofs10 , one of the most important results of signal processing, known as
the convolution theorem: a convolution in the time domain is a multiplication in the
frequency domain. For continuous and discrete systems, this means
(∫ )
t
Y (s) := L{y(t)} = L h(τ) u(t − τ)dτ = H(s)U(s), (4.34a)
0

( k
)
Õ
Y (k) := Z{y[k]} = Z h[l] u[k − l] = H(z)U(z) . (4.34b)
l=0

4.6.4 Differential and Difference Equations

Continuous LTI systems can be described in terms of (linear) differential equations,
while discrete LTI systems can be described in terms of (linear) difference equations.
The Laplace and the Z transforms are powerful tools to solve these equations because
of the properties in (4.29) and (4.31). In both cases, the equations have constant
coefficients and are often acronymized as LCCDE.

Differential Equations.
The general form of the LCCDE of a continuous SISO LTI system with input u(t) and
output y(t) reads
Nb
Õ Nf
Õ
an y (n)
(t) = bn u(n) (t), (4.35)
n=0 n=0

where N f ≥ Nb is the order of the system11 . The coefficients an are called feedback
coefficients; the coefficients bn are feedforward coefficients.
The LCCDE provides an implicit representation of a system since the input–
output relation can be revealed only by solving the equation. Introducing the Laplace
transform in a LCCDE is an operation similar to the Galerkin projection underpinning

10 See Wang (2009) for more details.

11 The condition N f ≥ Nb is necessary to ensure that the system is realizable, that is, both stable and
causal. More about this is provided in footnote 12.

h d i 0 0 9 8 088962 4 008 bli h d li b C b id i i

 76 Data-Driven Fluid Mechanics Chapter 4

reduced-order modeling (ROM, see Chapters 1 and 14). Recalling that the Laplace
transform is a projection onto the basis b L (t), (4.35) leads to:
Nb
DÕ Nf
E DÕ E
an y (n) (t), b L (t) = bn u(n) (t), b L (t)
n=0 n=0
Nb
Õ n o ÕNf n o
→ an L y (n) (t) = bn L u(n) (t)
n=0 n=0
(4.36)
Nb
Õ  Nf
Õ 
→ Y (s) an s (n)
= U(s) bn s (n)
,
n=0 n=0
ÍN f ÎN f
Y (s) b s(n)
n=0 n
bN f n=0
(s − zn )
H(s) = = ÍN = .
U(s) b
an s(n) a Nb Î N b
(s − pn )
n=0 n=0
The transfer function of LTI systems is a polynomial rational function of s, with the
coefficients of the polynomials being the coefficients of the LCCDE. In the factorized
form, zn and pn are, respectively, the zeros and the poles of the system12 . Note that
since the coefficients an , bn are real, these can either be purely real or appear in
complex conjugate pairs. These coefficients have a straightforward connection with
the LCCDE, which can immediately be recovered from the transfer function. The
zeros zn are associated to inputs ezn t in which the transfer function is null and thus
leads to no output; the poles pn corresponds to resonances, inputs e pn t in which the
transfer function is infinite and leads to the blowup of the system.
The poles are eigenvalues of the matrix A, advancing a linear system in its state-
space representation (see Chapters 10 and 12). In a stable system, poles are located in
regions of the s-plane that are “not accessible” by any input, that is, outside the ROC
of the transfer function. Defining the ROC of H(s) as R{s} > α, and observing that
the poles are by definition outside the ROC, stability is guaranteed if α = 0, that is, if
the ROC includes the imaginary axis. This is equivalent to imposing that all the poles
are located in the left side of the s-plane, that is, R{pn } < 0 ∀n ∈ [0, N f ]. This result
can also be derived from the BIBO stability condition in (4.27).

Difference Equations.
In the discrete case, the general form of LCCDE associated to SISO LTI systems with
input u[t] and output y[t] reads
Nb
Õ Nf
Õ Nb
Õ Nf
Õ
an y[k −n] = bn u[k −n], that is, y[k] = b∗n u[k −n]− an∗ y[k −n]. (4.37)
n=0 n=0 n=0 n=1

12 A transfer function that has more zeros than poles (i.e., N f > Nb ) is said to be improper. In this case,
lim s→∞ |H(s)| = +∞, which violates stability: this implies that at large frequencies, a finite input can
produce an infinite output. Moreover, after the polynomial division, the transfer function brings
polynomial terms in s. The inverse Laplace transform of these are (generalized) derivatives of the delta
functions; hence the corresponding impulse response h(t) = L −1 {H(s)} violates causality.

h d i 0 0 9 8 088962 4 008 bli h d li b C b id i i

 Mendez: Continuous and Discrete LTI Systems 77

The order of the system13 is max(Nb , N f ). The form on the right plays a fundamental
role in filter implementation, time-series analysis, and system identification and is
known as recursive form of the differnce equation. Note that the feedback and
feedforward coefficients in the recursive form are simply an∗ = an /a0 and b∗n = bn /a0 ,
respectively; hence the coefficient b∗0 is the static gain of the system.
As for the continuous case, projecting (4.37) onto the Z basis b Z (t) via the Z
transform and using (4.31) yields the transfer function of a discrete system:
Nb
DÕ Nf
E DÕ E
an y[k − n], b Z [k] = bn u[k − n], b Z [k]
n=0 n=0
Nb
Õ n o ÕNf n o
→ an Z y[k − n] = bn Z u[k − n]
n=0 n=0
(4.38)
Nb
Õ  Nf
Õ 
→ Y (z) an z −n
= U(z) bn z −n
,
n=0 n=0
Í N f −1 Î N f −1
Y (z) n=0
bn z−n
b0 n=0
(1 − ζn z−1 )
H(z) = = Í N −1 = z Nb −N f Î N ,
U(z) b
an z−n a0 b −1
(1 − πn z−1 )
n=0 n=0
where ζn and πn are, respectively, the zeros and poles of the discrete transfer function.
Observe that the factored form of the discrete transfer function is usually given in
terms of polynomials of z −1 rather than z.
The link between zero and poles in continuous and discrete domains is given by
the conformal mapping in Figure 4.3. In the absence of inputs, the poles control the
evolution of a linear system from its initial condition (i.e., the homogeneous solution
of the LCCDE). The Dynamic Mode Decomposition (DMD) introduced in Chapter 7
is a powerful tool to identify the poles πn of a system from data, and to build linear
reduced-order models by projecting the data onto the basis of eigenfunctions z πn k .
Finally, in analogy with the continuous case, a discrete system is stable if its poles
are outside the ROC of the transfer function. Defining the ROC of H(z) as |z| ≥ α,
one sees that this occurs if α = 1: the ROC includes the unit circle and hence all the
poles have |πn | < 1. This can be derived from the BIBO stability condition in (4.30).

Exercise 3: Transfer Function Analysis

Consider the system in Exercise 2. Compute the transfer function and the
system output from the frequency domain, then identify the LCCDE governing
the input–output relation. Then, assuming that a discrete system is obtained
sampling the continuous domain, derive a recursive formula that mimics the
input–output link of the continuous system. Test your result for a sampling
frequency of fs = 3 Hz and fs = 10 Hz.

13 Note that in (4.37) the restriction Nb ≥ N f is not needed to enforce causality: by construction, the
output y[k] only depends on past information.

h d i 0 0 9 8 088962 4 008 bli h d li b C b id i i

 78 Data-Driven Fluid Mechanics Chapter 4

4.7 Application I: Harmonic Analysis and Filters

BIBO stability guarantees that the output produced by a stationary input is also
stationary. It is thus interesting to consider only the portion of the complex planes s and
z associated to infinite duration signals, that is, s = jω and z = ejθ . These correspond
to harmonic eigenvalues of the LTI system, hence lead to a harmonic response. From
Laplace and Z transforms, we move to continuous and discrete Fourier transforms
in Section 4.7.1. A system that manipulates the harmonic content of a signal is a
filter; these are introduced in Section 4.7.2 along with their fundamental role in multi-
resolution decompositions.

4.7.1 From Laplace to Fourier

Consider the bilateral Laplace and Z transform of a signal along the imaginary axis
s = jω and the unitary circle z = ejθ , respectively,
∫ ∞
U(jω) = FC {u(t)} = u(t)e−jωt dt, (4.39a)
−∞

+∞
Õ
U(jθ) = FD {u[k]} = u[k]e−jθ k . (4.39b)
k=−∞

These are the continuous (CT) and the discrete (time) Fourier Transforms (DTFT).
Both are continuous functions, with the second being periodic of period 2π because
of the conformal mapping introduced in Figure 4.3. Comparing these to (4.28) and
(4.30) shows that the bilateral Laplace and Z transforms are the Fourier transforms
of u(t)e−σt and u[k]ρ−k . Without these exponentially decaying modulations, the
conditions for convergence are more stringent: signals must be absolutely integrable
and absolutely summable14 .
The main consequence is that infinite duration stationary signals do not generally
admit a Fourier transform. This explains why the manipulations of these signals by
an LTI system are better investigated in terms of some of their statistical properties,
such as autocorrelation or autocovariance, as illustrated in Section 4.8. A special
exception are periodic signals for which (4.39a) and (4.39b) lead to Fourier series,
and the problem of convergence becomes less stringent.
In stable continuous and discrete LTI systems, satisfying (4.30), the impulse
response always admits Fourier transform: these can be obtained by replacing s = jω
and z = ejθ in the transfer function. This leads to frequency transfer functions, which
are complex functions of real numbers15 (ω or θ), customarily represented by plotting
log(|H(x)|) and arg(H(x)) versus log(x) in a Bode plot, with x = ω or x = θ.
14 This condition is sufficient but not necessary: some non-square integral functions do admit a Fourier
transform. Important examples are the constant function u(t) = 1 or the step function uS (t). Moreover,
note that the Fourier transform can be obtained from the Laplace and Z transform, only for signals that
are absolutely integrable or summable. For instance, the Laplace transform of e αt with α > 0 has ROC
s > α, while the Fourier transform does not exist.
15 These are often called real frequencies as opposed to the complex frequencies s and z.

h d i 0 0 9 8 088962 4 008 bli h d li b C b id i i

 Mendez: Continuous and Discrete LTI Systems 79

The modulus of the frequency transfer function is the amplitude response; its argument
is the phase response.
If the Fourier transform (or series) exists for both inputs and outputs, the properties
of the Laplace and Z transform apply: the harmonic contents of the output is Y (jω) =
H(jω)U(jω) in the continuous domain; Y (jθ) = H(jθ)U(jθ) in the discrete one.
Discrete signals of finite duration u(tk ) = u ∈ Rnt ×1 = u[k], with k ∈ [0, nt − 1], are
usually extended to infinite duration signals assuming periodic boundary conditions.
The frequency domain is thus discretized into bins θ n = n∆θ with n ∈ [0, n f − 1]
and ∆θ = 2π/nF . The mapping to the continuous frequency domain, from Figure 4.3,
gives fn = 2πωn = n fs /n f , with fs = 1/∆t the sampling frequency. With both time
and frequency domain discretized, the Fourier pair are usually written as
nt −1 n f −1
1 Õ −2πj nn k 1 Õ 2πj n k
U[n] = √ u[k]e f ⇐⇒ u[k] =
√ U[n]e n f . (4.40)
nt k=0 nt n=0
The equations in (4.49) are, respectively, the discrete Fourier transform (DFT) and
√
its inverse. Note that the normalization 1/ nt is used for later convenience: we see
in Chapter 8 that (4.49) can be written as matrix multiplications with the columns
of the matrix being orthonormal vectors. Finally, if n f = nt and nt is a power of 2,
this multiplication can be performed using the famous FFT (fast Fourier transform)
algorithm (see Loan 1992), reducing the computational cost from nt2 to nt log2 (nt ).
An excellent review of the DFT is provided by Smith (2007b).

Exercise 4: Frequency Response Analysis

Consider the discrete system derived in Exercise 3, but now assume that the
static gain is unitary. Compute the frequency transfer function of this system
and show that this can be seen as a low-pass filter. Study how the frequency
response changes if the coefficients a1 or a2 are set to zero. Finally, derive
the system that should have the complementary transfer function and show its
amplitude response. Is this response also complementary?

4.7.2 Multiresolution Analysis and Digital Filters

Filters are at the center of most signal processing applications, and the theory behind
their design is a vast subject (see Smith 2007a). Among the essential applications
discussed in this book are feedback control design (Chapter 10), multiresolution
analysis (MRA) and wavelet decomposition (Chapter 5), and multiscale proper
orthogonal decomposition (Chapter 8).
In feedback control, a controller manipulates the feedback coefficients of an
LCCDE by introducing a control input, which is a function of the output. Therefore,
filter design methods can be used to design the actuation such that the transfer function
of the controlled system rejects certain disturbances (see Bode design methods
(Distefano 2013)).

h d i 0 0 9 8 088962 4 008 bli h d li b C b id i i

 80 Data-Driven Fluid Mechanics Chapter 4

In MRA, filters are used to decompose signals. While the DFT represents a signal as
a linear combination of harmonics, MRA represents it as a combination of frequency
bands called scales. A packet of similar frequencies can be assembled into bases called
wavelets, hence the connection to Chapter 5.
The MRA partitions the spectra of signal into n M scales, each taking a portion of
the signal’s content in bands [0, f1 ], [ f1, f2 ], . . . , [ fM−1, fs /2]. In Chapter 8, these will
be identified by a frequency splitting vector FV = [ f1, f2, f3, . . . , fM−1 ].
The MRA of a discrete signal can be written as
M
(M ) M
Õ Õ Õ
u[k] = sm [k] = F −1 Hm ( fn )U( fn ) with |Hm ( fn )| = 1 , (4.41)
m=1 m=1 m=1

where sm is the portion of the signal in the scale m, within the frequency range
fn ∈ [ fm−1, fm ] and Hm ( fn ) is the transfer function of the filter that isolates that
portion. Therefore, |Hm ( fn )| ≈ 1 for fn ∈ [ fm−1, fm ] and |Hm ( fn )| ≈ 0 otherwise.
The assumption on the right enables a lossless decomposition.
The MRA requires the definition of one low-pass filter for the range [0, f1 ], one
high-pass filter for the range [ fM−1, fs /2], and M − 2 band-pass filters. Because these
are complementary, all these filters can be obtained from a set of low-pass filters, as
described at the end of this section. Therefore, to learn MRA, one should first learn
how to construct a low-pass filter with a given cutoff frequency fc .
We now focus on the two main families of filters and the most common design
methods. Let us consider a specific example, with fs = 2k Hz and fc = 200 Hz. In the
digital frequency domain, we map the sampling frequency to θ s = 2π and the cutoff
to θ c = π/5.
The transfer function of the ideal low-pass filter is

e−jαd θ if |θ| ≤ θ c ,
(
Hid (jθ) = (4.42)
0 if θ c < |θ| ≤ π.

This leads to the impulse response

∫ π
1 sin[k − αd ]θ c
hid [k] = H(jθ)e−jkθ dθ = . (4.43)
2π −π π(k − αd )
The need for a delay α is evident after introducing FIR filters. Notice that having a
linear phase αθ delays the input without distorting its waveform. The modulus of this
frequency response and a portion of its transfer function are shown in Figure 4.4(a)
with continuous black curves.
Such an ideal filter is not realizable in the time domain because its impulse response
is not causal (h[k] , 0 for k < 0) and is not absolutely summable. The ideal
constraints must be relaxed. The most popular categories are IIR and FIR filters.

1) Infinite Impulse Response Filters (IIR)

These filters are based on a continuous function that mimics the ideal low-pass
filter. The most common are Butterworth, Chebyshev, and elliptic filters (Hayes 2011,

h d i 0 0 9 8 088962 4 008 bli h d li b C b id i i

 Mendez: Continuous and Discrete LTI Systems 81

1.0 0.2
Ideal Ideal
IIR IIR
0.8 FIR FIR
0.1
|H(fn )|
0.6

h(tk )
0.4
0.0
0.2

0.0
−0.1
0 100 200 300 400 0.00 0.01 0.02 0.03 0.04 0.05
f [Hz ] tk [s]

Figure 4.4 (a): Absolute value of the frequency response of an ideal (black continuous), an IIR
(red dash-dotted) and an FIR (blue dashed) low-pass filter. (b): impulse responses associated to
the frequency responses on the left. In the case of the FIR, the response is shifted backward by
αd .

Oppenheim & Schafer 2009). These filters have no zeros and N poles, with N the filter
order, equally spaced around the unit circle.
Once these poles are computed, the continuous frequency response function can
be readily obtained in its factor form and the last step consists in identifying the
associated recursive formula as in Exercise 4. However, note that mapping from s to z
is usually performed using the bilinear transform16 rather than the standard mapping
z = e∆ts that is used in Exercise 4, since this has the advantage of mapping fs /2 to
π and thus prevents aliasing. The nonlinearity in the bilinear transform results in a
wrapping of higher frequencies so the correct cutoff frequency should first be pre-
warped to account for the distortion in the frequency calculation17 .
Software packages such as S CI P Y in P YTHON or M ATLAB offer the functions
BUTTERWORTH to design a Butterworth filter with given order and cutoff frequency
(see P YTHON script E X 5. PY).
The red curves in Figure 4.4(a) show the amplitude response and the impulse
response of a Butterworth filter of order 11. The main advantage of these filters is their
capability of well approximating the ideal filter using a limited order, which requires
storing few coefficients in their recursive formulation. On the other hand, these filters
tend to become unstable as the order increases (and the poles approach the unit circle).
Moreover, their phase delay is generally not constant, and this potentially introduces
phase distortion. Finally, note that since the impulse response of these filters is infinite,
these cannot be implemented in the time domain via simple convolution, but via the
recursive solution of the filter’s LCCDE.
16 which reads
!
2 1 − z −1 1 + ∆t/2 s
s= ←→ z = .
∆t 1 + z −1 1 − ∆t/2 s

17 The pre-warp can be achieved using fc0 = fs /π tan π fc / fs . Therefore, if the desired cutoff frequency

is fc = 200 Hz with a sampling frequency fs = 1000 Hz, the filter should target a cutoff frequency of
fc0 = 231.26 Hz to compensate for the warping due to the bilinear transform.

h d i 0 0 9 8 088962 4 008 bli h d li b C b id i i

 82 Data-Driven Fluid Mechanics Chapter 4

2) Finite Impulse Response Filters (FIR)

These filters are constructed in the discrete domain and have no poles (no feedback
coefficients in their recursive formulation). This leads to a finite impulse response. The
classic design method is the windowing technique, which consists in multiplying the
impulse response of the ideal filter in (4.42) by a window w[k], which is zero outside
the interval 0 ≤ k ≤ N, with N the filter order.
Taking N as an odd number, these windows are symmetric about the midpoint, that
is, w[k] = w[N − k]; this results in the lag α = (N − 1)/2 in the output with respect to
the input. The need for a lag in the ideal filter in (4.42)–(4.43) is now clear: if α = 0,
the windowed impulse response is centered in 0, and the filter is noncausal. If the
filtering is performed “off-line,” it is possible to obtain a zero-phase filter by centering
the windowed impulse response in k = 0.
Common functions are the Hanning, Hamming, or Blackman and Kaiser windows.
The windowing in the time domain corresponds to a convolution in the frequency
domain between the ideal filter and the Fourier transform of the window function. This
smooths the transition from the band-pass to the band-stop region. Software packages
as S CI P Y in P YTHON or M ATLAB offer the functions FIRWIN and FIR 1 to design FIR
filters with a given order, cutoff frequency and window function.
In addition to the linear phase response, these filters are also always stable because
of the lack of poles. Moreover, the finite length of their impulse response enables their
implementation via convolution. Note, however, that FIR filters require much larger
order to achieve performances comparable with IIR filters. Figure 4.4(b) shows, in
the dashed blue line, the amplitude function and the corresponding impulse response
(shifted by (N − 1)/2) of an FIR filter designed using a Hamming window of order
N = 111. The higher the filter order, the larger the window multiplying the ideal
impulse response, the more this filter approach is the ideal one. On the other hand,
increasing the filter order increases the sensitivity of the filter to the finite duration of
the signal.
An FIR formulation makes the calculation of complementary filters particularly
simple thanks to the constant phase response, solely linked to the filter order. To
illustrate this, consider a signal u[k], and its DFT U[n]. Let us filter this signal
with a low-pass filter to obtain u L [k] using an FIR filter with frequency transfer
function H L = |H L |e−iαk . The filter operation in the frequency domain reads U L =
|H L |e−iαθ U. Let UH = |HH |e−iαθ U denote the high-pass filter that gives the signal
u H [k] having complementary spectra (i.e., |H L | + |HH | = 1) and same order and
thus same phase delay. Using the shifting properties of the Fourier transforms, the
link between the low-pass and the high-pass counter parts in the time and frequency
domain sets is
k Domain : u H [k] = u[k + α] − u L [k],
θ Domain : HH (iθ)U(iθ) = U(iθ)eiαθ − H L (iθ)U(iθ) (4.44)
→ HH (iθ) = e iαθ
− H L (iθ) .
Note that the backward shifting u[k + α] in the time domain cancels the phase delay
produced by the low-pass filter before performing the subtraction. It is easy to show

h d i 0 0 9 8 088962 4 008 bli h d li b C b id i i

 Mendez: Continuous and Discrete LTI Systems 83

that because of the linearity of the convolution, the impulse responses of complemen-
tary high-pass (h H ) and low-pass (h L ) filters are linked by18 δ[n] = h L [n] + h H [n].
Finally, we close with the practical implementation of MRA in “off-line” conditions
for which it is possible to release the constraints of causality and use zero-phase filters.
These are usually implemented by operating on the signal twice (first on u[k] and then
on u[−k]), to artificially cancel the phase delay of the operation. In S CI P Y and in
M ATLAB this is performed using the function FILTFILT.
If the phase delay is canceled, complementary filters can be computed by taking
differences of the frequency transfer functions (which become real functions). There-
fore, if the first scale with band-pass [0, f1 ] is identified by the frequency transfer
function H1 ( f ) = H L ( f , f1 ), the second scale with band-pass [ f1, f2 ] is identified by a
filter with transfer function H2 ( f ) = H L ( f , f2 ) − H L ( f , f1 ). The transfer function of
the general band-pass filter is Hm ( f ) = H L ( f , fm ) − H L ( f , fm−1 ), while the last scale
is identified by the high-pass filter with HM ( f ) = 1 − H L ( f , fM ).
This set of cascaded filters is known as a filter bank and is at the heart of the pyramid
algorithm for computing the discrete wavelet transform (Strang 1996, Mallat 2009),
where it is combined with sub-sampling at each scale. The general architecture of this
decomposition is summarized in Figure 4.4. Observe that at the limit at which all the
frequency bands become unitary, the MRA becomes a DFT.

Figure 4.5 Pyramid-like algorithm to compute the MRA of a signal. Each band-pass scale is
computed as the difference of two low-pass filters and the terms in blue are preserved to form
the summation in (4.41). The graph on the right shows a pictorial representation of the
partitioning of the signals spectra.

18 Note that this is not the only method of obtaining a high-pass filter from a low-pass filter: another
approach is to reverse the frequency response HL , flipping it from left to right about the frequency fs /4
for f > 0 and from right to left about − fs /4 for f < 0 (see Smith (1997)). The impulse response of the
resulting high-pass filter is hH [n] = (−1) n hL [n]. The two methods are equivalent if the cutoff
frequency separating the transition bands is fs /4. This is the case encountered when performing MRA
via dyadic wavelets as discussed in Chapter 5.

h d i 0 0 9 8 088962 4 008 bli h d li b C b id i i

 84 Data-Driven Fluid Mechanics Chapter 4

Exercise 5: Multi-Resolution Analysis

Consider the synthetic signal

−(t k −τ1 )2 −(t k −τ2 )2
u(tk ) = a1 sin(2π f1 tk )e b1
+ a2 sin(2π f2 tk )e b2
(4.45)
+ a3 sin(2π f3 tk ) + N (0, 0.2),
with a1 = 2, a2 = a3 = 1, f1,2,3 = 1, 20, 90 (in Hz), τ1 = 1, τ2 = 2.2 (in
s) and b1 = b2 = 0.05. N (0, 0.2) denotes Gaussian noise with zero mean
and standard deviation 0.2. The signal is sampled over nt = 4 096 points at a
sampling frequency of fs = 1k Hz.
(1) Compute “by hand” the impulse response h L of a low-pass FIR filter of
order NO = 5, with cutoff frequency fc = 10 Hz using a Hamming window.
Check your answers using Python’s function FIRWIN.
(2) Prepare a filter h L of order 511 using FIRWIN. Apply this to the signal using
four methods: (a) direct convolution, (b) FFT-based convolution with S CIPY’s
function FFTCONVOLVE, (c) FFT-based convolution using FFT and IFFT, and
(d) a recursive implementation solving the filter’s LCCDE.
(3) Construct a filter bank with frequency splitting vector FV =
[10, 70, 100, 300] and show the portions of the signal within the identified
scales. Use filters with Hamming window and an order of NO = 511.

4.8 Application II: Time-Series Analysis and Forecasting

LTI systems are the simplest model in time-series analysis and forecasting. In these
applications, treating signals and systems as fully deterministic is too optimistic, and
it is thus essential to consider stochastic signals: predictions have a certain probability
range (see Guidorzi 2003, Brockwell & Davis 2010). This section briefly reviews the
main features of stochastic signals and systems in Section 4.8.1. Section 4.8.2 reviews
the basic tools for forecasting, using classic linear regression. Only the discrete
domain is considered. More advanced techniques are discussed in Chapter 12.

4.8.1 Stochastic Linear Time-Invariant Systems

A stochastic signal (or the stochastic portion of a signal) is a member of an ensemble
of signals characterized by a set of probability density functions. For a comprehensive
review of stochastic signals, the reader is referred to classic textbooks (Ljung 1999,
Hsu 2013, Oppenheim 2015). Here, we briefly recall how LTI systems manipulate
stochastic signals.
The notion and the role of the impulse response remain the same as for deterministic
signals: given an input (stochastic) discrete signal, the response of the system is
governed by the convolution sum in (4.21). On the other hand, the notion of frequency

h d i 0 0 9 8 088962 4 008 bli h d li b C b id i i

 Mendez: Continuous and Discrete LTI Systems 85

spectra requires some adaptation, as stochastic signals do not generally admit a Fourier
transform and focus must be placed on properties that are deterministic also in a
stochastic signal. These are the statistical properties.
Accordingly, the time-invariance in LTI systems is extended in terms of invariance
of the statistical properties. This is linked to the notion of stationarity. Stationarity
can be weak or strong. A stochastic signal v[k] is stationary in a strict sense (strong
stationarity) if its distributions remain invariant over time. Weak stationarity (or
stationarity in a wide-sense) requires that only its time average µv and autocorrelation
rvv [m] of a signal v are time-invariant. These are defined as
µv = E{v[k]} and rvv [m] = E{v[k]v[k + m]} , (4.46)
with E the expectation operator.
In the analysis of the LTI system’s response to stochastic signals, the link between
a specific input and the corresponding output is not particularly interesting. Instead,
we focus on the link between the statistical properties of the input and the output. In
particular, we consider how the properties in (4.46) are manipulated. Let yv [k] be the
response of the system to the stochastic signal uv [k]. The expected (time average of
the) output µy is
( ∞ ) ∞ ∞
Õ Õ Õ
µy = E h[l]uv [k − l] = h[l]E{uv [k]} = µv h[l] . (4.47)
l=−∞ l=−∞ l=−∞

This is a direct application of the homogeneity (4.16) and superposition (4.17). We

thus see that in a BIBO stable system (satisfying (4.27)) the output average is finite if
the input average is finite19 . The input–output relation for the autocorrelation function
has a more involved derivation, here omitted (see Oppenheim et al. 1996). Given ruu
and ryy the input and the output autocorrelations, one retrieves
∞
Õ ∞
Õ
ryy [m] = ruu [m − l]rhh [l] where rhh [k] = h[k]h[l + k] . (4.48)
l=−∞ l=−∞

The sequence rhh [k] is the autocorrelation of the impulse response and the operation
on the left is a convolution. In words: the autocorrelation of the output is the
convolution of the autocorrelation of the input with the autocorrelation of the impulse
response. This equation extends the convolution link in (4.21) to the autocorrelation
functions. These functions admit Fourier transform, so the convolution theorem can
be used to see the link in the frequency domain:
Ryy (jω) = Chh (jω)Ruu (jω) , (4.49)
where Ryy (jω), Chh (jω) and Ruu (jω) are the Fourier transform of ryy [k], rhh [k], and
ruu [k], respectively. These are the power-spectral densities of y, h and u. Hence we
see that an LTI system acts on the frequency content of the autocorrelation function of
a stochastic signal.

19 We also see why a high-pass filter has

Í
h[k] = 0 while a low-pass filter has
Í
h[k] = 1.
k k

h d i 0 0 9 8 088962 4 008 bli h d li b C b id i i

 86 Data-Driven Fluid Mechanics Chapter 4

4.8.2 Time-Series Forecasting via Linear Time-Invariant Systems

Consider the explicit form of the LCCDE of an LTI in (4.37) and assume that the input
signal (u) has both a deterministic (ud ) and a stochastic (us ) part (i.e., u = ud + us ).
Because of homogeneity, the output of the LTI system also has a deterministic (yd )
and a stochastic part (ys ). We could split these as follows:
Nb
Õ Nf
Õ
y[k] = yd [k] + ys [k] = b∗n ud [k − n] − an∗ yd [k − n] + ys . (4.50)
n=0 n=1

Many models can be obtained depending on the assumptions on us , and hence

ys (see Guidorzi 2003, Brockwell & Davis 2010, Nielsen 2019). For example, the
stochastic part ys can be taken as white noise with zero average or as the output of
a moving average filtering20 of white noise. Any other filter can be used to allow
controlling and/or modeling the frequency content of the stochastic contribution using
(4.48) and (4.49).
To illustrate the main steps of time-series forecasting, let us consider the simplest
approach of ys being white noise. If the system is known (i.e., the coefficients
an∗ and b∗n are known), the recursive equation (4.50) can be written as a matrix
multiplication21 . Assume that we have collected nt samples of the input u. Let
u(−l)
d
, y(−l)
d
∈ Rnt ×1 be the vectors collecting the deterministic inputs and outputs
shifted backward by a lag l. Then, the matrix form of (4.50) is
 b0 
 . 
 
 .. 
 |   | | | | |   
 b Nb 
 
y  = u  + ys . (4.51)
 (0)   (0)
u(−1) ··· u(−N f ) y(−1) ··· y(−Nb ) 
 a1 
   
 |   | | | | | 
 .. 
    
 . 
a N F 
 
 
We define H := [u(0), . . . , u(−N f ), y(−1), . . . , y(Nb ) ] the Hankel matrix of the LTI
system and w = [b0, . . . , b Nb , a1, . . . , a N f ]T the vector of coefficients. The output of
the LTI system is
y(0) = Hw + ys . (4.52)
Time-series forecasting via LTI systems begins with system identification. An
excellent tutorial on the topic is provided by Semeraro and Mathelin (2016). The
first goal is to identify the set of coefficients w from the input/output vectors. The
stochastic part is considered as noise and the determinist part is our expectation. We
thus seek to solve the system y(0) ≈ Hw. Like most regression problems, this problem
is ill-posed: H is rectangular and there is no guarantee that a unique solution exists.
Like all linear regression problems, the solution is found by minimizing a regularized
cost function of the form
20 A “moving average” filter is a filter with constant impulse response.
21 This needs to be evaluated from the first to the last entry of y(0) .

h d i 0 0 9 8 088962 4 008 bli h d li b C b id i i

 Mendez: Continuous and Discrete LTI Systems 87

J(w) = ||y(0) − Hw||22 + αR(w) , (4.53)

with α ∈ R acting as a smoothing parameter and R(w) a regularizing function.22
Classic choices are R(w) = ||w||2 (l 2 penalty), or R(w) = ||w||1 (l 1 penalty), or a
combination of the two. The first is known as Tikhonov regularization, the second
as LASSO regularization, and the third as Elastic Net. These classic tools from
machine learning (see Vladimir Cherkassky 2008, Murphy 2012, Bishop 2016) are
also employed in Chapter 12.
The reader should notice that the regression method can be generalized easily:
one could replace the predictive equation (4.52) by a more complex model (e.g., an
artificial neural network or template of polynomial nonlinearities as in Chapter 12)
and minimize a cost function like (4.53) using an arsenal of optimization strategy.

Exercise 6: Time-Series Forecasting via LTI Systems

Generate a synthetic stochastic system with coefficients (a1, a2 ) = (−1/2, 0)

and (b0, b1, b2 ) = (2, 1.8, −1.2). Consider a deterministic input signal u[k] =
u(tk ) = sin(8πtk ) + 3 exp(−(tk − 3)2 /2) with tk = k∆t, ∆t = 0.02 and k ∈
[0, 1999]. The stochastic contribution ys [k] is white noise in [0, 1].
Use the set of 2 000 points in input and output to identify the set of coefficients
from the data, by solving (4.53) with a Tikhonov regularization. Consider α =
1, and recall that the minimization of J(w) in this kind of regression (known
also as Ridge regression) has the simple solution
w = (HT H + αI)−1 H y(0) ,
with I the identity matrix. Is the expected result recovered? Is the identified
LTI system capable of predicting what happens if tk > 5? What happens for
α = 0 (i.e., ordinary least square)?

4.9 What’s Next?

This chapter reviewed the fundamentals of signals and systems and presented LTI
systems in case of SISO. We have seen that the input–output relation can be derived
from knowledge of the impulse response of a system and via the convolution integral.
It was shown that complex exponentials are eigenfunctions of these systems and
that important transforms can be derived by projecting input and output signals onto
these eigenfunctions. In the eigenspace of the LTI systems, convolutions become
multiplications.

22 readers familiar with Lagrangian multipliers should recognize in (4.53) an augmented cost function
with α the Lagrangian multiplier.

h d i 0 0 9 8 088962 4 008 bli h d li b C b id i i

 88 Data-Driven Fluid Mechanics Chapter 4

Chapter 10 presents LTI systems, the so-called state-space representation, which

is more common in dynamical system theory and which allows for straightforward
generalization to MIMO systems. Chapter 11 reviews the analysis of nonlinear
systems, while Chapter 12 describes the system identification more broadly and also
considers nonlinear systems.
This chapter reviewed the link between the continuous and the discrete world and
the impact of the discretization on the eigenfunctions of an LTI system. Special values
of the complex frequencies, called poles, yield infinite response of a system and are
linked to the notion of stability, reviewed in Chapter 13. The reader should recognize
that the identification of these poles from large data sets is the essence of the DMD
described in Chapter 7.
Finally, this chapter also introduced the fundamentals of MRA, which complements
wavelet theory in Chapters 5 and 8 on the multiscale proper orthogonal decomposition.

h d i 0 0 9 8 088962 4 008 bli h d li b C b id i i

 5 Time–Frequency Analysis
and Wavelets
S. Discetti

Spectral analysis is the cornerstone of some of the most celebrated turbulent flow the-
ories. When focusing on the identification and modeling of coherent flow patterns, on
the other hand, the need of localizing the spectral content in time or space often arises.
Time–frequency analysis encompasses an arsenal of techniques aimed to localize the
spectral content of a signal in time (although for velocity fields we could clearly refer
to space and spatial frequencies). This chapter gives an overview of techniques for
time–frequency analysis. We start the journey with the windowed Fourier transform
(WFT) in continuous and discrete form, and introduce the concepts of shifting and
stretching to localize a variety of scales. The bounds imposed by the uncertainty
principle will clearly arise. Then, we adapt to those bounds introducing the continuous
wavelet transform (CWT) and its link to multi-resolution analysis (MRA). The
fundamentals of the discrete wavelet transform (DWT) and its intimate relation with
filter banks will be outlined. Finally, two applications are presented: a time–frequency
analysis of hot-wire data in a turbulent pipe flow using the WFT and the CWT; filtering
and compression of velocity fields using a filter based on the DWT. MATLAB codes
to practice with the provided examples are provided on the book’s webpage.1

5.1 Introduction

The Fourier transform (FT) of a function f (t) is an integral transform representing its
projection into harmonic functions. From the inner product introduced in Chapter 4,
the FT is defined as follows:
∫ +∞
F (ω) = f (t)e−jωt dt. (5.1)
−∞
The FT is a powerful and efficient technique to solve a wide variety of problems,
ranging from data analysis and filtering, image compression, communications, and
so on (Smith 2007b). Nonetheless, the FT has severe limitations when applied to
non-stationary signals. Indeed, the FT describes the signal’s frequency content, but
does not provide time localization, i.e., it does not determine the time in which a
given frequency occurs in the signal. This information is lost because of the infinite
integration bounds in (5.1) and the infinite support of the harmonic basis e−jωt .
1 www.datadrivenfluidmechanics.com/download/book/chapter5.zip

h d i 0 0 9 8 088962 4 009 bli h d li b C b id i i

 90 Data-Driven Fluid Mechanics Chapter 5

Consider, for example, the signal x(t) represented in Figure 5.1, with its (discrete)
FT normalized with respect to its maximum and represented as a function of the
frequency ν = ω/2π. The signal exhibits two different frequencies in well-localized
time intervals. Even though a clear signature of the two main frequencies is present in
the spectrum, no information is retained on the time localization of those frequencies.
Although this example might appear extreme at first glance, there is a wide range
of applications where the localization in time is relevant, for example, timescale
modification, sinusoidal modeling of musical compositions, cross-synthesis, and so
on. In fluid mechanics, the localization of scales in space-time has strong potential
implications in the identification and modeling of the flow dynamics (Farge 1992,
Schneider & Vasilyev 2010). For such cases, different implementations are sought,
possibly maintaining the appeal of computationally efficient algorithms such as the
fast Fourier transform (FFT, Cooley & Tukey 1965).
Chapter 4 introduced the limits of the FT and the MRA. This chapter focuses
on techniques for time–frequency analysis of signals, including a discussion of the
capabilities and limits of the WFT, and the wavelet-based multi-resolution methods.
The relation between the DWT and filter banks is also briefly discussed. The objective
is to introduce the fundamental concepts of time–frequency analysis, making no pre-
tence of being exhaustive. For a more comprehensive mathematical background, the
reader is referred to several excellent literature contributions on the topic (Strang 1989,
Daubechies 1990, Daubechies 1992, Cohen 1995, Strang & Nguyen 1996, Torrence
& Compo 1998, Van den Berg 2004, Mallat 2009, Kaiser 2010, Chui 1992).

5.2 Windowed Fourier Transform

5.2.1 Windowed Fourier Transform and Gabor Transform

The most intuitive workaround for the time localization limits of the FT is to split
the signal into shorter time segments and compute the Fourier spectrum on each of
them. This allows localizing in time the frequency content of the signal, with time and
frequency resolution depending on the segment size. Recall that the smallest resolved
frequency and the frequency resolution of the spectrum are inversely proportional
to the signal’s time duration (in this case, the segment). Similarly, shorter or larger
segments enable more or less precise time localization. The delicate compromise
between time and frequency resolution is further discussed in this section.
The signal-splitting method outlined earlier is equivalent to pre-multiplying the
signal by a top-hat window, equal to 1 in the selected time segment and 0 elsewhere.
The window is then shifted along the signal to obtain a time–frequency description.
This is the underlying idea of the Gabor transform (Gabor 1946) in which the kernel
of the FT is multiplied by a sliding window function. This transform is defined as
∫ ∞
G(τ, ω) = f (t)g(t − τ)e−jωt dt, (5.2)
−∞

h d i 0 0 9 8 088962 4 009 bli h d li b C b id i i

 Discetti: Time–Frequency Analysis and Wavelets 91

Figure 5.1 Example of non-stationary signal (top) and corresponding normalized Fourier
spectrum (bottom).

where the overbar indicates the complex conjugate. The window function g(t) is often
selected to be a Gaussian function:

1 1 t 2
g(t) = √ e− 2 ( α ) . (5.3)
α 2π

The parameter α can be used to stretch the window along the time axis. The window
can be translated over the time sequence with the shifting parameter τ to obtain the
temporal description.
The Gabor transform shares the same linearity properties of the FT and can be
inverted with the relation
∫ +∞ ∫ +∞
1
f (t) = G(τ, ω)g(t − τ)ejωτ dωdτ (5.4)
2π −∞ −∞

if the L 2 norm of g(t) is equal to 1.

Figure 5.2 shows the Gabor transform of the signal described by the following
equations:



 sin(2πν1 t) if 1 ≤ t ≤ 3,
x(t) = 0.5sin(2πν2 t)


if 4 ≤ t ≤ 5, (5.5)

0

otherwise,


with ν1 = 3 and ν2 = 10 being the frequencies of the two portions of the signal.

h d i 0 0 9 8 088962 4 009 bli h d li b C b id i i

 92 Data-Driven Fluid Mechanics Chapter 5

Figure 5.2 Example of application of the Gabor transform on the signal in (5.5). In both cases
α = 0.5, while τ = 2.0, 4.5 for the left and the right column, respectively. From top to bottom:
raw signal with superposed sliding window of the Gabor transform in thick black line;
premultiplied signal; Fourier spectrum of the signal premultiplied by the Gaussian kernel at
τ = 2.0, 4.5.

This is the same signal analyzed with FT in Figure 5.1. A Gaussian kernel with
α = 0.5 is here chosen. The selected kernel is expected to have sufficient width to
capture the low-frequency oscillations and to be reasonably well localized to identify
the high-frequency oscillations occurring on a shorter time segment. The Gabor kernel
is then centered at τ = 2.0 and τ = 4.5, i.e., at the center of the regions interested
by the two sinusoidal oscillations. Observing the spectrum of the Gabor transform
coefficients, the two frequencies are well localized both in the frequency and time
domain, with a single sharp peak in the spectrum at ν = 3 and ν = 10 at the two
selected time instants, respectively.

5.2.2 Discrete Gabor Transform

In real applications the signals are discretized, thus needing the definition of a discrete
Gabor transform. In addition to the discretization of time and frequencies introduced
in Chapter 4, we have to include also a discretization of the time shift τ = nτ0 , with

h d i 0 0 9 8 088962 4 009 bli h d li b C b id i i

 Discetti: Time–Frequency Analysis and Wavelets 93

Figure 5.3 Graphical explanation of the different description of the signal in the time domain,
frequency domain, and with the spectrogram obtained by the Windowed Fourier Transform.

n ∈ N, with τ0 being a fundamental time unit. With the sampling frequency being νs ,
the time grid is t = k∆t, with ∆t = 1/νs and k = 0, . . . , N − 1. The corresponding
discrete Gabor transform is thus given by
N
Õ −1
Gm,n = fk g(k∆t − nτ0 )e−j2πmν0 , (5.6)
k=0

with fk being the discrete values of f (t) sampled at t = k∆t.

A careful choice of the fundamental time τ0 and frequency ν0 for discretization
must be taken. A proper sampling is achieved with τ0 = ∆t and ν0 = νs /N. In the time
domain, boundary treatment (e.g., zero-padding, see Chapter 4) might be beneficial,
while in the frequency domain the same equally spaced bins of the discrete FT are
generally adopted (i.e., m = 0, . . . , N − 1).
Once the lattice of time and frequency is selected, (5.6) can be used to compute
a spectrogram. A conceptual sketch to exemplify the differences between the rep-
resentation in the time–frequency domain and the spectrogram is shown in Figure
5.3. A signal can be described purely in the time domain (i.e., the raw signal), with
no frequency resolution. On the other hand, a description in the frequency domain
can be obtained via the FT at the expense of withdrawing the time resolution in the
integration process. The spectrogram obtained with a time-shifting windowing FT is
able to achieve a temporal description of the frequency content by partly giving up
frequency resolution. While in Figure 5.3 an equal share between time and frequency
resolution is shown for simplicity of illustration, the real balance is tuned according
to the properties of the kernel used for the windowing.
This concept is made clear by inspection of different spectrograms computed on
the signal defined by (5.5), reported in Figure 5.4. For this computation, t0 = ∆t and
ν0 = νs /N are selected. The parameter α, i.e., the “time-width” of the windowing
function g(t), has a significant impact on the transform. According to (5.3), a
small value of α corresponds to a relatively narrow window, thus having high

h d i 0 0 9 8 088962 4 009 bli h d li b C b id i i

 94 Data-Driven Fluid Mechanics Chapter 5

Figure 5.4 Spectrogram of the signal described by (5.5) for two different values of the scaling
parameter α = 0.1 (left) and α = 10.0 (right).

temporal-localization capability (i.e., good temporal resolution) but poor frequency

resolution. Intuitively, a narrow window isolates short sequences during the time-
shifting process, thus hindering the identification of the low-frequency component
of the signal and lacking in spectral resolution. On the other hand, a large value of
α delivers a broader window, which in turn has poor temporal localization, but better
frequency resolution and capability of identifying the low-frequency component of
the signal.
Clearly, if α → 0, the spectrogram tends to perfect temporal localization with no
frequency resolution (i.e., a pure time-domain description). On the other hand, for
α → ∞, the kernel of the Gabor transform tends to the kernel of the FT, thus giving
up completely the time resolution in favor of full frequency resolution.
The backbone of this limit of the Gabor transform is the Heisenberg uncertainty
principle. In quantum mechanics, according to the uncertainty principle, position and
momentum of a particle cannot be simultaneously ascertained with arbitrary precision.
In more general terms, it is not possible to arbitrarily localize both a function and its
FT. If a signal is sharply localized (for instance a delta function), its FT is broadband;
vice versa, a Fourier spectrum with only one non-null component corresponds to a
periodic continuous sinusoidal function.
In strict mathematical terms, the uncertainty principle in the framework of time–
frequency analysis can be formulated as follows:
∫ +∞ ∫ +∞
1 1 1
  
2 2 2 ˆ 2
2
t | f (t)| dt 2
ω | f (ω)| dω ≥ . (5.7)
k f k −∞ 2π k f k −∞ 4

The integrals in parentheses are also known as dispersions.

The choice of the Gaussian kernel in the Gabor transform is not accidental. It can be
shown that Gaussian functions have an optimal dispersion product, thus minimizing
the combined uncertainty in time–frequency localization. The reader can indeed easily
verify that, if
1 − t
2
f (t) = √ e 2α2 , (5.8)
α 2π

h d i 0 0 9 8 088962 4 009 bli h d li b C b id i i

 Discetti: Time–Frequency Analysis and Wavelets 95

with its FT being

1 2 ω2
F (ω) = e− 2 α , (5.9)

then the left-hand side of (5.7) results in 1/4.

Furthermore, the integrals on the left-hand side have an interesting interpretation in
terms of the so-called Heisenberg’s boxes, i.e., ideal rectangles in the time–frequency
plane indicating the uncertainty in time–frequency definition. These “rectangles” are
often referred to as time–frequency atoms, already indicated in the conceptual sketch
highlighted in Figure 5.3. Interestingly enough, the dispersions for a Gaussian function
in (5.8) are proportional to α2 and 1/α2 , respectively; for increasing α, the Gaussian
function becomes less localized in space but more localized in frequency, and vice
versa.

Example 1: DFT versus WFT

Consider the signal x(t), t ∈ [0, 2π] described by (5.5) and sampled with N =
2 048 points.
1. Compute the spectrum of the signal using the DFT.
2. Compute the spectrogram of the signal using the WFT with Gaussian kernel
(as in (5.3)). Explore the effect of the standard deviation α of the kernel on
the time and frequency resolution.

Solution
The solution is provided in the form of MATLAB code in the supplementary
material. The lattice for time and frequency is set as τ0 = ∆t, ν0 = νs /N. The
reader can modify the script accordingly to generate Figures 5.1 and 5.4.

5.3 Wavelet Transform

5.3.1 Fundamentals
As discussed in Section 5.2, the main shortcoming of the WFT is the compromise to
be sought between time and frequency resolution, i.e., the capability of localizing
features in the time and frequency domains. The resolution can be maximized by
choosing windows with the minimal area of the corresponding Heisenberg box
in the time–frequency plane. However, if our analysis is restricted to one single
window choice, then only one scale is selected (within the uncertainty of time–
frequency localization from the Heisenberg principle). To achieve a well-resolved
description in a set of scales, a process that uses a library of windows spanning
the time–frequency domain should be conceived. “Broad” windows, with excel-
lent frequency localization but poor temporal resolution, can be used to describe
the low-frequency part of the signal. For such scales, the time localization is

h d i 0 0 9 8 088962 4 009 bli h d li b C b id i i

 96 Data-Driven Fluid Mechanics Chapter 5

(a) (b)
1 1
0.5 0.5
0 0
–0.5 –0.5
–1 –1
–2 0 2 4 –2 0 2 4

(c) (d)
1
0.5
0
–0.5
–1
–2 0 2 4

Figure 5.5 Haar mother wavelet (a), with examples of scaled (b), shifted (c), and
scaled–shifted (d) versions.

less relevant since low-frequency signals are by definition less localized in time.
“Narrow” windows, on the other hand, can deliver excellent time localization for the
high-frequency scales. This multi-resolution approach is the cornerstone of wavelet
theory.
The main ideas behind the wavelet concept are the scaling and shifting processes,
already introduced in the WFT. In wavelet theory, the starting point is a function
referred to as “mother wavelet” ψ(t), which is shifted along the signal to provide time
localization, and scaled to capture scales of different size. The family of wavelets is
thus generated as follows:

1 t−b
 
ψa,b (t) = √ ψ . (5.10)
a a

The variables a and b are, respectively, the scaling and shifting parameters. The
√
term 1/ a is introduced to obtain functions with unitary norm. The wavelet is then
shifted across the time domain and progressively scaled to create a collection of time–
frequency descriptions of the signal, i.e., a multi-resolution analysis.
An illustrative example of this process is shown in Figure 5.5. The earliest example
of a wavelet was the Haar wavelet, introduced in 1911 by Alfred Haar (1911) as the
Haar sequence (the term wavelet appeared only several decades later).

h d i 0 0 9 8 088962 4 009 bli h d li b C b id i i

 Discetti: Time–Frequency Analysis and Wavelets 97

(a) (b)

(c) (d)

Figure 5.6 Haar (top row) and Mexican-hat (bottom row) wavelets: time representation (left
column) and corresponding frequency spectra (right column).

It is defined as follows:



 1 if 0 ≤ t < 1/2,
ψ(t) = −1 if 1/2 ≤ t < 1,


(5.11)

0

otherwise.


Figure 5.5 includes the mother wavelet ψ1,0 , a scaled Haar wavelet ψ2,0 (i.e., with
scaling parameter a = 2, thus enlarging the support of the wavelet and, consequently,
the corresponding scale), a shifted Haar wavelet ψ1,1 (i.e., centered in t = 3/2, being
b = 1) and a scaled-shifted Haar wavelet ψ2,−1 .
The Haar wavelet has the disadvantage of being noncontinuous (although this is not
an issue for the discrete formulation); nonetheless, this turns out to be an advantage for
the description of signals with sharp changes. This feature is particularly appreciated,
for instance, for edge detection in image processing.
The Haar wavelet is highly localized, thus giving good time localization but poor
frequency resolution. This is a direct consequence of the uncertainty principle outlined
previously: a compact support in the time-domain results in a broadband frequency
spectrum (see Figure 5.6(b)).
Depending on the application and on the desired properties of time–frequency
resolution, there is a vast variety of mother wavelets already available in the literature
and in software toolboxes. A classic example is the Mexican-hat wavelet:

h d i 0 0 9 8 088962 4 009 bli h d li b C b id i i

 98 Data-Driven Fluid Mechanics Chapter 5

" 2#
2 t−b 2

− (t −b)
ψa,b (t) = √ 1− e a 2 . (5.12)
3aπ 1/4 a

The Mexican-hat wavelet and its FT are shown in Figure 5.6(c,d). The Gaussian
kernel, which has optimal time–frequency bandwidth product in (5.7), enables the
best compromise for the localization in the time and the frequency domains. From the
comparison, it is clear that while the spectrum of the Haar wavelet decays with ν −1 ,
the frequency spectrum of the Mexican-hat wavelet is much sharper.

5.3.2 The Continuous Wavelet Transforms

The CWT (Grossmann & Morlet 1984) is defined as the following integral transform:
∫ +∞
Wψ [ f (t)](a, b) = f (t)ψ a,b (t)dt. (5.13)
−∞

Notice that, similarly to the Gabor transform (5.2), the CWT depends on two
parameters, the scale a and the shift b. Additionally, as in the WFT, a wide variety
of CWTs can be defined by selecting the proper wavelet ψ for the desired application.
The main constraint for the selection of the wavelet is the admissibility condition, i.e.,
∫ +∞
| ψ̂(ω)| 2
Cψ = dω < ∞, (5.14)
−∞ |ω|

with ψ̂(ω) being the FT of the wavelet function.

The interpretation of the CWT is simple. The signal is processed by a wavelet
translated with the translation parameter b, and the result is stored; then the process is
repeated after scaling the wavelet with the scaling parameter a to extract a description
of the signal at a different scale. This is then carried out in a range of scales, thus
covering both the time and the frequency domain.
The CWT is linear (i.e., Wψ [α f (t) + βg(t)](a, b) = αWψ [ f (t)] + βWψ [g(t)]) and
invertible with the relation
∫ +∞ ∫ +∞
1 1
f (t) = Wψ [ f ](a, b)ψa,b (t)db da. (5.15)
Cψ 0 −∞ a
2

The existence of the inverse CWT clearly relies on the admissibility of the wavelet,
i.e., Cψ < ∞.
It is important to underline here that we shifted from the time–frequency repre-
sentation of the Gabor transform to the timescale representation of the CWT. Here
two important remarks are needed. First, the relation between scales and frequency
might not be immediate, considering that often wavelets have irregular shape (see
for instance the Daubechies wavelets, Daubechies 1992). A simple method to extract
such relation has been proposed by Meyers et al. (1993). It is based on convolution
of the wavelet with a cosine wave, and searching for the scale a that maximizes
the correlation. This is a useful exercise for wavelets where a dominant frequency

h d i 0 0 9 8 088962 4 009 bli h d li b C b id i i

 Discetti: Time–Frequency Analysis and Wavelets 99

Figure 5.7 Graphical explanation of the difference between the time/frequency description of
the windowed Fourier transform (a) and the timescale description of the wavelet analysis (b).

correlation is expected, such as the Morlet wavelet, which is a harmonic function

multiplied by a Gaussian. In this case, a linear relation between scale and wavelength
is observed (Meyers et al. 1993). Second, the shape of the time–frequency atoms
depends on the scaling parameter a. A conceptual sketch of the timescale description
obtained in this framework is illustrated in Figure 5.7. Interestingly, it can be easily
shown (see chapter 4 of Mallat (2009)) that the sides of the time–frequency atoms
scale respectively with a along the time axis, and 1/a along the frequency axis,
respectively. The area of the time–frequency atom remains constant, and only depends
on the wavelet function. It is thus clear that increasing the scaling parameter a
“broadens” the atom in time domain and “narrows” it in the frequency domain.
A comparison of the performances of the Gabor transform and the CWT is carried
out in Section 5.4.1 using hot-wire data from a pipe-flow experiment in the facility
CICLoPE (Centre for International Cooperation in Long Pipe Experiments).

5.3.3 Wavelet Series: the Path to Discrete Wavelet Transform

In practical applications, the data are normally available on a discrete grid, in time (for
instance the sampling instants of a hot-wire probe) and/or in space (e.g., the vector grid
of Particle Image Velocimetry, or of numerical simulations). It is thus natural to define
the wavelet transform in a discrete environment.
Furthermore, it should be remarked that, similarly to the Gabor transform, the CWT
description is fundamentally “redundant”, i.e., depending on the choice of the shifting
and scaling parameters the timescale space can be oversampled. We often search
instead for a different representation of our data (i.e., a different basis), with the aim of
highlighting certain features (see Part III). Thus, if we have N samples, we search for
a set of N numbers in a different space, which allow us to extract patterns or features
otherwise hidden in the time (or space) representation.

h d i 0 0 9 8 088962 4 009 bli h d li b C b id i i

 100 Data-Driven Fluid Mechanics Chapter 5

The wavelet family can be written as

1
ψm,n = p m ψ(a0−m x − nb0 ), (5.16)
a0
with a0 and b0 being positive parameters and m, n being integer numbers. Typically
a0 = 2 and b0 = 1, thus leading to
1
ψm,n = √ ψ(2−m (x − n2m )). (5.17)
2m
This results in progressive binary dilations, also called voices or octaves, and
translations of n2m . As will be made clear in Section 5.3.5, one of the main advantages
of this choice is the computational efficiency of its implementation.
A prominent role is played by orthogonal wavelets, i.e., fulfilling the condition
∫ +∞
hψm,n ψ p,q i = ψm,n (t)ψ p,q (t)dx = δm,n δ p,q , (5.18)
−∞

where h·, ·i is the inner product in L 2 and δi, j is the Kronecker delta
(
1 i = j,
δi, j = (5.19)
0 otherwise.
In case of decomposing with orthogonal wavelets, a given function f (t) can be
uniquely defined by its Discrete Wavelet Transform, i.e.,

Wm,n = ( f (t), ψm,n (t)), (5.20)

with the inverse DWT being

+∞ Õ
Õ +∞
f (t) = Wm,n ψm,n (t). (5.21)
m=−∞ n=−∞

This important property opens the question of how to generate a family of

orthogonal wavelets. Section 5.3.4 highlights the link between a general procedure
to generate wavelets and the multi-resolution analysis introduced in Chapter 4.

5.3.4 Multi-Resolution Analysis

The concept of multi-resolution approximation (or MRA) forms the fundamental
cornerstone for the design of discrete orthogonal wavelets. The concept is here briefly
introduced; for a more extensive treatment, the reader is referred to Mallat (2009).
The MRA computes the approximation of a generic function f as a set of
orthogonal projections on a sequence of spaces {Vm }m∈Z .
Consider the sequence of closed subspaces {Vm }m∈Z in L 2 (R). This can be used
for multi-resolution approximations if the following properties are satisfied:

• Vm ⊂ Vm+1, ∀m ∈ Z
• ∪+∞ 2
m=−∞ Vm spans L (R)

h d i 0 0 9 8 088962 4 009 bli h d li b C b id i i

 Discetti: Time–Frequency Analysis and Wavelets 101

• ∩+∞
m=−∞ Vm = {0}, i.e., the only intersection between the subspaces is the null
function
• if f (x) ∈ Vm , then f (2x) ∈ Vm+1, ∀m ∈ Z
• there exists a function φ in V0 , called a scaling function or father wavelet, such
that φ(x − n) with n ∈ Z is an orthonormal basis of V0 with respect to the inner
product in L 2 (R).

A simplified view of a 3-level MRA in the form of a Venn diagram is shown

in Figure 5.8. It can be demonstrated that an orthogonal basis for each subspace
{Vm }m∈Z can be constructed from the basis of V0 as follows:
√
{φm,n (x) = 2m φ(2m x − n), n ∈ Z}, (5.22)

since φ ∈ V0 ⊂ V1 ⊂ V2 ⊂ . . . Vm . For example,

√ an orthonormal basis for V1 can
be written from the basis of V0 as {φ1,n (x) = 2φ(2x − n), n ∈ Z}. Thus the scaling
function φ can be expressed as a linear combination of {φ1,n (x)}, i.e.,
Õ √ Õ
φ(x) = cn φ1,n (x) = 2 cn φ(2x − n), (5.23)
n∈Z n∈Z

referred to as the scaling equation or refinement equation.

We need now to determine the set of orthonormal wavelets corresponding to the
defined MRA. Consider, for this purpose, the orthogonal complement W0 of V0 in
V1 , i.e., the set of functions in V1 that are orthogonal to any function in V0 :

W0 = { f ∈ V1 : h f , gi = 0 ∀g ∈ V0 }. (5.24)

Our candidate to build a set of orthogonal wavelets is a function ψ ∈ W0 that

is analogous to the scaling function, i.e., it generates a basis of W0 from integer
translation ψ(x − n). Since ψ ∈ W0 ⊂ V1 and φ1,n is a basis for V1 , then
Õ
ψ(x) = an φ1,n . (5.25)
n∈Z

Recall that ψ ∈ W0 and φ ∈ V0 , then hψ, φi = 0 from which we obtain that

Õ
hψ, φi = an cn = 0. (5.26)
n∈Z

One simple solution for the previous equation is to set an = (−1)n c1−n . We can now
define the corresponding mother wavelet
Õ
ψ(x) = (−1)n c1−n φ(2x − n), (5.27)
n∈Z

which generates the family of orthogonal wavelets corresponding to the MRA defined
earlier. Interestingly enough, if we interpret the scaling function as a low-pass filter,
multiplying by (−1)n its impulsive response leads us to define its complementary high-
pass filter.
As discussed in Chapter 4, this is true because the dyadic construction in (5.22)
gives to these complementary filters equal portions of each scale’s spectra.

h d i 0 0 9 8 088962 4 009 bli h d li b C b id i i

 102 Data-Driven Fluid Mechanics Chapter 5

Figure 5.8 Conceptual sketch with Venn diagram of a three-level MRA. The spaces V0, V1 ,
and V2 of the MRA are indicated with the corresponding symbol located on their boundary.
The respective orthogonal complements W0 and W1 are indicated with the symbol located
within the corresponding domain.

Example: the Haar wavelet

The simplest example of construction of orthogonal wavelets via MRA is the Haar
wavelet. Consider a generic function f ∈ L 2 , and suppose we want to approximate it
with a piecewise constant approximation at different levels of resolution fn of length
2n−1 . For example, f0 is a piecewise constant approximation over segments of length
1, f1 on length 1/2, and so on. This is equivalent to selecting a scaling function in V0 ,

φ0,n (x) = I[n,n+1) (x), (5.28)

with I A being the indicator function, equal to 1 for all elements belonging to A and 0
elsewhere, and n ∈ Z.
It can be easily shown that the functions φ0,n are orthogonal, and that they span the
entire V0 with a set of integer translation of φ0,n = φ(x − n). This function is the Haar
scaling function.
If we consider now the piecewise approximation at level V1 , this is carried out by
functions that are piecewise constant over length √ 1/2. In this space, an orthonormal√
basis is constituted by the functions φ1,n (x) = 2I[n/2,(n+1)/2) (x) (with the factor 2
to account now for the shorter support of the functions). Intuitively, these functions
can be obtained by scaling and shifting the function φ0,n , thus obtaining
√
φ1,n (x) = 2φ(2x − n). (5.29)

This is the same result of (5.23) outlined earlier. From (5.23), for the Haar system
we obtain
( √
1/ 2 if n = 0, 1,
cn = (5.30)
0 otherwise.
From (5.27), we can simply derive the corresponding mother wavelet, which is the
discrete form of the Haar wavelet ψ reported in (5.11). It is straightforward to show
that φ and ψ are orthogonal, as was expectable since φ ∈ V0 and ψ ∈ W0 , which is
the orthogonal complement of V0 .

h d i 0 0 9 8 088962 4 009 bli h d li b C b id i i

 Discetti: Time–Frequency Analysis and Wavelets 103

Figure 5.9 Sketch of the computation process of the fast wavelet transform using filter banks.

Now, if we consider the piecewise approximation of a generic function f at the level

f0 and f1 , and compute the difference between them, we discover that it resembles a
set of scaled and translated versions of the Haar wavelet ψ. This simple result is just
a consequence of the fact that W0 is the orthogonal complement of V0 , i.e., the part
of f that is “discarded” in the projection on V0 belongs to W0 and, as such, can be
expressed as a linear combination of the wavelet function ψ.

5.3.5 The Discrete Wavelet Transform and Filter Banks

Consider
√ the generic function f1 (t) ∈ V1 . It can be written as a linear combination of
φ1,n = 2φ(2x − n), or alternatively it can be expressed as the sum of its projections
onto V0 and its orthogonal complement W0 , i.e.,
Õ Õ Õ
f1 (t) = α1,n φ1,n (t) = α0,n φ0,n (t) + β0,n ψ0,n (t). (5.31)
n∈Z n∈Z n∈Z

This operation is essentially a change of basis. The first term on the RHS is referred
to as approximation (it is indeed the best approximation of f (t) at the resolution level
of V0 ), and the second term is the detail, i.e., the part of f that is missing to obtain f1 .
As a general result, extending this procedure to the subsequent levels,
Õ ÕÕ
f (t) = α0,n φ0,n (t) + βm,n ψm,n (t), (5.32)
n∈Z m∈N n∈Z

i.e., each function f (t) can be expressed as the sum of the approximation at the
resolution level V0 and a sequence of details at the levels m ≥ 0.
This process is normally implemented as a two-channel multi-rate filter bank
(Mallat 1989), which convolves the signal with a low-pass filter g[n] and a high-
pass filter h[n]. The high-pass filter retains the details at the scale m, while the low-
pass filter contains the approximation. Both descriptions are subsampled by a factor
of 2; then the low-pass filtered signal is again passed through the high- and low-pass
filters to obtain the approximation and details at the subsequent scale. At the end of

h d i 0 0 9 8 088962 4 009 bli h d li b C b id i i

 104 Data-Driven Fluid Mechanics Chapter 5

the process, the DWT delivers an approximation at the level 0 and a set of details at
different scales. A sketch of this process is shown in Figure 5.9.
In order to establish the connection between DWT and filter banks, consider a
signal x[n] of N samples, and the simplest two-channel filter bank, composed by
a low-pass filter that performs a moving average of the values of the signal, and
a high-pass filter that computes the differences. The low-pass filter has coefficients
g[0] = g[1] = 1/2, while the high-pass filter has coefficients h[0] = −h[1] = 1/2. The
two filters separate the frequency components of the signal x into two bands, low and
high frequencies. In matrix form:

1 1   x0 
1  1 1  x1 
   

Gx =   x2  , (5.33)
  
2 1 1 
.
  
 .. 
 .. 

 . 
  
1 −1   x0 
1 1   x1 
  
−1
Hx =  . (5.34)

  x2 

2  1 −1
  .. 
  

 ... .

The convolution of x with each of the filter produce 2N components. Recall

that we are searching for a new representation, i.e., a new basis, thus we aim to
obtain N components. The filtering operation is thus followed by downsampling by
a factor of two (indicated with the symbol ↓ 2), which removes the odd components
of the filtered signals. This can be expressed simply
√ in matrix form as an identity
matrix with missing odd rows, and multiplying by 2 for energy normalization (since
downsampling is dropping half of the components). Combining the operations of
filtering and downsampling in matrix form, we obtain:

1 1 
1 1 1
 
(↓ 2)G = √ , (5.35)
 

2

 ... 

 
 
1 −1 
1 1 −1
 
(↓ 2)H = √ . (5.36)
 

2

 ... 

 
 

This operation can be applied recursively, as outlined in Figure 5.9, i.e., the
low-pass filtered part is passed again through the low- and high-pass filter, and
downsampled, and so on. But how does this relate to wavelets and MRA? The answer
is simple: the MRA is performing exactly the same operation, filtering (through φ)
and downsampling. The reader should now start to glimpse the relation with the

h d i 0 0 9 8 088962 4 009 bli h d li b C b id i i

 Discetti: Time–Frequency Analysis and Wavelets 105

Haar wavelet and the simple two-channel filter bank proposed here. From (5.23),
considering the coefficients in (5.30),
φ(t) = φ(2t) + φ(2t − 1), (5.37)
and, in general, for discrete signals and for the levels j and j − 1,
1
φ j−1,n = √ [φ j,2n + φ j,2n+1 ]. (5.38)
2
For the signal x of N samples, this would result in N/2 coefficients, with the same
output of applying the process of filtering and downsampling (it can be checked easily
that the operation can be written in matrix form exactly as in (5.35)).
Similarly, from (5.27),
1
ψ j−1,n = √ [φ j,2n − φ j,2n+1 ], (5.39)
2
which is equivalent to (5.36).
This is the celebrated Mallat’s pyramid algorithm, which is based on decomposing
a function at a certain resolution level as the sum of approximation and detail at a
coarser resolution. One of the main advantages of this algorithm is the efficiency of
its implementation. The fast wavelet transform requires a number of operations that
scales linearly with the number of samples of the signal, thus overcoming in efficiency
the FFT (whose complexity scales with N log2 (N)).
The inverse process (referred as synthesis or reconstruction) is based on upsampling
the signal filling the gaps with zeros, and on convolution with the same filters.

Example 2: Calculate the DWT of a signal

Consider the discrete signal:

x = [4, 6, 9, 5, 1, 3, 4, 4].
1. Compute the 1-level DWT with Haar wavelet and express it in the form of
a matrix transformation.
2. Compute the 3-level DWT with Haar wavelet and express it in the form of
a matrix transformation.
3. Compute the inverse DWT of the wavelet decomposition of the previous
point.
Solution
1. The father wavelet is given by (5.28), with the corresponding mother
wavelet being given by (5.27). After application of the filters and down-
sampling according to the matrix form of (5.35) and (5.36), we obtain
√ √ √ √ √ √ √
α(1) = [5 2, 7 2, 2 2, 4 2]; β(1) = [− 2, 2 2, − 2, 0]. (5.40)
In matrix form, this is equivalent to
(1)
x = W (1) xW , (5.41)

h d i 0 0 9 8 088962 4 009 bli h d li b C b id i i

 106 Data-Driven Fluid Mechanics Chapter 5

(1)
with xW = [α(1), β(1) ]T and
1 0 0 0 1 0 0 0 
1 0 0 0 0 0 0 

−1
0 1 0 0 0 1 0 0 

1 0 1 0 0 0 −1 0 0
W (1) =√  . (5.42)
2 0 0 1 0 0 0 1 0
0 0 1 0 0 0 −1 0 
0 0 0 1 0 0 0 1 
0 0 0 1 0 0 0

 −1
Notice that the columns of W (1) are orthonormal, and that by construction,
W (1) = [(↓ 2)GT (↓ 2)HT ] .
It is also worth evaluating the energy of the signal and the share of it between
the coefficients:
8
xi2 = 200,
Õ
E=
n=1
Eα(1) = 188,
Eβ(1) = 12,
i.e., 94% of the energy is contained in the approximation coefficients. This
is also a consequence of the signal having a nonzero mean. This means that
we can preserve 94% of the information of the signal by retaining only half
of the wavelet coefficients. This is the cornerstone of signal compression
using DWT.
2. To achieve a 3-level DWT, we have to apply recursively the same procedure
of the previous point to the approximation at each level. This leads to
α(2) = [12, 6]; β(2) = [−2, −2], (5.43)
√ √
α(3) = [9 2]; β(3) = [3 2]. (5.44)
The set of coefficients and the corresponding DWT matrix are as follows:
(3)
√ √ √ √ √
xW = [α(3), β(3), β(2), β(1) ] = [9 2, 3 2, −2, −2, − 2, 2 2, − 2, 0], (5.45)

 √1 √1 1
0 √1 0 0 0 
 8 8 2 2
1
√1 1
0 − √1 0 0 0 

√ 2
 8 8 2 
 √1 √1 − 21 0 0 √1 0 0 
 8 8 2
1
√1 − 21 0 0 − √1 0 0 

√
W (3) =  18 8 2
. (5.46)

√ − √1 0 1
2 0 0 √1 0 
 8 8 2 
 √1 − √1 0 1
0 0 − √1 0 
 8 8 2 2
1
− √1 0 − 12 0 0 0 √1 

√
 8 8 2 
1
√1 0 − 21 0 0 0 − √1 

√
 8 8 2

h d i 0 0 9 8 088962 4 009 bli h d li b C b id i i

 Discetti: Time–Frequency Analysis and Wavelets 107

Again, we notice that 81% of the energy is in the approximation and 90%
in the level 3 (i.e., in only 2 coefficients). Similarly, we retain 94% of the
energy by√retaining only the 3 coefficients with magnitude larger than or
equal to 2 2.
3. The inverse transform can be carried out simply by inverting the previous
matrices. Nonetheless, it is instructive to carry out the process according to
the filter bank implementation, i.e., recursive upsampling and convolution.
For the 3-level decomposition, we first upsample including zeros, and then
use the filters.
(3)
√ (3)
√
α↑2 = [9 2, 0] β↑2 = [3 2, 0], (5.47)
(3) (3)
α̂↑2 = [9, 9] β̂↑2 = [3 − 3], (5.48)
(3) (3)
α(2) = α̂↑2 + β̂↑2 = [12, 6]. (5.49)
At the next level,
(2) (2)
α↑2 = [12, 0, 6, 0] β↑2 = [−2, 0, −2, 0], (5.50)
(2)
√ √ √ √ (2)
√ √ √ √
α̂↑2 = [6 2, 6 2, 3 2, 3 2] β̂↑2 = [− 2, 2, − 2, 2], (5.51)
(2) (2)
√ √ √ √
α(1) = α̂↑2 + β̂↑2 = [5 2, 7 2, 2 2, 4 2]. (5.52)
Finally, for the last level,
(1)
√ √ √ √ (1)
√ √ √
α↑2 = [5 2, 0, 7 2, 0, 2 2, 0, 4 2, 0], β↑2 = [− 2, 0, 2 2, 0, − 2, 0, 0, 0],
(5.53)
(1) (1)
α̂↑2 = [5, 5, 7, 7, 2, 2, 4, 4], β̂↑2 = [−1, 1, 2, −2, −1, 1, 0, 0], (5.54)
(1) (1)
x = α̂↑2 + β̂↑2 = [4, 6, 9, 5, 1, 3, 4, 4]. (5.55)
The interesting exercise of carrying out the DWT after thresholding the
wavelet coefficients is left to the reader.

5.4 Tutorials and Examples

Time–frequency analysis, and particularly MRA with wavelets, has found numerous
applications in fluid mechanics, especially in turbulence modeling and simulation (see
e.g., Meneveau 1991, Farge 1992, Schneider & Vasilyev 2010). In this section, simple
examples of application of the concepts outlined earlier are proposed, with the target
aimed toward direct application on data analysis, rather than modeling.

h d i 0 0 9 8 088962 4 009 bli h d li b C b id i i

 108 Data-Driven Fluid Mechanics Chapter 5

Hz

Figure 5.10 Velocity fluctations for Example 3 (top) and corresponding Fourier spectrum
(bottom), normalized by the maximum value.

5.4.1 Time–Frequency Analysis of Hot-Wire Data: Continuous Wavelet

Transform versus Windowed Fourier Transform
In this example, a signal is analyzed using the WFT with Gabor kernel and the CWT
to obtain a time–frequency description. The signal is a hot-wire data set captured in
the CICLoPE facility at the Universitá Alma Mater Studiorum di Bologna (Talamelli
et al. 2009). The full data set comprises synchronized flow-field measurements with
Particle Image Velocimetry (PIV) at low repetition rate, and hot-wire measurements
with good temporal resolution. A full description of the experiment is reported in
Discetti et al. (2019). The data are available in the Turbase repository (https://
turbase.cineca.it/init/routes/#/logging/view_dataset/81/tabmeta).
In this section, we consider the data corresponding to Case 1 of the data set. The file
contains the variable ut, which corresponds to a 20-minute recording at 10 kHz on 6
channels. The first 5 channels correspond to the simultaneous recording of a rake of 5
hot-wires at the same streamwise and azimuthal position but increasing wall distance
from row 1 to 5. The 6th channel stores the recording time instant of the PIV system.
For the purpose of this analysis, we consider only data from the first row, i.e., from
the closest probe to the wall, within the time segment 1s < t ≤ 3s, subsampled at 1
kHz. The velocity fluctuations (i.e., after removing the time-average velocity) and the
corresponding Fourier spectrum are reported in Figure 5.10 for reference.
By visual inspection of the raw signal, we can notice the signature of relatively
low-frequency fluctuations, superposed with a range of high-frequency fluctuations of
lower intensity. We aim to localize these events through a time–frequency analysis.

h d i 0 0 9 8 088962 4 009 bli h d li b C b id i i

 Discetti: Time–Frequency Analysis and Wavelets 109

Figure 5.11 Spectrogram for Example 3, with α = 0.1 s (left) and α = 0.4 s (right). The
spectrograms have been normalized with their corresponding maximum value.

Figure 5.12 Continuous wavelet transform for Example 3. The wavelet coefficients are
reported in absolute value and normalized with respect to the maximum.

The quantitative analysis, with computation of the spectrogram and the scalogram, is
presented in the following example.

Example 3: Computing the Gabor spectrogram and the CWT scalogram

of hot-wire data
Consider the hot-wire experimental data presented before. Extract the data
from the closest probe to the wall, within the time segment 1 s < t ≤ 3 s, and
subsample the sequence to 1 kHz.
1. Compute the Gabor spectrogram of the signal with different kernel size (for
instance α = 0.1 s and α = 0.4 s).
2. Compute the CWT of the same signal and discuss the differences between
the scalogram and the spectrogram.

h d i 0 0 9 8 088962 4 009 bli h d li b C b id i i

 110 Data-Driven Fluid Mechanics Chapter 5

Solution
1. The spectrogram, computed with Gabor kernel with α = 0.1 s and α = 0.4 s
is illustrated in Figure 5.11. Once again, it is evident that in the case of the
narrow window (α = 0.1 s) a good time localization is achieved, at the
expense of the frequency resolution. On the other side, with α = 0.4 s, it
is possible to identify some low-frequency features (for example, a peak of
spectral energy is observed at t ≈ 2 s −ν ≈ 2 Hz, which is corresponding
to an intense fluctuation that is clearly identifiable in the raw signal and
with period of approximately 0.5 s). This feature was evidently distorted
for α = 0.1 s, which is not capable of identifying low-frequency features
due to the poor frequency resolution. Indeed the width of the kernel is 0.2 s
if measured at ±α, thus all frequencies below ≈ 5 Hz are not detectable.
This qualitative consideration should be of course quantitatively supported
by computing the impulse response of the selected Gabor kernel.
2. The corresponding scalogram from the CWT is shown in Figure 5.12.
This can be easily computed with the command cwt in MATLAB and is
generally implemented in libraries for the vast majority of programming
languages. In MATLAB the Morse wavelet is set by default. At low
frequencies, it can be observed that there is a good frequency resolution
but relatively poor time localization. For low frequencies, indeed, wavelets
with large support are used. Flowing along the spectrum of scales, we
clearly see that the temporal localization progressively improves, at the
expense of frequency resolution. This is in line with the conceptual sketch
of the wavelet transform reported in Figure 5.7. Since the signal is of finite
length, a cone of influence bounds the region where edge effects become
relevant, i.e., a region where the scaled wavelet partially extends outside of
the time domain of the raw signal.
Interestingly enough, a visual comparison of Figures 5.11 and 5.12
exemplifies the multi-resolution capabilities of the CWT. In the low-
frequency range, the CWT scalogram and the spectrogram with α = 0.4 s
share significant similarities; for frequencies in the range 5 − 20 Hz, there
is more agreement with the spectrogram with α = 0.1 s. This comparison
highlights the potential of CWT over the WFT: using a scaled version of
the kernel function at different frequencies, it is possible to carry out a
tunable MRA of the signal, depending on the scale to be observed.
The solution is provided in the supplementary material in the form of
MATLAB code.

5.4.2 Filtering and Compression

A great amount of literature on wavelets is devoted to denoising and/or compressing
data. For the case of additive noise, the noise contribution spreads uniformly over

h d i 0 0 9 8 088962 4 009 bli h d li b C b id i i

 Discetti: Time–Frequency Analysis and Wavelets 111

Figure 5.13 Representation of the two-level DWT coefficients with the Haar wavelet. The
coefficients are arranged in a 500 × 1000 matrix, i.e., with the same format of the data in input.

the set of wavelet coefficients; nonetheless, the coefficients of the signal are typically
rapidly decreasing in magnitude along the spectrum of scale, thus paving to way to
data compression and filtering by retaining only the most important ones. One of the
advantages of using wavelets for this purpose is their capability to preserve edges,
which is one of the most desired features in image compression.
In this example, DWT is used to analyze a flow field and synthesize it after a
thresholding procedure to remove the noise.
The test case is the flow in the wake of three cylinders with diameter D, arranged
with the axis on the vertices of an equilateral triangle with side length equal to 3D/2.
The downstream side of the triangle is orthogonal to the freestream flow, and it is
located at x = 0, centered on the y-axis. This configuration is known as fluidic pinball
(Deng et al. 2020), and is a very interesting test case for flow control applications
(Raibaudo et al. 2020).
An instantaneous flow field from DNS data at Re = 130 (referred to as a chaotic
regime (Deng et al. 2020)) is considered as a reference signal for the following
example. For simplicity, only the streamwise component of the velocity field is
analyzed.

Example 4: Filtering and compression of velocity fields

Consider the instantaneous flow field provided in the file pinball 000001.mat,
available in the supplementary material. The file contains the matrices U and
V, being respectively the streamwise and crosswise velocity fields on a 500 ×
1000 points grid (obtained after interpolation of the data from the original
DNS grid). The grid is stored in the file GridStruc.mat in the form of matrices
X and Y.
1. Superpose to the streamwise velocity field an additive Gaussian noise,
with standard deviation equal to 0.05U∞ (with U∞ being the freestream

h d i 0 0 9 8 088962 4 009 bli h d li b C b id i i

 112 Data-Driven Fluid Mechanics Chapter 5

Figure 5.14 Comparison of the reference field of a streamwise velocity component, a field
contaminated with noise, and fields filtered via DWT and FFT.

velocity). Compute a 2-level decomposition of the perturbed streamwise

velocity field with the Haar wavelet, and visualize the approximation and
detail coefficients of the two levels.
2. Compute a 4-level DWT with a Daubechies wavelet of the 4th order, and
threshold the wavelet coefficients to retain only the most significant 2%.
Compare the results with an FFT-based filtering of your choice.

Solution
1. The wavelet coefficients can be easily computed using standard packages
for wavelet calculations, available for a wide variety of programming lan-
guages. The solution presented here uses the MATLAB functions wavedec2,
appcoef2, and detcoef2, which compute, respectively, the full set of coeffi-
cients, the approximation, and the detail coefficients. The approximation
and detail coefficients have been scaled to 8-bit level representation, and
arranged in the same image for visualization purposes. The coefficients of
the 2-level decomposition of the perturbed streamwise velocity field are
reported in Figure 5.13. The image is divided into quadrants: the top-right
and bottom-left are the coefficients representative of the relevant horizontal
and vertical edges, respectively, at the first level; the bottom-right includes
the coefficients of the relevant scales on the diagonal direction at level 1; the

h d i 0 0 9 8 088962 4 009 bli h d li b C b id i i

 Discetti: Time–Frequency Analysis and Wavelets 113

top-left quadrant includes details and approximation of level 2, following

the same scheme as level 1.
It is clear that the detail coefficients of level 1 are mostly including noise,
and in part the regions of strong gradients close to the cylinders. At the
subsequent level, the details start including also some features of the shear
layers in the wake. The approximation in level 2 is a low-pass filtered
version of the original field (included in Figure 5.14 for reference).
2. The noisy field has been filtered using a 4-level approximation with a 4th-
order Daubechies wavelet, retaining only 2% of the coefficients with a soft-
thresholding procedure (Donoho & Johnstone 1994, Donoho 1995). An
FFT-based filtering procedure with a similar amount of retained coefficients
is carried out and included in Figure 5.14. The wavelet-based filter provides
slightly better results, with the relatively small margin possibly due to the
simplicity of this low-Reynolds number flow field.
The full solution is provided in the supplementary material in the form of
MATLAB code.

5.5 Conclusions

In this chapter, the fundamentals of time–frequency analysis have been reviewed. The
simple, yet powerful, concepts of scaling and shifting of windows to localize in time
the frequency content of the signal have been introduced in the framework of the WFT.
The consequences of Heisenberg’s uncertainty principle on the localization precision
of features in the time and frequency domains have been outlined.
The MRA with wavelets allows storing approximations of the signal at differ-
ent scales, thus configuring tunable time–frequency resolution. In this chapter, an
overview of the basic concepts behind the CWT and the DWT has been outlined,
together with the description of a procedure to determine a suitable family of
orthogonal wavelets. Some hints for the implementation of the DWT using two-
channel multirate filter banks have also been included.
The multi-resolution capabilities of wavelets are exploited in data compression
(e.g., the JPEG2000 algorithm is based on wavelets), denoising, speech recognition,
finance, and several areas of physics (Van den Berg 2004). In general, all signals
arising from multiscale processes are prone to effective use of wavelets. Turbulence
is clearly part of this paradigm. In the first age of wavelet use in turbulence, the
main applications were mostly (but not exclusively) targeted to turbulence model-
ing, efficient methods for computational fluid mechanics, and coherent structures
segmentation. In the coming years, it is foreseeable that the interest in multiscale
time–frequency methods will be further fostered by the current trend toward modal
decompositions blending energy optimality and spectral analysis (see e.g., Towne et al.
2018, Mendez, Balabane & Buchlin 2019, Floryan & Graham 2021).

h d i 0 0 9 8 088962 4 009 bli h d li b C b id i i

h d i 0 0 9 8 088962 4 009 bli h d li b C b id i i
 Part III
Data-Driven Decompositions

h d i 0 0 9 8 088962 4 0 0 bli h d li b C b id i i
h d i 0 0 9 8 088962 4 0 0 bli h d li b C b id i i
 6 The Proper Orthogonal
Decomposition
S. Dawson

The proper orthogonal decomposition (POD) is one of the most ubiquitous data
analysis and modeling techniques in fluid mechanics. Since many of the properties of
the POD are inherited from the singular value decomposition (SVD), we start with a
discussion of the SVD and describe those of its properties that are particularly useful
for understanding the POD. Our discussion of the POD starts by characterizing the
POD as a decomposition that is specific to a given data set before discussing how the
same concept arises when considering dynamical systems that are continuous in space
and time. We will describe several variants of the POD that have emerged over the
last half a century and how they are related, such as spectral and space-only POD. As
well as giving a broad overview, we will discuss some of the technical details often
omitted and/or taken for granted when POD is applied in practice, such as how to
incorporate nonstandard inner product weights. We finish by using a simple example
to demonstrate properties and methods of implementation for the POD.

6.1 Introduction

The POD is one of the most widely used techniques for data analysis in fluid
mechanics. Over the past 50 years, its usage in the community has grown and evolved
alongside developments in experimental measurement techniques, the introduction
and rapid development of computational methods for simulating fluid flows, theoreti-
cal developments in dynamical systems, and the ability to store and process increasing
quantities of data feasibly. We now discuss the broad motivation for POD and related
techniques. Suppose we have data that is a function of both space and time y(x, t).
A logical first step toward understanding the salient features of the data, and perhaps
the dynamics of the underlying system responsible for its generation, is to perform a
separation of variables:

m
Õ
y(x, t) = φ j (x)a j (t). (6.1)
j=1

That is, we decompose the data into a sum of spatial modes φ j (x), and their time-
varying coefficients (or equivalently, temporal modes) a j (t). There are many choices

h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
 118 Data-Driven Fluid Mechanics Chapter 6

for such a decomposition. For example, one might choose to perform a Fourier
transform in space or time, thus obtaining a Fourier basis. Rather than using a
predefined basis, POD chooses a decomposition based on the data itself, though we
will see later that in certain cases such data-driven modes can actually converge to
Fourier modes. Note that we will also see that this decomposition can be further
generalized to allow for time-dependency in the spatial structures.
Since many of the properties of POD are inherited from the SVD, we start with a
discussion of the SVD and its properties more generally, before discussing how it can
be utilized to define and compute POD of a given data set.

6.2 The Singular Value Decomposition

In short, POD can be viewed as the result of taking the SVD of a suitably arranged data
matrix. Because of this, and as many of the properties of POD are inherited directly
from those of the SVD, we start with a discussion of the SVD and its properties more
generally, before discussing how it can be utilized to define and compute POD upon a
given data set. This section might seem like a lot of rather dry mathematics, but it will
all end up being useful later in this chapter.

6.2.1 The Singular Value Decomposition: Definition

Suppose we have a matrix X that has n rows and m columns. Most of the time, if X is
a data matrix then its entries are real, though here for generality we assume that they
may be complex. Suppose that X has rank q ≤ min(m, n). The (reduced) SVD of X
gives
X = ΦΣΨ∗, (6.2)
where Φ ∈ Cn×q , Ψ ∈ Cm×q , and Σ ∈ Rq×q is a diagonal matrix with diagonal entries
consisting of the singular values σ1 ≥ σ2 ≥ · · · ≥ σq > 0 (the σq > 0 comes from
the fact that we have truncated the SVD to the rank of X). Additionally, the columns
of Φ and Ψ are orthonormal, so Φ∗ Φ = Ψ∗ Ψ = Iq , where Iq is a q × q identity
matrix.1 Note in particular that the SVD exists for all matrices. This is in contrast
with, for example, an eigendecomposition, which only exists for diagonalizable square
matrices.
Thus far, we have not explained the meaning of ·∗ , which denotes the adjoint of a
matrix. Most typically, this denotes the conjugate transpose (or just the transpose in
the case of a real matrix), but for reasons that will become apparent later in the chapter,
we want to keep its definition more general. To define the adjoint, we can first think
about the matrix X as mapping from a vector in Cm to a vector in Cn through matrix-
vector multiplication. Throughout this chapter, all vector quantities can be thought of
as column vectors, unless otherwise specified. That is, z ∈ Cm maps to y ∈ Cn via
1 Note that the definition of a reduced SVD given here differs from the full SVD, which defines Φ and Ψ
as being square unitary matrices, with Σ appropriately padded with zeros.

h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
 Dawson: The Proper Orthogonal Decomposition 119

y = X z. If we assume that we have inner products defined on Cm and Cn by hz 1, z 2 i m

and hy1, y2 i n , respectively, then the adjoint of a matrix A is defined as the matrix that
satisfies
hAz 1, y2 i n = hz 1, A∗ y2 i m . (6.3)

That is, the adjoint is what you get if you move an operator to the other side of an
inner product. As an aside, note that (6.3) can be very useful in its own right: if one
of m or n is much smaller than the other, then it can be substantially easier to evaluate
this inner product in the lower-dimensional space. Note also that this definition of an
adjoint reduces to the conjugate transpose matrix in the case where the inner products
are the “standard”, that is,

hz 1, z 2 i m = z 2† z 1, hy1, y2 i n = y2† y1, (6.4)

where ·† denotes the conjugate transpose.

6.2.2 The Singular Value Decomposition: Properties

This section lists various properties of the SVD that will be utilized in this chapter. We
do not give proofs for all of these properties, and interested readers are encouraged to
consult standard references in linear algebra (e.g., Van Loan & Golub 1983, Strang
1988, Horn & Johnson 2012) for further details.

Optimality
The best rank-r approximation to a matrix X (in a least-squares sense) is obtained by
taking the first r components of the SVD of X. That is, if Xr is a rank-r approximation
of X, we have
argmin k X − Xr kF = Φr Σr Ψr∗ , (6.5)
Xr

where Φr and Ψr are the first r columns of Φ and Ψ in the SVD of X, and Σr is
a diagonal matrix consisting of the first r singular values of X. Here the subscript F
denotes the Frobenius norm, given by
v v
u
u
tÕ n Õ m tmin(n,m)
u
2 Õ
k X kF = X jk = σj2, (6.6)

 
j=1 k=1 j=1

which is the same as the Euclidean norm if we were to squeeze the entries of the
matrix X into a single long vector. As will be seen later, this property is responsible
for the “optimality” of the POD. We also observe from the right inequality in
(6.6) a direct connection between the Frobenius norm and the singular values of a
matrix. Intuitively, this connection exists because the singular values contain all of
the information about the size of the components in the matrix X, with the Φ and
Ψ matrices containing orthonormal vectors. Looking again at the quantity that is

h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
 120 Data-Driven Fluid Mechanics Chapter 6

minimized on the left-hand side of (6.5), this connection between the Frobenius norm
and the SVD also leads to
v
u
tmin(n,m)
u
Õ
min k X − Xr kF = σj2 . (6.7)
Xr
j=r+1

As an additional aside, note that we could also consider the L2 operator norm of a
matrix, defined by
k X k2 = max k Xψk2 = σ1, (6.8)
kψ k2 =1

with the second equality arising because the unit vector ψ that maximizes k Xψk2 is
the first right singular vector of X, ψ 1 . Moreover, note that (6.5) also holds when the
Frobenius norm is replaced by the operator L2 norm. In this case, the equivalent of
(6.7) is
min k X − Xr k2 = σr+1 . (6.9)
Xr

Relationship between left and right singular vectors

Let φ j and ψ j denote the jth left and right singular vectors (i.e., the jth columns of Φ
and Ψ) of a matrix X, respectively, with corresponding singular value σj . These are
related by
σj φ j = Xψ j , (6.10)
σj ψ j = X φ j .∗
(6.11)
These relationships follow from the fact that XΨ = ΦΣΨ Ψ = ΦΣ, and similarly ∗

X ∗ Φ = ΨΣΦ∗ Φ = ΨΣ.

Relationship to eigendecompositions
The left and right singular vectors are eigenvectors of the matrices X X ∗ and X ∗ X
respectively, with
X X ∗ φ j = σj2 φ j , (6.12)
X Xψ j =
∗
σj2 ψ j . (6.13)
These relationships readily follow from (6.10) and (6.11). Note also that the cor-
relation matrices X X ∗ and X ∗ X are both Hermitian,2 since (X X ∗ )∗ = X X ∗ and
(X ∗ X)∗ = X ∗ X.

Permutation of rows and columns

Suppose we want to rearrange the rows of X. We can represent this rearrangement via
left multiplication of X by a permutation matrix P (i.e., P is a matrix with exactly one
in each row and column, and zeros elsewhere). An SVD of PX is given by
PX = (PΦ)ΣΨ∗ . (6.14)
2 Recall that a Hermitian (equivalently, self-adjoint) matrix A satisfies A = A∗ .

h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
 Dawson: The Proper Orthogonal Decomposition 121

Similarly, if Q is a permutation matrix that rearranges the columns of X via right

multiplication, then XQ has the SVD

XQ = ΦΣ(ΨQ)∗ . (6.15)

Dyadic expansion
The SVD can be expanded as a sum of rank-1 matrices via
q
Õ
X = ΦΣΨ = ∗
σj φ j ψ ∗j . (6.16)
j=1

Computing the SVD with a weighted inner product

In practice, the numerical computation of the SVD is most typically done using inbuilt
routines (e.g., numpy.linalg.svd in Python). Such routines implicitly assume that the
SVD is to be computed with standard inner products for both the columns and rows of
the matrix, as in (6.4). This section details how such standard routines may be utilized
to perform the SVD with a nonstandard inner product. In particular, suppose that we
replace the standard inner products by

hz 1, z 2 i m = z 2†Wm z 1, hy1, y2 i n = y2†Wn y1, (6.17)

where as before z j and y j are rows and columns of the matrix, respectively. In order
for these to satisfy the definition of an inner product, the weight matrices Wm and
Wn must be positive definite. Note that the adjoint of X with these inner products is
given by
X ∗ = Wm−1 X † Wn, (6.18)

which comes from the fact that

hy, X z, i n = z † X † Wn y
= z † Wm Wm−1 X † Wn y
= (Wm−1 X † Wn )y, z m



:= hX ∗ y, zi m .

Wn and Wm are often positive diagonal matrices, where the diagonal entries contain
integration weights. Henceforth, we assume that this is the case, which makes the
definition of their square roots Wm1/2 and Wn1/2 unambiguous. The SVD of X with
these inner products can be computed as follows:

1. Form the weighted matrix Xw = Wn1/2 XWm−1/2 .

2. Compute the SVD of the weighted matrix Xw = Φw ΣΨ†w using standard routines.
3. The singular vectors (which are orthonormal with respect to their respective norms)
are given by columns φ j and ψ j of Φ = Wn−1/2 Φw and Ψ = Wm−1/2 Ψw , respectively,
with corresponding singular values as given in Σ.

h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
 122 Data-Driven Fluid Mechanics Chapter 6

With this computation, it is easy to verify that the identified singular vectors are
orthonormal with respect to their respective inner products (Φ† Wn Φ = I, Ψ†Wm Ψ =
I), and that they satisfy the eigenvalue problems discussed in Section 6.2.2 for the
general inner products and adjoints:
Xw Xw† (φw ) j = σj2 (φw ) j ,
(Wn1/2 XWm−1/2 )(Wm−1/2 X † Wn1/2 )(Wn1/2 φ j ) = σj Wn1/2 φ j ,
X(Wm−1 X † Wn )φ j = σj2 φ j ,
X X ∗ φ j = σj2 φ j ,
and similarly
Xw† Xw (ψ w ) j = σj2 (ψ w ) j ,
(Wm−1/2 X † Wn1/2 )(Wn1/2 XWm−1/2 )(Wm1/2 ψ j ) = σj Wm1/2 ψ j ,
(Wm−1 X † Wn )Xψ j = σj2 ψ j ,
X ∗ Xψ j = σj2 ψ j ,
where here (φw ) j and (ψ w ) j are the jth columns of Φw and Ψw , respectively.

6.3 The Proper Orthogonal Decomposition of Discrete Data

Suppose that we measure data that is a function of both space and time, and assemble
all data into a matrix, such that each column of the matrix represents a “snapshot” of
all data that is measured at a given instance in time, while each row consists of all
of the measurements of a given quantity across all times. For example, if we measure
two components of the velocity of a fluid (u and v) at spatial locations x1, x2, . . . , xn ,
and at times t1, t2, . . . , tm , then the data matrix becomes the n × m matrix
 u(x1, t1 ) u(x1, t2 ) ··· u(x1, tm ) 

 u(x2, t1 ) u(x2, t2 ) ··· u(x2, tm ) 
 . .. ..

 .. .. 
 . . . 

u(xn, t1 ) u(xn, t2 ) ··· u(xn, tm )
 
X= , (6.19)
 v(x1, t1 ) v(x1, t2 ) ··· v(x1, tm ) 
 v(x2, t1 ) v(x2, t2 ) v(x2, tm ) 
 
···
 .. .. .. ..

 . . . .


 
v(xn, t1 ) v(xn, t2 ) ··· v(xn, tm )

where n = 2nx , since there are two measurements for each spatial location. Note
that we are not making any assumptions concerning the spatial or temporal resolution
of our data, or in what quantity or quantities are being measured. The only implicit
assumption that we are making in forming X is that we measure the same quantities
for each snapshot. We next subtract from each column of X a base condition y0 . Most
typically, this is the mean of all of the columns in X, though it need not always be.

h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
 Dawson: The Proper Orthogonal Decomposition 123

If X represents the transient response of a system to a given input, it would make sense
to consider the steady or average state as t → ∞ instead of the mean of the early-time
data. For notational convenience, we will refer to this base state that is subtracted from
the data as the “mean” in any case. Denoting the kth column of X by yk (x), we now
form the mean-subtracted matrix

Y = y10 y20 · · · ym 0 , (6.20)

 

where y j0 = y j − y0 is the jth mean-subtracted snapshot. Most simply, the POD can be
found from taking the SVD of Y . In particular, if we have

Y = ΦΣΨ∗, (6.21)

then POD modes are given by the columns φ j (x) of Φ, which evolve in time via
the corresponding coefficients a j (t) = σj ψ j (t)∗ (and for the case of real data with
a standard inner product, ψ j (t)∗ = ψ j (t)T ). That is, with reference to (6.1), a POD
expansion of the data is given by the dyadic expansion of the SVD as described in
Section 6.2.2:
Õq
yk (x) = φ j (x)σj ψ ∗j (tk ). (6.22)
j=1

Since we have arranged our data such that spatial location depends only on the row
number, and temporal location depends only on the column, the left and right singular
vectors depend only on space and time, respectively, allowing this direct use of the
SVD to obtain this separation of variables in our data. The optimality property of the
SVD described in Section 6.2.2 means that a truncation of this sum to r terms gives
the closest rank-r approximation to the full data set.
When assembling the data matrix in (6.19), we chose to place the measurements of
the v-component of velocity below all of the measurements of u, though we could have
also arranged the data in other ways, for example, with alternating rows containing the
u and v components of velocity. From the permutation properties of the SVD discussed
in Section 6.2.2, changing the arrangement does not affect the decomposition, aside
from permuting the entries in the POD modes to the appropriate locations. Similarly,
we can rearrange the columns of the data matrix without affecting the POD modes
at all. Indeed, this fact can be used to demonstrate that POD does not require time-
resolved data, given that shuffling the order of time-resolved snapshots can give a
non-time-resolved sequence of snapshots.
From Section 6.2.2, the right singular vectors of Y are eigenvectors of the matrix
Y Y . This m × m matrix can be interpreted as the correlation matrix between all
∗

snapshots:
 y10 , y10 x y2, y1 x · · · y , y1 x 


0 0
0 0

m
 y ,y y2, y2 x · · · 0 , y0

0 0
0 0  
 1 2 x ym 2 x

Y Y=
∗
. . .. . , (6.23)
. .. . ..

. 
 

 y 0, y 0 y2, y m x · · ·

0 0
y m, y m x 

0 0 
 1 m x

h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
 124 Data-Driven Fluid Mechanics Chapter 6

where h·, ·i x denotes a spatial inner product. Similarly, if we let z 0j (t) be a row vector
that is the jth row of Y , we have
 z 10 , z 10 t z ,z z ,z 


0 0
0 0
···

0 0 
10 20  t
10 n0  t 
 z ,z z 2, z 2 t · · · z 2, z n t 
 2 1 t
YY ∗ =  .. .. .. .. , (6.24)

. 
. 
 . . 

 z0 , z0 z n0 , z 20 t · · · z n, z n t 

0 0 
 n 1 t
where h·, ·i t is an inner product in the time dimension. Again from Section 6.2.2,
the POD modes φ j are eigenvectors of the matrix YY ∗ . This n × n matrix can
be interpreted as the time-correlation matrix between pairwise rows in Y .
Note that the two matrices YY ∗ and Y ∗ Y are generally of different sizes, with
YY ∗ being n × n and Y ∗ Y being m × m. Since the SVD of Y is related to the
eigendecompositions of these square matrices, it might be easiest to compute and work
with the smaller of these two matrices. For example, if one had very large snapshots
but comparatively few of them, then m  n, and so it can be more tractable to compute
the (full or partial) eigendecomposition of Y ∗ Y to obtain the POD coefficients a j (t).
The POD modes can then be computed using φ j = σj−1Y a j , following the properties
of the SVD in Section 6.2.2. This is precisely the “method of snapshots” for computing
POD introduced in Sirovich (1987). Conversely, if n  m, then one can instead start
by computing an eigendecomposition of YY ∗ .

6.4 The Proper Orthogonal Decomposition of Continuous

Systems

Up until this point, we have viewed the POD as being something that is computed
from a given data set. In reality, however, if our data comes from measurements of
a given system, then the POD should ideally be independent of the exact data that
we have collected and used for analysis. That is to say, we can alternatively think of
POD as being a property of a system, which we hope to accurately compute from data.
Looking back at the inner products of discrete data that feature in (6.23) and (6.24),
these can be viewed as discrete approximations to the integrals
D E ∫
y j0, yk0 = y10 (x, t j )y20 (x, tk )dx, (6.25)
x Ωx
D E ∫
z 0j , z k0 = z 10 (x j , t)z 20 (xk , t)dt, (6.26)
t Ωt

where the first equation is the spatial integral over the appropriate number of spatial
dimensions, and Ωx and Ωt define the spatial and temporal domains of interest.
Note that a desire to correctly approximate these integrals can lead to the need to
include integration weights for discrete data as described in Section 6.2.2.3 Note that,

3 To be more general, we also could have included an additional weight function in these integrals, but we
omit this for simplicity.

h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
 Dawson: The Proper Orthogonal Decomposition 125

depending on the context, sometimes it is most convenient to divide the integrals in

(6.25) and (6.26) by the size of the domains of integration, so that the inner products
become averages/expectations. More generally, we can think about (6.25) and (6.26)
as continuous functions of time and space, respectively (rather than being computed
for the discrete indices j and k). We can also be slightly more general, and think of
these integrals as computing an expected value of some stochastic process. With this
in mind, we can define the spatial correlation function

C(x, x 0) = E[ ȳ 0(x, t)y 0(x 0, t)], (6.27)

where E is the expectation, taken over time and/or ensembles. The POD in a
continuous setting amounts to finding modes φ j (x) that are eigenfunctions of the
integral equation
∫
C(x, x 0)φ j (x 0)dx 0 = σj2 φ j (x). (6.28)
Ωx

As is perhaps indicated by the notation used, the eigenvalues σj2 are real and positive,
and this equation is the continuous equivalent to (6.12). Note that it can be shown that
many of the same properties of the POD discussed in the discrete, finite-dimensional
context, such as orthogonality and optimality, also extend to equivalent concepts in
the continuous setting. In cases where the system is spatially homogenous (i.e., is
invariant under translations), C(x, x 0) = C(x − x 0) is only a function of x − x 0, and it
can readily be shown from (6.28) that the POD modes become Fourier modes. To see
this, if we change the variable of integration to r = x − x 0, (6.28) can be expressed as
φ j (x + r)
∫
C(r) dr = σj2 . (6.29)
Ωr φ j (x)

The right-hand side of (6.29) is independent of x, so eigenfunctions φ j must possess

φ (x+r)
the property that φj (x) is independent of x. We can verify that Fourier modes
j
(i.e., exp(2πik x)) satisfy this property. In the case where the domain is finite in the
direction on homogeneity, periodic boundary conditions determine the permissible
values of the spatial wavenumbers, k. Similarly, if the system is homogeneous in only
certain directions (such as the streamwise and spanwise directions in channel flow),
then POD modes become Fourier modes in these directions of spatial homogeneity.
Rather than only looking at the spatial correlation tensor, we can also consider the
space-time correlation tensor

C(x, x 0, t, t 0) = E[ ȳ 0(x, t)y 0(x 0, t 0)], (6.30)

where now the expectation can be thought of as an ensemble average across different
realizations of a stochastic field for all pairs of space-time coordinates (x, t) and (x 0, t 0).
An equivalent to (6.28) in this space-time setting is then given by
∫
C(x, x 0, t, t 0)φ̃ j (x 0, t 0)dx 0 dt 0 = σj2 φ̃ j (x, t). (6.31)
Ω x ,Ωt

h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
 126 Data-Driven Fluid Mechanics Chapter 6

Note in particular that the eigenfunctions of this equation are now functions of
both space and time, which we distinguish from the space-only modes that we have
dealt with so far using a ˜·. Most typically, systems modeled using this approach are
stationary in time, meaning that C(x, x 0, t, t 0) = C(x, x; , t − t 0) is only a function of the
difference in time τ = t − t 0. Similar to the case of spatial homogeneity, this means
that the temporal dependence on these modes become Fourier modes. This allows for
a temporal Fourier transform of (6.31), giving
∫
S(x, x 0, f )φ̂ j (x 0, f )dx 0 = σj ( f )2 φ̂ j (x, f ), (6.32)
Ωx

where the cross-spectral density tensor S(x, x 0, f ) is obtained from the temporal
Fourier transforms:
∫ ∞
S(x, x , f ) =
0
C(x, x 0, τ) exp(−i2π f τ)dτ. (6.33)
∞
This space-time or spectral POD was, in fact, the original formulation of POD as
described in Lumley (1967) and Lumley (1970). However, the “space-only” POD
has been much more commonly used since then, likely due to two main reasons:
it is easy to formulate and compute from data, and it is readily amenable for the
construction of projection-based reduced-order models (e.g., Holmes et al. 2012), as
will be discussed in Section 6.5. Indeed, the difference between the original and most
commonly used formulations has perhaps been under appreciated in the community
(including by the author). Recently, the work of Towne et al. (2018) has brought the
“spectral” version of POD back to prominence, though an exposition of the method
and its close connections to other modeling approaches, and in the development of
practical algorithms for its implementation (Schmidt & Towne 2019).

6.5 The Proper Orthogonal Decomposition for Projection-Based

Modeling

Perhaps the most common usage of the POD involves using a truncated basis of POD
modes to obtain a subspace that differential equations describing the physics of a
system may be projected onto. For example, if we consider the incompressible Navier–
Stokes momentum equation
∂u
= −(u · ∇)u + Re−1 ∆u − ∇p, (6.34)
∂t
then taking the spatial inner product of both sides of this equation with a POD mode
φ j (x) gives
∂u
 
, φ j = − (u · ∇)u, φ j x + Re−1 ∆u, φ j x − ∇p, φ j x . (6.35)




∂t x
Approximating the full velocity field with a set of POD modes (as in (6.1)), we obtain
aÛ = L a + Q(a, a) + f , (6.36)

h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
 Dawson: The Proper Orthogonal Decomposition 127

where L is a linear operator, Q is a bilinear operator, and f is a vector, each defined

based on the identified spatial POD modes by
Li j = ν ∆φ j , φi x , Qi jk = − (φ j · ∇)φk , φi x , f = − ∇p, φi x . (6.37)





This gives a means of approximating the Navier–Stokes equations by a set of nonlinear

ordinary differential equations, which is typically of much lower dimension than the
full discretized equations.

6.6 Extensions of the Proper Orthogonal Decomposition

There are many modifications and extensions of the basic POD definition and
algorithm that can improve on its suitability, accuracy, and efficiency in various
applications. If being used for reduced-order modeling as described in Section 6.5,
a subspace based on POD modes is optimal from an energetic sense, but is not
necessarily the best choice for retaining all features of dynamic importance. For the
case of linear systems, one can seek to define the subspace and direction of projection
such that they best preserve the system dynamics (quantified by the observability and
controllability Gramians). This is known as balanced truncation (Moore 1981), and it
was shown by Rowley (2005) that these subspaces can be identified from POD using
inner products that are weighted by the observability and controllability Gramians (see
also the similar method developed in Willcox and Peraire (2002)). Further discussion
of notions of these concepts are given in Chapter 10.
Several other variants of POD account for other desired qualities in spatio-
temporal decompositions, such as finding a balance between energetic optimality and
frequency localization (Sieber et al. 2016), and seeking scale separation (Mendez
et al. 2019). Additional variants aim to reparameterize the spatio-temporal domain
with translations, rotations, and/or rescalings to make certain data/systems (often
those dominated by convection) more amenable to low-dimensional approximation
(Rowley & Marsden 2000, Rowley et al. 2003, Mojgani & Balajewicz 2017, Mowlavi
& Sapsis 2018, Reiss et al. 2018, Black et al. 2019, Mendible et al. 2020).
Aside from these variants, computation of POD can require overcoming several
practical challenges involving the acquisition and processing of data of sufficient
quantity and quality to ensure converged results. For further discussion of such
practical considerations for experimental data in particular, see Chapter 9.

6.7 Examples

This section will present two simple examples to show how the POD can be applied,
and its outputs interpreted. Both of these examples come with accompanying Python
code, available on the book’s website.4 The example considered here is similar to an

4 www.datadrivenfluidmechanics.com/download/book/chapter6.zip

h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
 128 Data-Driven Fluid Mechanics Chapter 6

y(x, t)
2.0 1.6
1.2
Figure 6.1 Surface plot of
1.5 0.8
data from (6.38).
0.4
1.0 0.0
t

–0.4
0.5 –0.8
–1.2
0.0 –1.6
–2 –1 0 1 2
x

103

100 Figure 6.2 Singular values

corresponding to each POD
Singular value

10 –3 mode for data from (6.38).

10 –6

10 –9

10 –12

0 2 4 6 8 10
Mode index

example considered in (Tu et al. 2014). We consider data that is sampled from the
function
" # " #
(x − x1 )2 (x − x2 )2
y(x, t) = c1 exp sin(2π f1 t) + c2 exp sin(2π f2 t), (6.38)
2s12 2s22

where we choose the parameters c1 = 1, x1 = 0.5, s1 = 0.6, f1 = 1.3, c2 = 1.2,

x2 = −0.5, s2 = 0.3, and f2 = 4.1. We collect data in the range x ∈ [−2, 2] and
t ∈ [0, 50]. A visualization of this function over four time units is shown in Figure 6.1.
We perform POD by compiling the data in a matrix and performing a SVD, using
the procedure discussed in Section 6.3. The resulting singular values, spatial POD
modes, and corresponding temporal coefficients are shown in Figures 6.2 through 6.4,
respectively. The leading two singular values are much larger than the rest (which are
essentially zero), due to the fact that the data comes from a function that is the sum
of two terms (with each the product of a spatial function and a temporal function).

h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
 Dawson: The Proper Orthogonal Decomposition 129

0.10 Mode 1
POD mode shapes Mode 2 Figure 6.3 POD mode
shapes computed from data
0.05 from (6.38).

0.00

–0.05

–2 –1 0 1 2
x

0.10 Mode 1
Mode 2 Figure 6.4 POD mode
Mode coefficients

coefficients for data from

0.05
(6.38).

0.00

−0.05

−0.10
0 1 2 3 4
t

However, the leading two POD modes do not cleanly partition the two component
functions, and instead each mode is a mix of the two. This is due to the fact that the
SVD is “greedy” in the sense that the first set of singular vectors captures as much
energy as possible, which in this case means taking part of each of the two terms in
(6.38).
In Figure 6.5, we show the results of applying spectral POD to the same data set,
which we see is able to distinguish between the two spatial functions corresponding to
different temporal frequencies, both of which show up as peaks in the corresponding
spectrum, shown in Figure 6.6. Note that here a simple Fourier transform in time
would also suffice, since the data is entirely deterministic and free of noise. Here we
could also separate out the frequency content using dynamic mode decomposition on
appropriately arranged data.
The next example that we consider involves data from a fluid simulation, consid-
ering two-dimensional flow over a circular cylinder at a Reynolds number of 60.

h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
 130 Data-Driven Fluid Mechanics Chapter 6

Spectral POD modes Figure 6.5 Spectral POD

600
Mode 1 mode shapes computed from
Mode 2 data generated from (6.38).
400

200

0
–2 –1 0 1 2
y

Spectral content Figure 6.6 Spectral POD

6
energy content of data
5 generated from (6.38).
Frequency (rad/s)

0
–2 –1 0 1 2
x

A similar example has already been introduced and discussed in Chapter 1. We

consider 41 snapshots of data collected over approximately eight convective times
(being the time taken for fluid in the freestream to travel one cylinder diameter).
The data is collected when the system is in a transient state, close to the unstable
equilibrium solution, but with oscillations (corresponding to vortex shedding) that are
slowly growing in time.
The accompanying Python code shows how the data for both the streamwise
and transverse velocity components can be arranged as in (6.19); how POD can
be performed via an SVD; and how the resulting modes and mode coefficients can
be visualized. The first two modes are shown in Figures 6.7 and 6.8, with their
accompanying time-varying coefficients shown in Figure 6.9. We see that these
leading two modes are similar in structure, featuring oscillatory structures in the
streamwise (horizontal) direction, which have a phase difference between modes 1
and 2 (with a similar phase difference observed for the mode coefficients in time).

h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
 Dawson: The Proper Orthogonal Decomposition 131

Mode 1, u

Figure 6.7 Streamwise (u)

and transverse (v) velocity
components of the leading
Mode 1, v
POD mode for flow over a
cylinder in a transient
regime.

Mode 2, u

Figure 6.8 Streamwise (u)

and transverse (v) velocity
components of the second
POD mode for flow over a
Mode 2, v cylinder in a transient
regime.

This phase difference between both the streamwise-oscillating structures in space and
coefficients in time is approximately π/2. This can be reasoned from the fact that two
purely oscillatory signals at the same frequency will only be orthogonal if they are
exactly a quarter-cycle out of phase. These mode shapes, which can be thought of
as standing waves, allow for the representation of structures that convect downstream
over time.
Lower-energy modes (not shown) include modes that oscillate at harmonic frequen-
cies of these leading modes, as well as modes that capture the distortion of the base
flow as the system moves between the unstable equilibrium and limit cycle of vortex
shedding (as discussed in Chapter 1). Note that had we considered data for a different
section of the transient (or on the limit cycle), these leading two POD modes would
be slightly different, though would share the same qualitative features. In other words,
on this transient trajectory, the system is not statistically stationary, so the POD that
we obtain is a function of the data set that we use.

h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
 132 Data-Driven Fluid Mechanics Chapter 6

Mode 1
1.0
Mode coefficient s Mode 2

0.5

Figure 6.9 Time-varying

0.0
coefficients corresponding to
the POD modes shown in
–0.5
Figures 6.7 and 6.8.
–1.0
0 2 4 6 8
t

h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
 7 The Dynamic Mode
Decomposition: From Koopman
Theory to Applications
P. J. Schmid

The dynamic mode decomposition (DMD) is a data-driven method for flow-field

analysis, closely linked to a powerful mathematical framework for dynamical systems
known as Koopman theory. It provides approximations to the key elements of
Koopman analysis and has spawned many developments from its original formulation.
This chapter gives a brief derivation and outline of this decomposition method,
presents three applications to experimental and numerical data sequences, and touches
on some of the many exciting extensions and generalizations.

7.1 Introduction

Nonlinear dynamical systems are a common mathematical concept to describe many

evolution processes in the physical, biological, and social sciences. The motion of
fluids across all spatial and temporal scales is represented by a large-scale dynamical
system that arises from a discretization of a partial differential equation. This
differential equation, in turn, is derived from first principles, such as the conservation
of mass, momentum, energy, and other appropriate flow markers. Our understanding
of physical processes is thus closely linked to the analysis of nonlinear dynamical
systems and to the tools available for this task. Over many decades, the mathematical
methods at our disposal have been dominated by linear techniques. Relying on linear
superposition, a complex process can be deconstructed into subprocesses that can be
studied in isolation. Furthermore, linear techniques can be applied locally near special
points of the solution space and furnish information about the disturbance behavior in
their vicinity.
At the same time, with the rise of new experimental techniques and massively
parallel computers, the amount of data within our reach is staggering, and the detail
of this data is impressive. The data-rich environment has spawned many methods
to extract essential information from large data files, to describe physical processes
and to quantify output variables of common interest. Besides statistical methodology,
data decomposition techniques have dominated the field of data-driven analysis.
These techniques aim at a factorization of the full spatio-temporal observations into
subunits of decreasing coherence. For each of the available decompositions, the

h d i 0 0 9 8 088962 4 0 2 bli h d li b C b id i i
 134 Data-Driven Fluid Mechanics Chapter 7

precise meaning of “coherence” has to be defined anew, and often has to be tailored to
the specific circumstances of the data and/or flow configuration.
The data-centric approach carries both advantages and disadvantages. It is simpler
to justify, particularly in an experimental setting, since the gathered data – by
definition – is observable; model-based computational approaches, on the other hand,
can suffer from robustness issues. Still, even in data-driven analysis, we can encounter
hidden (and sometimes high-dimensional) variables that are essential from a physical
point of view, but are difficult to measure or absent in our data. Data are often
noisy, which can pose challenges to the subsequent analysis (see also Chapter 13).
The uncertainty in the data has to be taken into account and balanced against both
further approximations in the decomposition. In particular, the uncertainty in the
output variables has to be determined and accounted for in the final interpretation
of the results.
Many common techniques in describing processes, based on the data they produce,
are based on spectral analysis (in the most general sense of the word). This involves a
transformation into another basis, during which some of the independent variables
(time, space) will be replaced by their dual equivalents (frequency, wavenumber).
Ideally, this transformation introduces a sparse, localized, and hierarchical description
of the physical process (contained in the data sequence) in the new basis – and allows
a more concentrated representation of a potentially complex process.
Among various decompositions, modal decompositions (in a general sense) are
most common (Lumley 1970, Berkooz et al. 1993, Holmes et al. 2012, Taira
et al. 2017). They are the data-driven equivalent of a separation-of-variables approach:
we factorize the data into purely spatial components (the modes), purely temporal
components (the dynamics), and the associated amplitudes (the spectrum); see also
Chapter 8 for additional material on linear decompositions. This procedure is predi-
cated on a linear assumption: we can recover the original signals by a superposition
of (some of) the identified mode–dynamics–amplitude triplets. For this hierarchical
breakup and reassembly, a great many decompositions are available; for nonlinear
systems, or a nonlinear analysis, the options are far more limited.
A promising way forward seeks to postulate and design a coordinate transforma-
tion, under which the nonlinear dynamics reduces to a linear one. This approach
has been studied from a mathematical point of view for many decades, but as an
exact transformation has found only limited applicability to a few nonlinear partial
differential equations within a model-based setting.
Koopman analysis falls within this latter category, as it proposes a nonlinear
coordinate transformation that embeds a nonlinear finite-dimensional dynamical
system into an equivalent linear system, albeit of infinite dimensions. The remainder
of this chapter will motivate data analysis from this point of view (Lasota & Mackey
1994, Mezić 2005, Budišić et al. 2012, Mezić 2013) and present a computational
framework for the decomposition of data sequences generated by a nonlinear system.
The link to Koopman theory is also developed in Chapter 10, and the reader is urged
to consult this material.

h d i 0 0 9 8 088962 4 0 2 bli h d li b C b id i i
 Schmid: The Dynamic Mode Decomposition 135

F F
qn qn+1 qn+2 (State space)

K K
φn φn+1 φn+2 (Observable space)

Figure 7.1 Schematic of the Koopman idea. We have used the abbreviaton φn = φ(qn ).

7.2 The Koopman Idea

Koopman analysis is a mathematical formalism for dynamical systems theory. Its

core principle is a coordinate transformation that embeds a nonlinear dynamical
system in an equivalent linear dynamical system of observables. As such it presents
an alternative to the classical viewpoint of dynamical systems (Guckenheimer &
Holmes 1983), given by a state vector q and a governing equation for its evolution
f or a discrete map F which together give the dynamical system
dq
= f(q) (continuous), (7.1a)
dt
qn+1 = F(qn ) (discrete). (7.1b)
Dynamical systems of this form have been analyzed by a variety of mathematical
techniques and are part of the standard mathematical curriculum (Guckenheimer &
Holmes 1983). Equilibrium points and their stability can be computed and analyzed,
limit cycles, Poincaré maps, attractors of various dimensions, and bifurcations
between them can be defined and determined. This type of analysis is based on the
geometry of phase space, but encounters limitations when the dimensionality of the
state vector becomes exceedingly large.
Koopman analysis takes a different and data-driven approach, arguing that the
observed data are simply function evaluations of an underlying state. Concentrating
on the observed data, Koopman analysis provides tools and a mathematical foundation
for the analysis of the state-vector evolution. To this end, we introduce an operator K
that maps the observables over one time-step (for the discrete case). We have
Kφ(qn ) = φ(qn+1 ) = φ(F(qn )), (7.2)
defining the Koopman operator K. It is straightforward to show that K is a linear
operator, and satisfies a semi-group property, that is, closedness and associativity
under its binary operation. Further advancement in time is given by a multiple
application of the Koopman operator K or powers thereof. We can think of the
Koopman formulation as a lifting procedure of the dynamics from a state-space
description to an observable-space description. Figure 7.1 gives a sketch of the two
viewpoints. While the dynamics in state-space is finite-dimensional but governed by
a nonlinear operator F, the dynamics in the observable space can be described by a
linear operator K, but is infinite-dimensional in nature.

h d i 0 0 9 8 088962 4 0 2 bli h d li b C b id i i
 136 Data-Driven Fluid Mechanics Chapter 7

The choice of observables under which our nonlinear dynamical system is trans-
formed into a linear system is the key issue in Koopman analysis. In effect, we
are looking for a coordinate transformation or an embedding φ(q) that renders the
evolution in φ linear. Two examples shall illustrate the main ideas of finding these
Koopman embeddings (Tu et al. 2014, Brunton & Kutz 2019).
The first example (Tu et al. 2014) is given by the two-dimensional nonlinear map
xn+1 = axn, (7.3a)
yn+1 = b(yn − xn2 ). (7.3b)
We choose the observable vector φn = (xn, yn, xn2 ).
Simple algebra confirms that a
linear mapping K links the observable vector φn to its successive realization φn+1
according to

© xn+1 ª ©a 0 0 ª © xn ª
­ yn+1 ® = ­0 b −b® ­ yn ® . (7.4)
2 2 2
­ ® ­ ®­ ®
« x n+1 ¬ « 0 0 a ¬ « xn ¬
| {z } | {z } |{z}
φ n+1 K φn

We see that a three-dimensional (rather than infinite-dimensional) vector of observ-

ables, inspired by the nonlinearity in the original mapping, transforms the nonlinear
two-dimensional system into a linear one.
A second, even simpler, example (Brunton & Kutz 2019) is the logistic map
xn+1 = βxn (1 − xn ). Based on the success of the previous example, we choose the
observable φn = (xn, xn2 ). When deriving a mapping for this choice of observable, we
see that terms of cubic and quartic order in xn appear. These terms are not part of
the observable vector but, as we subsequently include them, even higher-order terms
will appear – leading to a truly infinite system. In essence, by choosing powers of
the state vector xn , we outrun the span of the observable basis. Consequently, closure
(as accomplished in the first example) cannot be reached. A more systematic manner
of computing Koopman embeddings is needed. Simply considering higher powers
of the observables, a technique related to Carleman linearization (Carleman 1932), is
generally not ensuring a mapping onto a linear dynamical system. See also the chapter
on nonlinear dynamics for additional details.
As the Koopman operator is defined as the linear mapping between two successive
observations, it seems prudent to consider its eigenfunctions, that is,
KΦ = ΦΛ, (7.5)
and express our observables in this Koopman eigenvector basis according to φ = Φξ.
Applying discrete time-stepping then produces a sequence of observables as
φ(qk ) = Kk φ(q0 ) = Kk Φξ = ΦξΛk . (7.6)
The choice of the Koopman eigenvector basis facilitates closure in the observable
basis, as it represents the invariant subspace of K. A truncation of the eigenvector
basis has to be considered with care, as the full closure relies on the infinite nature of
the observable basis. We also note that Φ is dependent on q0 .

h d i 0 0 9 8 088962 4 0 2 bli h d li b C b id i i
 Schmid: The Dynamic Mode Decomposition 137

The above representation holds the key to establishing a computational technique

to compute approximations to the Koopman eigenfunctions directly from a snapshot
sequence of observables.
Using expression (7.6) and the abbreviation φi = φ(qi ), we can establish the
snapshot matrix of observables as

φ0 φ1 · · · φn−1 = Φ I Λ · · · Λn−1 (I ⊗ ξ) = Φ diag(ξ) C,

 

where ⊗ denotes the Kronecker product, I stands for the n × n identity matrix, and the
matrix C represents the (discrete) temporal dynamics.
In what follows, we consider the original data sequence as observables. Juxtaposing
the previous final expression with a general decomposition of a data matrix D,
consisting of a temporal sequence of measurements/observable of our dynamic
process arranged in columns, we seek

time time

= B
Amplitudes
C
Dynamics

D A
Data Modes

from which we can identify the matrix A as containing the Koopman eigenvectors
Φ, the matrix B as containing the amplitudes, and the matrix C as containing the
temporal dynamics. The diagonality of B decouples the decomposition into rank-one
components. In addition, expression (7.6) establishes the matrix C as a Vandermonde
matrix, containing the eigenvalues of K, with diag({λ}) = Λ. We have

©1 λ1 λ12 ··· λ1n−1 ª

­1 λ2 λ22 ··· λ2n−1 ®®
C = ­­ . .. ® , (7.7)
­
.. ..
®
­ .. . . . ®®
­
«1 λn λn2 ··· λnn−1 ¬

h d i 0 0 9 8 088962 4 0 2 bli h d li b C b id i i
 138 Data-Driven Fluid Mechanics Chapter 7

where n denotes the number of snapshots (columns) in the data matrix D. In other
words, given a data matrix D ∈ Cm×n we have to find a decomposition into three
factors, a general matrix A ∈ Cm×n , a diagonal matrix B ∈ Cn×n, and a Vandermonde
matrix C ∈ Cn×n .

Computing Koopman Eigenvectors from Data

The decomposition of a rectangular matrix into three factors is a common procedure
in numerical analysis. Constraints on the three factors have to be imposed to arrive at a
unique decomposition. For example, the constraints of an orthogonal A- and C-matrix
and a diagonal B-matrix yield the familiar singular value decomposition (SVD) which
is related to the proper orthogonal decomposition from Chapter 6. Other constraints
will result in less familiar factorizations.
To advance on our decomposition, we use a relationship between Vandermonde and
companion matrices: Vandermonde matrices diagonalize companion matrices. We can
write
S = C−1 BC, (7.8)

with S as a companion matrix, that is, a matrix consisting of a subdiagonal of ones

and a nonzero final column:
a0
­1 a1
© ª
®
­
S=­ 1 a2
®
®. (7.9)
­ ®
.. ..
.
­ ®
­
­ . ®
®
« 1 an−1 ¬
We can replace the product BC by the product CS and recast our decomposition as
D = ACS. We rename the product AC as D0 and write

D = D0S. (7.10)

We take advantage of the companion structure of S to deduce the matrix D0 and the
last column of matrix D. The subdiagonal of ones links the j + 1th column of D0 to
the jth column of D for j = 1, . . . , n − 1. This part of the companion matrix S is
thus responsible for an index shift for the columns of the two matrices D and D0 . In
other words, the columns in D consist mostly of columns of D0 . The last column of S
(containing the coefficients a j ) states that the nth column of D is a linear combination
of all the columns of D0 . In the end, the columns of D are the backshifted columns of
D with the final column a linear combination of all columns of D0 .

7.3 The Dynamic Mode Decomposition

The transformation of our Koopman decomposition from a Vandermonde-based

into a companion-based factorization now provides a first algorithm to compute

h d i 0 0 9 8 088962 4 0 2 bli h d li b C b id i i
 Schmid: The Dynamic Mode Decomposition 139

approximations of the Koopman eigenvectors and eigenvalues. This decomposition

is known as the DMD (Schmid 2010, Schmid et al. 2011, Kutz et al. 2016).
Starting with a sequence of n + 1 snapshots, sampled (for simplicity) at equispaced
intervals ∆t in time, we break the full sequence into two subsequences, each with n
columns, according to

 D1 ,
 n


D1n+1 → (7.11)
 Dn+1,

 2
j
where we have introduced the notation Di to represent a sequence of measurement
j
from snapshot i to snapshot j, that is, Di = {di , di+1, . . . , d j }, with d denoting the
measurement (observable) vector. Equation (7.10) then suggests to express the last
column of D2n+1 , that is, the observable vector dn+1 as a linear combination of all
columns of D1n . We accomplish this by formulating a least-squares problem, and using
a QR-decomposition of D1n, we arrive at

S = R−1 Q H D2n+1 with QR = D1n, (7.12)

that is, the companion matrix S. The superscript H denotes the Hermitian (conjugate
transpose) operation. We then have the relation

KD1n = D2n+1 ≈ D1n S. (7.13)

This latter expression states that the action of the Koopman operator K on the data
set of observables D1n, pushing each snapshot forward over one time step ∆t, can
be expressed on the basis of all gathered snapshots using the companion matrix S
(Rowley et al. 2009). While K scales with the (high) dimensionality of the observer
vectors, the matrix S scales with number of gathered snapshots. Equation (7.13)
is reminiscent of an Arnoldi iteration, where the action of a large matrix on an
orthonormal set of vectors (from a Krylov subspace sequence) is expressed on the
same basis. As a consequence, any spectral information we wish to gather about
the operator K can be gathered, in an approximate manner, from the operator S. For
example, a subset of eigenvalues of K can be approximated by the eigenvalues of S.
Determining S via a least-squares approximation, as proposed earlier, can quickly
become ill-conditioned. For long data sequences, or data sequences that are charac-
terized by a noticeable amount of noise or uncertainty, the matrix D1n can become
rank-deficient. This is particularly true when data redundancy or near data redundancy
arises. For example, sampling a wake flow past a bluff body will eventually fail to
produce new data; instead, the repeated shedding of coherent structures will cause the
matrix to quickly decline in rank, once a few full shedding cycles have been sampled.
In this case, the matrix R from the QR-decomposition of D1n is no longer invertible.
For this reason, a more robust algorithm must be implemented.
While keeping with the general premise of the above least-squares problem, we
switch to solving it for the rank-deficient case, using a singular-value decomposi-
tion (Schmid 2010). We then have

h d i 0 0 9 8 088962 4 0 2 bli h d li b C b id i i
 140 Data-Driven Fluid Mechanics Chapter 7

1 given a snapshot sequence {d1, d2, . . . , dn+1 } sampled equispaced in time with ∆t

2 D1n ← {d1, d2, . . . , dn }

3 D2n+1 ← {d2, d3, . . . , dn+1 }

4 UΣV H ← D1n

5 S̃ ← U H D2n+1 VΣ+


6 [X, Λ] ← eig(S̃)
7 µ j ← log(Λ j j )/∆t

8 Φ j ← UX(:, j)

Figure 7.2 SVD-based dynamic mode decomposition algorithm. The matrix Φ represents the
matrix A in our general decomposition D = ABC. The eigenvalues Λ can be recast into the
Vandermonde matrix C of our general decomposition.

1/σj σj ≥ 
for
 
S̃ = U H
D2n+1 VΣ+
with UΣV = H
D1n, Σ+
= diag ,
0 σj < 
for
(7.14)
which applies a Moore–Penrose pseudo-inverse to the least-squares problem. We have
introduced a threshold  in the pseudo-inverse to signify a cutoff value for considering
the processed data matrix as rank-deficient. In expression (7.14) for S̃ we have also
projected the dynamics on the singular vectors contained in U, which are equivalent to
the POD-modes of D1n .
We can then state the simple algorithm for computing the DMD from a sequence
of data snapshots, and with it the approximate spectral properties of the Koopman
operator K.
This algorithm accommodates data sequences with a potential degree of redundancy
or near-redundancy, leading to a rank deficiency when processing the data matrix in
step 4 of the algorithm (see Figure 7.2). When forming the pseudo-inverse in step 5, a
threshold value  has to be specified and principal vectors of the SVD associated with
singular values below this threshold can be discarded. The logarithmic mapping in
step 7 is commonly applied in hydrodynamic applications to transform from a discrete
to a continuous time variable. The above algorithm is a rather robust technique for
extracting dynamic modes (and approximate Koopman eigenvectors) directly from
a sequence of observable data. It should be mentioned that an additional SVD of
D2n+1 will produce exact eigenvectors of the high-dimensionial operator for the case
of D1n and D2n+1 not spanning the same vector space (Tu et al. 2014). Extensions and
generalizations of the above core algorithm (regarding robustness, accuracy, efficiency
and applicability) will be discussed in Section 7.4.

h d i 0 0 9 8 088962 4 0 2 bli h d li b C b id i i
 Schmid: The Dynamic Mode Decomposition 141

7.4 Sparsity Promotion for Finding Amplitudes

Once the dynamic structure and their time dependence (contained in the Koopman
eigenvalues λ j ) have been identified, we are still left with the computation of the
amplitudes, that is, the diagonal matrix B of our original decomposition. These
amplitudes will provide information about the importance and dominance of the
identified structures, as they are present in the processed data sequence. Ideally,
structures and their dynamics that are key components in the overall data sequence
should be identified by a larger amplitude, while smaller amplitudes should be
attached to less dominant processes or flow features. For data from high-performance
simulations, this desirable relation between amplitude and relevance can often be
realized; for experimental data and data with a significant amount of noise, it cannot
be guaranteed. In particular, when the processed data sequence is marred by outliers
or other inaccuracies that are fairly localized in time, we may encounter intermittent
structures that are characterized by a large decay rate. This large decay rate accounts
for the fact that the associated flow feature disappears quickly over the time horizon of
the sampled data. If the intermittent structure is of sufficient size, a simple amplitude
algorithm (matching the identified dynamic modes to the original data sequence)
would assign a sizeable amplitude to it, thus giving it importance among the extracted
modes and processes. For a robust and physically meaningful assignment of ampli-
tudes to identified dynamic processes, a more subtle balance between the reproduction
of the original data sequence and the number of participating modes must be struck.
Mathematically and computationally, the recovery of the amplitudes is a nontrivial
step. It is best formulated as a mixed-norm optimization problem. The first component
of the optimization is responsible for enforcing data fidelity of the recovered process.
In other words, we choose the amplitudes of the modes and their dynamics to
best reproduce the original data sequence. This is accomplished by minimizing the
Frobenius norm of the modal expansion and the data sequence (Jovanović et al. 2014).
We denote the vector of amplitudes by b, that is, the diagonal elements of the matrix
B above. We have
kD1n − Φdiag(b)CkF → min, (7.15)

with Φ as the matrix containing the computed dynamic modes Φ j and C as the
Vandermonde matrix containing powers of the identified eigenvalues λ j = Λ j j . Using
a SVD of D1n, the definition of the matrix Φ, and some algebra, the optimization
problem above can be reformulated into the quadratic form

J (b) = b H Pb − q H b − b H q + s → min, (7.16)

with

P = (X H X)  (CC H )∗, (7.17a)

q ∗
= diag(CVΣX), (7.17b)
s = trace(Σ2 ), (7.17c)

h d i 0 0 9 8 088962 4 0 2 bli h d li b C b id i i
 142 Data-Driven Fluid Mechanics Chapter 7

where ∗ denotes the complex conjugate operation and  stands for the elementwise
(Hadamard) product. The amplitude vector b∗ that optimizes the match between the
dynamic mode expansion and the original data sequence is simply given as the solution
to this quadratic optimization problem, namely,

b∗ = P−1 q. (7.18)

This amplitude distribution only minimizes the reconstruction error of the DMD,
taking all available modes – whether physically relevant or linked to outliers in the data
– into account. For a more robust and physically meaningful reconstruction, we have
to restrict the number of modes being considered. We will do this by juxtaposing the
sparsity of the amplitude vector and the achieved data reconstruction error (Jovanović
et al. 2014). These two objectives are in conflict with each other: a very sparse
amplitude vector consisting of only a few nonzero amplitudes will not produce a small
data reconstruction error, while a minimal reconstruction error will not be achieved by
only a few amplitudes. A compromise has to be found that produces an acceptable
reconstruction error with as sparse an amplitude vector as can be managed.
The sparsity of a vector is commonly measured by its `0 -norm or cardinality. The
`0 -norm of a vector simply counts the number of its nonzero components, and is
strictly speaking not a norm (but a quasi-norm). Optimizations based on `0 -norms
are combinatorial in nature and difficult to implement. To this end, we substitute the
`1 -norm of the amplitude vector (i.e., the sum of the absolute values of its components)
as a proxy for its sparsity, and augment the cost functional J (b) to read

J (b) + γkbk1 (7.19)

with a user-specified parameter γ that quantifies the trade-off between the reconstruc-
tion error J and the sparsity of the amplitude vector kbk1 . This regularization step
is reminiscent of Lasso regression, where a least-squares fit is combined with an
L1 -penalty (Tibshirani 1996). Minimizing the above cost functional yields a lower-
dimensional representation of the full data matrix of the form.
In the augmented cost functional, γ is a positive parameter that quantifies our focus
on the sparsity of the amplitude vector b. Larger values of γ produce sparser solutions,
at the expense of a larger data reconstruction error (see Figure 7.3).
As the value of γ increases, the amplitude vector b becomes progressively
sparser. There is no straightforward link between the number of nonzero amplitudes
(cardinality of b) and the required value of γ to reach this outcome. In practice, one has
to experiment with different values of γ to determine the value for a desired number
of nonzero amplitudes. We notice, however, that the optimization procedure involved
only matrices that scale with the number of snapshots; multiple optimizations with
different values of γ should therefore be computationally feasible.
After we have reached the desired number of amplitudes, while keeping the
reconstruction error minimal, we lock the sparsity structure of the amplitude vector
and solve, once again, an optimization problem, where we optimize the reconstruction

h d i 0 0 9 8 088962 4 0 2 bli h d li b C b id i i
 Schmid: The Dynamic Mode Decomposition 143

Time Time

≈ B
Amplitudes
C
Dynamics

D Φ
Data Modes

J (b, γ)

Figure 7.3 Cost functional versus the user-specified parameter γ. Low values of γ result in
lower reconstruction error, but fuller amplitude vectors; larger values of γ lead to sparser
amplitude vectors at the expense of larger reconstruction errors.

error subject to enforcing the identified sparsity structure explicitly. This results in the
convex optimization problem (Jovanović et al. 2014)
J (b) → min, (7.20a)
subject to E b = 0,
H
(7.20b)
where the columns of E consist of unit vectors that identify the zero entries of the final
amplitude vector b, found from the previous mixed-norm optimization.
The mixed-norm optimization problem can be solved by common numerical
techniques, such as ADMM (alternating direction method of multipliers) (Boyd

h d i 0 0 9 8 088962 4 0 2 bli h d li b C b id i i
 144 Data-Driven Fluid Mechanics Chapter 7

et al. 2011), split Bregman iteration (Goldstein & Osher 2009), or IRLS (iteratively
reweighted least-squares) (Daubechies et al. 2010).
In the supplemental material, a three-step MATLAB R
code will analyze a snapshot
sequence of PIV data (flow through a cylinder bundle) and extract relevant flow
processes via a sparsity-promoting DMD.

7.5 Applications

A few selected applications shall demonstrate the DMD on experimental and numeri-
cal data sets. The examples are meant to illustrate the algorithm and showcase different
manners of utilizing the flexibility of the methodology in analyzing fluid behavior.

Example 1: Tomographic TR-PIV of a water jet

As a first example, we consider experimental data from tomographic time-resolved
PIV measurements of a water jet (Schmid et al. 2012). The jet emerges from a round
nozzle of diameter 10 mm into an octagonal water tank that allows optimal access
for tomographic imaging. The exit velocity of the jet is U = 0.5 m/s, resulting in
a Reynolds number of Re = 5000. Neutrally buoyant particles are used to produce
light-scatter images, which are recorded and processed by the tomographic imaging
system. Snapshots are acquired at a rate of 1 kHz in a three-dimensional measurement
domain of 50 mm × 50 mm × 32 mm. A representative snapshot from the experiment
is shown in Figure 7.4, visualized by velocity vectors in the axial center plane.
The data matrix D is formed by a sequence of snapshots (sampled at a frequency
of 1 kHz), each consisting of 107 × 62 × 62 three-dimensional velocity vectors. We
process n = 40 snapshots. The DMD, with sparsity promotion for the amplitude
recovery, is then applied to the data matrix.
The DMD-spectrum Λ j j is shown in Figure 7.4, superimposed on the unit disk.
Four eigenvalues are associated with modal structures of significant amplitude; they
represent two distinct structure with Strouhal number St = 0.374 and St = 0.691.
The amplitude distribution versus the frequencies confirms this finding: two marked
peaks at the aforementioned Strouhal numbers are clearly visible above a background
amplitude distribution. It appears that the jet dynamics captured by the tomo-PIV
data sequence is well described by two principal structures. These two structures are
displayed in Figure 7.5, in a two-dimensional slide through the center plane, and as
iso-contours of the Q-criterion. We identify the higher Strouhal number (St = 0.691)
as the primary instability of the jet, as its spatial support is closer to the nozzle exit.
At the lower Strouhal number (St = 0.374), an instability is identified that results in
the merging of nearly axisymmetric vortex rings, halving the characteristic axial scale
and halving the temporal frequency (Strouhal number) in the process. In the three-
dimensional visualizations, the onset of an azimuthal breakdown of the vortex rings
(via a secondary instability) is clearly visible.

h d i 0 0 9 8 088962 4 0 2 bli h d li b C b id i i
 Schmid: The Dynamic Mode Decomposition 145

(a)

(b)

(c)

Figure 7.4 Tomographic TR-PIV of a water jet. (a) Representative snapshot (visualized in a
two-dimensional slice through the center plane). (b) The DMD spectrum (superimposed on the
unit disk). (c) Amplitude distribution versus identified frequencies. (a) and (b) from Schmid,
Violato & Scarano (2012), reprinted by permission from Springer Nature.

Example 2: Composite analysis of an acoustically active jet

The second example is based on data from numerical simulations of the axisymmetric
compressible Navier–Stokes equations. We consider a jet exiting from a round nozzle
in quiescent air. With a Mach number of Ma = 0.9, the shear layers forming between
the exiting jet and the quiescent freestream cause acoustic radiation. We are interested
in analyzing the jet flow and identify structures that are responsible for the sound
radiated into the freestream.
To this end, we generate and process observables that account for each of the
dominant physical processes: hydrodynamic instabilities and acoustic sound radiation.
The hydrodynamics is represented by the vorticity ∇ × u of the flow field, acting
as a substitute for instability structures. The acoustic component is captured by the
dilatation, that is, the divergence ∇ · u of the velocity field. Other combinations or
markers for the respective physical processes are conceivable. Our two choices are
then combined into an observable vector according to
 
 
∇ × u
 
 
di =   , (7.21)
 
 
 
∇ ·u
 
 
 i
and the data matrix is formed by equispaced snapshots of synchronized vorticity-
dilatation fields. The underlying premise of this composite data processing is that the

h d i 0 0 9 8 088962 4 0 2 bli h d li b C b id i i
 146 Data-Driven Fluid Mechanics Chapter 7

(a) (b)

DM2 DM3

DM2

DM3

Figure 7.5 Tomographic TR-PIV of a water jet. Visualization of the two dominant dynamic
modes, (a) visualized by velocity vectors in a two-dimensional slide through the center plane,
and (b) visualized by iso-contours of the Q-criterion.

Koopman operator propagating this observed flow state will contain information about
the interplay between the two processes encapsulated in the chosen data. Eigenvectors
of the Koopman operator that have significant activity in both components may point
toward structures that are “synced” within our data stream and hence may have a
physical link, as they emerge with an identical temporal dynamics.
This processing of composite data is a rather compelling tool to analyze data from
multi-physics processes. While it cannot establish a cause–effect relationship between
the processed fields, it can identify structures (and their frequencies) that are “in
resonance” or arise conjointly. For example, fluid-structure problems, shock-boundary
layer interaction, or the link between coherent structures and their wall-shear stress
footprint can be treated using this composite multivariable DMD technique.
In our case, a DMD analysis of the gathered data sequence from axisymmetric
simulations of the compressible jet reveals a spectrum of eigenvalues (not shown)
with two characteristic, but broad, peaks in frequency. The first, lower-frequency
peaks are dominated by modes with mostly hydrodynamic (vorticity) components,
which concentrate in the center of the jet or on the developing instabilities of the
downstream outer shear layer (see Figure 7.6); the same structures also underscore the
importance of the collapse of the potential core, which appears on the centerline at
an axial distance of x ≈ 12 in the simulations. No significant or very little acoustic
activity in the freestream is visible.
Sampling from the higher-frequency peak, a different picture emerges (see
Figure 7.7). While the hydrodynamic component still concentrates on the center of

h d i 0 0 9 8 088962 4 0 2 bli h d li b C b id i i
 ∇×u

Schmid: The Dynamic Mode Decomposition 147

∇·u

Figure 7.6 Composite analysis of a compressible axisymmetric jet at Ma = 0.9. Two dynamic
modes (left and right) from the lower-frequency peak, with the top corresponding to the
hydrodynamic vorticity component and the bottom visualizing the acoustic dilatation
component.

the jet, the sound component of the dynamic modes is clearly visible and captures the
acoustic activity in the freestream. While the first mode (on the left) is dominated by
a nearly omnidirectional radiation of sound, the second mode (on the right) displays a
preferred directionality toward the upstream region. Especially these latter structures
are often responsible for feedback loops that magnify and sustain acoustic activity
and cause concentrated peaks in pressure sound spectrograms.
This example demonstrates the versatility of a Koopman perspective on dynamical
systems, as well as the flexibility of the DMD to explore physical processes from
a data-driven viewpoint. A traditional Poincaré approach based on a state-space
(rather than observable-space) formulation would be difficult to carry out with the
same ease.

Example 3: Streamline-based analysis of a jet in cross-flow

The final example further extends the capabilities of Koopman/DMD analysis to treat
configurations where the appropriate evolution direction is not the time coordinate. In
our derivation of the DMD, we have assumed the time coordinate as our evolution
direction. Consequently, we have stacked our data matrix such that each column
represents a different snapshot in time. The Vandermonde matrix C and the approx-
imate eigenvalues of the Koopman operator thus represent a temporal dynamics; the
associated modes, captured in Φ, describe spatial structures. While this division into

h d i 0 0 9 8 088962 4 0 2 bli h d li b C b id i i
 ∇×u

148 Data-Driven Fluid Mechanics Chapter 7

∇·u

Figure 7.7 Composite analysis of a compressible axisymmetric jet at Ma = 0.9. Two dynamic
modes (left and right) from the higher-frequency peak, with the top corresponding to the
hydrodynamic vorticity component and the bottom visualizing the acoustic dilatation
component. Significant acoustic activity is observed in both cases.

temporal dynamics and spatial structure may be appropriate for many applications,
there are just as many configurations where an alternative dynamics-structure division
seems more fitting. An example is given by a jet in cross-flow.
A jet in cross-flow (or transverse jet) is generated when a flow exiting an orifice on
a wall is subjected to a second flow parallel to the wall. Situations like this arise, for
example, in smoke stacks or every time a fluid is injected into a cross-stream, such as
in fuel injectors. The corresponding flow is rather complex and develops a great many
instabilities that interact in a nontrivial manner. Among the main instabilities is the
roll-up of the vortex sheet, which initially axisymmetric at the orifice quickly deforms
into two counterrotating vortex pairs that follow a bent trajectory. Along this bent,
rolled up vortex sheet instabilities develop quickly: initiated a short distance from the
orifice, vortex ring-like structures amplify before they break down into smaller and
less coherent fluid elements.
The description of the principal instability for a jet in cross-flow clearly shows
that the evolution of the instability is more aptly expressed as a spatial evolution
along the bent vortex sheet of the base flow. This realization immediately sways the
analysis of this physical process toward a data-driven (Koopman) approach, rather

h d i 0 0 9 8 088962 4 0 2 bli h d li b C b id i i
 Schmid: The Dynamic Mode Decomposition 149

Figure 7.8 Streamline-based analysis of a jet in cross-flow. A base streamline, emanating

from the jet nozzle, serves as the spatial evolution direction. Flow fields are sampled in n plane
orthogonal to the streamline and equispaced in arclength along the streamline. Figure
reproduced with permission from Henningson (2010).

than a traditional state-space description that would require the formulation of the
governing equations for the spatial evolution of disturbances along curved paths.
In our data-driven approach, we simply have to sample the flow fields in planes
orthogonal to a chosen evolution direction and stack the data matrix accordingly.
To this end, we determine a streamline emanating from the center of the orifice and
following the base flow. This streamline will serve as our spatial evolution direction.
We then partition the streamline into n = 110 equispaced (in arclength) intervals
between the location of instability onset and the location of ultimate instability
breakdown. For each point along the streamline, we define a plane normal to its
tangent, and interpolate the three-dimensional velocity field into this plane. These
planes (when reshaped into vectors) will form the columns of our data matrix D.
Figure 7.8 illustrates the aforementioned procedure.
Once we have established the evolution direction of our observables and formed the
data matrix, a standard DMD/Koopman analysis can be performed. It is worth noting
that the “spatial” direction, that is, the row direction of the data matrix, contains the
coordinates in the cross-planes (normal to the base-flow streamline tangent) as well as
the time coordinate. In order to reduce the size of the problem, we Fourier transform
in time and focus on the Strouhal number (nondimensional frequency) associated with
the Kelvin–Helmholtz instability of the vortex sheet.
Evaluating the dynamic modes from the analysis of our data matrix, we capture the
spatial breakdown of the vortex sheet. The identified instabilities are localized on the
base-flow counterrotating vortex sheet and describe vortical structures of increasing
complexity and scale; see Figure 7.9.

h d i 0 0 9 8 088962 4 0 2 bli h d li b C b id i i
 150 Data-Driven Fluid Mechanics Chapter 7

Figure 7.9 Streamline-based analysis of a jet in cross-flow. Dominant dynamic modes

displaying the structures causing the spatial breakdown of the base vortex sheet along the base
streamline. The approximate location of the base vortex sheet is indicated in red. Reproduced
with permission from Henningson (2010).

The dynamic modes consist of vortical structures that align with the counterrotating
vortex sheet of the base flow. Higher-order modes display more vortical elements that
progressively concentrate on the flanks of the base-flow vortex sheet. This is consistent
with the findings from direct numerical simulations as well as physical experiments.
The amplitude distribution of the modes (not shown) shows a broad peak, which
suggests that a superposition of a few dominant modes may best capture the pertinent
dynamics of the spatial instability and may provide a reduced-order description of the
Kelvin–Helmholtz-type breakdown of the base-flow’s counterrotating vortex sheet.
This final example demonstrates the capability of Koopman/DMD analysis to iden-
tify coherent fluid motion directly from data snapshots in situations where the relevant
evolution direction is not aligned with a coordinate direction, but is instead given by
a curved base-flow streamline. With a projection and rearrangement of the data fields,
the dynamic characteristics of complex instabilities can be readily computed.

h d i 0 0 9 8 088962 4 0 2 bli h d li b C b id i i
 Schmid: The Dynamic Mode Decomposition 151

7.6 Extensions and Generalizations

Since its introduction, the DMD and the underlying Koopman analysis have gained in
popularity as a tool for quantitative flow analysis. Both fields have spawned a great
deal of improvements, extensions, and generalizations – both from an algorithmic and
conceptual point of view. Only a few of them shall be mentioned here.
The choice of observables is a key issue in Koopman analysis. In the above
exposition, we use the raw snapshots, assuming that their evolution contains the
necessary information to embed the dynamics in a linear fashion. However, extended
DMD (Williams et al. 2015) aims at explicitly including observables that are higher
powers of the original snapshots. This follows the original idea of supplementing
the observables by selected nonlinearities. Alas, extended DMD quickly becomes
computationally expensive; a reduction in dimensionality using a kernel trick is
possible (Williams et al. 2015), but only provides temporary relief.
It has quickly been noted that, while equispaced snapshots are advantageous in
the motivation of the DMD algorithm, they are not required for a more general
formulation. In fact, only the snapshots in equal column position in D1n and D2n+1 have
to be separated by a constant ∆t. Within either snapshot sequence, non-equispaced
data can be accommodated.
A SVD is a core element of the DMD algorithm, which allows for a great deal
of improvements from an algorithmic standpoint. An incremental formulation of the
SVD allows the design of an efficient DMD algorithm for streaming data (Hemati
et al. 2017), where new data snapshots get incorporated into an already existing
decomposition by a perturbative adjustment of the modes and spectrum. This type
of algorithm permits the analysis of experimental data in real time. In addition, recent
developments in randomized algorithms for the SVD can also be incorporated into the
overall DMD algorithm and lead to higher efficiencies, while sacrificing little in terms
of accuracy.
Using delay-embedding techniques (related to the Ruelle-Takens methodology),
the predictive horizon of a reduced description of a dynamical system via DMD can
be substantially enhanced. In essence, the data sequence is transformed into a phase-
space description and decomposed there. The data matrix is block-Hankelized and
processed by the standard DMD algorithm (Arbabi & Mezić 2017, Brunton et al.
2017).
Finally, the recent popularity of machine learning and artificial neural network
techniques has also influenced data decompositions and model reduction methods.
By proposing a neural net in the form of an auto-encoder, a nonlinear Koopman
embedding can be learned directly from the data (Takeishi et al. 2017, Lusch
et al. 2018, Mardt et al. 2018, Otto & Rowley 2019, Yeung et al. 2017), by imposing
linearity of the mapping over one time step between the final step of the encoder
portion and the first layer of the decoder part. More advances in this direction are
expected over the coming years.

h d i 0 0 9 8 088962 4 0 2 bli h d li b C b id i i
 152 Data-Driven Fluid Mechanics Chapter 7

7.7 Summary and Conclusions

DMD is a simple, yet effective data-driven tool to extract quantitative information

about fluid flows from sequences of measurements. In its simplest form, it is straight-
forward to implement and use. Nonetheless, it is based on a sound, sophisticated
and rich mathematical formalism – Koopman analysis – that provides guidance and
support for further developments, applications, and interpretations. Many extensions
and improvements of the original algorithm are available, and a wide range of
applications (to numerical and experimental data) can be found in a growing body
of literature.

h d i 0 0 9 8 088962 4 0 2 bli h d li b C b id i i
 8 Generalized and Multiscale
Modal Analysis
M. A. Mendez

This chapter describes modal decompositions in the framework of matrix factor-

izations. We highlight the differences between classic space-time decompositions
and 2D discrete transforms and discuss the general architecture underpinning any
decomposition. This setting is then used to derive simple algorithms that complete
any linear decomposition from its spatial or temporal structures (bases). Discrete
Fourier transform (DFT), proper orthogonal decomposition (POD), dynamic mode
decomposition (DMD), and eigenfunction expansions (EF) are formulated in this
framework and compared in a simple exercise. Finally, this generalization is used
to analyze the impact of spectral constraints on the classical POD, and to derive the
multiscale proper orthogonal decomposition (mPOD). This decomposition combines
multi-resolution analysis (MRA) and POD. This chapter contains four exercises and
two tutorial test cases. The P YTHON scripts associated to these are provided in the
book’s website.1

8.1 The Main Theme

Modal analysis aims at decomposing a data set as a linear combination of elementary

contributions called modes. This provides the foundations to many areas of applied
mathematics, including pattern recognition and machine learning, data compression,
filtering, and model order reduction. A mode is the representation of the data along an
element of a basis, and the decomposition is the projection of the data set onto the basis
(see Chapter 4). A basis maps the data onto a different space from the one in which
it is sampled, with the hope that this new space is capable of better highlighting the
features of interest for the task at hand. This is the central theme of this chapter: modal
decompositions are ways of representing the same information in different bases.
There is no “perfect” basis. Although some decompositions are more flexible than
others, they all have their merit for a specific task. In other words, every basis can
highlight or “isolate” a specific pattern in the data. Depending on how these features
are used, one enters into various areas of applied mathematics – all relevant in fluid
mechanics – as discussed in different chapters of this book.
In data compression or model order reduction, for example, one aims at represent-
ing the data with the least number of modes, to distill its “most important” features.
1 www.datadrivenfluidmechanics.com/download/book/chapter8.zip

h d i 0 0 9 8 088962 4 0 3 bli h d li b C b id i i
 154 Data-Driven Fluid Mechanics Chapter 8

We thus look for the smallest basis, that is, the smallest subspace that can still handle
all the information we need. Yet, the definition of “most important” strongly depends
on the application and the experience of the user. In some settings, it is essential to
identify (and retain) information contained within a specific range of frequencies. In
others, it might be necessary to focus on the information that is localized within a
particular location in space or time; in still others, it might be essential to extract some
“coherency” or “energy” or “variance” contribution. More generally, one might be
interested in a combination of the previous, and the tools described in this chapter
give the reader full control over the full spectrum of options.

8.1.1 The Scope of Modal Decompositions

Data-driven modal decompositions were historically driven in fluid mechanics by the
quest for identifying (and objectively define) coherence structures in turbulent flows
(see Chapter 2). In parallel to this, another critical driver has always been reduced-
order modeling, that is the quest for identifying the “right” basis onto which to project
partial differential equations (PDEs). If such a basis consists of a few carefully chosen
modes, the projected system preserves the relevant dynamics and can be simulated
at a fraction of the computational cost. This opens the path to fast analysis and
prediction, optimization, system identification, control, and more modeling: reducing
the dimensionality of a system, we facilitate regression problems such as, for example,
the derivation of new models for turbulence.
In a book on data science and machine learning, this chapter could be titled linear
dimensionality reduction; this is a fundamental tool to simplify classification or regres-
sion problems. A complementary chapter should include nonlinear decompositions
based on kernels or artificial neural networks (ANN), or cluster-based methods. These
methods were initially developed in computer vision, data science, and statistics, and
are now revolutionizing the toolbox of the fluid dynamicists, as discussed in Chapters
1, 3, and 14.
This chapter is about linear decompositions in which a linear combination of
modes describes the data. Each mode might feature a nonlinear dynamic,2 but
their combination is linear, with coefficients determined using inner products (see
Chapter 4). In this chapter, we assume that the data are sampled in space and
time; hence, every mode has a spatial structure and a temporal structure. These
structures represent, respectively, a basis for the discrete space and the discrete time.
Furthermore, the decompositions presented consider data sets composed of a single
signal. Extended methods that construct hybrid bases combining different signals
(e.g., velocities and temperatures) are described in Chapter 9.

8.1.2 How Many Decompositions?

Modal decompositions can be classified into data-independent and data-driven. In the
first category, the basis of the decomposition is constructed regardless of the data set
2 In the sense that its evolution might not be described by a linear equation.

h d i 0 0 9 8 088962 4 0 3 bli h d li b C b id i i
 Mendez: Generalized and Multiscale Modal Analysis 155

at hand and is defined by the size of the data and the basis construction criteria. This
is the case of the DFT (see Chapter 4) or the discrete wavelet transforms (DWT, see
Chapter 5). In the first, structures are harmonics with a frequency that is an integer
multiple of a fundamental one; in the second, structures are obtained by scaling and
shifting a template basis (mother and father wavelets).
In data-driven decompositions, the basis is tailored to the data. The most classic
examples are the POD (see Chapter 6) and the DMD (see Chapter 7). Both POD and
DMD have many variants from which we can identify two categories of data-driven
decomposition: those arising from the POD are “energy-based;” those arising from
the DMD are “frequency-based.”
The POD basis is obtained from the eigenvalue decomposition of the temporal or
the spatial correlation matrices. This is dictated by a constrained optimization problem
that maximizes the energy (i.e., variance) along its basis, constrained to be orthogonal,
such that the error produced by an approximation of rank r̃ < R is the least possible.
Variants of the POD can be constructed from different choices of the inner product
or in the use of different averaging procedures in the computation of the correlations.
Examples of the first variants are proposed by Maurel et al. (2001), Rowley et al.
(2004), Lumley and Poje (1997), where multiple quantities are involved in the inner
product. Examples of the second variants are proposed by Citriniti and George (2000)
and Towne et al. (2018), where the correlation matrix is computed in the frequency
domain using time averaging over short windows, following the popular Welch’s
periodogram method (Welch 1967).
Within the frequency-based decompositions, the DMD basis is constructed
assuming that a linear dynamical system can well approximate the data. The DMD
is thus essentially a system identification procedure that aims at extracting the
eigenvalues of the linear system that best fit the data. Variants of the DMD propose
different methods to compute these eigenvalues. Examples are the sparsity promoting
DMD (Jovanović et al. 2014), the optimized DMD (Chen et al. 2012), or the
randomized DMD (Erichson et al. 2019), while higher-order formulations have been
proposed by Le Clainche and Vega (2017). Although the DMD represents the most
popular formalism in fluid mechanics for such linear system identification process,
analogous formulations (with slightly different algorithms) were introduced in the
late 1980s in climatology under the names of principal oscillation patterns (POP,
see Hasselmann 1988, von Storch & Xu 1990) or linear inverse modeling (LIM, see
Penland & Magorian 1993, Penland 1996).
Both “energy-based” and “frequency-based” methods have limits, illustrated in
some of the proposed exercises of this chapter. These limits motivate the need for
hybrid decompositions that mix the constraints of energy optimality and spectral
purity. Examples of such methods for stationary data sets are the spectral POD
proposed by Sieber et al. (2016), the multiresolution DMD (Kutz et al. 2016), the
recursive DMD (Noack et al. 2016), or the Cronos–Koopman analysis (Cammilleri
et al. 2013). A hybrid method that does not hinge on the stationary assumption is the
mPOD (see Mendez et al. 2018, Mendez et al. 2019).

h d i 0 0 9 8 088962 4 0 3 bli h d li b C b id i i
 156 Data-Driven Fluid Mechanics Chapter 8

All the decompositions mentioned thus far have a common underlying architecture.
This chapter presents this architecture and the formulation of the mPOD. This
decomposition offers the most general formalism, unifying the energy-based and the
frequency-based approaches.

8.2 The General Architecture

All the modal decomposition introduced in Section 8.1 can be written as a special kind
of matrix factorization. This view allows for defining a general algorithm for modal
decomposition that is presented in Section 8.3. First, Section 8.2.1 briefly reviews
the notation followed throughout the chapter while Sections 8.2.2 and 8.2.3 put this
factorization in a more general context of 2D transforms. Section 8.3 briefly discusses
the link between discrete and continuous domain, which is essential to render all
decompositions statistically convergent to grid-independent results.

8.2.1 A Note on Notation and Style

The matrix to be factorized in this chapter is denoted as D(xi, tk ) = D[i, k] and
is assumed to be a collection of samples of a real quantity (e.g., gray scale levels
in a set of images, pressure fields in a CFD simulation or deformation fields in
solid mechanics) along a spatial discretization xi and a time discretization tk . The
notation introduced in Chapter 4 is maintained for continuous and discrete signals.
The boldface notation in the spatial discretization indicates that this can be high-
dimensional, for example, xi = (xi , y j ) in a 2D domain and the boldface index i denotes
a linear matrix index.
The matrix linear index is important when we transform a spatial realization which
has the form of a matrix (e.g., a scalar pressure field p[i, j] with i ∈ [0, nx − 1] and
j ∈ [0, ny − 1]) into a vector (i.e., p[i] with i ∈ [0, ns − 1], being ns = nx ny ). This index
accesses entries of a matrix in a different way depending of whether the flattening
is performed column-wise or row-wise. For example, consider the case of a matrix
A ∈ C3×3 . The column-wise and the row-wise matrix indices are3
0 3 6 0 1 2
column-wise i : A = 1 4 7 , row-wise i : A = 3 4 5 .
 
2 5 8 6 7 8
 
In what follows, we consider a column-wise flattening.
All the material presented in this chapter assumes a 2D space domain, the
generalization to higher or lower dimensions being trivial. Moreover, we assume
that space and time domains are sampled on uniform meshes; the generalization
to nonuniform mesh requires extra care in the normalization process, as discussed

3 Recall that here we use a Python-like indexing. Hence the first entry is 0 and not 1.

h d i 0 0 9 8 088962 4 0 3 bli h d li b C b id i i
 Mendez: Generalized and Multiscale Modal Analysis 157

in Section 8.2.5. The space domain is sampled over a grid (xi , yi ) ∈ Rn x ×ny , with
xi = i∆x, i ∈ [0, nx − 1] and y j = j∆x, y ∈ [0, ny − 1].
In a vector quantity, for example, a velocity field U[u(xi ), v(xi )], we consider that
the reshaping stacks all the components one below the other producing a state vector
of size ns = nC nx ny , where nC = 2 is the number of velocity components. Therefore,
the snapshot of the data at a time tk is a vector dk [i] ∈ Rns ×1 , while the temporal
evolution of the data at a location i is a vector di [k] ∈ Rnt ×1 . We then have D[i, k =
c] = dc [k] ∈ Rns ×1 and D[i = c, k] = dc [l] ∈ R1×nt .

8.2.2 Projections in Space or Time

Every linear operation on a discrete 1D signal, represented as a column vector, can be
carried out via matrix multiplication. This is true for convolutions, changes of bases
(i.e., linear transforms), and filtering. Understanding the matrix formalism of these
operations brings at least three benefits. First, the notation is simplified by replacing
summations on sequences with matrix-vector multiplications. Second, the efficiency
and compactness of computer codes for signal processing are largely augmented,4 by
avoiding the nested “for” loops implied in the summation notation. Third, the linear
algebra representation of these operations offers a valuable geometrical interpretation.
Consider the problem of decomposing a column vector u ∈ Rnu ×1 into a linear
combination of basis vectors b1, b2, . . . , bnb of equal dimensions. This means


nu
Õ
u[k] = u = cr br ⇐⇒ u = B uB , (8.1)
r=1

where B = [b1, b2, . . . bn B ] ∈ Cnu ×nb is the basis matrix having all the elements
of the basis along its columns, and uB = [c1, c1, . . . , cnb ]T is the set of coefficients
in the linear combination, that is, the representation of the vector in the new basis.
Computing the transform of a vector with respect to a basis B means solving a linear
system of algebraic equations. Such a system can have no solution, one solution, or
infinite solutions depending on nb and nu .
If nb < nu , as it is the case in model-order reduction, the system is overdetermined
and there is no solution.5 In this case, we look for the approximated solution ũB that
is obtained by projecting u onto the column space of B. This is provided by the well-
known least-squares approximation, which gives6
u = B uB =⇒ ũB = (B† B)−1 B† u =⇒ ũ = B(B† B)−1 B† u = PB u . (8.2)
The least-squares solution minimizes ||u − B uB ||2 = ||e||2 ; the minimization
imposes that the error vector e is orthogonal to the column space of B. In the machine
learning terminology, the matrix PB = B(B† B)−1 B† is an autoencoder that maps a

4 Especially in interpreted languages such as MATLAB or Python.

5 Unless the u lays in the range of the column space of B in which case a unique solution exists. This is
rarely the case in practice.
6 Multiply the system by B† and then invert the resulting B† B.

h d i 0 0 9 8 088962 4 0 3 bli h d li b C b id i i
 158 Data-Driven Fluid Mechanics Chapter 8

signal in Rnu ×1 to Rnb ×1 (this is an encoding) and then back to Rnu ×1 (this is the
decoding).
Because the underlying linear system has no solution, the linear encoding does
not generally admit an inverse if nb < nu : it is not possible to retrieve u from ũB –
that is, the autoencoding loses information. A special case occurs if nb = nu . Under
the assumption that the columns are linearly independent– that is, B−1 exists – there is
only one solution. It is easy to show that7 this yields PB = I: in this case ũ = u, and the
tilde is removed because the autoencoding is lossless. This is fundamental in filtering
applications for which the basis matrix is usually square: a signal is first projected
onto a certain basis (e.g., Fourier or wavelets), manipulated, and then projected back.
The last possibility, that is nb > nu , results in an underdetermined system. In this
case, there are infinite solutions. Among these, it is common practice to consider the
one that leads to the least energy in the projected domain, that is, such that min(||uB ||).
This approach, which also yields a reversible projection, is known as the least norm
solution and reads
u = B uB ⇐⇒ uB = B† (BB† )−1 u . (8.3)
It is now interesting to apply these notions to decompositions (projections) in space
and time domains. Considering first the projection in the space domain, let the signal
in (8.2) be a column of the data set matrix, that is, u := dk [i] ∈ Rns ×1 . Let Φ =
[φ1, φ2, . . . , φ nφ ] ∈ Cns ×nφ denote the spatial basis. Because the matrix multiplication
acts independently on the columns of D, (8.1) is
dk = Φ dφ ⇒ d̃kφ = (Φ† Φ)−1 Φ† dk ; D = ΦDφ ⇒ D̃φ = (Φ† Φ)−1 Φ† D. (8.4)
Here the transformed vector is d̃kφ while the matrix D̃φ collects all the trans-
formed vectors, that is, the coefficients of the linear combinations of basis elements
{φ1, φ2, . . . , φ R } that represents a given snapshot dk .
The same reasoning holds for transforms in the time domain. In this case, let
the signal be a row of the data set matrix, that is, u := dTi [k] ∈ Rnt ×1 . Defining
the temporal basis matrix as Ψ = [ψ1, ψ2, . . . , ψnψ ] ∈ Cnt ×nψ and handling the
transpositions with care,8 the analogous of (8.4) in the time domain reads
dTi = Ψ dTψ ⇒ d̃Tiψ = (Ψ† Ψ)−1 Ψ† dTi ; D = Dψ ΨT ⇒ D̃ψ = DΨ(Ψ† Ψ)−1 . (8.5)

Here the transformed vectors are d̃iψ and the matrix D̃ψ collects in its columns all
the transformed vectors, that is, the coefficients of the linear combinations of basis
elements {ψ1, ψ2, . . . , ψR } that represent the data evolution at a location di [k].
Note that approximations can be obtained along space and time domains as
D̃φ = Φ(Φ† Φ)−1 Φ† D and D̃ψ = DΨ(Ψ† Ψ)−1 ΨT . (8.6)

7

−1
Use the distributive property of the inversion to show that PB = B(B† B)−1 B† = BB−1 B† B = I .
8 More generally, in a matrix multiplication AB, the matrix A is acting on the columns of B. The same
action along the rows of B is obtained by ABT .

h d i 0 0 9 8 088962 4 0 3 bli h d li b C b id i i
 Mendez: Generalized and Multiscale Modal Analysis 159

These approximations are exact (the projections are reversible) if the bases are
complete, that is, the matrices Φ and Ψ are square9 (nφ = ns and nψ = nt ). In these
cases, autoencoding is loseless: D̃φ = D and D̃ψ = D.
It is left as an exercise to see how (8.4)–(8.6) simplify if the bases in the space and
time are also orthonormal, that is, the inner products yield Φ† Φ = I and Ψ† Ψ = I
regardless of the number of basis elements10 nφ and nψ . On the other hand, from the
fact that (8.6) is exact only for complete bases, one can see that ΦΦ† = I only if
nφ = ns and ΨΨ† = I only if nψ = nt .

8.2.3 Projections in Space and Time

Let us now consider 2D transforms of the data set matrix. These are the discrete
version of 2D transforms, which in this case act on the columns and the rows of
the matrix independently. Let Dφψ be the 2D transform of the matrix D taking
Φ = [φ1, φ2, . . . , φ nφ ] ∈ Cns ×nφ and Ψ = [ψ1, ψ2, . . . , ψnψ ] ∈ Cnt ×nψ as basis matrices
for the columns and the rows, respectively. The general 2D transform pair is
D = ΦDφψ ΨT ⇔ D̃φψ = (Φ† Φ)−1 Φ† D Ψ(Ψ† Ψ)−1 . (8.7)
A classic example of this kind of transforms is the 2D Fourier transform in which
case the bases matrices are
1 1
Φ F [m, n] = √ e2πj m n/ns and Ψ F [m, n] = √ e2πj m n/nt . (8.8)
ns nt
These are the Fourier matrices, which are orthonormal. If these are also square (i.e.,
the number of frequency bins equals ns and nt , respectively), these matrices are also
symmetric. Then, (8.7) is reversible and the 2D discrete Fourier pair is11
D = Φ F DΨ b = ΦF D ΨF .
b F ⇐⇒ D (8.9)
The (discrete) Fourier transform plays an essential role in Section 8.4. Let us use
a wide hat b to indicate the Fourier transform of a matrix, whether this is carried out
along its rows, columns, or both.
Different decompositions can be derived by taking different bases for the rows
and for the columns. This is often done in the 2D discrete wavelet transform
(see Chapter 5), as one might desire to highlight different features along different
directions. For example, in the literature of image processing, we define horizontal
details as those features that have high frequencies over the column space and low
frequencies over the row space. We can thus construct a basis (i.e., design a filter) that
enhances or reduces these details.
Decompositions of this form are common and useful in image processing. For
reduced-order modeling, however, these have a substantial disadvantage: the number

9 And, of course, have full rank.

10 If Ψ† Ψ = I, then taking the conjugation on both sides gives ΨT Ψ = I while taking the transposition on
both sides gives ΨΨT = I.
11 Observe that transpositions are removed because of the symmetry of the bases.

h d i 0 0 9 8 088962 4 0 3 bli h d li b C b id i i
 160 Data-Driven Fluid Mechanics Chapter 8

of basis elements (i.e., its “modes”) is ns × nt . While most of the data have a sparse
representation in a 2D Fourier or wavelet bases,12 this representation is still inefficient.
The inefficiency stems from the lack of a “variable separation”: the entry Dφψ [m, n]
measures the correlation with a basis matrix constructed from the mth basis element of
the column space and the nth basis element of the row space. Even in the ideal case of
orthonormal bases on both columns and rows (i.e., space and time), the decomposition
in (8.7) requires a double summation:
s −1 n
nÕ Õ t −1

D[i, k] = Dφψ [m, n]φ m [i] ψn [k]. (8.10)

m=0 n=0

In modal analysis, we seek a transformation that renders Dφψ diagonal and that has
no more than R = min(ns , nt ) modes. We seek separation of variables and hence enter
data-driven modal analysis from Section 8.2.4.

8.2.4 The Fundamental Factorization

We impose that Dφψ is diagonal. In this case, the transformed matrix is denoted as
Σ = diag(σ1, . . . , σR ) ∈ RR×R and the entries σr are referred to as mode amplitudes.
The summation in (8.10) and the factorization in (8.7) become
R−1
Õ R−1
Õ
D = Φ Σ ΨT ⇐⇒ D[i, k] = σr φ m [i] ψn [k] ⇐⇒ D = σr φr ψrT . (8.11)
r=0 r=0

Each of the terms σr φr ψrT produces a matrix of unitary rank, which we call mode.
Note that the rank of a (full rank) rectangular matrix is rank(D) = min(ns , nt ), but
in general the number of relevant modal contributions is not associated to the rank.13
We obtain approximations of the data set by zeroing some of the entries along the
diagonal of Σ. As in Section 8.2.3 we use tildes to denote approximations. Moreover,
since infinite decompositions could be obtained by dividing the diagonal entries σr =
Σ[r, r] by the length of the corresponding basis matrices, we here assume that ||φr || =
||ψr || = 1 for all r ∈ [0, . . . , R − 1].
The reader has undoubtedly recognized that the factorization in (8.11) has the same
structure of the singular value decomposition (SVD) introduced in Chapter 6. The
SVD is a very special case of (8.11), but there are infinite other possibilities, depending
on the choice of the basis. Nevertheless, such a choice has an important constraint that
must be discussed here.
Assume that the spatial structures Φ are given. Projecting on the left gives
 σ1 ψ1 [1] σ1 ψ1 [2] ... σ1 ψ1 [nt ] 
 σ2 ψ2 [1] σ2 ψ2 [2] ... σ2 ψ2 [nt ] 

(Φ Φ) Φ D = Σ Ψ = Dφ = 
† −1 † T
.. .. .. ..  . (8.12)


 . . . . 

σ ψ [1] σR ψR [2] ... σR ψR [nt ]
 R R

12 That is, most of the entries in D φ , Dψ or D φψ are zero or almost zero.

13 For example, we might add five modes and obtain a matrix of rank four. The number of modes equals
the rank of the matrix only if these modes are orthogonal in both space and time, that is, only if these
are POD modes.

h d i 0 0 9 8 088962 4 0 3 bli h d li b C b id i i
 Mendez: Generalized and Multiscale Modal Analysis 161

This is equivalent to (8.4): this is a transform along the columns of D. The transformed
snapshot, that is, the set of coefficients in the linear combination of φ’s, changes from
snapshot to snapshot – they evolve in time. The key difference with respect to a 2D
transform is that the evolution of the structure in the basis element φr only depends on
the corresponding temporal structure ψr . Moreover, notice that σr = ||σr ψr ||, since
||ψr || = 1. This explains why the amplitudes are real quantities by construction: they
represent the norm of the rth column Dφ , which we denote as Dφ [:, r] following a
P YTHON notation.
The same observation holds if the temporal structures Ψ are given. In this case,
projecting on the right gives

 σ1 φ1 [1] σ2 φ1 [1] ... σR φ R [1] 

 σ1 φ2 [2] σ2 φ2 [2] ... σR φ R [2] 

DΨ(Ψ† Ψ)−1 = ΦΣ = Dψ =  .. .. .. ..  . (8.13)


 . . . . 

σ φ [n ] σ2 φ R [ns ] ... σR φ R [ns ]
 1 R s

This is equivalent to (8.5): this is a transform along the rows of D. The set of
coefficients in the linear combination of ψ’s are spatially distributed according to
the corresponding spatial structures φ’s. As before, we see that the amplitudes of the
modes can also be computed as σr = ||σR φr || = ||Dψ [:, r]||.

Figure 8.1 The “space view” and the “time-view” of modal analysis. In the “space view,”
every snapshot dk [i] is a linear combination of basis elements φ’s. The coefficients of this
combination evolve in time according to the associated ψ’s. In the “time-view,” every temporal
evolution di [k] is a linear combination of basis elements ψ’s. The coefficients of this
combination are spatially distributed according to the associated φ’s.

The space-time symmetry is further elucidated in Figure 8.1, considering also the
√ √
cases in which Φ := I/ ns or Ψ := I/ nt . In the “space view” on the left, we
follow the data in time from a specific spatial basis. If this basis is the set of impulses
δk [i], the temporal evolutions along each element of the basis is given by the time
evolution di [k]. In the “time view” on the right, we analyze the spatial distribution
from a specific temporal basis. If this basis is the set of impulses δi [k], the spatial
distribution of each member of the basis (i.e., every instant) is given by the snapshot
dk [i] itself.

h d i 0 0 9 8 088962 4 0 3 bli h d li b C b id i i
 162 Data-Driven Fluid Mechanics Chapter 8

With these views in mind, the reader should understand the most important
observations of this section: in the decomposition in (8.11) it is not possible to impose
both basis matrices Φ and Ψ– given one of the two, the other is univocally determined.
This observation also leads to the formulations of two general algorithms to compute
this factorization given one of the two bases. These are listed below:

Algorithm 1: Projection from Φ

Input: D ∈ Rns ×nt and Φ ∈ Rns ×R
Output: Σ ∈ RR×R and Ψ ∈ Rnt ×R
1 Project D̃φ = (Φ† Φ)−1 Φ† D (eq. 12)
2 From (4.12) it is Dφ = ΣΨT , hence :
3 for r <= R, r=r++, σr = ||Dφ [r, :]||2 ; ψr = Dφ [r, :]/σr
4 Assembly Σ = diag(σ0, σ1, . . . σR ) ∈ RR×R and Ψ = [ψ0, ψ1, . . . ψR ] ∈ Cnt ×R

Algorithm 2: Projection from Ψ

Input: D ∈ Rns ×nt and Ψ ∈ Rnt ×R
Output: Σ ∈ RR×R and Φ ∈ Rns ×R
1 Project D̃ψ = DΨ(Ψ† Ψ)−1 = ΦΣ (eq. 13)
2 From (4.12) it is Dψ = ΦΣ, hence :
3 for r <= R, r=r++, σr = ||Dψ [r, :]||2 ; φr = Dφ [r, :]/σr
4 Assembly Σ = diag(σ0, σ1, . . . σR ) ∈ RR×R and Φ = [φ0, φ1, . . . φ R ] ∈ Cns ×R

Finally, note that these algorithms are the most general ones to complete a
decomposition from its spatial or its temporal structures. Such a level of generality
highlights the common structure but also leads to the least efficient approach: every
decomposition offers valuable shortcuts that we discuss in Section 8.3.

8.2.5 Amplitudes and Energies

The normalization of a data-driven decomposition cannot be performed beforehand,
because the basis is initially unknown. We here analyze the consequences of imposing
a unitary norm for the bases vectors φ and ψ and the link of this operation with the
continuous domain. Using the continuous inner product (see Chapter 4), the energies
of continuous basis element in space and time are
1 1
∫ ∫
2 2
||φr (x)|| = φr (x) φr (x) dΩ and ||ψr (t)|| = ψr (t)ψ r (t) dT , (8.14)
Ω Ω T T
having taken the spacial basis in a domain14 Ω, and the temporal basis in a domain
T. If a Cartesian and uniform mesh is used in space and time, and if a right end point
integration scheme is considered, these are approximated by their discrete equivalent
14 This could be an area in a 2D domain or a volume in a 3D domain. In this sense, the bases in the
continuous domain are always considered in terms of “densities.”

h d i 0 0 9 8 088962 4 0 3 bli h d li b C b id i i
 Mendez: Generalized and Multiscale Modal Analysis 163

1 1 1 1
||φr (x)|| 2 = ||φr || 2 = φr† φ and ||ψr (t)|| 2 = ||ψr || 2 = ψr† ψ , (8.15)
ns ns nt nt
where the norms are simply Euclidean norms ||a|| = a † a. For nonuniform meshes or
more sophisticated integration schemes, weighted inner products must be introduced
(see Chapter 6).
Consider now an approximation of the data using only one mode. This is a matrix
D̃[i, k] = σr φr [i]ψr [k] of unitary rank. The total energy associated to this mode,
assuming that this is the discrete version of a continuous space-time evolution, is
ns Õnt
1 1 Õ
∫ ∫
2
E{ D̃[x, t]} = D̃ (x, t)dΩdt ≈ D̃2 [i, k] . (8.16)
ΩT T Ω nt ns i=0 k=0

The summation on the right is the Frobenious norm,15 hence we have

1 σ2 σ2
E{ D̃[x, t]} ≈ tr{ D̃† D̃} = r tr{ψr φr† φr ψr† } = r . (8.17)
nt ns nt ns nt ns
Observe that φr† φr = 1 because of the unitary length of the spatial structures and
tr{ψr ψrT } = ψr† ψr = 1 for the same constraint on the temporal structures. We thus
√
conclude that the square of the normalized amplitude σ̂r = σr / nt ns is mesh-
independent approximation of the energy contribution of each mode.
This statement is extremely important but requires great care: in general, summing
the energy contribution of each mode does not equal the energy in the entire data.
This occurs only if both the temporal and the spatial structures are orthonormal, that
is, only for the POD (see Chapter 6). For any other decomposition, the energy of an
approximation with r̃ < R modes is
1 n † †
o 1 n † †
o
E{ D̃[x, t]} = tr Ψ̃ Σ̃ Φ̃ Φ̃ Σ̃ Ψ̃ = tr Φ̃Σ̃ Ψ̃ Ψ̃ Σ̃ Φ̃ . (8.18)
ns nt ns nt
The convergence of a decomposition indicates how efficiently this can represent
the data set, that is, how quickly the norm of the approximation converges to the norm
of the data as r̃ → R. Convergence can be either measured in terms of l 2 norm or in
terms of Frobenius norm. Using the l 2 norm, the convergence can be measured as
||D − D̃(r̃)|| ||D − Φ̃Σ̃ Ψ̃||2
E(r̃) = = . (8.19)
||D|| ||D||

8.3 Common Decompositions

We here describe how common decompositions fit in the factorization framework.

We adopt a “time-view,” that is, we identify every decomposition from its temporal
structure Ψ and complete the factorization following Algorithm 2 from Section 8.2.

15 Recall: the Frobenious norm of a matrix A ∈ C n s ×n t can be written as | | A| | F = tr{ A† A} = tr{ AA† }.

h d i 0 0 9 8 088962 4 0 3 bli h d li b C b id i i
 164 Data-Driven Fluid Mechanics Chapter 8

8.3.1 The Delta Decomposition

This decomposition is useless, but conceptually important. Take the basis of impulses
√
Ψ δ = I/ nt as the temporal basis. This basis is orthonormal and we distinguish this
decomposition from the others using the subscript δ.
This decomposition generalizes the convolution representation of a signal intro-
duced in Chapter 4. The general temporal structure is ψr [k] = δr [k] = δ[k − r]. The
decomposition of the snapshot dk [i] becomes
R−1 R−1
Õ Õ δ[k − r] σδ k φ δk [i]
dk [i] = σδ r φ δr [i]ψδr [k] = σδ r φ δr [i] √ = √ . (8.20)
r=0 r=0
nt nt

Therefore, the spatial basis corresponds to the normalized snapshots:

dk [i] √ √
φ δk [i] = and σδ k = || dk [i]|| nt = || dk (xi )|| nt ns . (8.21)
||dk [i]||
This decomposition yields the highest possible temporal localization: large ampli-
tudes σδr correspond to a snapshot with large energy content. A strong amplitude
decay indicates that some snapshots have much lower energy content than others. In
a data set in which a strong impulsive event occurs, the POD temporal structure may
approach a delta-like form: when this happens, the associated spatial structure tends
to the corresponding normalized snapshot. A similar limit is obtained in a DWT (see
Chapter 5) if the finest scale is pushed to its limit. In a Haar wavelet transform,
√ for
example, the finest scale corresponds to elements of the form wr = (δr +δr+1 )/ 2. The
spatial structures associated to these are normalized averages of the snapshots dk and
√
dk+1 . Similarly, the spatial structure associated with the basis of the form ψr = 1/ nt
is the normalized time average of the data.

8.3.2 The Discrete Fourier Transform

We here refer to the DFT of the data set in time, that is, this is equivalent to performing
the DFT of D row-wise as in (8.5), followed by a normalization to compute the
amplitudes. In the DFT we set Ψ = Ψ F and we use the subscript F to distinguish
it from the others. This matrix is defined in (8.8). This basis is orthonormal and
symmetric. Using (8.13) and the general Algorithm 2 gives
b ⇐⇒ σFr = ||φr σr || = ||Dψ r || ,
DΨ F = Φ F Σ F = D that is, Φ F = DΣ
b −1 .
F
(8.22)
The first step on the left is the row-wise Fourier transform of D, denoted as D.b The
second step is the normalization, which can be computed along the rows of D, b and in
the last step we see that the spatial structures correspond to the normalized distribution
of a given entry of the Fourier spectra. We thus see that each entry of the spatial basis
tells how much the data at a specific location correlates with the harmonic associated
with the discrete frequency r. The computation of the digital frequencies from the
sampling frequency of the data is described in Chapter 4. In the same section, it was
recalled that the DFT implicitly assumes that the data is periodic of period T = nt ∆t.

h d i 0 0 9 8 088962 4 0 3 bli h d li b C b id i i
 Mendez: Generalized and Multiscale Modal Analysis 165

From a computational point of view, the multiplication by Ψ F can be efficiently

carried out using the FFT algorithm. Note that FFT routines from M ATLAB or P YTHON
define the Fourier basis matrix as the conjugate of (8.8). This decomposition is,
by definition, the one with highest frequency resolution. As described in Chapter
6, this comes at the price of no time localization, since this temporal basis is of
infinite duration. An important property of the Fourier basis is that its elements are
eigenvectors of circulant matrices (see Gray 2005).
This property links the DFT and the POD in stationary data, and it is worth a brief
discussion. An intuitive derivation follows from the notions introduced in Chapter 4,
specifically the convolution theorem in the discrete domain and the link between
impulse response of an LTI system and frequency transfer functions.
Hence, let u := u[k] ∈ Rnt ×1 with k ∈ [0, nt − 1] be a discrete signal, represented
by a column vector of appropriate size, which is introduced as an input to a SISO
system. Let y := y[k] ∈ Rnt ×1 be the corresponding output. Let h[k] be the impulse
response of the LTI system, assuming that this is of finite duration and that it has at
most nt nonzero entries. The input–output link in the time domain is provided by the
convolution sum (see Chapter 4). If both input and output are periodic, the convolution
can be written using a circulant matric Ch and a matrix multiplication:
l=∞
Õ
y[k] = u[k]h[k − l] → y = Ch u,
l=−∞

 y[0]   h[0] h[n − 1] h[nt − 2] ... h[1]   u[0] 

 y[1]   h[1] h[0] h[nt − 1] ... h[2]

 u[1]  (8.23)
    
.. ..

 u[2] 
   
 y[2]  
 =  h[2] h[1] h[0] . .  .


.. ..
 
.. ..
     
. . .

. h[2] h[1] h[n − 1]
    
    
 y[n − 1]  u[n − 1]
   h[n − 1] ... h[2] h[1] h[0]   

On the other hand, we know from the convolution theorem that the same can be
achieved in the frequency domain using an entry-by-entry multiplication of the Fourier
transform of the signals and the one of the impulse response. In terms of matrix
multiplications, these are, respectively,

u = ΨFu ,
b y = Ψ F y , and H
b b = ΨFh , (8.24)

where h := h[k] ∈ Rnt ×1 is the impulse response arranged as a column vector. To

perform the entry-by-entry matrix multiplication, we define a diagonal matrix H =
diag(H).
b Hence the operation in the frequency domain becomes b y = Hb u. Introducing
the Fourier transforms in (8.24) and moving back to the time domain, we get

b u → Ψ F y = H Ψ F u → y = Ψ F H Ψ F u hence Ch = Ψ F H Ψ F .
y = Hb (8.25)

The last equation on the right is obtained by direct comparison with (8.24) and is
extremely important: this is an eigenvalue decomposition. Each H[n] b entry of the
frequency response vector is an eigenvalue of Ch and the Fourier basis element ψ Fn
(corresponding to the frequency fn ) is the associated eigenvector.

h d i 0 0 9 8 088962 4 0 3 bli h d li b C b id i i
 166 Data-Driven Fluid Mechanics Chapter 8

Of great interest is the case in which the impulse response is even, that is, h[k] =
h[−k]. In this case the circulant matrix Ch becomes also symmetric and its eigenvalues
are real: this is the case of zero-phase (noncausal) filters. In this special case, also
the eigenvectors can be taken as real harmonics: they could be either sinusoidals or
cosinusoidals, that is, the basis of the discrete sine transform (DST) or the discrete
cosine transform (DCT) (see Strang 2007). In a completely different context, this
eigenvalue decomposition also appears in the POD of a special class of signals.
Because of its importance, the reader is encouraged to pause and practice with the
following exercise.
Exercise 1: The DFT and the Diagonalization of Circulant Matrices

Consider the discrete signal u from Exercise 5 in Chapter 4. First, compute

the DFT via matrix multiplication and compare that result with the one from
numpy’s FFT routine. Then prepare the impulse response h[k] of a low-
pass filter of order N = 211 using the windowing method and a Hamming
window. Then implement this filter using the matrix multiplication form of the
convolution in (8.23) and the matrix multiplication in the frequency domain
in (8.25). Compare the results with any of the previous implementations from
Chapter 4. Finally, construct the convolution matrix from (8.25) and test the
validity of the eigenvalue decomposition previously introduced.

8.3.3 The Proper Orthogonal Decomposition

The POD is introduced in Chapter 6; we here only focus on its special properties in
relation to Algorithm 2.
The temporal structures of the POD are eigenvectors of the temporal correlation
matrix, which in the notation of this chapter reads
t −1
nÕ
K = D† D = Ψ P ΛΨTP , that is, K [i, j] = σP2 ψ P [i]ψ P [ j]. (8.26)
r=0

The subscript P is used to distinguish the POD. The notation on the right is based
on an outer product representation of the eigenvalue decomposition of symmetric
matrices; this will be useful in Section 8.4. The first key feature of the decomposition
is that the eigenvalues of K are linked to the POD amplitudes as Λ = Σ2P . Hence the
diagonalization in (8.26) also provides the POD amplitudes and the normalization
in line 4 of Algorithm 2 is not needed. Introducing Ψ P in this algorithm gives
the Sirovinch formulation16 described in Chapter 6. Observe that since the POD
amplitudes are the singular values of the data set matrix, it is possible to compute
the convergence in (8.19) without computing norms:

16 Actually a much less efficient version: the projection step in line 1 of the algorithm could simply be
D̃ = DΨ P .

h d i 0 0 9 8 088962 4 0 3 bli h d li b C b id i i
 Mendez: Generalized and Multiscale Modal Analysis 167

v
u
2
σPr
t ÍR−1
||D − D̃(r̃)||2 r=r̃
E(r̃) = = . (8.27)
||D||2
ÍR−1 2
σPr
r=0

For the following discussion, two observations are of interest. The first is that
introducing the POD factorization (i.e., the SVD) in (8.26), we have
K = Ψ P Σ−1 T −1 T
P ΦP ΦP Σ P ΨP i.e., Λ = Σ−1
P ΦP ΦP Σ P .
T −1
(8.28)
The last step arises from a direct comparison with (8.26). Because Λ is diagonal, we
see that we must have ΦTP Φ P = I, that is, the spatial structures are also orthonormal.
We know this from Chapter 6, where it was shown that these are eigenvectors of
the spatial correlation matrix. The key observation is that the reverse must also be
true: every decomposition that has orthonormal temporal and spatial structure is a
POD. This brings us to the second observation on the uniqueness of the POD. In
a data set that leads to modes of equal energetic importance, the amplitudes of the
associated POD modes tend to be equal. This means repeated eigenvalues of K and
thus nonunique POD. In the extreme case of a purely random data set, it is easy to
see that the POD modes are all equal (see Mendez et al. 2017), and there are infinite
possible PODs. The impact of noise in a POD decomposition is further discussed in
Chapter 9.
Finally, observe that the POD is based on error minimization (or, equivalently,
amplitude maximization) and has no constraints on the frequency content in its
temporal structures Ψ P . However, a special case occurs in an ideally stationary
process. In such a process, the temporal correlations are invariant with respect to time
delays and solely depend on the time lag considered in the correlation. Hence the
correlation K[1, 4] = dT1 d4 is equal to K[4, 7] = dT4 d7 or K[11, 14] = dT11 d14 , for
example.
In other words, the temporal correlation matrix K becomes circulant, like the
matrix Ch in (8.23) for the convolution. Therefore, its eigenvectors are harmonics:
we conclude that the POD of an ideally stationary data set is either a DCT or a
DST.17 This property is the essence of the spectral POD proposed by Sieber et al.
(2016), which introduces an ingenious FIR filter along the diagonal of K to reach
a compromise between the energy optimality of the POD modes and the spectral
purity of Fourier modes. Depending on the strength of this filter, the SPOD offers
an important bridge between the two decompositions.

8.3.4 The Dynamic Mode Decomposition

The DMD is introduced in Chapter 7 and we here focus on its link with the
Algorithm 2. The DMD aims at fitting a linear dynamical system to the data, writing
every snapshot as

17 Depending on whether the temporal average has been removed. The mean is accounted for in a DCT but
not in a DST.

h d i 0 0 9 8 088962 4 0 3 bli h d li b C b id i i
 168 Data-Driven Fluid Mechanics Chapter 8

s −1
nÕ s −1
nÕ
di [k + 1] = ar ψ Dr [i]e−pr tk = ar ψ Dr [i]λrk−1 . (8.29)
r=0 r=0

The subscript D is used to distinguish the DMD. In Chapter 4, we have seen that
the complex exponentials are eigenfunctions of linear dynamical systems; Chapter
10 describes their state-space representation. Observe that the DMD is a sort of Z-
transform in which the basis elements only include the poles of the systems (see
Chapter 4), that is, the eigenvalues of the propagating matrix A in the state-space
representation in Chapter 10. Such decomposition is natural for a homogeneous
linear system, while the extension of the DMD for the forced system is proposed by
Proctor et al. (2016). This extension makes the DMD an extremely powerful system
identification tool that can be combined with the linear control methods introduced in
Chapter 10.
Many algorithms have been developed to compute the eigenvalues (λr ’s) from the
data set (see Chapter 7). Once these are computed, the matrix containing the temporal
structures of the DMD can be constructed,
 1
 1 1 ... 1 
λ λ2 λ3 ... λnt 
 1
 2
=  λ1 λ22 λ32 ... λn2 t  ∈ Cnt ×nt ,

ΨD (8.30)

 . .. .. ..
 ..

 . . . 

 nt
λ λ2nt λ3nt ... λnntt 

 1
and the decomposition completed following Algorithm 2.
From a dynamical system perspective, we know from Chapter 4 that a system is
stable if all its poles are within the unit circle, hence if all the λ’s have modulus
|λ| ≤ 1. From an algebraic point of view, we note that Ψ D differs from the temporal
structures of the other decompositions in two important ways. First, it is generally not
orthonormal: the full projection must be considered in line 1 of Algorithm 2. Second,
its inverse might not exist: convergence is not guaranteed.
Different communities have different ways of dealing with this lack of convergence.
In the fluid dynamics community, more advanced DMD algorithms, such as the
sparsity promoting DMD (Jovanović et al. 2014) or the optimized DMD (Chen
et al. 2012), have been developed to enforce that all the λ’s have unitary modulus. This
is done by introducing an optimization problem in which the cost function is defined
on the error minimization of the DMD approximation. Consequently, vanishing or
diverging decompositions are penalized. Observe that having all the modes on the
unit circle does not necessarily imply that these are DFT modes: the DMD does not
impose any orthogonality condition, which would force the modes to have frequencies
that are multiples of a fundamental one. If orthogonality is enforced as a constraint to
the optimization, then the only degree of freedom distinguishing DMD and DFT is
in the choice of the fundamental tone: while this is T = nt ∆t for the DFT, the DMD
can choose different values and bypass problems like spectral leakage or windowing
(Harris 1978).

h d i 0 0 9 8 088962 4 0 3 bli h d li b C b id i i
 Mendez: Generalized and Multiscale Modal Analysis 169

In the climatology community in which the DMD is known as POP

(see Hasselmann 1988, von Storch & Xu 1990) or LIM (see Penland &
Magorian 1993, Penland 1996), the lack of convergence is seen as a natural
limit of the linearization process. Accordingly, the results of these decompositions are
usually presented in terms of e-folding time, which is the time interval within which
the temporal structures decrease or increase by a factor e. In time-series analysis,
modes with a short e-folding time are modes with shorter predictive capabilities.
Variants of the POP/LIM have been presented in the fluid dynamics community
under the name of oscillating pattern decomposition (OPD, see Uruba 2012). The
link between DMD, POP, and LIM is also discussed by Schmid (2010) and Tu et al.
(2014).

Exercise 2: Eigenfunction Expansion and Modal Analysis

Consider a 2D pulsating Poiseuille flow within two infinite plates at a distance

2b over a length L, as shown in the left panel in the following figure. Assume
that the flow is fully developed (∂x u = 0) and incompressible (∂x u + ∂y v = 0).
It is easy to show that the pressure gradient must thus be uniform (∂x p =
const). If a sinusoidal pressure pulsation is introduced at the inlet, the Navier–
Stokes equation governing the velocity profile in the streamwise direction is
∆p(t)
ρ∂t u = − + µ∂yy u with ∆p(t) = p M + p A cos(ωt) , (8.31)
L
where ρ and µ are the density and the dynamic viscosity of the fluid.
Introducing the reference quantities
[y] = b [x] = L [p] = p M [t] = 1/ω , (8.32)
the problem can be written in dimensionless form as follows:
1 p̂ A
∂t û − ∂ŷ ŷ û = − 2 cos(tˆ), −1 < ŷ < 1, t ≥ 0,






 W 2 W
û(−1, t
ˆ ) = û(1, t
ˆ ) = 0, t ≥ 0, (8.33)


 û( ŷ, 0) = u0 ( ŷ), −1 < ŷ < 1 .



Here the hat denotes dimensionless variables, that is, â = a/[a] is the
dimensionless scaling of a variable a with respect to the reference quantity
[a]. The dimensionless number p controlling the response of the profile is the
Womersley number W = b ωρ/µ.
Equation (8.33) is a linear parabolic PDE with homogeneous Dirichlet bound-
ary conditions on both sides of the domain. A variable separated solution
can be found by expanding the velocity profile in terms of eigenfunctions
of the Laplacian operator. Given the boundary conditions in (8.32), these

h d i 0 0 9 8 088962 4 0 3 bli h d li b C b id i i
 170 Data-Driven Fluid Mechanics Chapter 8

are cosine functions. The solution via EF expansion (see Mendez & J.M-
Buchlin 2016) is
∞
Õ
û( ŷ, tˆ) = φ En ( ŷ) σEn ψ En (tˆ), (8.34)
n=1

with
π
 
φ En = cos (2n − 1) ŷ ,
2
16 p̂ A
σEn = , (8.35)
(2n − 1)π 16W 4 + (2n − 1)4 π 4
p

  2
W
ψ En = (−1) cos t − tan
n ˆ −1
.
[(2n − 1)π]
Note that the spatial structures are orthogonal but the temporal ones are not.
The amplitudes of the modes decay as ∝ 1/(2n − 1)3 when W → 0 and
as ∝ 1/(2n − 1) when W → ∞. Finally, note that all the amplitudes tend
to σEn → 0 as W → ∞: as the frequency of the perturbation increases,
the oscillations in the velocity profile are attenuated. Consider a case with
W = 10 and p̂a = 60. Several dimensionless velocity profiles during the
oscillations are shown in the left panel of the figure.
In this exercise, the reader should compare the DFT, the DMD, and the POD
with the eigenfunction solution. Assume that the space discretization consists
of ny = 2000 while the time discretization consists of nt = 200 with a
dimensionless sampling frequency of fˆs = 10.
First, construct the discrete data set from (8.34) to (8.35) by setting it in
terms of the canonical factorization in (8.11). Then, prepare a function that
implements Algorithm 2 described in Section 8.2.4 and use this algorithm to
compute the DFT, POD, and DMD. Finally, plot the amplitude decay of all the
decompositions and show the first three dominant structures in space and time
for each.

0.50

0.25

0.00
û

−0.25

−0.50

−1.0 −0.5 0.0 0.5 1.0

ŷ

(a) Schematic of the 2D pulsating Poiseuille flow considered in this exercise.

(b) Snapshots of the velocity profile during the pulsation.

h d i 0 0 9 8 088962 4 0 3 bli h d li b C b id i i
 Mendez: Generalized and Multiscale Modal Analysis 171

8.4 The Multiscale Proper Orthogonal Decomposition

This section is composed of three parts. We first analyze in Section 8.4.1 the impact
of frequency constraints on the POD. In particular, we are interested in what happens
if we filter a data set before computing the POD. We will see that if a frequency is
removed from the data, it is also removed from the temporal structures of its POD.
We extend this result to a decomposition of the data via MRA (see Chapters 4 and 5).
The MRA uses a filter bank to break the data into scales, and we here see how to keep
the PODs of all scales mutually orthogonal. Then, the POD bases of each scale can be
assembled into a single orthonormal basis. That is the basis of the mPOD – the mPOD,
presented in Section 8.4.2. Finally, the mPOD algorithm is described in Section 8.4.3.

8.4.1 Frequency-Constrained Proper Orthogonal Decomposition

We compute the POD of a data set that has first been filtered. Are the POD modes also
filtered? Can we obtain the same modes by filtering the data or filtering the temporal
correlation matrix? How are these linked to the POD modes of the original data? This
section answers these questions.
First, we need some definitions. Let the filter have a finite impulse response18
h ∈ Rnt ×1 . Following Section 8.3.2, the frequency response in the discrete domain
is a complex vector H b = Ψ F h ∈ Cnt ×1 . This vector can be arranged in two
possible matrices, both useful for the following discussion. The first is linked to
the diagonalization of circulant matrices, and is introduced in (8.25). The second
consists of a row-wise replication of the same vector. These are H ∈ Cnt ×nt and
H =1·H bT ∈ Cns ×nt , with 1 ∈ Rns ×1 a column vector of ones.19 To avoid confusion,
we here show both matrices20 :
H[0]
b 0 0 0  H[0]
b H[1]
b . . . H[n
b t ]
 0 H[1] 0 0  H[0] H[1] . . . H[n
b t ]
 b   b b
H= . .. .. ..  H= . .. .. ..  . (8.36)
 ..  ..
  
. . .  . . . 
H[0] H[1] . . . H[nt ]
 0 0 0 H[nt ]
   
b   b b b 

Recall that H is the eigenvalue matrix of the circulant convolution matrix Ch in
(8.23). Assuming that the filtering in the time domain is performed in all the spatial
location equally, the associated matrix multiplications are
D
bH D
bH
z }| { z }| {
= D Ch = DΨ F H Ψ F or = D Ψ F H Ψ F , (8.37)


DH DH
|{z} | {z }
D
b D
b

18 Which thus implies that the filter cannot be “ideal,” as discussed in Chapter 4.
19 √
To be normalized by 1/ ns to keep it of unitary length.
20 Note that the right multiplication by H is equivalent to the Hadamard product by H.

h d i 0 0 9 8 088962 4 0 3 bli h d li b C b id i i
 172 Data-Driven Fluid Mechanics Chapter 8

where  is the Hadamard product, that is, the entry-by-entry multiplication between
two matrices. Observe that D bH Ψ F is the inverse Fourier transform of DbH (i.e., the
frequency spectra of the filtered data) while D b = DΨ F is the Fourier transform of D
along its rows (i.e., in the time domain).
The representation on the left is a direct consequence of the link between convolu-
tion theorem and eigenvalue decomposition of circulant matrices introduced in (8.25).
The representation on the right opens to an intuitive and graphical representation of
the filtering process that is worth discussing briefly.
First, we introduce the cross-spectral density matrix K F . This collects the inner
product between the frequency spectra of the data evolution; it is the frequency domain
analogous of the temporal correlation matrix K . These matrices are linked:

KF = Db† Db = Ψ F D† D Ψ F = Ψ F K Ψ F ⇐⇒ K = Ψ F K F Ψ F . (8.38)
 

Since Ψ F Ψ F = I, the equation on the right is a similarity transform, hence K F and

K share the same eigenvalues. Introduce the eigenvalue decomposition of K in (8.26):

† t −1
nÕ
K F = Ψ F Ψ P Σ2P ΨTP Ψ F = Ψ
b P Σ2 Ψ
b P, that is, K F [i, j] = σP2 ψ
bP [i]ψ
bP [ j].


P
r=0
(8.39)
We thus see that the eigenvectors of K F are the conjugate of the Fourier transform
of the eigenvectors of K. The outer product notation on the right shows that the
diagonal of this matrix contains the sum of the power spectra of the temporal structures
of all the modes. This is the sum of positive real quantities.
Consider now the cross-spectral density matrix of the filtered data in (8.36) and use
the distributive property of the Hadamard product to get

†
b† D
KF = D b = D bH = D b  H †  H = K F  H,
b† D (8.40)




bH D
H H

having introduced the 2D symmetric frequency transfer function H = H † H. The

figure in Exercise 2 gives a pictorial representation of the magnitude of such 2D
frequency transfer function for ideal low-pass, band-pass, and high-pass filters. In
a low-pass filter, for example, the 2D frequency transfer function is unitary within a
square H[i, j] , 0, for all i, j ∈ [−ic , ic ] × [−ic , ic ] and approximately zero everywhere
else.
We thus see that the filtering of the data constraints the spectral content of the
POD modes: writing the cross-spectral density of the filtered data in terms of its
eigenvectors as in (8.38), we find

t −1
nÕ
K F H = K F  H ⇐⇒ K F H [i, j] = σP2 H ψ
bP H [i]ψ
bP H [ j]. (8.41)
r=0

Since K F H ≈ 0 outside the band-pass range of the 2D filter, and since the
entries along its diagonals are a summation of positive quantities, we conclude that
frequencies removed from the data cannot be present in any of the POD modes.
Moreover, we can compute the POD modes of the filtered data from the filtered

h d i 0 0 9 8 088962 4 0 3 bli h d li b C b id i i
 Mendez: Generalized and Multiscale Modal Analysis 173

Low-pass filtering

Band-pass filtering

High-pass filtering

Figure 8.2 Pictorial representation of the magnitudes of the frequency transfer functions of the
filters considered in this section. The figure on the left shows the magnitude of the 1D
frequency response H. b On the right, the figure shows how 2D filters H = H † H are
constructed from the extended response H. The red region represents the passband portion,
that is, within which |H| ≈ 1 while in the gray area it is |H| ≈ 0.

cross-spectral matrix K F H or, more conveniently, from the temporal correlation matrix
K H . The link between these matrices can be written as
K H = D†H D H = Ψ F K F H Ψ F = Ψ F K bH = Pπ K F H ,
bH Ψ F ⇐⇒ K (8.42)
where the definition of 2D Fourier transform in (8.8) is introduced and Pπ = Ψ F Ψ F
is the permutation matrix obtained by multiplying the Fourier matrix twice21 :
1 0 ... 0 0
0 0 0 1
 
.. 

Pπ = Ψ F Ψ F = Ψ F Ψ F = 0 0 1 .  . (8.43)

. . . .. 
 .. .. .. .

0 1
 0 ... 0
Finally, to answer the last question from this section, use the multiplication by the
diagonal matrix H in (8.39) instead of the Hadamard product:
K F H = DHb † DH = H†Db† DH = H † K F H . (8.44)



b
21 A funny question arose after reading this: what happens if the Fourier transform is performed four
times?

h d i 0 0 9 8 088962 4 0 3 bli h d li b C b id i i
 174 Data-Driven Fluid Mechanics Chapter 8

Since H † H , I, this is not a similarity transform. Hence the eigenvalues of these

two matrices are different: there is no simple link between σP and the σP H , nor
between the POD of the data and the POD of the filtered data. We do not attempt to
build such a connection: it suffices to notice that the POD of the filtered data is the
optimal basis for the portion of data whose spectra fall within the band-pass region of
the filter. In the terminology of MRA (see Chapter 4), we call this a scale.

Exercise 3: The POD of Filtered Data

Consider the test case from Exercise 2 but assume that the forcing is composed
of two terms. These are sinusoids associated to a Womersley number of W =
1 and W = 4, modulated by two windows We refer to these terms as F1
and F2 , respectively. As the problem is linear, we can sum the response of
the velocity profile to each term. The response of the velocity profile at the
centerline, for each of the terms, is shown in Figure 8.4.
In this exercise, we want to analyze the response of the velocity profile to
the perturbation at the highest frequency. First, perform a standard POD
and analyze the resulting structures. Then, filter the data using a high-pass
filter that isolates the forcing F2 and compute the POD again. Compute this
decomposition by filtering the data and by filtering the correlation matrix.
Compare the results. As an optional exercise, perform DMD and DFT using
the codes from Exercise 2: is it possible to separate the contribution of each
forcing term with these decompositions? The answer is no. Show it!

F1
1.0 F2
0.5
û(ŷ = 0, t)

0.0

−0.5

−1.0

0 100 200 300 400 500 600

t̂

Time evolution of the velocity profile at the center of the channel when each
of the two forcing terms F1 and F2 is active.

8.4.2 From Frequency Constraints to Multi-resolution Analysis

Section 8.4.1 provided the background to generalize the study of the frequency-
constrained POD to the mPOD. The mPOD constructs a set of orthogonal modes that
have no frequency overlapping, that is, no common frequency content within specific
frequency bandwidths.
To illustrate the general idea, let us consider a data set D ∈ Rns ×nt , with nt < ns .
We now use MRA to break the data set into scales. With no loss of generalities, let us

h d i 0 0 9 8 088962 4 0 3 bli h d li b C b id i i
 Mendez: Generalized and Multiscale Modal Analysis 175

consider three scales; these are identified by three filters with band-pass bandwidths
∆ f1 = [0, f1 ], ∆ f2 = [ f1, f2 ] and ∆ f3 = [ f2, fs /2]. We lump this information in a
frequency splitting vector FV = [ f1, f2, f3 ]. With three scales, the temporal correlation
matrix has nine partitions:
3 Õ
3

†
Õ
K = D † D = D1 + D2 + D3 D1 + D2 + D3 = Di† D j . (8.45)
i=1 j=1

Following an MRA formulation, we assume that these scales are isolated by filters
with complementary frequency response, that is,
Õ
Di = D Ψ F  H i Ψ F with H i ≈ 1 and H i  H j ≈ 0. (8.46)



This ensures a lossless decomposition of the data. We now use (8.39) and (8.45) in
(8.44) to analyze which portion of the cross-spectral density K F is taken by each of
the contributions in (8.44). Figure 8.3 gives a pictorial view of the partitioning, which
can be constructed following the graphical representation in Exercise 2. In wavelet
terminology, the term H †1 H 1 is the approximation term at the largest scale. The other
“pure” terms H †2 H 2 and H †3 H 3 are diagonal details of the scales 2 and 3. The terms
H †i H j with i > j are horizontal details while those with j < i are vertical details.

Figure 8.3 Repartition of the cross-spectral density matrix into the nine contributions
identified by three scales. The origin is marked with a red circle. The “mixed terms” H †i H j
with i , j are colored in light gray, while the “pure terms” H †i H i are colored following the
legend on the left, where the 1D transfer functions are also shown.

We have seen in Section 8.4.1 that removing a frequency from the data set removes
it from the temporal structures of its POD. There is thus no frequency overlapping22
between the eigenvectors of the contributions K1 = D1† D1 , K2 = D2† D2 , K3 = D3† D3 ,
which we denote as Km = Ψ Hm Λ Hm ΨTHm . However, the same is not necessarily true
for the correlations of the “mixed terms” Di† D j with i , j: the filters H †i H j with
i , j leave the diagonals of K F unaltered and thus the spectral constraints are much

22 At least in case of ideal filters, which are perfectly complementary.

h d i 0 0 9 8 088962 4 0 3 bli h d li b C b id i i
 176 Data-Driven Fluid Mechanics Chapter 8

weaker: it is possible that eigenvectors from D2† D3 , for example, have frequencies that
are already present among the eigenvectors of D2† D2 or D3† D3 .
We opt for a drastic approach: remove all the “mixed terms” Di† D j with i , j and
consider the correlation as the sum of M “pure terms”,
M
Õ M
Õ M
Õ
†
Dm = b  H †m H m Ψ F = Ψ Hm Λ Hm ΨTHm , (8.47)
 
K≈ Dm ΨF K
m=1 m=1 m=1

having introduced (4.41) and the eigenvalue decomposition of each term.

This is a fundamental step to link the mPOD to POD and DFT. Such simplification
results in a loss of information in the temporal correlation matrix, but it still ensures a
complete reconstruction of the data set. It remains to be seen if a basis constructed
by combining all the eigenvectors of different scales can form a complete and
orthonormal basis for nt .
Consider first each of these scales independently. Let us assume that the number of
nonzero frequencies in each scale, namely the number of frequency bins in the band-
pass region, is nm . If the frequency transfer functions are complementary, we have
nm = nt . If each scale contains at most nm frequencies, every temporal evolution of
Í
the data within that specific scale can be written as a linear combination of nm Fourier
modes. Therefore, every contribution Dm is at most of rank(Dm ) = nm and hence has
at most nm POD modes. These are clearly orthonormal, that is, ΨTHm Ψ Hm = I and
form a basis for that specific scale.
(1) (4)
Consider now modes from different scales, say for example ψ H2 and ψ H3 , that is,
the first eigenvectors from scales 2 and the fourth eigenvector from scale 3. Because
there is no frequency overlapping, their Fourier transforms are orthogonal, and thus
they must also be orthogonal:

ψ b(4) = 0 = ψ (1)† Ψ F Ψ F ψ (4) = ψ (1)† ψ (4) .

b(1)† ψ (8.48)
H2 H4 H2 H4 H2 H4

We thus conclude that a complete and orthonormal basis for Rnt can be constructed
from the eigenvectors of the various scales. That is the mPOD basis.
We close this section by highlighting how the mPOD connects POD and DFT or
DMD. In case of no frequency partitioning, the mPOD is a POD: the temporal struc-
tures can span the entire frequency range and are derived, as described in Section 8.2.4
and Chapter 6, under the constraint of optimal approximation for any given number
of modes. Introducing the frequency partitioning and the approximation in (8.46), we
identify modes that are optimal only within the frequency bandwidth of each scale.
The optimality of the full basis is lost as modes from different scales are not allowed
to share the same frequency bandwidth. As we introduce finer and finer partitioning,
each mode is limited within a narrower frequency bandwidth. At the limit for nm = 1
and M = nt , every mode is allowed to have only one frequency. The approximation
in (8.46) forces the spectra of the temporal correlation matrix to be diagonal, that is,
the correlation matrix is approximated as a Toeplitz circulant matrix. Accordingly, the
mPOD tends toward the DFT or DMD depending on the boundary conditions used
in the filtering of K. If periodicity is assumed, the DFT is produced. If periodicity

h d i 0 0 9 8 088962 4 0 3 bli h d li b C b id i i
 Mendez: Generalized and Multiscale Modal Analysis 177

is not assumed, the decomposition selects harmonic modes whose frequency is not
necessarily a multiple of the observation time, and a DMD is recovered.

8.4.3 The Multiscale Proper Orthogonal Decomposition Algorithm

The pseudo-code for computing the mPOD listed in Algorithm 3. This is composed
of eight steps.

Algorithm 3: Algorithm for computing the multiscale POD.

Input: D ∈ Rns ×nt , FV ∈ R M×1
Output: Ψ M ∈ Rnt ×R , Σ M ∈ RR×R , and Φ M ∈ Rns ×R
1 Compute temporal correlation matrix K = D† D.
2 Prepare the filter bank H1, H2, . . . , HM .
3 Decompose K into K1, K2, . . . , Km contributions.
4 Compute the temporal structures of each scale K = Ψm Λm ΨTm .
5 Assembly basis Ψ0 = [Ψ1, . . . , Ψ M ] and λ 0 = [diag(Λ1 ), . . . , diag(Λ M )].
6 Prepare permutation PΛ , sorting λ 0 . Then shuffle columns Ψ1 = Ψ0 PΛ .
7 Enforce orthogonality Ψ1 = Ψ M R → Ψ M = Ψ1 R−1 .
8 Complete the decomposition using Algorithm 2 Φ M → from D , Ψ M .

The first step computes the temporal correlation matrix, as in the POD. The second
step prepares the filter bank, according to the introduced frequency vector FV . The
third step computes the MRA of the temporal correlation matrix; the fourth step
computes the eigenvectors of each.23
At this stage, we can proceed with the preparation of a single basis. First, in
step 5, all the temporal structures and associated eigenvalues are collected into a single
matrix Ψ0 and a single vector of eigenvalues λ 0 . These eigenvalues give an estimation
of the relative importance of each term. This is sorted in descending order and the
information is used to permute the columns of Ψ0 using an appropriate permutation
matrix PΛ .
If the filters were ideal, the resulting temporal basis Ψ1 = Ψ0 PΛ would be com-
pleted. However, to compensate for the nonideal frequency response of the filters, a
reduced QR factorization is used in step 7 to enforce the orthonormality of the mPOD
basis. The result of the orthogonalization procedure is the matrix of the temporal
structure of the mPOD. The final step is the projection to compute the spatial structures
that is common to every decomposition and that can be computed using24 Algorithm 2.
Note that special attention should be given to the filtering process in the third
step. If this is done as in Exercise 4, it is implicitly assumed that the matrix K is
23 Can we by-pass this step? Can we compute all the Ψ m ’s from one single diagonalization (of a properly
filtered matrix)? These were two brilliant questions by Bo B. Watz, development engineer at Dantec
Dynamics, who attended the lecture series. The answer is that it is possible, in principle, to compute the
basis from one clever diagonalization. However, the filtering is challenging: that was the starting point
of new exciting developments toward the fast mPOD.
24 Observe that a simplified version could be used since the temporal structure is orthonormal!

h d i 0 0 9 8 088962 4 0 3 bli h d li b C b id i i
 178 Data-Driven Fluid Mechanics Chapter 8

periodic. When such an assumption is incorrect, edge effects could appear after the
filtering process. The cure for these effects is described in classical textbooks on image
processing (see Gonzalez & Woods 2017) and include reflections or extrapolations. To
limit the scope of this chapter, we do not address these methods here.25 The reader
should nevertheless be able to construct the mPOD algorithm using all the codes
developed for the previous exercises.

Exercise 4: Your own mPOD Algorithm

Combine the codes from the previous exercises to build your own function for
computing the mPOD, following Algorithm 3. This function should compute
Ψ M , while the decomposition can be completed using Algorithm 2.
Assume that filters are constructed using Hamming windows, and the filter
order is an input parameter. Test your function with the previous exercise and
show the structures of the first two mPOD modes. Compare their amplitudes
with the ones of the POD modes.

8.5 Tutorial Test Cases

In addition to the coding exercises, this chapter includes two tutorial test cases from
time-resolved particle image velocimetry measurements. The first data set collects the
velocity field of a planar impinging gas jet in stationary conditions; the second is the
flow past a cylinder in transient conditions. These are described in Mendez, Balabane
and Buchlin (2019) and Mendez et al. (2020). Other examples of applications of the
mPOD on experimental data can be found in Mendez et al. (2018), Mendez et al.
(2019) and Esposito et al. (2021). More exercises and related codes can be found
in https://github.com/mendezVKI/MODULO, together with an executable with
graphical user interface (GUI) developed by Ninni and Mendez (2021).

8.5.1 Test Case 1

To download the data set and store it in a local directory, run the following script
import urllib . request
print ( ’ Downloading Data for Tutorial 1... ’)
url = ’ https :// osf . io / c28de / download ’
urllib . request . urlretrieve ( url , ’ Ex_4_TR_PIV_Jet . zip ’)
print ( ’ Download Completed ! I prepare data Folder ’)
# Unzip the file
from zipfile import ZipFile
String = ’ Ex_4_TR_PIV_Jet . zip ’
zf = ZipFile ( String , ’r ’) ; zf . extractall ( ’ ./ ’) ; zf . close ()

25 Interested readers can find the related codes at https://github.com/mendezVKI/MODULO

h d i 0 0 9 8 088962 4 0 3 bli h d li b C b id i i
 Mendez: Generalized and Multiscale Modal Analysis 179

The data set consists of nt = 2000 velocity fields, sampled at fs = 2 kHz over a grid
of 60 × 114 points. The spatial resolution is approximately ∆x = 0.33 mm. The script
provided in the book’s website26 describes how to manipulate this data set, plot a time
step, and more. For the mPOD of this test case, the frequency splitting vector used
in Mendez, Balabane and Buchlin (2019) is constructed in terms of Strouhal number
St = f H/UJ , that is, the dimensionless frequency computed from the advection time
H/UJ , with H the standoff distance and UJ the mean velocity of the jet at the nozzle
outlet. These are H = 4mm and UJ = 6.5 m/s.
The reader is encouraged to compare the mPOD results with those achievable from
other decompositions. The spatial structures and the spectra of the temporal structures
in the first five modes are produced by the script T UT 1. PY. These are shown in Figure
8.4 for the dominant mPOD mode in the scale St = 0.1 − 0.2. This mode isolates the
roll-like structures produced by the evolution of the shear layer instability downstream
the potential core of the jet. Recall that the velocity components are stacked into a
single vector.

Figure 8.4 On the left: example snapshot from the first tutorial test case on the TR-PIV of an
impinging gas jet. Colormap in m/s. On the right: example of spatial structure (top) and
frequency content (bottom) of the temporal structures of a mPOD mode.

8.5.2 Test Case 2

To download this data set and store it into a local directory, run the following script
import urllib . request
print ( ’ Downloading Data for Tutorial 2... ’)
url = ’ https :// osf . io / qa8d4 / download ’
urllib . request . urlretrieve ( url , ’ Ex_5_TR_PIV_Cylinder . zip ’)
print ( ’ Download Completed ! I prepare data Folder ’)
26 www.datadrivenfluidmechanics.com/chapter8

h d i 0 0 9 8 088962 4 0 3 bli h d li b C b id i i
 180 Data-Driven Fluid Mechanics Chapter 8

# Unzip the file

from zipfile import ZipFile
String = ’ E x _ 5 _ T R _ P I V _ C y l i n d e r . zip ’
zf = ZipFile ( String , ’r ’) ;
zf . extractall ( ’ ./ DATA_CYLINDER ’) ; zf . close ()
The experimental setup and configuration are extensively described in Mendez et al.
(2020). This data set is computationally much more demanding than the previous, and
consists of nt = 13,200 velocity fields sampled at fs = 3 kHz over a grid of 71 × 30
points. The spatial resolution is approximately ∆x = 0.85 mm. A plot of a time step is
shown in Figure 8.5

13

12
U∞ [m/s ]

11

10

8
0 1 2 3 4
t[s]

Figure 8.5 Top left: a snapshot of the velocity field past a cylinder, obtained via TR-PIV.
Colormap in m/s. Bottom left: evolution of the free-stream velocity as a function of time,
sampled from the top left corner of the field. Right: example of spatial structure (top) and
frequency content (bottom) of the temporal structures of an mPOD mode.

As for the previous tutorial, a function is used to extract all the information about
the grid (in this case stored on a different file) and the velocity field.
This test case is characterized by a large-scale variation of the free-stream velocity.
A plot of the velocity magnitude in the main stream is shown in Figure 8.5. In the first
1.5 s, the free-stream velocity is at approximately U∞ ≈ 12 m/s. Between t = 1.5 s
and t = 2.5 s, this drops down to U∞ ≈ 8 m/s. The variation of the flow velocity is
sufficiently low to let the vortex shedding adapt, and hence preserve an approximately
constant Strouhal number of St = f dU∞ ≈ 0.19, with d = 5 mm the diameter of the
cylinder. Consequently, the vortex shedding varies from f ≈ 459 Hz to f ≈ 303 Hz.
Interestingly, the POD assigns the entire evolution to a single pair of modes, and it is
hence not possible to analyze the vortex structures in the shedding for the two phases
at approximately constant velocity, nor to distinguish them from the flow organization
during the transitory phase.
The mPOD can be used to identify modes related to these three phases. Three
scales are chosen in the exercise. The first, in the range ∆ f = [0 − 10] Hz, is designed
to isolate the large-scale motion of the flow, hence the variation of the free-stream

h d i 0 0 9 8 088962 4 0 3 bli h d li b C b id i i
 Mendez: Generalized and Multiscale Modal Analysis 181

velocity. The scale with ∆ f = [290–320] Hz is designed to capture the vortex shedding
when the velocity is at U∞ ≈ 8m/s, while the scale with ∆ f = [430–470] Hz identifies
the shedding when the velocity is at U∞ ≈ 12 m/s.
The spatial structure and the temporal evolution of an mPOD mode are also shown
in Figure 8.5. This mode is clearly associated to the second stationary conditions.
Other modes capture the first part while others are focused on the transitory phase:
besides allowing for spectral localization, the MRA structure of the decomposition
also allows for providing time localization capabilities.

8.6 What’s Next?

This chapter presented a general matrix factorization framework to formulate any

linear decomposition. This general framework allowed for deriving the mPOD, a
formulation that allows us to bridge energy-based and frequency-based formulations.
The proposed exercises and the extensive collection of codes listed in the book’s
website should allow the reader to efficiently use these decompositions on his/her
own data and provide a good starting point to develop computational proficiency on
this subject.
It is important to note that this chapter has only focused on the mathematical
framework of linear decompositions: Every decomposition provides its own way of
representing the data with respect to a certain basis. From here, the reader might
be interested in two possible directions. On the representation side, while linear
dimensionality reduction has been at the heart of an extensive body of literature in
fluid mechanics, nonlinear methods are now offering a new set of tools as showcased
in Chapter 1. On the reduced-order modeling side, whether a low dimensional
representation is capable of reducing the complexity of a dynamical system depends
on whether the projection (or the nonlinear mapping) is able to preserve the most
dynamically relevant phenomena in the new representation. These challenges are
described in Chapter 14.

h d i 0 0 9 8 088962 4 0 3 bli h d li b C b id i i
 9 Good Practice and Applications
of Data-Driven Modal Analysis
A. Ianiro

This chapter1 develops a theoretical background for the definition of good practice
(number of samples and sampling time versus the effect of noise) for data-driven
modal analysis and shows an example application to a channel-flow data set. Snapshot
proper orthogonal decomposition (POD) is employed as a benchmark allowing to
move the discussion toward the convergence of the flow statistics and the effect of
measurement noise on the temporal correlation matrix.
Beyond the estimation of the effect of noise on the estimated modes, this chapter
also presents several application examples showing how hidden information can be
extracted from a well-converged POD. In particular, it is shown how temporal modes
of non-time-resolved data can provide detailed phase information and how extended
POD modes can provide a linear stochastic estimation of correlated events. This
is useful for applications that involve flow sensing and for the study of convection
problems.

9.1 Introduction

When analyzing flow data sets, both from experiments and simulations, the researcher
is challenged by important questions about the needed size and completeness of the
data set.
The basic question that the researcher might ask him/herself is how many samples
do I need to store to obtain a reliable modal analysis? Considering, for instance, the
shedding phenomenon of the wake of a cylinder as discussed in Chapter 1, it is rather
intuitive that it is needed to correctly sample several phases of the shedding period and
to acquire enough samples per phase to achieve a satisfactory convergence of the flow
statistics and thus of the correlation matrix. Unfortunately, data storage, simulation or
experiment duration, and processing time might limit the amount of samples stored.
The question previously identified can be further subdivided into two questions: while
acquiring a data set, what should be the acquisition frequency? Moreover, given a
1 Andrea Ianiro acknowledges the support of his colleagues and friends Prof. S. Discetti and Prof. M.
Raiola. The majority of the original concepts summarized in the present chapter is the result of a
fruitful collaboration at UC3M in the last eight years. The application of the turbulent heat transfer in a
pipe is the result of a collaboration with Dr. A. Antoranz, Prof. O. Flores and Prof. M. Garcı́a-Villalba.
A. Antoranz, O. Flores and M. Garcı́a-Villalba are kindly acknowledged for sharing the turbulent-pipe
data set.

h d i 0 0 9 8 088962 4 0 4 bli h d li b C b id i i
 Ianiro: Good Practice and Applications 183

certain acquisition frequency, how many samples should be acquired? Answers to

these questions can be obtained the from basic statistical data characterization theory
described in books on experimental methods (Discetti & Ianiro 2017) and sits on the
theoretical arguments presented in Chapters 4 and 5.
While the two previous questions are of rather straightforward answer for numerical
simulations (which have only negligible truncation errors), experimental data are often
affected by uncertainties. Measurement uncertainty includes bias and random parts.
While the bias contribution mostly affects the estimated values of mode amplitude
and energy content, random uncertainties have a more complicated effect on the
correlation matrix. This fact raises a more intriguing question, that is, how does
measurement noise affect the convergence of the modes? A seminal work on the topic
was presented by Venturi (2006); following Venturi’s work, a simple practical criterion
for POD truncation was proposed by Raiola et al. (2015). Although the results are
derived for POD, all the modal decomposition approaches based on the singular value
decomposition of a snapshot matrix are affected by the noise perturbation of the time-
correlation matrix.
With a special focus on particle image velocimetry (PIV) data, despite the
continuous improvement of the available hardware, it is often not possible to perform
measurements at frame rates sufficiently high to obtain a characterization of the
flow dynamics. The researcher should wonder whether it could be possible to
extract information about the modes dynamics. Although temporal sampling will not
satisfy the Nyquist theorem, the phase relation between modes might be preserved,
and Lissajous curves of the temporal modes can allow retrieving this information.
Knowing dynamical information about some specific modes, both from theory or
from additional experiments, it can then be possible to identify the flow dynamics.
The suggested approach is explained by means of an application to the wake behind a
couple of tandem cylinders in ground effect (Raiola et al. 2016).
The temporal modes, as shown in for example Chapters 6 and 8, are a basis
for the temporal evolution of the data and can be used as a basis to project and
decompose other synchronized data sets. This decomposition, referred to as extended
proper orthogonal decomposition (herein EPOD) and first proposed by Borée (2003),
allows us to ascertain correlations between different flow quantities and/or flow
measurements (Picard & Delville 2000, Duwig & Iudiciani 2010).
This results in an excellent opportunity to retrieve temporal information from
synchronized time-resolved point measurements (for instance, from hot wires or
other high-repetition-rate pressure probes) and non-time-resolved flow-field measure-
ments (Tinney et al. 2008, Baars & Tinney 2014, Hosseini et al. 2015, Discetti
et al. 2018, Discetti et al. 2019). Even if recent works have proven the superior
performances of neural networks for the purpose of sensor-based flow reconstructions
(Güemes et al. 2019, Guastoni et al. 2020), the author believes that EPOD represents a
founding brick of machine learning-based flow sensing, and a description is provided
herein.
For what concerns the correlation between different synchronized flow quantities,
EPOD is especially well suited for the analysis of combustion or heat transport

h d i 0 0 9 8 088962 4 0 4 bli h d li b C b id i i
 184 Data-Driven Fluid Mechanics Chapter 9

problems (Antoranz et al. 2018), allowing us to obtain composite modes, for example,
modes of temperature and velocity that provide a clear description of the scalar
transport mechanisms. A MATLAB R
exercise focusing on the analysis of a data set
from the work by Antoranz et al. (2018) is proposed at the end of the chapter.

9.1.1 A Brief Recap of the Snapshot Proper Orthogonal

Decomposition Procedure
Even if already introduced in Chapter 6, it is here worth briefly recalling the procedure
employed for the derivation of the snapshot POD. Consider a data set of ns snapshot
of discrete measurements of a certain quantity a evaluated at nx equally spaced
measurement points xi . Snapshot POD, introduced in Chapter 6, requires us to
assemble the data set into a matrix A in which each column represents a snapshot
of all the data measured at a given instance in time t j , and each row consists of all the
instances of a given spatial point xi :

 ax1 ,t1 ax1 ,t2 ··· ax1 ,tn s 

 ax2 ,t1 ax2 ,t2 ax2 ,tn s
 
··· 
A=  . .. .. .. . (9.1)
 
 .. . . . 
 
a
 xn x ,t1 axn x ,t2 ··· axn x ,tn s .

Following the notation of Chapter 6, it is then useful to subtract a base condition y0

to each snapshot, commonly chosen to be the mean of all the columns of A, obtaining
a matrix Y composed of vectors y j 0. The POD can be found from the singular value
decomposition of Y = ΦΣΨ∗ . If nx > ns , a computationally more efficient approach
requires the solution of the eigenvalue problem for the temporal correlation matrix
Y ∗Y ,
 y10, y10

0 0 
y10, y20  · · ·
y10, yn0 s  


0 0
0 0 

 y ,y y2, y2 ··· y2, yn s 
 2 1
Y ∗Y =  .. .. .. ..  = ΨΛΨ∗, (9.2)

. 
. 
 . . 

 y0 , y0 n s , y20 ··· yn s , yn0 s 

0 
 ns 1
D E
where Y ∗ is the conjugate transpose of Y and yi0, y j0 is the scalar product of the
vectors yi and y j estimated at time instants ti and t j . The solution of the eigenvalue
problem of the cross-correlation matrix Y ∗ Y returns as right singular vectors the rows
of Ψ∗ , that is, the temporal modes of the POD, while the diagonal eigenvalue matrix
Λ = Σ2 contains the variance of each mode. The matrix Ψ can be used as a basis
for the projection of the matrix Y in order to obtain the spatial modes, which are the
columns of the matrix Φ. Spatial modes are thus a weighted average of the snapshots
through the matrix product:

Φ = Y ΨΣ−1 . (9.3)

h d i 0 0 9 8 088962 4 0 4 bli h d li b C b id i i
 Ianiro: Good Practice and Applications 185

9.2 Data Set Size and Richness

9.2.1 Effect of the Number of Samples on the Convergence of the Statistics

The answer to the question “how many samples are needed to obtain a well-converged
modal decomposition?” depends on the data set. It is rather intuitive that the statistical
properties of laminar flows can be estimated with much fewer samples than those of
a turbulent flow. The solution of the eigenvalue problem of the correlation matrix,
employed for the snapshot POD, provides modes that are ordered by their contribution
to the flow-field variance. It is thus intuitive that a reliable modal decomposition
requires us to have access to a data set that is sufficiently large and rich to achieve
a good statistical convergence for what concerns both the estimation of the mean
quantities and of their variance (i.e., turbulent kinetic energy if we are analyzing a
velocity field).
Considering n independent samples of a flow field u(x, y), without noise content, it
is possible to estimate the mean and the variance of a flow field as
ui
Ín
u ≈ i=1 ,
n
2 (9.4)
i=1 (ui − u)
Ín
u ≈
02 .
n−1
The two relations in (9.4) tend to equalities for large numbers of samples. According to
Everitt and Skrondal (2002), the standard error of a statistic is the standard deviation σ
of its sampling distribution. Following Kenney and Keeping (1951), it can be derived
that the standard errors of the mean and the variance are equal to
p
u 02
σu = √ ,
n
s (9.5)
2
(n − 1) (n − 1) (n − 3)
σu02 = u 02 γu − ,
n3 n3
where γu is the kurtosis, that is, the fourth standardized statistical moment
q of u. For
a normal distribution, γu is equal to three that results in having σu02 = 2 2(n−1)
n2
u 02 .
Both mean and variance reach statistical convergence with a dependence to the square
root of n; thus doubling the number of elements employed for the calculation of the
statistics allows us to reduce the convergence error by approximately 30%. For what
concerns the number n:
• For the estimation of the standard error of the mean-flow field, when averaging over
time, n is equal to the number of snapshots ns .
• For the estimation of the standard error of the temporal correlation matrix, n is the
number of element of a column of Y , nx .
In fact, the elements of the diagonal of the correlation matrix contain the variance of a
certain column of Y, which is evaluated along all this elements. Considering aver-
age values of the turbulent intensity below 10%, this results in an uncertainty of
approximately 1% of the mean velocity and in a 1% of relative uncertainty for the

h d i 0 0 9 8 088962 4 0 4 bli h d li b C b id i i
 186 Data-Driven Fluid Mechanics Chapter 9

elements of the covariance matrix if a total of 100 independent snapshots of a size of

100 × 100 elements is employed. The trace of the matrix Λ contains the cumulative
variance of the flow field along the entire data set; thus, its uncertainty should be
√
further reduced of a factor ns ; however, it has to be remarked that the variance
content associated to any mode is a fraction of the trace of lambda and that modes
are ordered by decreasing variance content. The uncertainty of the variance associated
to each mode thus affects in a different way the modes depending on their larger or
lower energy content. It is intuitive that flow field with a sparse eigenspectrum, that is,
with few modes containing a large amount of the variance, enables a modal analysis
with a very low relative uncertainty even with a small number of snapshots while the
uncertainty increases significantly for more spectrally rich data sets.
Regarding the question, while acquiring a data set, what should be the acquisition
frequency and how many periods should be acquired? it is a bit more complicated to
provide a rigorous answer. In case of a purely periodic phenomenon, such as the wake
of a cylinder at a relatively low Reynolds number, exactly sampling an integer number
of periods with any sampling frequency should provide a statistically converged modal
analysis; statistically converged modes should be also obtained sampling one shedding
period at frequency which is an integer multiple of the shedding frequency. The
application of such a criterion, however, is not possible for the majority of the flows,
especially when the information about the spectral content of the flow is not known a
priori.
A more general derivation about the required number of samples (and thus the
required sampling time in case of time-resolved data sets) can be obtained considering
the effect of the data autocorrelation on the standard errors. The analysis that leads to
(9.5) is based on the assumption that the snapshots used for the estimation of the mean
and of the correlation matrix are statistically independent. However, when dealing
with fluid-flow data sets, sometimes the data sequences are time resolved or spatial
information is somehow correlated (consider the typical case of PIV interrogation
windows with overlap, which will have a high degree of correlation, typically up to
75%). Under these conditions of correlated data, for a moderate to large data set, the
standard errors in (9.5) are larger, of a factor f (Bence 1995):
s
1+ρ
f = ,
1−ρ
(9.6)
i=2 ui−1 · ui
Ín
where ρ = Ín 2 .
i=2 ui

The autocorrelation coefficient of the data ρ is an indicator of how much the number
of samples should be increased in order to attain statistical convergence, for example,
for time-resolved fields in which ρ is likely to be of the order of 0.8, a nine-time-
larger number of samples is required to obtain the same statistical convergence as if
the samples were statistically independent. This last statement is equivalent to saying
that, if we observe a convective flow with a certain convective velocity, to obtain a
good convergence we should make sure that the total sampling time is at least one

h d i 0 0 9 8 088962 4 0 4 bli h d li b C b id i i
 Ianiro: Good Practice and Applications 187

order of magnitude larger than the time to convect from one measurement point to
another (in order to obtain at least 100 independent samples for each measurement
point). However, due to spatial correlation of the observed flow features this sampling
time should be, in principle, even larger and should be one or two orders of magnitude
larger than the characteristic convective time of the largest flow features analyzed.

9.2.2 The Effect of Noise

When dealing with experimental data, measurements are affected by errors, typically
classified into bias and random error. Considering the case of PIV, while the bias
error is mainly responsible for a signal attenuation (in particular, PIV interrogation
windows behave as a low-pass filter) random noise affects almost in the same way low-
frequency and high-frequency features in a certain measured flow (velocity, pressure,
or temperature) field. With a special focus on PIV, several works in the literature have
attempted to quantify the effect of noise on the modes, and notable examples are
certainly the works by Venturi (2006), Raiola et al. (2015), and Epps and Krivitzky
(2019).
Raiola et al. (2015) provide a simple yet useful estimation criterion to quantify
which POD modes are mostly affected by noise and to choose a suitable number of
modes to use in an optimal low-order reconstruction. The analysis is based on the
assumption that random error is Gaussian and the snapshot matrix containing the
measured field Y is the sum of the “true” snapshot matrix Ỹ without noise plus the
error snapshot matrix E, that is, Y = Ỹ + E. From this assumption, it is possible to
build the snapshot covariance matrix Y ∗ Y as the sum of the four terms:
Y ∗ Y = Ỹ ∗ Ỹ + E ∗ Ỹ + Ỹ ∗ E + E ∗ E. (9.7)
Assuming that the noise has a Gaussian distribution with standard deviation σe and
is identically distributed among the snapshots, it is possible to approximate the noise
covariance matrix as diagonal matrix with the diagonal elements equal to the square of
the noise standard deviation times the number of the elements of the snapshots (Huang
et al. 2005), that is, E ∗ E = nx σe2 I.2 It is also reasonable (although not always true) to
assume that the noise and the true field are not correlated, so that E ∗ Ỹ + Ỹ ∗ E can be
neglected in (9.7). This yields
Y ∗ Y ≈ Ỹ ∗ Ỹ + nx σe2 I . (9.8)
According to Marčenko and Pastur (1967), this approximation is still valid when
the matrix Y is rectangular, even for a non-Gaussian error that is independent on the
flow field and is identically distributed among the snapshots so that all the eigenvalues
of E ∗ E are in a neighborhood of nx σe2 . This last assumption is equivalent to the
assumption that the eigenvalues of the covariance matrix of the random error have
a constant spectra distribution.
rÍ
2
nx
e2
Note that according to (9.4) the unbiased estimation of σe ≈ n x −1 . However, when dealing with
i=1 i

the total energy corresponding to a certain eigenvector, this scales with n x .

h d i 0 0 9 8 088962 4 0 4 bli h d li b C b id i i
 188 Data-Driven Fluid Mechanics Chapter 9

Comparing the solution of the eigenvalue problem of the measured and true
covariance matrices, it is possible to write that
 
Ψ∗ Y ∗ Y = Ψ∗ Ỹ ∗ Ỹ + nx σe2 I = Λ̃Ψ∗,
(9.9)
∗ ∗
Ψ̃ Ỹ ∗ Ỹ = Λ̃Ψ̃ .
If the random-error eigenvalues are small with respect to the difference between two
successive eigenvalues λ̃i and λ̃i+1 of Ỹ ∗ Ỹ , it is possible to write that λi = λ̃i + nx σe2
and that the ith temporal mode of Y is approximately equal to the ith temporal mode
of Ỹ (Venturi 2006). The perturbation of the eigenvectors increases with an increasing
mode number as the variance content of the ith mode λ̃i approaches the value of nx σe2 .
These relationships can be accurately derived from matrix perturbation theory, along
with their bounds, and are a common assumption in perturbed principal component
analysis (PCA) applications (Huang et al. 2005). However, such a simplified descrip-
tion can be useful and is certainly sufficient to draw some conclusions and identify
some practical insights:

• If a data set can be represented with a small number of modes, it is possible to

obtain very high signal-to-noise ratios employing a low-order reconstruction with
a reduced number of modes, that is, Yk = ΦΣk Ψ∗ . Σk is the singular value matrix
keeping only the first k elements of the diagonal. Consider, for instance, a data
set of 100 snapshots in which the standard deviation of the noise is 1% of the
reference velocity while the turbulent fluctuations are 10% and assuming that
90% of the turbulent fluctuations are reconstructed with 10 modes. A low-order
reconstruction with only 10 modes allows us to reconstruct fluctuating fields with
a signal-to-noise ratio of 90, clearly larger than the initial SNR = 10.
• Since the eigenvalues of the covariance matrix are representative of the total
variance in the data set, they will increase linearly with an increasing number
of samples, assuming that the flow is statistically stationary. The measurement
noise, instead, spreads almost uniformly among all the modes keeping a constant
contribution equal to nx σe2 . It is thus possible to increase the ratio n λ̃σi 2 and
x e
improve the quality of the modal decomposition simply increasing the number
of snapshots employed. This result is very appealing since this relation is linear!
• While it is possible to obtain a theoretical demonstration of the existence of an
optimum number of modes (Raiola et al. 2015), often the noise variance is
unknown and cannot be determined a priori without previous knowledge of
the spectral content of the noise-free data set. The optimality property of POD
ensures that the most energetically relevant features are contained in the first
modes and that the energy content rapidly decays along the eigenspectrum.
As discussed here, the contribution of the noise should instead be practically
constant. After a certain mode number, thus, the variance content will remain
almost constant, being mainly ascribed to noise. This idea returns an operative
criterion, inspired by the scree plot test (Cattell 1966), which consists in looking
for an elbow in the plot of the cumulative variance content versus the mode
number.

h d i 0 0 9 8 088962 4 0 4 bli h d li b C b id i i
 Ianiro: Good Practice and Applications 189

9.2.3 An Example: Noisy Measurements in a Turbulent Channel

The effect of noise on POD modes, along with the choice of a suitable number of
modes to obtain an optimal low-order reconstruction, was assessed against noisy fields
obtained from a turbulent channel flow (Raiola et al. 2015). A total of 8 000 turbulent
fields were extracted from the Johns Hopkins turbulence database (Li et al. 2008) in
streamwise/wall-normal planes with a size of h × h with h being the channel half
height. These fields were employed to generate synthetic PIV particle image couples,
which were then processed to obtain “measured” flow fields as in a PIV experiment.
For this data set, PIV images allowed us to obtain a known level of random noise
while the amount of snapshots and measurement points were sufficiently large to
ensure statistical convergence. It was shown that employing the scree plot criterion
would allow us to find an optimum number of modes maximizing the true signal and
removing most of the noise. The reconstruction error was defined as the Frobenius
norm of the difference between the low-order reconstruction of the data set Yi with
respect to the measured data set Y . The scree plot method was employed looking
for F ≈ 1 with F being the ratio of the backward and forward derivative of the
reconstruction error versus the number of modes; that is,
1
δRM (i) = ||Y − Yi ||F ,
nx
(9.10)
δRM (i + 1) − δRM (i)
F(i) = .
δRM (i) − δRM (i − 1)
Figure 9.1 reports the evolution of the reconstruction error and of the parameter
F versus the number of modes and shows that for F > 0.999 the reconstruction
error with respect to the true fields δRT (i) = n1x ||Ỹ − Yi ||F is minimum, thus this
criterion can be employed to identify an optimal number of modes k F to mitigate noise
contamination of PIV data. This result is further confirmed from a visual inspection of
Figure 9.2, which shows that two vorticity patterns of similar magnitude are measured
by PIV, one present in the real flow field and the other one being an artifact. The
low-order reconstruction with the optimal number of modes conserves the true vortex
while the artifact is removed.

9.3 Extracting Phase Information

When dealing with flow-field measurements, experimental data sets are often sampled
at frame rates that are not sufficient to correctly describe the dynamics of the problem
under study. However, even when it is not possible to actually measure the frequency
of the flow unsteady phenomena, the POD temporal modes allow us to determine
relevant phase information. In case of phenomena characterized by dominant periodic
features (such as the shedding in the wake of a bluff body), these periodic features
can be usually represented with a compact subset of modes. In fact, a plot of the
correlation-matrix eigenvalues λi versus mode number usually shows a quite clear
spectral separation between first modes, accounting for the dominant periodic features,

h d i 0 0 9 8 088962 4 0 4 bli h d li b C b id i i
 190 Data-Driven Fluid Mechanics Chapter 9

Figure 9.1 Reconstruction errors with respect to true and measured fields (δRT and δRM ) (left
axis) and F (right axis) versus the number of modes used in the reconstruction. (From Raiola,
Discetti & Ianiro (2015), reproduced by permission from Springer Nature.)

and the following modes that describe the smaller-scale turbulent features. This
separation in the eigenspectra can be exploited to effectively extract periodic flow
features (Perrin et al. 2007) as detailed in the following.
Considering a temporally homogeneous flow or a temporally periodic phenomenon
(with period τ), the temporal correlation matrix is a function of solely the temporal
separation between two snapshots, that is, the temporal correlation between the time
t and the time t 0 is only a function of t − t 0, that is, Rt (t, t 0) = Rt (t − t 0). The
eigenfunctions of the temporal correlation matrix must thus be Fourier modes as
discussed in Chapter 6. In case of a shedding-dominated phenomenon in which
the most energetic flow features are characterized by a statistical periodicity with
a dominant frequency, it is safe to assume that the first POD modes, apart from
being orthogonal, also align to a Fourier decomposition of the field and show a
strong harmonic relation. Therefore, the POD temporal modes with larger eigenvalues,
obtained solving the eigenvalue problem of the temporal correlation matrix, can unveil
their relative phase information.
If we consider cases such as that of the vortices shed in the wake of a bluff body
or the vortices developed due to shear-layer instabilities in a jet, a traveling wave
is described by two high-energy modes (which are often found to be the first two
modes). Both modes have to share the same periodicity, that is, the shedding period
τ, thus, according to the orthogonality of the POD temporal modes, it is possible to
assume that the low-order reconstruction employing only the first two POD modes
would coincide with the decomposition (Raiola et al. 2016)

a2 (t) = y0 + ψ1 (ϑ)σ1 φ1 + ψ2 (ϑ)σ2 φ2, (9.11)

√
where σi = λi and ϑ = 2πt τ is the period phase. The functions ψ1 (ϑ) and ψ2 (ϑ) will
likely be two sinusoidal functions and must be in phase quadrature, that is, ψ1 ∝ sin(ϑ)
and ψ2 ∝ sin(ϑ + π/2). The scatter plot of the temporal modes ψ1 and ψ2 distributes

h d i 0 0 9 8 088962 4 0 4 bli h d li b C b id i i
 Ianiro: Good Practice and Applications 191

Figure 9.2 Instantaneous fluctuating vorticity (ωz ) field. (a) The DNS field used for this
benchmark. Magnified view of the: (b) DNS field, (c) measured field, (d) field reconstructed
with 1300 modes, (e) field reconstructed with 3000 modes. (From Raiola, Discetti & Ianiro
(2015), reproduced by permission from Springer Nature.)

in the neighborhood of a circle if ψ1 (ϑ) and ψ2 (ϑ) are sinusoidal functions. Since
temporal-mode vectors have a unitary norm and zero mean, the scatter plot will be in
the neighborhood of a goniometric circle (with radius 1) if the temporal modes are
multiplied by the factor ns /2.
p

In general, the scatter plot of the time coefficients might unlock information on the
phase and frequency relation between the first and other modes, often corresponding
to higher-order harmonics, thus shedding light on the interconnection between the

h d i 0 0 9 8 088962 4 0 4 bli h d li b C b id i i
 192 Data-Driven Fluid Mechanics Chapter 9

different flow features highlighted by the modal analysis. Assuming that the ith POD
mode is harmonically related and phase shifted with respect to the first mode, it is
possible to write that
p
ns /2ψi (ϑ) = sin (αi ϑ + δi ) , (9.12)

where αi is a positive integer and δi is the phase shift of the ith mode with respect to
the first mode.
In order to ascertain these frequency relations, the scatter plot of the POD temporal
modes ψ1 and ψi should be observed in search of Lissajous curves. A Lissajous curve
is the graph of a system of parametric equations of the type x = A sin(αt + δ) and
y = B sin(βt), which can describe even very complex harmonic motions. Visually, the
ratio α/β determines the number of ”lobes” of the figure. The procedure to identify the
harmonic relation and the phase shift for higher order modes relies on the simplifying
assumption that the first two modes are at the same frequency with a π/2 phase shift.
It is possible to extract the period phase from the time coefficients of the first two
modes, that is, for any snapshot tan(ϑ) = ψ1 /ψ2 . Subsequently, the positive integer αi
and the phase shift δi , which characterize the harmonic relation, can be extracted from
the solution of the optimization problem
p 
argmin ns /2 ψi − sin (αi ϑ + δi ) , (9.13)
αi ∈N
δi ∈R

where αi and the phase shift δi are the free parameters to identify. An example of
this procedure is given in the following, where it is applied to the analysis of the flow
features in the wake of two tandem cylinders located near a wall.

9.3.1 An Example: The Wake of Two Cylinders in Ground Effect

The work by Raiola et al. (2016) reports a PIV wind tunnel study in the wake of
cylinders near a wall. The experimental arrangement is sketched in Figure 9.3. The
cylinders had a diameter of 32 mm. The acquisition was performed with a 2Mpixels
camera (1600 × 1200 pixels array) with a spatial resolution of about 7.2 pixels/mm
while the sampling frequency of the PIV measurements was 10 Hz and the Reynolds
number was equal to 4.9 × 103 , thus the sampling frequency was not sufficient to
sample the characteristic frequencies involved in the problem. An ensemble of 2 000
image couples was acquired for each experiment, obtaining a total of 2 000 flow fields.
The interrogation strategy was an iterative multistep image deformation algorithm,
with final interrogation windows of 16 × 16 pixels with 50% overlap (i.e., final vector
spacing of 8 pixels, corresponding to 0.035D). The ensemble size both in term of
snapshots and number of points enables a sufficient convergence of the statistics
(errors well below 1% of the variance both in time and space).
The results of the modal analysis are reported herein for the case of two cylinders
with a relative distance L = 1.5D and a gap from the wall G = 1D. Under these
conditions the vortex shedding for the upstream cylinder is completely suppressed,
and the two cylinders act as a single bluff body (Zdravkovich 1987). At the same
time, the ground affects the pressure distributions, as well as the flow development.

h d i 0 0 9 8 088962 4 0 4 bli h d li b C b id i i
 Ianiro: Good Practice and Applications 193

As reported in Figure 9.4, the first two modes account for almost
p 60% of the variance,
and the scatter plots of the temporal modes (multiplied times ns /2) return a perfect
goniometric circle confirming that these two modes are representative of a traveling
pattern. The asymmetry of the first POD mode highlights the strong interaction of the
wall boundary layer with the shedding in the wake, evident from the presence of near-
wall vorticity visualized in the first (and second mode, not reported here for brevity).
Higher-order modes are mostly characterized by intermittent release of vorticity at a
spatial and temporal frequency that is a multiple of the principal shedding frequency
as shown from the contour plot of the fourth spatial mode and outlined from the
Lissajous curves in the scatter plot of the fourth modes against the first one. This
is an asymmetric mode that models the cross-wise oscillation of the shedding wake.

Figure 9.3 Experimental setup for PIV measurements in the wake of two tandem cylinders in
ground effect. (Reprinted from Raiola, Ianiro & Discetti (2016), copyright 2016, with
permission from Elsevier.)

9.4 Extended Proper Orthogonal Decomposition

Until this point, this chapter has been centered on the analysis of a given data set
A. However, the availability of multiple data sets, for example, A and B, are rather
frequent. Subtracting the base conditions y0,A to A and y0,B to B, it is possible to
obtain YA and YB . Given these two data sets, the most straightforward task would
be to perform two separate modal analyses, as for instance performed by Mallor et al.
(2019). However, if the measured quantities are synchronized, it is tempting to perform
a combined analysis to extract information not only about the modes of a certain data
set, for example, A, but also about the features of the second data set (B) that correlates
with these modes.
To this purpose, the reader must remember that the POD spatial modes can be
obtained from a projection of the snapshot matrix onto the temporal-mode matrix as
in (9.3). This projection approach can be used to extend the POD to other quantities,
in a generalized approach named Extended POD (Borée 2003). The columns of Ψ A ,
that is, the POD temporal modes, form a basis in the Rns vector space; thus, it is
possible to use Ψ A as a basis for the projection also of the matrix YB , provided that

h d i 0 0 9 8 088962 4 0 4 bli h d li b C b id i i
 194 Data-Driven Fluid Mechanics Chapter 9

Figure 9.4 Modal analysis results for two tandem cylinders with L/D = 1.5 and G/D = 1. On
the top, the correlation matrix eigenvalues are reported along with the scatter plots of the first
temporal mode against the second and the fourth modes (temporal modes are multiplied by
ns /2). At the bottom of the figure, the spatial modes 1 and 4 are reported, highlighting the
p

fact that the first and the fourth modes have an even spatial frequency ratio with a phase shift
of π/4. (Adapted from Raiola, Ianiro & Discetti (2016), copyright 2016, with permission from
Elsevier.)

the available number of snapshots of A and B is the same. Consequently, while in

(9.3) the projection was used to estimate the spatial modes Φ A of A, it is possible to
define the EPOD modes of B as

YB = ΦB A Σ B A Ψ∗A → ΦB A Σ B A = YB Ψ A, (9.14)

where the subscripts A and B refer to the quantities A and B, respectively. The columns
of ΦB A are thus the spatial modes of YB obtained from the projection on the temporal
modes of YA . The data sets A and B have to be captured/generated in the same time
reference frame; however, they might represent different physical quantities, and the
snapshots of A and B can have different numbers of elements.
While the POD temporal modes ψ Ai are estimated from the solution of the
eigenvalue problem of the correlation matrix YA∗ YA and are optimal for YA , the
extended spatial modes φ B A,i are not optimal nor are necessarily ordered by their
variance content σB2 i . The value of σB2 i that can be computed as the square root of the
variance of the columns of ΦB A Σ B A in order to obtain φ B A,i vectors of unitary norm.
It has to be remarked that the projection of the snapshot matrix B on the temporal
mode matrix Ψ A is energy preserving since Ψ A is a basis in the Rns vector space.
Due to the orthogonality of the temporal modes ψ Ai this has the consequence that
EPOD decomposition, although not being optimal is orthogonal. The orthogonality
principle has the advantage to guarantee that each EPOD mode σB A,i φ B A,i contains

h d i 0 0 9 8 088962 4 0 4 bli h d li b C b id i i
 Ianiro: Good Practice and Applications 195

only the part of B that is correlated with the contribution of the ith POD mode to
A. This concept is closely connected with the idea of the linear stochastic estimation
(LSE), first developed by Adrian (1975), a technique that is typically employed to
estimate an unknown quantity given a known quantity. As the name suggests, this
technique attempts to statistically draw a linear relation between a known quantity
and another quantity that has to be estimated. The EPOD can be considered as
the LSE of the spatial modes σB A, i φ B A,i , given the spatial modes φ Ai . All the
properties of the LSE can therefore be extended to the EPOD. In particular, it must
be noted that, while the relation between velocity fields and other quantities may
be formally nonlinear, the LSE may still prove to be adequate due to the small
magnitude of second-order terms, as shown by Adrian et al. (1989) for the case
of homogeneous turbulence. Several attempts to determinate the velocity modes
correlated to the POD modes of other quantities have been reported in the literature.
Two examples are reported in the following: the first making use of the extended POD
to reconstruct time-resolved fields from time resolved hot-wire measurements and
non-time-resolved PIV measurements and the second one that analyzes the thermal
transport in a pipe with nonhomogeneous heating.

9.4.1 An Application: Time-Resolved Flow Sensing with the Extended

Proper Orthogonal Decomposition
In the works by Discetti et al. (2018) and Discetti et al. (2019), we have employed
a combination of non-time-resolved field measurements and time-resolved point
measurements for the estimation of time-resolved turbulent fields, as a reduced-order
reconstruction with a limited subset of modes. The theoretical concepts behind this
approach have been developed through a vast literature (Ewing & Citriniti 1999,
Tinney et al. 2008, Baars & Tinney 2014, Hosseini et al. 2015, Discetti et al.
2018, 2019) and essentially grounded in the fact that EPOD can be employed for
linear stochastic estimation, that is, we can estimate the mode time coefficients
of a certain quantity with the known time coefficients of the modes of another
synchronized quantity.
Consider the two synchronized data sets: one at low repetition rate with PIV and
the other one at a high repetition rate with hot-wire anemometry. Due to the different
repetition rates, the number of time samples for the hot wire is much larger than that
for PIV, although fewer points are available in space. Employing only the time instants
at which both PIV and hot-wire measurements are available, it is possible to build two
matrices with synchronized snapshots: one named YP (where P stands for PIV) and
another named YH (where H stands for hot wire). The eigenvalue decomposition of
the correlation matrices of both quantities allows us to identify the matrices of the
temporal modes Ψ P and Ψ H . The product Ψ∗P Ψ H returns a new matrix Ξ = Ψ∗P Ψ H
that contains in its element i, j the correlation between the ith mode of the first data
set (PIV) and the jth mode of the second data set (HW).
The correlation information contained in the matrix Ξ can be used to estimate the
PIV flow fields at those time instants when hot-wire measurements are available, even
if PIV was not sampling. At any time instant t, the availability of a hot-wire snapshot

h d i 0 0 9 8 088962 4 0 4 bli h d li b C b id i i
 196 Data-Driven Fluid Mechanics Chapter 9

y H (t) allows us to estimate the vector containing the time coefficients of the hot-wire
spatial modes, ψ H (t), through a simple projection:

ψ H (t) = Σ−1 ∗
H Φ H y H (t). (9.15)

Given ψ H (t), it is possible to estimate the time coefficients of the PIV modes. The
vector ψ P (t)+ , containing the estimated time coefficients of all the PIV spatial modes,
is obtained following (9.16), and the PIV snapshot can be reconstructed as in (9.17):

ψ P+ (t) = Ξψ H (t), (9.16)

+
yP (t) = Φ P Σ P ψ P+ (t). (9.17)

The reader might notice that the success of such a procedure would require
a significant number of hot-wire probes in order to avoid a significant difference
between the dimensionality of PIV and hot-wire snapshot matrices. Even if some
works in the literature have employed large amounts of probes (Kerhervé et al. 2017),
this is not easily available. The most common approach passes through the definition
of pseudo-snapshots a Hi in which several measurements are taken at the same
“instant” as for PIV adding “virtual” probes. Virtual probes are obtained using time-
shifted probe data with the support of the Taylor hypothesis of uniform convection,
converting the hot-wire temporal information into spatial information.
It is worth noticing that the columns of both Ψ P and Ψ H are orthonormal vectors
and form two bases in the Rns vector space; consequently Ξ is also composed by
columns forming a basis. This implies that the process is energy preserving and that
the higher-order modes that might be mutually uncorrelated are taken into account
for the estimation in (9.16). Since all rows/columns of Ξ have unitary norm, if a
certain ith probe mode ( jth field mode) is uncorrelated with all the field modes (the
snapshot modes), the ith row ( jth column) of Ξ has to be composed of randomly
distributed elements, with unitary norm and zero mean, thus standard deviation equal
√
to 1/ nt . This reasoning allows us to filter the matrix Ξ removing all the elements
√
|Ξi, j | < 1/ nt (Discetti et al. 2018). This approach has allowed us to obtain a
reconstruction of the behavior of large-scale and very-large-scale motions high-Re
turbulent flows. An example of experiment is reported in Figure 9.5 in which PIV
measurments are synchronized with five hot-wire probes in the large-scale pipe-flow
facility CICLoPE (Univ. of Bologna, Italy). Figure 9.5, right-hand side, reports a
comparison between the PIV estimated fields and the hot-wire measurements that
shows a significant coherence and low noise level, despite the significant spectral
richness of the flow (Discetti et al. 2019).
As a further extension of this work, aiming at the identification of a suitable sensing
tool for the detection of large-scale and very-large-scale motions in a turbulent wall-
bounded flow, we have recently explored the performances of convolutional neural
networks (Güemes et al. 2019, Guastoni et al. 2020), obtaining better results than
EPOD. This highlights that machine learning tools have the capability to enhance and
eventually replace traditional data analysis tools in the field of fluid mechanics.

h d i 0 0 9 8 088962 4 0 4 bli h d li b C b id i i
 Ianiro: Good Practice and Applications 197

Figure 9.5 Left: Image of the experimental arrangement with PIV image plane and hot-wire
rake located downstream. Right: Contour of the evolution of the streamwise velocity
component obtained using Taylor’s hypothesis, Reτ = 9500. Top: hot-wire data. Bottom:
Estimated fields. The velocity contours are shown in inner units, that is, u+ = u/uτ . (Reprinted
from Discetti et al. (2019), copyright 2019, with permission from Elsevier.)

9.4.2 An Exercise: The Turbulent Heat Transfer in a Pipe Flow

In the work by Antoranz et al. (2018), we analyzed the thermal transport in a turbulent
pipe flow with nonhomogeneous heating. This configuration is representative of the
receivers of thermal solar power systems in which mirrors concentrate solar radiation
toward a receiver made of pipes through which a molten nitrate salt is flowing. In this
application, it is of foremost importance that the received thermal input is convected
along the whole pipe section in order to avoid that uneven temperature distributions
cause thermal stresses in the pipe and/or chemical reactions in the molten nitrate salts
employed as working fluid.
The data set analyzed in the paper was generated with direct numerical simulations
of a fully developed turbulent pipe flow, with circumferentially varying heat flux
boundary conditions. Details about the simulations are available in the works by
Antoranz et al. (2015) and Antoranz et al. (2018). The boundary conditions at the
pipe wall were nonslip conditions for the velocity while the heat flux had a sinusoidal
distribution for azimuthal cohordinate θ between 0 and π and was set to zero for
π < θ < 2π. Exploiting the statistical homogeneity of the fully developed pipe flow, a
total of 4 500 snaphsots were extracted taking radial planes at 50 streamwise locations
of the pipe at 90 time instants. Available quantities were three velocity components
(u, v, w), and the temperature T. In the following, the case at Reτ = 180 and Pr = 0.7
is analyzed.
Snapshot POD is performed on the temperature snapshot matrix, and extended
velocity modes (projected on the temperature temporal basis ΨT ) are estimated.
Figure 9.6, on the top, reports the normalized variance content of the temperature
modes and of the velocity extended modes. Although it should be clear to the reader
that the POD temperature modes are optimal for the reconstruction of the temperature
variance, it is still surprising that the majority of the thermal variance (and thus of
the turbulent thermal transport) is correlated only with a little amount of the velocity

h d i 0 0 9 8 088962 4 0 4 bli h d li b C b id i i
 198 Data-Driven Fluid Mechanics Chapter 9

variance. Looking at the cumulative sum of the variance, 80% of the temperature
variance is correlated with only 40% of the velocity variance. The fact that the
turbulent thermal transport is ascribed only to a limited amount of the velocity variance
suggests possible approaches for heat transfer enhancement, such as the use of vortex
generators able to produce a flow pattern similar to the first extended velocity mode
(see Figure 9.6, bottom) to improve the temperature uniformity in the pipe section.

Figure 9.6 Results of the EPOD analysis in a turbulent pipe with nonhomogeneous heat flux.
Top: Variance content (cumulative variance content in the inset) of temperature modes (blue)
and velocity extended modes (red). Bottom left: First temperature mode. Bottom Right: First
extended velocity mode with contour of the streamwise velocity component and in-plane
velocity vectors.

As an exercise, a few lines of the MATLAB R

code used to generate the results
presented in Figure 9.6 are provided. For the present data set, the data from the
numerical simulation are not available on a uniform Cartesian mesh, hence requiring
us to take into account the different flow areas corresponding to each point of the
snapshot matrix. Since the spatial mesh is not spatially uniform in the x−y plane, the
temporal correlation matrix of a generic quantity g should be computed taking into
account the weights corresponding to the different areas of all grid points (see also
Chapter 6 on the weighted SVD).
A data set named case1.m is provided, containing temperature and velocity
snapshot matrices, along with geometrical information about the grid points of the

h d i 0 0 9 8 088962 4 0 4 bli h d li b C b id i i
 Ianiro: Good Practice and Applications 199

snapshots. A total amount of ns = 2275 snapshots is available with a mesh made of

nr = 69 radial locations r and nθ = 100 equally spaced tangential locations. It has
to be noted, here, that a limited number of snapshots is provided, in order to enable
quick calculations on the laptops of the students attending the lecture series; however,
the turbulent pipe flow is characterized by a rich spectrum that causes a relatively poor
convergence of the modal analysis for higher mode numbers.
function C = corr_matrix_polar (g , r )
nr = size (g ,1) ;
ntheta = size (g ,2) ;
ns = size (g ,3) ;
% weights calculation
% The first point is in the center of the domain and
,→ has
% area dr * dr
dr = diff ( r ) ;
weight = 0* r ;
weight (1) = dr (1) /2;
weight (2: nr -1) = ( dr (1: nr -2) + dr (2: nr -1) ) /2;
weight ( nr ) = dr ( end ) /2;
temp = g .* repmat ((2* pi / ntheta * r .* weight ) .^0.5 ,[1 , nt ,
,→ ns ]) ;
% All the points have area r * dr *2* pi / ntheta .
temp = reshape ( temp , nr * ntheta , ns ) ;
% Area weighted correlation matrix
C = temp ’* temp / pi ;
% Correlation matrix has to be divided by
% the total area pi * R ^2 ( here R =1) .

Once the correlation matrix has been computed, it is possible to estimate the
temperature modes and the extended velocity modes for u, v, and w based on the
temperature temporal modes ΨT :
% building snapshot matrices
nzplanes = size (T ,3) ;
T = reshape (T , nr * ntheta , ns ) ;
u = reshape (u , nr * ntheta , ns ) ;
v = reshape (v , nr * ntheta , ns ) ;
w = reshape (w , nr * ntheta , ns ) ;
% SVD of the temperature covariance matrix
[ Psi_T lmd_T Psistar_T ]= svd ( C_T ) ;
% Calculating spatial modes
SigmaPhi_T =( T ) * Psi_T ;
SigmaPhi_x_T =( u ) * Psi_T ;
SigmaPhi_y_T =( v ) * Psi_T ;
SigmaPhi_z_T =( w ) * Psi_T ;

To calculate the variance content σi2 of each mode φi , it is again necessary to take
into account the different flow areas corresponding to each point:
function sigma = get_sigma ( sigmaphi ,r , nr , nt )
sigmaphi = reshape ( sigmaphi , nr , ntheta ) ;
dr = diff ( r ) ;
weight = 0* r ;
weight (1) = dr (1) /2;
weight (2: nr -1) = ( dr (1: nr -2) + dr (2: nr -1) ) /2;

h d i 0 0 9 8 088962 4 0 4 bli h d li b C b id i i
 200 Data-Driven Fluid Mechanics Chapter 9

weight ( nr ) = dr ( end ) /2;

temp = sigmaphi .* repmat ((2* pi / ntheta * r .* weight )
,→ .^0.5 ,[1 , nt ]) ;
temp = reshape ( temp , nr * ntheta ,1) ;
sigma = ( temp ’* temp / pi ) ^0.5;

h d i 0 0 9 8 088962 4 0 4 bli h d li b C b id i i
 Part IV
Dynamical Systems

h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
 10 Linear Dynamical Systems
and Control
S. Dawson

This chapter introduces a suite of ideas and tools from linear control theory, which
can be used to modify the behavior of linear dynamical systems using feedback
control. We use a simple example problem to motivate these methods and explicitly
demonstrate how they can be applied. The system that we focus on is a linearization
of unstable fluid flow over a bluff body. This system has already been discussed in
Chapter 1, and here we start with the same data set used in Chapter 6 as an example
for computing the proper orthogonal decomposition (POD). We start with a brief
review of linear systems in the time and frequency domain before introducing the
example system and embarking on a journey through different approaches to control
it. In particular, we introduce proportional, integral, and derivative (PID) control, pole
placement, and full-state linear quadratic regulator (LQR) optimal control. Due to the
breadth of topics covered, which might typically take one or several semester-long
courses to cover, we will not be able to provide a full theoretical background behind
each method but will focus on the main ideas and the practical implementation of these
methods. We also briefly discuss further extensions and a number of applications of
linear control theory.

10.1 Linear Systems

In many applications in science and engineering, we are interested in analyzing the

properties of systems that evolve in time. These temporal dynamics are often due both
to the internal dynamics of the system itself, and to the effect of external inputs on the
system, which might be known or unknown, controllable or uncontrollable, and either
beneficial or harmful to the desired behavior of the system.
The simplest dynamical system is one that is linear, with time evolution that can be
represented using a system of linear ordinary differential equations,

Û = Ax(t),
x(t) (10.1)

where x(t) ∈ Rn is a column vector consisting of each state in the system, x(t) Û is the
time derivative of this quantity, and A is an n × n matrix, which we assume to have
real entries. Recall that this system is asymptotically stable if the eigenvalues of A
(also referred to as the poles of the system) are located in the left half of the complex

h d i 0 0 9 8 088962 4 0 6 bli h d li b C b id i i
 204 Data-Driven Fluid Mechanics Chapter 10

plane. Accounting for inputs u(t) and outputs y(t), the full system can be specified by
matrices (A, B, C, D), with
Û = Ax(t) + Bu(t),
x(t)
(10.2)
y(t) = C x(t) + Du(t).
The state-space system in (10.2) can be expressed as a transfer function in the Laplace
(frequency) domain by
P(s) = C(sI − A)−1 B + D. (10.3)
Y(s)
P(s) can also be interpreted as the ratio U(s) ,
where U(s) and Y (s) are the Laplace
transforms of the input, u(t) and output, y(t), respectively.1 The eigenvalues of A in
(10.2) correspond to the values of s where the term (sI − A)−1 in (10.3) (which is
known as the resolvent) is not defined, due to (sI − A) being non-invertible. Further
discussion of the stability of linear systems, and their analysis through the resolvent,
can be found in Chapter 13. A linear system of the form given by (10.2) or (10.3) is
the “plant” that we will ultimately be seeking to control. We can represent this system
and its inputs and outputs schematically as

u xÛ = Ax + Bu y
y = C x + Du

The properties of linear systems in both the time and frequency domain are discussed
in further detail in Chapter 4.
To illustrate the various concepts and ideas that will be covered in this lecture,
we will focus on a specific example, coming from a relatively simple system in
fluid mechanics. We will look at two-dimensional flow over a circular cylinder at a
Reynolds number of 60, particularly focusing on the dynamics of the system as it
evolves in time away from its unstable equilibrium solution, and toward its vortex
shedding limit cycle. The dynamics of this system are discussed in further detail in
Chapter 1. The specific data set that we use is the same one utilized in Chapter 6 for
an example for computing the POD.
To obtain a linear model of the form indicated in (10.1), we take data from a short
region in time, identify the two leading POD modes in this region (see Chapter 6
for a detailed discussion of the POD), and then identify a linear model that captures
the dynamics of the evolution of the corresponding POD coefficients during this short
region of time. This model identification is done via the dynamic mode decomposition
algorithm, which is discussed in detail in Chapter 7. For additional details concerning
the implementation of this system identification method, as well as the various control
techniques that will be applied to this identified system, see the accompanying Python
code (available on the book’s website2 ). The leading two POD modes, along with
the evolution of their corresponding coefficients in time, are shown in Figures 10.1
1 In the case where there are multiple inputs or outputs, the components of G(s) are given by the ratios
between the Laplace transforms of each pair of inputs and outputs.
2 www.datadrivenfluidmechanics.com/dow nload/book/chapter10.zip

h d i 0 0 9 8 088962 4 0 6 bli h d li b C b id i i
 Dawson: Linear Dynamical Systems and Control 205

Figure 10.1 The first two

POD modes (transverse
velocity component) for flow
over a circular cylinder at a
Reynolds number of 60, on a
trajectory going between the
unstable equilibrium and
limit cycle of the system.

2.5 Figure 10.2 Evolution of the

first two POD mode
x1

0.0 coefficients, corresponding to

Data
Linear model
modes shown in Figure 10.1.
−2.5 Shown are both the actual
0 20 40 60 80 100 120 coeffiicents, and those
2.5
predicted by the identified
linear dynamics (10.4). Only
x2

0.0 the data to the left of the

vertical line is used for the
−2.5 identification of the linear
0 20 40 60 80 100 120
t
model.

and 10.2, respectively. As well as showing the true evolution of the POD coefficients
(which will be our system states x1 and x2 ), we also show the evolution predicted by
the identified linear model. The linear model accurately captures the true behavior over
a region in time substantially larger than that used for system identification, though
is unable to capture the eventual saturation of the coefficient amplitudes as the true
system converges toward the limit cycle. Note that the portion of the data used for
system identification is the same as that used to compute POD for the example in
Chapter 6. Using feedback control, we might hope to be able to keep the oscillations
seen in Figure 10.2 to small enough amplitudes such that the linear model remains
accurate, and that controllers designed using this linear model remain effective.
For this system, we obtain the linearized dynamics
0.0438 −0.7439
 
A= . (10.4)
0.7373 0.0527
This system has two unstable eigenvalues at 0.0483 ± 0.7406i, which is expected,
since the envelope of the POD coefficient amplitudes plotted in Figure 10.2 is growing

h d i 0 0 9 8 088962 4 0 6 bli h d li b C b id i i
 206 Data-Driven Fluid Mechanics Chapter 10

in time. The fact that these eigenvalues come as a complex conjugate pair is also
expected, since the dynamics of the system are oscillatory.
Note that this linear system is a simplification of the true dynamics not just
because we are using a linear model, but also because we are neglecting other states
of the system, corresponding to lower-energy, truncated POD modes. For example,
identifying a linear system with more states could model (at least in a linear sense) the
slow modification of the “mean” as the amplitude of the oscillations increases, as well
as dynamics with frequency content at harmonics of the primary oscillation frequency.
These dynamics are discussed in further detail in Chapter 14.
We will explore various choices for the B and C matrices in Sections 10.2–10.5, but
will typically choose to directly observe and control one of the system states. This also
represents a simplification of what measurement and control might look like for the
true system in practice, where localized sensors and actuators might be more feasible
than directly measuring or controlling an entire POD mode.

10.2 Proportional, Integral, and Derivative Feedback Control

Suppose now that we want to manipulate our system to achieve a desired output, y(t).
For the example described, perhaps we want to stabilize the unstable system, and thus
eventually drive all components of y(t) to zero. For an unstable system such as this,
this objective is impossible without the availability of real-time information about
the state of the system, and how it is responding to attempts at control. It would be
equivalent, for example, trying to balance an inverted pendulum without any visual or
tactile measurements.
What we want to do then, is to develop a means for utilizing knowledge of the
system output y(t) in order to determine appropriate system inputs to achieve the
desired control objective. This general concept is known as feedback control. For
example, we could link the output to the input using the following arrangement:

r xÛ = Ax + Bu y
Σ
− y = C x + Du

Here we refer to K as the controller, which could simply be a constant, or might have
its own internal dynamics.
In this section, we will assume that we have a single input to our system (so B has
one column), and a single output (so C has one row), and will always assume that there
is no direct feedthrough from the input to the output (so D = 0). It will be helpful to
look at our system in the frequency domain, which we can obtain from a state-space

h d i 0 0 9 8 088962 4 0 6 bli h d li b C b id i i
 Dawson: Linear Dynamical Systems and Control 207

Figure 10.3 Effect of

0.4 applying proportional
feedback control with K = 1
0.2 to the two systems, P1 and
P2 .
0.0
y(t)

−0.2 Uncontrolled
K = 1 feedback, P1 (s)
−0.4 K = 1 feedback, P2 (s)
Control applied
−0.6
0 10 20 30 40 50
t

system using (10.3). We will explore two different input and output combinations.
First, we have the case where the input uses a different state variable as the output
(system P1 ), using the input and output matrices
1
 
B1 = , C1 = 0 1 , (10.5)
 
0
giving the transfer function (see (10.3))
0.7373
P1 (s) = . (10.6)
s2
− 0.09654s + 0.5508
Second, we use the input and output matrices
0
 
B2 = , C2 = 0 1 , (10.7)
 
1
giving the transfer function
s − 0.04381
P2 (s) = . (10.8)
s2 − 0.09654s + 0.5508
Note that the denominators of the transfer functions P1 (s) and P2 (s) are the same
quadratic polynomial. This is as expected, since the roots of this polynomial are the
poles of the system, which are in turn given by the eigenvalues of A, which is the same
for both systems.
To begin with, and before doing any further analysis of this system, suppose that
we try to control our systems with the simplest feedback control, where K from our
schematic diagram being a multiplicative constant. This is known as proportional con-
trol. The results of applying this control to our system with K = 1 is shown in Figure
10.3, where we have the reference set at r = 0. Here, we allow the system to evolve
without any inputs, before turning on control at a time t = 15. Note that here we are
applying control to the unstable linearized system, rather than the full nonlinear sys-
tem. We notice that the system P2 seems to be stabilized, while P1 remains unstable.

h d i 0 0 9 8 088962 4 0 6 bli h d li b C b id i i
 208 Data-Driven Fluid Mechanics Chapter 10

How might we have predicted this observed behavior? With r = 0, the system input
is given by law u = −K y = −KC x. This means that the system with feedback evolves
according to
xÛ = (A − BKC)x. (10.9)
The stability of this system is determined by the eigenvalues of the matrix A − BKC.
In the frequency domain, it is easy to show that the transfer function of the controlled
system is given by
P
GC L = , (10.10)
1 + PK
where K is the transfer function for the controller in the frequency domain (which is
a constant for proportional control). Note that more generally, a closed-loop transfer
function is given by
GOL (s)
GC L = , (10.11)
1 + G F B (s)
where GOL denotes the open-loop transfer function between r and y, and G F B is the
transfer function going around the feedback loop. Going back to (10.10), here the
poles of the closed-loop system will be the solutions to
1 + PK = 0, (10.12)
which can be expressed as the roots of a characteristic polynomial. To study the
systems P1 and P2 in more detail, consider a more general system of the form
a1 s + a0
P(s) = 2 , (10.13)
s + b1 s + b0
for which P1 (s) and P2 (s) are both special cases. From (10.12), the closed-loop poles
of this system for control with a constant K can be found to be the solutions to the
polynomial
s2 + (b1 + a1 K)s + (b0 + a0 K) = 0. (10.14)
These solutions are given by
(b1 + K)2 − 4(b0 + a0 K)
p
b1 + a1 K
s=− ± . (10.15)
2 2
From this, we can start to understand why proportional control with K = 1 worked for
the system P2 , but not for P1 . The fact that P2 has a nonzero a1 term means that the
control is able to shift the real component of the poles by −a1 K/2, thus making them
more stable (assuming that a1 > 0). Note also that when the argument of the square
root is negative, this term will not affect the real component of the poles. Conversely,
if a1 = 0 as is the case with P1 , the control only affects the square root term in
(10.15). If the argument of the square root is negative, then this does not change
the real component, and thus the stability, of the system. On the other hand, if the
argument of the square root is positive, then one pole will become more stable, but the
other becomes more unstable as the feedback gain either increases or decreases. From
this, we can conclude that proportional feedback is unable to stabilize P1 .

h d i 0 0 9 8 088962 4 0 6 bli h d li b C b id i i
 Dawson: Linear Dynamical Systems and Control 209

As an aside, note that the dynamics of this system can be related to those of a spring-
mass/damper system, which also has a quadratic denominator in its transfer function.
In this analolgy, the negative coefficient of the linear term in the denominator of P1
and P2 corresponds to a negative damping, which is what leads to the system being
unstable. Furthermore, (10.14) shows that K can only affect the damping term when
a is nonzero, and otherwise only affects the stiffness-to-mass ratio, giving another
interpretation for why this proportional control cannot stabilize P1 .
What sort of feedback would work for P1 ? From (10.12), we see that we could
achieve the same behavior of the poles of the closed-loop system if we added a term
to our controller to mimic the first-order term in the numerator of P2 . For example, we
could let

K = a1 s + a0 , (10.16)

to achieve the same closed-loop pole behavior as we found with P2 using K = 1.

These systems are not exactly equivalent due to the zero being in the feedback part
of the loop, but the control results are very similar to those shown for P2 in Figure
10.3. While this control strategy results in a stable system, the fact that the zero in P2
is in the right half-plane is typically not desirable, as it can result in phenomena such
as the system response initially moving in the wrong direction from the desired state.
A stable system with zeros in the right half-plane is known as a non-minimum-phase
system.
More generally, control of the form given in (10.16) is called proportional-
derivative (or PD) control. Our original objective was to implement control to drive
the system output to zero (or to any other desired output signal). While it looks like
we are successful in doing this visually in Figure 10.3, the only way that this will be
exact in reality is for the denominator of the closed-loop transfer function to not have a
constant term. To achieve this, we could consider adding an “integrator” to our control
design. Generally, combining proportional (P), integral (I), and derivative (D) control
is called PID, as is a very commonly used classical control technique, which relies on
being able to choose appropriate gains in the control law

ki
K(s) = k p + + k p s. (10.17)
s

Note that adding an integrator can also destabilize the system, so care must be taken
when choosing the gains k p , ki , and k d used in PID control.
The analysis that we have performed in this section is tractable for this relatively
simple system, but quickly becomes unmanageable for more complex systems with
higher order numerators and denominators in their transfer functions. This motivates
the formulation of general rules to predict the effect of feedback controllers without
requiring us to explicitly solve for the roots of high-order polynomial equations. Such
rules will be discussed in Section 10.3.

h d i 0 0 9 8 088962 4 0 6 bli h d li b C b id i i
 210 Data-Driven Fluid Mechanics Chapter 10

10.3 The Root Locus Plot

This section introduces a tool that will allow us to predict in advance what the effect of
a given control strategy will be. In particular, it turns out that there are several simple
rules that can allow us to sketch where the poles and zeros go as the controller gain
increases. The root locus (developed by Evans (1948)) is a tool to see graphically how
the poles of a closed-loop system move as a parameter k (which most typically is a
controller gain) is varied. In particular, the root locus plot shows the locations of all
poles of a system as a function of k, where we can write the characteristic equation
for the poles of the closed-loop system as
1 + kG(s) = 0. (10.18)
In the case of proportional control, G(s) is simply the plant P(s) and k is the
proportional controller gain, though note that root locus plots can be constructed for
other forms of control, so long as the closed-loop poles can be expressed in the form
given by (10.18). Note in particular the similarity between (10.18) and (10.12). While
it is easy to make a root locus plot in MATLAB or Python, it is also possible to draw
these plots relatively accurately by hand, just by following a few simple rules (and
importantly, without needing to explicitly solve (10.18) by hand for all values of k.
We start by assuming that we know where the n poles (denoted as p j ) and m zeros (z j )
of G are located. It is also important to know d = n − m, which is the relative degree
of G. From this, we have the following rules:
1. The branches of a root locus start at the poles of G (i.e., when no control is applied,
or equivalently where k = 0), and end either at zeros of the open-loop system, or at
infinity (with a total of d ending at ∞).
2. A point on the real axis is included in the root locus if there is an odd total number
of poles and zeros to the right of it.
3. As k goes to ∞, the d branches of the root locus that go to ∞ asymptote to straight
lines with equal angles between them, with those angles (from the positive real axis)
given for k = 0, 1, . . . , d − 1 by
π + 2kπ
θk =
. (10.19)
d
Moreover, the asymptotes originate at a common center, given by
n m
©Õ Õ
c0 = d −1 ­ pj − zj ® . (10.20)
ª

« j=1 j=1 ¬
If d = 0, then all branches of the root locus end up at the zeros of the open-loop
system as k goes to ∞.
These rules typically give enough information to roughly sketch how the poles of the
closed-loop system will move as the controller gain varies. This can be particularly
useful for determining the best choice of control law for a given system. Note that
there are more complex rules that can enhance the accuracy of a hand-drawn root

h d i 0 0 9 8 088962 4 0 6 bli h d li b C b id i i
 Dawson: Linear Dynamical Systems and Control 211

Root locus for System 1 Root locus for System 2

1.0
4
0.5
2
Imag. axis

0 0.0

–2 Figure 10.4 Root locus plots

–0.5 for P1 (s) and P2 (s) with
–4 proportional control.
–1.0
0.0 0.1 –1 0
Real axis Real axis

locus plot, but we will not discuss them here. Root locus plots of the systems P1 (s)
and P2 (s) are shown in Figure 10.4. These plots are consistent with our findings in
Section 10.2, where it was found that proportional control could stabilize P2 (s) but
not P1 (s).
Note also that there are many other tools available to assist in designing controllers
for linear systems, such as Bode plots, Nyquist plots, and other auxiliary systems that
can help predict the behavior of the controlled system when it is subject to sensor
noise and disturbances. Details for these methods can be found in textbooks such as
Aström and Murray (2010) and Franklin et al. (1994).

10.4 Controllability and Observability

So far, we have looked at a few different control strategies, and were fortunately able
to find an approach that succeeded in stabilizing both of our systems. It is natural to
wonder if there are cases where this is an entirely futile exercise, in the sense that no
possible control strategy could work, given the ways in which we can manipulate the
state of our system, x, through the input u. It turns out, there is a relatively simple way
to determine whether a system is controllable in this sense. This can be done through
the use of the controllability matrix

C= B A2 B An−1 B . (10.21)
 
AB ···

In particular, we say that a system is controllable if the controllability matrix has a

rank of n (i.e., is full rank). Note that the C has n rows and n × ni columns, where ni
is the number of inputs. Along with how controllable a system is, we can also analyze
whether the outputs from the system are sufficient for us to be able to estimate the full
state of the system, through the observability matrix

h d i 0 0 9 8 088962 4 0 6 bli h d li b C b id i i
 212 Data-Driven Fluid Mechanics Chapter 10

 C 
 
 CA 
2 
 
O =  CA  . (10.22)

 . 
 .. 
 
C An−1 
 

We say that a system is observable if the observability matrix has rank n. O has n
columns and n × no rows, where no is the number of outputs. It is easy for us to
verify that the state-space versions of systems P1 and P2 are both controllable and
observable, and indeed this is the case for any choice of input and output channels
(i.e., for B = [0 1]T or [1 0]T , and C = [0 1] or [1 0]).
It turns out that we can also talk about controllability and observability not just as
binary notions, but can also study how controllable a system is through related objects
called the controllability and observability Gramians (Zhou & Doyle 1998), which are
discussed in Chapter 12.

10.5 Full-State Feedback

Thus far, we have only considered control using the system output, which we have
taken to be a scalar quantity. For the following sections, it will be useful to consider the
case where we have access to (or somewhat equivalently, have a method of estimating)
the full state of the system x. This can be represented with the following control
diagram, which will be useful in the following sections:

r x y
Σ xÛ = Ax + Bu C
−

We now consider two different methods for controller design using full-state feedback:
pole placement in Section 10.5.1, and optimal control using LQR in Section 10.5.2.

10.5.1 Pole Placement

When using full-state feedback with a control law given by

u = −K x, (10.23)

h d i 0 0 9 8 088962 4 0 6 bli h d li b C b id i i
 Dawson: Linear Dynamical Systems and Control 213

the dynamics of the closed-loop system can be written as

xÛ = (A − BK)x, (10.24)

which is the same as (10.9) with C = I . If we have desired closed-loop poles, then this
can be achieved by finding the components of K such that A − BK has eigenvalues
at these locations. If the system is controllable, then this can always be achieved. For
example, if we wish to place poles at p = −0.5 ± 0.5i for our systems P1 or P2 with
full-state feedback, then we find that K = [1.0965 0.0095]T . Pole placement can be
computed using the command K = place(A,B,p) in either MATLAB or Python.

10.5.2 Optimal Control

So far, the design of controllers has seemed somewhat ad hoc, and has relied upon
tuning parameters based on insight and desired properties. This motivates the question:
is there a way to define a controller that is, in some sense, the “best”? At least
for some definition of “best,” then this is something that we can indeed do. For
simplicity, assume that we have full-state feedback. The idea is that, when applying
control, we want to balance a desire to drive the system to a specified controlled state
(e.g., to zero), while at the same time ensuring that we are limiting the exertion of
our controller (i.e., penalizing the use of too much input energy). Mathematically, one
way to account for both of these objectives is by constructing a cost function of the
form
∫ ∞ 
J= xT Qx + uT Ru dt, (10.25)
t=0

where Q and R are positive definite matrices that define the penalty associated
with nonzero states and control effort, respectively. A controller that is designed to
minimize (10.25) is called an LQR controller. Mathematically, this control law that
minimizes (10.25) is given by
K = R−1 BT M, (10.26)

where M is the unique positive definite matrix that satisfies the algebraic Riccati
equation
AT M + M A − M BR−1 BT M + Q = 0. (10.27)

This might look like a difficult problem to solve, but once again there are commands
in MATLAB and Python libraries that will output the gain matrix K that optimizes
(10.25). As well as being “optimal” in this sense, it turns out that the resulting
controller also satisfies certain robustness properties (e.g., Stengel 1994).
To test this method, we design LQR controllers for the system P1 , using Q = I
and three different choices of R (10, 1, and 0.1), which prescribe how aggressive
the resulting controller is in achieving the desired state. The characteristics of these
feedback systems are explored in Figures 10.5 and 10.6. We see in Figure 10.5 that all
three controllers stabilize the system, though the locations of the poles are noticeably

h d i 0 0 9 8 088962 4 0 6 bli h d li b C b id i i
 214 Data-Driven Fluid Mechanics Chapter 10

axis

Figure 10.5 Poles and system

response for implementing
full-state feedback control
with LQR using three
choices of control effort
penalties.

axis

0.6 Uncontrolled
LQR, R = 100
0.4 LQR, R = 1
LQR, R = 0.1
0.2 Control applied Figure 10.6 Poles and system
response for implementing
y

0.0 full-state feedback control

with LQR using three
−0.2
choices of control effort
−0.4 penalties.

0 10 20 30 40 50
t

different for each case. Roughly speaking, larger R values penalize large amounts of
controller work, which tends to result in closed-loop systems that are “less different”
from the initial uncontrolled system, which is consistent with what is observed.
Similarly, we see in Figure 10.6 that the larger R value results in a controlled system
that takes longer to converge to the desired steady state, while the more aggressive
controllers (with smaller R values) converge to this state more rapidly, though with
larger control effort (which is not shown explicitly).

10.6 Additional Control Techniques and Considerations

As mentioned at the outset, there are numerous methods in linear control theory that
we have not touched on in this lecture. For a more comprehensive treatment of linear
control theory, standard references such as Aström and Murray (2010), Franklin et al.
(1994), and Skogestad and Postlethwaite (2007) are a good place to start, with the

h d i 0 0 9 8 088962 4 0 6 bli h d li b C b id i i
 Dawson: Linear Dynamical Systems and Control 215

latter in particular covering more advanced topics. In this final section of this chapter,
we briefly discuss additional ideas and techniques related to feedback control.

10.6.1 State Estimation

We observed in Section 10.5.2 that control could be achieved “optimally” if using
measurements of the full state. But is any of this useful if we do not have access to
the full state? As it turns out, if our system is observable (as described in Section
10.4), then we can design an estimator that can reconstruct an estimate of the full state
from the system outputs, which can then be used for state feedback, such as with an
LQR controller. The combined estimator and controller system, known as a linear-
quadratic-Gaussian (LQG) control system, will have poles at the locations of the LQR
subsystem, as well as those from the estimator subsystem. This property leads to the
separation principle, which says that if we can separately design a full-state feedback
system that is stable, along with an estimator that is stable, then the overall system is
also guaranteed to be stable.

10.6.2 Robust Control

There are many additional methods for controller design that extend beyond the
scope of what has been discussed here. In practice, factors such as measurement
noise, external disturbances, time delays, and modeling uncertainties can significantly
influence the performance of feedback control strategies. There are several common
metrics for determining how robust a controller will be to such effects, the most
common of which are concepts such as the phase margin and gain margin of a
feedback system. Roughly speaking, these are measures of how close a system is
to becoming unstable. There exist more advanced controller design techniques that
explicitly seek to produce feedback systems that are robust to disturbances, such as
H∞ methods (e.g., Skogestad & Postlethwaite 2007).

10.6.3 Nonlinear Control

This chapter has only considered the control of systems that are (or are assumed
to be) linear. While almost all real-life systems are nonlinear, linear methods can
still be very effective, subject to caveats that might relate to how close we stay
to a location about which a nonlinear system is linearized. However, there are
certainly cases where linear methods are ineffective or undesirable, thus requiring
more sophisticated nonlinear techniques. Extensions include gain scheduling and
linear parameter varying control (White et al. 2013), feedback linearization, and
backstepping (Khalil & Grizzle 2002).

h d i 0 0 9 8 088962 4 0 6 bli h d li b C b id i i
 216 Data-Driven Fluid Mechanics Chapter 10

10.6.4 Application of Feedback Control in Fluids

The ability to manipulate fluid flows to achieve desired behavior has the potential to
substantially reduce energy requirements for transportation of fluids, transportation of
vehicles through fluids, and to improve the reliability and safety of such engineering
tasks. The past several decades have seen substantial growth in control-theoretic
methods (those discussed in this chapter, and beyond) to enable the design of control
strategies that achieve such purposes. For example, many of the methods discussed in
this section (and related methods) have been directly applied to fluids systems for the
control of boundary layers (Bagheri et al. 2009, Semeraro et al. 2011, 2013, Belson
et al. 2013, Leclercq et al. 2019), flow over backward-facing steps (Hervé et al. 2012)
and cavities (Rowley & Williams 2006, Barbagallo et al. 2009, Illingworth et al. 2012),
bluff-body wakes (Bergmann et al. 2005, Illingworth et al. 2014, Brackston
et al. 2016, Illingworth 2016, Flinois & Morgans 2016), airfoils (Magill
et al. 2003, Ahuja & Rowley 2010, Kerstens et al. 2011, Brunton et al. 2014),
and wall-bounded turbulence (Choi et al. 1994, Kim 2011, Luhar et al. 2014, Toedtli
et al. 2019). Beyond methods based on linear control theory, recent developments
in applications of cluster (Kaiser et al. 2017a, Nair et al. 2019), network (Nair
et al. 2018), reinforcement learning (Rabault et al. 2019, see also Chapter 18),
genetic programming (Debien et al. 2016, Minelli et al. 2020, see also Chapter 17),
and other machine learning control methodologies (Gautier et al. 2015, Duriez
et al. 2017) have expanded the set of tools available for controlling fluid flows. For
more comprehensive reviews on methods and applications of flow control, see, for
example, Kim and Bewley (2007), Gad-el Hak (2007), Noack et al. (2011), Brunton
and Noack (2015), Rowley and Dawson (2017), and Brunton et al. (2020).

h d i 0 0 9 8 088962 4 0 6 bli h d li b C b id i i
 11 Nonlinear Dynamical Systems
S. Brunton

Dynamical systems provide a mathematical framework to describe the evolution of

complex physical systems, such as unsteady fluid dynamics. The field of dynamical
systems lives at the interface of many fields of mathematics, including linear algebra,
differential equations, topology, geometry, and numerical analysis. Modern dynamical
systems began with the seminal work of Poincaré on the chaotic motion of planets, and
it may be viewed as the culmination of hundreds of years of mathematical modeling
with differential equations. This chapter presents a modern perspective on dynamical
systems in the context of current goals and open challenges in fluid dynamics. Data-
driven dynamical systems is a rapidly evolving field, and therefore, we focus on a
mix of established and emerging methods that are driving current developments. In
particular, we will focus on the three key challenges of (1) high-dimensionality, (2)
discovering dynamics from data, and (3) finding data-driven representations that make
nonlinear systems amenable to linear analysis.

11.1 Introduction

In recent decades, the field of dynamical systems has grown rapidly, with analytical
derivations and first principle models giving way to data-driven approaches. Thus,
machine learning and big data are driving a paradigm shift in the analysis and under-
standing of dynamical systems in science and engineering. This trend is particularly
evident in the field of fluid dynamics, which is one of the original big data fields.
In addition, the classical geometric and statistical perspectives on dynamical
systems are being complemented by a third operator-theoretic perspective, based
on the evolution of measurements of the system. This so-called Koopman operator
theory is poised to capitalize on the increasing availability of measurement data from
complex systems. Moreover, Koopman theory provides a path to identify intrinsic
coordinate systems to represent nonlinear dynamics in a linear framework. Obtaining
linear representations of strongly nonlinear systems has the potential to revolutionize
our ability to predict and control these systems (Figure 11.1). See Chapters 10 and 7
for overviews of linear systems and Koopman theory, respectively.

h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
 218 Data-Driven Fluid Mechanics Chapter 11

Linear Weakly nonlinear Chaotic

Increasing nonlinearity

Figure 11.1 Schematic depiction of increasing nonlinearity in dynamical systems. Behavior

goes from periodic to quasiperiodic, and eventually becomes chaotic. Reproduced from
Brunton et al. (2017).

11.2 Dynamical Systems

11.2.1 Continuous-Time Systems

Throughout this chapter, we will consider dynamical systems of the form
d
x(t) = f(x(t), t; β), (11.1)
dt
where x is the state of the system and
Linear f isNonlinear
Weakly a vector field that possibly depends on the
Chaotic
state x, time t, and a set of parameters β.
Finite-dimensional representations of the Navier–Stokes equations take this form in
Increasing Complexity
which case the state x may contain the flow-field variables on a discretized grid, or a
finite number of Fourier coefficients. These dynamics are typically high-dimensional,
often requiring millions or billions of degrees of freedom. It is also possible to obtain
low-dimensional dynamical systems that describe a fluid flow in terms of the dynamics
of the amplitudes of a small set of modes, for example, by Galerkin projection of the
governing equations onto a truncated set of Proper Orthogonal Decomposition (POD)
modes. In these projected models, the dynamics in (11.1) have quadratic nonlinearities
due to the convective nonlinearity (u · ∇)u in the Navier–Stokes equations.
The Lorenz (1963) system

xÛ = σ(y − x), (11.2a)

yÛ = x(ρ − z) − y, (11.2b)
zÛ = x y − βz (11.2c)

is one of the earliest nonlinear reduced-order models of an unsteady fluid system,

as it was developed to model chaotic Rayleigh–Benard convection. The x, y, and z
variables denote the amplitudes of three spatial modes, posited by Saltzman (1962),
onto which the Navier–Stokes equations are Galerkin projected. In this case, the state
T T
vector is x = x y z and the parameter vector is β = σ ρ β . A trajectory
 

h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
 Brunton: Nonlinear Dynamical Systems 219

y
x
Figure 11.2 Chaotic trajectory of the Lorenz system from (11.2). From Brunton and Kutz
(2019), reproduced with permission of the Licensor through PLSclear.

Figure 11.3 Depiction of uncertainty in the Lorenz system. A cube of initial conditions is
evolved along the flow. Although the trajectories stay close together for some time, after a
while, they begin to spread along the attractor.

of the Lorenz system is shown in Figure 11.2, with parameters σ = 10, ρ = 28, and
β = 8/3.
The Lorenz system is among the simplest and most well-studied dynamical
systems that exhibit chaos, which is characterized as a sensitive dependence on
initial conditions. Two trajectories with nearby initial conditions will rapidly diverge
in behavior, and after long periods, only statistical statements can be made. This
sensitivity is depicted in Figure 11.3, where a cube of initial conditions is evolved
along the flow of the dynamical system, showing that after some time the initial
conditions become spread out along the attractor. This type of chaotic mixing is
characteristic of turbulence.

h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
 220 Data-Driven Fluid Mechanics Chapter 11

1 3.45

3.5

1.5 3.55

3.6
2
3.65

β β 3.7
2.5

r
r

3.75

3.8
3
3.85

3.5 3.9

3.95

4
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x x x
x

Figure 11.4 Attracting sets of the logistic map for varying parameter β. From Brunton and
Kutz (2019), reproduced with permission of the Licensor through PLSclear.

We will often consider the simpler case of an autonomous system without time
dependence or parameters:
d
x(t) = f(x(t)). (11.3)
dt

11.2.2 Discrete-Time Systems

In addition to the aforementioned continuous-time dynamics, we will also consider
discrete-time dynamical systems

xk+1 = F(xk ). (11.4)

Discrete-time dynamics are often more natural when considering experimental data
and simulations, as data is typically generated discretely in time.
Also known as a map, the discrete-time dynamics are more general than the
continuous-time formulation in (11.3), encompassing discontinuous and hybrid sys-
tems as well.
As an example, consider the logistic map, which is a simple model to explore the
complexity of population dynamics:

xk+1 = βxk (1 − xk ). (11.5)

As the parameter β is increased, the attracting set becomes increasingly complex,

shown in Figure 11.4. A series of period-doubling bifurcations occur until the
attracting set becomes fractal. The complexity of turbulence is also often understood
as the result of a sequence of period-doubling Hopf bifurcations.
Discrete-time dynamics may be induced from continuous-time dynamics, where xk
is obtained by sampling the trajectory in (11.3) discretely in time, so that xk = x(k∆t).
The discrete-time propagator F∆t is now parameterized by the time step ∆t. For an
arbitrary time t, the flow map Ft is defined as

h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
 Brunton: Nonlinear Dynamical Systems 221

∫ t0 +t
Ft (x(t0 )) = x(t0 ) + f(x(τ)) dτ. (11.6)
t0

11.2.3 Linear Dynamics and Spectral Decomposition

As explored in Chapter 10, whenever possible it is desirable to work with linear
dynamics of the form
d
x = Ax. (11.7)
dt
Even for nonlinear systems, it is often possible to understand the behavior near fixed
points in terms of locally linearized dynamics.
Linear dynamical systems admit closed-form solutions, and there are a wealth of
techniques for the analysis, prediction, numerical simulation, estimation, and control
of such systems. The solution of (11.7) is given by

x(t0 + t) = eAt x(t0 ). (11.8)

The dynamics are entirely characterized by the eigenvalues and eigenvectors of the
matrix A, given by the spectral decomposition (eigendecomposition) of A:

AT = TΛ. (11.9)

When A has n distinct eigenvalues, then Λ is a diagonal matrix containing the

eigenvalues λ j and T is a matrix whose columns are the linearly independent
eigenvectors ξ j associated with eigenvalues λ j . In this case, it is possible to write
A = TΛT−1 , and the solution in (11.8) becomes

x(t0 + t) = TeΛt T−1 x(t0 ). (11.10)

More generally, in the case of repeated eigenvalues, the matrix Λ will consist of Jordan
blocks (Perko 2013). Note that the continuous-time system gives rise to a discrete-time
dynamical system, with Ft given by the solution map exp(At) in (11.8). In this case,
the discrete-time eigenvalues are given by eλt .
For non-normal operators A, the eigenvectors may be nearly parallel. In this case,
even with stable eigenvalues, the system may exhibit transient growth, where an initial
condition first grows before eventually converging to the origin. Transient growth
is observed in many shear flows, and the excursion from the origin often results
in perturbation amplitudes that are large enough to excite nonlinear dynamics. For
example, pipe flow is linearly stable at all Reynolds numbers, but non-normality and
transient energy growth results in sustained turbulence at finite Reynolds numbers.
The matrix T−1 defines a transformation, z = T−1 x, into intrinsic eigenvector
coordinates, z, where the dynamics become decoupled:
d
z = Λz. (11.11)
dt

h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
 222 Data-Driven Fluid Mechanics Chapter 11

y
0
1
0
-1 x
-0.2 0. 0.2 -1
β 0.4 0.6
Figure 11.5 Schematic of a Hopf bifurcation. Adapted from Brunton, Proctor and Kutz
(2016a).

In other words, each coordinate, z j , only depends on itself, with dynamics given by
d
zj = λj zj . (11.12)
dt
Thus, it is highly desirable to work with linear systems, since it is possible to
easily transform the system into eigenvector coordinates where the dynamics become
decoupled. No such closed-form solution or simple linear change of coordinates exists
in general for nonlinear systems, motivating many of the directions described in this
chapter.

11.2.4 Bifurcations
We have already seen bifurcations that occur in the logistic map. Bifurcations are par-
ticularly important in fluid dynamic systems. Small changes in physical parameters,
such as the Reynolds number, may drive qualitative changes in the behavior of the
resulting dynamics. For example, in the flow past a stationary circular cylinder, when
the Reynolds number increases past 47, the flow goes from a steady laminar solution
to unsteady, periodic vortex shedding (see Chapter 1). This Hopf bifurcation is shown
in Figure 11.5, and is given by the following dynamical system:
dx  
= βx − ωy − Ax x 2 + y 2 , (11.13a)
dt
dy  
= ωx + βy − Ay x 2 + y 2 . (11.13b)
dt
As β goes from negative to positive values, a single stable fixed point at the origin
becomes unstable, and a stable limit cycle emerges.
As a simpler example, we often consider the one-dimensional pitchfork bifurcation,
shown in Figure 11.6 and given by the following dynamical system:
dx
= βx − x 3 . (11.14)
dt
For negative values of the parameter β, there is a single stable fixed point at the
origin. As this parameter increases from negative to positive, this fixed point becomes
√
unstable and two other stable fixed points emerge at ± β.

h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
 Brunton: Nonlinear Dynamical Systems 223

Figure 11.6 Schematic of a pitchfork bifurcation.

It is interesting to note that if we write the Hopf normal form in (11.13) in polar
coordinates, with r 2 = x 2 + y 2 , then the equation for the amplitude r reduces to the
pitchfork normal form in (11.14).

11.3 Goals and Challenges in Modern Dynamical Systems

There are a number of goals and uses of dynamical systems models of fluid flows:

1. Future state prediction. In many cases, we seek predictions of the future state
of a system. Long-time predictions may still be challenging, especially for chaotic
systems, such as turbulence, where long-term statistics are a more reasonable goal.
2. Design and optimization. The dynamical system may be used as a surrogate model
to tune the parameters of a system for improved performance or stability. For
example, aircraft and automobile designers modify geometry and control surfaces
to improve aerodynamic performance.
3. Estimation and control. For many fluid systems, an ultimate goal is to actively
control the system through feedback, using measurements of the system to inform
actuation to modify the behavior. For high-dimensional systems, such as fluids, it is
often necessary to estimate the full state of the system from limited measurements.
4. Interpretability and physical understanding. Perhaps a more fundamental goal of
dynamical systems modeling is to provide physical insight and interpretability into
a system’s behavior through analyzing trajectories and solutions to the governing
equations of motion. Dynamical systems models of fluids will ideally yield some
insight into fundamental mechanisms that drive the flow.

These goals are often challenged by the multiscale and nonlinear nature of fluids.
There is also uncertainty in the boundary conditions, parameters, and measurements
of the system. These challenges may be summarized as follows:

h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
 224 Data-Driven Fluid Mechanics Chapter 11

1. High-dimensionality. Fluid flows, and the resulting measurement and simulation

data, are often exceedingly high-dimensional. Such systems are expensive to
simulate and analyze, especially for iterative design optimization, uncertainty
quantification, and real-time control, motivating accurate and efficient surrogate
models of lower dimension. Pattern extraction and modal decompositions (Taira
et al. 2017, Taira et al. 2019) provide powerful techniques to reduce the dimension
of fluid systems. These techniques include both classical linear dimensionality
reduction, such as POD and dynamic mode decomposition (DMD) (see Chapters
6 and 7), as well as emerging nonlinear dimensionality reduction techniques from
machine learning (see Chapter 3).
2. Nonlinearity. Nonlinearity remains a primary challenge in analyzing and control-
ling fluid dynamics. Even simple quadratic nonlinearity, such as the convection term
in the Navier–Stokes equations, significantly complicates analysis and control. Tra-
ditional linearization around fixed points make it possible to leverage linear mod-
eling and control techniques, even for nonlinear systems. However, global analysis
has remained largely qualitative and computational, limiting the theory of nonlinear
prediction, estimation, and control away from fixed points and periodic orbits.
To address the issue of nonlinearity, operator-theoretic approaches to dynamical
systems are becoming increasingly used. As we will show, it is possible to represent
nonlinear dynamical systems in terms of infinite-dimensional but linear operators,
such as the Koopman operator from Section 11.4 which advances measurement
functions, and the Perron–Frobenius operator which advances probability densities
and ensembles through the dynamics.
3. Unknown dynamics. Despite knowledge of the governing Navier–Stokes equa-
tions, the reduced-order physical mechanisms are often unknown for a particular
flow. Traditionally, systems in fluid dynamics were analyzed by making ideal
approximations and then deriving simple differential equation models via asymp-
totic expansions. Dramatic simplifications could often be made by exploiting
symmetries and clever coordinate systems. With increasingly complexity, the
paradigm is shifting from this classical approach to data-driven methods to discover
governing equations. As data become increasingly abundant, and we continue to
investigate systems that are not amenable to first-principles analysis, regression and
machine learning are becoming vital tools to discover dynamical systems from data.
This is the basis of many of modern techniques, including the DMD and data-driven
Koopman methods (see Chapter 7), the sparse identification of nonlinear dynamics
(SINDy) (see Section 12.3), as well as the use of genetic programming to identify
dynamics from data (Bongard & Lipson 2007, Schmidt & Lipson 2009).

11.4 Koopman Theory

Koopman operator theory has recently emerged as an alternative perspective for

dynamical systems in terms of the evolution of measurements g(x). In 1931, Bernard
O. Koopman demonstrated that it is possible to represent a nonlinear dynamical

h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
 Brunton: Nonlinear Dynamical Systems 225

system in terms of an infinite-dimensional linear operator acting on a Hilbert space

of measurement functions of the state of the system. Koopman analysis has recently
gained renewed interest with the pioneering work of Mezic and collaborators (Mezić
& Banaszuk 2004, Mezić 2005, Budišić & Mezić 2009, Budišić et al. 2012, Budišić
& Mezić 2012, Lan & Mezić 2013, Mezić 2013). The so-called Koopman operator
is linear, and its spectral decomposition completely characterizes the behavior of
a nonlinear system, analogous to (11.7). However, it is also infinite-dimensional,
as there are infinitely many degrees of freedom required to describe the space
of all possible measurement functions g of the state. This poses new challenges.
Obtaining finite-dimensional, matrix approximations of the Koopman operator is the
focus of intense research efforts and holds the promise of enabling globally linear
representations of nonlinear dynamical systems. Expressing nonlinear dynamics in a
linear framework is appealing because of the wealth of optimal estimation and control
techniques available for linear systems and the ability to analytically predict the future
state of the system. Obtaining a finite-dimensional approximation of the Koopman
operator has been challenging in practice, as it involves identifying a subspace spanned
by a subset of eigenfunctions of the Koopman operator.

11.4.1 Mathematical Formulation

The Koopman operator advances measurement functions of the state with the flow of
the dynamics. We consider real-valued measurement functions g : M → R, which are
elements of an infinite-dimensional Hilbert space. The functions g are also commonly
known as observables, although this may be confused with the unrelated observability
from control theory. Often, the Hilbert space is given by the Lebesgue square-
integrable functions on M; other choices of a measure space are also valid.
The Koopman operator Kt is an infinite-dimensional linear operator that acts on
measurement functions g as

Kt g = g ◦ Ft , (11.15)

where ◦ is the composition operator. For a discrete-time system with time step ∆t, this
becomes

K∆t g(xk ) = g(F∆t (xk )) = g(xk+1 ). (11.16)

In other words, the Koopman operator defines an infinite-dimensional linear dynam-

ical system that advances the observation of the state gk = g(xk ) to the next time
step:

g(xk+1 ) = K∆t g(xk ). (11.17)

Note that this is true for any observable function g and for any state xk .

h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
 226 Data-Driven Fluid Mechanics Chapter 11

The Koopman operator is linear, a property that is inherited from the linearity of
the addition operation in function spaces:
Kt (α1 g1 (x) + α2 g2 (x)) = α1 g1 (Ft (x)) + α2 g2 (Ft (x)) (11.18a)
= α1 Kt g1 (x) + α2 Kt g2 (x). (11.18b)
For sufficiently smooth dynamical systems, it is also possible to define the
continuous-time analogue of the Koopman dynamical system in (11.17):
d
g = Kg. (11.19)
dt
The operator K is the infinitesimal generator of the one-parameter family of trans-
formations Kt (Abraham et al. 1988). It is defined by its action on an observable
function g:
Kt g − g g ◦ Ft − g
Kg = lim = lim . (11.20)
t→0 t t→0 t
The linear dynamical systems in (11.19) and (11.17) are analogous to the dynamical
systems in (11.3) and (11.4), respectively. It is important to note that the original state
x may be the observable, and the infinite-dimensional operator Kt will advance this
function. However, the simple representation of the observable g = x in a chosen basis
for Hilbert space may become arbitrarily complex once iterated through the dynamics.
In other words, finding a representation for K x may not be simple or straightforward.

Koopman Eigenfunctions and Intrinsic Coordinates

The Koopman operator is linear, which is appealing, but infinite-dimensional, posing
issues for representation and computation. Instead of capturing the evolution of
all measurement functions in a Hilbert space, applied Koopman analysis attempts
to identify key measurement functions that evolve linearly with the flow of the
dynamics. Eigenfunctions of the Koopman operator provide just such a set of
special measurements that behave linearly in time. In fact, a primary motivation to
adopt the Koopman framework is the ability to simplify the dynamics through the
eigendecomposition of the operator.
A discrete-time Koopman eigenfunction ϕ(x) corresponding to eigenvalue λ
satisfies
ϕ(xk+1 ) = K∆t ϕ(xk ) = λϕ(xk ). (11.21)
In continuous-time, a Koopman eigenfunction ϕ(x) satisfies
d
ϕ(x) = K ϕ(x) = λϕ(x). (11.22)
dt
Obtaining Koopman eigenfunctions from data or from analytic expressions is a central
applied challenge in modern dynamical systems. Discovering these eigenfunctions
enables globally linear representations of strongly nonlinear systems.
Applying the chain rule to the time derivative of the Koopman eigenfunction ϕ(x)
yields
d
ϕ(x) = ∇ϕ(x) · xÛ = ∇ϕ(x) · f(x). (11.23)
dt

h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
 Brunton: Nonlinear Dynamical Systems 227

Combined with (11.22), this results in a partial differential equation (PDE) for the
eigenfunction ϕ(x):
∇ϕ(x) · f(x) = λϕ(x). (11.24)

With this nonlinear PDE, it is possible to approximate the eigenfunctions, either

by solving for the Laurent series or with data via regression, both of which are
explored below. This formulation assumes that the dynamics are both continuous and
differentiable. The discrete-time dynamics in (11.4) are more general, although in
many examples the continuous-time dynamics have a simpler representation than the
discrete-time map for long times. For example, the simple Lorenz system has a simple
continuous-time representation, yet is generally unrepresentable for even moderately
long discrete-time updates.
The key takeaway from (11.21) and (11.22) is that the nonlinear dynamics
become completely linear in eigenfunction coordinates, given by ϕ(x). As a simple
example, any conserved quantity of a dynamical system is a Koopman eigenfunction
corresponding to eigenvalue λ = 0. This establishes a Koopman extension of
the famous Noether’s theorem (Noether 1918), implying that any symmetry in the
governing equations gives rise to a new Koopman eigenfunction with eigenvalue
λ = 0. For example, the Hamiltonian energy function is a Koopman eigenfunction
for a conservative system. In addition, the constant function ϕ = 1 is always a trivial
eigenfunction corresponding to λ = 0 for every dynamical system.

Eigenvalue Lattices
Interestingly, a set of Koopman eigenfunctions may be used to generate more
eigenfunctions. In discrete time, we find that the product of two eigenfunctions ϕ1 (x)
and ϕ2 (x) is also an eigenfunction

Kt (ϕ1 (x)ϕ2 (x)) = ϕ1 (Ft (x))ϕ2 (Ft (x)) (11.25a)

= λ1 λ2 ϕ1 (x)ϕ2 (x), (11.25b)

corresponding to a new eigenvalue λ1 λ2 given by the product of the two eigenvalues

of ϕ1 (x) and ϕ2 (x).
In continuous time, the relationship becomes
d
K (ϕ1 ϕ2 ) = (ϕ1 ϕ2 ) (11.26a)
dt
= ϕÛ1 ϕ2 + ϕ1 ϕÛ2 (11.26b)
= λ1 ϕ1 ϕ2 + λ2 ϕ1 ϕ2 (11.26c)
= (λ1 + λ2 )ϕ1 ϕ2 . (11.26d)

Mathematically, the set of Koopman eigenfunctions forms a commutative monoid

under point-wise multiplication; a monoid has the structure of a group, except that the
elements do not necessarily have inverses. Thus, depending on the dynamical system,
there may be a finite set of generator eigenfunction elements that may be used to
construct all other eigenfunctions. The corresponding eigenvalues similarly form a

h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
 228 Data-Driven Fluid Mechanics Chapter 11

lattice, based on the product λ1 λ2 or sum λ1 + λ2 , depending on whether the dynamics

are in discrete time or continuous time. For example, given a linear system xÛ = λx,
then ϕ(x) = x is an eigenfunction with eigenvalue λ. Moreover, ϕα = x α is also an
eigenfunction with eigenvalue αλ for any α.
The continuous-time and discrete-time lattices are related in a simple way. If
the continuous-time eigenvalues are given by λ, then the corresponding discrete-
time eigenvalues are given by eλt . Thus, the eigenvalue expressions in (11.25b) and
(11.26d) are related as

eλ1 t eλ2 t ϕ1 (x)ϕ2 (x) = e(λ1 +λ2 )t ϕ1 (x)ϕ2 (x). (11.27)

As another simple demonstration of the relationship between continuous-time and

discrete-time eigenvalues, consider the continuous-time definition in (11.20) applied
to an eigenfunction:
Kt ϕ(x) − ϕ(x) eλt ϕ(x) − ϕ(x)
lim = lim = λϕ(x). (11.28)
t→0 t t→0 t

11.4.2 Koopman Mode Decomposition

Until now, we have considered scalar measurements of a system, and we uncovered
special eigen-measurements that evolve linearly in time. However, we often take
multiple measurements of a system. In extreme cases, we may measure the entire
state of a high-dimensional spatial system, such as an evolving fluid flow. These
measurements may then be arranged in a vector g:
 g1 (x) 
 
 g2 (x) 
g(x) =  .  . (11.29)
 
 .. 
 
g (x)
 p 
Each of the individual measurements may be expanded in terms of the eigenfunctions
ϕ j (x), which provide a basis for Hilbert space:
∞
Õ
gi (x) = vi j ϕ j (x). (11.30)
j=1

Thus, the vector of observables, g, may be similarly expanded:

 g1 (x) 
 
 g2 (x)  Õ∞
g(x) =  .  = ϕ j (x)v j , (11.31)
 
 .. 
  j=1
g (x)
 p 
where v j is the jth Koopman mode associated with the eigenfunction ϕ j .
For conservative dynamical systems, such as those governed by Hamiltonian
dynamics, the Koopman operator is unitary. Thus, the Koopman eigenfunctions are

h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
 Brunton: Nonlinear Dynamical Systems 229

orthonormal for conservative systems, and it is possible to compute the Koopman

modes v j directly by projection:
 hϕ j , g1 i 
 hϕ j , g2 i 
 
vj =  .  , (11.32)
 
 .. 
 
 hϕ , g i 
 j p 
where h·, ·i is the standard inner product of functions in Hilbert space. These modes
have a physical interpretation in the case of direct spatial measurements of a system,
g(x) = x in which case the modes are coherent spatial modes that behave linearly with
the same temporal dynamics (i.e., oscillations, possibly with linear growth or decay).
Given the decomposition in (11.31), it is possible to represent the dynamics of the
measurements g as follows:
∞
Õ
g(xk ) = K∆t
k
g(x0 ) = K∆t
k
ϕ j (x0 )v j (11.33a)
j=0
∞
Õ
= k
K∆t ϕ j (x0 )v j (11.33b)
j=0
Õ∞
= λ kj ϕ j (x0 )v j . (11.33c)
j=0

This sequence of triples {(λ j , ϕ j , v j )}∞

j=0 is known as the Koopman mode decom-
position, and was introduced in Mezić (2005). The Koopman mode decomposition
was later connected to data-driven regression via the dynamic mode decomposi-
tion (Rowley et al. 2009).

Invariant Eigenspaces and Finite-Dimensional Models

Instead of capturing the evolution of all measurement functions in a Hilbert space,
applied Koopman analysis approximates the evolution on an invariant subspace
spanned by a finite set of measurement functions.
A Koopman-invariant subspace is defined as the span of a set of functions
{g1, g2, · · · , g p } if all functions g in this subspace

g = α1 g1 + α2 g2 + · · · + αp g p (11.34)

remain in this subspace after being acted on by the Koopman operator K:

Kg = β1 g1 + β2 g2 + · · · + β p g p . (11.35)

It is possible to obtain a finite-dimensional matrix representation of the Koopman

operator by restricting it to an invariant subspace spanned by a finite number of
p
functions {g j } j=0 . The matrix representation K acts on a vector space R p , with the
coordinates given by the values of g j (x). This induces a finite-dimensional linear
system, as in (11.17) and (11.19).

h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
 230 Data-Driven Fluid Mechanics Chapter 11

Any finite set of eigenfunctions of the Koopman operator will span an invari-
ant subspace. Discovering these eigenfunction coordinates is, therefore, a central
challenge, as they provide intrinsic coordinates along which the dynamics behave
linearly. In practice, it is more likely that we will identify an approximately invariant
p
subspace, given by a set of functions {g j } j=0 , where each of the functions g j is well
Íp
approximated by a finite sum of eigenfunctions: g j ≈ k=0 αk ϕk .

11.4.3 Examples of Koopman Embeddings

Nonlinear System with Single Fixed Point and a Slow Manifold
Here, we consider an example system with a single fixed point, given by
xÛ1 = µx1, (11.36a)
xÛ2 = λ(x2 − x12 ). (11.36b)
For λ < µ < 0, the system exhibits a slow attracting manifold given by x2 = x12 . It
is possible to augment the state x with the nonlinear measurement g = x12 , to define
a three-dimensional Koopman invariant subspace. In these coordinates, the dynamics
become linear:
 y   µ 0 0   y1   y1   x1 
d  1  
= 0 λ for  y2  =  x2  . (11.37a)
     
y 2 −λ   y2

dt   
   
y  x2
   
y
 3  0 0 2µ  y 
  3  3  1 
The full three-dimensional Koopman observable vector space is visualized in
Figure 11.7. Trajectories that start on the invariant manifold y3 = y12 , visualized by
the blue surface, are constrained to stay on this manifold. There is a slow subspace,
spanned by the eigenvectors corresponding to the slow eigenvalues µ and 2µ; this
subspace is visualized by the green surface. Finally, there is the original asymptotically
attracting manifold of the original system, y2 = y12 , which is visualized as the red
surface. The blue and red parabolic surfaces always intersect in a parabola that is
inclined at a 45◦ angle in the y2 –y3 direction. The green surface approaches this 45◦
inclination as the ratio of fast to slow dynamics becomes increasingly large. In the full
three-dimensional Koopman observable space, the dynamics produce a single stable
node, with trajectories rapidly attracting onto the green subspace and then slowly
approaching the fixed point.

Intrinsic Coordinates Defined by Eigenfunctions of the Koopman Operator

The left eigenvectors of the Koopman operator yield Koopman eigenfunctions (i.e.,
eigenobservables). The Koopman eigenfunctions of (11.37a) corresponding to eigen-
values µ and λ are
λ
ϕµ = x1, and ϕλ = x2 − bx12 with b= . (11.38)
λ − 2µ
The constant b in ϕλ captures the fact that for a finite ratio λ/µ, the dynamics only
shadow the asymptotically attracting slow manifold x2 = x12 , but in fact follow
neighboring parabolic trajectories. This is illustrated more clearly by the various
surfaces in Figure 11.7 for different ratios λ/µ.

h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
 Brunton: Nonlinear Dynamical Systems 231

Figure 11.7 Visualization of three-dimensional linear Koopman system from (11.37a) along
with projection of dynamics onto the x1 –x2 plane. The attracting slow manifold is shown in
red, the constraint y3 = y12 is shown in blue, and the slow unstable subspace of (11.37a) is
shown in green. Black trajectories of the linear Koopman system in y project onto trajectories
of the full nonlinear system in x in the y1 –y2 plane. Here, µ = −0.05 and λ = 1. Reproduced
from Brunton et al. (2016b).

In this way, a set of intrinsic coordinates may be determined from the observable
functions defined by the left eigenvectors of the Koopman operator on an invariant
subspace. Explicitly,
ϕα (x) = ξα y(x), where ξα K = αξα . (11.39)
These eigen-observables define observable subspaces that remain invariant under the
Koopman operator, even after coordinate transformations. As such, they may be
regarded as intrinsic coordinates (Williams et al. 2015) on the Koopman-invariant
subspace.

h d i 0 0 9 8 088962 4 0 bli h d li b C b id i i
 12 Methods for System Identification
S. Brunton

Fluid dynamics are characterized by high-dimensional, nonlinear dynamical systems

that exhibit multiple scales in space and time. Simulating these equations is currently
too expensive for real-time optimization and control (Bagheri et al. 2009, Fabbiane
et al. 2014, Brunton & Noack 2015), and there are efforts to obtain reduced-order
models (ROMs) that capture the salient dynamics with a lower-dimensional system.
There are two broad approaches to obtain ROMs: first, it is possible to start with
a high-dimensional system, such as the discretized Navier–Stokes equations, and
project the dynamics onto a low-dimensional subspace identified, for example, using
proper orthogonal decomposition (POD; Berkooz et al. 1993, Holmes et al. 2012)
and Galerkin projection (Quarteroni & Rozza 2013, Benner et al. 2015). The second
approach is to collect data from a simulation or an experiment and identify a low-
rank model using data-driven techniques. This approach is typically called system
identification, and is often preferred for control design because of the relative ease
of implementation. The dynamic mode decomposition (DMD) from Chapter 7 is an
example of system identification that is widely used in fluid mechanics.
Most of the system identification techniques discussed here, including DMD from
Chapter 7, result in linear models (see Chapter 10). Even for nonlinear systems,
such as the Navier–Stokes equations, it is often possible to use linear approaches
for closed-loop control. There are numerous examples of successful linear model-
based flow control (Bagheri et al. 2009, Fabbiane et al. 2014, Brunton & Noack 2015).
Control generally relies on fast, low-latency models for robust real-time performance,
especially for fluids where the dynamics often evolve on fast timescales. There-
fore, system identification also relies on dimensionality reduction to obtain low-
dimensional approximations of the system for use in real-time feedback control.
This chapter explores several system identification techniques. First, we introduce
balanced truncation and balanced POD (BPOD), which have been used extensively
to obtain linear ROMs in fluid systems. Next, we develop the eigensystem realization
algorithm (ERA), which has deep connections to both balanced POD and DMD.
Finally, we introduce the sparse identification of the nonlinear dynamics (SINDy)
algorithm.

h d i 0 0 9 8 088962 4 0 8 bli h d li b C b id i i
 Brunton: Methods for System Identification 233

12.1 Balanced Model Reduction

Balanced model reduction refers to an alternative control-oriented procedure that

results in low-dimensional models for input–output systems. POD is the most common
low-dimensional basis used for model reduction, and it orders modes based on
the kinetic energy content of the approximated state. Instead of energy, balanced
modes are ordered by how much input–output energy they capture, making them
useful for feedback control. This section will explore balanced model reduction in
the broader context of reduced-order modeling and system identification. Balanced
modeling of fluids is based on both data and equations, and is typically used to reduce
high-dimensional fluid systems. As with POD-based modeling, the general idea is
to determine a hierarchical modal decomposition of the system state that may be
truncated at some model order, only keeping the coherent structures that are most
important for control.

12.1.1 The Goal of Model Reduction

Consider a high-dimensional system, depicted schematically in Figure 12.1,
d
x = Ax + Bu, (12.1a)
dt
y = Cx + Du, (12.1b)
for example from a spatially discretized simulation of a PDE. The primary goal of
model reduction is to find a coordinate transformation x = Ψx̃ giving rise to a related
system (Ã, B̃, C̃, D̃) with similar input–output characteristics,
d
x̃ = Ãx̃ + B̃u, (12.2a)
dt
y = C̃x̃ + D̃u, (12.2b)
in terms of a state x̃ ∈ Rr with reduced dimension, r  n. Note that u and y are the
same in (12.1) and (12.2) even though the system states are different. Obtaining the
projection operator Ψ will be the focus of this section.
As a motivating example, consider the following simplified model:
d x1 −2 0 x1 1
      
= + u, (12.3a)
dt x2 0 −1 x2 10−10
 x1
 
y = 1 10 −10 . (12.3b)

x2
In this case, the state x2 is barely controllable and barely observable. Simply choosing
x̃ = x1 will result in a ROM that faithfully captures the input–output dynamics.
Although the choice x̃ = x1 seems intuitive in this extreme case, many model
reduction techniques would erroneously favor the state x̃ = x2 , since it is more lightly
damped. Throughout this section, we will investigate how to accurately and efficiently
find the transformation matrix Ψ that best captures the input–output dynamics.

h d i 0 0 9 8 088962 4 0 8 bli h d li b C b id i i
 234 Data-Driven Fluid Mechanics Chapter 12

u y
System

Figure 12.1 Input–output system. A control-oriented reduced-order model will capture the
transfer function from u to y.

The POD (Berkooz et al. 1993, Holmes et al. 2012) provides a transform matrix
Ψ, the columns of which are modes that are ordered based on energy content.1 POD
has been widely used to generate ROMs of complex systems, many for control, and it
is guaranteed to provide an optimal low-rank basis to capture the maximal energy or
variance in a data set. However, it may be the case that the most energetic modes are
nearly uncontrollable or unobservable, and therefore may not be relevant for control.
Similarly, in many cases the most controllable and observable state directions may
have very low energy; for example, acoustic modes typically have very low energy,
yet they mediate the dominant input–output dynamics in many fluid systems. The
rudder on a ship provides a good analogy: although it accounts for a small amount of
the total energy, it is dynamically important for control.
Instead of ordering modes based on energy, it is possible to determine a hierarchy of
modes that are most controllable and observable, therefore capturing the most input–
output information. These modes give rise to balanced models, giving equal weighting
to the controllability and observability of a state via a coordinate transformation
that makes the controllability and observability Gramians equal and diagonal. These
models have been extremely successful, although computing a balanced model using
traditional methods is prohibitively expensive for high-dimensional systems. In this
section, we describe the balancing procedure, as well as modern methods for efficient
computation of balanced models. A computationally efficient suite of algorithms for
model reduction and system identification may be found in Belson et al. (2014).
A balanced ROM should map inputs to outputs as faithfully as possible for a given
model order r. It is therefore important to introduce an operator norm to quantify how
similarly (12.1) and (12.2) act on a given set of inputs. Typically, we take the infinity
norm of the difference between the transfer functions G(s) and Gr (s) obtained from
the full system (12.1) and reduced system (12.2), respectively. This norm is given by
kGk∞ , max σ1 (G(iω)) . (12.4)
ω

To summarize, we seek a ROM (12.2) of low order, r  n, so the operator norm

kG − Gr k∞ is small.

1 When the training data consists of velocity fields, for example from a high-dimensional discretized fluid
system, then the singular values literally indicate the kinetic energy content of the associated mode. It is
common to refer to POD modes as being ordered by energy content, even in other applications,
although variance is more technically correct.

h d i 0 0 9 8 088962 4 0 8 bli h d li b C b id i i
 Brunton: Methods for System Identification 235

12.1.2 Change of Variables in Control Systems

The balanced model reduction problem may be formulated in terms of first finding a
coordinate transformation

x = Tz, (12.5)

that hierarchically orders the states in z in terms of their ability to capture the input–
output characteristics of the system. We will begin by considering an invertible
transformation T ∈ Rn×n , and then provide a method to compute just the first r
columns, which will comprise the transformation Ψ in (12.2). Thus, it will be possible
to retain only the first r most controllable/observable states, while truncating the
rest. This is similar to the change of variables into eigenvector coordinates, except
that we emphasize controllability and observability rather than characteristics of the
dynamics.
Substituting Tz into (12.1) gives

d
Tz = ATz + Bu, (12.6a)
dt
y = CTz + Du. (12.6b)

Finally, multiplying (12.6a) by T−1 yields

d
z = T−1 ATz + T−1 Bu, (12.7a)
dt
y = CTz + Du. (12.7b)

This results in the transformed equations

d
z = Âz + B̂u, (12.8a)
dt
y = Ĉz + Du, (12.8b)

where  = T−1 AT, B̂ = T−1 B, and Ĉ = CT. Note that when the columns of T are
orthonormal, the change of coordinates becomes

d
z = T∗ ATz + T∗ Bu, (12.9a)
dt
y = CTz + Du. (12.9b)

Gramians and Coordinate Transformations

The controllability and observability Gramians each establish an inner product on
state-space in terms of how controllable or observable a given state is, respectively.
As such, Gramians depend on the particular choice of coordinate system and will
transform under a change of coordinates. In the coordinate system z given by (12.5),
the controllability Gramian becomes

h d i 0 0 9 8 088962 4 0 8 bli h d li b C b id i i
 236 Data-Driven Fluid Mechanics Chapter 12

∫ ∞
eÂτ B̂B̂∗ e τ dτ,
∗
Ŵc = (12.10a)
∫0 ∞
−1 ATτ ∗ A∗ T−∗ τ
= eT T−1 BB∗ T−∗ eT dτ, (12.10b)
∫0 ∞
T−1 eAτ TT−1 BB∗ T−∗ T∗ eA τ T−∗ dτ,
∗
= (12.10c)
0
∫ ∞ 
τ
= T−1 eAτ BB∗ eA dτ T−∗, (12.10d)
0
= T−1 Wc T−∗ . (12.10e)

∗
Note that here we introduce T−∗ := T−1 = (T∗ )−1 . The observability Gramian
transforms similarly

Ŵo = T∗ Wo T, (12.11)

which is an exercise for the reader. Both Gramians transform as tensors (i.e., in terms
of the transform matrix T and its transpose, rather than T and its inverse), which is
consistent with them inducing an inner product on state-space.

Simple Rescaling
This example, modified from Moore (1981), demonstrates the ability to balance a
system through a change of coordinates. Consider the system

d x1 0 x1 10
       −3 
−1
= + u, (12.12a)
dt x2 0 −10 x2 103
 x1
 
y = 103 10−3 . (12.12b)

x2

In this example, the first state x1 is barely controllable, while the second state is barely
observable. However, under the change of coordinates z1 = 103 x1 and z2 = 10−3 x2 ,
the system becomes balanced:

d z1 0 z1 1
      
−1
= + u, (12.13a)
dt z2 0 −10 z2 1
 z1
 
y= 1 1 . (12.13b)

z2

In this example, the coordinate change simply rescales the state x. For instance,
it may be that the first state had units of millimeters while the second state had
units of kilometers. Writing both states in meters balances the dynamics; that is, the
controllability and observability Gramians are equal and diagonal.

h d i 0 0 9 8 088962 4 0 8 bli h d li b C b id i i
 Brunton: Methods for System Identification 237

12.1.3 Balancing Transformations

Now we are ready to derive the balancing coordinate transformation T that makes the
controllability and observability Gramians equal and diagonal:

Ŵc = Ŵo = Σ. (12.14)

First, consider the product of the Gramians from (12.10) and (12.11):

Ŵc Ŵo = T−1 Wc Wo T. (12.15)

Plugging in the desired Ŵc = Ŵo = Σ yields

T−1 Wc Wo T = Σ2 =⇒ Wc Wo T = TΣ2 . (12.16)

The latter expression in (12.16) is the equation for the eigendecomposition of

Wc Wo , the product of the Gramians in the original coordinates. Thus, the balancing
transformation T is related to the eigendecomposition of Wc Wo . Equation 12.16 is
valid for any scaling of the eigenvectors, and the correct rescaling must be chosen to
exactly balance the Gramians. In other words, there are many such transformations
T that make the product Ŵc Ŵo = Σ2 , but where the individual Gramians are not
equal (for example diagonal Gramians Ŵc = Σ c and Ŵo = Σ o will satisfy (12.16) if
Σ c Σ o = Σ2 ).
Below, we will introduce the matrix S = T−1 to simplify notation.

Scaling Eigenvectors for the Balancing Transformation

To find the correct scaling of eigenvectors to make Ŵc = Ŵo = Σ, first consider the
simplified case of balancing the first diagonal element of Σ. Let ξu denote the unscaled
first column of T, and let ηu denote the unscaled first row of S = T−1 . Then

η u Wc ηu∗ = σc , (12.17a)
ξu∗ Wo ξu = σo . (12.17b)

The first element of the diagonalized controllability Gramian is thus σc , while the first
element of the diagonalized observability Gramian is σo . If we scale the eigenvector
ξu by σs , then the inverse eigenvector η u is scaled by σs−1 . Transforming via the new
scaled eigenvectors ξ s = σs ξu and η s = σs−1 ηu , yields

η s Wc η s∗ = σs−2 σc , (12.18a)
ξ s∗ Wo ξ s = σs2 σo . (12.18b)

Thus, for the two Gramians to be equal,

 1/4
σc

σs−2 σc = σs2 σo =⇒ σs = . (12.19)
σo
To balance every diagonal entry of the controllability and observability Gramians,
we first consider the unscaled eigenvector transformation Tu from (12.16); the
subscript u simply denotes unscaled. As an example, we use the standard scaling in

h d i 0 0 9 8 088962 4 0 8 bli h d li b C b id i i
 238 Data-Driven Fluid Mechanics Chapter 12

most computational software so that the columns of Tu have unit norm. Then both
Gramians are diagonalized, but are not necessarily equal:

u Wc Tu = Σ c ,
T−1 −∗
(12.20a)
T∗u Wo Tu = Σo . (12.20b)

The scaling that exactly balances these Gramians is then given by Σ s = Σ1/4
c Σo
−1/4
.
Thus, the exact balancing transformation is given by
T = Tu Σ s . (12.21)
It is possible to directly confirm that this transformation balances the Gramians:
1/2 1/2
(Tu Σ s )−1 Wc (Tu Σ s )−∗ = Σ−1
s Tu Wc Tu Σ s = Σ s Σ c Σ s = Σ c Σ o ,
−1 −∗ −1 −1 −1
(12.22a)
(Tu Σ s )∗ Wo (Tu Σ s ) = Σ s T∗u Wo Tu Σ s = Σ s Σ o Σ s = Σ1/2 1/2
c Σo . (12.22b)
The manipulations rely on the fact that diagonal matrices commute, so that Σ c Σ o =
Σ o Σ c , etc.

Example of the Balancing Transform and Gramians

Before confronting the practical challenges associated with accurately and efficiently
computing the balancing transformation, it is helpful to consider an illustrative
example.
The following example illustrates the balanced realization for a two-dimensional
system. First, we generate a system and compute its balanced realization, along with
the Gramians for each system. Next, we visualize the Gramians of the unbalanced and
balanced systems in Figure 12.2.
To visualize the Gramians in Figure 12.2, we first recall that the distance the system
can go in a direction x with a unit actuation input is given by x ∗ Wc x. Thus, the
controllability Gramian may be visualized by plotting W1/2 c x for x on a sphere with
k x k = 1. The observability Gramian may be similarly visualized.
In this example, we see that the most controllable and observable directions may
not be well aligned. However, by a change of coordinates, it is possible to find a
new direction that is the most jointly controllable and observable. It is then possible
to represent the system in this one-dimensional subspace, while still capturing a
significant portion of the input–output energy. If the red and blue Gramians were
exactly perpendicular, so that the most controllable direction was the least observable
direction, and vice versa, then the balanced Gramian would be a circle. In this case,
there is no preferred state direction, and both directions are equally important for the
input–output behavior.

12.1.4 Balanced Truncation

We have now shown that it is possible to define a change of coordinates so that the
controllability and observability Gramians are equal and diagonal. Moreover, these
new coordinates may be ranked hierarchically in terms of their joint controllability

h d i 0 0 9 8 088962 4 0 8 bli h d li b C b id i i
 Brunton: Methods for System Identification 239

Wo

Wc
k xk = 1 Wb

x2

x1

Figure 12.2 Illustration of balancing transformation on Gramians. The reachable set with unit
1/2
control input is shown in red, given by Wc x for k x k = 1. The corresponding observable set
is shown in blue. Under the balancing transformation T, the Gramians are equal, shown in
purple. From Brunton and Kutz (2019), reproduced with permission of the Licensor through
PLSclear.

and observability. It may be possible to truncate these coordinates and keep only the
most controllable/observable directions, resulting in a ROM that faithfully captures
input–output dynamics.
Given the new coordinates z = T−1 x ∈ Rn , it is possible to define a reduced-order
state x̃ ∈ Rr as
 z1 
 . 
 
 ..  
  x̃
 z 
 
z =  r  (12.23)
 zr+1 
 .. 
 
 . 
 zn 
 
 
in terms of the first r most controllable and observable directions. If we partition the
balancing transformation T and inverse transformation S = T−1 into the first r modes
to be retained and the last n − r modes to be truncated,

Φ∗
 
T= Ψ Tt , S= , (12.24)
 
St

h d i 0 0 9 8 088962 4 0 8 bli h d li b C b id i i
 240 Data-Driven Fluid Mechanics Chapter 12

then it is possible to rewrite the transformed dynamics in (12.7) as

d Φ∗ AΨ Φ∗ ATt
      ∗ 
x̃ x̃ ΦB
= + u, (12.25a)
dt zt St AΨ St ATt zt St B
 
 x̃
y= + Du. (12.25b)

CΨ CTt
zt

In balanced truncation, the state z t is simply truncated (i.e., discarded and set equal to
zero), and only the x̃ equations remain:

d
x̃ = Φ∗ AΨx̃ + Φ∗ Bu, (12.26a)
dt
y = CΨx̃ + Du. (12.26b)

Only the first r columns of T and S∗ = T−∗ are required to construct Ψ and Φ, and
thus computing the entire balancing transformation T is unnecessary. The computation
of Ψ and Φ without T will be discussed in the following sections. A key benefit of
balanced truncation is the existence of upper and lower bounds on the error of a given
order truncation:
n
Õ
Upper bound: kG − Gr k∞ ≤ 2 σj , (12.27a)
j=r+1

Lower bound: kG − Gr k∞ > σr+1, (12.27b)

where σj is the jth diagonal entry of the balanced Gramians. The diagonal entries of
Σ are also known as Hankel singular values.

12.1.5 Computing Balanced Realizations

In the previous section, we demonstrated the feasibility of obtaining a coordinate
transformation that balances the controllability and observability Gramians. However,
the computation of this balancing transformation is nontrivial, and significant work
has gone into obtaining accurate and efficient methods, starting with Moore (1981),
and continuing with Lall et al. (2002), Willcox and Peraire (2002), and Rowley (2005).
For an excellent and complete treatment of balanced realizations and model reduction,
see Antoulas (2005).
In practice, computing the Gramians Wc and Wo and the eigendecomposition of
the product Wc Wo in (12.16) may be prohibitively expensive for high-dimensional
systems. Instead, the balancing transformation may be approximated from impulse-
response data, utilizing the singular value decomposition (SVD) for efficient extrac-
tion of the most relevant subspaces.
We will first show that Gramians may be approximated via a snapshot matrix from
impulse response experiments/simulations. Then, we will show how the balancing
transformation may be obtained from this data.

h d i 0 0 9 8 088962 4 0 8 bli h d li b C b id i i
 Brunton: Methods for System Identification 241

Empirical Gramians
In practice, computing Gramians via the Lyapunov equation is computationally
expensive, with computational complexity of O(n3 ). Instead, the Gramians may
be approximated by full-state measurements of the discrete-time direct and adjoint
systems:

direct: xk+1 = Ad xk + Bd uk , (12.28a)

adjoint: xk+1 = A∗d xk + C∗d yk . (12.28b)

Equation (12.28a) is the discrete-time dynamic update equation, and (12.28b) is the
adjoint equation. The matrices Ad , Bd , and Cd are the discrete-time system matrices.
Note that the adjoint equation is generally nonphysical, and must be simulated;
thus the methods here apply to analytical equations and simulations, but not to
experimental data. An alternative formulation that does not rely on adjoint data, and
therefore generalizes to experiments, will be provided in Section 12.2.
Computing the impulse response of the direct and adjoint systems yields the
following discrete-time snapshot matrices:

 Cd 
 
h i  Cd Ad 
C d = Bd Ad Bd ··· Am c −1
Bd Od =  .. . (12.29)
 
d 
 . 

C Amo −1 
 d d 

Note that when mc = n, C d is the discrete-time controllability matrix and when

mo = n, O d is the discrete-time observability matrix; however, we generally consider
mc , mo  n. These matrices may also be obtained by sampling the continuous-time
direct and adjoint systems at a regular interval ∆t.
It is now possible to compute empirical Gramians that approximate the true
Gramians without solving a Lyapunov equation:

Wc ≈ Wec = C d C ∗d , (12.30a)
Wo ≈ Weo = O∗d O d . (12.30b)

The empirical Gramians essentially comprise a Riemann sum approximation of the

integral in the continuous-time Gramians, which becomes exact as the time step of
the discrete-time system becomes arbitrarily small and the duration of the impulse
response becomes arbitrarily large. In practice, the impulse response snapshots should
be collected until the lightly damped transients die out. The method of empirical
Gramians is quite efficient, and is widely used (Moore 1981, Lall et al. 1999, Lall
et al. 2002, Willcox & Peraire 2002, Rowley 2005). Note that p adjoint impulse
responses are required, where p is the number of outputs. This becomes intractable
when there are a large number of outputs (e.g., full state measurements), motivating
the output projection below.

h d i 0 0 9 8 088962 4 0 8 bli h d li b C b id i i
 242 Data-Driven Fluid Mechanics Chapter 12

Balanced POD
Instead of computing the eigendecomposition of Wc Wo , which is an n × n matrix, it
is possible to compute the balancing transformation via the SVD of the product of the
snapshot matrices,
Od C d, (12.31)
reminiscent of the method of snapshots (Sirovich 1987). This is the approach taken by
Rowley (2005).
First, define the generalized Hankel matrix as the product of the adjoint (O d ) and
direct (C d ) snapshot matrices from (12.29), for the discrete-time system:
 Cd 
 
 Cd Ad  h i
H = Od C d =  ..  B d A d B d · · · Am c −1
Bd (12.32a)
 
d

 . 

C Amo −1 
 d d 
c −1
 C B
 d d Cd Ad Bd · · · C d Amd
Bd 
 Cd Ad Bd Cd A2d Bd · · · C d Am c
Bd 
 
d
= .. .. .. .. . (12.32b)
. . . .

 
Cd Amo −1 Bd Cd Amo Bd · · · Cd Amc +mo −2 Bd 
 
 d d d 
Next, we factor H using the SVD:
Ṽ∗
  
 Σ̃ 0
H = UΣV = Ũ ∗
≈ ŨΣ̃Ṽ∗ . (12.33)

Ut
0 Σt V∗t
For a given desired model order r  n, only the first r columns of U and V are
retained, along with the first r ×r block of Σ; the remaining contribution from Ut Σt V∗t
may be truncated. This yields a bi-orthogonal set of modes given by
−1/2
Direct modes: Ψ = C d ṼΣ̃ , (12.34a)
−1/2
Adjoint modes: Φ = O∗d ŨΣ̃ . (12.34b)
The direct modes Ψ ∈ Rn×r and adjoint modes Φ ∈ Rn×r are bi-orthogonal, Φ∗ Ψ =
Ir×r , and Rowley (2005) showed that they establish the change of coordinates that
balance the truncated empirical Gramians. Thus, Ψ approximates the first r-columns
of the full n × n balancing transformation, T, and Φ∗ approximates the first r-rows of
the n × n inverse balancing transformation, S = T−1 .
Now, it is possible to project the original system onto these modes, yielding a
balanced ROM of order r:
à = Φ∗ Ad Ψ, (12.35a)
B̃ = Φ Bd ,
∗
(12.35b)
C̃ = Cd Ψ. (12.35c)
It is possible to compute the reduced system dynamics in (12.35a) without having
direct access to Ad . In some cases, Ad may be exceedingly large and unwieldy, and

h d i 0 0 9 8 088962 4 0 8 bli h d li b C b id i i
 Brunton: Methods for System Identification 243

instead it is only possible to evaluate the action of this matrix on an input vector.
For example, in many modern fluid dynamics codes the matrix Ad is not actually
represented, but because it is sparse, it is possible to implement efficient routines to
multiply this matrix by a vector.
It is important to note that the ROM in (12.35) is formulated in discrete time, as it is
based on discrete-time empirical snapshot matrices. However, it is simple to obtain the
corresponding continuous-time system. In this example, D is the same in continuous
time and discrete time, and in the full-order and ROMs.
Note that a BPOD model may not exactly satisfy the upper bound from balanced
truncation (see (12.27)) due to errors in the empirical Gramians.

Output Projection
Often, in high-dimensional simulations, we assume full-state measurements, so that
p = n is exceedingly large. To avoid computing p = n adjoint simulations, it is possible
instead to solve an output-projected adjoint equation (Rowley 2005),

xk+1 = A∗d xk + C∗d Ũy, (12.36)

where Ũ is a matrix containing the first r singular vectors of C d . Thus, we first identify
a low-dimensional POD subspace Ũ from a direct impulse response, and then only
perform adjoint impulse response simulations by exciting these few POD coefficient
measurements. More generally, if y is high-dimensional but does not measure the full-
state, it is possible to use a POD subspace trained on the measurements, given by the
first r singular vectors Ũ of Cd C d . Adjoint impulse responses may then be performed
in these output POD directions.

Data Collection and Stacking

The powers mc and mo in (12.32) signify that data must be collected until the matrices
C d and O∗d are full rank after which the controllable/observable subspaces have been
sampled. Unless we collect data until transients decay, the true Gramians are only
approximately balanced. Instead, it is possible to collect data until the Hankel matrix is
full rank, balance the resulting model, and then truncate. This more efficient approach
is developed in Tu and Rowley (2012) and Luchtenburg and Rowley (2011).
The snapshot matrices in (12.29) are generated from impulse response simulations
of the direct (12.28a) and adjoint (12.36) systems. These time-series snapshots are
then interleaved to form the snapshot matrices.

Historical Note
The balanced POD method described earlier originated with the seminal work of
Moore (1981), which provided a data-driven generalization of the minimal realization
theory of Ho and Kalman (1965). Until then, minimal realizations were defined in
terms of idealized controllable and observable subspaces, which neglected the subtlety
of degrees of controllability and observability.

h d i 0 0 9 8 088962 4 0 8 bli h d li b C b id i i
 244 Data-Driven Fluid Mechanics Chapter 12

Moore’s paper introduced a number of critical concepts that bridged the gap from
theory to reality. First, he established a connection between principal component
analysis (PCA) and Gramians, showing that information about degrees of control-
lability and observability may be mined from data via the SVD. Next, Moore showed
that a balancing transformation exists that makes the Gramians equal, diagonal, and
hierarchically ordered by balanced controllability and observability; moreover, he
provides an algorithm to compute this transformation. This sets the stage for principled
model reduction, whereby states may be truncated based on their joint controllability
and observability. Moore further introduced the notion of an empirical Gramian,
although he didn’t use this terminology. He also realized that computing Wc and
Wo directly is less accurate than computing the SVD of the empirical snapshot
matrices from the direct and adjoint systems, and he avoided directly computing the
eigendecomposition of Wc Wo by using these SVD transformations. Lall et al. (2002)
generalized this theory to nonlinear systems.
One drawback of Moore’s approach is that he computed the entire n × n balancing
transformation, which is not suitable for exceedingly high-dimensional systems.
Willcox and Peraire (2002) generalized the method to high-dimensional systems,
introducing a variant based on the rank-r decompositions of Wc and Wo obtained
from the direct and adjoint snapshot matrices. It is then possible to compute the
eigendecomposition of Wc Wo using efficient eigenvalue solvers without ever actually
writing down the full n × n matrices. However, this approach has the drawback of
requiring as many adjoint impulse response simulations as the number of output
equations, which may be exceedingly large for full-state measurements. Rowley
(2005) addressed this issue by introducing the output projection, discussed earlier,
which limits the number of adjoint simulations to the number of relevant POD modes
in the data. He also showed that it is possible to use the eigendecomposition of the
product O d C d The product O d C d is often smaller, and these computations may be
more accurate.
It is interesting to note that a nearly equivalent formulation was developed 20
years earlier in the field of system identification. The so-called ERA, introduced by
Juang and Pappa (1985), obtains equivalent balanced models without the need for
adjoint data, making it useful for system identification in experiments. This connection
between ERA and BPOD was established by Ma et al. (2011).

12.2 System Identification

In contrast to model reduction, where the system model (A, B, C, D) was known,
system identification is purely data-driven. System identification may be thought of
as a form of machine learning, where an input–output map of a system is learned
from training data in a representation that generalizes to data that was not in the
training set. There is a vast literature on methods for system identification (Juang
1994, Ljung 1999), and many of the leading methods are based on a form of dynamic
regression that fits models based on data, such as the DMD. For this section, we

h d i 0 0 9 8 088962 4 0 8 bli h d li b C b id i i
 Brunton: Methods for System Identification 245

consider the ERA and observer-Kalman filter identification (OKID) methods because
of their connection to balanced model reduction (Moore 1981, Rowley 2005, Ma
et al. 2011, Tu et al. 2014) and their successful application in complex systems such
as vibration control of aerospace structures and closed-loop flow control (Bagheri,
Hoepffner, Schmid & Henningson 2009, Bagheri, Brandt & Henningson 2009,
Illingworth et al. 2010). The ERA/OKID procedure is also applicable to multiple-
input, multiple-output (MIMO) systems. Other methods include the autoregressive-
moving average (ARMA) and autoregressive moving average with exogenous inputs
(ARMAX) models (Whittle 1951, Box et al. 2015), the nonlinear autoregressive mov-
ing average with exogenous inputs (NARMAX) (Billings 2013) model, and the SINDy
method.

12.2.1 Eigensystem Realization Algorithm

The ERA produces low-dimensional linear input–output models from sensor mea-
surements of an impulse response experiment, based on the “minimal realization”
theory of Ho and Kalman (1965). The modern theory was developed to identify
structural models for various spacecraft (Juang & Pappa 1985), and it has been shown
by Ma et al. (2011) that ERA models are equivalent to BPOD models.2 However,
ERA is based entirely on impulse response measurements and does not require prior
knowledge of a model. ERA may be equally well applied to model high-dimensional
data, such as a fluid vector field, or to model low-dimensional data, such as a scalar
time-series of the lift coefficient. In the latter case, where the model is based on partial
or incomplete measurements of a high-dimensional state, the effect of latent or hidden
variables is included by considering time delays of the measurement. The following
mathematical treatment encompasses both cases.
We consider a discrete-time system:
xk+1 = Ad xk + Bd uk , (12.37a)
yk = Cd x k + Dd u k . (12.37b)
A discrete-time delta function input in the actuation u,
I, k = 0,

ukδ , u δ (k∆t) = (12.38)
0, k = 1, 2, 3, · · · ,
gives rise to a discrete-time impulse response in the sensors y:
Dd , k = 0,

δ δ
yk , y (k∆t) = (12.39)
Cd Ak−1
d
Bd , k = 1, 2, 3, · · · .
In an experiment or simulation, typically q impulse responses are performed, one
for each of the q separate input channels. The output responses are collected for each
impulsive input, and at a given time step k, the output vector in response to the jth

2 BPOD and ERA models both balance the empirical Gramians and approximate balanced
truncation (Moore 1981) for high-dimensional systems, given a sufficient volume of data.

h d i 0 0 9 8 088962 4 0 8 bli h d li b C b id i i
 246 Data-Driven Fluid Mechanics Chapter 12

impulsive input will form the jth column of ykδ . Thus, each of the ykδ is a p × q matrix
CAk−1 B. Note that the system matrices (A, B, C, D) don’t actually need to exist, as the
following method is purely data-driven.
The Hankel matrix H from (12.32) is formed by stacking shifted time-series of
impulse response measurements into a matrix, as in the Hankel alternative view of the
Koopman (HAVOK) method (Brunton et al. 2017):

 yδ y2δ ··· ymδ 

 1 c 
 yδ y3δ ··· δ
ymc +1 
 2
H=  . (12.40a)

.. .. ..
 ..
 
. . .


 δ δ δ 
 y mo y m
 o +1
··· ym c +mo −1 

c −1
 C B
 d d Cd Ad Bd · · · C d Am
d
Bd


 Cd Ad Bd Cd A2d Bd · · · mc
Cd Ad Bd
 
=  . (12.40b)

.. .. .. ..
. . . .

 
Cd Amo −1 Bd mc +mo −2
C d Am
 o

 d d
Bd ··· Cd Ad Bd 

The matrix H may be constructed purely from measurements y δ , without separately

constructing O d and C d . Thus, we do not need access to adjoint equations.
Taking the SVD of the Hankel matrix yields the dominant temporal patterns in the
time-series data:

Ṽ∗
  
 Σ̃ 0
H = UΣV = Ũ ∗
≈ ŨΣ̃Ṽ∗ . (12.41)

Ut
0 Σt V∗t

The small singular values in Σt are truncated, and only the first r singular values in Σ̃
are retained. The columns of Ũ and Ṽ are eigen-time-delay coordinates.
Until this point, the ERA algorithm closely resembles the BPOD procedure from
Section 12.1. However, we do not require direct access to O d and C d or the system
(A, B, C, D) to construct the direct and adjoint balancing transformations. Instead, with
sensor measurements from an impulse response experiment, it is also possible to create
a second, shifted Hankel matrix H0:

 y2
 y3δ ··· δ
ym c +1 

 yδ y4 δ
··· δ
ymc +2 
 3
H = .
0
.. .. ..  (12.42a)
 ..

. . . 
 δ δ δ

 mo +1 ymo +2 · · · ymc +mo 
y 
C A B
 d d d Cd A2d Bd ··· C d Am
d
c
Bd 
+1
 Cd A2d Bd Cd A3d Bd m
··· Cd Ad c Bd 
 
=
 .. .. .. ..  = O d AC d . (12.42b)
. . . .

 
 m m +1
C d A o B d C d A o B d · · · C d A c o B d m +m −1 
 d d d 

h d i 0 0 9 8 088962 4 0 8 bli h d li b C b id i i
 Brunton: Methods for System Identification 247

Based on the matrices H and H0, we are able to construct a ROM as follows:
−1/2 −1/2
à = Σ̃ Ũ∗ H0ṼΣ̃ , (12.43a)
 
1/2 I 0
B̃ = Σ̃ Ṽ∗ p , (12.43b)
0 0
 
Iq 0 1/2
C̃ = ŨΣ̃ . (12.43c)
0 0

Here I p is the p × p identity matrix, which extracts the first p columns, and Iq is the
q×q identity matrix, which extracts the first q rows. Thus, we express the input–output
dynamics in terms of a reduced system with a low-dimensional state x̃ ∈ Rr :

x̃k+1 = Ãx̃k + B̃u, (12.44a)

y = C̃x̃k . (12.44b)

H and H0 are constructed from impulse response simulations/experiments, without

the need for storing direct or adjoint snapshots, as in other balanced model reduction
techniques. However, if full-state snapshots are available, for example, by collecting
velocity fields in simulations or PIV experiments, it is then possible to construct direct
modes. These full-state snapshots form C d , and modes can be constructed by
−1/2
Ψ = C d ṼΣ̃ . (12.45)

These modes may then be used to approximate the full-state of the high-dimensional
system from the low-dimensional model in (12.44) by

x ≈ Ψx̃. (12.46)

If enough data is collected when constructing the Hankel matrix H, then ERA
balances the empirical controllability and observability Gramians, O d O∗d and C ∗d C d .
However, if less data is collected, so that lightly damped transients do not have time to
decay, then ERA will only approximately balance the system. It is instead possible to
collect just enough data so that the Hankel matrix H reaches numerical full rank (i.e.,
so that remaining singular values are below a threshold tolerance), and compute an
ERA model. The resulting ERA model will typically have a relatively low order, given
by the numerical rank of the controllability and observability subspaces. It may then
be possible to apply exact balanced truncation to this smaller model, as is advocated
in Tu and Rowley (2012) and Luchtenburg and Rowley (2011).

12.2.2 Observer Kalman Filter Identification

OKID was developed to complement the ERA for lightly damped experimental sys-
tems with noise (Juang et al. 1991). In practice, performing isolated impulse response
experiments is challenging, and the effect of measurement noise can contaminate
results. Moreover, if there is a large separation of timescales, then a tremendous

h d i 0 0 9 8 088962 4 0 8 bli h d li b C b id i i
 248 Data-Driven Fluid Mechanics Chapter 12

1
u 2
u 0.5 u = δ(t)
0
-2
0
0 20 40 60 80 100
OKID 0 20 40 60 80 100
8
6 2
y 4
2 y 1
0
-2
-4 0
0 20 40 60 80 100 0 20 40 60 80 100
k k

Figure 12.3 Schematic overview of OKID procedure. The output of OKID is an impulse
response that can be used for system identification via ERA. From Brunton and Kutz (2019),
reproduced with permission of the Licensor through PLSclear.

amount of data must be collected to use ERA. This section poses the general problem
of approximating the impulse response from arbitrary input–output data (Figure 12.3).
Typically, one would identify ROMs according to the following general procedure:

1. Collect the output in response to a pseudorandom input.

2. This information is passed through the OKID algorithm to obtain the de-noised
linear impulse response.
3. The impulse response is passed through the ERA to obtain a reduced-order state-
space system.

The output yk in response to a general input signal uk , for zero initial condition
x0 = 0, is given by

y0 = D d u 0, (12.47a)
y1 = C d B d u 0 + D d u 1, (12.47b)
y2 = C d A d B d u 0 + C d B d u 1 + D d u 2, (12.47c)
···
d Bd u0 + C d Ad Bd u1 + · · · + Cd B d u k−1 + D d u k . (12.47d)
yk = Cd Ak−1 k−2

Note that there is no C term in the expression for y0 since there is zero initial condition
x0 = 0. This progression of measurements yk may be further simplified and expressed
in terms of impulse response measurements ykδ :

 u0
 u1 ··· u m 
0 u0 ··· u m−1 
ym = y0δ y1δ δ 
..  . (12.48)
   
y0 y1 ··· ··· ym . .. ..
| {z } | {z }  .. . . . 
S Sδ ··· u0 
0 0

| {z }
B

It is often possible to invert the matrix of control inputs, B, to solve for the Markov
parameters S δ . However, B may either be un-invertible, or inversion may be ill-
conditioned. In addition, B is large for lightly damped systems, making inversion

h d i 0 0 9 8 088962 4 0 8 bli h d li b C b id i i
 Brunton: Methods for System Identification 249

computationally expensive. Finally, noise is not optimally filtered by simply inverting

B to solve for the Markov parameters.
The OKID method addresses each of these issues. Instead of the original discrete-
time system, we now introduce an optimal observer system:
x̂k+1 = Ad x̂k + K f (yk − yˆk ) + Bd uk , (12.49a)
ŷk = Cd x̂k + Dd uk , (12.49b)
which may be rewritten as:
 
 uk
x̂k+1 = (Ad − K f Cd ) x̂k + Bd − K f Dd , . (12.50)

Kf
| {z } | {z } yk
Ā d B̄ d

Recall from above that if the system is observable, it is possible to place the poles
of Ad − K f Cd anywhere we like. However, depending on the amount of noise in the
measurements, the magnitude of process noise, and uncertainty in our model, there
are optimal pole locations that are given by the Kalman filter. We may now solve for
δ
the observer Markov parameters S̄ of the system in (12.50) in terms of measured
inputs and outputs according to the following algorithm from Juang et al. (1991):
1. Choose the number of observer Markov parameters to identify, l.
2. Construct the data matrices:
S = y 0 y 1 · · · yl · · · y m , (12.51)
 

 u0 u1 · · · ul · · · u m 

 0 v0 · · · vl−1 · · · vm−1 
V=. .. .. .. .. ..  , (12.52)
 
 .. . . . . . 

0
 0 ··· v0 · · · vm−l 
T
where vi = uTi yiT .


The matrix V resembles B, except that has been augmented with the outputs
yi . In this way, we are working with a system that is augmented to include a
Kalman filter. We are now identifying the observer Markov parameters of the
δ δ
augmented system, S̄ , using the equation S = S̄ V. It will be possible to
identify these observer Markov parameters from the data and then extract the
impulse response (Markov parameters) of the original system.
δ δ δ
3. Identify the matrix S̄ of observer Markov parameters by solving S = S̄ V for S̄
using the right pseudo-inverse of V (i.e., SVD).
δ
4. Recover system Markov parameters, S δ , from the observer Markov parameters, S̄ :
δ
(a) Order the observer Markov parameters S̄ as follows:
δ
S̄0 = D, (12.53)
δ δ δ
h i
S̄ k = (S̄ )(1)
k
(S̄ )(2)
k
for k ≥ 1, (12.54)
δ δ δ
where (S̄ )(1)
k
∈ Rq×p , (S̄ )(2)
k
∈ Rq×q , and y0δ = S̄0 = D.

h d i 0 0 9 8 088962 4 0 8 bli h d li b C b id i i
 250 Data-Driven Fluid Mechanics Chapter 12

(b) Reconstruct system Markov parameters:

k
δ δ
ykδ = (S̄ )(1) δ
Õ
k
+ (S̄ )(2)
i yk−i for k ≥ 1. (12.55)
i=1

Thus, the OKID method identifies the Markov parameters of a system augmented with
an asymptotically stable Kalman filter. The system Markov parameters are extracted
from the observer Markov parameters by (12.55). These system Markov parameters
approximate the impulse response of the system, and may be used directly as inputs
to the ERA algorithm.
ERA/OKID has been widely applied across a range of system identification tasks,
including to identify models of aeroelastic structures and fluid dynamic systems.
There are numerous extensions of the ERA/OKID methods. For example, there are
generalizations for linear parameter varying (LPV) systems and systems linearized
about a limit cycle.

12.3 Sparse Identification of Nonlinear Dynamics

Discovering dynamical systems models from data is a central challenge in mathemat-

ical physics, with a rich history going back at least as far as the time of Kepler and
Newton and the discovery of the laws of planetary motion. Historically, this process
relied on a combination of high-quality measurements and expert intuition. With vast
quantities of data and increasing computational power, the automated discovery of
governing equations and dynamical systems is a new and exciting scientific paradigm.
Typically, the form of a candidate model is either constrained via prior knowledge
of the governing equations, as in Galerkin projection (Noack et al. 2003, Rowley
et al. 2004, Schlegel et al. 2004, Carlberg et al. 2011, Noack et al. 2011, Wang
et al. 2012, Balajewicz et al. 2013, Carlberg et al. 2017), or a handful of heuristic
models are tested and parameters are optimized to fit data. Alternatively, best-fit
linear models may be obtained using DMD or ERA. Simultaneously identifying
the nonlinear structure and parameters of a model from data is considerably more
challenging, as there are combinatorially many possible model structures.
The SINDy algorithm (Brunton et al. 2016a) bypasses the intractable combinatorial
search through all possible model structures, leveraging the fact that many dynamical
systems
d
x = f(x) (12.56)
dt

have dynamics f with only a few active terms in the space of possible right-hand side
functions; for example, the Lorenz equations only have a few linear and quadratic
interaction terms per equation.

h d i 0 0 9 8 088962 4 0 8 bli h d li b C b id i i
 Brunton: Methods for System Identification 251

We then seek to approximate f by a generalized linear model,

p
Õ
f(x) ≈ θ k (x)ξk = Θ(x)ξ, (12.57)
k=1

with the fewest possible nonzero terms in ξ. It is then possible to solve for the relevant
terms that are active in the dynamics using sparse regression (Tibshirani 1996, Zou &
Hastie 2005, Hastie et al. 2009, James et al. 2013), which penalizes the number of
terms in the dynamics and scales well to large problems.
First, time-series data are collected from (12.56) and formed into a data matrix:
T
X = x(t1 ) x(t2 ) · · · x(tm ) . (12.58)


A similar matrix of derivatives is formed:

T
Û = xÛ (t1 ) xÛ (t2 ) . (12.59)

X · · · xÛ (tm )

In practice, this may be computed directly from the data in X; for noisy data,
the total-variation regularized derivative tends to provide numerically robust deriva-
tives (Chartrand 2011). Alternatively, it is possible to formulate the SINDy algorithm
for discrete-time systems xk+1 = F(xk ), as in the DMD algorithm, and avoid
derivatives entirely.
A library of candidate nonlinear functions Θ(X) may be constructed from the data
in X:

Θ(X) = 1 X X2 · · · Xd · · · sin(X) · · · . (12.60)

 

Here, the matrix Xd denotes a matrix with column vectors given by all possible time-
series of dth degree polynomials in the state x. In general, this library of candidate
functions is only limited by one’s imagination.
The dynamical system in (12.56) may now be represented in terms of the data
matrices in (12.59) and (12.60) as
Û = Θ(X)Ξ.
X (12.61)

Each column ξk in Ξ is a vector of coefficients determining the active terms in the kth
row in (12.56). A parsimonious model will provide an accurate model fit in (12.61)
with as few terms as possible in Ξ. Such a model may be identified using a convex
`1 -regularized sparse regression:

ξk = argminξ0 k X
Û k − Θ(X)ξ 0 k2 + λkξ 0 k1 .
k k (12.62)
k

Here, XÛ k is the kth column of X, Û and λ is a sparsity-promoting knob. Sparse

regression, such as the LASSO (Tibshirani 1996) or the sequential thresholded least-
squares (STLS) algorithm used in SINDy (Brunton, Proctor & Kutz 2016a), improves
the numerical robustness of this identification for noisy overdetermined problems, in
contrast to earlier methods (Wang et al. 2011) that used compressed sensing (Donoho
2006, Candès, 2006, Candès et al. 2006a, 2006b, Candès & Tao 2006, Baraniuk 2007,
Tropp & Gilbert 2007). We advocate the STLS to select active terms.

h d i 0 0 9 8 088962 4 0 8 bli h d li b C b id i i
 252 Data-Driven Fluid Mechanics Chapter 12

Figure 12.4 Schematic of the sparse identification of nonlinear dynamics (SINDy)

algorithm (Brunton et al. 2016a). Parsimonious models are selected from a library of candidate
nonlinear terms using sparse regression. This library Θ(X) may be constructed purely from
measurement data. Reproduced from Brunton et al. (2016a).

The sparse vectors ξk may be synthesized into a dynamical system:

xÛk = Θ(x)ξk . (12.63)

Note that xk is the kth element of x, and Θ(x) is a row vector of symbolic functions of
x, as opposed to the data matrix Θ(X). Figure 12.4 shows how SINDy may be used to
discover the Lorenz equations from data.
The result of the SINDy regression is a parsimonious model that includes only the
most important terms required to explain the observed behavior. The sparse regression
procedure used to identify the most parsimonious nonlinear model is a convex
procedure. The alternative approach, which involves regression onto every possible
sparse nonlinear structure, constitutes an intractable brute-force search through the
combinatorially many candidate model forms. SINDy bypasses this combinatorial
search with modern convex optimization and machine learning. It is interesting to
note that for discrete-time dynamics, if Θ(X) consists only of linear terms, and if we
remove the sparsity promoting term by setting λ = 0, then this algorithm reduces
to the DMD (Rowley et al. 2009, Schmid 2010, Tu et al. 2014, Kutz et al. 2016). If a
least-squares regression is used, as in DMD, then even a small amount of measurement
error or numerical roundoff will lead to every term in the library being active in the
dynamics, which is nonphysical. A major benefit of the SINDy architecture is the
ability to identify parsimonious models that contain only the required nonlinear terms,
resulting in interpretable models that avoid overfitting.

Applications, Extensions, and Connections to Fluid Dynamics

The SINDy algorithm has recently been applied to discover models in fluid dynam-
ics (El Sayed M et al. 2018, Loiseau & Brunton 2018, Loiseau et al. 2018, Loiseau
2019, Guan et al. 2020, Deng et al. 2020). Because of the high-state dimension,
SINDy models of fluids are typically based on POD coefficients or low-dimensional
measurements such as the lift and drag coefficients (Brunton et al. 2016a, Loiseau

h d i 0 0 9 8 088962 4 0 8 bli h d li b C b id i i
 Brunton: Methods for System Identification 253

Figure 12.5 Schematic overview of nonlinear model identification from high-dimensional data
using the sparse identification of nonlinear dynamics (SINDy) (Brunton et al. 2016a). This
procedure is modular, so that different techniques can be used for the feature extraction and
regression steps. In this example of flow past a cylinder, SINDy discovers the model of Noack
et al. (2003). Reproduced from Brunton et al. (2016a).

& Brunton 2018, Loiseau et al. 2018). Recent studies have also leveraged SINDy
for turbulence modeling (Beetham & Capecelatro 2020, Schmelzer et al. 2020).
Figure 12.5 illustrates the application of SINDy to the flow past a cylinder, where
the generalized mean-field model of Noack et al. (2003) was discovered from data.
SINDy has also been applied to identify models in nonlinear optics (Sorokina
et al. 2016), plasma physics (Dam et al. 2017), chemical reaction dynamics (Hoffmann
et al. 2019), numerical algorithms (Thaler et al. 2019), and structural modeling (Lai &
Nagarajaiah 2019), among others (Narasingam & Kwon 2018, de Silva et al. 2019, Pan
et al. 2020).
Because SINDy is formulated in terms of linear regression in a nonlinear library, it
is highly extensible. The SINDy framework has been recently generalized by Loiseau
and Brunton (Loiseau & Brunton 2018) to incorporate known physical constraints
and symmetries in the equations by implementing a constrained sequentially thresh-
olded least-squares optimization. In particular, energy-preserving constraints on the
quadratic nonlinearities in the Navier–Stokes equations were imposed to identify
fluid systems (Loiseau & Brunton 2018), where it is known that these constraints
promote stability (Majda & Harlim 2012, Balajewicz et al. 2013, Carlberg et al. 2017).
This work also showed that polynomial libraries are particularly useful for building
models of fluid flows in terms of POD coefficients, yielding interpretable models
that are related to classical Galerkin projection (Brunton et al. 2016a, Loiseau &
Brunton 2018). Loiseau et al. (2018) also demonstrated the ability of SINDy to
identify dynamical systems models of high-dimensional systems, such as fluid flows,
from a few physical sensor measurements, such as lift and drag measurements on
the cylinder in Figure 12.5. For actuated systems, SINDy has been generalized to
include inputs and control (Brunton et al. 2016b), and these models are highly effective
for model predictive control (Kaiser et al. 2018). It is also possible to extend the
SINDy algorithm to identify dynamics with rational function nonlinearities (Mangan

h d i 0 0 9 8 088962 4 0 8 bli h d li b C b id i i
 254 Data-Driven Fluid Mechanics Chapter 12

et al. 2016), integral terms (Schaeffer & McCalla 2017), and based on highly corrupt
and incomplete data (Tran & Ward 2016). SINDy was also recently extended to
incorporate information criteria for objective model selection (Mangan et al. 2017),
and to identify models with hidden variables using delay coordinates (Brunton
et al. 2017). Finally, the SINDy framework was generalized to include partial
derivatives, enabling the identification of partial differential equation models (Rudy
et al. 2017, Schaeffer 2017).

h d i 0 0 9 8 088962 4 0 8 bli h d li b C b id i i
 13 Modern Tools for the Stability
Analysis of Fluid Flows
P. J. Schmid

The response behavior of a fluid flow to small perturbations is a key measure of

its dynamic properties and provides insight into the prevalence and dominance of
coherent structures and their evolution in time or space. This chapter gives a brief
introduction to the principal concepts of stability analysis of fluid flows, advocating
a general mathematical framework based on finite-time or frequency gains, coupled
to an optimization environment. This framework is particularly flexible in describing
complex fluid behavior characterized by multi-physics and multiscale mechanisms
interacting in a feedforward or feedback manner. Special cases of the same framework
yield more familiar stability and receptivity concepts based on linear-algebra tools.
Various implementation details will be discussed, and an example from computational
aeroacoustics will be presented to showcase the mathematical techniques and illustrate
the merits of the proposed approach.

13.1 Introduction

Stability analysis is a key discipline in fluid dynamics and is ubiquitous in the fluid
dynamics literature, either as a way of analyzing complex fluid behavior or as a means
to utilize it to manipulate intrinsic fluid motion. Instabilities are often postulated as the
driving force for pattern formation, for the rise of specific scales, or for the bifurcation
into a different flow regime.
Instabilities can be observed all around us. The flow patterns forming behind bluff
objects, the breakup of water jets and drops, the clustering of stars in rotating galaxies,
the formation of thermal plumes, the shaping of glaciers and stalactites/stalagmites in
caves, the sand ripples in river beds, and desert dunes fall among applications where
stability analysis of the governing equations has contributed to our understanding of
the dominant physical processes at play (Figure 13.1).

h d i 0 0 9 8 088962 4 0 9 bli h d li b C b id i i
 256 Data-Driven Fluid Mechanics Chapter 13

Buoyancy-driven

Vortical

Interfacial

Rotational

Shear-driven

Stratification-driven

Figure 13.1 Examples of instabilities.

As much as instabilities are responsible for the observable features in deformable

media around us, they are at the same time undesirable in many technological appli-
cations, or key players in natural phenomena. They often place limitations on the safe
and efficient operation of many fluid devices; in flames, they lead to incomplete and
nonstoichiometric combustion, increased soot formation and significant NOx output;

h d i 0 0 9 8 088962 4 0 9 bli h d li b C b id i i
 Schmid: Modern Tools for the Stability Analysis of Fluid Flows 257

instabilities in the inlet of a jet engine lead to non-smooth operation, while acoustic
instabilities lead to premature material fatigue; and strong magnetic instabilities are
one of the key reasons why nuclear fusion has not yet matured into a reliable technol-
ogy. In natural settings, instabilities in the atmospheric boundary layer are responsible
for weather abnormalities, and in lakes, rivers, and oceans they influence nutrient
transport.
The underlying mechanisms for instabilities can be rather multifaceted. Yet, it is
common to classify instabilities by their principal physical mechanism responsible for
the growth (or decay) of disturbances: buoyancy-driven, shock-induced, rotational,
magnetic, shear-driven, morphological, vortical, interfacial, thermal, reactive, acous-
tical, chemical, and so on. Instabilities have been studied intensively over the past
decades. Along with this categorization, each effect is characterized and parameterized
by a nondimensional number, such as the Reynolds number, Rayleigh number, Rossby
number, Weber number, Mach number, Damkoehler number, and so on. These num-
bers quantify the relative importance of the involved processes and act as bifurcation
parameters that, after passing a critical value, establish the presence of an instability.
Besides the key physical processes and their nondimensional number, the size of the
disturbance background or of an initial condition, necessary to induce an instability, is
important. Infinitesimal disturbances describe the early departure from an established
equilibrium state. They have the added mathematical advantage that linearization is
easily justified. The same is not true for finite-amplitude disturbances whose analysis
requires more sophisticated mathematical techniques. In what follows, we will outline
techniques for a general stability analysis, but will focus on common techniques for a
linear analysis.

Hydrodynamic Stability Theory: A Brief Look Back

Owing to its central position within the broad field of fluid mechanics, hydrodynamic
stability theory has a long history that is worth recalling.
While it is difficult to name a precise scientific event or discovery that acts as the
defining moment of hydrodynamic stability theory, it is commonly acknowledged
that the experiments of Osbourne Reynolds in the 1880s have, for the first time,
solely focused on stability issues of shear flows and proposed a criterion (the
Reynolds number) for the distinction of laminar, turbulent, and transitional flow
regimes (Reynolds 1883). At the same time, Lord Rayleigh has given a formal
mathematical framework for the evolution of infinitesimal perturbations in an inviscid
fluid (Rayleigh 1887). Around the same time, the definition of stability has been
influenced by the mathematical work of A. Lyapunov, who gave a precise meaning
to the asymptotic fate of perturbations from an equilibrium point (Lyapunov 1892).
At the beginning of the twentieth century, the stability formulation by Orr and
Sommerfeld, a viscous extension of the Rayleigh equation, has moved stability issues
in the focus of fluid dynamicists and mathematicians and is still in widespread use
today (Orr 1907a, Sommerfeld 1908). Many subsequent investigations focused on the
associated Orr–Sommerfeld eigenvalue problem, and determined instability modes for
a wide range of fluid configurations, but in particular for wall-bounded unidirectional

h d i 0 0 9 8 088962 4 0 9 bli h d li b C b id i i
 258 Data-Driven Fluid Mechanics Chapter 13

shear flows. The incorporation of nonlinear saturation effects has further extended the
range of applicability of stability results to finite-amplitude perturbations.
The advent of the computers and numerical algorithms has also greatly influenced
hydrodynamic stability theory, and it did not take long before the central equations of
stability theory have been solved numerically to high precision and for increasingly
complex flows, such as high-speed, compressible flows.
At the same time, transition to turbulence and hydrodynamic stability have been
linked in an effort to predict the onset of turbulent fluid motion using stability
calculations. Secondary instability theory, a two-stage stability concept, has been
introduced and proposed as the route of many flows to a highly disordered state
(Orszag & Patera 1983, Herbert 1988, Koch et al. 2000).
A new development emerged in the 1990s whereby the original Lyapunov stability
concept has been called into question, as it does not contain any notion of a time
horizon. In this vein, eigenvalues have been increasingly superseded by more abstract
operator concepts (Butler & Farrell 1992, Reddy & Henningson 1993, Trefethen
et al. 1993, Schmid & Henningson 2001, Schmid 2007). This non-eigenvalue-
based stability theory has succeeded in uncovering and explaining a great deal of
experimentally observed phenomena and is often a key component of instability and
amplification processes.
The past decade has continued to produce a great many tools related to stability
theory: direct numerical simulations produce high-fidelity fluid solutions and give
unprecedented insight into all aspects of transport processes and instabilities; the
parabolized stability equations (PSE) provide a powerful and efficient tool for the
calculation of stability characteristics in configurations that go far beyond the early
simple geometries (Bertolotti et al. 1992). The stability of flow can also be deter-
mined globally using high-performance computers and iterative eigenvalue algorithms
(Theofilis 2011). Following these techniques, stability properties have been computed
for such complex flows as airfoil sections, turbomachinery stages, or wing-tip vortices.
Moreover, variational techniques (Hill 1995), which recast stability concepts into the
form of an optimization problem, have resulted in novel and powerful tools that extract
stability information (even in the nonlinear case) directly from simulation software. It
is these latter optimization techniques – in a rudimentary form – that will be the focus
of this chapter.

A Broader Viewpoint on Instabilities

Besides the rise of sophisticated tools to address stability issues for complex flows, the
definition of stability has also changed over the years. Starting from a rather binary
definition (stable or unstable equilibrium), based on the existence of eigenvalues in
a specific half-plane (Drazin & Reid 1981), we have brought back a timescale into
the definition of stability, thus dismissing the asymptotic definition of Lyapunov,
and reformulated stability-related issues under a more general sensitivity concept
(sensitivity to initial conditions, to forcing, to internal changes, etc.). Section 13.2
gives a brief summary of stability notion, starting with the mathematical definition

h d i 0 0 9 8 088962 4 0 9 bli h d li b C b id i i
 Schmid: Modern Tools for the Stability Analysis of Fluid Flows 259

q0 δ q0 δ q0 δ
ε ε
q(t)
qe qe qe

q(t) q(t)

Lyapunov stable Asymptotically stable Exponentially stable

(a) (b) (c)

Figure 13.2 Stability concepts, illustrating (a) Lyapunov stability, (b) asymptotic stability, and
(c) exponential stability.

of Lyapunov stability and progressing to more adapted concepts for hydrodynamic

stability calculations (Schmid & Henningson 2001).

13.2 Definitions of Stability

The stability of a dynamical system has been formalized mathematically by Lyapunov

in his landmark paper at the end of the nineteenth century (Lyapunov 1892). This
definition has stood the test of time and is still principally applied in a wide range of
situations from fluid dynamics to other dynamical systems in the engineering or life
sciences.
The first definition of Lyapunov stability states that, given a general autonomous
dynamical system q = N(q) for the state vector q with q(0) = q0, an equilibrium point
qe satisfying N(qe ) = 0 is Lyapunov stable, if for every ε > 0, there exists a δ > 0
such that, if kq0 − qe k < δ, then for every t > 0 we have kq(t) − qe k < ε.
Lyapunov stability thus relates an initial condition to the long-term fate of the state
vector following the governing equations. If the initial condition q0 is δ-close to the
equilibrium point qe, a Lyapunov-stable system will remain in a neighborhood of size
ε of this equilibrium for all times. This condition has to hold true for any ε one may
choose. The previous definition ensures boundedness of the solution as long as the
initial condition is sufficiently close to the equilibrium state.
A more stringent definition of stability, extensively used in fluid dynamics appli-
cation, is asymptotic stability. It adds to the definition of Lyapunov stability the
supplementary condition that, for an infinite time horizon, the equilibrium state qe
is approached. In other words, the solution to the dynamical system does not only stay
bounded in a neighborhood of the equilibrium point qe, it approaches the equilibrium
point as time tends to infinity.
A third definition of stability adds yet more restrictions to asymptotic stability
by specifying the manner in which the equilibrium point is approached. For initial

h d i 0 0 9 8 088962 4 0 9 bli h d li b C b id i i
 260 Data-Driven Fluid Mechanics Chapter 13

conditions within a δ-neighborhood of the equilibrium point, we require that the

equilibrium point qe is approached exponentially according to kq(t) − qe k < αkq0 −
qe k exp(−βt) for all times, with α and β as positive constants. Systems of this type
are referred to as exponentially stable and return to the equilibrium state within
monotonically decreasing bounds when started in a δ-neighborhood of qe .
Figure 13.2 illustrates the concept of Lyapunov, asymptotic and exponential
stability schematically for a dynamical system with two dependent variables.

For Hydrodynamic Applications

The definitions of Lyapunov have been adopted for hydrodynamic stability calcula-
tions; only minor modifications have been added to quantify concepts observed in fluid
systems. We will follow Joseph (1976) and introduce four variants of hydrodynamic
stability. These definitions require the introduction of a disturbance measure, which
we take as the perturbation kinetc energy E.
The first definition (asymptotic stability) states that an equilibrium solution is
asymptotically stable, if the perturbation energy satisfies
E(t)
lim → 0. (13.1)
t→∞ E(0)
This definition coincides with Lyapunov’s asymptotic stability definition, where the
distance to the equilibrium state is measured by the kinetic energy of the perturbation.
This definition, however, does not specify an ε-neighborhood bounding the distance
from equilibrium as the equilibrium state is approached.
The second definition (conditional stability) accounts for the fact that an initial
energy threshold value has to be respected for asymptotic stability. It states that an
equilibrium is conditionally stable, if there exists a positive energy level δ such that
asymptotic stability is observed as long as the initial energy of the perturbation falls
below δ, that is, E(0) < δ.
A further restriction on this concept yields the third definition (global stability):
for an infinite threshold value δ, the equilibrium is said to be globally stable. In other
words, the equilibrium state is attracting trajectories, no matter how far we start from
the equilibrium point.
Finally, even further constraints lead to the fourth definition (monotonic stability).
We state that an equilibrium is monotonically stable, if the energy of the perturbation
decreases monotonically for all times. Mathematically, this condition requires
dE
<0 for all t > 0. (13.2)
dt

The Role of Time

It is interesting to note that none of the previous definitions, whether set forth by
Lyapunov or proposed by Joseph (1976), contains a characteristic timescale. Instead,
time is either absent in the definition, or an infinite time horizon is envisioned. This
lack of restriction on the time it takes to return to the equilibrium point and thus to
confirm stability, suggests that such-defined stability measures (such as growth rates,

h d i 0 0 9 8 088962 4 0 9 bli h d li b C b id i i
 Schmid: Modern Tools for the Stability Analysis of Fluid Flows 261

A Supercritical A Subcritical

(a) p p (b)
pcrit pcrit

Figure 13.3 Bifurcation behavior, illustrated in the amplitude-parameter plane. In the

supercritical bifurcation scenario (a), finite-amplitude states exist only for parameter values p
larger than the critical one, pcrit . In the subcritical bifurcation scenario (b), finite-amplitude
states exist at parameter values p where the base state has not yet gone unstable to infinitesimal
perturbations, that is, for p < pcrit .

frequencies, modal shapes) may misrepresent the dynamic features of a flow that is
characterized and dominated by fluid processes on finite timescales.

Critical Parameters
The equations governing the temporal evolution of the state variable, denoted
symbolically by the function f, commonly contain a variety of physical parameters
often expressed as nondimensional numbers (such as the Reynolds number, Rayleigh
number, Mach number, etc.) or parameters linked to the shape of the disturbances
(such as their wavenumbers in homogeneous coordinate directions). The stability
properties thus depend on these parameters as well, and a specific perturbation may
be stable at one parameter setting, but become unstable at another. The parameter
value(s) at which this transition from stability to instability occurs are known as
critical parameters. Different critical parameters can be computed, depending on the
chosen definition of stability; for example, the critical Reynolds number for monotonic
stability of plane channel flow is Rec = 49.6, while the critical Reynolds number
for asymptotic stability of the same flow is Rec = 5772.2, where the Reynolds
number is based on the centerline base velocity and the half-channel height (Joseph &
Carmi 1969, Orszag 1971).

Bifurcation Behavior
Once we surpass the critical parameter, an instability ensues and the disturbance
grows in amplitude/energy until nonlinearities saturate the growth and establish a new,
nonlinear equilibrium state of finite amplitude. Finite-amplitude states, however, can
also exist (due to a conditional stability; see the earlier definition) at parameter values
below the critical one.
The existence of these finite-amplitude states determines the bifurcation behavior
of the flow (see also Chapter 10 for a discussion of bifurcations in dynamical
systems). Two cases have to be distinguished: supercritical and subcritical bifurcation
behavior. In the supercritical case, finite-amplitude states exist only past the point
(in parameter space) where the equilibrium state has gone asymptotically unstable to
infinitesimal perturbations. This situation is illustrated in Figure 13.3(a). Infinitesimal

h d i 0 0 9 8 088962 4 0 9 bli h d li b C b id i i
 262 Data-Driven Fluid Mechanics Chapter 13

perturbations below pcrit are asymptotically stable; for p > pcrit, finite-amplitude states
exist owing to an asymptotic instability of the infinitesimal state (dashed line). In
the subcritical case (see Figure 13.3(b)), we have a parameter regime where finite-
amplitude states coexist with asymptotically stable infinitesimal states. In this regime,
the infinitesimal state is conditionally stable: for an initial energy below a critical
value, the perturbation returns to the infinitesimal state (open circle); while for an
initial energy surpassing a threshold value (indicated by the dashed blue curve),
a higher-energy state is approached at the same parameter value p (closed circle).
After passing pcrit, this threshold value is zero, indicating that infinitesimal energy is
necessary to approach the higher-energy state via an asymptotic instability.
Examples of supercritical bifurcation behavior in fluid systems include Raleig-
Bénard convection and Taylor–Couette flow (flow between two differentially rotating
coaxial cylinders) within certain parameter regimes, while wall-bounded shear flows
such as plane Poiseuille, plane Couette, or pipe flow are governed by subcritical
bifurcations.

Temporal, Spatial, Spatio-temporal Evolution

While many stability calculations use time as the evolution coordinate and treat
disturbances as temporally evolving entities, there are configurations where a different
description is more appropriate. For example, disturbance evolution in flow past a
roughness element is more aptly described as a spatial problem. In this case, the
proper mathematical formulation is a signaling problem, rather than a temporal initial-
value problem. In a signaling problem, a harmonic or steady point source drives
the flow perpetually, and the instability manifests itself as amplitude growth as we
move downstream from the disturbance source. The evolution direction is the spatial
direction given by a representative streamline of the base flow. In this case, the time
coordinate is often taken as harmonic and can be Fourier transformed.
Besides the purely temporal and the purely spatial approach, we can also formulate
a general spatio-temporal approach, where we postulate a spatially and temporally
localized initial condition and track its development in space and time, without
assumptions on either evolution direction. Figure 13.4 illustrates the three concepts of
perturbation evolution: temporal, spatial, and spatio-temporal. Each concept has to be
motivated by the specific flow configuration and requires corresponding mathematical
tools for its analysis.

13.3 General Formulation

Most treatises on hydrodynamic stability theory start with many simplifying assump-
tions about the flow (steady, parallel, unidirectional, incompressible base flow of
Newtonian fluid), guiding the reader toward a set of mathematical techniques for the
solution of the resulting stability equations. Only in a second step are some of the
initial assumptions relaxed or eliminated, leading to additional mathematical compli-
cations. This simple-to-complex route has also been the historic route: simplifications

h d i 0 0 9 8 088962 4 0 9 bli h d li b C b id i i
 Schmid: Modern Tools for the Stability Analysis of Fluid Flows 263

2π/α
t0 t > t0

x x
Single wavenumber Growth in time

ω ω
t0 t > t0

x x
Single frequency Growth in space

t0 t > t0

x x
Multiple wavenumbers and frequencies Growth in space and time

Figure 13.4 Stability formalisms for a temporally evolving perturbation (top), a spatially
evolving disturbance (middle), and a general spatio-temporal evolution from a point source
(bottom).

have been necessary, since the mathematical tools have been restricted to analytical,
approximate, or asymptotic methodology, such as, for example, perturbation and
asymptotic methods. With the advent of computational and data-driven techniques,
we are now in a position to address directly – with few imposed restriction and forced
assumptions – the hydrodynamic stability behavior of complex flows. The analysis of
unsteady, separated, and multiscale flows are now within our range, as is the nonlinear
evolution of disturbances in these configurations.
For this reason, we will advocate a reverse route of exposition: starting from the
general, unrestricted stability problem to the more confined, but perhaps more familiar
setup. We will describe a rather general computational framework for hydrodynamic
stability problems based on the optimization of a cost functional subject to constraints,
among them our nonlinear partial differential equation governing the fluid motion.
This formalism is very flexible in describing and quantifying the perturbation dynam-
ics under minimal restrictions and limitations. After having established this general
framework, we will then make progressive assumptions – either about the base flow
or the perturbation dynamics – to find reduced descriptions of the general formulation.
This will naturally introduce familiar techniques for the analysis of a fluid system’s
stability characteristics.
The general framework for hydrodynamic stability analysis relies on a math-
ematical formalism that uses calculus of variations, optimization techniques, and
concepts from linear algebra. A basic knowledge of these disciplines is helpful
in understanding the derivations later; it can be gained from excellent textbooks,
monographs, and review articles, such as Kot (2015) and Cassel (2013) for calculus of
variations, Gunzburger (2003) and Magri (2019) for adjoint techniques, Kochenderfer

h d i 0 0 9 8 088962 4 0 9 bli h d li b C b id i i
 264 Data-Driven Fluid Mechanics Chapter 13

and Wheeler (2019) and Nocedal and Wright (2006) for optimization, and Trefethen
and Bau (1997) for linear algebra, among many other resources.

An Optimization Formulation for Hydrodynamic Stability

We assume a general form of the governing equations according to
d
q = N(q, f), (13.3)
dt
where the components of the state vector q ∈ Cnq represent all variables required
to fully describe the flow state. The vector f ∈ Cn f stands for the control variable.
The exact form of the governing equations is given by the operator N, where we
additionally have assumed an appropriate spatial discretization of the state vector
q. We consider nq degrees of freedom for the state vector, comprising the product
of the total number of grid points (for example, nx × ny × nz in three Cartesian
dimensions) and the number of tracked variables (for example, three velocities
and pressure for three-dimensional incompressible flow). The control vector has n f
degrees of freedom, indicating the number of grid points and variables we have chosen
to manipulate. Boundary conditions are enforced within the spatial discretization
schemes, and the initial condition will be imposed via the control variable f (see below
for details).
Next, we introduce a cost functional J (q, f), consisting of a real scalar that depends,
in a yet unspecified way, on the variables q and f. We recall that a functional is
a mapping of functions onto a (generally complex) scalar. Thus, J takes the time-
dependent input functions q and f and assigns a scalar value to them. More specifically,
this cost functional represents the objective of the optimization, for example, a
maximization of energy amplification (for stability analyses) or a minimization of
instabilities (for flow control applications; see also the control framework developed
in Chapter 9). During the optimization, the governing equations (13.3) have to be
satisfied, which renders our optimization setup as a constrained optimization problem.
Invoking standard techniques, we convert this constrained optimization problem into
an unconstrained equivalent by adding the constraints via Lagrange multipliers to the
cost functional; this results in the augmented Lagrangian:

† d
 
L(q, q , f) = J (q, f) − q , q − N(q, f) .
†
(13.4)
dt
This mathematical step required the introduction of Lagrange multipliers q† and
the choice of a scalar product. It follows from expression (13.4) (more specifically,
from term with the scalar product) that the Lagrange multipliers q† have the same
dimensionality as q, that is, q† ∈ Cnq . We thus realize that the Lagrange multipliers
represent time-dependent “flow fields.” At this stage, however, we do not yet have an
equation governing their temporal evolution. Proceeding, we define a scalar product as
∫ τ
ha, bi ≡ a H b dt. (13.5)
0

h d i 0 0 9 8 088962 4 0 9 bli h d li b C b id i i
 Schmid: Modern Tools for the Stability Analysis of Fluid Flows 265

The superscript H stands for the transpose conjugate operation. In the previous
definition, we have introduced the time horizon τ, which has to be user-specified and
adapted to the relevant timescales of the flow under consideration. In what follows,
it will prove mathematically advantageous to recast the cost functional J in terms of
the scalar product (13.5). We write

J (q, f) = hw(t), j(q, f)i , (13.6)

where we included a weight function w(t) that is responsible for the enforcement of
j within the time interval [0, τ]. As an example, specifying a cost functional j over
the entire time horizon, we choose w = 1; for focusing only on the value of j at the
end of the time horizon, we choose w = δ(t − τ).
With the augmented Lagrangian L fully specified, we seek an optimum of L
with respect to all independent variables, that is, q, q†, f. This mathematical problem
requires calculus of variations: we determine three time-dependent functions (q, q†, f)
that optimize a scalar value (our augmented Lagrangian L). Analogous to computing
an optimum for a function of three independent variables f (x, y, z) by setting the first-
order partial derivatives with respect to the three variables to zero, fx = fy = fz = 0,
we have to take the first-order variations of the scalar L (see (13.4)) with respect
to the three functions q, q†, f and set them to zero simultaneously (see, e.g., Cassel
(2013) or Kot (2015)). The resulting three conditions are referred to as the Karush–
Kuhn–Tucker (KKT) system. Formally, we have

δL δL δL
= 0, = 0, = 0, (13.7)
δq† δq δf

where each of the above conditions furnishes a governing equation.

The first condition reads, more explicitly,
* +
d
δq†, q − N(q, f) = 0, (13.8)
dt
| {z }
=0

a condition that has to hold true for all variations δq† . This expression thus recovers
our original governing equation for q.
The second condition is slightly more involved, as the state variable q appears in
the governing equation as well as in the cost functional. We have
 * H H +
∂j d ∂N ∂ j d ∂N
   
w, δq − q†, δq − δq = w + q† + q†, δq = 0,
∂q dt ∂q ∂q dt ∂q
| {z }
=0
(13.9)
where integration by parts (in time) is required to transfer the d/dt-operator onto the
adjoint variable q† . After isolating the first variation δq in the scalar products, we can
extract an evolution equation for the adjoint variable q† .

h d i 0 0 9 8 088962 4 0 9 bli h d li b C b id i i
 266 Data-Driven Fluid Mechanics Chapter 13

Finally, the third condition yields the following explicit expression:

 * H H +
∂j ∂N ∂j ∂N
   
w, δf − q , − †
δf = w+ q , δf = 0,
†
(13.10)
∂f ∂f ∂f ∂f
| {z }
=0

from which we can extract an algebraic equation for the cost functional gradient with
respect to the control vector f.
In summary, we obtain from the three optimality conditions the following three
equations:
δL d
=0: q = N(q, f), (13.11a)
δq† dt
H  H
δL d ∂N ∂j

=0: − q† = q† + w, (13.11b)
δq dt ∂q ∂q
 H H
δL ∂j ∂N

=0: w=− q† . (13.11c)
δf ∂f ∂f

By design, this set of equations provides an extremum (q, f) of the chosen cost
functional J . Analogously to finding the extremum of a multivariate function, the
system (13.11) has to be solved simultaneously. In practice, however, we approach
the solution in an iterative manner. We solve the two evolution equations (13.11a,b)
exactly, while we use the algebraic optimality condition (13.11c) iteratively to
converge toward an optimal control variable f. The above optimization problem is the
basis for our analysis of fluid problems as to their stability, receptivity, and sensitivity
to internal or external changes. At the same time, the optimization framework (with
only minor modifications) can also be used to compute optimal flow control strategies
or optimal designs of fluid devices.

An Iterative Process: Direct-Adjoint Looping

The full iterative process unravels as follows: we commence with an initial control-
variable estimate f (0) and solve the governing equation (13.11a) over the chosen
time interval [0, τ]. This direct solution will then act as the forcing term for the
adjoint equation (13.11b), which is solved in a time-reversed manner from t = τ
to t = 0. The time-reversed integration of the adjoint stems from the integration by
parts (in time) during the derivation of the adjoint equation; the change in sign for the
first-order time derivative requires a reverse integration or, alternatively, a change of
time-coordinate in the form t 0 = τ − t. The resulting adjoint solution q† then is used
in expression (13.11c) to determine the cost functional gradient with respect to the
control variable f. This gradient information is provided to a standard optimization
routine, such as an algorithm from the conjugate gradient or quasi-Newton family,
and a new control variable f (1) is determined. The iterative process then repeats
with this improved control strategy, until a user-specified convergence criterion is

h d i 0 0 9 8 088962 4 0 9 bli h d li b C b id i i
 Schmid: Modern Tools for the Stability Analysis of Fluid Flows 267

f (0)
d
q = N(q, f (i) )
dt
i →i+1

∂j (i)
 
f (i+1) = opt ,f t
∂f
0 Until convergence τ
H H
∂j ∂N
 
w=− q†
∂f ∂f

∂N H †
 H
d † ∂j
 
− q = q + w
dt ∂q ∂q

Figure 13.5 Schematic of adjoint looping. An initial guess f (0) for the control variable is used
to solve the governing equations over a time horizon t ∈ [0, τ]. This is followed by a solution
of the adjoint equations backward in time (from t = τ to t = 0), accounting for the driving by
the forward solution q. The adjoint solution q† is then substituted into the cost functional
gradient ∇f j, which in turn is passed to an optimization algorithm opt. A new and improved
control variable f is determined, and another iteration starts. The iterations are continued until
convergence is reached.

satisfied, or the computational resources are exhausted. Figure 13.5 gives a schematic
representation of this iterative process, often referred to as adjoint looping.
A few observations about the full system (13.11) and the iterative scheme are worth
pointing out.
(i) While the direct problem (13.11a) may be nonlinear, the adjoint equation (13.11b)
is always linear in q† . This fact results from the linear appearance of q† in the
augmented Lagrangian. The linearity of the adjoint equation can be exploited in
parallel-in-time algorithms for the adjoint part of the loop (Skene et al. 2020);
increased performance and computational efficiency can be gained in this manner.
(ii) Expression ∂N/∂q is recognized as the Jacobian of the governing equation (13.3),
and constitutes a matrix of size nq × nq . For nonlinear N, the Jacobian appears in
complex conjugate form in the adjoint equation and is evaluated at the time-local flow
field q. Details on how to compute this Jacobian – or more specifically its action on a
given state vector – are given below.
(iii) The system matrix (∂N/∂q) H for the adjoint equation (13.11b) does not
dependent on the cost functional J . The influence of J in the adjoint equation only
stems from the external forcing term (∂ j/∂q) H w. Different cost functionals produce
different adjoints via this forcing term, even though the inherent system dynamics is
independent of our choice of cost functional.
(iv) Given a nonlinear N or a non-convex cost functional J, any optimization scheme
based on gradient information can only reach a local extremum. No guarantee can

h d i 0 0 9 8 088962 4 0 9 bli h d li b C b id i i
 268 Data-Driven Fluid Mechanics Chapter 13

be given for convergence toward a global optimum. The type of optimum (minimum
or maximum) can be determined by monitoring the cost functional; a sign change in
applying the gradient-based update may be necessary.
Despite some obvious limitations, the general optimization formalism is versatile
and flexible, and can be brought to bear on a wide range of fluid problems.

13.4 Reduced Formulation for Special Cases

The above system of equations (13.11a-b-c) can be applied to general configurations

and fluid systems, including nonlinear governing equations. In this section, we will
make common assumptions about the fluid system and link the general framework
to more familiar concepts of flow analysis. To this end, we take the right-hand side
of the governing equations in the special form N(q, f) = Lq + f, with L representing a
linear(ized) set of equations (see also the material in Chapter 9). With this assumption,
we can recover common stability and receptivity results (Schmid 2007).

Stability of Linear, Time-Invariant Systems

The linear stability of fluid flow is very common type of flow analysis. It rests on the
assumption of a steady equilibrium state, which (i) can be computed exactly, such as,
for example, in channel or pipe flow, or (ii) can be argued for as an approximation
following a timescale separation. Linearization about this equilibrium state yields a
linear, time-invariant (LTI) system, and the stability of the equilibrium state depends
on the dynamics of infinitesimal perturbations about it.
Stability can be defined in a variety of ways. We will use a gain concept: the
amplification of the size of a perturbation over its initial size, optimized over all
initial conditions. This gain definition runs counter to Lyapunov’s definition of
stability, as it retains the timescale over which we consider the amplification. In
addition, our gain definition requires a choice of measure of disturbance size. This
choice is nontrivial in many cases and crucially depends on the flow configuration
as well as the fluid properties. For incompressible flows, the choice is bit more
straightforward: the kinetic perturbation energy is commonly used as a measure of
disturbance size. If additional physical effects, such as, for example, buoyancy, surface
tension, compressibility, more thought has to be put into an appropriate measure.
Mathematically, we define the norm of the state vector according to
kqkQ2 = q H Qq, (13.12)
in which Q is a symmetric, positive definite weight matrix. Factoring Q into Q = F H F
(using a Cholesky factorization), we can establish a link to the more standard L2 -
norm as
kqkQ2 = q H F H Fq = kFqk22 . (13.13)
We state our cost objective as the quotient of the state-vector norm over the control-
vector norm, evaluated over a time interval [0, τ]. The norms for the state vector and

h d i 0 0 9 8 088962 4 0 9 bli h d li b C b id i i
 Schmid: Modern Tools for the Stability Analysis of Fluid Flows 269

the control vector do not necessarily have to coincide; for this reason, we choose two
different weight matrices in the respective norms. Furthermore, we evaluate the gain
at the end of the time horizon, at t = τ, which coincides with more traditional notions
of (in)stability. This choice of evaluation leads to w = δ(τ − t) in (13.6). We have
kqkQ2 q H Qq
j(q, f) = = , (13.14)
kf kR2 f H Rf
with R denoting the weight matrix for the control variable, which in our case is taken
as the initial condition, that is, f = q(t = 0). Consequently, the evolution operator
takes on the form N(q, f) = Lq + fδ(t).
With this setup, we can reconsider the general iterative optimization framework and
use the linearity of our governing equations to make simplifications. The first step in
the direct-adjoint loop consists of solving the governing equations (13.11a) over the
time interval [0, τ]. The solution of this problem can formally be written as
q(τ) = exp(Lτ) f, (13.15)
introducing the matrix exponential that transforms the initial condition f into the
output perturbation q(τ). The second step, based on equation (13.11b), simplifies for
our LTI-case1
d 2Q
− q† = L H q† + wq, (13.16)
dt kf kR2
which again can be solved formally according to
2Q
q† (0) = exp(L H τ) q(τ). (13.17)
kf kR2
The third and final step uses the cost functional gradient
 H
∂j
w = −δ(t)q†, (13.18)
∂f
where the right-hand side term, δ(t), is equivalent to the gradient ∂N/∂f for our special
case.
At this stage, we would use the aforementioned gradient in a user-specified
optimization routine (such as conjugate gradient, for example) to arrive at an
improved initial condition f. In this special case, however, we further simplify the
direct-adjoint loop (Figure 13.5) by choosing a steepest-descent procedure as our
gradient-based optimization routine. This results in an explicit expression for the new
initial condition. We have
kqkQ2
 H
∂j
w=− 2Rf w. (13.19)
∂f kf kR4
Combining (13.15), (13.17), (13.18), and (13.19), we obtain
kq(τ)kQ2
R−1 exp(L H τ) Q exp(Lτ) f = f. (13.20)
kf kR2
1 Recall that the derivative of the quadratic form x H Ax with respect to x is 2Ax for symmetric A.

h d i 0 0 9 8 088962 4 0 9 bli h d li b C b id i i
 270 Data-Driven Fluid Mechanics Chapter 13

This last formula can be further manipulated into an expression for f

exp(L̄τ) exp(L̄τ) f̄ = σ 2 f̄,

H
(13.21)


describing a full iteration of the direct-adjoint loop. We have introduced the energy
gain σ 2 = kq(τ)kQ2 /kf kR2 , together with the transformed matrix L̄ = FLG−1 and the
transformed initial condition f̄ = Gf. The matrix R = G H G has been decomposed into
its Cholesky factors G.
Equation (13.21) is an eigenvalue problem for the energy gain σ 2 . It then follows
that the largest eigenvalue produces the optimal energy growth, and the associated
eigenvector f̄ gives the optimal initial condition f. The two matrix exponentials
in (13.21) are conjugate transpose to each other; the product yields real gains σ 2 .
This configuration also suggests that the eigenvalue problem can be transformed to
a singular value problem for exp(L̄τ): the largest singular value (i.e., the L2 -norm
of the matrix exponential) produces σ, the principal left singular vector corresponds
to f and the principal right singular vector gives the corresponding optimal output
q(τ). This latter relation is at the core of nonmodal stability theory: the norm of
the matrix exponential measures the maximum transient energy amplification. The
above demonstrates that it can now be thought of as a special case of the direct-adjoint
looping procedure for LTI systems. The iterative optimization procedure is equivalent
to a power iteration applied to the composite matrix in (13.21), and the optimization
problem can be solved by straightforward linear algebra operations.

Receptivity of Linear, Time-Invariant Systems

Section 13.3 addressed the stability of an equilibrium state by computing the maximal
gain of a perturbation over a specified time horizon. Alternatively, amplification of
perturbations can also be accomplished by constant forcing. This alternative setup
addresses the issue of receptivity, that is, the susceptibility of the flow to its disturbance
environment (e.g., free-stream turbulence or wall roughness). Due to the linearity of
the underlying equations, we can recast the general, forced problem into a frequency
response analysis to harmonic forcing and use superposition to capture more general
disturbance environments. An analysis of this type is referred to as a linear receptivity
analysis.
As before, we consider an LTI system and simplify the right-hand side in (13.11)
to N(q, f) = Lq + f exp(iωt), where ω denotes the forcing frequency and f its spatial
shape. We define a frequency response gain as the energy (norm) of the response
divided by the energy (norm) of the forcing. Due to linearity, the response will adopt
the frequency of the forcing; hence, only the spatial shapes of forcing and response
are relevant to the analysis. We choose

kqw kQ2
j(q, f) = , (13.22)
kf kR2

where qw denotes the spatial shape of the forced response q = qw exp(iωt). We note
that j(q, f) is independent of time, and thus we do not need to specify w, but still

h d i 0 0 9 8 088962 4 0 9 bli h d li b C b id i i
 Schmid: Modern Tools for the Stability Analysis of Fluid Flows 271

assume it as a constant. The above analysis is linked to transfer functions, and the
reader is urged to compare with related material in Chapters 4 and 9.
The direct problem (13.11a) reduces to
∫ t
q(t) = exp(L(t − t 0)) f exp(iωt 0) dt 0, (13.23a)
0
= (iω − L)−1 [exp(iωt) − exp(Lt)] f. (13.23b)
We see that equation (13.23a) represents a convolution of the harmonic forcing with
the input response of the linear system (see also Chapter 5 for additional material).
Before moving forward, we stress that the above solution assumes an asymptotically
stable system, with all eigenvalues of L contained in the stable half-plane. In essence,
we only consider the long-term response of the system and ignore transient processes
of establishing this long-term response. As a consequence, the term in (13.23b)
containing the matrix exponential can be neglected. We have
q = (iω − L)−1 f exp(iωt) = qw exp(iωt). (13.24)
Reformulating the adjoint equation in (13.11b) for our special case, we obtain
d † 2Q
− q = L H q† + qw exp(iωt) w, (13.25)
dt kf kR2
which has the formal solution
2Q
q† = (−iω − L H )−1 qw exp(iωt) w. (13.26)
kf kR2

From the optimality condition (13.19) and the expression for q† we get
kqw kQ2
H
∂N

− 2R f w = q† = exp(−iωt) q† . (13.27)
kf kR4 ∂f
Combining with (13.24) and (13.26) yields
R−1 (−iω − L H )−1 Q (iω − L)−1 f = σ 2 f, (13.28)
with σ 2 as the sought-after response gain, defined as kqw kQ2 /kf kR2 . Using the Cholesky
factorization of the weights Q and R (see above), we can write the above expression
more compactly as
H
(iω − L̄)−1 (iω − L̄)−1 f̄ = σ 2 f̄, (13.29)


with L̄ = FLG−1 and f̄ = Gf.

As before, we obtain an eigenvalue problem for the gain σ 2 : the optimal frequency
H
response gain for a given frequency ω is the largest eigenvalue of (iω − L̄)−1 (iω −
L̄)−1, or, equivalently, the first singular value of (iω − L̄)−1 . This latter operator is
referred to as the resolvent. The optimal forcing and output are given, respectively,
by the principal right and left singular vectors (Schmid & Henningson 2001, Schmid
2007). The resolvent has found widespread application in the investigation of fluid
behavior (see, e.g., McKeon & Sharma 2010, Moarref et al. 2014, Yeh & Taira 2019).

h d i 0 0 9 8 088962 4 0 9 bli h d li b C b id i i
 272 Data-Driven Fluid Mechanics Chapter 13

The Time-Asymptotic Limit

In some cases, we may be interested in the time-asymptotic limit τ → ∞. This may be
justified by the presence of a very large timescale in the fluid system, or by the fact that
the linear system matrix L is of normal type, that is, commutes with its adjoint. Fluid
systems in that category are, for example, Rayleigh–Benard convection or Taylor–
Couette flow in the small-gap limit. Other flows, while strictly nonnormal, may be
nearly normal, and thus justify a simplified analysis. In the normal or nearly normal
case, the direct-adjoint looping and the evaluation of the matrix functions exp(τ L̄) (for
the initial-value problem) or (iω − L̄)−1 (for the harmonically driven problem) further
reduced. The dynamics of a problem in this category is governed by the eigenvalues of
L̄. For the initial-value problem, the time-asymptotic fate of perturbations is given by
the least stable mode, that is, the eigenvector of L̄ corresponding to the eigenvalue
with the largest growth rate. This dominant structure will grow exponentially in
time, with the growth rate (and frequency) given by the least stable eigenvalue. For
the harmonically driven case, the resolvent analysis reduces to a simple resonance
argument: the response to harmonic forcing is inversely proportional to the distance
of the forcing frequency ω to the spectrum (eigenvalues) of L̄. The closeness of the
forcing frequency to the spectrum is the only mechanism to solicit a large response
from the fluid system. In the general, non-normal case, we can also generate a
substantial response, even though the forcing frequency is far from the spectrum of
L̄ – a phenomenon referred to as pseudo-resonance.

A Link Between the Optimization and Linear-Algebra Approach

We have seen that for linear time-invariant systems, there is a link between the general
direct-adjoint looping method, which finds optimal solutions to stability or receptivity
problems iteratively, and the direct (non-iterative) evaluation of matrix function
norms. These latter linear-algebra solutions are preferable when they apply, but the
flexibility and generality of the direct-adjoint approach can hardly be overstated.
The connection between linear algebra and optimization, as a solution technique
for hydrodynamic stability and receptivity problems, can be traced to the optimal
character of the L2 -norm of a linear operator (matrix): the L2 -norm of a matrix is
defined as the maximum ratio of input norm to output norm. If a mapping between the
input and output is available and given by a function f of L, we have
k f (L)qk2
k f (L)k2 = max . (13.30)
q,0 kqk2
In the previous sections, we have seen that f (z) = exp(tz) for LTI-stability problems,
and f (z) = 1/(iω − z) for LTI-receptivity analyses. Other mappings, not covered here,
include, for example, f (z) = az n for linear time-periodic (LTP) stability studies, also
known as Floquet analysis (Schmid 2007).
The norm of the mapping f (L) represents the maximum amplification over all
input vectors q. This latter statement is underlying all our gain optimizations of
the previous sections: we wish to optimize an output-to-input ratio (gain) where the
mapping between input and output is given by a linear operator (matrix function).

h d i 0 0 9 8 088962 4 0 9 bli h d li b C b id i i
 Schmid: Modern Tools for the Stability Analysis of Fluid Flows 273

The optimal solution is then given by a singular value decomposition that results
in the optimal gain (largest singular value), the optimal input (principal left singular
vector), and the associated output (principal right singular vector); see Chapter 6 for
additional material. The connection between optimization and linear algebra seems
rather attractive, but does come with its limitations, particularly when we consider
more complex flow configurations.
The foremost limitation is the restriction to linear and time-invariant systems.
For nonlinear or nonautonomous linear systems, we have to revert to the iterative
optimization techniques. Another limitation stems from the definition of the cost
functional as a gain or a quadratic form. The stability/receptivity of some fluid
problems require “exotic” norms that cannot be reduced to L2 -norms. In many other
circumstances, we may wish to augment the cost functional by terms and norms that
enforce additional constraints, for example, spatial sparsity. In other cases, the norm
may not include all components of the state vector, resulting in the so-called semi-
norm problem and requiring side constraints to ensure convergence of the iterative
optimization scheme.
The most obvious advantage of the optimization approach lies in its versatility to
adapt to/accommodate linear as well as nonlinear governing equations, to L2 -norms
as well as other, more exotic norms, to gains as well as to more general functionals,
and to additional side constraints.

13.5 Practical Issues and Implementation Details

The iterative optimization approach is a versatile concept to find optimal solutions

or gradients of flow characteristics with respect to user-defined control variables, but
an efficient implementation is crucial for making it a practical tool for quantitative
flow analysis, particularly for large-scale applications, multi-physics problems, and
nonlinear governing equations.

High Dimensionality
Whether using the iterative direct-adjoint looping or the special formulation as a
linear-algebra problem, any effort should be taken to reduce the number of degrees of
freedom and thus ensure a swift convergence to a solution. The most obvious reduction
for linear problems is the use of transforms in the homogeneous directions, which is
equivalent to a separation-of-variable approach. If our linear geoverning equations
are of constant-coefficient type in one (or more) of the spatial coordinates, we can
employ a Fourier transform in this directions and introduce an associated wavenumber.
We then have to solve the iterative or linear-algebra system for each wavenumber
(or wavenumber tuple) – a far more efficient undertaking than solving the global
governing equations. Other techniques to reduce the number of degrees of freedom
involve similarity transformations or the exploitation of symmetries.

h d i 0 0 9 8 088962 4 0 9 bli h d li b C b id i i
 274 Data-Driven Fluid Mechanics Chapter 13

Efficient Extraction of Linearized and Adjoint Information

The overall efficiency of the iterative optimization scheme (13.11) depends on
efficiency in determining all its computational ingredients. In particular, we need a
fast and convenient way to determine the key components of the direct-adjoint loop:

H H
∂N ∂j
 
, . (13.31)
∂q ∂q

We wish to determine these expressions directly from the computational program

used to solve the direct problem. To this end, we start with the numerical code
that time steps the direct problem dq/dt = N(q, f) over a specified interval [0, τ],
given an initial condition or an external driving term. The associated adjoint evolution
operator (∂N/∂q) H is required, but only necessary as a matrix-vector multiplication.
Assuming that the (generally nonlinear) direct problem is given as a sequence of
linear and nonlinear submodules that are encoded in our numerical program, we seek
to extract the necessary linearized and adjoint information using ideas from reverse-
mode automatic differentiation (Fosas de Pando et al. 2012).
Let us illustrate the key steps of this process on a simplified example. We choose

d
q = N(q) = N2 (Dx q, Dy N1 (Dx q)), (13.32)
dt

for which we determine the adjoint evolution matrix for the second leg of the direct-
adjoint iterative optimization. The functions N1 (q) and N2 (q, q) are assumed nonlinear
in their argument(s), but acting locally on the grid points. We next break the composite
right-hand side of (13.32) into procedural steps (similar to the modules/subroutines of
a computer program), introducing auxiliary variables for each step as needed. We
obtain

q0 = q, (13.33a)
q1 = N1 (Dx q0 ), (13.33b)
q2 = N2 (Dx q0, Dy q1 ), (13.33c)
d
q = q2 . (13.33d)
dt

The traversal from a given flow field q to its final time rate of change dq/dt is given
by a directed, acyclic graph (DAG) connecting the various modules. For simplicity,
we assume that numerical differentiation (indicated by multiplication by Dx or Dy ) is
a linear operation.
We first linearize the process of evaluating N(q). With differentiation assumed
linear, we only have to linearize the modules N1 and N2 . This is accomplished by
a simple Taylor-series expansion about the linearization state (generally, the base flow
denoted by q̄); we have

h d i 0 0 9 8 088962 4 0 9 bli h d li b C b id i i
 Schmid: Modern Tools for the Stability Analysis of Fluid Flows 275

dN1  N1 (q̄ + q) − N1 (q̄)


q ≈ = A1 q, (13.34a)
dq q̄
 
N2 (q̄ + q, q̄) − N2 (q̄, q̄)
= A2,0 q, (13.34b)

N2 (q̄, q̄ + q) − N2 (q̄, q̄)
= A2,1 q. (13.34c)

We see that the linearization of the two nonlinear modules is accomplished by two
function calls. If the linearization point remains constant, it even reduces to one
function call. The parameter  in the above expression has to be chosen sufficiently
small to ensure adequate accuracy of the linearization, but large enough to avoid
effects due to round-off errors.
Replacing the nonlinear procedural steps by their linearized equivalents then yields

q0 = q, (13.35a)
q1 = A1 Dx q0, (13.35b)
q2 = A2,0 Dx q0 + A2,1 Dy q1, (13.35c)
d
q = q2 . (13.35d)
dt
Combining all steps into one, results in the linearized governing equations. They
read
d
q = A2,0 + A2,1 Dy A1 Dx q. (13.36)


dt | {z }
dN
dq
The evolution equation for the adjoint variables involves the complex transposition of
this equation. We have to form
d †  
q = Dx H A2,0H
+ A1H Dy H A2,1
H
q†, (13.37)
dt
| {z }
dN H
!

dq
or, broken down into procedural steps,

q†0 = q†, (13.38a)

q†1 = H †
A2,1 q0 , (13.38b)
q†2 = A1 Dy H q†1,
H
(13.38c)
q†3 = H †
A2,0 q0 + q†2, (13.38d)
d †
q = Dx H q†3 . (13.38e)
dt
Comparing the linearized direct problem to the above adjoint problem, we notice that
the involved operators have to be invoked in reverse order and as their conjugate

h d i 0 0 9 8 088962 4 0 9 bli h d li b C b id i i
 276 Data-Driven Fluid Mechanics Chapter 13

transpose version. The involved operators are, however, readily available, in their
original and conjugate transpose form.
The aforementioned computational technique presents an efficient way of extract-
ing linearized and adjoint information required to perform the direct-adjoint optimiza-
tion (see Figure 13.5).
Since it is not linked to a specific set of equations, but rather to a simulation
code, it can easily be extended, without the need for further derivations. The gradient
information is numerically accurate to the precision of the chosen numerical scheme,
and requires a computational effort similar to the direct problem.

Krylov Time-Stepping
The previous sections have addressed the spatial discretization and the automated
extraction of linearized/adjoint information. The temporal evolution of the governing
equations also has to be considered carefully for an overall efficient optimization
procedure.
We advocate the use of Krylov time-stepping to advance the direct and adjoint
governing equations over the chosen time horizon [0, τ]. To this end, we approximate
the matrix exponential as the map over finite time interval, by forming a m-
dimensional Krylov subspace based on the evolution matrix L according to

Km = span{v, Lv, L2 v, . . . , Lm−1 v}, (13.39)

where v represents a starting vector. We continue by projecting the matrix L onto this
subspace to obtain the approximation
H
L ≈ Qm Hm Qm , (13.40)

where Qm denotes an m-dimensional, orthonormal basis for the subspace Km . The

m × m upper Hessenberg matrix Hm is a representation of the full matrix L. Any
evaluation of any matrix function involving L can then be replaced by a more efficient
evaluation using Hm instead. More specifically, for a matrix-vector operation involving
the function g(z), we have
H
g(L)b ≈ Qm g(Hm ) Qm b, (13.41)

requiring a far smaller effort, since the matrix function evaluation is performed on the
reduced matrix Hm .
Still, even with these approximations, the repeated computation of the matrix-
function-vector product constitutes the most costly part of our time-stepping method.
For this reason, we choose a time-stepping scheme that uses a minimum number of
Krylov projections for a desired performance (accuracy).
We then use this reduced representation of matrix functions to design an exponen-
tial time-stepping scheme and apply it to a nonlinear system of ordinary differential
equations of the general form:
d
q = N(q), q(t0 ) = q0 . (13.42)
dt

h d i 0 0 9 8 088962 4 0 9 bli h d li b C b id i i
 Schmid: Modern Tools for the Stability Analysis of Fluid Flows 277

We use a Taylor expansion about q0 to yield

d
q = N(q0 ) + ∇N| 0 (q − q0 ) + r(q), (13.43)
dt
with r as the (nonlinear) remainder term. We apply an integrating factor based on the
Jacobian ∇N| 0 , together with a variable change s = (t − t0 )/∆t, to arrive at
∫ 1
exp(L0 ∆t) − I
q(t0 + ∆t) = q0 + ∆t N(q0 ) + ∆t exp(sL0 ∆t) r(q(s)) ds, (13.44)
L0 ∆t 0

where we used the notation L0 = ∇N| 0 . This expression is the starting point of
any exponential time-stepping method. Various manifestations of exponential time-
stepping schemes distinguish themselves by how the integral term is treated and
by how the second term on the right-hand side is approximated (Hochbruck &
Ostermann 2010, Schulze et al. 2009). The following scheme (referred to as the
exponential Rosenbrock-type scheme) blends many of the advantages of exponential
time-stepping using Krylov techniques to approximate the various matrix functions
arising in the formula. We have
 
∆t ∆t
k1 = q0 + ϕ1 L0 N(q0 ), (13.45a)
2 2
k2 = q0 + ∆t ϕ1 (∆tL0 ) N(q0 ) + ∆t ϕ1 (∆tL0 ) r(k1 ), (13.45b)

q(t0 + ∆t) = q0 + ∆t ϕ1 (∆tL0 ) N(q0 ) + ∆t [16ϕ3 (∆tL0 ) − 48ϕ4 (∆tL0 )] r(k1 )

+ ∆t [−2ϕ3 (∆tL0 ) + 12ϕ4 (∆tL0 )] r(k2 ), (13.45c)

which requires three Krylov projections per time step. In the above time-stepping
scheme, we encounter higher exponential functions ϕk that can be derived from the
recurrence relation
1
ϕk (z) = + zϕk+1 (z), k = 0, 1, . . . , (13.46)
k!
with ϕk (0) = 1/k! and ϕ0 (z) = exp(z). They can be evaluated using the general matrix
function expression for a Krylov subspace framework.

Checkpointing
Using the direct-adjoint optimization scheme with a nonlinear governing equation or
a general form of the driving term (∂ j/∂q) H , the adjoint equation (13.11b) depends
not only on q†, but also on q. This link between the direct and adjoint equation leads
to additional computational challenges: we have to retain the flow fields q during the
temporal evolution of the direct problem, and inject them, in reverse order, into the
coefficients of the adjoint problem (13.11b). For small-size problems, we may have
sufficient memory to store all generated flow fields q during the direct part of the loop.
But even more moderately sized problems, we have to resort to a technique known as
checkpointing.

h d i 0 0 9 8 088962 4 0 9 bli h d li b C b id i i
 278 Data-Driven Fluid Mechanics Chapter 13

During the computation of the direct solution, we store the flow fields q only at
a given number of checkpoints {tc }. These flow fields will subsequently be used
as initial conditions to recompute the necessary flow fields between checkpoints as
needed by the adjoint equation. Figure 13.6 demonstrates the two techniques: (i)
storing all flow fields, and (ii) storing only at specific checkpoint and recovering the
required flow fields by additional direct simulations. In the first case, we store all
flow fields at all time steps within the interval [0, τ]. These fields are then injected
(in reverse order) into the coefficients of the adjoint equation as well as the driving
term of the adjoint equation. No computational overhead arises. In the second case,
we assume memory restrictions that allow us to only store a finite number of flow
fields in the interval [0, τ] (in the illustration, seven checkpoints have been chosen).
The checkpoints are distributed unevenly, with the final points stored for every time
step. These last, densely stored fields are immediately used in the coefficients of the
adjoint equation during the beginning adjoint loop. Once these stored fields have been
used up, we recompute the direct flow fields starting from the next earliest checkpoint.
Once these fields have been restored, we continue with the integration of the adjoint
equation. This direct-recovery adjoint-integration procedure continues until we reach
the initial time t = 0. It is obvious that we trade efficiency in using the available
memory with inefficiency in simulation time, as most direct fields are computed twice:
once during the direct loop and another time during the recovery part of the adjoint
loop. The extra work is hence the cost of an additional direct loop. The distribution of
the checkpoints is critical for the overall efficiency; libraries dealing with the storage
and recovery management are readily available (Griewank & Walther 2000, Wang
et al. 2009).

Miscellaneous
The previous sections outlined the most common computational tools to speed up
the optimization by direct-adjoint looping. More recently, additional techniques have
arisen and are currently validated and implemented.
For large-scale systems and moderate-to long-time horizons, a parallelization
technique in time may be advantageous. These techniques, which break the full-
time interval [0, τ] into p disjoint subintervals on which the governing equations are
solved in parallel and subsequently adjusted and merged into a global solution, have
been developed over the past five decades and have seen many applications in the
computational sciences. Parallel-in-time techniques are particularly attractive for the
solution of the adjoint equations, as the adjoint equation is linear by design, and can
thus be solved efficiently by splitting the homogeneous and inhomogeneous part of
the solution and treating them parallel in time.
Model reduction techniques to speed up the direct and adjoint loop of the iterative
optimization scheme are also being developed for large-scale applications. These
techniques provide approximate gradient and sensitivity information based on a
reduced description of the full dynamics.

h d i 0 0 9 8 088962 4 0 9 bli h d li b C b id i i
https://doi.org/10.1017/9781108896214.019 Published online by Cambridge University Press

Direct Adjoint
t¯
(a)

t
Direct adj dir adj dir adj dir adj direct adjoint
t1 t¯
(b)
1 1
2
3
4
t2 5
6
7
2 2
3
4
t3 5
6
7
3 3
4
t4 5
6
7
4 4
t5 5
t6 6
t7 7
τ 5
6
7 extra work
t
Figure 13.6 Schematic of checkpointing. (a) Checkpoints at all 25 time steps of the direct problem are stored and used as variable coefficients in
the adjoint part of the iteration. (b) Checkpoint at 7 time steps (filled blue circles) are stored during the direct solution; they are used as starting
points (initial conditions) for launching direct simulations to restore direct solutions for the entire interval [0, τ]. We note that exactly seven
checkpoints are stored at all stages of the process.
 280 Data-Driven Fluid Mechanics Chapter 13

Receptivity

Acoustics

Instab
ility

Instability

Figure 13.7 Sketch of tonal noise. Flow around a NACA-0012 airfoil at an angle of attack
develops a shear instability on the suction and pressure side. Noise radiates from the trailing
edge upstream and downstream, triggering via a receptivity process (indicated by red symbols)
instabilities in the pressure-side and suction-side boundary layer, hence closing the feedback
loop.

13.6 Tonal Noise Analysis: Quantifying a Complex Feedback

Mechanism

Tonal noise is an aeroacoustic process by which sound is radiated from a moving body
and maintained via a hydrodynamic-acoustic feedback mechanism. It has been studied
theoretically and experimentally over many years, and recently direct numerical
simulations have joined earlier investigations in trying to uncover the key elements
of the underlying feedback loop (Fosas de Pando et al. 2017). Tonal noise on airfoils
typically appears at chord-based Reynolds numbers of Re ∼ 105, with applications
to glider airplanes and wind turbines. Tonal noise in these latter applications is
undesirable, and control schemes – active or passive – are being developed to suppress
or eliminate radiated sound. In order to do so, a full understanding of the global tonal-
noise instability is required after which we can attempt to break the feedback loop at
the weakest link.
The optimization framework, outlined in this chapter, is particularly suited to
analyze the key elements of the tonal noise instability; the same framework can also be
used, in a second step, to design control schemes or to suggest geometric modifications
to weaken or suppress the undesired tonal effects.
Our understanding of tonal noise is based on a sequence of physical processes
as follows. The boundary layers on the suction and pressure side of the airfoil
support instabilities on either side. Convected downstream toward the trailing edge, the

h d i 0 0 9 8 088962 4 0 9 bli h d li b C b id i i
 Schmid: Modern Tools for the Stability Analysis of Fluid Flows 281

100
0.2
10−1
0.0
M1 10−2
−0.2
M
L
H 10−3
ωi −0.4

10−4
−0.6

−0.8 10−5

−1.0 10−6
0 10 20 30 40 50 60 70
ωr

Figure 13.8 Global spectrum of the tonal noise problem in the complex frequency plane. The
eigenvalues are colored according to the size of the relative residual with respect to the
linearized operator. The spectrum is divided into least stable modes (labeled M),
low-frequency modes (labeled L) and high-frequency modes (labeled H).

instability structures collide and generate a noticeable sound wave. This wave travels
radially outward and reinitiates the instabilities in the pressure-side and suction-side
boundary layers through a receptivity process. The feedback loop is thus closed.
Figure 13.7 presents a sketch of this scenario. Two feedback loops – one based on
the pressure-side boundary layer, one based on the suction-side boundary layer –
linked at the trailing edge characterize the overall tonal noise process. It seems
obvious that a local stability analysis, based on frozen boundary-layer velocity profile,
cannot account for the complexity of the hydrodynamic-acoustic loop and cannot
quantitatively capture the main features. Instead, a global perspective is necessary,
and a DNS-based analysis beyond local instabilities should be attempted.
We consider a chord-based Reynolds number of Rec = 200,000, a Mach number
of Ma = 0.4, and describe the flow by the two-dimensional compressible Navier–
Stokes equation. The airfoil has a NACA-0012 profile and a 2o angle of attack to the
free-stream flow direction. The specific heat γ and Prandtl number Pr are taken as
constant, with values 1.4 and 0.71, respectively. At these parameter values, the mean
flow separates and reattaches on both sides.
We can gain insight into the tonal noise problem by first looking at the global
spectrum of the governing equations, linearized about the mean state (Fosas de
Pando et al. 2014). This is accomplished using an Arnoldi algorithm coupled to
a linearized simulation code for the physical problem; alternatively, a dynamic
mode decomposition (DMD) could have been used. As outlined earlier, we are
less interested in the modal solutions that represent the time-asymptotic structures
(possibly) encountered as τ → ∞, we compute the global spectrum (see Figure 13.8)
to provide a first picture of preferred frequencies supported by the linear dynamics.
The global spectrum consists of various branches, showing equispaced, discrete
eigenvalues. We classify these branches according to their frequency range or the

h d i 0 0 9 8 088962 4 0 9 bli h d li b C b id i i
 282 Data-Driven Fluid Mechanics Chapter 13

0.010
M1, p 0.008
1.0

(a) 0.006

0.5 0.004

0.002

0.0 0.000
y

−0.002
−0.5 0.05 −0.004
0.00
−0.006
−0.05
−1.0 −0.008
0.5 0.6 0.7 0.8 0.9 1.0 1.1
−0.010
−1.0 −0.5 0.0 0.5 1.0 1.5 2.0
x

108
(b)
0.6
M1, |u∗ |
0.4 107

0.2

0.0 106
y

−0.2

−0.4 105

−0.6

104
−0.5 0.0 0.5 1.0 1.5
x

Figure 13.9 (a) Spatial structure of the global mode labeled M1 , visualized by the real part of
the associated near-field pressure levels, and the real part of the streamwise velocity levels in
the vicinity of the aerofoil surface (inset). The mode has been normalized by the maximum
value of the velocity field in the near wake 1 < x < 1.2. (b) Spatial structure of the associated
adjoint global mode (labeled M1 in Figure 13.8), visualized by the magnitude of the
streamwise velocity levels, and the real part in the insets.

structure of the corresponding modes. They are outlined in Figure 13.8. Inspecting
the modes across the full frequency range reveals that only the modes labeled M,
which are also among the least damped modes, exhibit significant acoustic activity.
A similar analysis shows that the low-frequency modes (designated L in the figure)
represent the separation bubble dynamics and the reattachment dynamics. The high-
frequency modes (marked by H in the figure) describe the Kelvin–Helmholtz-type
shear layer instabilities on the suction side. The modes labeled H or L do not show a
large acoustic component, and thus shall be ignored in our analysis of the tonal noise
problem. Instead, we concentrate on the M-modes.

h d i 0 0 9 8 088962 4 0 9 bli h d li b C b id i i
 Schmid: Modern Tools for the Stability Analysis of Fluid Flows 283

We then focus on the least stable mode, labeled as M1 in the figure. A visualization
of its spatial structure (see Figure 13.9(a)) demonstrates all components of the
hypothesized feedback loop: the boundary layer on the suction side clearly shows an
instability, and the collision of instabilities from the suction and pressure side creates
sufficient shear to trigger an acoustic wave that emanates from the trailing edge and
radiates omnidirectionally. Other modes from the M-branch show a similar structure,
at different frequencies, but with only minor differences in their spatial composition.
After the dominant structure, with a significant acoustic component, has been
identified, we need to find its structural sensitivity from the adjoint structure. In
essence, this undertaking asks the question of how to trigger or influence the M1 -
mode in an optimal manner. The answer to this question should point at regions of the
flow where a minimal effort (input) is necessary to stimulate the occurrence of tonal
noise. The flow field adjoint to the M1 -structure will give this information, since it
is directly proportional to the amplitude of a general perturbation as it projects onto
M1 . In other words, the adjoint mode associated with M1 will identify the location,
shape, and state-vector components where the feedback loop that sustains tonal noise
(based on M1 ) can most easily be broken. Figure 13.9(b) shows the adjoint M1 -mode;
it has been visualized by the magnitude of the adjoint streamwise velocity. A localized
support of this structure is found on the pressure side of the airfoil, between roughly
20% and 50% of chord. We can argue that the tonal noise structure associated with M1
has its “origin” at this location, and it is at this location where tonal noise can most
easily be manipulated. Early experiments and in more recent numerical simulations
have proposed this position, although without an associated analysis.
The above analysis shows that complex stability/receptivity issues require math-
ematical and computational tools that transcend local analyses and time-asymptotic
assumptions. An optimization-based framework using direct as well as adjoint
information about the flow is sufficiently flexible and effective (once implemented in
an efficient manner) to answer many of the questions of flow analysis and to provide
a more complete and satisfactory picture of fluid flow behavior.

13.7 Further Extensions

We have demonstrated the optimization framework based on the direct governing

equations and the adjoint equations (carrying sensitivity information) on a physically
complex, but still comparatively simple, application: the aeroacoustic phenomenon
of tonal noise. It is easy to recognize that by studying this example, we have
only scratched the surface of what can be accomplished by this type of analysis.
Many generalizations are within the range of an optimization-based computational
framework; many of them have been realized over the past years. Minimal-seed
computations for transition to turbulence, nonlinear harmonic analysis of driven fluid
systems, the exploration of regularizing constraints on the optimization are but a few
extensions that will bring new ideas to the quantitative analysis of fluid flows.

h d i 0 0 9 8 088962 4 0 9 bli h d li b C b id i i
 284 Data-Driven Fluid Mechanics Chapter 13

13.8 Summary and Conclusions

Hydrodynamic stability theory has played a central role in the development of

theoretical fluid mechanics. It has motivated the development of a great many
mathematical and computational tools to probe the behavior of perturbations about
a steady or quasi-steady flow state and has formed the foundation for subsequent
research fields, such as flow control. Starting in the beginning to determine the large-
time stability of generic configurations, it has matured into a sophisticated set of tools
applicable to large-scale systems and multiscale simulations. During this process,
many fundamental assumptions and limitations of early stability theory have been (and
will continue to be) challenged and abandoned. They will be replaced by a general and
flexible numerical platform based on complex optimization, fast solvers, and parallel
algorithms. In this manner and form, modern hydrodynamic stability theory will assert
its central position in the quantitative analysis of complex fluid systems.

h d i 0 0 9 8 088962 4 0 9 bli h d li b C b id i i
 Part V
Applications

h d i 0 0 9 8 088962 4 020 bli h d li b C b id i i

h d i 0 0 9 8 088962 4 020 bli h d li b C b id i i
 14 Machine Learning for
Reduced-Order Modeling
B. R. Noack, D. Fernex, and R. Semaan

We describe data-driven reduced-order modeling (ROM) approaches using powerful

methods of machine learning. The focus is gray-box models (GBMs) distilling coher-
ent structure dynamics from snapshot data. We highlight two generally applicable
methods: the proper orthogonal decomposition (POD) Galerkin method and cluster-
based models. The Galerkin method has deep roots in the mathematical investigation
of the Navier–Stokes equations and comes with illuminating physical insights. Yet,
the accurate data-driven variant requires a rich set of tuning tools and tends to
have a narrow range of validity. As an alternative, the recently developed cluster-
based network model is inherently robust, can easily integrate numerous operating
conditions, and its construction can fully be automated. However, we lose connection
to first principles. Of course, there are many different shades of gray, which will briefly
be reviewed. The chapter concludes with a tutorial of xROM, a freely available software
for POD and cluster modeling for different type grids and a large volume of data.

14.1 Introduction

ROM has a long tradition in fluid mechanics. The focus of this chapter is gray-box
modeling,1 which resolves the coherent structure flow dynamics. Over 500 years ago,
Leonardo da Vinci made the first paintings of wake vortices. von Helmholtz (1858)
laid the foundation of their dynamic modeling with his famous vortex laws. In the
years that followed, many vortex models were proposed. Examples are a pair of two
equal vortices rotating about their axis, a pair of two equal but opposite vortices
moving uniformly, leapfrogging ring vortices, the Föppl (1913) vortex model of the
near wake, the von Kármán (1911) vortex model for vortex shedding, a shear-layer
vortex model (Hama 1962), the recirculation zone model (Suh 1993), the vortex-
model-based feedback control (Noack et al. 2004, Protas 2004), just to name a few.
The reduced-order vortex models inspired high-dimensional simulations with vortex
blobs (Leonard 1980), and vortex filaments (Ashurst & Meiburg 1988). Vortex models
are closely tied to first principles and robustly model coherent structures and their
convection.

1 The term is adopted from Wiener (1948).

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 288 Data-Driven Fluid Mechanics Chapter 14

Another branch of reduced-order models can be traced back to stability analysis of

Orr (1907) and Sommerfeld (1908) and the Galerkin (1915) method. These founda-
tions corroborated the idea that fluid dynamics can be understood as a superposition of
global physical modes. Landau (1944) and Hopf (1948) even considered turbulence
as originating from a sequence of many Hopf bifurcations. The mean-field theory
of Stuart (1958) beautifully explained the nonlinear saturation of instability modes.
Subsequently, the Galerkin method was not only applied to instability modes but
also to mathematical modes from Hilbert space considerations (Busse 1991, Noack
& Eckelmann 1994). As with vortex models, the Galerkin method was increasingly
generalized and became a foundation of computational fluid mechanics.
Data-driven reduced-order models have arguably originated from Galerkin methods
with a POD expansion (Lorenz 1956, Lumley 1967). The POD of a given snapshot
ensemble yields a Galerkin expansion with minimal representation error for given
number of modes. The pioneering wall turbulence model of Aubry et al. (1988)
demonstrated their use in explaining complicated dynamics. In the last three decades,
a myriad of POD models have been proposed. The initial excitement of POD was
mitigated by the increasing realization of the fragility of the resulting Galerkin models
(Rempfer 2000) and their limited range of validity. Galerkin methods are detailed in
Section 14.2.
A recent development of cluster-based reduced-order models (CROM) was pio-
neered by Burkardt et al. (2006) and Hof et al. (2004). Here, the dynamics are modeled
as a stochastic transition between a small number of representative flow states. The
k-means clustering of snapshots yields centroids as representative flow states with
a minimal representation error. The dynamics can be framed in a Markov (Kaiser
et al. 2014) and network model (Li et al. 2021). In Section 14.3, this branch of ROM
is discussed including the latest innovations (Fernex et al. 2019, 2021).
Of course, there are different shades of gray with GBMs, which are outlined in
Section 14.4. The chapter concludes with a tutorial of POD modeling (Section 14.5)
with freely available software by the authors.

14.2 Proper Orthogonal Decomposition Galerkin Model

This section gives recipes and suggestions for the construction of POD Galerkin
models. First, the considered configurations are specified in Section 14.2.1. Then,
Sections 14.2.2 and 14.2.3 describe the kinematical and dynamical steps to a POD
model. Sections 14.2.4 and 14.2.5 provide closures for an existing POD model and
identification methods for experiment. Section 14.2.6 reviews common applications.

14.2.1 Flow Configuration

We assume an incompressible viscous flow in a steady domain Ω. A location in this
domain is denoted by x = (x, y, z) = (x1, x2, x3 ). The time is represented by t. The
velocity and pressure fields read u = (u, v, w) = (u1, u2, u3 ) and p, respectively. The
Newtonian fluid is described by the density ρ and dynamic viscosity µ.

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 Noack et al: Machine Learning for Reduced-Order Modeling 289

Let D and U be characteristic size and velocity scales of the configuration.

The flow properties are characterized by the Reynolds number Re = ρUD/µ, or,
equivalently by its reciprocal ν = 1/Re. In the sequel, we assume that all quantities
are nondimensionalized with D, U, ρ, and µ.
The continuity and Navier–Stokes equations read

∇ · u = 0, (14.1)
∂t u + ∇ · (u ⊗ u) = −∇p + ν4u. (14.2)

The Dirichlet boundary conditions at stationary walls x ∈ ∂Ω are

u(x, t) = 0. (14.3)

In addition, a free-stream, stress-free, or von Neumann condition may be used for open
flows.
The initial condition u(x, 0) at time t = 0 may be a perturbed steady solution u s (x)
as in Chapter 1 of this book.
In what follows, we assume to have gathered M snapshots u m , m = 1, . . . , M
from a post-transient flow solution. The snapshots are requested to be statistically
representative for the computation of first and second moments.

14.2.2 Proper Orthogonal Decomposition

POD is a data-driven realization of a Galerkin approximation
N
Õ
û(x, t) = u0 (x) + ai (t) ui (x) (14.4)
i=1

of a velocity field u(x, t), as detailed in Chapter 6. Here, u0 is the basic mode, for
example, the steady Navier–Stokes solution or the mean flow, and ui are expansion
modes with amplitudes ai . These approximations rely on a square-integrable Hilbert
space L 2 (Ω) equipped with an inner product between two elements v and w,
∫
(v, w)Ω := d x v · w. (14.5)
Ω

The corresponding norm reads

p
kuk Ω := (v, w)Ω . (14.6)

POD modes are an orthonormal set of expansion modes that minimize the in-
sample representation error
M
1 Õ m
E= k u − û m kΩ2 , (14.7)
M m=1

with respect to the snapshot ensemble.

The computation of the POD modes is performed in the following steps:

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 290 Data-Driven Fluid Mechanics Chapter 14

Computation of the basic mode: The basic mode is a time-averaged flow and
absorbs a potential inhomogeneity of the steady boundary condition,
M
1 Õ m
u0 (x) := u (x). (14.8)
M m=1
One example is the uniform free-stream condition u = U ê x in the far-field,
where ê x is the unit vector in streamwise direction. The remaining fluctuation
u 0 = u − u0 satisfies homogenized boundary conditions, for example,
u 0 = 0 in case of the uniform free-stream condition. This implies that any
expansion mode ui will satisfy these homogenized boundary conditions. In
other words, the expansion (14.4) satisfies the full boundary conditions for
arbitrary choices of amplitudes ai , i = 1, . . . , N.
Computation of the correlation matrix: The M × M correlation matrix R has the
elements
1
R mn := (u m − u0, u n − u0 )Ω , m, n = 1, . . . , M. (14.9)
M
Spectral analysis of the correlation matrix: This Gramian matrix R is symmetric
and positive semi-definite. This implies real and nonnegative eigenvalues as
well as orthogonal eigenvectors. Let
 αi1 
 .
 
αi :=  ..


 
 αM 
 i 
be the ith eigenvector of the correlation matrix, that is,
R αi = λi αi , i = 1, . . . , M. (14.10)
Without loss of generality, we assume the real eigenvalues to be sorted in
decreasing order
λ1 ≥ λ2 ≥ · · · ≥ λ M = 0,
and that the eigenvectors are orthonormal
M
Õ
αi · α j = αim α m
j = δi j , i, j = 1, . . . , M.
m=1

Note that the Mth eigenvalue must vanish, because M snapshots span at most
an M − 1-dimensional hyperplane.
Computation of the POD modes: The POD modes are given by
M
1 Õ m m
ui (x) := √ αi (u (x) − u0 (x)) , i = 1, . . . , N. (14.11)
M λi m=1
Here, N ≤ M − 1 is the number of expansion modes in (14.4). The POD
modes are orthonormal by construction:
ui , u j Ω = δi j , i, j = 1, . . . , N.

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 Noack et al: Machine Learning for Reduced-Order Modeling 291

Computation of the amplitudes: The mode amplitudes scale – not accidentally –

with the corresponding eigenmodes of the correlation matrix
r
λi m
ai =
m
α ,
M i
and satisfy
M
1 Õ m
ai = a = 0, i = 1, . . . , N, (14.12a)
M m=1 i
M
1 Õ m m
ai a j = a a = λi δi j , i, j = 1, . . . , N. (14.12b)
M m=1 i j

14.2.3 Galerkin Method

The projection of the Navier–Stokes equations (14.2) on the expansion (14.4) is
performed in two steps. First, the inner product between the Navier–Stokes equations
and the mode ui is taken. Second, the resulting equations are integrated over the
domain Ω. This Galerkin projection yields
N N
dai
liνj a j +
Õ Õ
=ν qicjk a j ak i = 1, . . . , N, (14.13)
dt j=0 j,k=0

with the Galerkin system coefficients

liνj = ui , 4u j Ω, (14.14a)

qicjk = ui , ∇ · u j ⊗ u k . (14.14b)



Ω

For analytical convenience, a0 = 1 following Rempfer and Fasel (1994b). The

pressure term vanishes for many boundary conditions. Otherwise, it leads to an
additional quadratic term (Noack et al. 2005). Again, the reader is asked to refer to
the original literature for the details.
Even for a 100% accurate Galerkin approximation with vanishing representation
error, there is no guarantee that the stability properties of the Navier–Stokes equations
are conserved in the Galerkin system. A stable Navier–Stokes solution may become
an unstable Galerkin solution (Noack et al. 2003, Rempfer 2000). In fact, Galerkin
systems are often observed to be fragile with unphysically long transient times or
even finite-time divergence to infinity. Schlegel and Noack (2015) provide necessary
and sufficient conditions for bounded solutions only based on the Galerkin system
coefficients. Section 14.2.4 suggests cures.

14.2.4 Closures and Stabilizers

For turbulent flows, the representation error of the POD expansion may well be 50%
or larger. This large neglected fluctuation needs to be accounted for in the Galerkin

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 292 Data-Driven Fluid Mechanics Chapter 14

a) b) c)
20 20
a1 a1
0 0
-10 -10
-20 -20
0 200 400 600 time 1000 0 200 400 600 time 1000

Figure 14.1 POD Galerkin model for the 3D mixing layer. For details see Noack et al. (2004).
(a) Snapshot; (b) a1 of LES; (c) a1 of POD model.

system. This residual has two effects: a high-frequency noise excitation and an energy
dissipation. Typically, only the energy dissipation is accounted for.

Mean-flow model: Aubry et al. (1988) account for the stabilizing coupling between
fluctuations and mean-flow by replacing the constant basic mode with
variable one computed from the Reynolds equation. The shift mode (Noack
et al. 2003) has the same purpose for an oscillatory flow.
Global eddy viscosity: Aubry et al. (1988) employs a Boussinesq ansatz by replacing
the kinematic viscosity ν with an effective one νeff = ν + νT , where the global
eddy viscosity νT is a tuning parameter. This ansatz will be dissipative but
the Galerkin model inaccurately implies that a high-Reynolds number flow
behaves like the laminar solution at a low-Reynolds number.
Modal eddy viscosities: Rempfer and Fasel (1994a) has refined this Boussinesq
ansatz by a more realistic mode-dependent eddy viscosity – inspired by
spectral turbulence theory.
Nonlinear eddy viscosities: Global and modal eddy viscosities imply that a nonlin-
ear dynamics can be approximated by a linear Galerkin system term. Östh
et al. (2014) have refined this ansatz by a fluctuation-dependent scaling. The
scaling has been justified with a finite-time thermodynamics closure (Noack
et al. 2008) and can be shown to guarantee boundedness (Cordier et al. 2013).

We keep the list of closures this short and simple. Figure 14.1 previews results of a
modal eddy viscosity closure for a turbulent mixing layer.

14.2.5 Model Identification

A Galerkin projection is also possible on subdomains. However, the Navier–Stokes
equations residual needs to be evaluated (Rempfer 1995). 2D particle image velocime-
try (PIV) field measurements of 3D flows do not provide the plane normal velocity
components, let alone their normal derivatives. In some cases, for example, non-
Newtonian fluids, the evolution equation may not be known. These are motivations
for model identification. We keep the ansatz of a constant-linear-quadratic dynamics,

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 Noack et al: Machine Learning for Reduced-Order Modeling 293

N N N
dai Õ ÕÕ
= fi (a)ci + li j a j + qi jk a j ak , (14.15)
dt j=1 j=1 k=j

with the state a := [a1, . . . , a N ]T .

A classical calibration method aims to minimize the state difference between
T
the computed/measured POD mode amplitudes a• (t) = a1• (t), . . . , a•N (t) and the
modeled one a◦ (t),
∫t1
1
E1 := dt k a• − a ◦ k 2 = min. (14.16)
t1 − t0
t0

This formulation requires the integration of the model in the time interval [t0, t1 ]. The
4D Var method is a powerful method for minimizing the model error (Semaan et al.
2015). Yet, the challenge is that a short interval may not provide enough information
about the dynamics while a long interval leads to inevitable “phase drifts” between the
full plant and the model. If the full plant and the model are “out of sync,” an overly
dissipative dynamics with a◦ ≡ 0 will become the best model.
An alternative optimization avoids this phase-drift problem. Now, the dynamics
residual is minimized along the full-plant trajectory t 7→ a• ,
∫t1 2
1 da
•
dt − f (a ) = min.
•
E2 := dt (14.17)


t1 − t0
t0

Here, a long time integral helps in the model accuracy as the whole attractor is
incorporated in the calibration. This optimization requires the temporal state derivative
and leads to an analytically solvable least-means-square problem. The challenge is that
a long-term accumulation of errors, like amplitude errors, can easily occur, because
they are weakly penalized by the error (14.17).
A large arsenal of methods have been developed for improving long-term asymp-
totics, for mitigating the effect of insufficient data (Cordier et al. 2010), and for sparsi-
fying the dynamics, that is, make it more human interpretable (Brunton et al. 2016a).
The reader is encouraged to visit the literature if needed.

14.2.6 Applications
Reduced-order models have many classical applications:

Understanding: A sufficiently low-dimensional sparse ROM may help to understand

the dynamics by distilling the human-interpretable modes and the key linear
and nonlinear terms.
Analysis: The Galerkin method provides powerful analytical frameworks for the
energy-flow analysis and for explanation of aerodynamic forces and torques
(Noack et al. 2011).

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 294 Data-Driven Fluid Mechanics Chapter 14

Estimation: A dynamic ROM may be the basis of a dynamic observer, that is, may
help to predict flow states based on one or few sensor signals.
Prediction: A dynamic ROM may help to predict future states. One example is an
undesirable flutter in an airplane experiment and an early shutdown of the
wind-tunnel. A second example is a weather forecast.
Exploration: A ROM may help to explore unobserved behavior for new initial
conditions or new parameters. Given the narrow range of validity of data-
driven ROM, this new behavior needs to be validated in the full plant.
Control: ROM may help in the control design of first- and second-order dynamics.
Closures: ROM may guide closure terms for unsteady dynamics in the spirit of Liu
(1989).
Response model: ROM may also provide a mapping from actuation or configuration
parameters to the time-averaged flow (Chapter 16) or the performance
(Albers et al. 2020).

14.3 Cluster-Based Reduced-Order Modeling

CROM can be considered as an antidote to POD Galerkin models. POD expansions

are replaced by discrete representative flow states, centroids. And the analytically
elegant POD Galerkin system is sacrificed for a dynamic transition model between the
centroids, loosing the possibility to perform stability analyses, compute energy flows,
and so on. The big advantage of CROM is the possibility to automate the complete
model identification from snapshot data and to arrive at an inherently robust dynamical
model. Unlike POD models, CROM do not require stabilizing auxiliary models, such
as shift modes, calibrated cubic terms, or tuned eddy viscosity parameters. In contrast
to POD, CROM requires time-resolving snapshots u m (x) = u m (x, t m ), t m = m∆t
with fixed time step ∆t. In the following, we outline the kinematic compression with
clustering (Section 14.3.1) and present two CROM, one focusing on statistics (Section
14.3.2) and one better resolving dynamics (Section 14.3.3). Section 14.3.4 outlines
generalizations.

14.3.1 Clustering
Following Chapter 1, we employ clustering as an autoencoder. Let ck (x), k = 1, . . . , K
be the centroids, that is, representative flow states. For a given flow field u, the encoder
provides the index of the closest centroid

k(u) = arg min ku − ci kΩ . (14.18)

i=1,...,K

The decoder H returns the corresponding centroid

û = H(u) = ck(u) . (14.19)

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 Noack et al: Machine Learning for Reduced-Order Modeling 295

a) b)
1 2 3 4 5 6 7 8 9 10
1
2
3
4
5
6
7
8
9
10

Figure 14.2 Cluster-based Markov model for the mixing layer. For the details, see Kaiser et al.
(2014). (a) Vorticity snapshot; (b) Markov matrix P.

In analogy to POD, the set of K centroids are chosen to minimize the in-sample
representation error
M
1 Õ m
E= k û − u m kΩ2 , (14.20)
M m=1

with respect to the snapshot ensemble. The number of centroids K is a design

parameter. Burkov (2019) provides a useful criterion based on an out-of-sample error.
Li et al. (2021) propose an alternative criterion based on the dynamic prediction error.

14.3.2 Cluster-Based Markov Models

The dynamics may be stroboscopically monitored at the sampling time step ∆t. Let ni
be the number of snapshots in cluster i and ni j be the number of observed transitions
from cluster j to cluster i in the next time step. Then, Pi j = ni j /ni estimates the
probability to go from cluster j to cluster i. Let p m = [p1, . . . , pK ]T comprise the
probabilities to be in any of the clusters at time t m . Then, the Markov model advances
this probability vector according to
p m+1 = P p m, P = (Pi j ). (14.21)
Under reasonably generic conditions and sufficient data, the iteration (14.21) will
converge against a fixed point approximating the cluster population of the snapshots
p∞
i ≈ qi := ni /M. In the sequel, the model will be referred to as a cluster-based
Markov model (CMM).
Kaiser et al. (2014) provides the mixing layer (see Figure 14.2) and the Ahmed
body wake as examples for CMM. The statistics of CMM is very accurate. On the
downside, the temporal evolution is too diffusive for a small number of clusters.

14.3.3 Cluster-Based Network Models

The diffusion effect has been mitigated by a recent network model (Li et al. 2021).
Let Qi j be the data-inferred probability of a trajectory to go from cluster j directly to
cluster i. Let Ti j be the corresponding
 averaged transition time. Both parameters are
comprised in K × K matrices Q = Qi j and T = Ti j . And, for concreteness, let us
  

assume we start at centroid k = 1 at time t0 = 0. Now, the cluster-based network model

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 296 Data-Driven Fluid Mechanics Chapter 14

(CNM) stochastically generates a sequence of times t0, t1, t2, . . . and cluster affiliations
k 0 = 1, k1 , k2, . . . consistent with the mentioned direct transition probabilities Qi j and
transition times Ti j . The trajectory for t ∈ [tm, tm+1 ] moves uniformly from centroid
ckm to ckm+1 ,
tm+1 − t t − tm
u ◦ (x, t) = ck (x) + ck (x). (14.22)
tm+1 − tm m tm+1 − tm m+1
In contrast to CMM, this model has a time-continuous flow representation u ◦ (x, t),
that is, does not jump discretely between centroids. Moreover, the model-based
autocorrelation function is found to be much more accurate. We refer to Li et al.
(2021) for a detailed discussion of the model and the shear-flow example. The price for
this increased dynamic resolution is a slightly larger error in the model-based cluster
population.

14.3.4 Generalizations
CROM, in particular CNM, can be significantly generalized. One model may describe
many operating conditions, if we enrich the set of centroid correspondingly and
include the parameters b of the operating conditions in the model parameters, for
example, P(b) for the CMM and Q(b), T (b) for the CNM (Fernex et al. 2021, Li
et al. 2021). In addition, the assumed motion from centroid to centroid may be relaxed
to a “flight” using the centroids and their transition characteristics as lighthouses. With
these generalizations, a CROM has successfully approximated a dragreduction study
of wall turbulence with dozens of spanwise traveling surface waves (to be published
soon by the authors). Summarizing, CROM is based on a local interpolation between
centroids as “collocation points” of the data set, and there is no limit to the structure of
the data. The simple nature of clustering makes it a solid foundation for a very general
modeling toolkit. The loss of analytical relationships to first principles is rewarded by
significant gain in human-interpretable model complexity.

14.4 General Principles: Different Shades of Gray

This section outlines general principles and alternative realizations of data-driven

GBMs. The notation is adopted from Wiener (1948) who distinguished between white-
box models (WBMs) for the full plant, black-box models (BBMs) for input–output
dynamics, and GBMs for low-dimensional state dynamics. Many GBMs follow the
diagram (14.23).
WBM GBM state estimate

Kinematics u(x) → a → û(x) (14.23)

↓ ↓
∂u da
Dynamics = N (u) ← = f (a)
∂t dt

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 Noack et al: Machine Learning for Reduced-Order Modeling 297

The first row defines an autoencoder, that is, a low-dimensional flow representation
with an encoder to and decoder from the low-dimensional feature vector a. The
dynamics of this feature vector (bottom, middle) approximates the Navier–Stokes
equations (bottom, left). The POD and cluster-based models follow this scheme. Also
vortex models fit into this approach.
GBM models differ in the chosen autoencoder and the chosen model identification
of the dynamics f (a). The autoencoder may be mathematical (Fourier modes,
Tchebychev polynomials, etc.), physical (stability modes, Stokes modes, etc.), or
data-driven (POD, DMD, centroids, etc.). The dynamics may be derived from first
principles (Galerkin projection, etc.), or from data (model identification), or a mixture
of both (model identification with a Thikonov penalization of the Navier–Stokes-based
model, etc.). We shall not pause to declinate the different shades of gray, but refer to
the excellent literature on the topic (Fletcher 1984, Holmes et al. 2012).
Instead, we mention two different approaches. Feature-based manifold models
(FeMM) (Loiseau et al. 2018, 2021) assume a typical experimental situation. A sensor
signal s is recorded in time, lifted to a dynamic feature a for which a dynamic model
can be identified. The feature a is employed for an estimator of the flow using PIV and
simultaneous sensor data. K-nearest neighbor interpolation is a simple yet effective
method for state estimation. The following equation outlines the data structure:
Kinematics s → a → û(x)
↓ (14.24)
da
Dynamics = f (a).
dt
Effectively, FeMM is a BBM with a sensor-based estimator of a manifold.
Another approach relies on brute-force data interpolation. Let us assume that the
data contains the snapshots u m and the time derivative ∂t u m at the same instants, for
example, from double-shot PIV. Now, the flow field u ◦ can be integrated by an Euler
scheme,
u ◦ (t + ∆t) = u ◦ (t) + ∆t ∂t u ◦ (t).
In the simplest 1-nearest neighbor realization, the time derivative is taken to be ∂t u m ,
where u m is the nearest neighbor to u ◦ . Obvious refinements, like higher-order
interpolation and smoothing are possible and probably advised.
We pause the discussion of the spectrum of GBM enabled by machine learning.
Evidently, the choice of the GBM is guided by the available data and by the purpose
of the model. It pays to start with a simple purely data-driven model before advancing
to more refined and targeted ones. Moreover, every mapping can be identified with a
rich set of tools – from Taylor’s expansion to neural networks.
We conclude with a few words of warning. Data analysis is easy, insightful, and fun.
Even with validation procedures, results are guaranteed. Dead data cannot complain
anymore. Dynamic models attempt to predict the unknown future, which is easy
for attractor dynamics, that is, transitions between similar recurrent states, but is
next to impossible for extrapolation into terra incognita. Data-driven control-oriented
dynamic models attempt to predict the unknown future with a rich set of possible

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 298 Data-Driven Fluid Mechanics Chapter 14

actuations. The available data for control-oriented modes will typically be sparse, that
is, data has to be replaced by good simplifying guesses.

14.5 Tutorial: xROM

This tutorial aims to give readers a hands-on experience with the learned Galerkin pro-
cesses. This includes Galerkin expansion, Galerkin projection, and model calibration.
The tutorial shall solely rely on the numerical package xROM (Semaan et al. 2020),
which is a freely available tool for ROM. One purpose of xROM is to perform a modal
decomposition from snapshot data and to carry out the Galerkin projection using the
Navier–Stokes equations. Moreover, xROM
(i) can analyze native CFD and PIV data;
(ii) can handle a large spectrum of 2D and 3D grids (Cartesian, structured, unstruc-
tured, etc.);
(iii) can create modal expansions and other modes (e.g., shift modes);
(iv) is user-friendly through a single configuration file;
(v) is parallelizable.
On the full capabilities of xROM, the reader is referred to the user manual, which is
included in the package.
For this tutorial, we shall employ direct numerical simulations (DNS) of a cylinder
flow at Re = 100 following Chapter 1 and duplicating the results presented in Noack
et al. (2003).
The tutorial is organized as follows. In Section 14.5.1, we shall get acquainted with
the numerical setup and the considered flow. General characteristics of xROM and how
to use it are introduced in Section 14.5.2. We shall deploy xROM to perform three tasks
in Section 14.5.3.

14.5.1 Numerical Setup

We consider the 2D cylinder wake at Re = 100. The DNS is performed with an
FEM of third-order accuracy in space and time. The computational domain ΩDNS is
bounded by the rectangle −5 < x < 15 and −5 < y < 5, excluding the cylinder
Ωcyl = (x, y) : x 2 + y 2 ≤ 1/2 . Details can be inferred from Noack et al. (2003).
Figures 14.3 and 14.4 provide an impression of the domain, the grid, and the periodic
flow. In the ROM analysis, the observation domain Ω = ΩDNS is identical to the
computational domain.

14.5.2 General Characteristics of xROM

xROM capabilities
xROM is mainly designed to generate POD Galerkin models for a wide range of flow
problems. To accomplish this, it has several additional preprocessing tools and options

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 Noack et al: Machine Learning for Reduced-Order Modeling 299

Figure 14.3 Unstructured grid for the FEM direct numerical simulation of cylinder wake.

Min=-04.998
Max= 05.027

Figure 14.4 Streamlines of a snapshot.

that extend its capabilities and configurability. The main features that make the full
version of xROM attractive are presented in the following.

Preprocessing: Two tools are available for the preprocessing phase. First, a PIV
converter that directly converts PIV snapshots from their original .txt or
.dat format to snapshots readable by xROM. Second, a tool to reduce the
domain according to user-defined specifications, to compute the POD modes
only in a specific region of interest.
Formats: xROM currently reads/writes two formats: CFD General Notation System
CGNS, an efficient binary open-source format, and the ASCII format from
Tecplot, .dat, which is widely used in the research community.
Case types: xROM supports a wide variety of mesh types: Cartesian equidistant,
Cartesian, structured, and unstructured grids, in 2D and 3D.
Parallelization: The software is fully parallelized using Massively Parallel Imple-
mentation (MPI), which offers several benefits.

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 300 Data-Driven Fluid Mechanics Chapter 14

Requirements and Download

xROM is distributed as a single binary executable file, which makes its usage very easy.
It should work on any Linux distribution with a C library version 2.19 or newer. To
check the C library version, run in a terminal:
$ ldd -- version
For this tutorial, you can download a lighter version of the program that excludes
CGNS capabilities (link).2 The archive contains three items: the xROM software, the
documentation, and the current sample case. Once the archive is extracted, open a
terminal and go to where the binary is stored. Make sure that the binary is executable.
If not, run:
$ chmod + x xROM
To run xROM, the program needs the path to the input folder as argument. There are
two possibilities to define it: by specifying it directly as command line argument or
by using the pop-up window. The two ways to define the input folder are shown in the
following:

(a) (b)

Figure 14.5 (a) Cylinder wake folder incorrectly selected. (b) Cylinder wake folder correctly
selected.

Input folder from the command line: In that case, simply specify the correct path
to the case folder when running the program, such as:
$ ./ xROM / absolute / path / to / folder
Please note that it has to be the absolute path, which can be found using:
$ pwd
Input folder from the pop-up: If no path is defined in the command line, a pop-up
window opens in which the user must select the right folder. In this case, the
correct command is:
$ ./ xROM
Be careful to correctly select the folder: it is valid only when the field
“Selection” of the window contains the full path, including the input folder.
In practice, the user has to click twice on the case folder (i.e., enter the
2 https://cloudstorage.tu-braunschweig.de/getlink/fi9HQ68borJL8dJ9t4Yxv2au/

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 Noack et al: Machine Learning for Reduced-Order Modeling 301

folder) to get the path, unlike other common GUIs where the user only
have to click once. This is exemplified in Figure 14.5, which shows how
the cylinder wake folder is (a) wrongly and (b) correctly selected.

Configuration file and execution steps

The configuration file contains all the processes and their options to be executed: the
tools that will be used (PIV conversion, domain reduction, . . . ), the steps that will be
executed, and the parameters for each of these steps. It is the only required control
input from the user. All capabilities of xROM can be set up in this configuration file.
The general form of an input is
<parameter> : <value>
The detailed specifications are the following:
• There is only one parameter and its value per line.
• The parameter and its value are separated by a colon “:.”
• The spaces in the line are irrelevant.
• The position of a line in the file does not matter: it is for instance the same if a
parameter is defined in the first line or the last line.
• To deactivate an option, the user must add the comment symbol “#” at the beginning
of the line.
The standard process of xROM is divided into three main steps:
1. Galerkin expansion,
2. Galerkin projection,
3. dynamical system.
Each of these steps contains its owns options and tools in the configuration file. The
user can selectively execute individual steps. It is possible, for instance, to compute
only the Galerkin expansion or the Galerkin expansion and the projection without the
dynamical system.

14.5.3 Exercises
As already mentioned in Section 14.5.1, we shall use DNS snapshots of a
cylinder flow, which are included in the package under CylinderWake/
InputData/Snapshots. Since all three exercises use the same data set, we
shall use the same case folder and simply generate (automatically) a different
run directory for each exercise. Before each run, we shall make the necessary
modifications to the configuration file.

Exercise 1: Galerkin Expansion of the Cylinder flow

In this first exercise, we perform a POD on the computed snapshots. The configuration
file for this exercise is located under CylinderWake/ConfigFile, which you can
open with any text editor. Most of the options are self-explanatory. For more details,
you are referred to the user manual.

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 302 Data-Driven Fluid Mechanics Chapter 14

As you can see in ConfigFile, only the first step GalerkinExpansion is enabled.
The rest are disabled as these options are not relevant for this first exercise. Besides
activating GalerkinExpansion, the only other option we need to concern ourselves
with is the BaseFlow selection, which we choose as MeanFlow.
After executing the program using ./xROM, a new folder called OutputData will
be generated. Inside, the POD results are found under GalerkinExpansion. Here
you should find the output of the entire Galerkin expansion, which includes:
• The base flow as a Tecplot-compatible .dat file.
• The correlation matrix.
• The eigenvalues.
• The POD mode coefficients as an array, and as .png images under
PODAmplitudesPlot.
• The POD modes as Tecplot-compatible .dat files under PODModes.
• The POD spectrum as a .png image.
If you wish to plot the eigenvalues and the mode coefficients yourself, you can simply
disable the option plotFigures. All generated field data are saved in ASCII format
and are Tecplot-compatible.
Examining the modes and the mode coefficients reveals the expected results for this
shedding flow:
• The turbulent kinetic energy is highly concentrated in the first modes and drops
quickly with higher mode numbers.
• A strong mode pairing between (at least) the first 8 modes.
• The mode pairs exhibit a phase shift between them.
• Modes and mode coefficients are each dominated by a single frequency that
increases with higher mode pair.

Exercise 2: POD Galerkin Model without Calibration

Now, as we have computed the Galerkin expansion, we shall perform a Galerkin
projection and generate our first reduced-order model. To that end, we shall activate
the second (GalerkinProjection) and third (DynamicalSystem) steps in the
ConfigFile as follows:
Ga le rkinProjection : yes # ( yes , no ) }
DynamicalSystem : yes # ( yes , no )

GalerkinProjection and DynamicalSystem perform a Galerkin projection on the

POD modes and solve the dynamical system, respectively. You should leave all other
parameters unchanged. We note here the deactivated model calibration option.
After executing the program, two new directories for the Galerkin projection
and the dynamical system will be generated under outputData/Run2. Under
GalerkiProjection, you should now have three data files: ConvectionTensor,
ViscosityMatrix, and MassMatrix, which are the linear, quadratic, and mass
matrix terms from the projection. Since there is no forcing and the POD modes are
orthonormal, the resultant mass matrix is (within machine accuracy) a unity matrix.

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 Noack et al: Machine Learning for Reduced-Order Modeling 303

These coefficients are then used to generate the dynamical system, which is
integrated between time tStartDS and tEndDS. Inspecting the reference POD
mode coefficient against those of the dynamical system informs us of the model
accuracy. All eight reference and generated mode coefficients are found under
DynamicalSystem/DSAmplitudesPlot. Comparing both mode coefficients shows
a good match for the initial modes and a gradual degradation toward an unstable
behavior for the higher modes. Without calibration, this behavior is expected.

Exercise 3: POD Galerkin Model with Calibration

In the second exercise, the Galerkin projection yielded an unstable model. This
instability can be attributed to truncation errors, to structural instability of the Galerkin
projection, or to insufficient numerical precision. There exists many techniques to
calibrate a dynamical model (Östh et al. 2014). In this exercise, we shall rely on the
easiest and very accurate modal nonlinear eddy viscosity. You can simply enable it in
the ConfigFile as follows:
Calibration : ModalNonLinear # ( no ,
,→ ModalNonLinear )

After rerunning the program, you can inspect the mode coefficients under
DynamicalSystem/DSAmplitudesPlot. The improvement should be clear. The
new model shows better accuracy and stability for all modes.

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 15 Advancing Reacting Flow
Simulations with Data-Driven
Models
1 2 3 4
K. Zdybal , G. D’Alessio , G. Aversano , M. R. Malik ,
5
A. Coussement, J. C. Sutherland, and A. Parente

The use of machine learning algorithms to predict behaviors of complex systems is

booming. However, the key to an effective use of machine learning tools in multi-
physics problems, including combustion, is to couple them to physical and computer
models. The performance of these tools is enhanced if all the prior knowledge and
the physical constraints are embodied. In other words, the scientific method must
be adapted to bring machine learning into the picture, and make the best use of the
massive amount of data we have produced, thanks to the advances in numerical com-
puting. The present chapter reviews some of the open opportunities for the application
of data-driven reduced-order modeling of combustion systems. Examples of feature
extraction in turbulent combustion data, empirical low-dimensional manifold (ELDM)
identification, classification, regression, and reduced-order modeling are provided.

15.1 Introduction
The simulation of turbulent combustion is a very challenging task for a number of
aspects beyond turbulence. Combustion is intrinsically multiscale and multi-physics.
It is characterized by a variety of scales inherently coupled in space and time through
1 Kamila Zdybał acknowledges the support of the Fonds National de la Recherche Scientifique
(F.R.S.-FNRS) through the Aspirant Research Fellow grant.
2 Giuseppe D’Alessio acknowledges the support of the Fonds National de la Recherche Scientifique
(F.R.S.-FNRS) through the FRIA fellowship.
3 Gianmarco Aversano acknowledges the funding from the European Research Council (ERC) under the
European Union’s Horizon 2020 research and innovation programme under grant agreement No.
714605.
4 Mohammad Rafi Malik acknowledges the funding from the European Research Council (ERC) under
the European Union’s Horizon 2020 research and innovation programme under grant agreement No.
714605.
5 Alessandro Parente acknowledges the funding from the European Research Council (ERC) under the
European Union’s Horizon 2020 research and innovation programme under grant agreement No.
714605.

h d i 0 0 9 8 088962 4 022 bli h d li b C b id i i

 Zdybal et al: Advancing Reacting Flow Simulations 305

thermochemical and fluid dynamic interactions (Pope 2013). Typical chemical mecha-
nisms describing the evolution of fuels consist of hundreds of species involved in thou-
sands of reactions, spanning 12 orders of magnitude of temporal scales (Frassoldati
et al. 2003). The interaction of these scales with the fluid dynamic ones defines the
nature of the combustion regime as well as the limiting process in determining the
overall fuel oxidation rate (Kuo & Acharya 2012). When the characteristic chemical
scales are much smaller than the fluid dynamic ones, the combustion problem becomes
a mixing one (i.e., mixed is burnt (Magnussen 1981)): combustion and chemistry are
decoupled, and the problem is highly simplified. Likewise, for chemical timescales
much larger than the fluid dynamic ones, the system can be described taking into
account chemistry only, neglecting the role of fluid dynamics altogether.
The intensity of interactions between turbulent mixing and chemistry is measured
using the Damköhler number, defined as the ratio between the characteristic mixing,
τm , and chemical, τc , timescales:
τm
Da = . (15.1)
τc

In terms of the Damköhler number, Da  1 indicates a mixing-controlled, fast

chemistry process. On the other hand, Da  1 denotes a chemistry-controlled, slow
chemistry process. Most practical combustion systems operate at conditions character-
ized by a non-negligible overlap between flow and chemical scales. This is particularly
true for novel combustion technologies, where the use of diluted conditions and the
enhanced mixing leads to a Da distribution close to unity. This grants some control
on the combustion process, thanks to the increase of the characteristic chemical scales
compared to the mixing ones. In particular, the operating conditions (temperature and
compositions) can be adjusted in such a way that the emissions are kept below the
required values (Wünning & Wünning 1997, Cavaliere & de Joannon 2004, Parente
et al. 2011). The condition Da ≈ 1 is generally referred to as finite-rate chemistry, to
indicate that combustion is not infinitely fast but of finite speed.
Modeling finite-rate combustion regimes is very challenging because both fluid
mechanics and chemistry effects must be accounted for accurately. In particular,
chemistry cannot be described using simplified global mechanisms, and this results
in a significant computational burden of combustion simulations. The resolution
of turbulent combustion problems requires the resolution of hundreds of transport
equations for (tightly coupled) chemical species on top of the conservation equations
for mass, momentum, and energy. Beside the high dimensionality of the problem (Lu
& Law 2009), the transport equations of reacting scalars require closure models, when
the reacting structures are not fully resolved on the numerical grid.
The challenges associated with turbulent combustion modeling make the use of
machine learning very attractive. While turbulent combustion models are spread
across combustion industries, their current predictive capabilities fall well short of
what would be needed in decision making for new design and regulation (Pope 2013).
High-fidelity, direct numerical simulations (DNS) of combustion systems are still lim-
ited to isolated aspects of a turbulent combustion process and simple building blocks.

h d i 0 0 9 8 088962 4 022 bli h d li b C b id i i

 306 Data-Driven Fluid Mechanics Chapter 15

Still, these high-fidelity simulations are rich in information that could help decode the
complexity of turbulence–chemistry interactions and guide the development of filtered
and lower-fidelity modeling approaches for faster evaluations.
The objective of the present chapter is to demonstrate the potential of data-driven
modeling in the context of combustion simulations. In particular, we present:
• The application of principal component analysis (PCA) and other linear and
nonlinear techniques to identify low-dimensional manifolds in high-fidelity
combustion data sets, and to reveal the key features of complex nonequilibrium
phenomena. Different techniques are compared to PCA, including nonnegative
matrix factorization (NMF), autoencoders, and local PCA in Section 15.3.
• The development of reduced-order models (ROMs), to be used in conjunction with,
or to replace high-fidelity simulation tools, to reduce the burden associated
with the large number of species in detailed chemical mechanisms. First, the
use of transport models based on PCA is presented in Section 15.4. Finally,
the application of the data-driven adaptive-chemistry approach based on the
combination of classification and chemical mechanism reduction is discussed
in Section 15.5.

15.2 Combustion Data Sets

The structure of combustion data sets differs from the one seen in pure fluid mechanics
applications (presented in Chapter 6). The data set encountered in multicomponent
reactive flows is stored in the form of a matrix X ∈ R N ×Q . Each column of X is tied
to one of the Q thermochemical state-space variables: temperature T, pressure p, and
Ns −1 chemical species6 mass (or mole) fractions, denoted by Yi for the ith species. For
open flames and atmospheric burners the pressure variable can be omitted. Each of the
N rows of X contains observations of all Q variables at a particular point in the physi-
cal space and/or time (and sometimes, a point in the space of other independent param-
eters, as briefly discussed later). This structure of the data matrix is presented below:
 .. .. .. .. ..
. . . . .


X = T p Y1 Y2 ... YNs −1  . (15.2)
 
. .. .. .. .. 
 .. . . . . 

For such a data set, Q = Ns + 1. We denote the ith row (observation) in X as xi ∈ RQ
and the jth column (variable) in X as X j ∈ R N . When the data set is only resolved
in space and not resolved in time, N represents the number of points on a spatial grid,
and X can be thought of as a data snapshot (a notion much like the one discussed in
Chapters 6–9). Typically, we can expect N  Q. However, the magnitude of Q will
strongly depend on the number of species, Ns , involved in the chemical reactions, and
can even reach the order of thousands for more complex fuels (Lu & Law 2009).
6 Since mass (or mole) fractions sum up to unity for every observation, out of Ns species only Ns − 1 are
independent.

h d i 0 0 9 8 088962 4 022 bli h d li b C b id i i

 Zdybal et al: Advancing Reacting Flow Simulations 307

Combustion data sets can be obtained from numerical solutions of combustion

models, or from experimental measurements of laboratory flames. When generating
a numerical data set, we incorporate information about the chemistry of the process
by selecting a chemical mechanism for combustion of a given fuel. The mechanism
determines which chemical species are involved in the specified chemical reactions,
and it can vary in complexity.
For the discussion of data sets, it is useful to introduce the distinction between
premixed and non-premixed combustion. In the premixed case, fuel and oxidizer are
first mixed before they are burned. In the non-premixed combustion, fuel and oxidizer
originate from separate streams and have to first mix for the combustion to occur.
In this latter case, the mixture fraction variable is an important scalar quantity that
specifies the local fuel to oxidizer ratio and is defined as
νYF − YO2 + YO2 ,2
Z= (15.3)
νYF ,1 + YO2 ,2
for any point in space, where YF is the local mass fraction of fuel, YO2 is the local
mass fraction of oxidizer in the mixture, and ν is their stoichiometric ratio (Bilger
et al. 1990). YF ,1 is the mass fraction of fuel in the fuel stream and YO2 ,2 is the mass
fraction of oxidizer in the oxidizer stream. The complete combustion of fuel happens
at the stoichiometric mixture fraction Zst . If Z < Zst , the mixture is called lean (fuel
deficient) and if Z > Zst , the mixture is called fuel rich.
Figure 15.1 presents an overview of frequently encountered numerical data sets,
ordered schematically by the amount of information about the combustion process
they contain. In particular:
• The zero-dimensional (0D) reactor model (e.g., perfectly stirred reactor [PSR])
assumes that combustion happens in a single point in space. This model can
thus only include information about chemistry and thermodynamics of the
combustion process evolving in time. It carries no information about spatial
gradients of the thermochemical variables. The 0D reactor describes an ideal
mixing process with Da  1. When the data set is formed from the 0D reactor
simulation, the N rows of X are linked to points in time only.
• The steady laminar flamelet (SLF) model (Peters 1984) assumes that fuel and
oxidizer are two impinging streams, originating from the fuel and the oxidizer
feed respectively. Combustion happens in infinitely thin sheets (flamelets) where
fuel and oxidizer meet at varying stoichiometric proportions. The stoichiometry
is specified by the mixture fraction variable, Z, and the thermochemical variables
vary along the axis between pure oxidizer (Z = 0) and pure fuel (Z = 1). By
varying the distance between the feeds (or the outlet velocities), a strain (speci-
fied by the strain rate χ) is imposed on the flamelet, which can displace it from
its equilibrium. The SLF model describes a non-premixed system with Da  1.
This is thus a two-parameter model with each row of X linked to one pair (Z, χ).
• The counterflow diffusion flame (CFDF) model (Law 2010) is similar to the SLF
model, as it assumes two separate streams of fuel and oxidizer diffusing into one
another (non-premixed system). However, in contrast to SLF, the flow equations

h d i 0 0 9 8 088962 4 022 bli h d li b C b id i i

 308 Data-Driven Fluid Mechanics Chapter 15

Figure 15.1 Examples of common numerical combustion data sets schematically presented on
the axis of increasing complexity.

for CFDF are solved in the physical space. Flow is parameterized by the velocity
gradient parameter a, which is an equivalent of the local strain rate. In the CFDF,
each row of X is linked to a point in the physical space and time.
• The one-dimensional turbulence (ODT) model (Kerstein 1999, Echekki et al. 2011)
can additionally incorporate the effects of turbulence by introducing eddy events
on a one-dimensional domain. It allows for both spatial and temporal evolution
of the flow. ODT can be used as a standalone model but can also serve as a
subgrid model in large eddy simulations (LES).
• The Reynolds-averaged Navier–Stokes (RANS) models (Ferziger et al. 2002) solve
the time-averaged equations describing fluid dynamics and hence any sources of
unsteadiness coming from turbulence are averaged.
• The LES (Ferziger et al. 2002) resolves large scales associated with the fluid
dynamics processes but a subgrid model is required to account for processes
occurring at the smallest scales. The choice for the subgrid model becomes
particularly important in reactive flows, since combustion is inherently tied to
the smallest scales.
• The DNS (Ferziger et al. 2002) allows for the most accurate description of the
coupled interaction between fluid dynamics and the thermochemistry since all
subgrid processes are resolved directly. The scarcity of the DNS data sets is due
to the large computational cost of performing DNS simulations, especially for
more complex fuels.

More information on combustion theory and combustion models can be found in

(Turns et al. 1996, Kee et al. 2005, Law 2010).
The selected data set is the starting point for applying a data science technique,
and we often refer to those data sets as the training data. From the description of
various data sets presented above, it is visible that the choice of the data set will be
pertinent to the type and quality of the analysis performed. For instance, we can expect
that features identified will depend on the amount of information about the coupled
phenomena of turbulence and chemistry that was initially captured in the data set.

h d i 0 0 9 8 088962 4 022 bli h d li b C b id i i

 Zdybal et al: Advancing Reacting Flow Simulations 309

Table 15.1 Few selected common data scaling

Scaling technique Scaling factor d j
criteria. σj is the standard deviation, max(X j )
Auto σ is the maximum value, min(X j ) is the
√j
Pareto σj minimum value, and X̄ j is the mean
Range max(X j ) − min(X j ) observation of the jth state-space variable X j .
VAST σj2 /X̄ j

15.2.1 Data Preprocessing

Combustion data sets typically contain variables of different numerical ranges (e.g.,
both temperature, which can range from hundreds up to thousands, and mass (or mole)
fractions of chemical species, which take values in the range Yi ∈ h0, 1i). Scaling
the data set is crucial in order to balance the importance of all state-space variables
(Parente & Sutherland 2013). The data set X can be centered and scaled as follows:
e = (X − X̄)D−1,
X (15.4)

where X̄ contains the mean observations of each variable, D is the diagonal matrix of
scales, where the jth element d j from the diagonal is the scaling factor corresponding
to the jth state-space variable X j . A few of the common scaling criteria are collected
in Table 15.1. The result is a centered and scaled data matrix X.
e Other preprocessing
means can include outlier detection or data sampling. For the remainder of this chapter,
we will assume that Xe represents the matrix that has been adequately preprocessed.

15.3 Feature Extraction Using Dimensionality Reduction

Techniques

Dimensionality reduction techniques offer a way to represent high-dimensional

combustion data sets in a new, lower-dimensional basis. Such data representations
are referred to as low-dimensional manifolds. Techniques such as PCA, NMF, or
autoencoders can extract those manifolds in an empirical way from the training data
sets. This approach belongs to the family of ELDMs (Pope 2013, Yang et al. 2013),
and it is based on the idea that compositions occurring in combustion systems lie close
to a low-dimensional manifold. In addition, these techniques offer a way to extract
meaningful features by exploiting the fact that the new basis can be better suited to
represent certain physical phenomena underlying the original data. In this section,
we review a few popular dimensionality reduction techniques, and we present their
potential to detect features in combustion data sets.

15.3.1 Description of the Analyzed Data Set

In this section, we apply various dimensionality reduction techniques on a numerical
data set from a DNS simulation of a CO/H2 turbulent jet (Sutherland et al. 2007)

h d i 0 0 9 8 088962 4 022 bli h d li b C b id i i

 310 Data-Driven Fluid Mechanics Chapter 15

using the chemical mechanism consisting of 12 chemical species and 33 reactions

(Yetter et al. 1991). Additional information regarding the numerical simulation can
be found in Sutherland et al. (2007). The data set consists of a two-dimensional slice
extracted from the 3D domain arranged in a matrix X. Each row of X corresponds to
a point on the two-dimensional grid, and therefore the matrix X can be thought of as a
single-time snapshot. Several such snapshots from different times in the simulation are
available. The matrix X has 13 columns corresponding to temperature and 12 species’
mass fractions7 .

15.3.2 Extracting Features with Data Reduction Techniques

Principal Component Analysis
PCA (Jolliffe 2002) projects the preprocessed data set X
e onto a new basis represented
by the orthonormal matrix of modes A ∈ R Q×Q . The matrix A can be obtained as
the eigenvectors resulting from the eigendecomposition of the data covariance matrix
S ∈ RQ×Q ,
1 e> e
S= X X, (15.5)
N −1
or from the singular value decomposition (SVD) of the data set,
e = UΣA>,
X (15.6)
where > denotes matrix transpose. This formulation makes PCA equivalent to proper
orthogonal decomposition (POD) (see Chapter 6). The preprocessed data matrix Xe can
then be projected onto A to obtain the principal components (PCs) matrix Z ∈ R N ×Q :
Z = XA.
e (15.7)
PCs represent the original observation in the new coordinate system defined by A. It is
worth noting that the columns of Z are linear combinations of the original state-space
variables. The linear coefficients of that combination are specified in the entries of A,
and we will refer to them as weights. Adopting the general notion of the data set as
presented in (15.2), the jth PC, Z j (the jth column of Z), is computed as
Q
Õ
Zj = ei = a1, j · T
ai j X e + a2, j · pe + a3, j · Ye1 + · · · + aQ, j · YeNs −1, (15.8)
i=1

where ai j are the elements (weights) from the jth column of A and tildes represent
the preprocessing applied to each state-space variable. Since the basis matrix is
orthonormal, ai j ∈ [−1, 1] and A−1 = A> .
By keeping a reduced number of the q < Q first PCs, we obtain the closest8 rank-q
approximation of the original matrix,

7 For the purpose of the feature extraction analysis presented here, all Ns species were included in the
data set.
8 in terms of L2 , Frobenius, or trace norms, which follows from the Eckart–Young–Mirsky theorem
(Eckart & Young 1936).

h d i 0 0 9 8 088962 4 022 bli h d li b C b id i i

 Zdybal et al: Advancing Reacting Flow Simulations 311

Figure 15.2 The eigenvector weights ai j

associated with the first PC found by
PCA on a data set preprocessed with
Range scaling.

Figure 15.3 The eigenvector weights ai j

associated with the first PC found by
PCA on a data set preprocessed with
Pareto scaling.

X ≈ Xq = Zq Aq > D + X̄, (15.9)

where the index q denotes the truncation from Q to q components. Note that the
reverse operation to the one defined in (15.4) has to be applied using matrices D and X̄.
We can assign physical meaning to the PCs by looking at the linear coefficients
(weights) ai j from the basis matrix A (Parente et al. 2011, Bellemans et al. 2018).
High absolute weight for a particular variable means that this variable is identified
by the PCA as important in the linear combination from (15.8). Moreover, since the
vector A j identifies the same span as −A j , only the relative sign of a particular weight
with respect to the signs of other weights is important. In addition, PCs are ordered
so that each PC captures more variance than the following one. Thus, we can expect
the most important features identified by PCA to be visible in the first few PCs. This
property of PCA can also guide the choice for the value q. Since PCs are decorrelated,
increasing q in the PCA reconstruction from (15.9) guarantees an improvement in the
reconstruction errors.
Preprocessing the data set, prior to applying a dimensionality reduction technique,
can have a significant impact on the shape of the low-dimensional manifold and on the
types of features retrieved from the data set (Parente & Sutherland 2013, Peerenboom
et al. 2015). Figures 15.2 and 15.3 present the weights ai j associated with the first PC
resulting from Range and Pareto scaling (see Table 15.1) of the original data set. First,
we observe that the structure of the PC can change significantly with the choice of
the scaling technique. If we further consider the mixture fraction variable as defined
in (15.3) as a linear combination of fuel YF and oxidizer YO2 mass fractions, it can
be observed that the coefficients in front of YF and YO2 are of opposite signs in that

h d i 0 0 9 8 088962 4 022 bli h d li b C b id i i

 312 Data-Driven Fluid Mechanics Chapter 15

Figure 15.4 Mean, minimum (worst

reconstruction), and maximum (best
reconstruction) R2 values for
reconstructing three snapshots separated
by ∆t = 5 ms (blue triangles) using
modes from the snapshot at time
t = 2.0 · 10−3 s (red circles). Shown for
increasing q in the PCA reconstruction.

Figure 15.5 Mean, minimum (best

reconstruction), and maximum (worst
reconstruction) NRMSE values for
reconstructing three snapshots separated
by ∆t = 5 ms (blue triangles) using
modes from the snapshot at time
t = 2.0 · 10−3 s (red circles). Shown for
increasing q in the PCA reconstruction.

definition. Using Range scaling (Figure 15.2), the first PC can be attributed to the
mixture fraction variable, where the only high weights are for the fuel (H2 and CO) and
oxidizer components (O2 and N2 ), and the two have opposite signs. The correlation
between the mixture fraction variable and the first PC is in this case 99.96%. It is
worth noting that the mixture fraction was not among the variables in the original
data set X and PCA identifies it automatically. With Pareto scaling (Figure 15.3), the
first PC is almost entirely aligned with the temperature variable and carries almost no
information about the mass fractions of the chemical species.
In a previous study (Biglari & Sutherland 2012), the authors have demonstrated that
PCA can identify PCs that are independent of the filter width on a fully resolved jet
flame. PCA was performed on the state-space variables filtered using a top-hat filter
of varying widths. To test the capability of PCA to extract time-invariant features of
the data set, we can also use a fixed set of modes to reconstruct new, unseen data,
such as data from future time steps of the same temporally evolving system. If the
reconstruction process leads to errors that are comparable with the ones obtained
for the training data, we can expect that PCA captures the essence of the physical
processes underlying the system. We demonstrate this using 2D slices from the DNS
data set from four time intervals separated by ∆t = 5 ms. In Figures 15.4 and 15.5,
three future data snapshots (blue triangles) are reconstructed using q first eigenvectors
found on the initial snapshot (red circles). Figure 15.4 shows the coefficients of
determination, which can be computed for the jth variable X j in the data set as

h d i 0 0 9 8 088962 4 022 bli h d li b C b id i i

 Zdybal et al: Advancing Reacting Flow Simulations 313

(xi j − x̂i j )2
ÍN
R2j = 1 − Íi=1 , (15.10)
2
i=1 (xi j − x̄ j )
N

where xi j is the ith observation of the jth variable, x̂i j is the PCA reconstruction of that
observation, and x̄ j is the average observation of X j . The coefficient of determination
R2 measures the goodness of the model fit with respect to fitting the data with the
mean value x̄ j . Values R2 ∈ (−∞, 1], where R2 = 1 means a perfect fit. The smaller
the R2 value gets, the worse the model fit. Figure 15.5 shows the normalized root-
mean-squared errors (NRMSE) on a logarithmic scale. NRMSE can be computed for
the jth variable X j in the data set as
s
2
i=1 (xi j − x̂i, j )
ÍN
1
NRMSE j = · . (15.11)
x̄ j N

In Figures 15.4 and 15.5, the markers represent the mean values of R2 or NRMSE
averaged over all variables in the data set. The bars range from the minimum and
the maximum value achieved for any variable in the data set. It can be observed
that for all the reconstructed snapshots, the error metrics show comparable values.
This indicates the capability of PCA to capture generalized, time-invariant features of
temporally evolving systems. It can also be observed that the mean errors grow for
future snapshots, which indicates that there might be a limit on the time separation ∆t
for which we can extend the applicability of the features found.

Local Principal Component Analysis

PCA can also be applied locally (LPCA), on portions of the original data set. In LPCA,
a clustering algorithm is first applied to the data set to partition observations into local
clusters. Next, PCA is performed separately in each cluster. LPCA can not only allow
for further reduction in dimensionality by adjusting to the potential nonlinearities of
the data set, but it can also aid in detecting local features of the data. Local PCA
can also guide the decision on how to partition the data into clusters, since it grants
control of the data reconstruction errors. Figure 15.6 shows the difference between the
global and the local approach on a synthetic data set that is visibly composed of two
distinct clusters. In the global case, PCA is performed on the entire data set, ignoring
the apparent clusters. When the data set is partitioned and PCA is performed locally in
each cluster, a different set of eigenvectors and PCs is tied to each identified cluster of
data. Those local PCs can better represent the physics or the underlying features of a
particular cluster by capturing the local variance instead of the global one. In LPCA,
the ith observation xi is reconstructed from the local, q-dimensional manifold in an
analogous way to (15.9):

xi ≈ xq,i = zq,i (A(k)

q ) D+c ,
> (k)
(15.12)

where k is the index of the cluster to which the ith observation belongs, zq,i is the
ith observation represented in the local, truncated basis A(k)
q identified on the kth

h d i 0 0 9 8 088962 4 022 bli h d li b C b id i i

 314 Data-Driven Fluid Mechanics Chapter 15

Globally applied PCA Locally applied PCA

Figure 15.6 Schematic distinction between global and local PCA on a synthetic
two-dimensional data set. The arrows represent the two global/local modes from the matrix A,
defining the directions of the largest variance in the global/local data.

Initialization Iterating

Figure 15.7 Diagram presenting schematically the VQPCA clustering algorithm.

cluster. Each cluster is centered separately using the centroid c(k) of the kth cluster
and typically the global diagonal matrix of scales D is applied in each cluster.
Data clustering, prior to applying local PCA, can be performed with any algorithm
of choice. One of the techniques discussed in this chapter is the vector quantization
PCA (VQPCA) algorithm (Kambhatla & Leen 1997, Parente et al. 2009), presented
schematically in Figure 15.7. This is an iterative algorithm in which the observations
are assigned to the cluster for which the local PCA reconstruction error of that
observation is the smallest. The hyperparameters of the algorithm include the number
of clusters k to partition the data set, the number of PCs q used to approximate the
data set at each iteration, and the initial cluster partitioning. The latter is predefined
by setting the initial cluster centroids. The most straightforward way is to initialize
centroids randomly, but another viable option is to use partitioning resulting from a
different clustering technique such as K-means (MacQueen et al. 1967). The VQPCA
algorithm iterates until convergence of the centroids and of the reconstruction error is

h d i 0 0 9 8 088962 4 022 bli h d li b C b id i i

 Zdybal et al: Advancing Reacting Flow Simulations 315

Figure 15.8 Conditioning variable

partitioning into k = 2 clusters in the
mixture fraction-temperature space.

Figure 15.9 VQPCA partitioning into

k = 4 clusters in the mixture
fraction-temperature space.

reached. The reconstruction error is measured between the centered and scaled data
set X
e and the approximation X ei,q using the q first PCs computed from the eigenvectors
found on the ith cluster. More details on the VQPCA algorithm can be found in Parente
et al. (2009) and Parente et al. (2011).
Another possible way of clustering the combustion data sets is to use a conditioning
variable, such as a mixture fraction, and partition the observations based on bins of that
variable. If the mixture fraction is used, observations are first split into fuel-lean and
fuel-rich parts at the stoichiometric mixture fraction Zst . If more than two clusters are
requested, lean and rich sides can then be further divided. The approach of performing
local PCA on clusters identified through binning the mixture fraction vector is referred
to as FPCA.
Local PCA was investigated on the benchmark DNS data set (Section 15.3.1) using
two clustering algorithms. Figures 15.8 and 15.9 show a comparison between two
clustering results in the space of temperature and the mixture fraction variable. Figure
15.8 presents clustering into k = 2 clusters using bins of mixture fraction as the
conditioning variable. This partitioning can be thought of as hardcoded in a sense that
the split into two clusters will always be performed at Zst . The features retrieved on
local portions of data can thus only be attributed to the lean and rich zones. In contrast,
Figure 15.9 presents clustering into k = 4 clusters using the VQPCA algorithm.
VQPCA could distinguish between the oxidizer (cluster k1 ), the fuel (cluster k3 ), and
the region where the two meet close to stoichiometric conditions (clusters k 2 and k4 ).

h d i 0 0 9 8 088962 4 022 bli h d li b C b id i i

 316 Data-Driven Fluid Mechanics Chapter 15

Figure 15.10 Temperature profile of the DNS data set (left) and the result of partitioning the
data set into k = 4 clusters using the VQPCA algorithm (right).

This is even more apparent if we plot the result of the VQPCA clustering on a spatial
grid in Figure 15.10. The space is clearly divided into the inner fuel jet (k3 ), the outer
oxidizer layer (k1 ), and the two thin reactive layers (k2 , k4 ) for which the temperature
is the highest.
The success of local PCA in extracting features depends on the clustering technique
used. In a previous study (D’Alessio et al. 2020), VQPCA has been compared to other
clustering algorithms, and better results in terms of clustering quality and algorithm
speed have been obtained. An unsupervised clustering algorithm based on the VQPCA
partitioning has recently been proposed (D’Alessio et al. 2020b) to perform data
mining on a high-dimensional DNS data set. If an algorithm such as VQPCA is used,
the types of features found depend on data preprocessing and can additionally depend
on the hyperparameters. By changing the number of clusters or the number of PCs in
the approximation, the user can potentially retrieve different features. This has been
also investigated in a previous study (D’Alessio et al. 2020a) for a more complex,
high-dimensional DNS data set. This is in contrast with clustering strategies based
on binning a single physical quantity, such as mixture fraction or heat release rate.
For instance, taking the heat release rate as an example, we might anticipate that the
cluster formed from the high heat release rate bin will identify the chemically reacting
region.

Nonnegative Matrix Factorization

NMF (Paatero & Tapper 1994) is an algorithm for factorizing data matrices whose
elements are nonnegative. This technique can thus be applied to the data set as defined
in (15.2) since the thermochemical state-space variables are nonnegative, provided that
e = XD−1 ). This is to
at the preprocessing step the data set is scaled but not centered (X
ensure that we are not introducing negative elements in the matrix X e through centering
by mean values9 as is done in (15.4). Given the nonnegative data set X ∈ R N ×Q , the
scaling matrix D and the value for q, NMF aims to find two nonnegative matrices
W ∈ R N ×q and F ∈ Rq×Q such that

9 Alternatively, one can also subtract minimum values from each variable, thus making the range of each
state-space variable start at 0.

h d i 0 0 9 8 088962 4 022 bli h d li b C b id i i

 Zdybal et al: Advancing Reacting Flow Simulations 317

Figure 15.11 First nonnegative factor f1 .

NMF performed on Range-scaled data
with q = 2.

Figure 15.12 Second nonnegative factor

f2 . NMF performed on Range-scaled data
with q = 2.

X ≈ Xq = WFD. (15.13)

The matrix of nonnegative factors F can be regarded as the one containing a basis
(analogous to the matrix A found by PCA). The matrix W represents the compressed
data, namely the NMF scores and is thus analogous to the PCs matrix Z. The
factorization to W and F is not unique, and various optimization algorithms exist
(Berry et al. 2007, Lin 2007). In this chapter, we use the MATLAB R
routine nnmf,
which minimizes the root-mean-squared residual (Berry et al. 2007),

1
D= √ ||X − WF||F , (15.14)
N ·Q

starting with random initial values for W and F, where the subscript F denotes the
Frobenius norm. This optimization might reach local minimum and thus repeating the
algorithm can yield different factorizations.
Similarly to what was done in PCA, we can look at the nonnegative factor weights
(the elements of F) to assign physical meaning to the factors. With the nonnegative
constraint on F, only nonnegative weights can be found. Figures 15.11 and 15.12 show
the first two nonnegative factors that together represent the mixture fraction variable
(compare with Figure 15.2). The oxidizer components are included in the first factor
and the fuel components in the second. If NMF is compared to PCA, the latter can
be thought of as more robust since NMF required two modes to capture the same
information (the mixture fraction variable) that was included in a single PCA mode.

h d i 0 0 9 8 088962 4 022 bli h d li b C b id i i

 318 Data-Driven Fluid Mechanics Chapter 15

Figure 15.13 Two-dimensional manifold

obtained using an Autoencoder with five
hidden layers and SELU activation
function. Colored by the result of
VQPCA partitioning into k = 4 clusters.

Autoencoders
The autoencoder (Goodfellow et al. 2016, Wang et al. 2016) is a type of an
unsupervised artificial neural network (ANN) whose aim is to learn the q-dimensional
representation (embedding) of the Q-dimensional data set such that the reconstruction
error between the input and the output layer is minimized. The standard form of an
autoencoder is the feedforward neural network having an input layer and an output
layer with the same number of neurons, and one or more hidden layers. Given one
hidden layer, the encoding process takes as an input the preprocessed data matrix
e ∈ R N ×Q and maps it to H ∈ R N ×q , with q < Q:
X

H = f (XG
e + B), (15.15)

where the columns of H are referred to as the codes, f is the activation function such
as sigmoid, rectified linear unit (ReLU), or squared exponential linear unit (SELU),
G ∈ RQ×q is the matrix of weights and B ∈ R N ×q is the matrix of biases. At the
decoding stage, H is mapped to the reconstruction Xeq ,

eq = f 0(HG0 + B0),
e≈X
X (15.16)

where f 0, G0, and B0 may be unrelated to f , G, and B. The encoding/decoding process

can thus be summarized as
encoding decoding
X eq ∈ R N ×Q .
e ∈ R N ×Q −−−−−−→ H ∈ R N ×q −−−−−−→ X (15.17)

In this section, we use an autoencoder with five hidden layers and SELU activation
function and generate a two-dimensional embedding (q = 2) of the original data
set. Figure 15.13 shows the two-dimensional manifold obtained after the autoencoder
compression (represented by the matrix H). The manifold is colored by the previously
obtained result of partitioning via the VQPCA algorithm. From Figure 15.13, it can be
observed that the result of VQPCA partitioning still uniformly divides the autoencoder
manifold. Clusters k 1 and k 3 , representing the oxidizer and fuel respectively, are
located at the opposing ends of the manifold. Thus, it is possible to think of that
manifold as describing the progress of the combustion process, with the fuel and
oxidizer meeting in the center (k2 and k4 ) where they finally react.

h d i 0 0 9 8 088962 4 022 bli h d li b C b id i i

 Zdybal et al: Advancing Reacting Flow Simulations 319

Figure 15.14 The two-dimensional

manifold obtained from local PCA using
VQPCA clustering algorithm and
increasing the number of clusters,
overlayed with the Autoencoder manifold
using the Procrustes analysis. VQPCA
performed with k = 4 clusters.

Figure 15.15 The two-dimensional

manifold obtained from local PCA using
VQPCA clustering algorithm and
increasing the number of clusters,
overlayed with the Autoencoder manifold
using the Procrustes analysis. VQPCA
performed with k = 128 clusters.

Linear Operations That Aid in Feature Interpretation

Several linear operations can aid the interpretation of PCs or low-dimensional
manifolds. One such technique is the rotation of modes/factors (Abdi 2003, Bellemans
et al. 2018) with the varimax orthogonal rotation (Kaiser 1958) used most commonly.
The varimax rotation first selects which of the n factors should be rotated together.
It then maximizes the sum of variances of the squared weights within each of the n
rotated factors

(wi2j − w̄i2j )2,

Õ
V= (15.18)

where wi j is the ith weight on the jth factor and w̄i2j is the mean of the squared weights
(Abdi 2003). After rotation, those n factors explain the same total amount of variance
as they did together before the rotation, but variance is now redistributed differently
among the selected factors. Varimax-rotated factors typically have high weights on
fewer variables than the original factors that can aid in their physical interpretation.
The rotated factors can then be used as the new basis to represent the original data
set. Varimax and several other rotation methods are available within the MATLAB R

routine rotatefactors.
Another interesting technique is the Procrustes analysis (Seber 2009), which
is a series of linear operations that allow translation, rotation, or scaling of the
low-dimensional manifold. This can be particularly useful when manifolds obtained
from two different dimensionality reduction techniques should be compared.

h d i 0 0 9 8 088962 4 022 bli h d li b C b id i i

 320 Data-Driven Fluid Mechanics Chapter 15

Figures 15.14 and 15.15 present the Procrustes transformation (using the MATLAB R

routine procrustes) of the LPCA manifold onto an autoencoder manifold. With

the series of linear transformations the manifolds overlay each other, and the match
between them becomes closer to exact as the number of clusters with which VQPCA
was performed is increased. This suggests that the data reconstruction error for the
two-dimensional approximation from LPCA will approach the one from autoencoder
decoding when the number of clusters is high enough.

15.4 Transport of Principal Components

In Section 15.2, we have seen that the number of thermochemical state-space variables
Q determines the original dimensionality of the data set. This number also reflects
how many transport equations for the state-space variables should be solved in
a numerical simulation. The general transport equation for the set Φ = X> =
[T, p,Y1,Y2, . . . ,YNs −1 ]> of state-space variables is
DΦ
= −∇ · (jΦ ) + SΦ,
ρ (15.19)
Dt
where jΦ is the mass-diffusive flux of Φ relative to the mass-averaged velocity, and
SΦ is the volumetric rate of production of Φ (also referred to as the source of Φ).
Performing detailed simulations with large chemical mechanisms with significant
number of chemical species is still computationally prohibitive. Sutherland and
Parente (2009) proposed to use PCA to reduce the number of transport equations
that solve a combustion process. Instead of solving the original set of Q transport
equations, the original variables are first transformed to the new basis identified by
PCA on the training data set X. Next, the truncation from Q to q first PCs is performed.
Transport equations for the q first PCs can be formulated from (15.19) using the
truncated basis matrix Aq ,
Dz
ρ = −∇ · (jz ) + Sz, (15.20)
Dt
where z = Zq > (with Zq = (X − X̄)D−1 Aq ), jz = Aq jΦ , and Sz = Aq SΦ D−1 (also
referred to as the PC-sources). This is the discrete analogous of the Galerkin projection
methods described in Chapters 6 and 14. Note that the source terms of the PCs, Sz , are
scaled (but not centered) using the same scaling matrix D as applied on the data set X.
The challenge associated with the resolution of the PC-transport equation is related
to the PC-source terms. The latter is highly nonlinear functions (based on Arrhenius
expressions) of the state-space variables. The nonlinearity of the chemical source
terms strongly impacts the degree of reduction attainable using the projection of
the species transport equations onto the PCA basis (Isaac et al. 2014, 2015). A
solution to this problem is to use PCA to identify the most appropriate basis to
parameterize the ELDM and then, both the thermochemical state-space variables and
the PC-source terms can be nonlinearly regressed onto the new basis. This allows
us to overcome the shortcomings associated with the multilinear nature of PCA and

h d i 0 0 9 8 088962 4 022 bli h d li b C b id i i

 Zdybal et al: Advancing Reacting Flow Simulations 321

to reduce the number of components required for an accurate description of the

state-space.
The construction of an appropriate low-dimensional manifold requires training
data. This might be seen as a limitation of the approach since all system states
are required before applying model reduction. Although initial studies on PCA
models involved DNS data of turbulent combustion (Sutherland & Parente 2009,
Pope 2013), recent studies have demonstrated (Biglari & Sutherland 2015, Echekki
& Mirgolbabaei 2015, Coussement et al. 2016, Malik et al. 2018, Dalakoti et al. 2020)
that PCA-based models can be trained on simple and inexpensive systems, such as
zero-dimensional reactors and one-dimensional flames (see Section 15.2), and then
applied to model complex systems, such as the flame–vortex interaction (Coussement
et al. 2013), the flame–turbulence interaction (Isaac et al. 2015) as well as the turbulent
premixed (Coussement et al. 2016) and non-premixed flames (Malik et al. 2020).
Up-to-date, the PC-transport approach has been demonstrated for a wide range of
problems involving ideal reactors, including batch and PSR reactors.

15.4.1 Nonlinear Regression Models

The thermochemical state-space variables Φ and the PC-source terms SZq can be
mapped to the q-dimensional PC space using a nonlinear regression function F , such
that
φ ≈ F (Zq ),
where φ represents any of the dependent variables (T, p,Y1,Y2, . . . ,YNs −1 , or SZq ). The
models used in the literature to generate the function F include, but are not limited to
• multivariate adaptive regression splines (MARS) (Friedman 1991);
• artificial neural networks (ANNs) (Pao 2008);
• Gaussian process regression (GPR) (Rasmussen 2006).
In a previous study (Isaac et al. 2015), the authors compared different regression
models in their ability to accurately map highly nonlinear functions (such as the
chemical source terms) to the PCA manifold. In this section, we focus on the use
of GPR for state-space and source term parameterization. The choice of GPR is due to
its semi-parametric nature that increases the generality of the approach. GPR employs
Gaussian mixtures to capture information about the relation between data and input
parameters, making predictions of non-observed system states more reliable than in
fully parametric approaches.

15.4.2 Validation of the PC-Transport Approach in Large Eddy

Simulation Simulations
To demonstrate the application of the PC-transport approach, we show the results
of the LES of a multidimensional flame (Malik et al. 2020). A piloted methane–
air diffusion flame (referred to as Flame D) (Barlow & Frank 1998) for which

h d i 0 0 9 8 088962 4 022 bli h d li b C b id i i

 322 Data-Driven Fluid Mechanics Chapter 15

high-fidelity experimental data is available is selected. Flame D is fueled by a mixture

of CH4 and air (25% / 75% by volume), at Re = 22 400 and 294K, respectively. The
fuel jet is surrounded by pilot flames that stabilize the flame, which is surrounded by
a co-flow.
The development of a reduced-order model for subsequent application in LES
requires the availability of training data. The most critical aspect when generating a
training data set is to make sure that the generated state-space includes all the possible
states accessed during the actual simulation. At the same time, that data cannot
come from expensive simulations. Parametric calculations on inexpensive canonical
configurations can provide an effective solution in this context, although the nature of
the flame archetype chosen is an open subject of discussion. Considering the non-
premixed nature of Flame D, it was decided to rely on unsteady one-dimensional
laminar counter diffusion flames. The use of unsteady inlet velocity profiles allows
us to explore transient flame behavior, including extension and reignition phenomena.
The calculations were performed using the OpenSMOKE++ suite developed in
Politecnico di Milano (Cuoci et al. 2015). The GRI 3.0 (Gardiner et al. 2012)
mechanism, involving 35 species and 253 reactions (excluding NOx ), was used. The
inlet conditions were set equal to the experimental ones (Barlow & Frank 1998).
Multiple simulations were carried out by varying the strain rate, from equilibrium
to complete extinction. The unsteady solutions were saved on a uniform grid of 400
points over a 0.15 m domain. All of the unsteady data from the various simulations
was used collectively for the PCA analysis. The final data set consisted of ∼ 80, 000
observations for each of the state-space variables.
Table 15.2 shows the basis matrix weights obtained from the PCA analysis on the
five major chemical species only. It can be seen that Z1 has a large positive weight
for CH4 and a large negative value for the oxidizer (O2 and N2 ). This can be linked
to the definition of mixture fraction from (15.3). Figure 15.16 shows a plot of the first
PC highly correlated with the mixture fraction (correlation factor exceeding 99%).
Therefore, in the numerical simulation, the first PC is directly replaced by the mixture
fraction, to avoid transporting a reactive scalar. The weights for Z2 also show an
interesting pattern: a positive correlation for H2 O and CO2 , and a negative correlation
for CH4 , O2 and N2 . This can be linked to a progress variable, where products have
positive stoichiometric coefficients and reactants negative ones. Figure 15.17 shows
a plot of the second PC highly correlated with the mass fraction of CO2 , a variable
that can be attributed to the progress of reaction. It is worth pointing out that PCA
identifies these patterns without any prior assumptions or knowledge of the system of
interest.
LES simulations were then performed in OpenFOAM using the PC-transport
approach. The independent variables (PCs) are transported, and the state-space is
recovered from the nonlinear regression using GPR. The low-Mach Navier–Stokes
equations are solved on an unstructured grid, together with the PC-transport equations
from (15.20). Since the state-space was accurately regressed using q = 2 PCs, the
simulation was carried out using only Z1 and Z2 as transported PCs. More details
about the numerical setting are reported in Malik et al. (2020). As part of the validation

h d i 0 0 9 8 088962 4 022 bli h d li b C b id i i

 Zdybal et al: Advancing Reacting Flow Simulations 323

Table 15.2 The elements ai j of

Species A1 A2 A3 A4 A5
the basis matrix A identified by
H2 O −0.02 0.51 0.45 −0.73 0.02 PCA when the data set was
O2 −0.18 −0.67 −0.22 −0.60 0.30 composed of five major chemical
N2 −0.64 −0.01 −0.15 −0.09 −0.74 species only.
CH4 0.73 −0.14 −0.20 −0.27 −0.56
CO2 −0.02 0.50 −0.82 −0.14 0.21

Figure 15.16 Scatter plot of the first PC

(PC-1) versus the mixture fraction
variable. From Malik et al. (2020),
copyright 2020, with permission from
Elsevier.

of the method, we look at the temperature profiles conditioned on mixture fraction at

axial location x/D = 60 (x = 432 mm), shown in Figure 15.18. It can be observed that
the predicted temperature lies well inside the single shot experimental data points.
Figure 15.19 shows the original manifold obtained from the training data set,
while Figure 15.20 shows the manifold accessed during the simulation with major
species at t = 1 s. It can be observed that the simulation does not leave the training
manifold, which bounds all the points. It is also apparent that most of the solution
data is clustered toward the equilibrium solution, showing that the Flame D does not
experience significant extinction and reignition.
The PC–GPR model can accurately simulate a complex multidimensional flame,
using only two PCs instead of the 35 chemical species. The model can be built
using inexpensive one-dimensional simulations, as long as the training data cover the
potential state-space accessed during the actual simulation. The PCs stay bounded to

Figure 15.17 Scatter plot of the second

PC (PC-2) versus the mass fraction of
CO2 species, YCO2 . From Malik et al.
(2020), copyright 2020, with permission
from Elsevier.

h d i 0 0 9 8 088962 4 022 bli h d li b C b id i i

 324 Data-Driven Fluid Mechanics Chapter 15

Figure 15.18 Conditional average of the

temperature at x/D = 60. From Malik et
al. (2020), copyright 2020, with
permission from Elsevier.

Figure 15.19 Scatter plot of the

low-dimensional combustion manifold in
the space of the second PC (PC-2), and
the mixture fraction variable. Manifold
obtained from the training data set.
From Malik et al. (2020), copyright
2020, with permission from Elsevier.

the training manifold during the simulation, indicating that the choice of an unsteady
canonical reactor ensures to span all the potential chemical states accessed during
the simulation. The strength of the method resides in the fact that it does not require
any prior selection of variables. Instead, it automatically extracts the most relevant
variables to describe the system. From this perspective, the PC–GPR method can
be regarded as a generalization of tabulated chemistry approaches (Pope 2013),
particularly for complex systems requiring the definition of a larger number of
progress variables.

Figure 15.20 Scatter plot of the

low-dimensional combustion manifold in
the space of the second PC (PC-2), and
the mixture fraction variable. Manifold
accessed during the simulation. From
Malik et al. (2020), copyright 2020, with
permission from Elsevier.

h d i 0 0 9 8 088962 4 022 bli h d li b C b id i i

 Zdybal et al: Advancing Reacting Flow Simulations 325

15.5 Chemistry Acceleration via Adaptive-Chemistry

In Section 15.4, PCA was used to derive the reduced number of transport equations for
the new set of variables, PCs. Regression was introduced to handle the nonlinearity
of chemical source terms. In this section, we investigate the potential of local PCA
(Section 15.3.2) to classify the thermochemical state-space into locally homogeneous
regions and apply chemical mechanism reduction locally.

15.5.1 Description of the Approach

Ren and Pope (2008) proposed the operator-splitting strategy for application in
numerical algorithms used to solve the multicomponent reactive flows. With this
approach, chemistry acceleration techniques can be implemented to reduce the
computational cost related to the inclusion of detailed kinetic mechanisms. If we
consider Ψ as a vector representing the temperature and mass fractions of chemical
species, the transport equation for Ψ can be written as
dΨ
= C(Ψ, t) + D(Ψ, t) + S(Ψ), (15.21)
dt
where C(Ψ, t) and D(Ψ, t) are the convective and diffusive transport terms, respec-
tively, and S(Ψ) is the source term vector (representing the rates of change of Ψ due
to chemical reactions) (Cuoci et al. 2013). For N grid points in the simulation, we
obtain a system of N ordinary differential equations (ODEs). According to the Strang
splitting scheme (Strang 1968), (15.21) can be solved by grouping the contributions
of convection and diffusion and integrating them separately from the chemical source
term
( a
dt = S(Ψ ),
dΨ a
b (15.22)
dt = C(Ψ , t) + D(Ψ , t),
dΨ b b

where the indices a and b denote the separate contributions to Ψ. In particular, three
sub-steps are adopted for the numerical integration:

1. Reaction step: The ODE system corresponding to the source term S(Ψa ) is
integrated over ∆t
2 .
2. Transport step: The ODE system accounting for the convection and diffusion terms
C(Ψb , t) and D(Ψb , t) is integrated over ∆t.
2 .
3. Reaction step: The source term S(Ψa ) is again integrated over ∆t

Unlike the transport step, the two reaction steps do not require boundary conditions
and they do not have spatial dependence. The system of N ODEs from steps 1 and 3
can be solved independently from the system of N ODEs from step 2: this makes the
adaptive-chemistry techniques very effective and easy to implement. The idea behind
the adaptive-chemistry approach is that it is possible to locally consider only a subset
of the chemical species implemented in the detailed mechanism, while the remaining
subset consists of the species that locally have zero concentration, or result to be

h d i 0 0 9 8 088962 4 022 bli h d li b C b id i i

 326 Data-Driven Fluid Mechanics Chapter 15

not chemically active. It is thus possible to build a library of reduced mechanisms

in the preprocessing step, covering the composition space, which is expected to
be visited during simulation of the flame of interest (Schwer et al. 2003, Banerjee
& Ierapetritou 2006). To build the library of reduced mechanisms, the state-space
must be first partitioned into a prescribed number of clusters, and a reduced kinetic
mechanism is then created separately in each cluster. The high-dimensional space
can be partitioned via the VQPCA algorithm (Section 15.3.2), and for each cluster
a reduced kinetic mechanism is generated via a directed relation graph with error
propagation (DRGEP) (Pepiot-Desjardins and Pitsch 2008). At each time step of the
CFD simulation, the whole set of species in the detailed mechanism is transported.
Before the chemical step, the grid points are classified on-the-fly using the local PCA
reconstruction error metrics (Kambhatla & Leen 1997), and for each of them the
appropriate reduced mechanism is selected from the library. Although a direct relation
between the CPU time and the number of species implemented in the mechanism
is found for both the transport and the reaction sub-steps described in the (15.22),
alleviating the costs related to the reaction step appears to be more important. The
reaction step, in fact, results to be the most consuming part of the computation
requiring up to 80–85% of the CPU time (Cuoci et al. 2013). The full details of the
overall procedure, called SPARC, as well as its validation for steady and unsteady
laminar flames, can be found in D’Alessio et al. (2020).

15.5.2 Application of the Approach

The application of the adaptive-chemistry approach is briefly presented here for the
simulation of an axisymmetric, non-premixed laminar nitrogen-diluted methane co-
flow flame (Mohammed et al. 1998). The detailed simulation was first carried out
using the POLIMI C1C3 HT 1412 kinetic mechanism (84 species and 1698 reactions)
(Ranzi et al. 2014), and the adaptive simulations were then performed using four
degrees of chemical reduction, ranging from DRGE P = 0.03 to DRGE P = 0.005.
The average and the maximum number of species included in the reduced mechanisms
depending on the reduction tolerance DRGE P are reported in Table 15.3. An excellent
agreement was observed for all the cases, although for DRGE P = 0.03 and 0.02 a
negligible discrepancy was observed for the lift-off height. The inverse proportionality
between the simulation accuracy and the degree of reduction, that is, the tolerance
used for DRGEP, is due to a different number of species included in the reduced
mechanisms. Figure 15.21 shows a comparison between the temperature and species
profiles obtained from the detailed and the adaptive simulation carried out with
DRGE P = 0.005.
The inclusion of a higher number of chemical species in the reduced mechanisms
obviously entails an increase of the mean CPU time required to carry out the
chemical step (τ̄chem ). Nevertheless, the adaptive simulation speedup with respect to
the detailed simulation (Schem ) is large (Schem ∼ 4) even when the lowest degree of
chemical reduction is adopted, as reported in Table 15.4.

h d i 0 0 9 8 088962 4 022 bli h d li b C b id i i

 Zdybal et al: Advancing Reacting Flow Simulations 327

Figure 15.21 Temperature, O2 , OH and CO profiles obtained from the detailed and the
adaptive-chemistry simulation of an axisymmetric, non-premixed laminar nitrogen-diluted
methane coflow flame.

Table 15.3 The mean and maximum number of chemical

DRGE P nsp
mean nsp
max
species versus the uniformity coefficients for prescribed
0.03 31 38 tolerances  in DRGEP. The composition space was
0.02 34 42 partitioned via the VQPCA algorithm.
0.01 39 44
0.005 43 50

If multidimensional simulations of the same system are not available, the SPARC
model can be trained using lower-dimensional (0D or 1D) detailed simulations of
the same chemical system. Reduced mechanisms can then be generated based on such
training data sets and applied in the multidimensional (2D or 3D) adaptive simulations.
D’Alessio, Parente, Stagni and Cuoci (2020) also considered a data set consisting

Table 15.4 Performances of adaptive-chemistry algorithm:

DRGE P τ̄chem Schem
average CPU-time per cell (in ms) for chemical step
0.03 2.78 5.39 integration (τ̄chem ) and relative mean speed-up factor
0.02 3.03 4.94 (Schem ) for the steady-state laminar methane flame with
0.01 3.25 4.61 respect to the degree of chemistry reduction (DRGE P ).
0.005 3.78 3.97
detailed 15.02 -

h d i 0 0 9 8 088962 4 022 bli h d li b C b id i i

 328 Data-Driven Fluid Mechanics Chapter 15

of observations generated through steady-state CFDF simulations (see Section 15.2),

evaluated at different strain rates from 10s−1 to 330s−1 (the latter corresponding to
extinction conditions). This new data set consisted of ∼ 11 000 points, corresponding
to 30 different strain rates randomly chosen in the considered interval. Also in this
case, the accuracy using mechanisms obtained from a lower dimensional training data
set was high, with the speedup with respect to the detailed simulation between 4.5
and 6. More details can be found in D’Alessio, Parente, Stagni and Cuoci (2020).

15.6 Available Software

Apart from the many available MATLAB R

routines, some of which were mentioned
in this chapter, two recently developed Python libraries can be used to perform
dimensionality reduction using PCA and other techniques:

• OpenMORe (D’Alessio et al. 2020) is a collection of Python modules for reduced-

order modeling (ROM), clustering, and classification. It incorporates many ROM
techniques such as PCA, local PCA, kernel PCA, and NMF along with the
varimax rotation for factor analysis. It can also be used for data clustering with
VQPCA or FPCA algorithms, and it introduces utilities for evaluating the clus-
tering solutions. The software, along with detailed documentation and several
examples, is available at https://github.com/burn-research/OpenMORe.
• PCAfold (Zdybał et al. 2020) is a Python library that can be used to generate,
improve, and analyze low-dimensional manifolds obtained via PCA. Several
novel functionalities to perform PCA on sampled data sets and analyze
the improved representation of the thermochemical state-space variables
(Zdybał et al. 2021) were introduced. The effect of the training data
preprocessing on the topology of the low-dimensional manifolds can be
investigated with the novel approach to assess the quality of manifolds
(Armstrong & Sutherland 2021, Zdybal et al. 2022a, Zdybal et al. 2022b).
PCAfold also accommodates for treatment of the thermochemical source
terms for use in PC-transport approaches (Section 15.4). The user can find
numerous illustrative examples and the associated Jupyter notebooks in the
software documentation (available at https://pcafold.readthedocs.io/)
under Tutorials & Demos. The software is available at https://gitlab
.multiscale.utah.edu/common/PCAfold.

15.7 Summary

In this chapter, several examples of the application of data-driven techniques to multi-

component reactive flow systems have been shown. Data reduction techniques, such
as PCA, NMF, or autoencoders also offer a way to explore hidden features of data sets
(Section 15.3). These methods are capable of detecting information in an unsupervised

h d i 0 0 9 8 088962 4 022 bli h d li b C b id i i

 Zdybal et al: Advancing Reacting Flow Simulations 329

way and can thus aid in the process of selecting the best set of variables to effectively
parameterize complex systems using fewer dimensions. Moreover, it has been shown
how the aforementioned algorithms can be effectively coupled with algorithms that
are widely used in the combustion community, such as DRGEP, to obtain physics-
informed reduced-order models. Two applications of reduced-order modeling were
presented in this chapter: reduction of the number of transport equations (Section 15.4)
and reduction of large chemical mechanisms (Section 15.5). The main power of data-
driven techniques is that the modeling can be informed by applying the technique on
simple systems that are computationally cheap to obtain. Further improvement can be
achieved based on the feedback coming from the validation experiments.

h d i 0 0 9 8 088962 4 022 bli h d li b C b id i i

 16 Reduced-Order Modeling for
Aerodynamic Applications and
Multidisciplinary Design
Optimization
1
S. Görtz , P. Bekemeyer, M. Abu-Zurayk, T. Franz, and M. Ripepi

Reduced-order models (ROMs) for both steady and unsteady aerodynamic applica-
tions are presented. The focus is on compressible, turbulent flows with shocks. We
consider ROMs combining proper orthogonal decomposition (POD), Isomap, which
is a nonlinear manifold learning method, and autoencoder networks with interpolation
methods. Physics-based ROMs, where an approximate solution is found in the POD-
subspace by minimizing the corresponding steady or unsteady flow-solver residual,
are also being discussed. The ROMs are used to predict unsteady gust loads for rigid
aircraft as well as static aeroelastic loads in the context of multidisciplinary design
optimization (MDO) where the structural model is to be sized for the (aerodynamic)
loads. They are also used in a process where an a priori identification of the critical
load cases is of interest and the sheer number of load cases to be considered does not
allow for using high-fidelity numerical simulations. The different ROM methods are
applied to 2D and 3D test cases at transonic flow conditions where shock waves occur
and in particular to a commercial full aircraft configuration.

16.1 Introduction and Motivation

The multidisciplinary design of a civil transport aircraft is a highly iterative optimiza-

tion process, each design cycle requiring a large volume of computations to analyze
the current performance, handling qualities and loads. For conventional aircraft, a
load envelope may require on the order of 100 000 simulations to find all critical
load cases, comprising gust encounters, maneneuvers, and failure scenarios at various,
mostly off-design Mach numbers, flight levels, and trim states within the aircraft flight
envelope. For unconventional aircraft, where little or no engineering experience is
available, up to 10 million computations may be required. The use of high-fidelity

1 The author would like to thank Airbus for providing the XRF-1 testcase as a mechanism for
demonstration of the approaches presented in this chapter.

h d i 0 0 9 8 088962 4 023 bli h d li b C b id i i

 Görtz et al: ROMs for Aerodynamic Applications and MDO 331

computational fluid dynamics (CFD) in this context is at the horizon, but still too
costly and time-consuming to provide all the required aerodynamic data, that is, steady
and unsteady pressure and shear stress distributions on the aircraft surface, at any point
within this envelope. This motivates procedures and techniques aimed at reducing the
computational cost and complexity of high-fidelity simulations to provide accurate but
fast computations of, for example, the aerodynamic loads and aircraft performance.
A classical approach to reduce the numerical complexity is to simplify the physics.
An example of this is the common use of linear potential flow equations during
loads analysis. However, such physical model simplifications have the disadvantage
of neglecting significant effects such as transonic flow, stall, and friction drag in the
case of aerodynamics. This may be acceptable early on in the design process, while
more detailed analysis may be applied at a later stage when the design space has been
narrowed down sufficiently.
As an alternative to simplifying the physical model, reduced-order modeling
(ROM) provides another approach to reduce numerical complexity and computational
cost while providing answers accurate enough to support design decisions and to
perform quick trade studies. In general, the various ROM methods realize such a
goal by identifying a low-dimensional subspace or manifold based on an ensemble
of high-fidelity solutions “snapshot,” which sample a certain parametric domain of
interest. The number of degrees of freedom (DoF) is then reduced while retaining
the problem’s physical fidelity, thus allowing predictions of the required aerodynamic
data with lower evaluation time and storage than the original CFD model. ROMs
are anticipated to enable incorporating high-fidelity-based aerodynamic data earlier
into the design process. Methods for minimizing the number of expensive high-
fidelity simulations needed to extract a reduced-order model from simulation data
are sought after, including adaptive sampling strategies. Data classification methods
are of interest to gain more physical insight, for example, to identify (aerodynamic)
nonlinearities and to track how they evolve over the design space and flight envelope.
Finally, data-driven methods, including machine learning methods stemming from
other fields of research, that are able to extract relevant information from high-fidelity
numerical and experimental data and for merging heterogeneous data can help to
arrive at a more consistent, homogeneous digital aircraft model. This chapter focuses
on reduced-order modeling approaches based on high-fidelity CFD in the context
of aerodynamic applications and multidisciplinary analysis and optimization. The
emphasis is on the basic ideas behind the methods, their accuracy and computational
efficiency, and their application to industrial aircraft configurations. The mathematical
formulation of the methods can be found in the corresponding references.
The basic ideas behind reduced-order modeling based in POD and Isomap are
explained briefly in the next section, followed by a description of their implemen-
tation. The last section is concerned with their application to different two- and three-
dimensional use cases.

h d i 0 0 9 8 088962 4 023 bli h d li b C b id i i

 332 Data-Driven Fluid Mechanics Chapter 16

16.2 Reduced-Order Modeling

The goal of a ROM is to emulate an expensive process like a full-order CFD

simulation. Based on a given set of parameter-dependent CFD solutions, called
snapshots, a ROM can be built and exploited to predict solutions at untried parameter
combinations. For highly nonlinear functional dependencies of the underlying process,
a proper set of snapshots is necessary to obtain a sufficient accuracy of the predictions
(Franz et al. 2014). ROMs can be divided into intrusive and nonintrusive methods.
Intrusive methods make use of the underlying equations, whereas nonintrusive
ROMs are completely decoupled from the underlying equations, but coupled with an
interpolation method, which inherits the parameter dependency of the solutions.
ROMs for aerodynamic applications operate on parametrically generated data rep-
resented by either surface quantities (e.g., surface pressure and shear stress) or volume
quantities (e.g., the primitive variables). Here, the focus is on compressible, turbulent
flows with shocks as described by the Navier–Stokes equations. The DLR TAU code
(Kroll et al. 2014) is utilized as a CFD solver, employing hybrid unstructured grids,
to obtain the aerodynamic data snapshots by solving the Reynolds-Averaged Navier–
Stokes (RANS) equations. The parameters can be related to the flow (e.g., the angle of
attack, the Mach number), the geometry (e.g., wing span, taper ratio, and sweep angle),
the structure (e.g., Young’s and shear modulus of the beam representation of the wing
box), and the flight condition (e.g., load factor, altitude). The model order reduction
techniques used are hereafter briefly described, and their application to aerodynamic
problems and Multidisciplinary Design Analysis and Optimization (MDAO) is shown
in the following.

16.2.1 Proper Orthogonal Decomposition-Based ROMs

A widely used tool for reduced-order modeling is POD, also known as principal
component analysis (PCA) and Karhunen–Loéve expansion (KLE). In fluid dynamics,
POD is applied to steady problems and unsteady problems in the time as well as
frequency domain as illustrated in Chapter 6. The POD method generates a sequence
of orthogonal basis functions, called POD modes, through modal analysis of an
ensemble of snapshot flow solutions, where every snapshot corresponds to a different
combination of parameters, for example, different combinations of Mach number and
angle of attack. These solutions to the governing equations, here the RANS equations,
form the solution space. The POD modes span an optimal linear subspace for the
corresponding solution space. By choosing a subset of modes, the method seeks to
isolate the few main structures whose linear combination represents the system in
an optimal way. Several variants of POD-based ROMs have been developed that
primarily differ in the way they connect the retained modes to the parameter space
of the snapshot distribution. The POD may be embedded in a Galerkin projection
framework; it may be combined with a CFD flux residual minimization scheme;
or it may be coupled to an interpolation method (POD+I). Galerkin projection is
an example of an intrusive method. It projects the underlying spatially discretized

h d i 0 0 9 8 088962 4 023 bli h d li b C b id i i

 Görtz et al: ROMs for Aerodynamic Applications and MDO 333

partial differential equations (PDEs) onto the POD subspace to obtain a system of
ordinary differential equations (ODEs).
POD+I is a nonintrusive method as the interpolation technique does not require
any details on the underlying governing equations. ROM-predicted solutions are
determined by directly interpolating the coefficients of the POD modes, without
the need to solve any ODE system. It generally establishes a multidimensional
relationship between the modal coefficients or amplitudes and the parameter space,
for example, by fitting a radial basis function in the modal space to the set of snapshot
points in the parameter space. This requires simple interpolation of scalar-valued POD
basis coefficients to get a surrogate model of these coefficients as a function of the
parameters. To predict a flow solution at an untried combination of parameters, the
surrogate is evaluated for these parameters, and the predicted POD coefficients are
multiplied with the corresponding precomputed, invariant POD modes. This has the
advantage of simplicity of implementation and independence of the complexity of
the system and source of the modes being processed, which allows for application
to multidisciplinary problems and the combination of different data sources such
as CFD and experimental test results. The main disadvantage of nonintrusive POD
methods stems from their reliance on interpolation techniques to accurately reproduce
the possibly very nonlinear response surfaces of the modal coefficients.
Intrusive POD-based methods do better in this respect. This is done by solving
a nonlinear least-squares (LSQ) optimization problem for the modal coefficients
minimizing the steady (or unsteady) flow-solver residual of the governing equations
in POD subspace (Zimmermann & Görtz 2010). For the semi-discrete Navier–Stokes
equations,

∂w (t)
R̂ := R (w (t)) + = 0 ∈ RN , (16.1)
∂t

with w = [ρ, ρv, ρE t , νt ] ∈ R N and N corresponding to the order of CFD model

(number of variables times number of grid nodes), we search for an approximated
solution corresponding
r
Õ
w≈ ai Ui + w = Ur a + w (16.2)
i=1

in the POD subspace Ur ∈ R N x r, r  N, minimizing a subset of the unsteady

residual in the L2 norm,
2
min k R̂∗ (Ur a + w) k L2 , (16.3)
a

with a being the vector of the unknown POD coefficients ai and w the mean of the
snapshots set. A greedy missing point estimation is used to select the subset. This
approach is an alternative to POD-based ROMs based on Galekin–projection. In the
following, such approaches will be referred to as POD+LSQ or LSQ-ROM.

h d i 0 0 9 8 088962 4 023 bli h d li b C b id i i

 334 Data-Driven Fluid Mechanics Chapter 16

16.2.2 Reduced-Order Models Based on Isomap

The linear nature of POD makes the method attractive but also is a source of its
limitations. Highly nonlinear flow phenomena, such as shocks, are often insufficiently
reproduced, because of the underlying assumption that the full-order CFD flow
solutions lie in a low-dimensional linear subspace. Although the coefficients of
the POD basis can handle nonlinearities, the nonlinearity might not be sufficiently
accounted for in the POD basis due to insufficient sampling in the solution space.
One approach to improve the fidelity of linear ROMs is to substitute the POD
with a nonlinear manifold learning, or, more generally, dimensionality reduction (DR)
technique, which assumes that full-order data lie on a nonlinear manifold of low-
dimension. The manifold can be approximated by sampling the full-order model.
IsomapTenenbaum (2000), which is a nonlinear DR method based on multidimen-
sional scaling (MDS), can be employed to extract low-dimensional structures hidden
in a given high-dimensional data set. The Isomap method only provides a mapping
from the high-dimensional input space onto a lower-dimensional embedding space
for a fixed finite set of given snapshots. For any ROM of the Navier–Stokes equations,
however, it is an essential requirement that the approximate reduced-order flow
solutions are of the same type and dimension N as the CFD snapshots. Hence, once the
set of low-dimensional vectors is obtained, a back-mapping from the reduced-order
embedding to the high-dimensional solution space is mandatory. Coupled with an
interpolation model formulated between the parameter space and the low-dimensional
space, a ROM is obtained, which is capable of predicting full-order solutions at untried
parameter combinations. This method will be referred to as Isomap+I. Furthermore,
another back-mapping from the low-dimensional space to the high-dimensional space
may be performed based on residual optimization. Its objective is to obtain a CFD-
enhanced prediction by minimizing the discretized flux residual of the interpolated
solution. This method will be referred to as Isomap+LSQ.
A schematic representation of the procedure for creating parametric ROMs for
steady flow problems based on POD and Isomap in combination with interpolation
or residual minimization is shown in Figure 16.1.

16.3 Implementation

The methods described herein are implemented in DLR’s Surrogate Modeling

for AeRo Data Toolbox (SMARTy) [Bekemeyer et al, 2022]. SMARTy is a
modular, object-oriented Python package for data-driven modeling, simulation, and
optimization. It is designed as an application programming interface (API) that
provides a set of techniques that can be used out of the box, utilizing the provided
application scripts. For more advanced scenarios, different methods can easily be
combined in “user-defined” scripts and workflows.
Key features include design of experiments methods, adaptive sampling strategies,
dimensionality reduction, and data fusion techniques (Mifsud et al. 2019), as well
as methods for surrogate modeling, variable-fidelity modeling (Han & Görtz 2012a),

h d i 0 0 9 8 088962 4 023 bli h d li b C b id i i

 Görtz et al: ROMs for Aerodynamic Applications and MDO 335

Training input 1 Training output 2 3

a) POD model
Angle of a�ack

b) ISOMAP

Mach number
Define training Compute training Perform order
points outputs with CFD reduction
4 5
a) Interpolation
b) Residual minimization
c) Combined approach

Find coefficients for Flow prediction at

model prediction untried condition

Figure 16.1 Steady ROMs in a nutshell.

and intrusive and nonintrusive reduced-order modeling (Zimmermann et al. 2014).

In terms of interpolation, thin-plate spline (TPS) interpolation, Kriging, gradient-
extended Kriging (Han et al. 2013), and hierarchical Kriging (Han & Görtz 2012b)
are available. Various kernels (or correlation functions) are implemented and differen-
tiated and can be used with these interpolation methods. Computationally demanding
tasks of the interpolation module are implemented in Cython. Moreover, it has
an interface to the FlowSimulator (Reimer et al. 2019) that enables easy access
to the DLR CFD code TAU as well as allows using SMARTy as a plugin in
complex workflows such as surrogate-based optimization. SMARTy is parallelized
in order to handle distributed CFD data on multicore architectures. SMARTy can
be used for rapidly predicting aerodynamic data based on high-fidelity CFD (Mifsud
et al. 2014), fusing numerical and experimental data (Franz & Held 2017), surrogate-
based optimization (Han et al. 2018), uncertainty quantification (Kroll et al. 2015),
and robust design (Sabater et al. 2020).

16.4 Applications

The applications demonstrate how reduced-order modeling and dimensionality reduc-

tion methods can be employed to provide a digital description of an aircraft in terms

h d i 0 0 9 8 088962 4 023 bli h d li b C b id i i

 336 Data-Driven Fluid Mechanics Chapter 16

Figure 16.2 XRF-1 generic long-range transport aircraft (left) and NASA’s Common Research
Model (CRM) (right).

of its aerodynamic characteristics and properties based on a limited set of high-fidelity

simulations. This digital description is sometimes referred to as “virtual aircraft” or
“digital aircraft” (Kroll et al. 2015).
Below, several 2D and 3D test cases are used. The 2D test cases are used to validate
the general feasibility of the methods. The 3D use cases demonstrate their applicability
to problems of interest to industry. Three 3D use cases are employed, an isolated
wing and two complete aircraft configurations. The first one is the XRF-1 long-range
transport aircraft configuration. The XRF-1 is an Airbus provided industrial standard
multidisciplinary research test case representing a typical configuration for a long-
range wide body aircraft. The XRF-1 research test case is used by Airbus to engage
with external partners on development and demonstration of relevant capabilities/tech-
nologies. The second one is the NASA Common Research Model (CRM) (Vassberg
et al. 2008). The CRM consists of a contemporary supercritical transonic wing and
a fuselage that is representative of a wide-body commercial transport aircraft. The
geometry of both aircraft configurations is shown in Figure 16.1.

16.4.1 Parametric Reduced-Order Models Based on Proper Orthogonal

Decomposition and Isomap for LANN Wing
A parametric ROM for steady CFD data was developed based on Isomap, a nonlinear
“manifold learning” method. It is assumed that the space of all CFD solutions
forms a nonlinear manifold, which in turn constitutes a sub-manifold of Rn of
lower intrinsic dimensionality. This is a more elaborate approach compared to POD
method in which a linear subspace of Rn is assumed. Isomap is comparable to
the kernel-PCA, which uses kernels derived from simulation data. ML methods
apply various approaches to identify the manifold geometry and to represent it in
a low-dimensional Euclidean space. For this purpose, Isomap uses the pairwise
geodesic distances between previously generated CFD solutions at selected parameter
combinations. After approximating these distances, Isomap computes a data set of
low-dimensional vectors in the so-called embedding space, whose pairwise Euclidean
distances correspond to the approximated geodesic distances. A mapping from the
embedding space back onto the manifold in the high-dimensional CFD solution space
was developed to make predictions of CFD solutions at any point in the parameter

h d i 0 0 9 8 088962 4 023 bli h d li b C b id i i

 Görtz et al: ROMs for Aerodynamic Applications and MDO 337

Figure 16.3 Steady Euler computation for the LANN wing compared to intrusive and
nonintrusive ROM predictions based on Isomap and POD, M = 0.81, α = 2.6◦ .

space. When combined with an interpolation model between the parameter space and
the embedding space, similar to POD with interpolation (POD+I), an Isomap-based
ROM (Franz et al. 2014), called Isomap+I, is obtained.
Parametric ROMs for the steady transonic flow around the LANN wing were gen-
erated. The LANN Wing is a well-known supercritical research wing model, built by
the Lockheed-Georgia Company for the U.S. Air Force for testing at NLR and NASA
Langley in 1983; the best reference is: LANN (Lockheed, AFWAL, NASA-Langley,
and NLR) Wing Test Program: Acquisition and Application of Unsteady Transonic
Data for Evaluation of Three-Dimensional Computational Methods (dtic.mil). Here,
the ROMs were parametric with respect to the flow conditions. The parameter space,
defined by variations of the angle of attack, α and the Mach number, Ma, was
previously sampled in the range [α × Ma] = [1◦, 5◦ ] × [0.76, 0.82] using a “design of
experiment” (DoE) based on a Latin-Hypercube sampling approach with 25 different
α-Mach-combinations. At these parameter combinations, the corresponding CFD
solutions were computed and used as input data for both ROMs. Figure 16.3 provides
a comparison of the predictions of Isomap+I and POD+I with the computed TAU
reference solution for the pressure coefficient at three wing sections for the LANN
wing in inviscid flow. The two different curves for the pressure coefficient distributions
correspond to the pressure and suction sides. The required parameters for Isomap+I
were determined automatically, yielding a two-dimensional embedding space and
seven nearest neighbors to be ideal for the embedding. Isomap+I yields better
predictions than a POD-based interpolation method, in particular for the location and

h d i 0 0 9 8 088962 4 023 bli h d li b C b id i i

 338 Data-Driven Fluid Mechanics Chapter 16

magnitude of the shock wave at transonic flow conditions. Isomap+LSQ is able to

further improve the prediction around the shock wave, where Isomap+LSQ refers to
the intrusive ROM approach based on residual minimization.

16.4.2 Parametric Reduced-Order Model Based on Proper Orthogonal

Decomposition and Isomap for XRF-1
The validation of the POD and Isomap-based ROMs was performed using the XRF-
1 wing fuselage configuration. Here, the focus was on how accurate the different
ROMs predict shocks in the transonic flow regime and on evaluating the impact of the
accuracy of the prediction on downstream analyses. In this specific case, the ROMs
were used to predict the surface pressure distribution acting on the wing at a particular
flow condition, which in turn was used to size the wing structural model of the (rigid)
XRF-1 subject to this one load case. The wing of the XRF-1 was parameterized with
five free-form deformation parameters in order to vary the twist of the wing. About
100 TAU simulations were performed for 100 differently twisted wings at a Mach
number of 0.83, a Reynolds number of 43.3 · 106 and a target lift coefficient of 0.5.
An Isomap-based adaptive sampling algorithm was used to select combinations of
parameters in the five-dimensional parameter space where to compute snapshots with
TAU (Franz et al. 2017).
The fully turbulent calculations were performed with TAU using the Spalart–
Allmaras negative turbulence model. The hybrid unstructured numerical grid had
784,384 grid points. Using the 100 snapshots, ROMs for the surface pressure dis-
tribution as a function of the geometry variations were constructed. For this purpose,
POD and Isomap were combined with an interpolation scheme to make predictions
at 10 untried parameter combinations in the five-dimensional parameter space. These
10 parameter combinations were selected at the center of the 10 simplexes with the
largest volume of the Delaunay triangulation of the parameter space, that is, where
the largest “gap” between snapshots occurs in the five-dimensional parameter space.
Predicting the pressure distribution at these points is most challenging for the ROMs,
because the prediction points are the farthest away from the snapshots.
The prediction of the surface pressure distribution at one of the 10 untried parameter
combinations, or test points, is shown in Figure 16.4. The twisting of the wing using
the free-form deformation is shown in the upper right corner of the figure. There are
slight differences in the contour distribution between the two ROM predictions, but
the shock position and strength are sufficiently well reproduced compared to the TAU
reference solution.
These aerodynamic loads were used to size the structural model of the XRF-1 wing,
which consists of 2,167 elements. Normally, several critical load cases are used to
size a structure; however, it should first be investigated how the accuracy of a ROM
prediction of the aerodynamic load distribution affects the sizing of the structure.
The result can be seen in Figure 16.5. Here only results for POD-based predictions
are compared with TAU, but they are also available for Isomap. The picture shows
the thickness distribution of the structural elements of 350 optimization regions (opt
region), where one optimization region consists of several finite elements.

h d i 0 0 9 8 088962 4 023 bli h d li b C b id i i

 Görtz et al: ROMs for Aerodynamic Applications and MDO 339
90 7. Applications

Figure 7.19. Prediction of the surface C p -distribution of the XRF-1 by

16.4 Comparison
FigureIsomap+I and POD+Iof the
atsurface pressure
an untried distribution
parameter predicted by
combination p̃3 ∈ P \ P.
Isomap+interpolation and POD+interpolation for the rigid XRF-1 with TAU reference solution
for an untried combination of geometry parameters at four different spanwise wing sections.
Reprinted from Ripepi et al. (2018) by permission from Springer Nature.

The structural sizing leads to a wing mass of 3 761 kg when sizing with TAU
reference loads, 3 776 kg when sizing with the POD-based predictions, and 3,755 kg
when sizing the wing with Isomap-based predictions. The difference in wing mass is
therefore only 0.4% and 0.1%, showing that ROM-based predictions of aerodynamic
loads are of sufficient accuracy in the context of structural sizing.

h d i 0 0 9 8 088962 4 023 bli h d li b C b id i i

 1

Thickn
340 0
Data-Driven Fluid Mechanics Chapter 16
240 250 260 270 280 290 300 310 320 330 340 350
Opt region
·10−2
2 TAU
POD+I
Thickness (m)
1

0
240 250 260 270 280 290 300 310 320 330 340 350
Opt region

Figure 7.24. Thicknesses at the wing spar optimization regions for the
16.5 Comparison
predictions
Figure of the (top)
of Isomap+I resultsand
optimizing
POD+Ithe differentatregions
(bottom) of the wing structure
p3 . Emphasized are of
the rigid XRF-1 with
the regions withthe aerodynamic
largest gap. loads predicted by the POD-based ROM (red bars) and
with the reference loads computed with the CFD code TAU (blue bars).

there are three wing spars, the thickness distribution plotted against the optimization
16.4.3 region does not showModeling
Reduced–Order an decreasingfor behaviour as usual, butProblems
Static Aeroelastic for each separate spar it
does. However, close to the optimization region 320, there is an outlier which may
Steady
be due toROMs were
the fact thatused to predict
the inner spar, the
withstatic aeroelasticoptimization
corresponding loads for structural
regions sizing
288
to 303, ends and the loads are distributed to the two remaining spars. 2018).
within a multidisciplinary design and optimization context (Ripepi et al. They
As it can
were also used in a process where an a priori identification
be seen, Isomap and POD-based ROMs provide good predictions of the thickness of the critical load cases
is of interest at
distributions andthe
thewing
sheerspar
number of load cases
optimization to beParticularly,
regions. consideredthe cannot be computed
detailed views
with high-fidelity
emphasize CFD.
that there is almost no mismatch between the computed thicknesses of
the Within
predictedthesolutions
framework andofthe
MDO, the CFD-code
computed thicknessesTAU of and a computational
the corresponding structural
reference
solutions.
mechanicsThus, (CSM)thesesolver
ROMsare are used
suitable
to in the context
perform of MDO
coupled and should lead
fluid–structure to a
interaction
speed-up of the optimization process as mentioned above.
(CFD–CSM) simulations to compute aerodynamic loads acting on the wing, which
are in turn used to design or optimize the structure (structural sizing). Since TAU
is repeatedly called for different structural models and different wing geometries to
evaluate the load cases during the optimization, the flow solver should be replaced by
a steady ROM for the elastic aircraft. Replacing the flow solver with a ROM has the
advantage that the snapshots necessary to build the ROM can be calculated offline, that
is, before the actual optimization begins. Once the snapshots have been computed, the
ROM itself can also be generated offline. During the optimization, the much-faster
ROM replaces the flow solver. This will help to speed up the overall design and
optimization process and allow more load cases to be considered in the sizing of the
structure. To set up the ROMs, snapshots (TAU solutions) for different aerodynamic
shapes and structural modes are required. Here, a ROM for the static aeroelastic load
prediction of the XRF-1 wing-fuselage configuration was created, that is, the shape
was not varied.
For this purpose, 21 coupled computations with a structured mesh suitable for
RANS simulations with one million mesh points (Figure 16.6, left) and an ANSYS
finite element model with 2,167 elements (Figure 16.6, right) for different altitudes
and Mach numbers at a load factor of 2.5g were performed to create snapshots for the
wing deformation and the aerodynamic load in terms of the pressure distribution. The
21 points in the parameter space are shown in Figure 16.7 as blue symbols.

h d i 0 0 9 8 088962 4 023 bli h d li b C b id i i

 However, the integrated nature of the MDO process
involves complex interactions between the different disci-
plines, which are difficult to be represented with a single
global ROM, if not at the expense of a costly sampling of the
whole paramater design space with multidisciplinary high-
Görtz et al: ROMs
fidelity simulations. Taking notefor that Aerodynamic
such a global ROM Applications and MDO
Z 341
may be devised and sought in future works and projects,
namic applications, loads and MDO hereafter an efficient approach to manage this complexity is X
Y
shown. This approach exploits a divide-and-conquer strat-
egy and the MDO is decomposed in sub-processes, for which
geometrical parameters (wing
small parametric ROMs can be easily generated separately
pect ratio, taper ratio, swept
and used for fast system-level analysis.
or selected airfoil sections) and Fig. 4 Detailed view of the surface of the CFD structured grid of the
First, Sect. 4.1 describes how to construct a ROM of generic long-range transport aircraft
box stiffness and mass). The
coupled, static aeroelastic solutions for a given (flexible)
d using Design of Experiment
aero-structural configuration and different load cases (in
Fig. 3, ROM of the Aeroelastic Model block, w/o Sizing
nature of the MDO process Process, and parameters the flight conditions). Then Sect.
ns between the different disci- 4.2 describes how to select critical load cases based on an
o be represented with a single aeroelastic ROM built for a given aerodynamic shape, but
ense of a costly sampling of the structure sized over different load cases (in Fig. 3, ROM of
ce with multidisciplinary high- the Aeroelastic Model block,Z with Sizing Process, and
note that such a global ROM parameters the load cases). Section 4.3 shows a ROM for
in future works and projects, steady aerodynamic loads predictions of a rigid model of
ch to manage this complexity is the generic transport aircraft X
Y
configuration subject to
its a divide-and-conquer strat- geometry variations (in Fig. 3, ROM of the Aerodynamic
osed in sub-processes, for which Model block and parameters the Geometry), which is
be easily generated separately afterwards used (Sect. 4.4), for a given aerodynamic shape
el analysis. and
Fig. 4load case,view
Detailed in the structural
of the sizing
surface of sub-process.
the CFD structured grid of the Fig. 5 Structural finite element model of the wing, showing the front,
how to construct a ROM of generic long-range transport aircraft rear and middle spars, and the ribs
olutions for a given (flexible) Figure 16.6 Structured CFD
4.1 Static aeroelastic loads of a generic long- mesh for XRF-1 (left), finite element model of the wing (right).
n and different load cases (in Reprinted from Ripepi
range transport aircraftet al. (2018) by permission from
is used Springer
for the Nature.
finite element analysis. The coupling between
astic Model block, w/o Sizing the aerodynamic and theM. Ripepimodel
structural et al.is obtained through
e flight conditions). Then Sect. In the following, the performances and effectiveness of the the use of Radial Basis Functions (RBFs) [8] to transfer the
critical load cases based on an ROM approach in predicting aerodynamic loads to the structural model, and a linear
SM (ANSYS) static aeroelastic 5,000static aeroelastic loads for a interpolation of the structural displacements onto the aero-
given aerodynamic shape, but generic long-range transport aircraft fuselage-wing con-
t of
load2430 wall-clock
cases (in Fig. 3, ROM of seconds.
figuration in the transonic flow regime are presented. dynamic grid.
DoE Samples
ck, with Sizing Process, and The3 geometry of the configuration and the Prediction underlying Points The aircraft empty mass is 117,888.2 kg. A flight con-
rium and the trim
Section 4.3 shows a ROM for
correction are
structured grid featuring 784,384 grid points, including 19,211 figuration with 55,000 kg of payload mass and 61,000 kg of
with twoofnested loops. 4,000
redictions a rigid model of Insurface
ordergrid topoints, are depicted in Fig. 4. The aerodynamic fuel mass is considered. Pull-up and pull-down maneuvers,
modeleach
is a Reynolds-Averaged Navier–Stokes (RANS) model of respectively load factor 2.5 g and - 1 g, are analyzed.
craftequilibrium
configuration subject
(outerto loop) The computation of the coupled flow-structure solutions
3, ROM of the Aerodynamic with a negative Spalart–Allmaras type 1-equation turbulence
ation
ers the of the displacements
Geometry), which is modelfrom
[3] and solved using the computational fluid-dynamics were performed in parallel on the DLR C2 A2 S2 E-2 cluster
Altitude [m]

h,
for deformation
a given aerodynamic of shape
the CFD (CFD) DLR TAU solver 3,000
mesh [38, 60]. The computational struc- using two nodes1 with 24 cores. Computing a free-flight
ural sizing sub-process. tural
Fig. 5model (CSM),
Structural finitehowever, is spatially
element model discretized
of the wing, showingbythe
finite
front,
of the flow solution (with inner
rear and
element middle
shells spars, and thewith
and beams ribs 2167 elements (1914 shells,
dsstrategy),
of a genericinterpolation
long- 212ofbeamstheand 41 mass type elements) and 1725 nodes, as 1
InteÒ XeonÒ E5-2695 v2 Processors (30M Cache, 2.40 GHz, 12
is used for
showed the 5.
in Fig. finite
Theelement2,000
ANSYS analysis.
StructuralThe couplingsoftware
Mechanics between Cores).
raft
del to the CSM mesh, and com-
the aerodynamic and the structural model is obtained through
olution.
mances and effectiveness of the the use of Radial Basis Functions (RBFs) [8] to transfer the
123
g static outer
aeroelastic aerodynamic loads to the structural model, and a linear
loads for ais necessary
pling iterations interpolation of the 1,000
structural displacements onto the aero-
rt aircraft fuselage-wing con-
owconvergence.
regime are presented. In each of these
dynamic grid.
The aircraft empty mass is 117,888.2 kg. A flight con-
trimmed solution is obtained
nfiguration and the underlying
figuration with 55,000 kg of payload mass and 61,000 kg of
384 grid points, including 19,211
gy,
ted inwhere
Fig. 4. Thethe angle of attack
aerodynamic fuel mass is is considered. Pull-up 0 and pull-down maneuvers,
0.6g and - 1 0.65 0.7 0.75 0.8 0.85
dift balancing
Navier–Stokes aircraftof weight
the model
(RANS) respectively load factor 2.5 g, are analyzed.
aras type 1-equation turbulence The computation of the coupled flow-structure solutions Mach
to a given load factor. were performed in parallel on the DLR C2 A2 S2 E-2 cluster
e computational fluid-dynamics
stem is computational
, 60]. The solved bystruc- first using
using two anodes1 Fig. with 624Design
cores. Computing a free-flight
of experiment samples and prediction points
is spatially discretized by finite
Figure 16.7 Sample locations for ROM generation in the parameter space (blue) spanned by
gy, running 3500 iterations, fol-
ith 2167 elements (1914 shells, Mach number and altitude and prediction points (red). Reprinted from Ripepi et al. (2018) by
idual strategy,
e elements) and 1725 where
nodes, as thepermission
density
InteÒ XeonÒ from
1
E5-2695Springer
v2 ProcessorsNature.
(30M Cache, 2.40 GHz, 12
Sfour orders
Structural of magnitude
Mechanics software at the
Cores). retained. The performances of the ROM approach are
up to six orders of magnitude evaluated at flight conditions with Mach number 0.81 for
different altitudes. The 123 prediction points and the DoE
upled iterations. However, Thethecoupled calculations were carried out in such a way that the lift counteracts
y is bounded by a maximum of and inertial forces due to theFig.
the weight
sample points are shown in 6.
load factor, which corresponds to the ratio
Before computing the ROM predictions, a leave-one-out
of the lift of cross-validation
an aircraft to its strategyweight, hasand been
represents a global
performed measure of the stress or
to understand
ented, the ROM is parametrized
“load” to which the structure of the aircraft is
if the set of sample points were enough to cover subjected. The snapshots
the were used to
s, i.e., altitude and Mach number.
set up a ROMparameter
based on space. POD orTherefore,
Isomap and thin-plate spline (TPS)
following this strategy, alter- interpolation. The
conditions and configurations
calculation ofnatelya snapshot
one of with TAU and ANSYS
the high-fidelity snapshotstook
of 60
theminutes
DoE sampleon 48 cores, while
or the snapshot computation is a
building the ROM. In this case, the generation set ofhasthe ROM
been left took
out one
from minute
the ROM on one core
generation and the
procedure ROM prediction
elity simulations have been per- of the surface pressure
(which is distribution
then built and
using deformation
the remaining were
21 in
DoE real time
sample on one core.
h numbers ranging from 0.65 to snapshots as the training set, retaining all the 21 POD
ween 0 and 5000 m. The payload modes). In the corresponding flight condition of the left-out
t fixed The sample points are snapshot (i.e., the validation point), the ROM prediction
h d i 0 0 9 8 088962 4 023 bli h d li b C b id i i
 342 Data-Driven Fluid Mechanics Chapter 16

M. Ripepi et al.

Cp Z

Y HiFi: TAU & ANSYS

X
ROM: POD & TPS

Cp
y/b = 25%

0 0.2 0.4 0.6 0.8 1

x/c

Cp
Cp
y/b = 50%

0 0.2 0.4 0.6 0.8 1

Cp

y/b = 25%
y/b = 75%
y/b = 50% Y Z

X
0 0.2 0.4 0.6 0.8 1 y/b = 75%
x/c

Fig. 12 Comparison of the pressure distribution at the 25, 50 and 70% spanwise airfoil sections, for Mach 0.81 and altitude 4000 m

Figure 16.8 Comparison of the pressure

Load cases distributions for the flexible XRF-1 at different wing mass case: MTOW (Max Fuel)
computed stall speed limit (n=-1g)
computed stall speed limit (n=2.5g)

sections for a Mach number of 0.81 and 4 000 m altitude (visualization on jig-shape).
Hi-Fi predicted sizing load cases
ROMs predicted sizing load cases

it
Reprinted from Ripepi
CFD et al. (2018) by
Structure permission from Springer Nature.
Sizing ed
lim
Reduced-order models for aerodynamic applications, loads and MDOve s
pe
di
altitude

g
-1
400 Snapshots 1 km d,
ee
l sp
360
st
al
mass cases has been
g
.5

predictions
2

ROMs Sizing
d,

aerodynamic load ca
e
pe
ls
al
st

Fig. 13 The load case identification process mass cases considered

0.2 Mach

the complete aerodynamic load case identification process.

Ma maximum take-off m
As an example Fig. 14 shows the flight envelope as com- Fig. 14 Design load cases identified with the coupled high-fidelity/
puted with TAU for one particular mass case (MTOW) and ROM procedure and once with maxim
the corresponding identified design load cases.
This process guarantees an efficient search and selection order models after all are just a linear combination of the payload and maximum
approximation of such snapshots. Therefore, it may be
for new critical load cases. However, it should be pointed
out that the prediction capability of the ROMs depends on possible that the ROMs could fail to provide additional
and maximum payloa
the high-fidelity snapshots used to generate them. Reduced- candidates for the design load cases, which might turn out five mass cases was c

and of 1000 m in altit

123 After computing th
flexible aircraft (i.e., t
of the static aeroelasti
to the sizing tool S-
output to the designe
Fig. 11 Comparison of the wing structural displacement between the sized structure able t
Figure 16.9 Comparison
coupled high-fidelity of steady ROM model with coupled CFD/CSM
(TAU-ANSYS) and the reduced simulation. order Reprinted from
iterated, until conver
Ripepi et al. (2018)modelby permissionatfrom
(POD-TPS), Springer Nature.
the validation point h ¼ 1000 m, Ma ¼ 0.75,
n ¼ 2:5 g
namic loads associat
with the currently siz
In Figure 16.8, Nevertheless
the ROM prediction of the surfacewhen
these discrepancies, pressure ROM is (red)Since
the distribution for a
this process
POD-based ROMs
Mach numberemployed of 0.81 and within
4 000the m MDO
altitudeprocess,
is compared it would withnot affect the
a TAU-ANSYS reference
resultThe of the optimization in terms of final SMARTy toolbox. H
solution (black). agreement is very good; only design,
the gradient as will of be the pressure
parameter space with
distribution atshown the shock for the sizing slightly.
deviates process For in Sect. 4.4.
all prediction points, the relative error
tudes. The ROM pred
As final remark, a different physical modelling (e.g.,
in the aerodynamic coefficients is below 0.5%. An excellent agreement sizing betweentool, which d
unsteady RANS or Hybrid RANS/LES) would be neces-
the elastic wing deformation predicted by the ROM and the coupled high-fidelity Whenever a newly pre
sary in those sample points of the flight envelope where the
reference TAU-ANSYS computation is observed in Figure 16.9. potentially critical, th
aeroelastic solution is not converged, to properly model the
unsteadiness of the separated flow and the effect of the puted with the high-fi
time-varying aerodynamic load on the structure and the and checked with the
h d i 0 0 9 8 088962 4 023 bli h d lli tib Ct i b id i i
As an example tw
 Görtz et al: ROMs for Aerodynamic Applications and MDO 343

16.4.4 Reduced-Order Models in the Context of Multidisciplinary

Design Optimization
A gradient-based optimization algorithm can efficiently spot local optima in high-
dimensional design problems. A gradient-based approach, which can tackle hundreds
of design parameters controlling the wing profile shapes and the structural thicknesses
with a single optimizer, was used for aero-structural optimization of the XRF-1 (Görtz
et al. 2016). Such optimization problems comprise the computation of disciplinary as
well as interdisciplinary gradients, that is, gradients of, for example, aerodynamic
quantities with respect to shape parameters as well as gradients of aerodynamic
quantities with respect to structural parameters. The adjoint approach was deemed
suitable for computing the aerodynamic gradients since the number of design variables
is considerably higher than the number of constraints (Abu-Zurayk et al. 2017). On
the structural side, a parallelized finite differences (FD) approach was used to compute
the gradients. FD was feasible because the structural problem is linear, and its solution
costs much less than that of the aerodynamic problem. Disciplinary gradients were
compared with interdisciplinary gradients, and only the most significant gradients
were employed during the optimization.
Aiming to an aeroelastic structural optimization of the generic long-range transport
aircraft able to withstand the critical loads, the Mach-altitude envelope for five
design mass cases has been computed for the baseline XRF-1 configuration using
high-fidelity coupled CFD–CSM methods, and the relative design aerodynamic load
cases have been determined. The Mach-altitude envelope for the five mass cases was
computed in intervals of 0.02 in Mach number, and of 1 000 m in altitude for two load
factors (−1.0, 2.5 g). After computing the aerodynamic load envelope of the flexible
aircraft (i.e., the aerodynamic pressure distribution of the static aeroelastic solution),
all the loads were passed to a sizing tool. This tool provides as an output to the designer
the critical loads and the sized structure able to sustain such loads. The process is
iterated, until convergence, by recomputing the aerodynamic loads associated to the
static aeroelastic solution with the currently sized structure.
Since this process is computationally very expensive, POD-based ROMs were
employed using the DLR’s SMARTy toolbox. Here ROMs were used to explore the
parameter space with a finer sampling at in-between altitudes. The ROM-predicted
aeroelastic loads are sent to the sizing tool, which determines if the loads are critical.
Whenever a newly predicted aeroelastic load is found to be potentially critical, the
corresponding load case is recomputed with the high-fidelity coupled CFD–CSM
methods and checked with the sizing tool if it is really critical or not.
As an example, two of the five critical mass cases were used to generate 400
sized high-fidelity aeroelastic snapshots. A parametric reduced-order model has been
generated using such snapshots, and then used to compute 360 additional load
predictions. Three of these 360 predictions were found to be additional candidates
for design load cases, and by computing them with the high-fidelity tools, one case
was found to be actually critical.
After identifying the sizing load cases for the baseline configuration, their definition
was frozen, assuming that a gradient-based optimization only results in small changes

h d i 0 0 9 8 088962 4 023 bli h d li b C b id i i

 344 Data-Driven Fluid Mechanics Chapter 16
S. GÖRTZ, Č. ILIC, M. ABU-ZURAYK, R. LIEPELT, J. JEPSEN, T. FÜHRER, R. BECKER, J. SCHERER,
T. KIER, M. SIGGEL

feasible SQP algorithm may fail to satisfy all

the constraints, or take too many iterations to do
so. With the FSQP algorithm, on the other hand,
an inexact gradient will result in stopping before
the true optimum has been reached, but the final
design will be guaranteed to be feasible and
might have a sufficient improvement in the
objective over the initial design.

Fig. 17. Gradient-based optimization convergence. Only

accepted designs after each line search are shown.
Figure 16.10 Gradient-based, high-fidelity aero/structure optimization; baseline aircraft
configuration (left), course of optimization in terms of objective of
The comparison off-design
function performance
and quantities of of
the ROM-based
interest as a function of number of iterations after optimized and baseline
selection of designs is given
sizing load cases.by
the aeroelastic lift-to-drag polar in Fig. 18 at
design Mach number. There is a small peak in
the objective improvement at the cruise design
to the wing shape and the structure. The objective of the gradient-based MDO was to
point, due to its higher weight of 0.40 in the
maximize the range factor for the same transport aircraft
multi-point at thebut
objective, design
not atcruise point as
the expense of
well as at 4 off-design points. The optimization resultedoff-design
the nearby in an increase in thewhich
points, range are
by
a factor of 6%, see Figure 16.10. improved over the baseline as well.

16.4.5 Aeroelastic Reduced-Order Model Based on Synthetic Modes and

Gappy Proper Orthogonal Decomposition
Recently, a derivative of the constraint gappy POD method (Franz & Held 2017)
has been developed and applied for the rapid prediction of aerodynamic loads while
taking
Fig. 16. Drag elastic
(top) and massdeformations into account.
(bottom) gradients with This is of interest in steady fluid–structure
respect to interaction
aerodynamic and analysis.
structuralFirst,
design a set of synthetic surface deformations around an estimated
variables

elasticthe
This is indeed axis is determined,
result that was reached the inso-called
the synthetic modes, which are an established
present case,
conceptas illustrated
for unsteady by the optimization
aeroelastic Fig. 18. Comparison of aeroelastic polars of baseline and
investigations including low-fidelity aerodynamic
convergence plot in Fig. 17. The objective is the optimized designs
models (Voss
range coefficient, whichetisal.equal
2011).to the product
Second,
of the lift-to-drag samples
ratio are computed
and inverse of the mass with TAU for the different synthetic mode shapes
coefficient.
based Theon objective was increased
a Halton sampling strategy 7 Conclusions
by (Kuipers & Niederreiter 2005), and a POD basis
about 6% from the which
is computed, baseline, which the
contains was deformations
DLR has corresponding
developed atocollaborative
the mode shapes MDO
achieved both by increase in lift-to-drag ratio platform based on DLR standards and more than
and by as well as
decrease the mass
in the surface forces computed
coefficient. The with TAU. Finally, surface forces for
40 established disciplinary tools. A multi-level
a
total rungiventimedeformation of interestwas
for this optimization can 80
then rapidly be obtained by solving a least-squares
MDO architecture has been implemented,
hours, using 96 computing
problem, minimizing cores
the of an HPC between
difference the target
including an surface
overall deformation
aircraft synthesisgiventool
by
cluster. theIn current
total 230 CFD-CSM
structural modelsimulations
of interest and theandeformation
chain, entries
automatic loads within
process basedtheonPOD
mid-
(load cases and performance cases) were fidelity aerodynamics to identify configuration-
performed. modes to find a set of POD coefficients. The resulting POD coefficients can then be
specific load cases, aerodynamic performance
used to compute the surface force by a simple matrix-vector product. A more
computations for the aeroelastic aircraft with
in-depth
theoretical introduction, including synthetic mode generation and a formulation
automatic hybrid unstructured meshing for of
the constraint least-squares optimization problem, can be found in Bekemeyer et al.
(2019). Here results are presented for a long-range passenger aircraft with a semi- 10

wingspan of approximately 29.0 m shown in Figure 16.11. During each analysis, an

iterative fluid–structure coupling process is performed including a structural optimiza-

h d i 0 0 9 8 088962 4 023 bli h d li b C b id i i

 Görtz et al: ROMs for Aerodynamic Applications and MDO 345

Figure 16.11 Assessment of an aeroelastic reduced-order model based on synthetic mode

shapes; mode shapes and computational grid (top), global deformation behavior (bottom left)
and optimized skin thicknesses (bottom right).

tion in each iteration step to minimize weight while respecting stress constraints. The
final deformation behavior of the wing for a transonic load at M = 0.784 is compared
between the proposed reduced-order model approach and a fully coupled CFD–CSM
analysis with excellent agreement. Also final skin thicknesses show only minor devia-
tions around the engine pylon wing junction. However, the online cost for performing
the ROM-based analysis is a factor of nearly 2 000 faster than its full-order equivalent.

16.4.6 Nonlinear Dimensionality Reduction by Autoencoder Networks

The POD-based ansatz is still considered as the state-of-the-art method in reduced-
order modeling, but also nonlinear dimensionality reduction methods like Isomap and
autoencoders have been used to build ROMs. An autoencoder (Goodfellow et al. 2016)
is a neural network used for unsupervised learning of efficient codings. The aim of
an autoencoder is to learn a representation (encoding) for a set of data, typically for
the purpose of dimensionality reduction or representing data in a lower-dimensional
space. The decoder is used for mapping back to the original dimension of the
problem (reconstruction), see Figure 16.12. There exist several modifications of the
autoencoder; however, only the standard method was employed here (Hinton 2006).
The method was used in the context of reduced-order modeling of distributed
aerodynamic data. Instead of computing a reduced-order representation of the data
via POD, the autoencoder was employed to compute the reduced-order representation.
The method directly yields a mapping from the reduced-order representation to the
high-dimensional space; however, for huge input data (large dimension) it is difficult

h d i 0 0 9 8 088962 4 023 bli h d li b C b id i i

Ms > Franz • 30. April 2013
346 Data-Driven Fluid Mechanics Chapter 16

ncoders
Decoder

Reduction, i.e.
: x ∈ Rn 7→ x̃ ∈ Rñ ,
n1

multilayer artificial
k n2

truct the input after

n3
eneck Code layer

ood results for

n2
es with sharp edges
ies) were observed in
n1

k with linear activation

ers is quite similar to
Encoder

Figure 16.12 Schematics of an autoencoder network.

and time-consuming to train a deep network. Training times can be significantly

reduced on GPUs when compared with CPUs.
The potential of autoencoder networks for nonlinear dimensional reduction was
evaluated using a two-dimensional test case from the “virtual aircraft” application
area. Thirty steady flow solutions were computed for the NACA 64A010 airfoil at
transonic flow conditions with the DLR flow solver TAU by varying the angle of attack
and the Mach number as shown by the black symbols in Figure 16.13. Predictions
were made for three different flow conditions (red crosses in Figure 16.13) and
compared with TAU reference solutions in terms of the surface pressure distribution.
The autoencoder predictions were also compared with ROM predictions based on
Isomap, POD and cluster POD (C-POD). K-means clustering was used to cluster
a given set of parameter-dependent flow solutions in k clusters. The clustering was
performed based on either the flow parameters or the reduced representation of the
solutions given by Isomap. On each cluster, a local POD-based reduced-order model
was built, which is similar to local principal component analysis (LPCA) described in

h d i 0 0 9 8 088962 4 023 bli h d li b C b id i i

 www.DLR.de • Folie 27 > ROMs > Franz •
Görtz et al: ROMs for Aerodynamic Applications and MDO 30. April 2013
347
Test case
Latin hypercube
0.82 sampling
(30 snapshots ∈ R400 )

0.8

Mach
0.78
0.5

0.76

Z
0

-0.5

0.74
0 0.5 1
X
4 5 6 7 8 9 10
α
www.DLR.de • Folie 31 > ROMs > Franz • 30. April 2013 NACA64A010 airfoil
• Snapshots
Results × Prediction points

Tau
POD
C-POD
−1
AE
Isomap

α = 8.85
Cp

0
Mach = 0.79

0 0.2 0.4 0.6 0.8 1

x

Figure 16.13 Autoencoder network for reduced-order modeling of distributed aerodynamic

data: sample locations (black) and trial locations (red) in parameter space spanned by angle of
attack and Mach number (top) and surface pressure predictions (bottom) with different
interpolation-based ROMs at an untried parameter combination compared to TAU reference
solution.

Chapter 15, where PCA is applied separately on clusters of observations. To predict a

solution at an untried parameter combination, the model belonging to the cluster with
the nearest mean was exploited.
Figure 16.13 compares the different ROM predictions with the reference solution
at one of the trial points. The quality of the autoencoder (AE) prediction exceeds that
of the established linear dimension reduction method (POD), which is the case for all
three trial flow conditions; however, Isomap is seen to be superior to the autoencoder
in terms of predicting the shock location and shock strength, while C-POD turns out
to be only slightly better than POD. Note that in this study, convolutional autoencoder
networks in particular were found to reproduce the strong nonlinearities in the quantity
of interest very well, even for a small number of snapshots.
First, training data is generated by simulating a training maneuver (e.g., the 1-cos
gust response of an aircraft), which is step 4 in Figure 16.14. This data is then used

h d i 0 0 9 8 088962 4 023 bli h d li b C b id i i

 348 Data-Driven Fluid Mechanics Chapter 16

Gust amplitude vg
Training input 2 Training output 3
1
𝐂L (Lg, vg)1
(Lg, vg)X (Lg, vg)2
(Lg, vg)3

Gust Length Lg time

Compute steady Define training Compute training
solution 𝒘𝟎 points outputs with uRANS

4 5 6 ROM predictions
POD model
𝒘 𝒕 = 𝚽𝐚
𝚽:POD-modes
Use residual min. for 𝐚
𝑚𝑖𝑛 ∥ 𝐑෡ 𝚽𝐚 + 𝐰 ∥2L
𝐚 2

Store full flow field Build residual Predict response

solutions minimization model behavior

Figure 16.14 Unsteady ROMs in a nutshell.

to obtain a subspace representation using POD. Second, a subset of the grid points
is computed by combining the classical Discrete Empirical Interpolation Method
method and a greedy missing point estimation. This subset of grid points is used
to restrict the evaluation of the unsteady flow solver residual to a small number of
points during the online prediction phase for reasons of computational efficiency
(Bekemeyer et al. 2019) (step 5 in Figure 16.14). The resulting model can then be
used to simulate gust response behavior at varying parameter combinations such as
gust length and amplitude by minimizing the unsteady residual on the subspace at
each time step. ROM results are compared to a time-marching reference solution
for a 1-cos gust with length of 213 m and amplitude as defined by international
certification authorities at cruise conditions for NASA’s CRM, compare Figure 16.2.
In addition, the linearized frequency domain solver (LFD) based on the TAU code
(Thormann & Widhalm 2013) was adopted to efficiently predict the dynamic behavior
of the CRM assuming transonic and mildly separated flow and dynamically linear
response behavior. The LFD was originally developed, offering a significant reduction
in computational effort for small-perturbation problems while retaining the fidelity
of the RANS flow solutions. The lift coefficient, pitching moment coefficient, and
the differences in surface pressure distribution at the time step corresponding to
the maximum lift coefficient are shown in Figure 16.15. Throughout, the unsteady
ROM accurately predicts the full-order model (FOM) behavior, whereas the linearized
approach deviates once flow separation occurs. Significant discrepancies between the
three approaches are observed in the surface pressure distribution.

h d i 0 0 9 8 088962 4 023 bli h d li b C b id i i

 Görtz et al: ROMs for Aerodynamic Applications and MDO 349

Figure 16.15 Comparison of gust response behavior of the CRM for the full-order model
(FOM), unsteady nonlinear ROM and linear frequency domain (LFD) solver. Lift and pitching
moment coefficient (left), surface pressure differences at maximum lift coefficient (right).

16.4.7 Unsteady Physics-Based Nonlinear Reduced-Order Modeling

The investigation of unsteady, gust induced air loads is a crucial part of aircraft design,
and time-marching CFD simulations are infeasible due to their high computational
cost. However, nonlinear aerodynamic effects should not be neglected as is the case
for lower fidelity methods. Thus, based on previous work on steady transonic flows
using a combination of POD and nonlinear least-squares-based residual minimization
techniques (Zimmermann & Görtz 2010, 2012), an unsteady nonlinear reduced-
order model (LSQ-ROM) was developed to predict dynamically nonlinear maneuver
and gust response behavior in the transonic flow regime (Bekemeyer et al. 2019).
The schematics of the unsteady physics-based nonlinear reduced-order modeling
procedure are sketched in Figure 16.14.

16.5 Conclusions

Parametric ROMs based on both POD, cluster POD, and Isomap were developed and
successfully applied to a variety of two- and three-dimensional test cases, including
two industrial aircraft configurations, to predict aerodynamic loads over a range of
flow or operating conditions as well as for changing geometry parameters. Isomap
is a dimensionality reduction technique, which is based on a nonlinear “manifold
learning” method. In this method, it is assumed that the CFD solutions form a
nonlinear manifold, which in turn constitutes a sub-manifold of lower intrinsic
dimensionality. Isomap yields better predictions than a POD-based ROM method,
in particular for the location and magnitude of the shock wave at transonic flow
conditions. This is due to the fact that Isomap makes predictions based on a local
subspace, similar to cluster POD, while POD-based predictions make use of a global
subspace.

h d i 0 0 9 8 088962 4 023 bli h d li b C b id i i

 17 Machine Learning for Turbulence
Control
B. R. Noack, G. Y. Cornejo Maceda, and F. Lusseyran

Closed-loop turbulence control has current and future engineering applications of

truly epic proportions, including cars, trains, airplanes, jet noise, air-conditioning,
medical applications, wind turbines, combustors, and energy systems. A key feature,
opportunity and technical challenge, is the inherent nonlinearity of the actuation
response.
This chapter outlines model-based and model-free control taming the nonlinear
dynamics. Model-based control employs the POD Galerkin method (see Chapter 14:
Noack et al.: Machine learning for Reduced-Order Modeling). Many flow control
results will be explained with a set of few different sparse human-interpretable models.
Artificial intelligence (AI)/machine learning (ML) has opened another game-changing
new avenue: the automated model-free discovery and exploitation of unknown
nonlinear actuation mechanisms directly in the plant. Variants of this machine
learning control (MLC) will be discussed. The chapter concludes with a tutorial of
MLC on a simple dynamical system with a freely available MATLAB/Octave code.

17.1 Introduction

Controlling turbulence for engineering goals is one of the oldest and most fruitful
academic and technological challenges. Engineering applications have epic propor-
tion. The infrastructure of any industrial nation requires flawless pipe systems with
turbulent flows of drinking water, gas, oil, and sewage. Drag reduction in these pipe
flows spells energy saving for operation. Transport-related examples include drag
reduction of road vehicles, airborne transport, ships, and submarines, lift increase
of airfoils, and associated noise reduction. Energy and production-related examples
include efficiency increase of harnessing wind and water energy, heat transfer,
chemical, and combustion processes.
Closed-loop active control is increasingly investigated for performance improve-
ments in industrial applications. Several trends support this development. First,
aerodynamic shapes and passive control have been optimized for over 100 years.
Thus, further optimization efforts may yield diminishing returns. Second, in many
cases the engineering challenge is not the cruise or normal operating condition but

h d i 0 0 9 8 088962 4 024 bli h d li b C b id i i

 Noack et al: Machine Learning for Turbulence Control 351

short-term and rare events. For wind turbines, the effect of gusts needs to be mitigated.
Aeroengines need to be safe during 60 to 90 seconds of takeoff, where they can
produce six times the thrust required during cruise. The large nacelle is only required
for this takeoff and may be replaced by a smaller version with active control. Third,
actuators and sensors become increasingly cheaper, more powerful, more reliable, and
thus more economical. Finally, modern control methods can harness increasingly more
complex dynamics, last but not least through the powerful augmentations by ML.
Goals, tools, and principles of turbulence control are reviewed in Section 17.2.
This chapter focuses on the control logic. The classical paradigm is from under-
standing to dynamic modeling to model-based control design to controller tuning in
the full plant. Key applications are first (relaxational) and second-order (oscillatory)
dynamics. The nonlinear dynamics need to be linearized or at least sufficiently
understood. In Section 17.3, we present simple human-interpretable POD-based
models showing the spectrum from linear to strongly nonlinear dynamics explaining
a large range of current experiments.
Building a control-oriented model is a challenge for shear turbulence with a
large range of temporal and spatial scales and their nonlinear interactions. The
challenge becomes even larger for spatially distributed control with many actuators
and many sensors. Here, powerful methods of ML may fully invert the classical
control paradigm. First, the near-optimal control law is automatically learned in the
full plant without any model. Then, simpler control laws with similar performance
are identified. The performance-winning mechanism is then dynamically modeled,
leading to a deeper understanding of the flow. In Section 17.4, different realizations
of this MLC are described. The chapter concludes with a MLC tutorial for the control
of nonlinearly coupled oscillators (Section 17.5). For more information, we refer to
our recent reviews (Brunton & Noack 2015, Brunton et al. 2020) and our textbooks
on model-based control (Noack et al. 2011) and MLC (Duriez et al. 2017) and the
literature cited therein.

17.2 Goals, Tools, and Principles

This section provides a high-level perspective on turbulence control, reviewing typical

goals (Section 17.2.1), hardware tools (Section 17.2.2), and control principles (Section
17.2.3). Subsequent sections assume that the configuration with goals, actuators, and
sensors is given and the control logic needs to be developed.

17.2.1 Goals
Turbulent flows around cars, trucks, trains, airplanes, helicopters, ships, and sub-
marines affect their performance. An important engineering goal is drag reduction
or, more precisely, the reduction of net propulsion power and thus energy saving.
Airplanes during takeoff and landing require high lift at low velocities. Here, a
goal is to achieve lift increase with minimal means, for example, low extra weight

h d i 0 0 9 8 088962 4 024 bli h d li b C b id i i

 352 Data-Driven Fluid Mechanics Chapter 17

and small material fatigue. During cruise, the increase of lift-to-drag ratio leads to
reduced fuel consumption, and hence reduced environmental impact and increased
profitability. Ground and airborne traffic produce undesirable noise. Here, noise reduc-
tion is another important goal. Chemical production, combustion processes, and heat
exchangers profit from mixing increase. Another important goal is the stabilization
of flow oscillations. Examples include flutter and buffeting of airfoils, cavity noise
(Rossiter modes) associated with landing gears of airplanes and detrimental vortex
shedding around buildings, chimneys, bridges, the pillars of oil-producing platforms,
underwater pipelines, and heat-exchanging pipes in nuclear plants.
The goals of drag reduction, lift increase, mixing increase, noise reduction, and
oscillation mitigation may be served by separation prevention, skin-friction reduction,
and manipulation of free shear flows. As a rule of thumb, in most flow control strate-
gies the transverse mixing of free and wall-bounded shear flows is increased (drag
reduction, mixing increase) or decreased (skin friction reduction, noise reduction,
oscillation prevention). We refer to the excellent discussion of underlying flow physics
by Gad-el Hak (2007).

17.2.2 Tools
Flow control performance during cruise or standard operating conditions is well
served by shape optimization under given constraints. An ubiquitous example is cars,
which have become increasingly more aerodynamic over the past hundred years.
The performance may be further improved by passive devices, that is, small changes
that require no energy input. Feathers at the end of arrows stabilize the flight and
increase the range – perhaps the first example of man-made passive control. Another
example is vortex generators on the wing of a Boeing 737 to prevent early separation
with associated lift decrease and drag increase. The efficiency of wind turbines may
also profit by up to 7% from SmartBlade vortex generators. Riblets reduce skin-
friction drag by up to 11%. The chevron-shaped engine nozzles reduce jet noise by 1–3
dB. Spoilers at the trailing edge of a car stabilize aerodynamic forces and reduce drag.
Vanes in ducts and at airplanes redirect the flows in a performance increasing way.
Passive devices may come with parasitic drag during cruise or standard operating
conditions. In many cases, passive devices, like vortex generators, may be emulated by
active (energy-consuming) devices, like fluidic jets. Active control has the advantage
that it can be turned on or off and even operate at a large range of amplitudes and fre-
quencies. This dynamic range often yields performance benefits in contrast to steady
operation. Active control may be operated open-loop, that is, be blind to the flow
state. Yet there may be significant potential in combining unsteady active control with
sensor-based flow monitoring. In the sequel, we will focus on this closed-loop control.
The choice of the actuators, their kind, amplitude, frequency range, and placement
has a large effect on control efficiency. Actuators are often placed at high-receptivity
locations, for example, at the trailing edge of a bluff body, where the separating shear
layer can be easily manipulated. The coherent structures associated with boundary-
layer transition and skin friction can be directly opposed with wall motion or wall

h d i 0 0 9 8 088962 4 024 bli h d li b C b id i i

 Noack et al: Machine Learning for Turbulence Control 353

actuation. The actuation may be based on blowing, suction, and zero-net mass flux,
each having its distinct advantages and disadvantages.
Yet, this choice is largely based on engineering experience with little guidance from
mathematical and physical methods. The closest to a first-principle based method is
the volume force optimization of linearized Navier–Stokes dynamics. Making a good
choice of actuation is a challenge. The scaling properties to full-scale engineering
configurations is the next task, again with no or little guidance from available
theories.
Optimizing sensor placement is easier and may be guided by a direct numerical
simulation or particle image velocimetry measurements. Yet, the choice of the
optimization criterion is far from trivial. The chosen sensors may, for instance, be
placed at locations to optimize (1) linear flow estimation, (2) linear-quadratic flow
estimation, (3) Kalman filter-based flow estimation, (4) a dynamic observer based
on a reduced-order model, (5) the estimate of the observable subspace (balanced
truncation), or, in the most general case, (6) the feedback control performance for
all considered control laws. In the following, we assume a specified control goal and
a given actuator and sensor configuration.

17.2.3 Principles
The choice of the actuators and sensors is – lacking a computationally accessible
mathematical method – always guided by a conceptual idea of the control mechanism.
A long list of hypotheses can be produced. Skin-friction reduction relies on the
suppression of sweeps and ejections (opposition control). Separation can be delayed
by increasing the turbulence level (destabilizing control). Heat exchangers also rely
on mixing increase. Vortex shedding behind bluff bodies can be mitigated with phasor
control, or exciting larger or smaller frequencies and blowing the wake away (Coanda
blowing, base bleeding). A few general dynamics principles can be distilled.

Kill the monster while it is little! The ideal scenario of stabilizing control is that the
instability is successfully fought in statu nascendi. This requires vanishing
energy in a noise-free environment and a noise-dependent level otherwise.
Examples are mitigation of a Tollmien–Schlichting instability with local
opposition control, the reduction of skin-friction drag by opposing the wall-
normal velocity of sweeps and ejections and Coanda blowing / flow vectoring
to symmetrize the wake.
Support the enemy of your enemy! In some cases, the actuation may be too weak
or distant for a direct stabilization. We might exploit that it is much easier
to excite a coherent structure than to prevent an instability. Vortex shedding
behind a cylinder may be mitigated by exciting shear-layer vortices at high
frequency. In this case, the shear-layer structures decrease the gradient of the
mean flow and hence the growth rate of vortex shedding. The energized shear
layer is the enemy of the enemy (vortex shedding).

h d i 0 0 9 8 088962 4 024 bli h d li b C b id i i

 354 Data-Driven Fluid Mechanics Chapter 17

Blow the instability away! A not-so-subtle control is based on blowing the instabil-
ity away. The vortex shedding behind a D-shaped cylinder is successfully
reduced by Coanda forcing, that is, redirecting the flow into the near wake.
Support your friend! In some cases, naturally occurring coherent structures serve
the control goal, that is, they are your friends. Kelvin–Helmholtz vortices
over a backward-facing step reduce the recirculation zone. An amplification
of these vortices reduces the length of this zone even further.
Manipulate the sociology of coherent structures! Many aerodynamic goals
have a well-founded conceptual picture of the beneficial and detrimental
coherent structures to be augmented or mitigated. For mixing and noise of
broadband turbulence, such a picture is largely missing. Mixing and noise is
based on a longer-term integration of many different structures. This is the
classic case for model-free exploration and exploitation of control.
To illustrate these principles, let us consider following dynamical system coupling
a self-amplified amplitude limited oscillator (a1, a2 ) and a stable linear oscillator at
10-fold frequency (a3, a4 ) inspired by several wake and airfoil models (Luchtenburg
et al. 2009, Semaan et al. 2015):
da1 /dt = σa1 − a2, da3 /dt = −0.1a3 − 10a4,
da2 /dt = σa2 + a1 + b1, da4 /dt = −0.1a4 + 10a3 + b2, (17.1)
σ = 0.1 − a12 − a22 − a32 − a42 − b3 .
Without control, b1 = b2 = b3 = 0, the first oscillator converges to the limit cycle
a12 +a22 = 0.1 with unit frequency, while the second one vanishes, a3 = a4 = 0. We will
ignore the stable oscillator (a3, a4 ) unless it is excited with b2 . The first oscillator can
be stabilized with phasor control b1 = −0.4 a2 . This oscillator can also be indirectly
stabilized by exciting the second oscillator b2 = ka4 to fluctuation level above 0.1 so
that σ < 0. The constant blowing is mimicked by b3 = 0.2, leading to σ ≤ −0.1, that
is, stabilization.

17.3 Nonlinear Low-Dimensional Models for Control

Despite the powerful tools for linear model reduction and control, the assumption
of linearity is often overly restrictive for real-world flows. Turbulent fluctuations are
inherently nonlinear, and often the goal is not to stabilize an unstable fixed point but
rather to change the nature of a turbulent attractor. Moreover, it may be the case that
the control input is either a bifurcation parameter itself, or closely related to one, such
as the control surfaces on an aircraft.
The degree of nonlinearity is most easily characterized in a Galerkin modeling
framework discussed in Chapters 1 and 15. We assume a Galerkin model with the
steady solution u s as basic mode and N expansion modes ui , i = 1, . . . , N,
Na
Õ
u(x, t) ≈ û(x, t) = u s (x) + ai (t) ui (x). (17.2)
i=1

h d i 0 0 9 8 088962 4 024 bli h d li b C b id i i

 Noack et al: Machine Learning for Turbulence Control 355

Thus, the state is approximated by the vector a = [a1, . . . , a N ]T comprising the modal
amplitudes. The Galerkin system reads
N N
dai Õ Õ
= fi (a, b) = li j a j + qi jk a j ak + gi b. (17.3)
dt j=1 j,k=1

Here, the constant term vanishes identically, because the steady Navier–stokes solution
u s corresponds to the fixed point a s = 0 in the Galerkin framework. Following the
example of Chapter 1, a single linear control input gi b is added.
In the following sections, three prototypic examples are discussed: (1) an oscillation
around the fixed point (Section 17.3.1); (2) a self-excited amplitude-limited oscillation
(Section 17.3.2); (3) frequency cross-talk with two different frequencies (Section
17.3.3); over the base-flow deformation. More details and the remaining irreducible
cases are elaborated in Brunton and Noack (2015).

17.3.1 Linear Dynamics

First, a small oscillatory fluctuation around a steady solution is considered. Examples
include the flow over a backward-facing step (Hervé et al. 2012) at subcritical
Reynolds number with noise, transition delay of a boundary layer (Semeraro et al.
2013), or stabilization of a cylinder wake (Roussopoulos 1993, Weller et al. 2009).
In the following, a control-oriented oscillator model is presented as a least-order
description. Then energy-based control design is exemplified as a powerful method
of linear and nonlinear models.
The considered flows can be described by

u(x, t) = u s (x) + u u (x, t), (17.4a)

u (x, t) = a1 (t) u1 (x) + a2 (t) u2 (x).
u
(17.4b)

Here, ui , i = 1, 2 may correspond to the real and imaginary part of the unstable
complex eigenmode or are distilled from the oscillatory data. Higher harmonics are
neglected. By construction, the stable or unstable fixed point is a s = 0.
The linearized version of the dynamics (17.3) reads
d σ 0
 u
−ωu
  
a = A0 a + B b, where A0 = , B= , (17.5)
dt ωu σ u g
where λ = σ u ± ı ωu denotes the complex conjugated eigenvalue pair of the linear
system. Without loss of generality, the modes can be rotated so that the gain in the
first component vanishes. A similar equation holds for the measurement equation. The
matrices A0 and B depend on the fixed point a s . The linear dynamics (17.5) may also
be an acceptable approximation for turbulent flows with dominant oscillatory behavior
with small modifications (Brunton & Noack 2015)
Control design of the linear system (17.5) can be performed with one of many
control theory methods (Aström & Murray 2010). Here, we illustrate the idea of
energy-based control, which is particularly suited for nonlinear dynamics. Moreover,

h d i 0 0 9 8 088962 4 024 bli h d li b C b id i i

 356 Data-Driven Fluid Mechanics Chapter 17

energy-based control has a kinematic relation to phasor control and an energetic

relation to opposition control.
The growth rate σ u is assumed positive and small enough so that the timescale of
amplitude growth is small compared to the timescale of oscillation. In this case, the
state can be approximated by the polar coordinates r, θ with a1 = r cos θ, a2 = r sin θ.
Here, r and ω = dθ/dt are slowly varying functions of time. The energy equation is
given by
dr 2 da1 da2
= 2a1 + 2a2 = 2σ u r 2 + 2g a2 b.
dt dt dt
The control goal is an exponential decay of the amplitude with σ c < 0, that is,
dr 2 !
= 2σ c r 2 .
dt
Eliminating the time derivative in both equation yields
dr 2
= 2σ c r 2 = 2σ u r 2 + 2g a2 b.
dt
The control command b increases (decreases) the energy r 2 /2 if it has the same
(different) sign as a2 . To prevent wasting actuation energy with the wrong phase,
the linear ansatz b = −k a2 is made. The gain k > 0 is determined by substituting
a2 = r cos θ in the energy equation and averaging over one period. The resulting
control law reads
σ c − σu
b=2 a2 . (17.6)
g
The gain increases with the difference between the natural and design growth rate σ c −
σ u and decreases with the forcing constant g in the linear dynamics. The factor 2 arises
from the fact that the actuation is only effective in the [0, 1]T direction. The control
design is very simple and immediately reveals the physical mechanism. The method
is easy to generalize for nonlinear systems, particularly if the fluctuation is composed
of clean frequency components. Related approaches are the Lyapunov control design
(Khalil 2002) and harmonic balance (Jordan & Smith 1988).

17.3.2 Weakly Nonlinear Dynamics

As a refinement to the linearization, mean-field theory is recapitulated (Stuart 1958,
Stuart 1971), providing an important nonlinear amplitude selection mechanism. The
onset of vortex shedding behind a cylinder is one well-investigated example fitting
this description (Schumm et al. 1994). This section and Section 17.3.2 outline the
dynamical model and the corresponding control design, respectively.
Qualitatively, mean-field theory describes the feedback mechanism between the
fluctuations and the base flow. The fluctuation gives rise to a Reynolds stress, which
changes the base flow. The base-flow deformation generally reduces the production
of fluctuation energy with increasing fluctuation level until an equilibrium is reached.
The resulting evolution equations are also referred to as weakly nonlinear dynamics,
as they describe a mild form of nonlinearity.

h d i 0 0 9 8 088962 4 024 bli h d li b C b id i i

 Noack et al: Machine Learning for Turbulence Control 357

The fluctuation has the same representation as in Section 17.3.1. However, the base
flow is allowed to vary by another mode u3 , called the 0th, base deformation, or shift
mode (Noack et al. 2003). This mode can be derived from the (linearized) Reynolds
equation and is assumed to be slaved to the fluctuation level. The resulting velocity
field ansatz reads

u(x, t) = u s (x) + u u (x, t) + u ∆ (x, t), (17.7a)

u (x, t) = a1 (t) u1 (x) + a2 (t) u2 (x),
u
(17.7b)
u ∆ (x, t) = a3 (t) u3 (x). (17.7c)

The evolution equation is given by

d a1 a1
   
= A(a3 ) + B b, A(a3 ) = A0 + a3 A3, (17.8a)
dt a2 a2
 
a3 = αu a12 + a22 , (17.8b)

where

σu −ωu −βu −γ u 0
     
A0 = , A3 = , B= . (17.9)
ωu σu γu −βu g

Without loss of generality, αu > 0. Otherwise, the sign of the mode u3 must be
changed. A nonlinear amplitude saturation requires the constant βu > 0 to be positive.
For a3 ≡ 0, (17.8) is equivalent to (17.5). However, (17.8) has a globally stable
limit cycle with radius r ∞ = σ u /α u βu in the plane a = a∞ = σ u /βu , and with
p
3 3
center 0, 0, a3∞ .
 

In the framework of weakly nonlinear stability theory, the growth rate is considered
a linear function of the Reynolds number σu = κ(Re − Rec ). Here, Rec corresponds to
its critical value where the steady solution with damped oscillations (σu < 0) becomes
unstable (σu > 0) in a Hopf bifurcation, giving rise to self-amplified oscillatory
fluctuations. The other parameters are considered to be constant. This yields the
famous Landau equation for the amplitude dr/dt = σ u r − βr 3 , β = αu βu and a
corresponding equation for the frequency. The Landau equation explains the famous
√
square root amplitude law r ∝ Re − Rec for supercritical Reynolds numbers
assuming a soft bifurcation. We refer to the literature for a discussion of the hard
subcritical bifurcation with quintic nonlinearity (Stuart 1971). Intriguingly, even
turbulent flows with dominant periodic coherent structures may be described by (17.8)
(Bourgeois et al. 2013).
The mean-field model and variants thereof have been successfully used for the
model-based stabilization of the cylinder wake at Re = 100 (Gerhard et al. 2003,
Tadmor et al. 2011). In these studies, an energy-based control design as discussed in
Section 17.3.1 has been used to prescribe a fixed decay rate of the model.

h d i 0 0 9 8 088962 4 024 bli h d li b C b id i i

 358 Data-Driven Fluid Mechanics Chapter 17

17.3.3 Moderately Nonlinear Dynamics

Some oscillatory flows may be tamed by direct mitigation with models and methods
described in Sections 17.3.1 and 17.3.2. Not all plants, particularly turbulent flows,
have the actuation authority for such a stabilization. However, periodic forcing at high
frequency has mitigated periodic oscillations in a number of experiments. Examples
are the wake stabilization with a oscillatory cylinder rotation (Thiria et al. 2006),
suppression of Kelvin–Helmholtz vortices in transitional shear layers (Parezanovic
et al. 2016), reduction of the separation zone in a high-lift configuration (Luchtenburg
et al. 2009), and elongation of the dead-water region behind a backward facing step
(Vukasonivic et al. 2010). Also a periodic frequency at 60–70% of the dominant
shedding frequency may substantially delay the vortex formation in wall-bounded
shear-layers1 and D-shaped cylinders (Pastoor et al. 2008). Sections 17.3.1 and 17.3.2
outline a corresponding modeling and control strategy.
Here, a generalized mean-field model for such frequency cross-talk phenomena is
reviewed from Luchtenburg et al. (2009). Let u u denote the natural self-amplified
oscillation represented by two oscillatory modes u1 , u2 . Analogously, the actuated
oscillatory fluctuation u a is described by two modes u3 , u4 . The base-flow defor-
mation due to the unstable natural frequency ωu and stable actuation frequency ω a
are described by the shift-modes u5 and u6 , respectively. The resulting velocity
decomposition reads
u(x, t) = u s (x) + u ∆ (x, t) + u u (x, t) + u a (x, t), (17.10a)
u (x, t) = a1 (t) u1 (x) + a2 (t) u2 (x),
u
(17.10b)
u (x, t) = a3 (t) u3 (x) + a4 (t) u4 (x),
a
(17.10c)
u ∆ (x, t) = a5 (t) u5 (x) + a6 (t) u6 (x). (17.10d)
Generalized mean-field arguments yield the following evolution equation:
 a1   a1 
d  a2   a2 
   
= , a  + B b, (17.11a)
 
A(a 5 6 ) 
dt  a3   a3 

 a   a 
 4   4 
A(a5, a6 ) = A0 + a5 A5 + a6 A6, (17.11b)
   
a5 = αu a12 + a22 , a6 = α a a32 + a42 , (17.11c)

 σu −ωu 0 0   −βuu −γ uu 0 0 
ωu σu 0 0 γ uu 0 0
   
−βuu
where A0 =   , A5 =  ,
   
 0 0 σa −ω a 


 0 0 −β au −γ au 


 0 0 ωu σu 


 0 0 γ au −β au 

 0   −βua −γ ua 0 0 
0 γ ua 0 0
   
−βua
=  , A6 =  .
   
B
 0 
  0 0 −β aa −γ aa 


 g 


 0 0 γ aa −β aa 

1 J.-L. Aider, Private communication.

h d i 0 0 9 8 088962 4 024 bli h d li b C b id i i

 Noack et al: Machine Learning for Turbulence Control 359

It should be noted that a vanishing actuation implies a3 = a4 = a6 = 0 and yields the

mean-field model of (17.8) modulo index numbering. In the sequel, we assume that
the fixed point is unstable (σ u > 0) and the limit cycle is stable (βuu > 0). Similarly,
a vanishing natural fluctuation a1 = a2 = a5 yields another mean-field model, again
modulo index numbering. In the sequel, we assume that the actuated structures vanish
after the end of actuation, implying σ aa < 0 and β aa > 0 in the model.
An interesting aspect of (17.11) is the frequency cross-talk. The effective growth
rate for the natural oscillation reads
   
A11 = σ u − βuu αu a12 + a22 − βua α a a32 + a42 . (17.12)

The forcing stabilizes the natural instability if and only if βua > 0. Complete
stabilization implies A11 ≤ 0. From (17.12), such complete stabilization is achieved
with a threshold fluctuation level at the forcing frequency
σu
a32 + a42 ≥ .
α a βua
Thus, increasing the forcing at higher or lower frequency can decrease the natural
frequency.

a) Natural wake with vortex shedding b) Actuated partially stabilized wake

Figure 17.1 Flow visualization of the experimental wake behind a D-shaped body without (a)
and with symmetric low-frequency actuation (b), reproduced with permission from Pastoor
(2008). The D-shaped body is indicated in gray; the red squares mark the location of the
pressure sensors, and the blue arrows indicated the employed ZNMF actuators.

The generalized mean-field model has been fitted to numerical URANS simulation
data of a high-lift configuration (Luchtenburg et al. 2009) with high-frequency forcing.
This model also accurately describes the experimental turbulent wake data with a
stabilizing low-frequency forcing (Aleksic et al. 2010) as shown in Figure 17.1. The
model the generalized mean-field model as expressed by (17.10) and (17.11) may
guide in-time control (Duriez et al. 2017) and adaptive control design, providing the
minimum effective actuation energy (Luchtenburg et al. 2010). In principle, (17.11)
can be generalized for an arbitrary number of frequencies.

17.4 Model-Free Machine Learning Control

Model-based control has success stories in turbulence experiments for first- and
second-order dynamics as outlined in the previous section. Control-oriented models

h d i 0 0 9 8 088962 4 024 bli h d li b C b id i i

 360 Data-Driven Fluid Mechanics Chapter 17

for broadband turbulence are quite a challenge. Hence, most experimental turbulence
control approaches are model-free, albeit typically restricted to tuning one or few
actuation parameters. This section outlines the recently discovered MLC approaches,
which solve a much more general problem: the optimization of general nonlinear
multiple-input multiple-output (MIMO) control laws in an automated manner in
the full plant without a model. First (Section 17.4.1), cluster-based control (CBC)
is described allowing the rapid learning of simple smooth nonlinear control laws.
Second (Section 17.4.2), linear genetic programming control is presented, allowing
to optimize rather complex control laws with more testing.

17.4.1 Cluster-Based Control

Cluster-based control (Nair et al. 2019) is based on a low-dimensional empirical
parameterization of MIMO control laws. The parameters are optimized with a
gradient-based search. Let b comprise Nb actuation commands. Let Ns instantaneous
sensor signals represent the output s. The goal is to minimize a cost J, for example,
drag or noise, with associated actuation penalty. The corresponding regression
problem for the sensor-based control law
b = K (s) (17.13)
reads
K ? = arg min J(K). (17.14)
K

Typically, the optimization (17.14) leads to a practically unsolvable variational

problem in experiment and in simulations because of limited testing and simulation
time.
The trick is to employ sensor data s m , m = 1, . . . , M for unforced and several forced
flows. The sensor data is clustered to K centroids ck , k = 1, . . . , K. The actuation
vector at each centroid, bk = K(ck ), k = 1, . . . , K, is considered as tunable parameter.
The CBC law (17.13) uses the actuation parameters at the centroids to interpolate
between them, for example, with K-nearest neighbors. The optimization of this CBC
law is performed with a robust downhill simplex method for the K × Nb actuation
parameters. The centroids may be improved with added sensor data from the CBC law.
Nair et al. (2019) employs this approach for optimizing stall prevention over
an actuated airfoil with large-eddy simulations in few dozen runs. Several other
CBC variants have been proposed from brute-force testing (Kaiser et al. 2017b) to
model-based control (Fernex et al. 2019). No experimental applications have been
reported yet.

17.4.2 Linear Genetic Programming Control

Another approach to solving the regression problem (17.14) is facilitated by an
evolutionary algorithm, more precisely linear genetic programming (LGP). In contrast
to CBC, no advance data are assumed, and the control laws can, in principle, be rather
complex.

h d i 0 0 9 8 088962 4 024 bli h d li b C b id i i

 Noack et al: Machine Learning for Turbulence Control 361

Figure 1 of Li et al. (2018) illustrates the approach. A new MIMO control law
(17.13) is tested for a given time in a fast control loop. The typical timescale from
sensing to actuation is of the order of milliseconds in experiments. The performance
J of the control is optimized in a slow outer loop. In aerodynamic experiments, the
testing of one control law typically takes 5 to 10 seconds. After a few hundred up to
one thousand control tests, the optimization is typically converged for O(10) sensors
and up to O(10) actuators.
The optimization is based on LGP starting with the first generation of Ni = 100
random control laws Ki1 , i = 1, . . . , Ni . All these laws are tested in the experiment
Ji1 and sorted according to performance J11 ≤ J21 ≤ · · · ≤ JN 1 . Then, the N = 1
i
e
1
best control law K1 is directly copied into the new generation. The other laws are
obtained from genetic operations that preserve memory (replication), tend to breed
better laws (cross-over), and explore new minima (mutation). The new generation is
again tested and sorted, The iteration stops after a prescribed convergence criterion is
reached. Typically up to Ng = 10 generations are required. We refer to the exquisite
detailed description of LGP in Wahde (2008) and for control applications to Duriez
et al. (2017).

Figure 17.2 LGPC learning curve of jet mixing optimization. For details see Zhou et al.
(2020).

h d i 0 0 9 8 088962 4 024 bli h d li b C b id i i

 362 Data-Driven Fluid Mechanics Chapter 17

Figure 17.2 shows the learning of a distributed actuation at a nozzle exit for jet
mixing optimization. In 11 generations, the automated LGPC learning has discovered
axisymmetric forcing, helical forging, and flapping forcing before discovering a much
better performing hitherto unknown combined actuation. For details, the reader is
referred to Zhou et al. (2020).

17.5 Tutorial: xMLC for Nonlinear Dynamics

In this section, MLC is exemplified for a dynamical system using the xMLC software.
The tutorial begins with a short description of the generalized mean-field model
(GMFM), a benchmark dynamical system of Duriez et al. (2017). Then, the download
and the installation of the xMLC software are explained. xMLC capabilities and
functions are illustrated by control design for the GMFM. Section 17.4 gives advice
for user-defined problems.

17.5.1 The Generalized Mean-Field Model: A Nonlinear Benchmark Problem

The GMFM is a four-dimensional mean-field Galerkin model that describes the non-
linear interaction between two oscillatory fluctuations (see Section 17.3.3 and Chapter
5 of Duriez et al. (2017)). Here, for simplicity, specific values for all coefficients are
assumed. It consists of two oscillators, (a1, a2 ) (unstable) and (a3, a4 ) (stable), coupled
by a nonlinear growth rate σ of the (a1, a2 ) oscillator. The control objective is to stabi-
lize the first oscillator. This can only be achieved via excitation of the second oscillator.
Let a = [a1, a2, a3, a4 ]T ∈ R4 be the state variable and b ∈ R the actuation command.
The controlled GMFM is described by (17.1) with b1 ≡ 0, b2 = b(a1, . . . , a4 ) and
b3 ≡ 0. A full-state control b = K(a1, a2, a3, a4 ) is performed by adding a forcing
term b in the fourth equation of the system. The cost function J comprises the mean
distance to the fixed point Ja and a penalization term for the actuation Jb :

J = Ja + γJb ,
1
∫ Tmax
Ja = (a12 + a22 ) dt, (17.15)
Tmax
1
∫0Tmax
Jb = Tmax 0
b2 dt.
The penalization parameter γ is taken to be 0.01. The cost function is computed
for Tmax = 100T1 where T1 = 2π √is the period of the first oscillator. The initial
√ is set to [a1, a2, a3, a4 ] = [ 0.1, 0, 0, 0] , that is, on the unforced
condition limit cycle
T T

(radius 0.1) and the second oscillator is at the fixed point [0, 0]T . The dynamics of
the GMFM without control is illustrated in Figure 17.3.

17.5.2 Installation
In this section, the necessary steps to install the xMLC software are presented.

h d i 0 0 9 8 088962 4 024 bli h d li b C b id i i

 Noack et al: Machine Learning for Turbulence Control 363

1 1

0 0

–1 –1
0 50T 100T 0 50T 100T

0.5 0.4

r
1
0.2

0 0

–0.2
–r 1

–0.5 –0.4
–0.5 –r 1 0 r1 0.5 –0.4 –0.2 0 0.2

Figure 17.3 Unforced dynamics of the GMFM: actuation command b (top left), growth rate σ
(top right), phase space for the unstable oscillator (bottom left), and phase space for stable
oscillator (bottom right). The last period (T1 ) is depicted in red. The circle (◦) and the dot (•)
represent, respectively, the initial condition and the final state for the oscillators. For the non
controlled system, they overlap. This figure has been plotted with MATLAB with 100 points
per period and 100 periods.

Requirements
The xMLC software has been coded in MATLAB version 9.2.0.556344 (R2017a).
Natick, Massachusetts: The MathWorks Inc., 2017 and Octave version 4.2.2 (2018).
John W. Eaton and others. Thus, any later version should be compatible with the
software. No particular MATLAB package is needed for the functioning of the
software. For Octave, the package communications is needed. It is freely available
on the official webpage https://octave.sourceforge.io. It can also be downloaded
directly from the graphic user interface with the pkg function.

Installation
The xMLC software can be download from the following link:

https://BerndNoack.com/programs/xMLC v0.9.2.tar.gz

Once downloaded, untar the xMLC v0.9.2.tar.gz file and copy the xMLC v0.9.2
folder where it is needed. Installation is then complete.
To assure the compatibility with MATLAB or Octave, some specific files must be
used. Those files are stored in the Compatibility/ folder. To change from MATLAB
to Octave and conversely, execute the adequate bash file MatlabCompat.sh or
OctaveCompat.sh in the Compatibility/ folders. If no Linux system is used, copy
the files from Compatibility/MATLAB/ or Compatibility/Octave/ to the corresponding
folders.

h d i 0 0 9 8 088962 4 024 bli h d li b C b id i i

 364 Data-Driven Fluid Mechanics Chapter 17

For further information about the content of the xMLC v0.9.2 folder please consult
the README.md file.

17.5.3 Execution
To use the xMLC software, launch a MATLAB or Octave session in the xMLC v0.9.2
folder. Then run the Initialization.m script having replaced the files for com-
patibility as outlined in Section 17.5.2. This script loads all the necessary paths and
creates a xMLC class object with the GMFM benchmark problem. This object is stored
under the variable mlc by default. To create a new xMLC class object with the default
parameters, use the following command: MyMLC = MLC;.
The following section describes some of the basic xMLC parameters and functions.
All the commands are compatible with MATLAB and Octave except the saving and
loading functions that are specific for each software.

xMLC parameters
Once the Initialization.m script is launched, a small description of the problem to
be solved is printed. It contains information about the number of inputs (controllers),
the number of outputs (sensors), the population size, and the strategy (genetic
operators probabilities). This is shown each time a new instance of a xMLC class object
is created. To show it again use the show problem method by using the command
mlc.show problem;. The mlc object has four properties:

• population: contains all the information and database references for each genera-
tion of individuals;
• parameters: where all the parameters are defined, be it for the problem, the control
law description or the xMLC parameters;
• table: is the database, it contains all the individuals explored, not all of them are
evaluated thanks to screening options;
• generation: integer referring to the actual generation. It is set to 0 by default when
the population is empty.

You can access and modify xMLC parameters with the following commands:

% To display all the parameters

mlc.parameters % MATLAB only, for Octave see default_parameters.m
% To display parameters related to the problem definition
mlc.parameters.ProblemParameters % MATLAB only
% To display parameters related to the control law definition
mlc.parameters.ControlLaw % MATLAB only
% To display parameters related to the MLC algorithm
mlc.parameters.CrossoverProb
% To modify some MLC parameters
mlc.parameters.Elitism = 0;
mlc.parameters.CrossoverProb = 0.5;

h d i 0 0 9 8 088962 4 024 bli h d li b C b id i i

 Noack et al: Machine Learning for Turbulence Control 365

mlc.parameters.MutationProb = 0.5;
mlc.parameters.ReplicationProb = 0;
mlc.parameters.PopulationSize = 50;
Problem parameters and control law parameters should not be modified before
starting a run; only xMLC parameters should be considered. To change them, see
Section 17.5.4.

My first run
Once the xMLC parameters are appropriately set, the optimization process can be
launched. The evolution of the population of individuals is done with the go method.
A generation can be given as an argument to perform several evolution iterations.
With no arguments one evolution step is performed leading to the evaluation of the
generated individuals. Once the process is over, the xMLC object can be saved to be
used later. All runs are saved in the save runs/ folder with their name and with their
compatibility (MATLAB or Octave). The folders are created automatically.
% To compute the next generation and evaluate the new individuals.
mlc.go;
% To compute the next 5 generations :
mlc.go(5);
% To continue with the following 5 next generations :
mlc.go(10);
% Change the name of my run and save
mlc.parameters.Name = ’MyRun’;
mlc.save_matlab; % for MATLAB
mlc.save_octave; % for Octave
% Load my run
mlc.load_matlab(’MyRun’); % for MATLAB
mlc.load_octave(’MyRun’); % for Octave
for loops in Octave are not as optimized as in MATLAB. Hence, the evaluation of
the individuals may take much more time with Octave as compared to MATLAB. The
reader is invited to optimize the evaluation of the individuals following ones needs by
adjusting the evaluation parameters in mlc.parameters.ProblemParameters or in
the GMFM problem.m file, see Section 17.5.4.

Post-Processing and Analysis

Once the optimization process is done, the best individual can be accessed with
the following method: mlc.best individual;. This command plots information
concerning the best individual. The ID number represents its index in the database. For
the GMFM problem, a visualization of the best control law is plotted, see Figure 17.4.
The first oscillator has been successfully stabilized whereas the second oscillator has
been destabilized. Moreover the fixed point has been reached only after three periods.
The growth√ rate σ oscillates around 0, and the second oscillator is on the limit cycle
of radius 0.1.

h d i 0 0 9 8 088962 4 024 bli h d li b C b id i i

 366 Data-Driven Fluid Mechanics Chapter 17

For more information about the best control law, use the command
mlc.table.individuals(IDN); where IDN is the ID number of the best
individual. Other features can be extracted such as the learning process or the Pareto
front, thanks to the command mlc.convergence;, see Figure 17.5. Figure 17.5(a)
shows that the performance of the best control increases throughout the generations
as well as the overall performance of the population. From first to last generation
the cost function is dropped by one order of magnitude. In Figure 17.5(b) the latest
generations are the ones pushing forward the Pareto front optimizing Ja and Jb even
though the penalization parameter is only γ = 0.01.

1
0
0 –0.2
–0.4
–1 –0.6
0 50T 100T 0 50T 100T

0.5

r1 0.5

0 0

–r 1 –0.5

–0.5
–0.5 –r 1 0 r1 0.5 –0.5 0 0.5

Figure 17.4 Same as Figure 17.3 for the best control law computed after the evolution of 100
individuals through 10 generations.

17.5.4 My MLC Problem

The definition of an MLC problem and its parameters is done by two files:
MyProblem problem.m and MyProblem parameters.m, respectively. An
example is provided in the PlantnMyProblem folder. The problem parameters
and control law parameters should be modified only by creating an appropriate
MyProblem parameters.m file. Those parameters include the penalization
parameter γ (gamma), the actuation limitation, the number of constants, the operators
(+, −, ×, ÷, . . . ), the register initialization, and so on. The user is free to do any
modification in the MyProblem files.

h d i 0 0 9 8 088962 4 024 bli h d li b C b id i i

 Noack et al: Machine Learning for Turbulence Control 367

(a)
(b)
10 -3
10
6
9
5
8

4 7

6
3
5
2
4

1 3

2
0

1
0 0.05 0.1 0.15 0.2 0.25

(a) (b)

Figure 17.5 Two figures obtained thanks to the convergence method for the 100 × 10
GMFM runs. Those plots show the distributions of the individuals following their performance
and cost throughout the generations. Plotted with MATLAB. (a) Cost distribution for each
generation. The color represents how each individual has been created. The red solid line
represents the best individual for each generation, and the green dashed line follows the
median performance. (b) Zoom-in on the Pareto front (black dashed line) for the 1000
individuals. Each dot represents one individual, the color codes the generation.

Any malfunction in the evolution process should be reported to Guy Y. Cornejo

Maceda ([email protected]) with the error message.
A cheat sheet containing more information about other features is also provided in
the xMLC v0.9.2 folder.

h d i 0 0 9 8 088962 4 024 bli h d li b C b id i i

 18 Deep Reinforcement Learning
Applied to Active Flow Control
1
J. Rabault and A. Kuhnle

This chapter reviews the recent applications of deep reinforcement learning (DRL) to
the control of fluid mechanics systems. DRL algorithms attempt to solve complex
nonlinear, high-dimensionality control, and optimization problems using a direct
exploration approach, that is, trial and error. Here, we start with an in-depth review
of current state-of-the-art DRL algorithms, and we provide the background necessary
to understand how these are able to perform learning through direct close-loop
interaction with the system to control. Then, we offer an overview of the applications
of DRL to active flow control (AFC) that can be found in the literature, and we
use this review to draw concrete guidelines for the application of DRL to fluid
mechanics. Finally, we discuss both outstanding challenges and directions for further
improvement in the use of DRL methods for AFC. The chapter is supported by several
examples of DRL application codes, and offers both a theoretical and a practical guide
to new practitioners.

18.1 The Promise of Deep Reinforcement Learning

in Fluid Mechanics

Reinforcement learning (RL) – often referred to as DRL when it is combined with

deep neural networks (DNNs) as function approximators – has seen rapid progress in
recent years. As discussed briefly in Chapter 3, while supervised learning is focused
on learning a function such as a classification or regression task from observed
input–output data points, and unsupervised learning is focused on identifying patterns
without explicit target labels, RL is concerned with optimizing decision-making in
interactive problems for which no solution is known a priori. Instead of target labels,
a reward signal, which may be sparse and/or time-delayed, acts as feedback to reflect
overall performance and to consequently guide the learning process. This implies that
RL training takes place through direct interaction with the system to control in a trial-
and-error, closed-loop fashion.
The general RL framework presents many similarities with the control framework
presented in Chapter 10. However, while traditional proportional integral differential
1 Jean Rabault was funded through the Research Council of Norway, by the Petromaks II project DOFI,
grant number 280625.

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 Deep Reinforcement Learning and Active Flow Control 369

Machine learning

Supervised learning Unsupervised learning Reinforcement learning

data data environment interaction
labels no labels reward signal

Model-free RL Model-based RL
treat environment as blackbox learn model of environment
learn only control policy use model to find policy

Value-based methods Policy-based methods Imitation learning &

Q-learning Policy gradient inverse RL
Bellman equation / Monte Carlo estimate Infer policy and/or reward
temporal difference from examples

DQN Policy-based methods REINFORCE

Actor: policy-based
Critic: value-based
NAF PPO
ACER
... SAC
...
...
Figure 18.1 Overview of the different machine learning approaches related to reinforcement
learning. The focus of this chapter is on model-free RL algorithms or, more specifically, on
value-based / Q-learning algorithms and policy-based / policy gradient methods, which are
discussed in Sections 18.2.2 and 18.2.3, respectively. Popular recent example algorithms like
deep Q-learning (DQN) or proximal policy optimization (PPO) are introduced in the
corresponding section.

(PID) controllers or linear quadratic regulators (LQR) rely on a local analysis of the
system and assume properties such as linearity, RL aims at performing control without
imposing requirements on the underlying system. In particular, the RL algorithms
discussed in this chapter are able to effectively map nonlinear systems through direct
exploration, and to find complex nonlinear control strategies adapted to each system.
As a consequence of its generality, applications of the RL paradigm are many,
ranging from playing a wide range of games, such as Atari, Chess, Go, or Poker
(Silver et al. 2017, 2018), to providing high levels of dexterity to robots (Hwangbo
et al. 2019), to controlling complex industrial systems in order to optimize, for
example, power consumption (Knight 2018). All these applications illustrate that
properly designed RL systems are able to perform a wide range of decision-making
and control tasks involving complex systems. This makes RL methods a promising
candidate for solving AFC problems in the context of fluid mechanics. AFC is well
known to be very difficult due to the many challenges presented by the underlying
Navier–Stokes equations, and the inherent nonlinearity and high dimensionality of
their solutions. In this chapter, AFC strategies under consideration are of a closed-
loop nature, as an observation of the state of the flow is provided to the RL agent.
This chapter provides an overview of RL applications for fluid mechanics problems.
First, we introduce basic concepts of RL and briefly describe the most common classes
of modern RL algorithms. Next, we focus on how RL can be applied to fluid mechanics
and review recent work on this topic. Finally, we discuss best practices and caveats to

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 370 Data-Driven Fluid Mechanics Chapter 18

consider when applying RL to fluid mechanics problems, as well as future prospects

and the potential of RL for the field. The content of this chapter closely follows the
material that was presented at the 2019 “Flow/Interface School on Machine Learning
and Data Driven Methods (Rabault 2019).”

18.2 A Brief Introduction to Deep Reinforcement Learning

18.2.1 Learning through Trial and Error

RL is the area of machine learning concerned with learning optimal behavior through
interaction and trial and error. The standard RL framework conceptualizes this
problem as a sequence of interaction episodes in which an agent repeatedly observes
the state of its environment, decides what action to perform next, and receives a
reward signal as performance feedback. Typically, one distinguishes between model-
based and model-free RL: model-based RL methods incorporate knowledge about
the problem domain to learn a functional model of the environment simultaneously to
optimizing its decision-making, whereas model-free RL considers the environment
to be a black-box and instead focuses entirely on learning to maximize the attained
reward through trial and error. Since these classes are very different in practice,
and since the large majority of deep-learning-based RL methods are model-free, we
will focus exclusively on model-free RL for the purpose of this chapter. Figure 18.1
presents an overview of the different classes of RL methods together with some key
characteristics and popular example algorithms.
More formally, the agent is usually represented as a stochastic decision policy π
that, given an observation of the environment state s, outputs an action distribution
π(a | s), from which a decision is sampled on which action a to take. The general
framework is summarized in Figure 18.2. Similarly, the environment is modeled
as an (unknown) stochastic process W(r, snext | s, a) that provides the likelihood
of next state snext and reward r, given that the system is currently in state s and
action a has been performed. An interaction episode – also referred to as trajectory
or rollout – can thus be represented as a sequence of state–action–reward triples
τ = (s0, a0, r0 ), (s1, a1, r1 ), . . . . In addition to this general framework, a few important

points will be assumed in all of the following:

• Strictly speaking, the environment state s is a discrete, potentially noisy and
incomplete observation of the true internal (and continuous) state of the
environment. In some cases, the state can be captured perfectly as, for instance,
the position of stones in the game of Go, whereas in most cases the observation
comes from noisy sensors, providing partial information about the environment,
like pressure values at a range of sensor locations on the surface of a cylinder
immersed in a flow. The literature nonetheless generally refers to it as the
environment “state” s, instead of a more appropriate term like “observation,”
which we will do here as well.
• The action a may be either a decision from a fixed finite set of choices or a
continuous control signal. Taking the previous examples, the action space in

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 Deep Reinforcement Learning and Active Flow Control 371

Go is a set of positions where the next stone can be placed at each time step,
while an action in the cylinder control case could correspond to the discretized
opening of valves, that is, the mass flow rate ejected, at some actuation slots on
the surface of the cylinder.
• The reward r loosely reflects the performance of the agent and enables the agent to
learn an effective policy from trial-and-error interaction with the environment.
Importantly, the impact of a (sequence of) action(s) may not become apparent
immediately, resulting in delayed and/or sparse rewards, which is why they
cannot be used as independent supervised signal per time step. For instance, in
the previous examples, an agent may only receive a final reward at the end of
a round of Go depending on its success, and mass flow rate control decisions
have longer-term impact on stability and drag of the cylinder system. Therefore,
instead of optimizing the instantaneous reward directly, RL algorithms optimize
the cumulative reward over an episode trajectory τ, R(τ) = t=0 γ · rt ,
Í∞ t
discounted by a factor 0 ≤ γ ≤ 1 (usually γ ∈ [0.95, 0.999]). On the one hand,
the discount factor encourages shorter-term over longer-term gains while on the
other hand, it ensures that the cumulative reward converges in the case of very
long or unbounded episodes.
• The following reward-related value abstractions are used by different algorithms to
make sense of the environment and its reward structure, and to provide guidance
for the learning process:
1. the state-value V π (s) = Eτ∼π [R(τ) | s0 = s] represents the “expected value”
of being in state s, that is, the expected return obtained through a sequence of
interactions following the policy π when starting in environment state s;
2. the state-action- or Q-value2 Q π (s, a) = Eτ∼π [R(τ) | s0 = s, a0 = a] that,
similarly, captures the “expected value” of taking action a in state s, that is, the
expected value of taking action a when starting in state s and continuing to act
according to π afterward;
3. the advantage Aπ (s, a) = Q π (s, a) − V π (s) that indicates how much “better”
(in terms of expected value) it is to take action a when in state s, compared to
the expected return when randomly sampling an action from π.
RL is not a new field, and the framework represented in Figure 18.2 can be
traced back to at least the 1980s. However, the learning algorithms used for training,
and in particular the function approximators being trained, have undergone rapid
development since the popularization of deep learning. The key idea behind DRL
is to leverage the universal approximator properties of DNNs (Hornik et al. 1989),
together with effective training methods such as stochastic gradient descent and
backpropagation of errors (see Chapter 3), in order to make RL algorithms able to
cope with complex systems. In the following sections, we will introduce the two main
classes of RL algorithms – Q-learning and policy gradient methods – together with
recent examples of DRL algorithms for each class.

2 Q originally comes from “quality”, but nowadays is just known as Q-value and not used as an
abbreviation anymore.

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 372 Data-Driven Fluid Mechanics Chapter 18

Figure 18.2 Generic reinforcement learning setup. Iteratively, an agent reacts to the observed
environment state with an action, which in turn influences how the environment state evolves.
Moreover, the agent receives a reward signal, which it aims to optimize over the course of
multiple training episodes, that is, rollouts of agent–environment interactions.

18.2.2 Q-Learning
Q-learning comprises a range of RL algorithms, dating back to at least the work
of Watkins (1989), though it mainly relies on the Bellman equation, which was
formulated much earlier (Bellman 1957). A method belonging to this category, Deep
Q-Network (DQN) (Mnih et al. 2015), was among the early breakthroughs of DRL.
DQN combines DNNs and Q-learning, and its success in learning to play Atari games
solely from vision and game score feedback triggered a renewal of interest in DRL.
This success is made possible by the combination of RL for learning through trial and
error, with DNNs for their ability to extract features and identify patterns in data.
Q-learning is a value-based method, meaning that the algorithm focuses on
learning a (state-action) value function, Q(s, a), which represents the overall reward
obtained by taking action a in state s. Given this Q-function, Q-learning defines the
resulting policy as π(s) = argmaxa Q(s, a). Note that this definition crucially relies
on the fact that the action space is finite, as otherwise the computation of argmax
is not straightforwardly possible. The Q-value function is optimized via temporal
difference (TD) learning in which its own estimates are used as optimization targets.
More precisely, the TD target corresponding to an experience time step st , at , rt , st+1
can be recursively estimated as Qtarget = rt + γ · maxa0 Q(st+1, a 0). Note that the value
of the next state is maxa0 Q(st+1, a 0) since the policy always picks the max action. The
Q-function is then updated by incrementally shifting its predictions in the direction of
the TD residual Qtarget − Q, with the update step size being controlled by a learning
rate η > 0 (n indicates the current optimization step):

Q n+1 (st , at ) := Q n (st , at ) + η · rt + γ · max Q n (st+1, a 0) − Q n (st , at ) . (18.1)



0
a

In the context of deep RL, the value function Q is parametrized by an artificial

neural network Q θ with learned weights θ, which is optimized via gradient descent.
By minimizing the quadratic loss function L(Qtarget, Q θ ) = 0.5 · (Qtarget − Q θ (s, a))2

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 Deep Reinforcement Learning and Active Flow Control 373

Figure 18.3 Illustration of the learning process of an RL agent in a grid-world with one
positive-reward goal state G and two negative-reward failure states T. Starting with no prior
knowledge, the agent learns to estimate the value of states more accurately over the course of
training solely via interaction. Initially, the negative reward of the failure states propagates to
neighboring states, so the agent learns to avoid moving there. As a consequence, it
subsequently comes across the goal state whose positive reward propagates back to the starting
point over time, yielding a first successful strategy. Ultimately, the agent learns to trade off
avoiding failure states with quicker strategies of getting to the goal state.

between Qtarget and the Q θ to be optimized, we arrive at the equivalent update rule
for DNN-parametrized Q-functions (note that Qtarget is fixed with respect to the
differentiation):
target
 
θ n+1 = θ n + η · ∇θ L Q θn , Q θ (st , at )
 2
= θ n + η · ∇θ 0.5 · rt + γ · max Q , a 0

Q , a

θ n (s t+1 ) − θ (s t t )
a0
= θ n + η · rt + γ · max Q θn (st+1, a 0) − ∇θ Q θ (st , at ) . (18.2)


0 a

Q-learning is typically illustrated at the example of a grid world, as shown in

Figure 18.3. Given a finite state and action space, tabular Q-learning simply represents
the Q-value function as an array with states on one axis and actions on the other.
This scenario is useful for grasping how the algorithm learns to improve its behavior:
during the process of repeatedly interacting with the world in a series of episodes,
the update in (18.1) iteratively transmits the reward information of a state to adjacent
states, ultimately communicating the signal from goal to start state(s) and thus arriving
at a successful policy. The approach follows the “dynamic programming” pattern to
make the optimization tractable, as is familiar from optimization methods in other
domains.
A fundamental problem with the Q-learning formulation is that the policy derived
from the learned value function as π(s) = argmaxa Q(s, a) is deterministic, and thus
can easily get stuck in local maxima during the training process, even if their corre-
sponding Q-value is only marginally higher. A common solution to this is to introduce
an exploration rate 0 ≤  ≤ 1, which determines how frequently the argmax policy
is ignored and a random action is sampled instead. Q-learning can still learn from
such off-policy data – that is, actions which do not follow the policy – since it can
still extract useful information to update its Q-value estimates. How best to trade off

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 374 Data-Driven Fluid Mechanics Chapter 18

exploration and exploitation is an important and problem-dependent hyperparameter

decision, and it is common to use decay schedules which over time decrease the
initially high exploration rate to leverage the accumulated knowledge. Note that
exploration is only useful during training, and is turned off for deployment.
DQN (Mnih et al. 2015) is considered a breakthrough, being the first Q-learning
algorithm to successfully leverage a DNN as a feature extractor instead of a table.
This allowed researchers to move beyond low-dimensional finite states and to handle
the complex visual state-space of Atari games. DQN introduced a range of technical
improvements, which were required to stabilize the learning optimization, most
notably: (a) an experience replay memory that stores a large amount of experience
time steps and from which random time steps are sampled for an update, thus relaxing
otherwise detrimental temporal correlation between update time steps; and (b) the use
of a target network to compute the TD target in (18.2), which is an infrequently synced
copy of the main network and thus reduces the problem of potentially detrimental
feedback loops in the recursive updates.
Since this milestone achievement, more improvements and novel techniques have
been introduced in follow-up papers: for instance, double DQN (Hasselt et al. 2016)
improves a bias in the original DQN formulation that tends to overestimate Q-
values. Dueling DQN (Wang et al. 2016) factorizes Q-values into state-value V(s)
and advantage A(s, a), which helps the network to learn coherent Q-values per state.
Categorical DQN (Bellemare et al. 2017) extends the Q-learning framework to work
with (quantized) reward/value distributions, which enables the agent to capture the
reward structure and uncertainty more faithfully. Normalized advantage function
(NAF) (Gu et al. 2016) derives a continuous variant of the Q-learning algorithm, to
be able to work with non-categorical action spaces. These are just a few examples
of extensions, and many more can be found in the literature, see for example Recht
(2019).

18.2.3 Policy Gradient Methods

Policy gradient algorithms are another popular class of RL approaches, dating back
to at least the work of Williams (1992). These are policy-based methods, meaning
that they attempt to optimize a parametrized (stochastic) policy πθ directly with
respect to its expected return Eτ∼πθ [R(τ)]. This is in contrast to Q-learning, where
Q is optimized to approximate the state-action-value function and the (deterministic)
policy is inferred via argmaxa Q(s, a). In deep RL, the policy is parametrized by a
DNN (the policy network) that outputs relevant distribution parameters, for instance,
the probability per action for discrete actions, or the mean and standard deviation for
normally distributed continuous actions.
Fundamental to this optimization is the policy gradient theorem, which uses the
“log-derivative trick” to establish the following equality for the derivative of the
expected return:

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 Deep Reinforcement Learning and Active Flow Control 375

∫
∇θ Eτ∼πθ [R(τ)] = R(τ) · ∇θ πθ (τ) dτ

∫
= R(τ) · πθ (τ) · ∇θ log πθ (τ) dτ

= Eτ∼πθ R(τ) · ∇θ log πθ (τ) . (18.3)

 

Using the conditional (in-)dependence between time steps, the fact that the
logarithm of the product probability πθ (τ) can be turned into a sum, and the fact that
the transition probabilities of the environment do not depend on the policy, a simple
Monte Carlo gradient estimator ĝ of this expectation can be derived as:

n
1 Õ
ĝ = R(τi ) · ∇θ log πθ (τi )


·
n i=0
n
"∞ #!
1 Õ Ö
(i) (i) (i) (i) (i)
= · R(τi ) · ∇θ log πθ (at |st ) · W(st+1 |st , at )
n i=0 t=0
n
"∞ ∞
#!
1 Õ Õ
(i) (i)
Õ
(i) (i) (i)
= · R(τi ) · ∇θ log πθ (at |st ) + log W(st+1 |st , at )
n i=0 t=0 t=0
n ∞
!
1 Õ Õ
(i) (i)
= · R(τi ) · ∇θ log πθ (at | st ) , (18.4)
n i=0 t=0

where W(st+1 |st , at ) is the probability that, starting from state st and taking action
at , the environment evolves into state st+1 . To sum it up, policy gradient methods
work by sampling a set of trajectories at each training step, and use the collected
on-policy data to shift the weights of the policy network in order to maximize its
expected return, that is, to increase the likelihood of trajectories with high return and
discourage decisions that have led to low episode returns. Note that, in contrast to Q-
learning, exploration is intrinsic to policy gradient methods since the latter is based on
a stochastic policy. Consequently, balancing exploration and exploitation is part of the
training objective, which means that only after repeated positive feedback the action
distribution narrows to the point where it is virtually deterministic. When deploying
the trained algorithm, sampling is turned off and replaced by deterministically picking
the maximum likelihood action.
The gradient estimator in (18.4) is also known as the REINFORCE algorithm
(Williams 1992). However, it is a high-variance estimate since it scales all time steps
of an episode by the same episode return R(τ). This in turn may impact convergence
speed of the gradient descent optimization. Lower-variance estimators, which instead
weight time steps according to the “reward to go”, that is, the future return Rt (τ) =
R(τ[t, t + 1, . . . ]) following the current timestep t can be derived similarly. Variance
can be further reduced by observing that a state-dependent baseline function b(s) can
be subtracted from the return without affecting the expectation value in (18.3). It can
be shown that b(s) ≈ V π (s) is the optimal choice with respect to variance reduction

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 376 Data-Driven Fluid Mechanics Chapter 18

for a state-dependent baseline. The following thus yields a lower-variance gradient

estimator for the policy gradient objective:
n ∞
1 ÕÕ
ĝ = Rt (τ (i) ) − b(st(i) ) · ∇θ log πθ (at(i) | st(i) ). (18.5)


·
nt i=0 t=0

Algorithms that focus on learning a baseline value function in addition to the actual
policy are also called actor-critic methods. In this case, the value function is typically
a second “critic” DNN besides the first “actor” policy DNN, trained for a regression
objective similarly to Q-learning in (18.2). The purpose of the critic is to modify the
return R(τ) in the original policy gradient equation (18.3) in order to improve algo-
rithm convergence. This can be performed in several ways, for example, by subtracting
a baseline as is done through using Rt (τ) − b(st ) in (18.5), or by replacing the return
altogether by the value output of a network trained through Q-learning, V(st ) ≈ Rt (τ).
A range of extensions and modifications of this approach to policy optimization
have been explored. ACER (Wang et al. 2017) is an off-policy actor-critic algorithm
that uses experience replay similar to DQN (see Section 18.2.2). Soft actor-critic
(SAC) (Haarnoja et al. 2018) incorporates an entropy-based loss component into the
policy optimization to encourage exploration. Deterministic policy gradient (DPG)
(Silver et al. 2014) derives a deterministic variant of the policy gradient theorem.
Trust-region policy optimization (TRPO) (Schulman et al. 2015) is a related class of
algorithms that additionally constrain the update steps to a trust region, preventing
the updated policy distribution from diverging too far from the old policy, thus
limiting the confidence in each single, potentially detrimental update. Proximal policy
optimization (PPO) (Schulman et al. 2017) is inspired by such trust-region methods
and loosely enforces conservative updates by clipping the objective accordingly.

18.2.4 Current Research Directions in Deep Reinforcement Learning

Since its early breakthroughs, DRL has evolved into a mature research field with
many research directions and open problems actively being worked on. The following
paragraphs briefly introduce some interesting new lines of research to improve DRL
algorithms and make them more practically applicable. Importantly, this review
emphasises that DRL is a fast-moving field with lots of yet untapped potential.

Parallelized Data Collection and Distributed Learning

DRL algorithms can be very data-hungry and, different from supervised machine
learning, do not learn from static data sets, but have to collect data via interaction.
Consequently, methods have been developed to parallelize data generation across
100–1000s of “worker” agents, each of which interacts with a copy of the simulated
environment. The collected experience is aggregated in either one central or multiple
distributed parameter servers that perform the actual RL optimization algorithm,
and the updated weights are synchronized with the workers and other servers.
For instance, Asynchronous Advantage Actor-Critics (A3C) (Mnih et al. 2016)
used around 16 workers, Importance-weighted actor-learner architectures (IMPALA)

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 Deep Reinforcement Learning and Active Flow Control 377

(Espeholt et al. 2018) pushed the number to around 200, and the evolution-strategy-
based system of Salimans et al. (2017) leveraged more than 1 000 cores in parallel.
Such a parallelization approach is particularly interesting for the application of DRL
to flow control because of its time-consuming simulations, as illustrated by Rabault
and Kuhnle (2019).

Inverse Reinforcement Learning

Many environments cannot easily be parallelized, particularly if they are not
simulation-based. In such cases it can be infeasible to collect enough interaction
data to train an RL agent from scratch. Inverse reinforcement learning comprises
techniques to instead make use of existing data that were not produced by the agent.
By inferring a suitable reward function from such trajectories, the agent is provided
a learning signal to bootstrap its own policy without the need for trial and error.
Two popular approaches are generative adversarial imitation learning (GAIL) (Ho &
Ermon 2016) and adverserial inverse reinforcement learning (AIRL) (Fu et al. 2018).

Model-Based Influences
The successes of DRL have mostly used model-free RL algorithms, which attempt
to solve a task solely based on interaction and reward. However, trying to understand
how the environment works can provide an additional learning signal that ultimately
also benefits the agent’s task-solving ability, and there has been work to augment
model-free approaches with model-based features. For instance, Jaderberg et al.
(2017) used auxiliary tasks such as next-frame prediction to encourage the network
to learn representations that capture basic world dynamics. More recently, Ha and
Schmidhuber (2018) presented an approach in which a full world model is learned
from (random) interaction data and subsequently used by the agent for training, instead
of interacting with the real environment.

Exploration and Curiosity

RL agents trained from scratch may either never observe any useful reward signal due
to sparse rewards, or get stuck in local optima early on in the learning process and
exploit this suboptimal solution instead of continuing to search for better strategies.
While random exploration is sufficient for simple problems, bigger and more complex
environments require more principled exploration approaches. A range of recent work
has attempted to incorporate a sense of “intrinsic curiosity” into the agent’s objective,
by comparing the agent’s expectation with the actual outcome and rewarding surprisal
(Pathak et al. 2017, Savinov et al. 2018, Burda et al. 2019). This encourages the agent
to explore uncertain areas of the problem space until they are sufficiently explored.
In addition, exploration problems can be alleviated by incrementally making the task
more difficult, for example, by starting episodes from a point closer to obtaining a
reward than the actual initial state (Salimans & Chen 2018).

From Simulation to Real World

Simulations address the problem of DRL methods being data-hungry by providing
a safe and infinite source of training data. However, the learned behavioral policy

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 378 Data-Driven Fluid Mechanics Chapter 18

ultimately has to be transferred from an imperfect simulation to the real-world

application. Some recent work has focused on techniques to bridge the simulation
reality gap, as in Hwangbo et al. (2019).

18.3 Recent Applications of Deep Reinforcement Learning to Fluid

Mechanics Problems

In this section, we focus on recent applications of DRL in the fluid mechanics

literature. DRL can be applied to several classes of problems in fluid mechanics. First,
it can be used to find high-level control strategies of complex systems, in a similar
way to how robots are controlled. This class of applications includes, for example,
swimming of fish schoolings, which we will review first. Second, we will discuss
applications of DRL to a more detailed control of flow instabilities and detachment
at small scales. Several such applications have been presented recently, such as the
control of vortex shedding control or of Rayleigh–Benard instabilities.
In addition to these applications, there is a range of work that in other ways
relates to fluid mechanics problems, or to the interaction between fluid mechanics and
other engineering challenges, such as controlling unmanned aerial vehicles (UAVs)
(Bohn et al. 2019), or optimizing the trajectory of gliders taking ascendences (Reddy
et al. 2016). However, these are out of scope of the focus of this chapter, and we refer
the reader directly to these papers for further information. Finally, another application
of DRL within fluid mechanics is shape optimization (Yan et al. 2019, Viquerat
et al. 2019). This last category is, so far, less mature than the use of DRL for control,
in particular since shape optimization problems are less naturally expressed as a
sequence of interactions with an environment. Therefore, we will not discuss this
application in more detail either, and the reader is again referred to the most recent
publications on the topic (Garnier et al. 2019, Viquerat et al. 2019, Yan et al. 2019).

18.3.1 Swimming and Locomotion

To the best of our knowledge, the first applications of DRL in fluid mechanics have
been to study the collective behavior of swimmers. This is due to the fact that such
behavior is relatively easy to simulate, especially if reduced-order models are used
(Gazzola et al. 2014, 2016), while the complexity of the system composed of several
swimmers makes it a challenging setup to investigate with other methods than via
empirical simulation. In this context, DRL made it possible to both understand and
reproduce a variety of schooling behaviors. In such studies, the environment state is
a representation of the flow and the relative effect of swimmers on each other, while
the actions describe the swimming effort of individual fishes, and reward is based on
either attaining maximum speed or reducing energy consumption.
The most recent studies in this field solve the full Navier–Stokes equations, either
in two or three dimensions (Novati et al. 2017, Verma et al. 2018). This is particularly
impressive, given both the computational cost of such simulations, and the large state

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 Deep Reinforcement Learning and Active Flow Control 379

and action spaces associated with multiple 3D-simulated fishes swimming together as
a group. In these studies, DRL is able to discover strategies that provide clear energy
benefits compared to fishes swimming individually, and complex strategies are found
that allow follower fishes to effectively extract energy from the vortex shedding alley
behind leading fishes. Such results open the way to experimental understanding of
fish schoolings and associated energy gains, as well as to applications in robotics. In a
similar fashion, applications to gliding have been proposed (Novati et al. 2019). This
category of applications can be seen as an effective way of finding optimal locomotion
strategies for active swimmers or gliders, in a direct analogy to the use of DRL for
legged robots locomotion.
These examples illustrate how DRL is used as a tool to experimentally study
complex phenomena where a purely theoretical approach faces challenges, since
describing a school of swimming fish is analytically intractable, and other experi-
mental or numerical approaches do not cope well with the high dimensionality and
analytical intractability of the problem.

18.3.2 Feedback Active Flow Control

The second category of fluid mechanics applications where DRL has attracted a lot
of attention recently is feedback flow control (Rabault et al. 2020, Ren et al. 2020).
Similar to the examples of the previous paragraph, feedback flow control is a natural
fit for the closed-loop framework of DRL. Properties of the flow such as pressure or
velocity maps make up the environment state, while the actions represent some form
of forcing on the flow. Depending on the use case, the objective defined by the reward
is either to reduce drag, or to damp unstable waves or modes.
The first application of DRL for such problems is arguably the control of the
Karman vortex street presented in Rabault et al. (2019).3 While this problem has
been considered at several occasions using traditional control theory techniques
(Roussopoulos 1993, Min & Choi 1999, Thiria et al. 2006), the wake dynamics
are increasingly challenging to control as the Reynolds number – that is, flow
nonlinearity – is increased. Here, DRL was used to control the separation and vortex
shedding behind a 2D cylinder, as illustrated in Figure 18.4, and is able to reduce
drag significantly. While the general situation is similar to what was presented in
Chapter 10, the DRL approach is able to control the wake also in cases when a linear
approximation does not suffice to capture the dynamics of the system. In a series of
follow-up experiments, speeding up the learning process via parallelization as well
as choosing an optimal reward function and DRL hyperparameters were investigated
(Rabault & Kuhnle 2019).4 The results in particular emphasized the sensitivity of

2 The full code is released open-source, and the results can be reproduced following the instructions
available at https://github.com/jerabaul29/Cylinder2DFlowControlDRL.
3 The code is available here: https://github.com/jerabaul29/Cylinder2DFlowControlDRL
4 The code is available here:
https://github.com/jerabaul29/Cylinder2DFlowControlDRLParallel

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 380 Data-Driven Fluid Mechanics Chapter 18

Figure 18.4 Illustration of active flow control of the von Karman alley by a deep
reinforcement learning Agent (PPO algorithm). This figure is a slightly edited reproduction of
the results presented in Rabault et al. (2019), reproduced with permission. The top panel shows
a representative snapshot of the 2D flow around a cylinder, in a setting similar to the
benchmark of Schäfer et al. (1996). The bottom panel shows the typical flow field obtained
when a trained DRL agent is used to control the vortex shedding to reduce drag. The black
dots indicate the location of a series of pressure probes, which form the state observation. The
red dots at the top and bottom of the cylinder indicate the position of two small synthetic jets,
whose mass flow rates are decided by the actions of the DRL agent. Here, the reward function
encourages to minimize drag while keeping lift fluctuation low. A drag reduction of about 8%
is obtained, close to the optimal drag reduction value predicted by the symmetric baseline flow
analysis of Bergmann et al. (2005).

the resulting strategy to the exact choice of reward function, and more generally the
importance of using physical insights when defining the reward.
Having established the potential of DRL for AFC, several questions remained to
enable the application of DRL to more realistic and sophisticated systems. First, DRL
needs to be able to cope with the large action space dimensionality of real-world
systems. This is challenging since the size of the control space to explore is much
larger when the number of actuation signals is increased (typically, the naive direct
exploration cost is a power function of the action space dimensionality, as it scales
with the combinatorial cost of exploring the effect of each combination of actuations).
Fortunately, most realistic systems exhibit invariants of some form, which can be
exploited to reuse the same policy across different parts of the system and, conse-
quently, reduce exploration and learning costs. This was demonstrated empirically
through the analysis of thin fluid film instability control (Belus et al. 2019),5 and
will be discussed in more details in Section 18.4.3, where we provide guidelines for
practitioners.
Second, robustness of the final control strategy is critical for real-world applications
employing DRL. Here again, recent results have been promising. Several robust
applications of DRL, even in complex environments, have been presented by the

5 The code is available here: https://github.com/vbelus/falling-liquid-film-drl

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 Deep Reinforcement Learning and Active Flow Control 381

Figure 18.5 Demonstration of the robust closed-loop AFC strategy found by a PPO agent, as
described in the work of Tang et al. (2020) (the figure is a slightly edited version of the results
presented there, used with the permission of AIP Publishing). A single PPO agent is trained on
simulations similar to the ones presented in Figure 18.4. However, in the present case, the
Reynolds number used for each episode is chosen at random, in the discrete set {100, 200,
300, 400}. After training, the PPO agent is found both to exhibit highly effective control
(comparable with the optimal control expected from the symmetric flow analysis suggested in
Bergmann et al. (2005)), and to be robust enough to control the flow at any Reynolds number
in the range [60, 400], including values that had not been seen during training. This illustrates
both the robustness and generalization ability of AFC strategies discovered through DRL.

ML community, for instance, for playing games or for robotic control (Hwangbo
et al. 2019). In the fluid mechanics community, the first investigation of the ability of
DRL to produce robust control strategies was presented by Tang et al. (2020).6 In this
work, the DRL agent was trained on a range of Reynolds numbers rather than a single
fixed number, as was done in previous studies. This means that the agent is trained on
a range of flow conditions that, while broadly exhibiting the same patterns, still differ
in the precise behavior and responsiveness of the wake. Experiments showed that the
DRL agent is able to learn a robust global control strategy, and to perform control over
a wide range of Reynolds numbers, including ones not seen during training. The main
findings of this work are visualized in Figure 18.5. Such findings are very promising
for applications in the real world, where one would expect a similar variation of flow
properties.

6 The code is available here: https://github.com/thw1021/Cylinder2DFlowControlGeneral

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 382 Data-Driven Fluid Mechanics Chapter 18

The examples presented so far still work with relatively low Reynolds numbers, that
is, moderate nonlinearity and complexity, owing to the computational cost associated
with increasing the number. More critically, all the AFC examples have been in
the laminar (though, chaotic) regime. Therefore, a final milestone to achieve in
order to truly qualify DRL as a possible AFC method for realistic applications is to
show successful application in turbulent flow conditions. An important first step in
this direction was recently achieved in 2D simulations of the Karman wake behind
a cylinder, at an intermediate Reynolds number of 1000 (Ren et al. 2020). The
setup is very similar to the one introduced by (Rabault et al. 2019), except for the
flow regime that is chaotic due to a much higher Reynolds number. The drag in
these conditions can be reduced by up to around 30% on average, and a transient
drag reduction of up to around 50% can be observed, which clearly demonstrates
the potential of DRL control even in the case of more complex, higher Reynolds
number flows.
In addition to the application of DRL for controlling instabilities in the wake
of a cylinder or at the surface of thin water films, at least three other applications
were presented recently. The first is concerned with the control of the Kuramoto–
Sivashinsky (KS) equation (Bucci et al. 2019), which is an example of a system
exhibiting spatiotemporal chaos. The authors find that a simple Q-learning method is
able to capture the dynamics of the system and control its behavior with only limited
actuation. The second example is focused on controlling spatio-temporal structures
happening during Rayleigh–Benard convection (Beintema et al. 2020). Again, the
authors find that DRL is able to drastically reduce the development of instabilities
in the system by controlling the distribution of heating power at the lower side of
a convection cell. The strategy obtained through DRL is compared with classical
control algorithms and found to significantly outperform them. Finally, Xu et al.
(2020) studied the control of the wake behind a cylinder using small contra-rotative
cylinders.
The work of Beintema et al. (2020) clearly illustrates the limitations of traditional
“optimal control” algorithms. While these methods are very attractive both from a
theoretical and mathematical point of view, in practice they rely on strong assumptions
about the underlying system for both convergence and stability. These assumptions
– for instance, some form of Lipschitz-continuity, convexity, or the existence of
extrema/poles – are often not satisfied by real-world, nonlinear, high-dimensional
systems. In such cases, DRL, which is based on learning to control through trial
and error and leverages the approximation capabilities of deep learning, is not only
more robust but it may be the only feasible approach to control systems that exhibit
strong nonlinearities. Indeed, if a system is highly nonlinear, any method based on lin-
earization or local-order reduction techniques will always have a very limited domain
in the phase space within which it is valid. Consequently, one needs to explore the
space and exploit the gathered knowledge, which is precisely the learning process for-
malized by the DRL framework, while relying on DNNs as general-purpose function
approximators.

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 Deep Reinforcement Learning and Active Flow Control 383

18.4 A Guide to Deploying Deep Reinforcement Learning

in Fluid Mechanics

In this section, we present a range of technical points to keep in mind when designing
a new application of DRL in fluid mechanics. This section is a synthesis of beneficial
practices found in papers referred to in the Sections 18.2 and 18.3. While the
discussion is focused on fluid mechanics problems, the techniques discussed here are
generally applicable to many DRL applications.

18.4.1 Problem Identification and Coupling to Deep Reinforcement

Learning Framework
The first critical requirement before attempting to leverage DRL in an application is
to make sure that the problem and its characteristics fit the DRL framework. First, the
problem needs to be expressible as a closed-loop interaction, as described in Section
18.2.1. Second, DRL is of particular interest since it imposes relatively few assump-
tions on the underlying system. so it is especially suitable to tackling problems with
high nonlinearity and few mathematical properties such as convexity or Lipschitz-
continuity, which would be required for analytical tools and traditional applied math-
ematical methods. This applies to most non-simplistic fluid mechanics applications.
Once these high-level requirements are met, the problem needs to be wrapped up
as an interactable “environment.” This is usually done by implementing a small set
of predefined interface methods. Code examples here are based on the applied DRL
library Tensorforce (Kuhnle et al. 2017) but similar patterns can be found in other DRL
frameworks as well. To use the environment interface of Tensorforce, the user needs
to provide a specification of the state and action space (states() and actions()),
a method to start a new interaction episode (reset()) and a function that, given the
agent actions, advances the environment by one time step (execute(...)):
# Inherit from the Environment base class
class MyEnvironment ( Environment ) :

def states ( self ) :

""" State space specification """
return dict (...)

def actions ( self ) :

""" Action space specification """
return dict (...)

def reset ( self ) :

""" Resets the environment to start a new episode """
# ... prepare environment ...
return initial_state

def execute ( self , actions ) :

""" Executes the given action ( s ) and advances the
,→ environment """
# ... advance environment ...
return next_state , is_terminal , reward

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 384 Data-Driven Fluid Mechanics Chapter 18

The flow of interaction between the agent and the environment closely follows that
of Figure 18.2. First, the environment is prepared for a new episode of interaction with
the agent and returns the initial state (reset()). Starting from here, alternately, the
agent reacts to the observed environment state, and the environment is advanced by
one time step based on the agent’s decision (execute(...)). In all this, the state
and action space specify the communication format, that is, what observations of
the environment look like (for instance, image size or number of sensor values) and
what type of actions the environment expects in return from the agent (for instance,
discrete or continuous decisions). Finally, the interaction either terminates once a
failure or success state is reached, or is aborted after a predefined number of time
steps, particularly if there are no natural terminal states (as is often the case for flow
control problems). Subsequently, a new training episode is initiated. This is illustrated
by the following code snippet:
# Train for 200 episodes
for _ in range (200) :

# Initialize environment
states = environment . reset ()
terminal = False

# Episode of agent - environment interaction

while not terminal :
actions = agent . act ( states )
states , terminal , reward = environment . execute (
,→ actions )
agent . observe ( terminal , reward )

While it may seem more flexible to reimplement a DRL algorithm from scratch
instead of having to interface with an existing framework, we discourage fluid
mechanics practitioners, particularly if they are new to DRL, from doing so, and highly
recommend relying on well-established open-source implementations such as Tensor-
force, Stable-Baselines or RLlib instead. The main reason is that it is a challenging
task to make sure that a DRL implementation actually works correctly, and debugging
it if it does not. This is due to a combination of factors, including their intrinsic non-
determinism on various levels (action sampling, random weight initialization, etc) and
the fact that internal representations are non-interpretable. Moreover, one frequently
finds that DRL algorithms learn to “compensate” for implementation errors, meaning
that the agent may be able to optimize the reward function reasonably well even in the
context of critical bugs since, to the learning algorithm, an erroneous implementation
is just yet another complex optimization problem to solve.

18.4.2 Problem Modeling: Invariances, Reward Shaping, and Parallelization

When modeling the problem environment, important practical considerations need to
be taken. While they may seem obvious at first, challenges when deploying DRL, in
the experience of the authors, often stem from inconsistent or lacking attention to these
points.

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 Deep Reinforcement Learning and Active Flow Control 385

First, one needs to make sure that the choice of the state input to the agent is
consistent with the problem to solve, and provides a sufficient amount of information
for the agent to identify potential causality relations between the state of the system,
the action taken, the evolution in time of the system, and the reward observed. At this
stage, physical understanding of the controlled system is very beneficial. In addition,
the general setup of the control must take advantage of the structure of the problem in
order to perform learning effectively. In particular, properties of invariance by transla-
tion, rotation, or any other form of symmetry should be exploited. Rabault et al. (2019)
and Belus et al. (2019) investigated several possible methodologies for exploiting sys-
tem invariants when designing DRL applications. Using a simple example application
(controlling the development of instabilities in a thin fluid film over an extensive
region in space using multiple jets), they found that failure to take into account
invariants can lead to prohibitive exploration costs which make learning impossible
when several control signals are needed. In contrast, successful exploitation of these
invariants made it possible to control an arbitrary number of actuators. The best DRL
deployment technique, shown in Figure 18.6, relies on applying the same DRL agent
to control different parts of the system, which are modeled as separate environments
but reuse the same agent DNN weights – and thus the distilled knowledge about the
system. The same approach can be adopted to enforce other 2D, 3D, or axis-symmetric
policy invariants, depending on the problem under consideration. A challenge with
this approach is to formulate a suitable reward function. The reward function of each
environment should, on the one hand, be as “local” as possible, so that each agent
instance gets direct feedback on the consequence of its actions, while on the other
hand, the overall optimization objective is usually a global problem that does not have
a simple “local” solution. One approach to address this dilemma is to formulate the
reward function as a weighted sum of two components, one limited to the local effects
of each agent instance and the other reflecting the global state of the system.
Second, the choice, shaping, and physical meaning of the reward function is critical
for obtaining efficient control strategies. This is well known both in RL where it
is often referred to as “reward shaping” (Ng et al. 1999), as well as in a variety of
applications within mechanics (Allaire & Schoenauer 2007). Indeed, a bad (or naive)
choice of reward function may lead to unexpected and detrimental behavior despite the
fact that an agent manages to achieve consistently high rewards. An example of this
problem is described in Rabault et al. (2019) and Rabault and Kuhnle (2019): when
using DRL to control the wake of a cylinder with the aim of reducing drag, a reward
function focusing purely on drag reduction leads to degenerate strategies with a strong
biased blowing. While this effectively reduces drag, it also creates a large lift bias. To
prevent this problem, one has to modify the reward function so as to both encourage
drag-reducing strategies while simultaneously discourage the use of detrimental tricks
to get there. As is the case with specifying good state and action spaces, designing
appropriate reward functions benefits greatly from engineering knowledge about the
system to control.
Another important point to consider when deploying DRL in fluid mechanics
applications is the parallelization of learning. Since most fluid mechanics simulations

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 386 Data-Driven Fluid Mechanics Chapter 18

Figure 18.6 Illustration of the environment-splitting method for taking advantage of physical
system invariants when performing AFC with DRL. The figures are reproduced from Belus
et al. (2019). Left: the general setup of the environment-splitting method. The neighborhood of
each actuator constitutes an individual environment, and all the environments are controlled by
clones of the same DRL agent. This way, invariance of the policy is enforced. Right: applying
this method translates in a training cost that is independent of the number of jets. This is in
stark contrast to methods that do not take advantage of these invariance properties. Indeed, as
described in Belus et al. (2019), failing to represent invariants inside the structure of the DRL
problem formulation leads to prohibitive exploration cost as the control space dimension
increases, and training fails completely in a setting without such invariant setup when as little
as five jets are involved.

are expensive, they are frequently the time bottleneck in the overall training process.
Fortunately, the DRL side of the training process can be parallelized simply by running
several copies of the environment simulation (e.g., on different cores or machines) and
enabling the agent’s learning algorithm to incorporate data from parallel experience
streams. Consequently, one can leverage parallelized CFD simulations – plus, where
applicable, multiple experience streams per simulation when implementing invariant
control, as described earlier – to effectively sample a large amount of training data,
given enough computational resources.

18.4.3 Other Techniques: Normalization, Actuation Frequency,

and Transfer Learning
So far, we have mostly discussed high-level considerations around problem modeling
when applying DRL to AFC. In this section, we focus on concrete technical points
that can have a large impact on the learning process and thus should be considered
“best practice”.
A first important detail for effective learning is to normalize state, action, and
reward values. Generally, DNNs perform best when their inputs and outputs are,
roughly, centered around zero and normally distributed or within the range [−1; 1].
This is due to several reasons, for instance, most typical nonlinear activation functions
change their “firing” state within that range, or most weight initialization schemes base
their effectiveness on assuming such input values. Therefore, both the input state and
the action output should be scaled accordingly to fall within this range. In addition,
the signals should be normalized in such a way that meaningful differences for the

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 Deep Reinforcement Learning and Active Flow Control 387

application are reflected in corresponding variation of the input/output values and

hence are “visible” to the DNN within that normalized range. For instance, if DRL
is supposed to avoid instabilities in a dynamic system, it may not be appropriate to
scale a signal linearly between -1 and 1 if a residual value of, say, 10−10 is to be
obtained. In such cases, it is probably better to first apply a logarithmic transformation
before scaling the results to [−1; 1]. As a “rule of thumb,” the signals for which a DNN
is supposed to react and behave differently should be visible to the naked eye in a plot
displaying the range from -1 to 1.
Similar reasoning applies to the reward function. However, since different methods
use the reward signal in different ways, this requires some knowledge about the type of
algorithm used (see the overview of different classes of algorithms in Section 18.2).
In particular, the normalization needs to consider the effect of the discount factor,
the episode length, and the fact that the actual training signal is not the per time
step reward, but is based on the discounted cumulative sum of rewards. Empirically,
zero-centered values not exceeding the range from −10 to 10 work well. Again, it
is important that meaningful performance differences for the application should be
reflected in the reward signal accordingly.
Another critical point for successful and stable learning is to choose an appropriate
action “frequency” or “granularity” (Rabault & Kuhnle 2019). Whereas other applica-
tions like games often naturally define the moments when an agent needs to act, this is
not obvious for physical control problems. The dynamics of a physical system usually
take place at a characteristic timescale, and the relation between this natural frequency
of the system, fs , and the frequency of agent decision time steps, fa , is important. If
fa < fs , the system is evolving too fast for the agent, meaning that, on the one hand, it
may miss important details between the unobserved action intervals and, on the other
hand, it is not able to react to system developments fast enough. However, fa  fs
is also problematic, which is probably best illustrated at the beginning of training
(but applies throughout training), when the agent is mainly exploring by randomly
picking actions. Each individual random action on its own will have little impact on
the system, and it becomes increasingly unlikely with growing fa / fs that the agent
randomly picks a long-enough sequence of coherent actions that results in meaningful
positive feedback. Consequently, the action frequency fa should be chosen such that
it is granular enough to control the system, but coarse enough to observe moments
of successful control accidentally during random exploration. In practice, after having
identified the approximate timescale fs for the system to control, fa / fs ∈ [5; 10] is
a reasonable initial choice, which can subsequently be tuned. This is illustrated in
Figure 18.7. A related technique in the RL literature is “frame skip,” which refers to
skipping a fraction of screen frames if the game simulation is rendered at a too high
frequency.
In some situations, the appropriate action frequency may be lower than either the
numerical time step in a simulation, or the maximum response frequency of sensors
and actuators in a real-world application. If such constraints are encountered, an
interpolation law should be used to obtain a smooth control law at the high frequency
corresponding to the numerical time step or capability of the hardware. The choice of

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 388 Data-Driven Fluid Mechanics Chapter 18

3.3

3.2
CD
3.1

3.0
fa / fs = 200 5
10 1
2.9
0 50 100 150 200
total nbr of episodes total CPU time

Figure 18.7 Illustration of the importance of choosing consistent timescales across the DRL
setup. The figure is reproduced with minor editing from Rabault and Kuhnle (2019), with the
permission of AIP Publishing. Depending on the choice of action frequency fa relative to the
typical timescale of the underlying system fs , the learning curves of a PPO agent attempting to
reduce drag show substantial variation. On the one hand, if the action frequency fa is too low
( fa / fs = 1), the DRL agents cannot control the system as it is unable to react fast enough to
affect the vortex shedding. On the other hand, if the action frequency fa is too high
( fa / fs = 100), learning is not successful either, since the agent’s random behavior at the
beginning of training simply adds high-frequency white noise to the system, which is
effectively smoothed out at the larger timescale of vortex shedding. However, between these
two extremes, there is a range of fa / fs ratios for which the agent successfully learns to control
the system.

the interpolation function may be important for the quality of control, as discussed in
Tang et al. (2020).
Another technique that can improve learning quality and stability is transfer
learning, as illustrated by recent work. First, Ren et al. (2020) initially trained a DRL
agent in a situation of reduced complexity – in this case, a flow at moderate Reynolds
number – before moving to the full problem complexity, a flow at a higher Reynolds
number. This approach stabilized learning and yielded better final results. Second,
Tang et al. (2020) trained a DRL agent to control a range of Reynolds numbers at once,
where the Reynolds number was randomly chosen at the beginning of each episode.
Here again, the learning process improved and the final strategies were more efficient
than comparable strategies trained to control flow just at a single Reynolds number.
This indicates that training agents on a broader range of similar control problems and
environment conditions help to find better and more robust strategies.

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 Deep Reinforcement Learning and Active Flow Control 389

18.5 Perspectives and Remaining Questions

DRL is a promising methodology for solving closed-loop control problems involving

high-dimensional nonlinear systems. In such conditions, traditional optimization
and control techniques often fail, whereas DRL with its approach of exploring the
dynamics of the system to control offers a solution to the challenges encountered by
traditional methods. Moreover, DRL is attractive for real-world applications because
of its computational requirements. While training an RL control policy can be
expensive in terms of the number of training episodes/interactions required, applying
a trained agent is not, as it simply requires a neural network inference step, whose
computational cost is usually linear in the input size. This is in stark contrast to, for
example, adjoint-based methods.
This explains why the fluid mechanics community has been investigating the use
of DRL to control increasingly complex systems. Recent work has evidenced the
fact that DRL can control complex flow dynamics in 2D configurations, even for
intermediate Reynolds number and chaotic conditions. Moreover, it has been shown
that the learned control strategies can be made robust to a range of flow regimes, which
is important for the prospect of real-world application. Consequently, next steps in
that direction involve, on the one hand, showcasing the ability of DRL to handle even
more challenging configurations and, on the other hand, investigating the application
of DRL methods to the control of 3D flows. The latter is a critical milestone for fully
qualifying DRL as an effective approach to AFC.
A more fundamental question is whether DRL can enable us to gain a better
understanding of the physical properties of the systems being controlled. No results
have been presented on this aspect yet, however, there is hope that the actuation
strategies discovered by DRL through trial and error can be combined with more
traditional analysis to gain insights into physical mechanisms. In this context, DRL
would be used as a tool to experimentally investigate the global nonlinear response
of a system, while more traditional techniques relying on, for example, linearization
or modal analysis, are employed to quantify the stability or robustness of the newly
discovered configurations.
Another promising direction for future research is to attempt to combine classical
control methods with strategies found via DRL. While traditional control methods
are often better understood and more robust, DRL by contrast is often able to find
more efficient control strategies, but with the downside that it is very challenging to
substantiate them with theoretical proofs of their stability and safety. The promise of a
combined approach is to “get the best of both worlds,” which would greatly facilitate
their adoption for practical and industrial applications. There are several possibilities.
One may, for instance, define a domain in the phase space where traditional controllers
are able to guarantee safe operation, and hence allow DRL algorithms to further
optimize the control while the system remains in this safe domain. If the system
gets close to the boundaries of guaranteed safety, DRL can be turned off and the
more interpretable traditional algorithm takes over control. Alternatively, DRL can be
applied on top of traditional control strategies, as illustrated by the residual learning

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 390 Data-Driven Fluid Mechanics Chapter 18

approach (Johannink et al. 2019), meaning that DRL learns how best to modify (within
a safe margin) the control given by the underlying system for further gains.
As a final thought, we want to emphasize the importance of reproducibility of
experimental results obtained by applying DRL, in particular with respect to open-
sourcing code and providing containerized environments. While it may be sufficient to
describe the algorithm and theoretical foundation in the case of traditional approaches,
the control strategies found by DRL cannot be described in the same way, since
they are encoded in the precise weights of the neural network. Consequently, it is
important to publish environment and training code together with the trained DRL
agent to be able to reproduce, analyze, and leverage the reported results. Moreover,
it is well known in the DRL community that reproducibility is an important problem
(Henderson et al. 2018), and that even seemingly irrelevant technical implementation
choices can play a significant role for the quality of the learned behavior (Engstrom
et al. 2020). We hope that the fluid mechanics community will follow the example
set by the ML community in this aspect, and will increasingly adopt the policy of
releasing code open source and enabling reproducibility by providing containerized
environments.

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 Part VI
Perspectives

h d i 0 0 9 8 088962 4 026 bli h d li b C b id i i

h d i 0 0 9 8 088962 4 026 bli h d li b C b id i i
 19 The Computer as Scientist
J. Jiménez

Using the identification of causally significant flow structures in two-dimensional

turbulence as an example, this chapter explores how far the usual procedure of
planning experiments to test hypotheses can be substituted by “blind” randomized
trials, and notes that the increased efficiency of computers is beginning to make
such a “Monte Carlo” approach practical in fluid mechanics. The processes of data
generation, model creation, validation, and verification are described in some detail.
Although the purpose of the chapter is to explore the procedure, rather than to develop
new models for two-dimensional turbulence, it is intriguing that the Monte Carlo
process naturally led to the consideration of vortex dipoles as building blocks of the
flow, on a par with the more classical individual vortex cores. Although not completely
novel, this “spontaneous” discovery supports the claim that an important advantage of
randomized experiments is to bypass researcher prejudice and paradigm lock. The
method can be extended to three-dimensional flows in practical times.

19.1 Introduction

We discussed in Chapter 2 the different meanings of the word “Science,” and how they
are related to the scientific method and to the idea of causality. We also mentioned
two approaches to determining causes: the observational one, which uses correlations
between past and future events to deduce rules that describe the future behavior
of the system, and the interventional one, in which the researcher monitors the
consequences of changing the initial conditions in the (often aspirational) hope of
eventually controlling it.
Most of the discussion in Chapter 2 was observational, because experiments in our
chosen field of turbulence are expensive. In this chapter, we explore the interventional
determination of causality in physical systems, and discuss whether recent technolog-
ical developments provide us with new ways of doing it. To fix ideas, we illustrate our
argument with the study of a particular flow, two-dimensional decaying homogeneous
turbulence, about which the general feeling is that most things are understood. As
such, we will be able to test our conclusions against accepted wisdom, although this
This work was supported by the European Research Council under the Coturb grant
ERC-2014.AdG-669505.

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 394 Data-Driven Fluid Mechanics Chapter 19

may lead some readers to think that the discussion is uninteresting, because nothing
new is likely to be discovered. We will see that this is not quite true. Some new
aspects can still be unearthed, or at least reemphasized, and the method for doing
so is interesting because it would have been impractical a decade ago.
We mentioned in Chapter 2 that fluid mechanics has only recently accumulated
enough data to even consider a data-driven approach, but the same has not been true
in other fields. Most medieval science was data-driven. Kepler’s laws (1618) were the
empirical results of observations, with no plausible explanation until the publication
of Newton’s Principia, 70 years later (Newton 1687). Engineering and medicine have
always been at least partly empirical and data-driven, and biochemistry has probably
been the first subject to reach the data-driven level in modern times, made possible
by the fast methods of DNA synthesis that became available in the 1980s. Much of
the modern discussion on raw empiricism relates to biochemistry (Voit 2019), and
so did the first article claiming to describe a “robot scientist” (King et al. 2009). In
fluid mechanics, the availability of enough data was made possible by the numerical
simulations of the 1990s, which appeared to promise that any question that could be
posed to a computer would eventually be answered (Brenner et al. 2019, Jiménez
2020a), but this promise was tempered by the limitation that the questions first had to
be put to the computer by a researcher.
In this chapter, we discuss to what extent this last roadblock can be removed. The
new enabling technology is the increased speed and memory of computers, which can
now do in minutes what used to take days a decade ago. We will argue that this enables
a new way of asking questions, not based on a plausible hypothesis, but randomly, in
the hope that some of them might turn out to be interesting. This procedure does not
avoid the necessity of answering the questions, although this can presumably be done
quickly by the computer, nor of evaluating how “interesting” the answers are, which
computers most probably cannot yet do. This “Monte Carlo” procedure does not point
towards a future of “human-less” research, but to a new level of partnership between
humans and computers. Similar steps have been taken before: we no longer dig canals
or throw spears by hand, except as a sport, nor do we integrate ordinary differential
equations using special functions. Most of these human–machine symbioses are
usually considered beneficial, although many created their own disruptions when they
were introduced. There is no reason to believe that this time will be different, but it is
fair to question what the new advantages and difficulties will be.
Relinquishing control of the questions to be asked is not an altogether new
experience to anybody who has trained graduate students or mentored postdocs, not
to speak of managers of large industrial or academic research groups. Any such
person knows the feeling that, at some point, the research is no longer theirs, although
most of us console ourselves by arguing that we are training our peers and that, in
the end, the overall direction is set by us. Monte Carlo science lacks both of these
(probably spurious) consolations. This may be its main advantage. It is inevitable
that our students or subordinates share some of our ideas and, most probably, some
of our prejudices. It is often argued that researcher prejudice is the main roadblock
to qualitative scientific advances, and that “paradigm shifts,” always hard to come

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 Jiménez: The Computer as Scientist 395

by, are delayed because of it (Kuhn 1970). The main advantage to be gained from a
Monte Carlo questioning algorithm is probably its lack of prejudice. If we can avoid
transferring our biases to it, such an algorithm would act as an unprejudiced, although
probably not yet very smart, scientific assistant.
The rest of this chapter is structured as follows. Section 19.2 describes how the
Monte Carlo ideas just described can be applied to the particular problem of two-
dimensional turbulence, and Section 19.3 briefly discusses the physics that can be
learned from them. Section 19.4 closes with a discussion of the connections that can
be established between the scientific method and modern data analysis or artificial
intelligence, in the light of the experience of Sections 19.2 and 19.3.

19.2 2-D Turbulence: A Case Study

Two-dimensional turbulence is an old problem in fluid mechanics, of importance in

geophysical flows (Maltrud & Vallis 1991), highly stratified situations (Voropayev
et al. 1991), and Bose–Einstein condensates (Neely et al. 2010). It differs from
three-dimensional turbulence in that the absence of vortex stretching inhibits the
amplification of velocity gradients. In the inviscid limit, energy and enstrophy are
thus conserved, and the usual energy cascade is disrupted (Betchov 1956). Early
papers centered on the statistical consequences of these conservation laws (Kraichnan
1967, 1971, Batchelor 1969), but it was soon realized that the formation of long-
lasting coherent vortices (Onsager 1949, Fornberg 1977, McWilliams 1984) interferes
with the regular statistical theory (McWilliams 1990a). Since then, two-dimensional
turbulence has mostly been analyzed in terms of the behavior of these vortices
(McWilliams 1990b, Carnevale et al. 1991). General reviews can be found in
Kraichnan and Montgomery (1980) and Boffetta and Ecke (2012).
Because two-dimensional turbulence is relatively cheap to simulate, it serves as
a good system in which to test how Monte Carlo science can be applied to a real
flow and, since a lot is already known about it, any conclusions drawn in this way
can be checked against received wisdom, both for accuracy and for originality. In this
section, we apply the ideas in Section 19.1 to test whether the vortex model mentioned
earlier is the only possible one for causality in two-dimensional decaying turbulence.
The problem, the general procedure, and some of the physical conclusions derived
from it are described in Jiménez (2018b, 2020a, 2020b). Here we center on the data-
processing aspects, and on how they can be related to the current interest in big data
analysis and artificial intelligence.

19.2.1 Data Generation and Analysis

The problem to be addressed can be summarized as follows: given a two-dimensional
field of decaying turbulence, can we identify which properties of a flow neighbor-
hood make it more or less likely to result in widespread flow changes when it is
perturbed? The underlying theme is turbulence control. For example, given a turbulent

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 396 Data-Driven Fluid Mechanics Chapter 19

Case Perturbation to cell

A ω ⇒ −ω
B u ⇒ −u

Table 19.1 Initial perturbations for the two experiments discussed in the text. In both
cases, the mean velocity and vorticity are zeroed over the full computational box after
the perturbation is applied. An extra “pressure” step is applied to enforce continuity
after modifying u in case B, and may substantially modify the intended perturbation.

atmosphere in which simulations tell us that a hurricane is likely to form after a certain
time, and assuming that we have no limitations on what can be done, what should we
do? Target existing storms? Target the jet stream?
Thus posed, the problem is to clarify causality, and our method is interventional:
that is, we change the initial conditions and classify the results (see the discussion
in the Section 2.1). A number (Nex p ) of experimental flow fields (“flows” from
now on) are prepared, and each of them is perturbed to create a variety of initial
conditions (“experiments” from now on). Each experiment is run for a prescribed
time, T, of the order of ω00 T ≈ 10 turnovers, where ω00 = hω2 i 1/2 is the root-
mean-square (r.m.s.) vorticity of the initial unperturbed flow, and the time-dependent
average h·i is taken over the full computational box. As the evolution of the perturbed
flow diverges from that of the unperturbed initial condition, the magnitude of the
evolving perturbation is defined as the norm, (t), of the difference between the
perturbed and unperturbed flow fields, evaluated over the entire computational box.
The experiments for which this magnitude is largest at some predetermined time,
Tref , are defined as most “significant.” For each experiment and test time, the nkeep
perturbations with the largest and smallest deviations (Tref ) are classified as most
or least significant, respectively, because, in common with many complex systems, it
is empirically found that the first few most- and least-significant experiments result
in fairly similar perturbation intensities (LeCun et al. 2015). Because significance is
defined as the effect on the flow at some future time, the most significant experiments
are also defined as being most “causally important” for the flow evolution. In our study,
the initial perturbation is applied by dividing each flow into a regular grid of Nc × Nc
square cells, each of which is in turn modified in a number of different ways. The two
experiments discussed here are listed in Table 19.1; several others are presented in
Jiménez (2020b). In most cases, the results are averaged over Nex p = 768 flows, with
nkeep = 5, on a grid with Nc = 10.
Simulations are performed in a doubly periodic square box of side L = 2π, using a
standard spectral Fourier code dealiased by the 2/3 rule. Time advance is third-order
Runge–Kutta. The flow is defined by its velocity field u = (u, v) as a function of the
spatial coordinates x = (x, y), and time. The scalar vorticity is ω = ∂x v − ∂y u, and
the rate-of-strain tensor is si j = (∂i u j + ∂j ui )/2, where the subindices of the partial
derivatives range over (x, y), and those of the velocity components over (u, v). The
rate-of-strain magnitude is S 2 = 2si j si j , where repeated indices imply summation.

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 Jiménez: The Computer as Scientist 397

Time and velocity are respectively scaled with ω00 , and with q00 = (u 02 + v 02 )1/2 , both
measured at the unperturbed initial time, t = 0. All the cases discussed here have
Fourier resolution 2562 , with Re = q00 L/ν = 2500, where ν is the kinematic viscosity.
Further details can be found in Jiménez (2018b, 2020b).
Figure 19.1(a,b) shows the typical initial vorticity and kinetic energy fields, with
the 10 × 10 grid overlaid, and Figure 19.1(c) shows the corresponding enstrophy
and energy spectra. This figure also includes the transfer function for a window
corresponding to the isolation of an individual cell. Because such a window acts as a
multiplicative factor in physical space, it acts as a convolution kernel for the spectra,
smoothing details narrower in wavenumber (or wider in wavelength) than the width
of the transfer function. The cell outlined in red in Figure 19.1(a,b) is one of the most
significant ones, and the one outlined in black is one of the less significant. Both are
modified as in case A in Table 19.1 (i.e., inverting the vorticity in the cell). These two
cells have been chosen so that they have similar initial perturbation intensities, but
different intensities at the classification time, ω00 Tref = 4.5. The evolution in physical
space of their perturbations is shown in Figure 19.1(d,e). It is clear that the more
significant perturbation rearranges the flow in its vicinity, and its effect eventually
spreads to the full field, while the less significant one stays localized and soon decays.
The geometric size of the perturbations can be estimated by the ratio between
their integral quadratic and point-wise maximum. It is shown in Jiménez (2020b)
that all perturbations start with sizes of the order of the cell size, Lc , and either
decay or spread to the size of the computational box, 10Lc , over times of the order
of ω00 t ≈ 5 turnovers. This can be taken as the time over which perturbations lose their
individuality, and over which causality can be usefully studied.
The purpose of our study is to determine which properties of the unperturbed
initial cells best correlate with the effect of modifying them. Several factors are
important, such as the modification method (Table 19.1), the norm used to measure
the perturbation intensity, and the time Tref at which the classification is done. The
following analysis uses the kinetic energy norm k uk2 as a measure of intensity,
and a reference time ω00 Tref = 4.5 for the classification, but similar experiments
were repeated using k uk∞ , kωk2 , and kωk∞ , with little difference in the results.
The choice of Tref is discussed in Jiménez (2020b), and its effect on the results
in Section 19.2.2. The cell properties used as diagnostic variables are the average
cell enstrophy, ωc2 = hω2 ic , defined over individual cells, the averaged kinetic
energy, qc2 = hu2 + v 2 ic , and the average magnitude of the rate-of-strain tensor, Sc2 .
Jiménez (2018b, 2020a) diagnosed significance using an optimal threshold computed
for each of these variables in isolation. Each modification of an individual cell was
classified as significant or not according to the flow behavior at Tref , and the resulting
labeled set was used to train the threshold in such a way that the number of mis-
classified cells was minimized. Jiménez (2020b) and the present notes use a version
of the same idea with the capacity of multivariable classification: support vector
machines (Cristianini & Shawe-Taylor 2000), implemented by the fitcsvm MAT-
LAB routine, which determines a separating hyperplane instead of a scalar threshold
(see Figure 19.2).

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 398 Data-Driven Fluid Mechanics Chapter 19

(a) (b) (c)

0.6

0.4

0.2

0
101 102

(d)

(e)

Figure 19.1 (a) Initial vorticity field, ω/ω00 , used for the evolutions in (d,e). Case A in Table
19.1. The cells outlined in black (less significant) and in red (more significant) have relatively
similar initial perturbation intensities, but very different later evolutions. (b) As in (a), for the
velocity magnitude, |u|/q00 . (c) Premultiplied spectra of the initial flow fields used in the
experiments, in terms of the wavenumber magnitude, k. , Enstrophy spectrum; ,
energy; , transfer functions of a box window corresponding to the cell size of the
experiments, vertically scaled to fit the plot. ◦, Nc = 4; 4, Nc = 6; 5, Nc = 10. (d) Evolution of
the perturbation vorticity for the less significant cell, marked in black in (a,b) 19.1. From left to
right: ω00 t = 0, 1.3, 2.6, 4.5. (e) As in (d) for the more significant cell.

The efficiency of the classifier, defined as the fraction of correctly classified events,
is collected in the tables in Figure 19.3 for various experimental perturbations and
combinations of diagnostic variables. It is clear that some linear combination of
enstrophy and kinetic energy is always able to separate the data almost perfectly,
but that Sc rarely helps. The best combination of vorticity and velocity used by the
classifier depends on the initial perturbation method, as shown in Figure 19.2. In
general, the enstrophy is the best diagnostic variable for perturbations that manipulate
the vorticity (case A in Figure 19.2(a)), while the kinetic energy is the dominant vari-
able for the cases that manipulate the velocity (case B in Figure 19.2(b–d)). The fifth
column in Figure 19.3 shows that some improvement can be achieved by including
contributions from both the enstrophy and the energy, but that the effect is marginal.
The classification efficiency depends on the size of the experimental cells. In gen-
eral, the efficiency degrades for larger cells, as shown in Figure 19.3 for Nc = 4 − 10.
The details of the degradation in case B can be followed by the increasing overlap
of the joint p.d.f.s in Figure 19.2(b–d). This dependence of the effectiveness of the

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 Jiménez: The Computer as Scientist 399

(a) (b) (c) (d)

3 3 3 3

2 2 2 2

1 1 1 1

0 0 0 0
0 2 4 0 2 4 0 2 4 0 2 4

Figure 19.2 Optimum classification lines for different initial perturbations, in terms of the
kinetic energy and of the enstrophy. (a) Case A, Nc = 10 (b) Case B, Nc = 10. (c) Case B,
Nc = 6. (d) Case B, Nc = 4. In all cases, 768 flows and nkeep = 5. The contour lines contain
50%, 70%, and 90% of the joint probability density functions (p.d.f.) of the diagnostic
variables for: , most significant cases; , least significant cases.

1
Case A 1.00 0.64 0.72 1.00 1.00 0.72
0.8 Nc = 10
B 0.69 0.65 1.00 0.70 1.00 1.00 0.6

1
Case A 0.89 0.58 0.72 0.92 0.89 0.72
0.8 Nc = 6
B 0.68 0.63 0.98 0.68 0.98 0.98 0.6

1
Case A 0.72 0.56 0.72 0.77 0.76 0.72
0.8 Nc = 4
B 0.64 0.60 0.89 0.64 0.90 0.89 0.6

c
Sc qc c
–Sc c
–qc Sc–qc

Figure 19.3 Efficiency of support vector machine classifiers. Unit efficiency is perfect
classification, and 0.5 is random choice. Rows are the cases in Table 19.1, and columns are the
different combinations of diagnostic variables.

enstrophy is consistent with the spectrum in Figure 19.1(c), which shows that the
cell size when Nc = 10 preserves details of the spectrum of the order of the vortex
size, but it is little surprising that small cells work so well for the kinetic energy,
whose spectrum peaks at the scale of the box, suggesting that, at least at this low
Reynolds number, the causality of the kinetic energy remains concentrated at the
scale of individual vortex pairs. In fact, the effectiveness of the enstrophy and of the
energy behave differently with the cell size. While the mean effectiveness of ωc2 as a
diagnostic variable in case A decays from 1 to 0.72 as Nc = 10 → 4, that of qc2 in case
B only decays from 1 to 0.89 in the same range.
The difference among perturbation methods is best displayed by constructing for
each case a conditional “template” for the immediate neighborhood of the significant

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 400 Data-Driven Fluid Mechanics Chapter 19

(a) (b)

Figure 19.4 Conditional vorticity and velocity distributions, normalized with their
unconditional r.m.s., in the neighborhood of the most significant cells. The test cell is outlined
in blue and, because of the invariances of the problem, the orientation and the vorticity sign are
immaterial. The vorticity contours are spaced by ω/ω00 = 0.3. Solid lines and positive, and
dashed ones are negative. (a) Case A in Table 19.1. (b) Case B. Compiled from 768 flows,
nkeep = 1.

cells of the initial unperturbed flow. Figure 19.4(a,b) includes templates for cases A
and B, built from the 3×3-cell neighborhood of the most significant cell in each exper-
iment. To take into account the reflection and rotational symmetries of the equations
of motion, the template is computed by averaging these flow patches after rotating and
reflecting them so that they mutually agree as much as possible. To compensate for the
effect of the magnitude of the templates, their intensity is scaled to match the global
intensity of the flow before comparing them to individual neighborhoods.
Figure 19.4(a), which represents the conditional structure in the vicinity of cells
which are most sensitive to inverting their vorticity, is an isolated vortex. This would
appear to support the classical view that two-dimensional turbulence is a system con-
trolled by the interactions among individual vortices (McWilliams 1990b, Carnevale
et al. 1991). But the template in Figure 19.4(b), which corresponds to cells that have
been perturbed by inverting their velocity and that are thus best diagnosed by the
magnitude of their kinetic energy, is a vortex dipole. This is a less expected result,
but a reasonable one, because the velocity field of a dipole contains a local jet, and
it makes sense that blocking it has a strong effect. In general, the templates for the
most significant structures in experiments that manipulate the vorticity are isolated
vortices, while the manipulation of the velocity results in dipoles (Jiménez 2020b).
The templates obtained from coarser experimental grids are similar to those for
Nc = 10, but they become progressively less well defined as Nc decreases.

19.2.2 Verification and Validation

Any result based on the exploitation of large data sets is statistical and, as such, needs
to be verified and validated. Verification refers to whether the experimental procedure
can be trusted, while validation tests whether the results are relevant to the problem at
hand. Further details for most of this section can be found in Jiménez (2020b).

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 Jiménez: The Computer as Scientist 401

Verification typically boils down to testing statistical significance, but it may

include other things. The simplest question is whether enough experiments have
been performed to justify trusting the averages obtained, but it is not always enough
to apply the theory for sums of independent variables. For example, consider
the templates in Figure 19.4. When they are computed using either 768 or 128
experiments, they differ by approximately kuT (128) − uT (768) k2 /kuT (768) k2 ≈ 0.2.
This appears high, because if we assume that most of the error is due to the
case with the fewest experiments, and treat the templates as simple averages, √ we
would not expect a variability with respect to the average higher than 1/ 128 ≈
0.09. However, a conditional average is more complicated than a simple sum of
independent variables. Templates are reoriented and rescaled during the conditioning,
and a relatively small misalignment of the final average can result in a large L2
difference even for essentially equivalent templates. A more relevant question is
whether our templates are more or less stable than those computed for randomly
chosen cells. Jiménez (2020b) repeated the template calculation for random cells and
for the least significant ones, and found differences of 40% between 128 and 768
random experiments, and 40–60% for the templates of the neighborhoods of the least
significant cells, compared with the 20% found above for the most significant tem-
plates cells. This suggests that the latter represents physical features, while the other
two do not.
Another important test is whether the distribution of relative approximation errors
between the flow and the template is similar when the template is tested against the
same set of experiments used to create it, or against an independent data set. This
is tested in Figure 19.5 in which the experiments have been divided into halves. The
templates are computed using the first half, and tested against either that half or against
the second one. The template-flow approximation error is displayed in the ordinates,
and both sets of results agree well, reducing the probability of template overfitting.
On the other hand, the average approximation errors are substantial, of the order of
k u − uT k2 /kuk2 ≈ 0.7. This is lower than the expected value of 1.4 between unrelated
flows of the same intensity, or than the experimental values of order unity obtained
by Jiménez (2020b) when repeating the experiments with randomly chosen cells, but
it underscores that, although templates contain information about the structure of the
flow in the neighborhood of significant cells, we should not expect to find many “pure”
circular vortices or dipoles in the flow (see Figure 19.1(a)).
Although the verification tests show that the templates in Figure 19.4 can be
statistically trusted, it remains to be shown that they represent flow properties rather
than experimental artifacts.
Consider first the sensitivity of the results to experimental conditions. We have
already seen that different experiments produce different templates, but this should be
considered a desirable property of the Monte Carlo exploration of possibilities. More
disturbing is the dependence on the testing time. The experiments above produce a
classification of flow features at time t = 0, based on the intensity of the perturbations
at a later time, Tref , and this classification is considered to be a property of the flow
at t = 0. Whether this is true or not needs to be tested, starting with whether cells

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 402 Data-Driven Fluid Mechanics Chapter 19

1.5 1.5
(a) (b)

1 1

0.5 0.5

0 0.2 0.4 0 1 2 3 4 5

Figure 19.5 (a) Joint p.d.f. of the relative approximation error, defined as the difference
between the templates and the unperturbed initial flow, ku − uT k2 /k uk2 , versus the measured
divergence at ω00 Tref = 4.5. Case B in Table 19.1, plotted against the dipole template in
Figure 19.4(b). The black lines use all the cells in 384 experiments, and the red ones only use
the most significant ones in each experiment. nkeep = 5 and Nc = 10. Contours contain 50%
and 95% of the probability. , Tested using the training set; , using an independent
test set. (b) As in (a), but plotted versus qc2 . The vertical dashed line is the discrimination
threshold in Figure 19.2(b).

classified as significant for a given Tref , remain significant when classified at a

different time.
This was checked by Jiménez (2020b), who showed that the fraction of cells
classified as significant at Tref = T1 and at T2 decays approximately as 1 − 0.04ω00 (T2 −
T1 ). Since we have seen that perturbations lose their individuality in ω00 T ≈ 5, this
means that approximately 80% of the cells classified as significant remain so during
the time of the experiment. This is substantially higher than random persistence. If
two independent sets of nkeep = 5 cells in a 10 × 10 grid are compared, their average
intersection is only expected to be about 0.25%.
More interesting is the question of whether the relation between templates and
significance can be reversed, so that not only can we say that significant cells look like
vortices or dipoles, but that vortices and dipoles are predominantly significant. This is
tested in Figure 19.5(a), which shows joint p.d.f.s of the flow-template approximation
error versus the perturbation magnitude at Tref . Contrary to Figure 19.2, the p.d.f.s
drawn in black in the figure are compiled over all the cells of the original flow, without
classification, and it is clear that there is an inverse relation between the two quantities:
Cells whose flow neighborhood is better approximated by a significant template are
also more causally significant, and vice versa. The p.d.f.s drawn in the figure with red
lines only include cells previously classified as most significant. They are the right-
most end of the unconditional distributions. As already mentioned, the figure displays
separately the result of using test data from the set used to train the template, and data
from a separate set of experiments, not used in the training. They agree well.
Figure 19.5(b) further tests the validity of the template by displaying the fitting
error against the kinetic energy of the central cell, which is the diagnostic variable of

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 Jiménez: The Computer as Scientist 403

Figure 19.6 Sample segmented vorticity field with pair identification. Blue vortices are
counterclockwise, and red ones are clockwise. Black connectors are corrotating pairs, magenta
are counterrotating dipoles, and yellow markers are unpaired vortices. Note that, because of
their irregular shape, some centers of gravity fall outside their vortex. Vorticity threshold,
|ω| = 0.9ω00 .

choice for the particular experiment in the figure. Both quantities are also correlated,
especially for badly fitting neighborhoods, whose kinetic energy is also low. As with
the perturbation magnitude, the significant cells are just the low-approximation-error
end of the joint p.d.f.

19.3 Vortices and Dipoles

Although a detailed discussion of the dynamics of two-dimensional turbulence

is beyond the scope of the present notes, which are mostly concerned with the
methodological aspects of how best to collect the information required for scientific
discovery, it is still useful to review the most interesting physical aspect of the previous
results, which is the classification of significant structures into vortices and dipoles.
Individual vortices are widely recognized to be the dominant coherent structures
of two-dimensional turbulent flows (McWilliams 1990b). But vortex dipoles (“mod-
ons”) have also been investigated as important components of rotating quasi-two-
dimensional flows, particularly in geophysics (Flierl et al. 1980, McWilliams 1980),
where they retain their individuality for long times in the atmosphere and in the
ocean. However, although their presence in nonrotating two-dimensional turbulence
is also well known, they have tended to be considered “transient” structures, whose
importance is usually not emphasized (McWilliams 1990b). The present results
suggest that they deserve a second look. Simulations of a point vortex gas, which can
be viewed as a non-dissipative model for two-dimensional turbulence, suggest that the
relevant property of dipoles in this context is that they carry no net circulation, so that
their interaction with the rest of the flow is relatively weak. They are therefore able

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 404 Data-Driven Fluid Mechanics Chapter 19

to move for long distances before they are destroyed by collisions with other objects,
and they stir the flow in the process.
To gain a better sense of the prevalence of dipoles in two-dimensional turbulence,
Figure 19.6 shows a segmentation of a sample flow field into individual positive and
negative vortices. Vortices are defined as connected regions in which |ω| ≥ Hω00 , with
H = 0.9, chosen to maximize the number of individual vortices (Moisy & Jiménez
2004). The vortices in Figure 19.6 have been grouped, whenever possible, into co and
counterrotating pairs. Two vortices are considered a potential pair if their areas differ
by less than a factor of m2 , which is an adjustable parameter. The figure uses m = 2,
but statistics compiled with m = 1.5 and m = 3 show no substantial differences.
Vortices are paired to the closest unpaired neighbor within their area class, and no
vortex can have more than one partner. Some vortices find no suitable partner, and are
left unpaired. Statistics compiled over approximately 8,500 independent flow fields
show that, out of approximately 5 × 105 vortices, 48% are paired to form dipoles, 24%
are in corotating pairs, and 28% are isolated. Most vortices in the flow are thus in the
form of pairs, mostly dipoles. The difference between corotating and counterrotating
pairs is interesting, but the reason is probably that corotating pairs tend to merge into
single cores (Meunier et al. 2005), while modons are long-lasting (Flierl et al. 1980,
McWilliams 1980).

19.4 Discussion and Conclusions

This chapter, together with Chapter 2, have explored two related but independent ideas
in the context of scientific discovery in turbulence research.
The first idea is the well-known distinction between correlation and causation. We
have noted that, because of the difficulty of doing experiments in turbulence (compu-
tational or otherwise), research has tended to center on correlations. Chapter 2 gave
examples of what can be learned from this approach, and noted that the wealth of data
generated by direct simulations in the 1990s probably contributed to that trend by cre-
ating the illusion that “we know everything,” and that all questions can be answered.
There are two problems with this illusion. The first one is that, although large data
sets contains many answers, the probability of finding them without active experi-
mentation is very small. Consider how difficult it would be to study large asteroid
strikes on Earth by observing the configuration of the solar system before spontaneous
strikes happen. These are rare, and it is unclear whether the precursor to one of them
applies to all the others. It is much more efficient to perturb the system (hopefully
computationally) and observe the result. The same is true of many intermittent but
significant processes in turbulence, especially if we are interested in controlling the
flow by introducing local out-of-equilibrium perturbations. Studying them requires
experiments that separate present causes from later effects. We noted in Chapter 2 that
correlation is the tool of prediction, while causation is the tool of control.
The second problem is how to choose which experiments to perform. The classical
search for scientific causation is hypothesis-driven. The entry point to the optimization

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 Jiménez: The Computer as Scientist 405

cycle of scientific research is that the researcher thinks of a model, and designs
experiments to test it. While this “hypothesis-driven search” has obvious theoretical
appeal, it risks limiting creativity. A model has to be conceived before it is tested, and
new ideas depend on the imagination of the researchers. However, we have noted at
the beginning of this chapter that faster experimental and computational methods open
the possibility of what we have called Monte Carlo searchers (perhaps describable
as “search-driven hypotheses”) in which experiments are performed randomly, and
evaluated a posteriori in the hope that some of them may be interesting. This is more
expensive than the classical procedure, but may be our best hope of avoiding ingrained
prejudice.
We have illustrated these ideas by the application to two-dimensional turbulence
in Sections 19.2 and 19.3, but the subject of these lectures is not turbulence itself
but the method, and we now briefly discuss how far the hopes expressed earlier have
been realized in the exercise. The first consideration is cost, which is always a pacing
item for turbulence research. The whole program in this paper took about a month
of computer time in a medium-sized department cluster, and was programmed by the
author during his spare time over a year. Considerably more time was spent discussing
with colleagues what should be done than in actually doing it. Since the effort was
conceived as a proof of principle, the problem was chosen small on purpose, but more
interesting problems are within reach. The basic simulation that had to be repeated
many times for the experiments in Section 19.2 runs in about 10 core-seconds, but
even a 2563 simulation of three-dimensional isotropic turbulence can be run in two
minutes in a modern GPU (Jiménez 2020a). A program to address causality in the
three-dimensional turbulence cascade, about which considerably less is known than in
two dimensions, could thus be performed in a modest GPU cluster in a few months.
Again, the main roadblock would be to decide what to do.
The second question is whether something of physical interest has been learned.
As mentioned at the beginning of Section 19.2, little was expected from a problem
that is usually considered to be essentially understood, but the actual results were
interesting. This chapter broadly follows the discussion in Jiménez (2020b), but two
early versions (Jiménez 2018b, Jiménez 2020a) missed the dipoles completely, and
concluded that the experiments confirmed the classical vortex-gas model of two-
dimensional turbulence. The dipole template in Figure 19.4(b) was a mildly surprising
result of further postprocessing and, if we admit that surprise is one of the defining
ingredients of discovery (Schaffer 1994), it was a minor discovery. Unfortunately, we
saw in Section 19.3 that dipoles were not completely unknown components of two-
dimensional flows, and that finding them was rather an instance of recalling something
that had been forgotten than of discovering something truly new. It is clear that the
present chapter needs a lot of extra work before it becomes an article of independent
interest in fluid mechanics, rather than in methodology, but the fact that something
unexpected by the author was found without “prompting” is encouraging for the future
of the Monte Carlo method in problems in which something is genuinely unknown.
The third question is what have we learned about the process of data exploitation.
The scientific method can be seen as an optimization loop to search for the “best”

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 406 Data-Driven Fluid Mechanics Chapter 19

Figure 19.7 The scientific method. Reproduced from Chapter 2 for easy of reference.

hypothesis, and it is natural to ask which parts of it can be automated. Our case
study in Sections 19.2 and 19.3 was conceived as an experimental test of how far
the automation process could be pushed. Several things can be concluded. The first
one is that step S1 in Figure 19.7 (observations) can be largely outsourced to the
computer, including the reference to “asking questions.” The experiments in Table
19.1 were decided (on purpose) with little thought about their significance, although
it is difficult to say how much they were influenced by the previous experience of the
author. The same can be said about the parts of step S3 that refer to testing predictions
against observations. Simulations of two-dimensional turbulence are by now trivial,
and the classification of the results was outsourced to library programs (Figures
19.2 and 19.3). An interesting, although not completely unexpected, outcome of the
experience has been the importance of verification and validation, and a rereading of
Section 19.2.2 and, up to a point, of Section 19.3, shows pitfalls that were avoided this
way. Such problems are expected in any project involving data analysis, but they are
especially dangerous in cases, like the present one in which the goal is to probe the
unknown.
The last point to consider is the model generation and evaluation step (S2) in
Figure 19.7, which is the core of the discovery process. Even here, something was
automated. The templates in Figure 19.4 are flow models, and they were obtained
automatically. Note that it is at this point that the transition from subsymbolic to
symbolic AI took place in this example. Figures 19.2 and 19.3 are primarily useful
for computer classification, but the templates in Figure 19.4 are intended for humans.
On the other hand, the interpretation in Section 19.3 of the templates was entirely
manual, and it is difficult to see at the moment how it could be outsourced to a
computer. This is not only because of the need for a level of intelligence somewhat
above present computer software, but because what are we trying to do is not well
defined. If the goal of a scientific project is to find a “good” model, we need to define
precisely what we consider a good hypothesis. We mentioned in Chapter 2 that one
of the goals of the scientific method was to produce “beautiful” theories, but we do
not know very well how to define scientific beauty. The ultimate models for two-
dimensional turbulence are the Navier–Stokes equations, and the ultimate causes of
any observation are some notional initial conditions. We have restricted ourselves

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 Jiménez: The Computer as Scientist 407

here to a time horizon shorter than the memory loss due to chaos, because of some
generalized interest in flow control. This is a choice that influences the results and
makes them less general, but it defines the metric that allows us to define templates.
Human supervision will probably still be required for some time to refine hypothe-
ses, but Monte Carlo science can contribute something even now. Consider again
the interpretation of the scientific method as an optimization loop to maximize
“understanding.” As such, it could conceivably be implemented as an automatic
optimization procedure (e.g., a neural network). However, most classical optimizers,
including humans, assume local convexity of the cost function, which is at the root of
our misgivings about researcher originality and prejudice. Monte Carlo is a different
way of looking for a maximum, which, in principle, bypasses the convexity constraint
and escapes local maxima by injecting noise in the process (Deb & Gupta 2006). The
best use of Monte Carlo science is probably as a partially randomized search and
classification step, followed by repeated human fine-tuning. The encouraging news in
fluid mechanics is that doing this is now becoming possible.

h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
h d i 0 0 9 8 088962 4 02 bli h d li b C b id i i
 References 409

References

Abdi, H. (2003), ‘Factor rotations in factor analyses’, Encyclopedia for Research Methods for
the Social Sciences, Sage, Thousand Oaks, CA, pp. 792–795.
Abraham, R., Marsden, J. E., and Marsden, J. E. (1978), Foundations of Mechanics, Benjam-
in/Cummings Publishing Company, Reading, MA.
Abraham, R., Marsden, J. E., and Ratiu, T. (1988), Manifolds, Tensor Analysis, and Applica-
tions, Vol. 75 of Applied Mathematical Sciences, Springer-Verlag, New York.
Abu-Mostafa, Y. S., Magndon-Ismail, M., and Lin, H.-T. (2012), Learning from Data: A Short
Course, https://amlbook.com/.
Abu-Zurayk, M., Ilic, C., Schuster, A., and Liepelt, R. (2017), Effect of gradient approximations
on aero-structural gradient-based wing optimization, in ‘EUROGEN 2017’, September 13–
15, 2017, Madrid, Spain.
Adrian, R. J. (1975), On the role of conditional averages in turbulence theory, in ‘Proceedings of
the 4th Biennial Symposium on Turbulence in Liquids, University of Missouri-Rolla, 1975’,
Science Press, Princeton, NJ.
Adrian, R. J. (1991), ‘Particle-imaging techniques for experimental fluid mechanics’, Annual
Review of Fluid Mechanics 23(1), 261–304.
Adrian, R. J. (2007), ‘Hairpin vortex organization in wall turbulence’, Physics of Fluids
19, 041301.
Adrian, R. J., Jones, B., Chung, M., Hassan, Y., Nithianandan, C., and Tung, A.-C. (1989),
‘Approximation of turbulent conditional averages by stochastic estimation’, Physics of
Fluids A: Fluid Dynamics 1(6), 992–998.
Ahuja, S. and Rowley, C. W. (2010), ‘Feedback control of unstable steady states of flow past a
flat plate using reduced-order estimators’, Journal of Fluid Mechanics 645, 447–478.
Albers, M., Meysonnat, P. S., Fernex, D., Semaan, R., Noack, B. R., and Schröder, W. (2020),
‘Drag reduction and energy saving by spanwise traveling transversal surface waves for flat
plate flow’, Flow, Turbulence and Combustion 105, 125–157.
Aleksic, K., Luchtenburg, D. M., King, R., Noack, B. R., and Pfeiffer, J. (2010), Robust
nonlinear control versus linear model predictive control of a bluff body wake, in ‘5th AIAA
Flow Control Conference’, June 28–July 1, 2010, Chicago, AIAA-Paper 2010-4833, pp.
1–18.
Allaire, G. and Schoenauer, M. (2007), Conception optimale de structures, Vol. 58 of
Mathématiques et Applications, Springer, Berlin.
Alsalman, M., Colvert, B., and Kanso, E. (2018), ‘Training bioinspired sensors to classify
flows’, Bioinspiration & Biomimetics 14(1), 016009.
Amsallem, D., Zahr, M. J., and Farhat, C. (2012), ‘Nonlinear model order reduction based on
local reduced-order bases’, International Journal for Numerical Methods in Engineering
92(10), 891–916.
Antoranz, A., Gonzalo, A., Flores, O., and Garcia-Villalba, M. (2015), ‘Numerical simulation of
heat transfer in a pipe with non-homogeneous thermal boundary conditions’, International
Journal of Heat and Fluid Flow 55, 45–51.
Antoranz, A., Ianiro, A., Flores, O., and Garcı́a-Villalba, M. (2018), ‘Extended proper
orthogonal decomposition of non-homogeneous thermal fields in a turbulent pipe flow’,
International Journal of Heat and Mass Transfer 118, 1264–1275.
Antoulas, A. C. (2005), Approximation of Large-Scale Dynamical Systems, Society for Indus-
trial and Applied Mathematics, Philadelphia.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 410 Data-Driven Fluid Mechanics

Arbabi, H. and Mezić, I. (2017), ‘Ergodic theory, dynamic mode decomposition, and computa-
tion of spectral properties of the Koopman operator’, SIAM Journal on Applied Dynamical
Systems 16(4), 2096–2126.
Armstrong, E. and Sutherland, J. C. (2021), ‘A technique for characterizing feature size and
quality of manifolds’, Combustion Theory and Modelling, 25, 646–668.
Arthur, D. and Vassilvitskii, S. (2007), k-means++: The advantages of careful seeding, in
‘Proceedings of the 18th Annual ACM-SIAM Symposium on Discrete Algorithms’, Society
for Industrial and Applied Mathematics, Philadelphia, pp. 1027–1035.
Ashurst, W. T. and Meiburg, E. (1988), ‘Three-dimensional shear layers via vortex dynamics’,
Journal of Fluid Mechanics 189, 87–116.
Aström, K. J. and Murray, R. M. (2010), Feedback Systems: An Introduction for Scientists and
Engineers, Princeton University Press, Princeton, NJ.
Aubry, N., Holmes, P., Lumley, J. L., and Stone, E. (1988), ‘The dynamics of coherent structures
in the wall region of a turbulent boundary layer’, Journal of Fluid Mechanics 192(1), 115–
173.
Baars, W. J. and Tinney, C. E. (2014), ‘Proper orthogonal decomposition-based spectral higher-
order stochastic estimation’, Physics of Fluids 26(5), 055112.
Babaee, H. and Sapsis, T. P. (2016), ‘A variational principle for the description of time-
dependent modes associated with transient instabilities’, Philosophical Transactions of the
Royal Society of London 472, 20150779.
Bagheri, S., Brandt, L., and Henningson, D. (2009), ‘Input–output analysis, model reduction
and control of the flat-plate boundary layer’, Journal of Fluid Mechanics 620, 263–298.
Bagheri, S., Hoepffner, J., Schmid, P. J., and Henningson, D. S. (2009), ‘Input–output analysis
and control design applied to a linear model of spatially developing flows’, Applied
Mechanics Reviews 62(2), 020803-1-020803-27.
Balajewicz, M., Dowell, E. H., and Noack, B. R. (2013), ‘Low-dimensional modelling of high-
Reynolds-number shear flows incorporating constraints from the Navier-Stokes equation’,
Journal of Fluid Mechanics 729, 285–308.
Baldi, P. and Hornik, K. (1989), ‘Neural networks and principal component analysis: Learning
from examples without local minima’, Neural Networks 2(1), 53–58.
Banerjee, I. and Ierapetritou, M. (2006), ‘An adaptive reduction scheme to model reactive flow’,
Combustion and Flame 144(3), 619–633.
Baraniuk, R. G. (2007), ‘Compressive sensing’, IEEE Signal Processing Magazine 24(4), 118–
120.
Barbagallo, A., Sipp, D., and Schmid, P. J. (2009), ‘Closed-loop control of an open cavity flow
using reduced-order models’, Journal of Fluid Mechanics 641, 1–50.
Barkley, D. and Henderson, R. (1996), ‘Three-dimensional Floquet stability analysis of the
wake of a circular cylinder’, Journal of Fluid Mechanics 322, 215–241.
Barlow, R. S. and Frank, J. H. (1998), ‘Effects of turbulence on species mass fractions in
methane/air jet flames’, Symposium (International) on Combustion 27, 1087–1095.
Batchelor, G. K. (1969), ‘Computation of the energy spectrum in homogeneous two dimensional
turbulence’, Physics of Fluids 12 (Suppl. II), 233–239.
Battaglia, P. W., Hamrick, J. B., Bapst, V., Sanchez-Gonzalez, A., Zambaldi, V., Malinowski,
M., Tacchetti, A., Raposo, D., Santoro, A., Faulkner, R. Gulcehre, C., Song, F., Ballard,
A., Gilmer, J., Dahl, G., Vaswani, A., Allen, K., Nash, C., Langston, V., Dyer, C., Heess, N.,
Wierstra, D., Kohli, P., Botvinick, M., Vinyals, O., Li, Y., and Pascanu, R. (2018), ‘Relational
inductive biases, deep learning, and graph networks’, arXiv preprint arXiv:1806.01261.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 References 411

Beerends, R. J., ter Morsche, H. G., van den Berg, J. C., and van de Vrie, E. M. (2003), Fourier
and Laplace Transforms, Cambridge University Press, New York.
Beetham, S. and Capecelatro, J. (2020), ‘Formulating turbulence closures using sparse regres-
sion with embedded form invariance’, arXiv preprint arXiv:2003.12884 .
Beintema, G., Corbetta, A., Biferale, L., and Toschi, F. (2020), ‘Controlling Rayleigh–Bénard
convection via reinforcement learning’, Journal of Turbulence 21(9–10), 585–605.
Bekemeyer, P., Ripepi, M., Heinrich, R., and Görtz, S. (2019), ‘Nonlinear unsteady reduced-
order modeling for gust-load predictions’, AIAA Journal 57(5), 1839–1850.
Bekemeyer, P., Wunderlich, T., Görtz, S., and Dähne, D. (2019), Effect of gradient approx-
imations on aero-structural gradient-based wing optimization, in ‘International Forum on
Aeroelasticity and Structural Dynamics (IFASD-2019)’, June 9–13, 2019, Savannah, GA.
Bekemeyer, P., Bertram, A., Chaves, D. A. H., Ribeiro, M. D., Garbo, A., Kiener, A., Sabater,
C., Stradtner, M., Wassing, S., Widhalm, M., Goertz, S., Jaeckel, F., Hoppe, R. & Hoffmann,
N. (2022), ‘Data-driven aerodynamic modeling using the DLR SMARTy toolbox’.
Bellemans, A., Aversano, G., Coussement, A., and Parente, A. (2018), ‘Feature extraction and
reduced-order modelling of nitrogen plasma models using principal component analysis’,
Computers & Chemical Engineering 115, 504–514.
Bellemare, M. G., Dabney, W., and Munos, R. (2017), A distributional perspective on
reinforcement learning, in ‘Proceedings of the 34th International Conference on Machine
Learning – Volume 70’, ICML 17, JMLR.org, pp. 449–458.
Bellman, R. (1957), ‘A Markovian decision process’, Journal of Mathematics and Mechanics
6(5), 679–684.
Belson, B. A., Semeraro, O., Rowley, C. W., and Henningson, D. S. (2013), ‘Feedback control
of instabilities in the two-dimensional Blasius boundary layer: The role of sensors and
actuators’, Physics of Fluids (1994–present) 25(5), 054106.
Belson, B. A., Tu, J. H., and Rowley, C. W. (2014), ‘Algorithm 945: Modred – A parallelized
model reduction library’, ACM Transactions on Mathematical Software (TOMS) 40(4), 30.
Belus, V., Rabault, J., Viquerat, J., Che, Z., Hachem, E., and Reglade, U. (2019), ‘Exploiting
locality and translational invariance to design effective deep reinforcement learning control
of the 1-dimensional unstable falling liquid film’, AIP Advances 9(12), 125014.
Bence, J. R. (1995), ‘Analysis of short time series: Correcting for autocorrelation’, Ecology
76(2), 628–639.
Benner, P., Gugercin, S., and Willcox, K. (2015), ‘A survey of projection-based model reduction
methods for parametric dynamical systems’, SIAM Review 57(4), 483–531.
Bergmann, M., Cordier, L., and Brancher, J.-P. (2005), ‘Optimal rotary control of the cylinder
wake using proper orthogonal decomposition reduced-order model’, Physics of Fluids
(1994–present) 17(9), 097101.
Berkooz, G., Holmes, P., and Lumley, J. L. (1993), ‘The proper orthogonal decomposition in
the analysis of turbulent flows’, Annual Review of Fluid Mechanics 23, 539–575.
Bernal, L. P. (1981), The coherent structure of turbulent mixing layers, PhD thesis, California
Institute of Technology, Pasadena.
Berry, M. W., Browne, M., Langville, A. N., Pauca, V. P., and Plemmons, R. J. (2007), ‘Algo-
rithms and applications for approximate nonnegative matrix factorization’, Computational
Statistics & Data Analysis 52(1), 155–173.
Bertolotti, F., Herbert, T., and Spalart, P. (1992), ‘Linear and nonlinear stability of the Blasius
boundary layer’, Journal of Fluid Mechanics 242, 441–474.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 412 Data-Driven Fluid Mechanics

Betchov, R. (1956), ‘An inequality concerning the production of vorticity in isotropic turbu-
lence’, Journal of Fluid Mechanics 1, 497–504.
Bieker, K., Peitz, S., Brunton, S. L., Kutz, J. N., and Dellnitz, M. (2019), ‘Deep model
predictive control with online learning for complex physical systems’, arXiv preprint
arXiv:1905.10094.
Biglari, A. and Sutherland, J. C. (2012), ‘A filter-independent model identification technique
for turbulent combustion modeling’, Combustion and Flame 159(5), 1960–1970.
Biglari, A. and Sutherland, J. C. (2015), ‘An a-posteriori evaluation of principal component
analysis-based models for turbulent combustion simulations’, Combustion and Flame
162(10), 4025–4035.
Bilger, R., Starner, S., and Kee, R. (1990), ‘On reduced mechanisms for methane-air combustion
in nonpremixed flames’, Combustion and Flame 80, 135–149.
Billings, S. A. (2013), Nonlinear System Identification: NARMAX Methods in the Time,
Frequency, and Spatio-Temporal Domains, John Wiley & Sons, Hoboken, NJ.
Bishop, C. M. (2016), Pattern Recognition and Machine Learning, Springer, New York.
Bistrian, D. and Navon, I. (2016), ‘Randomized dynamic mode decomposition for non-intrusive
reduced order modelling’, International Journal for Numerical Methods in Engineering 112,
3–25.
Black, F., Schulze, P., and Unger, B. (2019), ‘Nonlinear Galerkin model reduction for systems
with multiple transport velocities’, arXiv preprint arXiv:1912.11138.
Boffetta, G. and Ecke, R. E. (2012), ‘Two-dimensional turbulence’, The Annual Review of Fluid
Mechanics 44, 427–451.
Bohn, E., Coates, E. M., Moe, S., and Johansen, T. A. (2019), Deep reinforcement learning
attitude control of fixed-wing UAVs using proximal policy optimization, in ‘2019 Interna-
tional Conference on Unmanned Aircraft Systems (ICUAS)’, June 11–14, 2019, Atlanta,
GA, IEEE, pp. 523–533.
Bongard, J. and Lipson, H. (2007), ‘Automated reverse engineering of nonlinear dynamical
systems’, Proceedings of the National Academy of Sciences 104(24), 9943–9948.
Borée, J. (2003), ‘Extended proper orthogonal decomposition: A tool to analyse correlated
events in turbulent flows’, Experiments in Fluids 35(2), 188–192.
Bourgeois, J. A., Martinuzzi, R. J., and Noack, B. R. (2013), ‘Generalised phase average with
applications to sensor-based flow estimation of the wall-mounted square cylinder wake’,
Journal of Fluid Mechanics 736, 316–350.
Box, G. E., Jenkins, G. M., Reinsel, G. C., and Ljung, G. M. (2015), Time Series Analysis:
Forecasting and Control, John Wiley & Sons, Hoboken, NJ.
Boyd, S., Parikh, N., Chu, E., Peleato, B., and Eckstein, J. (2011), ‘Distributed optimization
and statistical learning via the alternating direction method of multipliers’, Foundations and
Trends in Machine Learning 3, 1–122.
Brackston, R., De La Cruz, J. G., Wynn, A., Rigas, G., and Morrison, J. (2016), ‘Stochastic
modelling and feedback control of bistability in a turbulent bluff body wake’, Journal of
Fluid Mechanics 802, 726–749.
Breiman, L. (2001), ‘Random forests’, Machine Learning 45(1), 5–32.
Brenner, M., Eldredge, J., and Freund, J. (2019a), ‘Perspective on machine learning for
advancing fluid mechanics’, Physical Review Fluids 4(10), 100501.
Bright, I., Lin, G., and Kutz, J. N. (2013), ‘Compressive sensing and machine learning strategies
for characterizing the flow around a cylinder with limited pressure measurements’, Physics
of Fluids 25(127102), 1–15.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 References 413

Brockwell, P. J. and Davis, R. A. (2010), Introduction to Time Series and Forecasting (Springer
Texts in Statistics), Springer, New York.
Brooks, R. (1990), ‘Elephants don’t play chess’, Robotics and Autonomous Systems 6, 3–15.
Brown, G. L. and Roshko, A. (1974), ‘On the density effects and large structure in turbulent
mixing layers’, Journal of Fluid Mechanics 64, 775–816.
Brunton, B. W., Brunton, S. L., Proctor, J. L., and Kutz, J. N. (2016a), ‘Sparse sensor
placement optimization for classification’, SIAM Journal on Applied Mathematics 76(5),
2099–2122.
Brunton, S., Brunton, B., Proctor, J., Kaiser, E., and Kutz, J. (2017), ‘Chaos as an intermittently
forced linear system’, Nature Communication 8(19), 1–9.
Brunton, S. L., Brunton, B. W., Proctor, J. L., and Kutz, J. N. (2016b), ‘Koopman invariant
subspaces and finite linear representations of nonlinear dynamical systems for control’, PLoS
ONE 11(2), e0150171.
Brunton, S. L., Dawson, S. T. M., and Rowley, C. W. (2014), ‘State-space model identification
and feedback control of unsteady aerodynamic forces’, Journal of Fluids and Structures
50, 253–270.
Brunton, S. L., Hemati, M. S., and Taira, K. (2020), ‘Special issue on machine learning and
data-driven methods in fluid dynamics’, Theoretical and Computational Fluid Dynamics 34,
333–337.
Brunton, S. L. and Kutz, J. N. (2019), Data-Driven Science and Engineering: Machine
Learning, Dynamical Systems, and Control, Cambridge University Press, Cambridge.
Brunton, S. L. and Noack, B. R. (2015), ‘Closed-loop turbulence control: Progress and
challenges’, Applied Mechanics Reviews 67(5), 050801:01–48.
Brunton, S. L., Noack, B. R., and Koumoutsakos, P. (2020), ‘Machine learning for fluid
mechanics’, Annual Review of Fluid Mechanics 52, 477–508.
Brunton, S. L., Proctor, J. L., and Kutz, J. N. (2016a), ‘Discovering governing equations from
data by sparse identification of nonlinear dynamical systems’, Proceedings of the National
Academy of Sciences 113(15), 3932–3937.
Brunton, S. L., Proctor, J. L., and Kutz, J. N. (2016b), ‘Sparse identification of nonlinear
dynamics with control (SINDYc)’, IFAC NOLCOS 49(18), 710–715.
Bucci, M. A., Semeraro, O., Allauzen, A., Wisniewski, G., Cordier, L., and Mathelin, L. (2019),
‘Control of chaotic systems by deep reinforcement learning’, Proceedings of the Royal
Society A: Mathematical, Physical and Engineering Sciences 475(2231), 20190351.
Budišić, M. and Mezić, I. (2009), An approximate parametrization of the ergodic partition using
time averaged observables, in ‘Proceedings of the 48th IEEE Conference on Decision and
Control, 2009 held jointly with the 2009 28th Chinese Control Conference. CDC/CCC 2009.
December 15–18, 2009, Shanghai, China, IEEE, Piscataway, NJ, pp. 3162–3168.
Budišić, M. and Mezić, I. (2012), ‘Geometry of the ergodic quotient reveals coherent structures
in flows’, Physica D: Nonlinear Phenomena 241(15), 1255–1269.
Budišić, M., Mohr, R. and Mezić, I. (2012), ‘Applied Koopmanism’, Chaos: An Interdisci-
plinary Journal of Nonlinear Science 22(4), 047510.
Burda, Y., Edwards, H., Storkey, A. J., and Klimov, O. (2019), Exploration by random network
distillation, in ‘7th International Conference on Learning Representations, ICLR 2019, New
Orleans, LA, USA, May 6–9, 2019’, OpenReview.net.
Burkardt, J., Gunzburger, M., and Lee, H.-C. (2006), ‘POD and CVT-based reduced-order mod-
eling of Navier–Stokes flows’, Computer Methods in Applied Mechanics and Engineering
196(1–3), 337–355.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 414 Data-Driven Fluid Mechanics

Burkov, A. (2019), The Hundred-Page Machine Learning Book, Andriy Burkov Quebec City.
Busse, F. (1991), ‘Numerical analysis of secondary and tertiary states of fluid flow and their
stability properties’, Applied Scientific Research 48(3), 341–351.
Butler, K. and Farrell, B. (1992), ‘Three-dimensional optimal perturbations in viscous shear
flow’, Physics of Fluids 4, 1637–1650.
Butler, K. M. and Farrell, B. F. (1993), ‘Optimal perturbations and streak spacing in wall-
bounded shear flow’, Physics of Fluids A 5, 774–777.
Callaham, J., Maeda, K., and Brunton, S. L. (2019), ‘Robust reconstruction of flow fields from
limited measurements’, Physical Review Fluids 4, 103907.
Cammilleri, A., Gueniat, F., Carlier, J., Pastur, L., Memin, E., Lusseyran, F., and Artana,
G. (2013), ‘POD-spectral decomposition for fluid flow analysis and model reduction’,
Theoretical and Computational Fluid Dynamics 27(6), 787–815.
Candès, E. J. (2006), Compressive sensing, in ‘Proceedings of the International Congress of
Mathematics, August 22–30, 2006’, European Mathematical Society, Zurich.
Candès, E. J., Romberg, J., and Tao, T. (2006a), ‘Robust uncertainty principles: Exact signal
reconstruction from highly incomplete frequency information’, IEEE Transactions on
Information Theory 52(2), 489–509.
Candès, E. J., Romberg, J., and Tao, T. (2006b), ‘Stable signal recovery from incomplete and
inaccurate measurements’, Communications in Pure and Applied Mathematics 8, 1207–
1223.
Candès, E. J. and Tao, T. (2006), ‘Near optimal signal recovery from random projections:
Universal encoding strategies?’, IEEE Transactions on Information Theory 52(12), 5406–
5425.
Cardesa, J. I., Vela-Martı́n, A., and Jiménez, J. (2017), ‘The turbulent cascade in five
dimensions’, Science 357, 782–784.
Carlberg, K., Barone, M., and Antil, H. (2017), ‘Galerkin v. least-squares Petrov–Galerkin
projection in nonlinear model reduction’, Journal of Computational Physics 330, 693–734.
Carlberg, K., Bou-Mosleh, C., and Farhat, C. (2011), ‘Efficient non-linear model reduction
via a least-squares Petrov–Galerkin projection and compressive tensor approximations’,
International Journal for Numerical Methods in Engineering 86(2), 155–181.
Carlberg, K., Tuminaro, R., and Boggs, P. (2015), ‘Preserving Lagrangian structure in nonlinear
model reduction with application to structural dynamics’, SIAM Journal on Scientific
Computing 37(2), B153–B184.
Carleman, T. (1932), ‘Application de la théorie des équations intégrales linéaires aux systémes
d’équations différentielles non linéaires’, Acta Mathematica 59(1), 63–87.
Carnevale, G. F., McWilliams, J. C., Pomeau, Y., Weiss, J. B., and Young, W. R. (1991),
‘Evolution of vortex statistics in two-dimensional turbulence’, Physical Review Letters
66, 2735–2737.
Cassel, K. (2013), Variational Methods with Applications in Science and Engineering,
Cambridge University Press, Cambridge.
Cattafesta III, L. N. and Sheplak, M. (2011), ‘Actuators for active flow control’, Annual Review
of Fluid Mechanics 43, 247–272.
Cattell, R. B. (1966), ‘The scree test for the number of factors’, Multivariate Behavioral
Research 1(2), 245–276.
Cavaliere, A. and de Joannon, M. (2004), ‘Mild combustion’, Progress in Energy and
Combustion Science 30, 329–366.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 References 415

Champion, K., Lusch, B., Kutz, J. N., and Brunton, S. L. (2019), ‘Data-driven discovery of
coordinates and governing equations’, Proceedings of the National Academy of Sciences
116(45), 22445–22451.
Champion, K., Zheng, P., Aravkin, A. Y., Brunton, S. L., and Kutz, J. N. (2020), ‘A unified
sparse optimization framework to learn parsimonious physics-informed models from data’,
IEEE Access 8, 169259–169271.
Chang, H., Yeung, D.-Y., and Xiong, Y. (2004), Super-resolution through neighbor embedding,
in ‘Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and
Pattern Recognition, 2004. CVPR 2004.’, Vol. 1, IEEE, pp. I–I.
Chartrand, R. (2011), ‘Numerical differentiation of noisy, nonsmooth data’, ISRN Applied
Mathematics 2011, article ID 164564.
Chen, K. K. and Rowley, C. W. (2011), ‘H2 optimal actuator and sensor placement in the
linearised complex Ginzburg-Landau system’, Journal of Fluid Mechanics 681, 241–260.
Chen, K. K., Tu, J. H., and Rowley, C. W. (2012), ‘Variants of dynamic mode decomposi-
tion: Boundary condition, Koopman, and Fourier analyses’, Journal of Nonlinear Science
22(6), 887–915.
Choi, H., Moin, P., and Kim, J. (1994), ‘Active turbulence control for drag reduction in wall-
bounded flows’, Journal of Fluid Mechanics 262, 75–110.
Chui, C. K. (1992), An Introduction to Wavelets, Academic Press, San Diego, CA.
Cimbala, J. M., Nagib, H. M., and Roshko, A. (1988), ‘Large structure in the far wakes of
two-dimensional bluff bodies’, Journal of Fluid Mechanics 190, 265–298.
Citriniti, J. H. and George, W. K. (2000), ‘Reconstruction of the global velocity field in the
axisymmetric mixing layer utilizing the proper orthogonal decomposition’, Journal of Fluid
Mechanics 418, 137–166.
Cohen, L. (1995), Time-Frequency Analysis, Prentice Hall, Englewood Cliffs, NJ.
Colabrese, S., Gustavsson, K., Celani, A., and Biferale, L. (2017), ‘Flow navigation by smart
microswimmers via reinforcement learning’, Physical Review Letters 118(15), 158004.
Colvert, B., Alsalman, M., and Kanso, E. (2018), ‘Classifying vortex wakes using neural
networks’, Bioinspiration & Biomimetics 13(2), 025003.
Cooley, J. W. and Tukey, J. W. (1965), ‘An algorithm for the machine calculation of complex
Fourier series’, Mathematics of Computation 19(90), 297–301.
Cordier, L., El Majd, B. A., and Favier, J. (2010), ‘Calibration of POD reduced-order models
using Tikhonov regularization’, International Journal for Numerical Methods in Fluids
63(2), 269–296.
Cordier, L., Noack, B. R., Daviller, G., Delvile, J., Lehnasch, G., Tissot, G., Balajewicz, M., and
Niven, R. (2013), ‘Control-oriented model identification strategy’, Experiments in Fluids
54, Article 1580.
Corino, E. R. and Brodkey, R. S. (1969), ‘A visual investigation of the wall region in turbulent
flow’, Journal of Fluid Mechanics 37, 1.
Cornejo Maceda, G. Y., Li, Y., Lusseyran, F., Morzyński, M., and Noack, B. R. (2021),
‘Stabilization of the fluidic pinball with gradient-based machine learning control’, Journal
of Fluid Mechanics 917, A42, 1–43.
Cornejo Maceda, G., Bernd R. Noack, Lusseyran, F., Deng, N., Pastur, L., and Morzyński,
M. (2019), ‘Artificial intelligence control applied to drag reduction of the fluidic pinball’,
Proceedings in Applied Mathematics and Mechanics 19(1), e201900268:1–2.
Corrsin, S. (1958), Local isotropy in turbulent shear flow, Res. Memo 58B11, National Advisory
Committee for Aeronautics, Washington, DC.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 416 Data-Driven Fluid Mechanics

Coussement, A., Gicquel, O., and Parente, A. (2013), ‘MG-local-PCA method for reduced order
combustion modeling’, Proceedings of the Combustion Institute 34(1), 1117–1123.
Coussement, A., Isaac, B. J., Gicquel, O., and Parente, A. (2016), ‘Assessment of different
chemistry reduction methods based on principal component analysis: Comparison of the
MG-PCA and score-PCA approaches’, Combustion and Flame 168, 83–97.
Coveney, P. V., Dougherty, E. R., and Highfield, R. R. (2016), ‘Big data need big theory too’,
Philosophical Transactions of the Royal Society A 374, 20160153.
Cox, T. F. and Cox, M. A. A. (2000), Multidimensional Scaling, Vol. 88 of Monographs on
Statistics and Applied Probability, 2nd ed., Chapman & Hall, London.
Cranmer, M. D., Xu, R., Battaglia, P., and Ho, S. (2019), ‘Learning symbolic physics with graph
networks’, arXiv preprint arXiv:1909.05862.
Cranmer, M., Greydanus, S., Hoyer, S., Battaglia, P., Spergel, D., and Ho, S. (2020),
‘Lagrangian neural networks’, arXiv preprint arXiv:2003.04630.
Cristianini, N. and Shawe–Taylor, J. (2000), An Introduction to Support Vector Machines and
Other Kernel-based Learning Methods, Cambridge University Press, Cambridge.
Cuoci, A., Frassoldati, A., Faravelli, T., and Ranzi, E. (2013), ‘Numerical modeling of laminar
flames with detailed kinetics based on the operator-splitting method’, Energy & Fuels
27(12), 7730–7753.
Cuoci, A., Frassoldati, A., Faravelli, T., and Ranzi, E. (2015), ‘OpenSMOKE++: An object-
oriented frame-work for the numerical modeling of reactive systems with detailed kinetic
mechanisms’, Computer Physics Communications 192, 237–264.
Dalakoti, D. K., Wehrfritz, A., Savard, B., Day, M. S., Bell, J. B., and Hawkes, E. R. (2020), ‘An
a priori evaluation of a principal component and artificial neural network based combustion
model in diesel engine conditions’, Proceedings of the Combustion Institute 38, 2701–2709.
D’Alessio, G., Attili, A., Cuoci, A., Pitsch, H., and Parente, A. (2020a), Analysis of turbulent
reacting jets via principal component analysis, in ‘Data Analysis for Direct Numerical
Simulations of Turbulent Combustion’, Springer, Cham, pp. 233–251.
D’Alessio, G., Attili, A., Cuoci, A., Pitsch, H., and Parente, A. (2020b), Unsupervised data
analysis of direct numerical simulation of a turbulent flame via local principal component
analysis and procustes analysis, in ‘15th International Workshop on Soft Computing Models
in Industrial and Environmental Applications’, September 16–18, 2020, Burgos, Spain,
Springer Nature, Switzerland, pp. 460–469.
D’Alessio, G., Cuoci, A., Aversano, G., Bracconi, M., Stagni, A., and Parente, A. (2020),
‘Impact of the partitioning method on multidimensional adaptive-chemistry simulations’,
Energies 13(10), 2567.
D’Alessio, G., Cuoci, A., and Parente, A. (2020), ‘OpenMORe: A Python framework for
reduction, clustering and analysis of reacting flow data’, SoftwareX 12, 100630.
D’Alessio, G., Parente, A., Stagni, A., and Cuoci, A. (2020), ‘Adaptive chemistry via pre-
partitioning of composition space and mechanism reduction’, Combustion and Flame
211, 68–82.
Dam, M., Brøns, M., Juul Rasmussen, J., Naulin, V., and Hesthaven, J. S. (2017), ‘Sparse
identification of a predator-prey system from simulation data of a convection model’, Physics
of Plasmas 24(2), 022310.
Daubechies, I. (1990), ‘The wavelet transform, time-frequency localization and signal analysis’,
IEEE Transactions on Information Theory 36(5), 961–1005.
Daubechies, I. (1992), Ten Lectures on Wavelets, Vol. 61, Society for Industrial and Applied
Mathematics, Philadelphia.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 References 417

Daubechies, I., Devore, R., Fornasier, M., and Güntürk, C. (2010), ‘Iteratively reweighted
least squares minimization for sparse recovery’, Communications on Pure and Applied
Mathematics 63, 1–38.
de Silva, B., Higdon, D. M., Brunton, S. L., and Kutz, J. N. (2019), ‘Discovery of physics from
data: Universal laws and discrepancy models’, arXiv preprint arXiv:1906.07906.
Deane, A. E., Kevrekidis, I. G., Karniadakis, G. E., and Orszag, S. A. (1991), ‘Low-
dimensional models for complex geometry flows: Application to grooved channels and
circular cylinders’, Physics of Fluids A 3, 2337–2354.
Deb, K. and Gupta, H. (2006), ‘Introducing robustness in multi-objective optimization’,
Evolutionary Computation 14, 463–494.
Debien, A., von Krbek, K. A., Mazellier, N., Duriez, T., Cordier, L., Noack, B. R., Abel, M. W.,
and Kourta, A. (2016), ‘Closed-loop separation control over a sharp edge ramp using genetic
programming’, Experiments in Fluids 57(3), 40.
del Álamo, J. C., Jiménez, J., Zandonade, P., and Moser, R. D. (2006), ‘Self-similar vortex
clusters in the logarithmic region’, Journal of Fluid Mechanics 561, 329–358.
Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L. (2009), Imagenet: A large-scale
hierarchical image database, in ‘2009 IEEE Conference on Computer Vision and Pattern
Recognition’, June 20–25, 2009, Miami, FL, IEEE, pp. 248–255.
Deng, N., Noack, B. R., Morzyński, M., and Pastur, L. R. (2020), ‘Low-order model for
successive bifurcations of the fluidic pinball’, Journal of Fluid Mechanics 884, A37.
Discetti, S., Bellani, G., Örlü, R., Serpieri, J., Vila, C. S., Raiola, M., Zheng, X., Mascotelli,
L., Talamelli, A., and Ianiro, A. (2019), ‘Characterization of very-large-scale motions in
high-Re pipe flows’, Experimental Thermal and Fluid Science 104, 1–8.
Discetti, S. and Ianiro, A. (2017), Experimental Aerodynamics, CRC Press, Boca Raton, FL.
Discetti, S., Raiola, M., and Ianiro, A. (2018), ‘Estimation of time-resolved turbulent fields
through correlation of non-time-resolved field measurements and time-resolved point
measurements’, Experimental Thermal and Fluid Science 93, 119–130.
Distefano, J. (2013), Schaum’s Outline of Feedback and Control Systems, McGraw-Hill
Education, New York.
Dong, C., Loy, C. C., He, K., and Tang, X. (2014), Learning a deep convolutional network for
image super-resolution, in ‘European Conference on Computer Vision’, September 6–12,
2014, Zurich, Switzerland, Springer, Cham, pp. 184–199.
Dong, S., Lozano-Durán, A., Sekimoto, A., and Jiménez, J. (2017), ‘Coherent structures
in statistically stationary homogeneous shear turbulence’, Journal of Fluid Mechanics
816, 167–208.
Donoho, D. L. (1995), ‘De-noising by soft-thresholding’, IEEE Transactions on Information
Theory 41(3), 613–627.
Donoho, D. L. (2006), ‘Compressed sensing’, IEEE Transactions on Information Theory
52(4), 1289–1306.
Donoho, D. L. and Johnstone, J. M. (1994), ‘Ideal spatial adaptation by wavelet shrinkage’,
Biometrika 81(3), 425–455.
Dracopoulos, D. C. (1997), Evolutionary Learning Algorithms for Neural Adaptive Control,
Perspectives in Neural Computing, Springer-Verlag, London.
Drazin, P. and Reid, W. (1981), Hydrodynamic Stability, Cambridge University Press,
Cambridge, UK.
Dullerud, G. E. and Paganini, F. (2000), A Course in Robust Control Theory: A Convex
Approach, Texts in Applied Mathematics, Springer, Berlin, Heidelberg.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 418 Data-Driven Fluid Mechanics

Duraisamy, K., Iaccarino, G., and Xiao, H. (2019), ‘Turbulence modeling in the age of data’,
Annual Reviews of Fluid Mechanics 51, 357–377.
Duriez, T., Brunton, S. L., and Noack, B. R. (2017), Machine Learning Control: Taming
Nonlinear Dynamics and Turbulence, Springer, Cham.
Duwig, C. and Iudiciani, P. (2010), ‘Extended proper orthogonal decomposition for analysis of
unsteady flames’, Flow, Turbulence and Combustion 84(1), 25.
Echekki, T., Kerstein, A. R., and Sutherland, J. C. (2011), The one-dimensional-turbulence
model, in T. Echekki & E. Mastorakos, eds., ‘Turbulent Combustion Modeling’, Springer,
Dordrecht, pp. 249–276.
Echekki, T. and Mirgolbabaei, H. (2015), ‘Principal component transport in turbulent combus-
tion: A posteriori analysis’, Combustion and Flame 162(5), 1919–1933.
Eckart, C. and Young, G. (1936), ‘The approximation of one matrix by another of lower rank’,
Psychometrika 1(3), 211–218.
Ehlert, A., Nayeri, C. N., Morzyński, M., and Noack, B. R. (2020), ‘Locally linear embedding
for transient cylinder wakes’. arXiv manuscript 1906.07822 [physics.flu-dyn].
El Sayed M, Y., Semaan, R., and Radespiel, R. (2018), Sparse modeling of the lift gains of a
high-lift configuration with periodic coanda blowing, in ‘2018 AIAA Aerospace Sciences
Meeting’, January 8–12, 2018, Kissimmee, FL, p. 1054.
Elman, J. L. (1990), ‘Finding structure in time’, Cognitive Science 14(2), 179–211.
Encinar, M. P. and Jiménez, J. (2020), ‘Momentum transfer by linearised eddies in turbulent
channel flows’, arXiv e-prints p. 1911.06096.
Engstrom, L., Ilyas, A., Santurkar, S., Tsipras, D., Janoos, F., Rudolph, L., and Madry,
A. (2020), Implementation matters in deep RL: A case study on PPO and TRPO, in
‘International Conference on Learning Representations’ April 26–30, 2020, Addis Ababa,
Ethiopia.
Epps, B. P. and Krivitzky, E. M. (2019), ‘Singular value decomposition of noisy data: Noise
filtering’, Experiments in Fluids 60(8), 126.
Erichson, N. B., Mathelin, L., Kutz, J. N., and Brunton, S. L. (2019), ‘Randomized dynamic
mode decomposition’, SIAM Journal on Applied Dynamical Systems 18(4), 1867–1891.
Erichson, N. B., Mathelin, L., Yao, Z., Brunton, S. L., Mahoney, M. W., and Kutz, J. N. (2020),
‘Shallow neural networks for fluid flow reconstruction with limited sensors’, Proceedings of
the Royal Society A 476(2238), 20200097.
Espeholt, L., Soyer, H., Munos, R., Simonyan, K., Mnih, V., Ward, T., Doron, Y., Firoiu,
V., Harley, T., Dunning, I., Legg, S., and Kavukcuoglu, K. (2018), IMPALA: Scalable
distributed deep-RL with importance weighted actor-learner architectures, in J. Dy and
A. Krause, eds., ‘Proceedings of the 35th International Conference on Machine Learning’,
Vol. 80 of Proceedings of Machine Learning Research, PMLR, Stockholmsmassan, Stock-
holm, Sweden, pp. 1407–1416.
Esposito, C., Mendez, M., Gouriet, J., Steelant, J., and Vetrano, M. (2021), ‘Spectral and modal
analysis of a cavitating flow through an orifice’, Experimental Thermal and Fluid Science
121, 110251.
Evans, W. R. (1948), ‘Graphical analysis of control systems’, Transactions of the American
Institute of Electrical Engineers 67(1), 547–551.
Everitt, B. and Skrondal, A. (2002), The Cambridge Dictionary of Statistics, Vol. 106,
Cambridge University Press, Cambridge.
Ewing, D. and Citriniti, J. H. (1999), Examination of a LSE/POD complementary technique
using single and multi-time information in the axisymmetric shear layer, in ‘IUTAM

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 References 419

Symposium on Simulation and Identification of Organized Structures in Flows’, May 25–29,

1997, Lyngby, Denmark, Springer, Dordrecht, pp. 375–384.
Fabbiane, N., Semeraro, O., Bagheri, S., and Henningson, D. S. (2014), ‘Adaptive and model-
based control theory applied to convectively unstable flows’, Applied Mechanics Reviews
66(6), 060801-1–060801-20.
Farge, M. (1992), ‘Wavelet transforms and their applications to turbulence’, Annual Review of
Fluid Mechanics 24(1), 395–458.
Fernex, D., Noack, B. R., and Semaan, R. (2021), ‘Cluster-based network models – From
snapshots to complex dynamical systems’, Science Advances 7(25).
Fernex, D., Semann, R., Albers, M., Meysonnat, P. S., Schröder, W., Ishar, R., Kaiser, E., and
Noack, B. R. (2019), ‘Cluster-based network model for drag reduction mechanisms of an
actuated turbulent boundary layer’, Proceedings in Applied Mathematics and Mechanics
19(1), 1–2.
Ferziger, J. H., Perić, M., and Street, R. L. (2002), Computational Methods for Fluid Dynamics,
Vol. 3, Springer, Berlin.
Feynman, R. P., Leighton, R. B., and Sands, M. (2013), The Feynman Lectures on Physics,
Vol. 2, Basic Books, New York.
Finzi, M., Wang, K. A., and Wilson, A. G. (2020), ‘Simplifying Hamiltonian and Lagrangian
neural networks via explicit constraints’, Advances in Neural Information Processing
Systems 33, 17605–17616.
Fleming, P. J. and Purshouse, R. C. (2002), ‘Evolutionary algorithms in control systems
engineering: A survey’, Control Engineering Practice 10, 1223–1241.
Fletcher, C. A. J. (1984), Computational Galerkin Methods, 1st ed., Springer, New York.
Flierl, G. R., Larichev, V. D., McWilliams, J. C., and Reznik, G. M. (1980), ‘The dynamics of
baroclinic and barotropic solitary eddies’, Dynamics of Atmospheres and Oceans 5, 1–41.
Flinois, T. L. and Morgans, A. S. (2016), ‘Feedback control of unstable flows: A direct mod-
elling approach using the eigensystem realisation algorithm’, Journal of Fluid Mechanics
793, 41–78.
Flores, O. and Jiménez, J. (2010), ‘Hierarchy of minimal flow units in the logarithmic layer’,
Physics of Fluids 22, 071704.
Floryan, D. and Graham, M. D. (2021), ‘Discovering multiscale and self-similar structure
with data-driven wavelets’, Proceedings of the National Academy of Sciences 118(1),
e2021299118.
Föppl, L. (1913), ‘Wirbelbewegung hinter einem kreiszylinder (Vortex movement behind a
circular cylinder)’, Bayerische Akademie der Wissenschaften, Munich.
Fornberg, B. (1977), ‘A numerical study of 2-D turbulence’, Journal of Computational Physics
25, 1–31.
Fosas de Pando, M., Schmid, P., and Sipp, D. (2014), ‘A global analysis of tonal noise in flows
around aerofoils’, Journal of Fluid Mechanics 754, 5–38.
Fosas de Pando, M., Schmid, P., and Sipp, D. (2017), ‘On the receptivity of aerofoil tonal noise:
An adjoint analysis’, Journal of Fluid Mechanics 812, 771–791.
Fosas de Pando, M., Sipp, D., and Schmid, P. (2012), ‘Efficient evaluation of the direct and
adjoint linearized dynamics from compressible flow solvers’, Journal of Computational
Physics 231(23), 7739–7755.
Franklin, G. F., Powell, J. D., and Emami-Naeini, A. (1994), Feedback Control of Dynamic
Systems, Vol. 3, Addison-Wesley, Reading, MA.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 420 Data-Driven Fluid Mechanics

Franz, T. and Held, M. (2017), Data fusion of CFD solutions and experimental aerodynamic
data, in ‘Proceedings Onera-DLR Aerospace Symposium (ODAS)’ June 7–9, 2017, Aussois,
France.
Franz, T., Zimmermann, R., and Görtz, S. (2017), Adaptive sampling for nonlinear dimension-
ality reduction based on manifold learning, in P. Benner, M. Ohlberger, A. Patera, G. Rozza,
and K. Urban, eds., Model Reduction of Parametrized Systems, Springer, Cham, pp. 255–
269.
Franz, T., Zimmermann, R., Görtz, S., and Karcher, N. (2014), ‘Interpolation-based reduced-
order modelling for steady transonic flows via manifold learning’, International Journal of
Computational Fluid Dynamics 28(3–4), 106–121.
Frassoldati, A., Faravelli, T., and Ranzi, E. (2003), ‘Kinetic modeling of the interactions
between NO and hydrocarbons at high temperature’, Combustion and Flame 135, 97–112.
Freeman, W. T., Jones, T. R., and Pasztor, E. C. (2002), ‘Example-based super-resolution’, IEEE
Computer Graphics and Applications 22(2), 56–65.
Freymuth, P. (1966), ‘On transition in a separated laminar boundary layer’, Journal of Fluid
Mechanics 25, 683–704.
Friedman, J. (1991), ‘Multivariate adaptive regression splines’, The Annals of Statistics
19(1), 1–67.
Fu, J., Luo, K., and Levine, S. (2018), Learning robust rewards with adverserial inverse
reinforcement learning, in ‘6th International Conference on Learning Representations, ICLR
2018, Vancouver, BC, Canada, April 30–May 3, 2018, Conference Track Proceedings’,
OpenReview.net.
Fukami, K., Fukagata, K., and Taira, K. (2018), ‘Super-resolution reconstruction of turbulent
flows with machine learning’, arXiv preprint arXiv:1811.11328.
Gabor, D. (1946), ‘Theory of communication. Part 1: The analysis of information’, Journal
of the Institution of Electrical Engineers-Part III: Radio and Communication Engineering
93(26), 429–441.
Gad-el Hak, M. (2007), Flow Control: Passive, Active, and Reactive Flow Management,
Cambridge University Press, Cambridge.
Galerkin, B. G. (1915), ‘Rods and plates: Series occuring in various questions regarding the
elastic equilibrium of rods and plates (translated)’, Vestn. Inzhen. 19, 897—908.
Galletti, G., Bruneau, C. H., Zannetti, L., and Iollo, A. (2004), ‘Low-order modelling of laminar
flow regimes past a confined square cylinder’, Journal of Fluid Mechanics 503, 161–170.
Gardiner, W. C., Lissianski, V. V., Qin, Z., Smith, G. P., Golden, D. M., Frenklach, M., Moriarty,
N. W., Eiteneer, B., Goldenberg, M., Bowman, C. T., Hanson, R. K., and Song Jr., S. (2012),
‘Gri-mech 3.0’.
Gardner, E. (1988), ‘The space of interactions in neural network models’, Journal of Physics A:
Mathematical and General 21(1), 257.
Garnier, P., Viquerat, J., Rabault, J., Larcher, A., Kuhnle, A., and Hachem, E. (2019), ‘A review
on deep reinforcement learning for fluid mechanics’, arXiv preprint arXiv:1908.04127.
Gaster, M., Kit, E., and Wygnanski, I. (1985), ‘Large-scale structures in a forced turbulent
mixing layer’, Journal of Fluid Mechanics 150, 23–39.
Gautier, N., Aider, J.-L., Duriez, T., Noack, B., Segond, M., and Abel, M. (2015), ‘Closed-loop
separation control using machine learning’, Journal of Fluid Mechanics 770, 442–457.
Gazzola, M., Hejazialhosseini, B., and Koumoutsakos, P. (2014), ‘Reinforcement learning and
wavelet adapted vortex methods for simulations of self-propelled swimmers’, SIAM Journal
on Scientific Computing 36(3), B622–B639.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 References 421

Gazzola, M., Tchieu, A. A., Alexeev, D., de Brauer, A., and Koumoutsakos, P. (2016), ‘Learning
to school in the presence of hydrodynamic interactions’, Journal of Fluid Mechanics
789, 726–749.
Gerhard, J., Pastoor, M., King, R., Noack, B. R., Dillmann, A., Morzyński, M., and Tadmor, G.
(2003), Model-based control of vortex shedding using low-dimensional Galerkin models, in
‘33rd AIAA Fluids Conference and Exhibit’, Orlando, FL, June 23–26, 2003. Paper 2003-
4262.
Germano, M., Piomelli, U., Moin, P., and Cabot, W. H. (1991), ‘A dynamic subgrid-scale eddy
viscosity model’, Physics of Fluids A: Fluid Dynamics 3(7), 1760–1765.
Giannetti, F. and Luchini, P. (2007), ‘Structural sensitivity of the first instability of the cylinder
wake’, Journal of Fluid Mechanics 581, 167–197.
Glaz, B., Liu, L., and Friedmann, P. P. (2010), ‘Reduced-order nonlinear unsteady aerodynamic
modeling using a surrogate-based recurrence framework’, AIAA Journal 48(10), 2418–2429.
Goldstein, T. and Osher, S. (2009), ‘The split Bregman method for L1 -regularized problems’,
SIAM Journal on Imaging Sciences 2(2), 323–343.
Gonzalez-Garcia, R., Rico-Martinez, R., and Kevrekidis, I. (1998), ‘Identification of distributed
parameter systems: A neural net based approach’, Computers & Chemical Engineering
22, S965–S968.
Gonzalez, R. C. and Woods, R. E. (2017), Digital Image Processing, 4th ed., Pearson, New
York.
Goodfellow, I., Bengio, Y., and Courville, A. (2016), Deep Learning, MIT Press, Cambridge,
MA.
Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville,
A., and Bengio, Y. (2014), Generative adversarial nets, in ‘Advances in Neural Information
Processing Systems’, December 8–13, 2014, Montreal, Canada, pp. 2672–2680.
Görtz, S., Ilic, C., Abu-Zurayk, M., Liepelt, R., Jepsen, J., Führer, T., Becker, R., Scherer,
J., Kier, T., and Siggel, M. (2016), Collaborative multi-level MDO process development
and application to long-range transport aircraft, in ‘Proceedings of the 30th Congress of
the International Council of the Aeronautical Sciences (ICAS)’, September 25–30, 2016,
Daejeon, Korea.
Gray, R. M. (2005), ‘Toeplitz and circulant matrices: A review’, Foundations and Trends R
in
Communications and Information Theory 2(3), 155–239.
Greydanus, S., Dzamba, M., and Yosinski, J. (2019), ‘Hamiltonian neural networks’, Advances
in Neural Information Processing Systems 32, 15379–15389.
Griewank, A. and Walther, A. (2000), ‘Algorithm 799: Revolve: An implementation of
checkpointing for the reverse or adjoint mode of computational differentiation’, ACM
Transactions on Mathematical Software 26(1), 19–45.
Grossmann, A. and Morlet, J. (1984), ‘Decomposition of hardy functions into square integrable
wavelets of constant shape’, SIAM Journal on Mathematical Analysis 15(4), 723–736.
Gu, S., Lillicrap, T., Sutskever, I., and Levine, S. (2016), Continuous deep Q-learning with
model-based acceleration, Vol. 48 of Proceedings of Machine Learning Research, PMLR,
New York, pp. 2829–2838.
Guan, Y., Brunton, S. L., and Novosselov, I. (2020), ‘Sparse nonlinear models of chaotic
electroconvection’, arXiv preprint arXiv:2009.11862.
Guastoni, L., Güemes, A., Ianiro, A., Discetti, S., Schlatter, P., Azizpour, H., and Vinuesa,
R. (2020), ‘Convolutional-network models to predict wall-bounded turbulence from wall
quantities’, arXiv preprint arXiv:2006.12483.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 422 Data-Driven Fluid Mechanics

Guckenheimer, J. and Holmes, P. (1983), Nonlinear Oscillations, Dynamical Systems, and

Bifurcations of Vector Fields, Springer Verlag, New York.
Güemes, A., Discetti, S., and Ianiro, A. (2019), ‘Sensing the turbulent large-scale motions with
their wall signature’, Physics of Fluids 31(12), 125112.
Guidorzi, R. (2003), Multivariable System Identification. From Observations to Models,
Bononia University Press, Bologna.
Gunzburger, M. (2003), Perspectives in Flow Control and Optimization, Society for Industrial
and Applied Mathematics, Philadelphia.
Ha, D. and Schmidhuber, J. (2018), ‘World models’, CoRR, arXiv:1803.10122.
Haar, A. (1911), ‘Zur theorie der orthogonalen funktionensysteme’, Mathematische Annalen
71(1), 38–53.
Haarnoja, T., Zhou, A., Abbeel, P., and Levine, S. (2018), Soft actor-critic: Off-policy maximum
entropy deep reinforcement learning with a stochastic actor, in J. Dy and A. Krause,
eds, ‘Proceedings of the 35th International Conference on Machine Learning’, Vol. 80
of Proceedings of Machine Learning Research, PMLR, Stockholmsmassan, Stockholm,
Sweden, pp. 1861–1870.
Halko, N., Martinsson, P.-G., Shkolnisky, Y., and Tygert, M. (2011), ‘An algorithm for the
principal component analysis of large data sets’, SIAM Journal on Scientific Computing
33, 2580–2594.
Halko, N., Martinsson, P.-G., and Tropp, J. A. (2011), ‘Finding structure with randomness:
Probabilistic algorithms for constructing approximate matrix decompositions’, SIAM Review
53(2), 217–288.
Hama, F. R. (1962), ‘Streaklines in a perturbed shear flow’, The Physics of Fluids 5(6), 644–
650.
Hamilton, J. M., Kim, J., and Waleffe, F. (1995), ‘Regeneration mechanisms of near-wall
turbulence structures’, Journal of Fluid Mechanics 287, 317–348.
Han, Z.-H., Abu-Zurayk, M., Görtz, S., and Ilic, C. (2018), Surrogate-based aerodynamic shape
optimization of a wing-body transport aircraft configuration, in ‘Notes on Numerical Fluid
Mechanics and Multidisciplinary Design’, October 13–14, 2015, Braunschweig, Germany,
Springer, Berlin, pp. 257–282.
Han, Z.-H. and Görtz, S. (2012a), ‘Alternative cokriging method for variable-fidelity surrogate
modeling’, AIAA Journal 50(5), 1205–1210.
Han, Z.-H. and Görtz, S. (2012b), ‘Hierarchical kriging model for variable-fidelity surrogate
modeling’, AIAA Journal 50(9), 1885–1896.
Han, Z.-H., Görtz, S., and Zimmermann, R. (2013), ‘Improving variable-fidelity surrogate mod-
eling via gradient-enhanced kriging and a generalized hybrid bridge function’, Aerospace
Science and Technology 25(1), 177–189.
Hansen, N., Niederberger, A. S., Guzzella, L., and Koumoutsakos, P. (2009), ‘A method for
handling uncertainty in evolutionary optimization with an application to feedback control of
combustion’, IEEE Transactions on Evolutionary Computation 13(1), 180–197.
Harris, F. (1978), ‘On the use of windows for harmonic analysis with the discrete Fourier
transform’, Proceedings of the IEEE 66(1), 51–83.
Hasselmann, K. (1988), ‘PIPs and POPs: The reduction of complex dynamical systems
using principal interaction and oscillation patterns’, Journal of Geophysical Research
93(D9), 11015.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 References 423

Hasselt, H. v., Guez, A., and Silver, D. (2016), Deep reinforcement learning with double
Q-learning, in ‘Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence’,
AAAI 16, February 12–17, 2016, Phoenix, AZ, AAAI Press, Palo Alto, CA, pp. 2094–2100.
Hastie, T., Tibshirani, R., and Friedman, J. (2009), The Elements of Statistical Learning, Vol. 2,
Springer, New York.
Hayes, M. (2011), Schaums Outline of Digital Signal Processing, McGraw-Hill Education, New
York.
Hemati, M., Rowley, C., Deem, E., and Cattafesta, L. (2017), ‘De-biasing the dynamic mode
decomposition for applied Koopman spectral analysis’, Theoretical and Computational
Fluid Dynamics 31(4), 349–368.
Henderson, P., Islam, R., Bachman, P., Pineau, J., Precup, D., and Meger, D. (2018), Deep
reinforcement learning that matters, in S. A. McIlraith and K. Q. Weinberger, eds.,
‘Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence, (AAAI-18),
the 30th Innovative Applications of Artificial Intelligence (IAAI-18), and the 8th AAAI
Symposium on Educational Advances in Artificial Intelligence (EAAI-18), New Orleans,
Louisiana, USA, February 2–7, 2018’, AAAI Press, Palo Alto, CA, pp. 3207–3214.
Henningson, D. (2010), ‘Description of complex flow behaviour using global and dynamic
modes’, Journal of Fluid Mechanics 656, 1–4.
Herbert, T. (1988), ‘Secondary instability of boundary layers’, Annual Review of Fluid
Mechanics 20, 487–526.
Hernán, M. A. and Jiménez, J. (1982), ‘Computer analysis of a high-speed film of the plane
turbulent mixing layer’, Journal of Fluid Mechanics 119, 323–345.
Hervé, A., Sipp, D., Schmid, P. J. and Samuelides, M. (2012), ‘A physics-based approach to
flow control using system identification’, Journal of Fluid Mechanics 702, 26–58.
Hill, D. (1995), ‘Adjoint systems and their role in the receptivity problem for boundary layers’,
Journal of Fluid Mechanics 292, 183–204.
Hinton, G. E. (2006), ‘Reducing the dimensionality of data with neural networks’, Science
313(5786), 504–507.
Hinton, G. E. and Sejnowski, T. J. (1986), ‘Learning and releaming in Boltzmann machines’,
Parallel Distributed Processing: Explorations in the Microstructure of Cognition 1(282–
317), 2.
Hinze, M. and Kunisch, K. (2000), ‘Three control methods for time-dependent fluid flow’, Flow,
Turbulence and Combustion 65, 273–298.
Ho, B. L. and Kalman, R. E. (1965), Effective construction of linear state-variable models from
input/output data, in ‘Proceedings of the 3rd Annual Allerton Conference on Circuit and
System Theory’, October 20–22, 1965, Monticello, IL, pp. 449–459.
Ho, J. and Ermon, S. (2016), Generative adversarial imitation learning, in D. D. Lee,
M. Sugiyama, U. V. Luxburg, I. Guyon, and R. Garnett, eds., ‘Advances in Neural
Information Processing Systems 29’, December 5–10, 2016, Barcelona, Spain, Curran
Associates, Red Hook, NY, pp. 4565–4573.
Hochbruck, M. and Ostermann, A. (2010), ‘Exponential integrators’, Acta Numerica 19, 209–
286.
Hochreiter, S. and Schmidhuber, J. (1997), ‘Long short-term memory’, Neural Computation
9(8), 1735–1780.
Hof, B., Van Doorne, C. W., Westerweel, J., Nieuwstadt, F. T., Faisst, H., Eckhardt, B., Wedin,
H., Kerswell, R. R., and Waleffe, F. (2004), ‘Experimental observation of nonlinear traveling
waves in turbulent pipe flow’, Science 305(5690), 1594–1598.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 424 Data-Driven Fluid Mechanics

Hoffmann, M., Fröhner, C., and Noé, F. (2019), ‘Reactive SINDy: Discovering governing
reactions from concentration data’, Journal of Chemical Physics 150, 025101.
Holland, J. H. (1975), Adaptation in Natural and Artificial Systems: An Introductory Analysis
with Applications to Biology, Control, and Artificial Intelligence, University of Michigan
Press, Ann Arbor.
Holmes, P., Lumley, J. L., Berkooz, G., and Rowley, C. W. (2012), Turbulence, Coherent
Structures, Dynamical Systems and Symmetry, 2nd paperback ed., Cambridge University
Press, Cambridge.
Hopf, E. (1948), ‘A mathematical example displaying features of turbulence’, Communications
on Pure and Applied Mathematics 1(4), 303–322.
Hopfield, J. J. (1982), ‘Neural networks and physical systems with emergent collective
computational abilities’, Proceedings of the National Academy of Sciences 79(8), 2554–
2558.
Horn, R. A. and Johnson, C. R. (2012), Matrix Analysis, Cambridge University Press,
Cambridge.
Hornik, K., Stinchcombe, M., and White, H. (1989), ‘Multilayer feedforward networks are
universal approximators’, Neural Networks 2(5), 359–366.
Hosseini, Z., Martinuzzi, R. J., and Noack, B. R. (2015), ‘Sensor-based estimation of the
velocity in the wake of a low-aspect-ratio pyramid’, Experiments in Fluids 56(1), 13.
Hou, W., Darakananda, D., and Eldredge, J. (2019), Machine learning based detection of flow
disturbances using surface pressure measurements, in ‘AIAA Scitech 2019 Forum’, January
7–11, 2019, San Diego, CA, p. 1148.
Hoyas, S. and Jiménez, J. (2006), ‘Scaling of the velocity fluctuations in turbulent channels up
to Reτ = 2003’, Physics of Fluids 18, 011702.
Hsu, H. (2013), Schaum’s Outline of Signals and Systems, 3rd ed. (Schaum’s Outlines),
McGraw-Hill Education, New York.
Huang, Z., Du, W., and Chen, B. (2005), Deriving private information from randomized data,
in ‘Proceedings of the 2005 ACM SIGMOD International Conference on Management of
Data’, June 14–16, 2005, Baltimore, MD, pp. 37–48.
Hwangbo, J., Lee, J., Dosovitskiy, A., Bellicoso, D., Tsounis, V., Koltun, V., and Hutter,
M. (2019), ‘Learning agile and dynamic motor skills for legged robots’, arXiv preprint
arXiv:1901.08652.
Illingworth, S. J. (2016), ‘Model-based control of vortex shedding at low Reynolds numbers’,
Theoretical and Computational Fluid Dynamics 30(5), 1–20.
Illingworth, S. J., Morgans, A. S., and Rowley, C. W. (2010), ‘Feedback control of flow reso-
nances using balanced reduced-order models’, Journal of Sound and Vibration 330(8), 1567–
1581.
Illingworth, S. J., Morgans, A. S., and Rowley, C. W. (2012), ‘Feedback control of cavity flow
oscillations using simple linear models’, Journal of Fluid Mechanics 709, 223–248.
Illingworth, S. J., Naito, H., and Fukagata, K. (2014), ‘Active control of vortex shedding: An
explanation of the gain window’, Physical Review E 90(4), 043014.
Ingle, V. K. and Proakis, J. G. (2011), Digital Signal Processing Using MATLAB, Cengage
Learning, Boston.
Isaac, B. J., Coussement, A., Gicquel, O., Smith, P. J., and Parente, A. (2014), ‘Reduced-order
PCA models for chemical reacting flows’, Combustion and Flame 161(11), 2785–2800.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 References 425

Isaac, B. J., Thornock, J. N., Sutherland, J., Smith, P. J., and Parente, A. (2015), ‘Advanced
regression methods for combustion modelling using principal components’, Combustion and
Flame 162(6), 2592–2601.
Jackson, C. P. (1987), ‘A finite-element study of the onset of vortex shedding in flow past
variously shaped bodies’, Journal of Fluid Mechanics 182, 23–45.
Jaderberg, M., Mnih, V., Czarnecki, W. M., Schaul, T., Leibo, J. Z., Silver, D., and Kavukcuoglu,
K. (2017), Reinforcement learning with unsupervised auxiliary tasks, in ‘5th International
Conference on Learning Representations, ICLR 2017, Toulon, France, April 24–26, 2017,
Conference Track Proceedings’, OpenReview.net.
James, G., Witten, D., Hastie, T., and Tibshirani, R. (2013), An Introduction to Statistical
Learning, Springer, New York.
Jho, H. (2018), ‘Beautiful physics: Re-vision of aesthetic features of science through the
literature review’, The Journal of the Korean Physical Society 73, 401–413.
Jiménez, J. (2002), Turbulence in ‘Perspectives in Fluid Dynamics’, Cambridge University
Press, Cambridge.
Jiménez, J. (2013), ‘How linear is wall-bounded turbulence?’, Physics of Fluids 25, 110814.
Jiménez, J. (2018a), ‘Coherent structures in wall-bounded turbulence’, Journal of Fluid
Mechanics 842, P1.
Jiménez, J. (2018b), ‘Machine-aided turbulence theory’, Journal of Fluid Mechanics 854, R1.
Jiménez, J. (2020a), ‘Computers and turbulence’, The European Journal of Mechanics -
B/Fluids 79, 1–11.
Jiménez, J. (2020b), ‘Monte Carlo science’, Journal of Turbulence 21(9–10), 544–566.
Jiménez, J., Kawahara, G., Simens, M. P., Nagata, M., and Shiba, M. (2005), ‘Characterization
of near-wall turbulence in terms of equilibrium and “bursting” solutions’, Physics of Fluids
17, 015105.
Jiménez, J. and Moin, P. (1991), ‘The minimal flow unit in near-wall turbulence’, Journal of
Fluid Mechanics 225, 213–240.
Jiménez, J. and Pinelli, A. (1999), ‘The autonomous cycle of near-wall turbulence’, Journal of
Fluid Mechanics 389, 335–359.
Jiménez, J. and Simens, M. P. (2001), ‘Low-dimensional dynamics of a turbulent wall flow’,
Journal of Fluid Mechanics 435, 81–91.
Jiménez, J., Wray, A. A., Saffman, P. G., and Rogallo, R. S. (1993), ‘The structure of intense
vorticity in isotropic turbulence’, Journal of Fluid Mechanics 255, 65–90.
Johannink, T., Bahl, S., Nair, A., Luo, J., Kumar, A., Loskyll, M., Ojea, J. A., Solowjow, E., and
Levine, S. (2019), Residual reinforcement learning for robot control, in ‘2019 International
Conference on Robotics and Automation (ICRA)’, May 20–24, 2019, Montreal, Canada,
IEEE, pp. 6023–6029.
Jolliffe, I. (2002), Principal Component Analysis, Springer-Verlag, New York.
Jordan, D. and Smith, P. (1988), Nonlinear Ordinary Differential Equations, Clarendon Press,
Oxford.
Joseph, D. (1976), Stability of Fluid Motions I, Springer Verlag, New York.
Joseph, D. and Carmi, S. (1969), ‘Stability of Poiseuille flow in pipes, annuli and channels’,
Quarterly of Applied Mathematics 26, 575–599.
Jovanović, M. R., Schmid, P. J., and Nichols, J. W. (2014), ‘Sparsity-promoting dynamic mode
decomposition’, Physics of Fluids 26(2), 024103.
Juang, J. N. (1994), Applied System Identification, Prentice Hall PTR, Upper Saddle River.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 426 Data-Driven Fluid Mechanics

Juang, J. N. and Pappa, R. S. (1985), ‘An eigensystem realization algorithm for modal parameter
identification and model reduction’, Journal of Guidance, Control, and Dynamics 8(5), 620–
627.
Juang, J. N., Phan, M., Horta, L. G., and Longman, R. W. (1991), Identification of
observer/Kalman filter Markov parameters: Theory and experiments, Technical Memoran-
dum 104069, NASA.
Kaiser, E., Kutz, J. N., and Brunton, S. L. (2018), ‘Sparse identification of nonlinear dynamics
for model predictive control in the low-data limit’, Proceedings of the Royal Society of
London A 474(2219), 20180335.
Kaiser, E., Li, R., and Noack, B. R. (2017), On the control landscape topology, in ‘The 20th
World Congress of the International Federation of Automatic Control (IFAC)’, Toulouse,
France, pp. 1–4.
Kaiser, E., Noack, B. R., Cordier, L., Spohn, A., Segond, M., Abel, M., Daviller, G., Osth, J.,
Krajnovic, S., and Niven, R. K. (2014), ‘Cluster-based reduced-order modelling of a mixing
layer’, Journal of Fluid Mechanics 754, 365–414.
Kaiser, E., Noack, B. R., Spohn, A., Cattafesta, L. N., and Morzyński, M. (2017a), ‘Cluster-
based control of a separating flow over a smoothly contoured ramp’, Theoretical and
Computational Fluid Dynamics 31(5–6), 579–593.
Kaiser, E., Noack, B. R., Spohn, A., Cattafesta, L. N., and Morzyński, M. (2017b), ‘Cluster-
based control of nonlinear dynamics’, Theoretical and Computational Fluid Dynamics 31(5–
6), 1579–593.
Kaiser, G. (2010), A Friendly Guide to Wavelets, Springer Science & Business Media, Boston.
Kaiser, H. F. (1958), ‘The varimax criterion for analytic rotation in factor analysis’, Psychome-
trika 23(3), 187–200.
Kambhatla, N. and Leen, T. K. (1997), ‘Dimension reduction by local principal component
analysis’, Neural Computation 9(7), 1493–1516.
Kaptanoglu, A. A., Morgan, K. D., Hansen, C. J., and Brunton, S. L. (2020),
‘Physics-constrained, low-dimensional models for MHD: First-principles and data-driven
approaches’, arXiv preprint arXiv:2004.10389 .
Kawahara, G. (2005), ‘Laminarization of minimal plane Couette flow: Going beyond the basin
of attraction of turbulence’, Physics of Fluids 17, 041702.
Kawahara, G., Uhlmann, M., and van Veen, L. (2012), ‘The significance of simple invariant
solutions in turbulent flows’, Annual Review of Fluid Mechanics 44, 203–225.
Kee, R. J., Coltrin, M. E., and Glarborg, P. (2005), Chemically Reacting Flow: Theory and
Practice, John Wiley & Sons, Hoboken, NJ.
Keefe, L., Moin, P., and Kim, J. (1992), ‘The dimension of attractors underlying periodic
turbulent Poiseuille flow’, Journal of Fluid Mechanics 242, 1–29.
Kenney, J. and Keeping, E. (1951), Mathematics of Statistics, Vol. II, D. Van Nostrand Co., New
York.
Kerhervé, F., Roux, S., and Mathis, R. (2017), ‘Combining time-resolved multi-point and
spatially-resolved measurements for the recovering of very-large-scale motions in high
Reynolds number turbulent boundary layer’, Experimental Thermal and Fluid Science
82, 102–115.
Kerstein, A. R. (1999), ‘One-dimensional turbulence: Model formulation and application
to homogeneous turbulence, shear flows, and buoyant stratified flows’, Journal of Fluid
Mechanics 392, 277–334.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 References 427

Kerstens, W., Pfeiffer, J., Williams, D., King, R., and Colonius, T. (2011), ‘Closed-loop
control of lift for longitudinal gust suppression at low Reynolds numbers’, AIAA Journal
49(8), 1721–1728.
Khalil (2002), Nonlinear Systems, 3rd ed., Dover, I, New York.
Khalil, H. K. and Grizzle, J. W. (2002), Nonlinear Systems, Vol. 3, Prentice Hall, Upper Saddle
River, NJ.
Kim, H. J., Jordan, M. I., Sastry, S., and Ng, A. Y. (2004), Autonomous helicopter flight via
reinforcement learning, in ‘Advances in Neural Information Processing Systems’, December
8–13, 2003, Vancouver and Whistler, British Columbia, Canada, pp. 799–806.
Kim, H. T., Kline, S. J., and Reynolds, W. C. (1971), ‘The production of turbulence near a
smooth wall in a turbulent boundary layer’, Journal of Fluid Mechanics 50, 133–160.
Kim, J. (2011), ‘Physics and control of wall turbulence for drag reduction’, Philosophical
Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences
369(1940), 1396–1411.
Kim, J. and Bewley, T. (2007), ‘A linear systems approach to flow control’, Annual Review of
Fluid Mechanics 39, 383–417.
King, P., Rowland, J., Aubrey, W., Liakata, M., Markham, M., Soldatova, L. N., Whelan, K. E.,
Clare, A., Young, M., Sparkes, A., Oliver, S. G., and Pir, P. (2009), ‘The robot scientist
Adam’, Computer 42(7), 46–54.
Kingma, D. P. and Ba, J. (2014), ‘Adam: A method for stochastic optimization’, arXiv preprint
arXiv:1412.6980.
Kline, S. J., Reynolds, W. C., Schraub, F. A., and Runstadler, P. W. (1967), ‘Structure of
turbulent boundary layers’, Journal of Fluid Mechanics 30, 741–773.
Knaak, M., Rothlubbers, C. and Orglmeister, R. (1997), A Hopfield neural network for flow
field computation based on particle image velocimetry/particle tracking velocimetry image
sequences, in ‘International Conference on Neural Networks, 1997’, Vol. 1, October 8–10,
1997, Lausanne, Switzerland, IEEE, pp. 48–52.
Knight, W. (2018), ‘Google just gave control over data center cooling to an AI’,
www.technologyreview.com/s/611902/google-just-gave-control-over-data-
center-cooling-to-an-ai/.
Koch, W., Bertolotti, F., Stolte, A., and Hein, S. (2000), ‘Nonlinear equilibrium solutions
in a three-dimensional boundary layer and their secondary instability’, Journal of Fluid
Mechanics 406, 131–174.
Kochenderfer, M. and Wheeler, T. (2019), Algorithms for Optimization, MIT Press, Cambridge,
MA.
Kolmogorov, A. N. (1941), ‘The local structure of turbulence in incompressible viscous fluid
for very large Reynolds numbers’, Doklady Akademii Nauk SSSR 30, 209–303.
Kot, M. (2015), A First Course in the Calculus of Variations, American Mathematical Society,
Providence, RI.
Koza, J. R. (1992), Genetic Programming: On the Programming of Computers by Means of
Natural Selection, The MIT Press, Boston.
Kraichnan, R. H. (1967), ‘Inertial ranges in two-dimensional turbulence’, Physics of Fluids
10, 1417–1423.
Kraichnan, R. H. (1971), ‘Inertial range transfer in two- and three-dimensional turbulence’,
Journal of Fluid Mechanics 47, 525–535.
Kraichnan, R. H. and Montgomery, D. (1980), ‘Two-dimensional turbulence’, Reports on
Progress in Physics 43, 547–619.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 428 Data-Driven Fluid Mechanics

Kramer, B., Grover, P., Boufounos, P., Benosman, M., and Nabi, S. (2015), ‘Sparse sensing and
DMD based identification of flow regimes and bifurcations in complex flows’, arXiv preprint
arXiv:1510.02831.
Krizhevsky, A., Sutskever, I., and Hinton, G. E. (2012), Imagenet classification with deep
convolutional neural networks, in ‘Advances in Neural Information Processing Systems’,
December 3–6, 2012, Lake Tahoe, NV, pp. 1097–1105.
Kroll, N., Abu-Zurayk, M., Dimitrov, D., Franz, T., Führer, T., Gerhold, T., Görtz, S., Heinrich,
R., Ilic, C., Jepsen, J., Jägersküpper, J., Kruse, M., Krumbein, A., Langer, S., Liu, D.,
Liepelt, R., Reimer, L., Ritter, M., Schwöppe, A., Scherer, J., Spiering, F., Thormann,
R., Togiti, V., Vollmer, D., and Wendisch, J.-H. (2015), ‘DLR project digital-x: Towards
virtual aircraft design and flight testing based on high-fidelity methods’, CEAS Aeronautical
Journal 7(1), 3–27.
Kroll, N., Langer, S., and Schwöppe, A. (2014), The DLR flow solver TAU – Status and recent
algorithmic developments, in ‘52nd Aerospace Sciences Meeting’, January 3–17, 2014,
National Harbor, MD, American Institute of Aeronautics and Astronautics, Reston, VA.
Kuhn, T. S. (1970), The Structure of Scientific Revolutions, 2nd ed., Chicago University Press,
Chicago.
Kuhnle, A., Schaarschmidt, M., and Fricke, K. (2017), ‘Tensorforce: a TensorFlow library for
applied reinforcement learning’, https://tensorforce.readthedocs.io/en/latest/#.
Kuipers, L. and Niederreiter, H. (2005), Uniform Distribution of Sequences, Dover, Mineola,
NY, p. 129.
Kuo, K. and Acharya, R. (2012), Fundamentals of Turbulent and Multi-Phase Combustion,
Wiley, New York.
Kutz, J., Brunton, S., Brunton, B., and Proctor, J. (2016), Dynamic Mode Decomposition: Data-
Driven Modeling of Complex Systems, Society for Industrial and Applied Mathematics,
Philadelphia.
Kutz, J. N. (2017), ‘Deep learning in fluid dynamics’, Journal of Fluid Mechanics 814, 1–4.
Kutz, J. N., Fu, X., and Brunton, S. L. (2016), ‘Multiresolution dynamic mode decomposition’,
SIAM Journal on Applied Dynamical Systems 15(2), 713–735.
Labonté, G. (1999), ‘A new neural network for particle-tracking velocimetry’, Experiments in
Fluids 26(4), 340–346.
Lagaris, I. E., Likas, A., and Fotiadis, D. I. (1998), ‘Artificial neural networks for solving ordi-
nary and partial differential equations’, IEEE Transactions on Neural Networks 9(5), 987–
1000.
Lai, Z. and Nagarajaiah, S. (2019), ‘Sparse structural system identification method for
nonlinear dynamic systems with hysteresis/inelastic behavior’, Mechanical Systems and
Signal Processing 117, 813–842.
Lall, S., Marsden, J. E., and Glavaški, S. (1999), ‘Empirical model reduction of controlled
nonlinear systems,’ International Federation of Automatic Control 32(2), 2598–2603.
Lall, S., Marsden, J. E., and Glavaški, S. (2002), ‘A subspace approach to balanced truncation
for model reduction of nonlinear control systems’, International Journal of Robust and
Nonlinear Control 12(6), 519–535.
Lan, Y. and Mezić, I. (2013), ‘Linearization in the large of nonlinear systems and Koopman
operator spectrum’, Physica D: Nonlinear Phenomena 242(1), 42–53.
Landau, L. D. (1944), ‘On the problem of turbulence,’ C.R. Acad. Sci. USSR 44, 311–314.
Landau, L. D. and Lifshitz, E. M. (1987), Fluid Mechanics, Vol. 6 in Course of Theoretical
Physics, 2nd English ed., Pergamon Press, Oxford.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 References 429

Lasota, A. and Mackey, M. (1994), Chaos, Fractals, and Noise: Stochastic Aspects of Dynamics,
Springer Verlag, New York.
Law, C. K. (2010), Combustion Physics, Cambridge University Press, Cambridge.
Le Clainche, S. and Vega, J. M. (2017), ‘Higher order dynamic mode decomposition to identify
and extrapolate flow patterns’, Physics of Fluids 29(8), 084102.
Leclercq, C., Demourant, F., Poussot-Vassal, C., and Sipp, D. (2019), ‘Linear iterative method
for closed-loop control of quasiperiodic flows’, Journal of Fluid Mechanics 868, 26–65.
LeCun, Y., Bengio, Y., and Hinton, G. (2015), ‘Deep learning’, Nature 521(7553), 436–444.
Lee, C., Kim, J., Babcock, D., and Goodman, R. (1997), ‘Application of neural networks to
turbulence control for drag reduction’, Physics of Fluids 9(6), 1740–1747.
Lee, J.-H. and Sung, H. J. (2013), ‘Comparison of very-large-scale motions of turbulent pipe
and boundary layer simulations’, Physics of Fluids 25, 045103.
Lee, Y., Yang, H., and Yin, Z. (2017), ‘PIV-DCNN: Cascaded deep convolutional neural
networks for particle image velocimetry’, Experiments in Fluids 58(12), 171.
Leonard, A. (1980), ‘Vortex methods for flow simulation’, Journal of Computational Physics
37(3), 289–335.
Li, H., Fernex, D., Semaan, R., Tan, J., Morzyński, M., and Noack, B. R. (2021), ‘Cluster-based
network model’, Journal of Fluid Mechanics 906, A21:1–41.
Li, Q., Dietrich, F., Bollt, E. M., and Kevrekidis, I. G. (2017), ‘Extended dynamic mode
decomposition with dictionary learning: A data-driven adaptive spectral decomposition
of the Koopman operator’, Chaos: An Interdisciplinary Journal of Nonlinear Science
27(10), 103111.
Li, R., Noack, B. R., Cordier, L., Borée, J., Kaiser, E., and Harambat, F. (2018), ‘Linear genetic
programming control for strongly nonlinear dynamics with frequency crosstalk’, Archives of
Mechanics 70(6), 505–534.
Li, Y., Perlman, E., Wan, M., Yang, Y., Meneveau, C., Burns, R., Chen, S., Szalay, A., and
Eyink, G. (2008), ‘A public turbulence database cluster and applications to study Lagrangian
evolution of velocity increments in turbulence’, Journal of Turbulence 9, N31.
Liberty, E., Woolfe, F., Martinsson, P.-G., Rokhlin, V., and Tygert, M. (2007), ‘Randomized
algorithms for the low-rank approximation of matrices’, Proceedings of the National
Academy of Sciences 104(51), 20167–20172.
Lin, C.-J. (2007), ‘Projected gradient methods for nonnegative matrix factorization’, Neural
Computation 19(10), 2756–2779.
Ling, J., Jones, R., and Templeton, J. (2016), ‘Machine learning strategies for systems with
invariance properties’, Journal of Computational Physics 318, 22–35.
Ling, J., Kurzawski, A., and Templeton, J. (2016), ‘Reynolds averaged turbulence modelling
using deep neural networks with embedded invariance’, Journal of Fluid Mechanics
807, 155–166.
Ling, J. and Templeton, J. (2015), ‘Evaluation of machine learning algorithms for prediction
of regions of high Reynolds averaged Navier Stokes uncertainty’, Physics of Fluids
27(8), 085103.
Liu, J. T. C. (1989), ‘Coherent structures in transitional and turbulent free shear flows’, Annual
Review of Fluid Mechanics 21(1), 285–315.
Ljung, L. (1999), System Identification: Theory for the User, Prentice Hall, Upper Saddle River,
NJ.
Ljung, L. (2008), ‘Perspectives on system identification’, IFAC Proceedings Volumes
41(2), 7172–7184.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 430 Data-Driven Fluid Mechanics

Ljung, L. and Glad, T. (1994), Modeling of Dynamic Systems, Prentice Hall, Englewood Cliffs,
NJ.
Loan, C. V. (1992), Computational Frameworks for the Fast Fourier Transform, Society for
Industrial and Applied Mathematics (SIAM), Philadelphia.
Loiseau, J.-C. (2019), ‘Data-driven modeling of the chaotic thermal convection in an annular
thermosyphon’, arXiv preprint arXiv:1911.07920.
Loiseau, J.-C. and Brunton, S. L. (2018), ‘Constrained sparse Galerkin regression’, Journal of
Fluid Mechanics 838, 42–67.
Loiseau, J.-C., Brunton, S. L., and Noack, B. R. (2021), From the POD-Galerkin method to
sparse manifold models, in P. Benner, S. Grivet-Talocaia, A. Quarteroni, R. G., W. Schilders,
and L. M. Silveria, eds., Handbook of Model-Order Reduction. Volume 3: Applications, De
Gruyter, Berlin, pp. 279–320.
Loiseau, J.-C., Noack, B. R., and Brunton, S. L. (2018), ‘Sparse reduced-order modeling:
Sensor-based dynamics to full-state estimation’, Journal of Fluid Mechanics 844, 459–490.
Lorenz, E. (1956), Empirical orthogonal functions and statistical weather prediction, Statistical
forecasting project, Scientific Report No. 1, MIT, Department of Meteorology, Cambridge,
MA.
Lorenz, E. N. (1963), ‘Deterministic nonperiodic flow’, Journal of the Atmospheric Sciences
20(2), 130–141.
Lozano-Durán, A., Flores, O., and Jiménez, J. (2012), ‘The three-dimensional structure of
momentum transfer in turbulent channels’, Journal of Fluid Mechanics 694, 100–130.
Lozano-Durán, A. and Jiménez, J. (2014), ‘Time-resolved evolution of coherent structures in
turbulent channels: Characterization of eddies and cascades’, Journal of Fluid Mechanics
759, 432–471.
Lu, L., Meng, X., Mao, Z., and Karniadakis, G. E. (2019), ‘Deepxde: A deep learning library
for solving differential equations’, arXiv preprint arXiv:1907.04502.
Lu, S. S. and Willmarth, W. W. (1973), ‘Measurements of the structure of the Reynolds stress
in a turbulent boundary layer’, Journal of Fluid Mechanics 60, 481–511.
Lu, T. and Law, C. K. (2009), ‘Toward accommodating realistic fuel chemistry in large-scale
computations’, Progress in Energy and Combustion Science 35(2), 192–215.
Luchtenburg, D. M., Günter, B., Noack, B. R., King, R., and Tadmor, G. (2009), ‘A generalized
mean-field model of the natural and actuated flows around a high-lift configuration’, Journal
of Fluid Mechanics 623, 283–316.
Luchtenburg, D. M. and Rowley, C. W. (2011), ‘Model reduction using snapshot-based
realizations’, Bulletin of the American Physical Society 56(18), 37pp.
Luchtenburg, D. M., Schlegel, M., Noack, B. R., Aleksić, K., King, R., Tadmor, G., and
Günther, B. (2010), Turbulence control based on reduced-order models and nonlinear
control design, in R. King, ed., Active Flow Control II, Vol. 108 of Notes on Numerical
Fluid Mechanics and Multidisciplinary Design, May 26–28, 2010, Springer-Verlag, Berlin,
pp. 341–356.
Luhar, M., Sharma, A. S., and McKeon, B. J. (2014), ‘Opposition control within the resolvent
analysis framework’, Journal of Fluid Mechanics 749, 597–626.
Lumley, J. (1970), Stochastic Tools in Turbulence, Academic Press, New York.
Lumley, J. L. (1967), The structure of inhomogeneous turbulent flows, in A. M. Yaglam and V. I.
Tatarsky, eds., ‘Proceedings of the International Colloquium on the Fine Scale Structure of
the Atmosphere and Its Influence on Radio Wave Propagation’, Doklady Akademii Nauk
SSSR, Moscow, Nauka.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 References 431

Lumley, J. L. and Poje, A. (1997), ‘Low-dimensional models for flows with density fluctua-
tions’, Physics of Fluids 9(7), 2023–2031.
Lusch, B., Kutz, J. N., and Brunton, S. L. (2018), ‘Deep learning for universal linear embeddings
of nonlinear dynamics’, Nature Communications 9(1), 4950.
Lyapunov, A. (1892), The General Problem of the Stability of Motion, The Kharkov Mathemat-
ical Society, Kharkov, 251pp.
Ma, Z., Ahuja, S., and Rowley, C. W. (2011), ‘Reduced order models for control of fluids using
the eigensystem realization algorithm’, Theoretical and Computational Fluid Dynamics
25(1), 233–247.
Mackie, J. (1974), The Cement of the Universe: A Study of Causation, Oxford University Press,
New York.
MacQueen, J. (1967), Some methods for classification and analysis of multivariate observations,
in ‘Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probabil-
ity’, Vol. 1, Oakland, CA, pp. 281–297.
Magill, J., Bachmann, M., and Rixon, G. (2003), ‘Dynamic stall control using a model-based
observer.’, Journal of Aircraft 40(2), 355–362.
Magnussen, B. F. (1981), On the structure of turbulence and a generalized eddy dissipation
concept for chemical reaction in turbulent flow, in ‘19th AIAA Aerospace Science Meeting’,
May 21–24, 2007, Williamsburg, VA.
Magri, L. (2019), ‘Adjoint methods as design tools in thermoacoustics’, Applied Mechanics
Reviews 71(2), 020801.
Mahoney, M. W. (2011), ‘Randomized algorithms for matrices and data’, Foundations and
Trends in Machine Learning 3, 123–224.
Majda, A. J. and Harlim, J. (2012), ‘Physics constrained nonlinear regression models for time
series’, Nonlinearity 26(1), 201.
Malik, M. R., Isaac, B. J., Coussement, A., Smith, P. J., and Parente, A. (2018), ‘Principal
component analysis coupled with nonlinear regression for chemistry reduction’, Combustion
and Flame 187, 30–41.
Malik, M. R., Obando Vega, P., Coussement, A., and Parente, A. (2020), ‘Combustion modeling
using Principal Component Analysis: A posteriori validation on Sandia flames D, E and F’,
Proceedings of the Combustion Institute 38(2), 2635–2643.
Mallat, S. (2009), A Wavelet Tour of Signal Processing, Elsevier, Oxford.
Mallat, S. G. (1989), ‘Multiresolution approximations and wavelet orthonormal bases of
L 2 (R)’, Transactions of the American Mathematical Society 315(1), 69–87.
Mallor, F., Raiola, M., Vila, C. S., Örlü, R., Discetti, S., and Ianiro, A. (2019), ‘Modal
decomposition of flow fields and convective heat transfer maps: An application to wall-
proximity square ribs’, Experimental Thermal and Fluid Science 102, 517–527.
Maltrud, M. E. and Vallis, G. K. (1991), ‘Energy spectra and coherent structures in forced two-
dimensional and beta-plane turbulence’, Journal of Fluid Mechanics 228, 321–342.
Mangan, N. M., Brunton, S. L., Proctor, J. L., and Kutz, J. N. (2016), ‘Inferring biological
networks by sparse identification of nonlinear dynamics’, IEEE Transactions on Molecular,
Biological, and Multi-Scale Communications 2(1), 52–63.
Mangan, N. M., Kutz, J. N., Brunton, S. L., and Proctor, J. L. (2017), ‘Model selection for
dynamical systems via sparse regression and information criteria’, Proceedings of the Royal
Society A 473(2204), 1–16.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 432 Data-Driven Fluid Mechanics

Manohar, K., Brunton, B. W., Kutz, J. N., and Brunton, S. L. (2018), ‘Data-driven sparse sensor
placement: Demonstrating the benefits of exploiting known patterns’, IEEE Control Systems
Magazine 38(3), 63–86.
Marčenko, V. A. and Pastur, L. A. (1967), ‘Distribution of eigenvalues for some sets of random
matrices’, Mathematics of the USSR-Sbornik 1(4), 457.
Mardt, A., Pasquali, L., Wu, H., and Noé, F. (2018), ‘VAMPnets: Deep learning of molecular
kinetics’, Nature Communications 9, 5.
Marsden, J. E. and Ratiu, T. S. (1999), Introduction to Mechanics and Symmetry, 2nd ed.,
Springer-Verlag, New York.
Maulik, R., San, O., Rasheed, A., and Vedula, P. (2019), ‘Subgrid modelling for two-
dimensional turbulence using neural networks’, Journal of Fluid Mechanics 858, 122–144.
Maurel, S., Borée, J., and Lumley, J. (2001), ‘Extended proper orthogonal decomposition:
Application to jet/vortex interaction’, Flow, Turbulence and Combustion 67(2), 125–136.
McKeon, B. and Sharma, A. (2010), ‘A critical-layer framework for turbulent pipe flow’,
Journal of Fluid Mechanics 658, 336–382.
McWilliams, J. C. (1980), ‘An application of equivalent modons to atmospheric blocking’,
Dynamics of Atmospheres and Oceans 5, 43–66.
McWilliams, J. C. (1984), ‘The emergence of isolated coherent vortices in turbulent flow’,
Journal of Fluid Mechanics 146, 21–43.
McWilliams, J. C. (1990a), ‘A demonstration of the suppression of turbulent cascades by
coherent vortices in two-dimensional turbulence’, Physics of Fluids A 2, 547–552.
McWilliams, J. C. (1990b), ‘The vortices of two-dimensional turbulence’, Journal of Fluid
Mechanics 219, 361–385.
Meena, M. G., Nair, A. G., and Taira, K. (2018), ‘Network community-based model reduction
for vortical flows’, Physical Review E 97(6), 063103.
Mendez, M. A., Hess, D., Watz, B. B., and Buchlin, J.-M. (2020), ‘Multiscale proper orthogonal
decomposition (MPOD) of TR-PIV data- A case study on stationary and transient cylinder
wake flows’, Measurement Science and Technology 80, 981–1002.
Mendez, M. A. and Buchlin, J.-M. (2016), ‘Notes on 2D pulsatile Poiseuille flows: An
introduction to eigenfunctions and complex variables using MATLAB’, Technical Report
TN 215, von Karman Institute for Fluid Dynamics, Sint-Genesius-Rode, Belgium.
Mendez, M. A., Scelzo, M., and Buchlin, J.-M. (2018), ‘Multiscale modal analysis of an
oscillating impinging gas jet’, Experimental Thermal and Fluid Science 91, 256–276.
Mendez, M., Balabane, M., and Buchlin, J.-M. (2019), ‘Multi-scale proper orthogonal decom-
position of complex fluid flows’, Journal of Fluid Mechanics 870, 988–1036.
Mendez, M., Gosset, A., and Buchlin, J.-M. (2019), ‘Experimental analysis of the stability of the
jet wiping process, part II: Multiscale modal analysis of the gas jet-liquid film interaction’,
Experimental Thermal and Fluid Science 106, 48–67.
Mendez, M., Raiola, M., Masullo, A., Discetti, S., Ianiro, A., Theunissen, R., and Buchlin, J.-
M. (2017), ‘POD-based background removal for particle image velocimetry’, Experimental
Thermal and Fluid Science 80, 181–192.
Mendible, A., Brunton, S. L., Aravkin, A. Y., Lowrie, W., and Kutz, J. N. (2020), ‘Dimension-
ality reduction and reduced-order modeling for traveling wave physics’, Theoretical and
Computational Fluid Dynamics 34(4), 385–400.
Meneveau, C. (1991), ‘Analysis of turbulence in the orthonormal wavelet representation’,
Journal of Fluid Mechanics 232, 469–520.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 References 433

Meneveau, C. and Katz, J. (2000), ‘Scale-invariance and turbulence models for large-eddy
simulation’, Annual Review of Fluid Mechanics 32(1), 1–32.
Meunier, P., Le Dizès, S., and Leweke, T. (2005), ‘Physics of vortex merging’, C. R. Physique
6, 431–450.
Meyers, S. D., Kelly, B. G., and O’Brien, J. J. (1993), ‘An introduction to wavelet analysis in
oceanography and meteorology: With application to the dispersion of Yanai waves’, Monthly
Weather Review 121(10), 2858–2866.
Mezić, I. (2005), ‘Spectral properties of dynamical systems, model reduction and decomposi-
tions’, Nonlinear Dynamics 41(1–3), 309–325.
Mezić, I. (2013), ‘Analysis of fluid flows via spectral properties of the Koopman operator’,
Annual Review of Fluid Mechanics 45, 357–378.
Mezić, I. and Banaszuk, A. (2004), ‘Comparison of systems with complex behavior’, Physica
D: Nonlinear Phenomena 197(1), 101–133.
Mifsud, M., Vendl, A., Hansen, L.-U., and Görtz, S. (2019), ‘Fusing wind-tunnel measurements
and CFD data using constrained gappy proper orthogonal decomposition’, Aerospace
Science and Technology 86, 312–326.
Mifsud, M., Zimmermann, R., and Görtz, S. (2014), ‘Speeding-up the computation of high-
lift aerodynamics using a residual-based reduced-order model’, CEAS Aeronautical Journal
6(1), 3–16.
Milano, M. and Koumoutsakos, P. (2002), ‘Neural network modeling for near wall turbulent
flow’, Journal of Computational Physics 182(1), 1–26.
Min, C. and Choi, H. (1999), ‘Suboptimal feedback control of vortex shedding at low reynolds
numbers’, Journal of Fluid Mechanics 401, 123–156.
Minelli, G., Dong, T., Noack, B. R., and Krajnović, S. (2020), ‘Upstream actuation for
bluff-body wake control driven by a genetically inspired optimization’, Journal of Fluid
Mechanics 893, A1.
Mizuno, Y. and Jiménez, J. (2013), ‘Wall turbulence without walls’, Journal of Fluid Mechanics
723, 429–455.
Mnih, V., Badia, A. P., Mirza, M., Graves, A., Lillicrap, T., Harley, T., Silver, D., and
Kavukcuoglu, K. (2016), Asynchronous methods for deep reinforcement learning, in M. F.
Balcan and K. Q. Weinberger, eds., ‘Proceedings of the 33rd International Conference on
Machine Learning’, Vol. 48 of Proceedings of Machine Learning Research, PMLR, New
York, pp. 1928–1937.
Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, M. G., Graves,
A., Riedmiller, M., Fidjeland, A. K., Ostrovski, G., Petersen, S., Beattie, C., Sadik, A.,
Antonoglou, I., King, H., Kumaran, D., Wierstra, D., Legg, S., and Hassabis, D. (2015),
‘Human-level control through deep reinforcement learning’, Nature 518(7540), 529–533.
Moarref, R., Jovanovic, M., Tropp, J., Sharma, A., and McKeon, B. (2014), ‘A low-order
decomposition of turbulent channel flow via resolvent analysis and convex optimization’,
Physics Fluids 26, 051701.
Mohammed, R., Tanoff, M., Smooke, M., Schaffer, A., and Long, M. (1998), ‘Computational
and experimental study of a forced, time-varying, axisymmetric, laminar diffusion flame’,
Symposium (International) on Combustion 27(1), 693–702.
Moisy, F. and Jiménez, J. (2004), ‘Geometry and clustering of intense structures in isotropic
turbulence’, Journal of Fluid Mechanics 513, 111–133.
Mojgani, R. and Balajewicz, M. (2017), ‘Lagrangian basis method for dimensionality reduction
of convection dominated nonlinear flows’, arXiv preprint arXiv:1701.04343.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 434 Data-Driven Fluid Mechanics

Moore, B. C. (1981), ‘Principal component analysis in linear systems: Controllability, observ-

ability, and model reduction’, IEEE Transactions on Automatic Control AC-26(1), 17–32.
Morzyński, M., Afanasiev, K., and Thiele, F. (1999), ‘Solution of the eigenvalue problems
resulting from global non-parallel flow stability analysis’, Computer Methods in Applied
Mechanics and Engineering 169, 161–176.
Morzyński, M., Noack, B. R., and Tadmor, G. (2007), ‘Global stability analysis and reduced
order modeling for bluff-body flow control’, Journal of Theoretical and Applied Mechanics
45, 621–642.
Morzyński, M., Stankiewicz, W., Noack, B. R., Thiele, F., and Tadmor, G. (2006), Generalized
mean-field model for flow control using continuous mode interpolation, in ‘3rd AIAA Flow
Control Conference’, June 5–8, 2006, San Francisco, CA, Invited AIAA-Paper 2006-3488.
Mowlavi, S. and Sapsis, T. P. (2018), ‘Model order reduction for stochastic dynamical systems
with continuous symmetries’, SIAM Journal on Scientific Computing 40(3), A1669–A1695.
Murata, T., Fukami, K., and Fukagata, K. (2019), CNN-SINDY based reduced order modeling
of unsteady flow fields, in ‘Fluids Engineering Division Summer Meeting’, July–August
2019, San Francisco, CA, Vol. 59032, American Society of Mechanical Engineers, New
York, p. V002T02A074.
Murphy, K. P. (2012), Machine Learning, MIT Press, Cambridge.
Nagata, M. (1990), ‘Three-dimensional finite-amplitude solutions in plane Couette flow:
Bifurcation from infinity’, Journal of Fluid Mechanics 217, 519–527.
Nair, A. G., Brunton, S. L., and Taira, K. (2018), ‘Networked-oscillator-based modeling and
control of unsteady wake flows’, Physical Review E 97(6), 063107.
Nair, A. G. and Taira, K. (2015), ‘Network-theoretic approach to sparsified discrete vortex
dynamics’, Journal of Fluid Mechanics 768, 549–571.
Nair, A. G., Yeh, C.-A., Kaiser, E., Noack, B. R., Brunton, S. L., and Taira, K. (2019), ‘Cluster-
based feedback control of turbulent post-stall separated flows’, Journal of Fluid Mechanics
875, 345–375.
Narasingam, A. and Kwon, J. S.-I. (2018), ‘Data-driven identification of interpretable reduced-
order models using sparse regression’, Computers & Chemical Engineering 119, 101–111.
Neely, T. W., Samson, E. C., Bradley, A. S., Davis, M. J., and Anderson, B. P. (2010),
‘Observation of vortex dipoles in an oblate Bose–Einstein condensate’, Physical Review
Letters 104, 160401.
Newell, A. and Simon, H. A. (1976), ‘Computer science as empirical inquiry: Symbols and
search’, Communication of ACM 19(3), 113–126.
Newton, I. S. (1687), Philosophiae naturalis principia mathematica, Vol. 1, London.
Ng, A. Y., Harada, D., and Russell, S. J. (1999), Policy invariance under reward transformations:
Theory and application to reward shaping, in ‘Proceedings of the Sixteenth International
Conference on Machine Learning’, ICML 99, Morgan Kaufmann, San Francisco, CA,
pp. 278–287.
Ng, A. Y., Jordan, M. I., and Weiss, Y. (2002), On spectral clustering: Analysis and an algorithm,
in ‘Advances in Neural Information Processing Systems’, December 3–8, 2001, Vancouver,
British Columbia, Canada, pp. 849–856.
Nielsen, A. (2019), Practical Time Series Analysis, O’Reilly Media, Sebastopol, CA.
Nilsson, N. (1998), Artificial Intelligence: A New Synthesis, Morgan Kaufmann, San Francisco,
CA.
Ninni, D. and Mendez, M. A. (2021), ‘Modulo: A software for multiscale proper orthogonal
decomposition of data’, SoftwareX 12, 100622.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 References 435

Noack, B. R. (2019), Closed-loop turbulence control – From human to machine learning (and
retour), in Y. Zhou, M. Kimura, G. Peng, A. D. Lucey, and L. Hung, eds., ‘Fluid-Structure-
Sound Interactions and Control. Proceedings of the 4th Symposium on Fluid-Structure-
Sound Interactions and Control’, Springer, Berlin, pp. 23–32.
Noack, B. R., Afanasiev, K., Morzyński, M., Tadmor, G., and Thiele, F. (2003), ‘A hierarchy of
low-dimensional models for the transient and post-transient cylinder wake’, Journal of Fluid
Mechanics 497, 335–363.
Noack, B. R. and Eckelmann, H. (1994), ‘A global stability analysis of the steady and periodic
cylinder wake’, Journal of Fluid Mechanics 270, 297–330.
Noack, B. R., Mezić, I., Tadmor, G., and Banaszuk, A. (2004), ‘Optimal mixing in recirculation
zones’, Physics of Fluids 16(4), 867–888.
Noack, B. R., Morzynski, M., and Tadmor, G. (2011), Reduced-Order Modelling for Flow
Control, Vol. 528 of International Centre for Mechanical Sciences: Courses and Lectures,
Springer Science & Business Media, Dordrecht.
Noack, B. R., Papas, P., and Monkewitz, P. A. (2005), ‘The need for a pressure-term
representation in empirical Galerkin models of incompressible shear flows’, Journal of Fluid
Mechanics 523, 339.
Noack, B. R., Pelivan, I., Tadmor, G., Morzyński, M., and Comte, P. (2004), Robust low-
dimensional Galerkin models of natural and actuated flows, in ‘Fourth Aeroacoustics
Workshop’, RWTH, Aachen, February 26–27, 2004, pp. 0001–0012.
Noack, B. R., Schlegel, M., Ahlborn, B., Mutschke, G., Morzyński, M., Comte, P., and
Tadmor, G. (2008), ‘A finite-time thermodynamics of unsteady fluid flows’, Journal of Non-
Equilibrium Thermodynamics 33(2), 103–148.
Noack, B. R., Stankiewicz, W., Morzynski, M., and Schmid, P. J. (2016), ‘Recursive dynamic
mode decomposition of a transient cylinder wake’, Journal of Fluid Mechanics 809, 843–
872.
Nocedal, J. and Wright, S. (2006), Numerical Optimization, Springer Verlag, New York.
Noé, F. and Nuske, F. (2013), ‘A variational approach to modeling slow processes in stochastic
dynamical systems’, Multiscale Modeling Simulation 11(2), 635–655.
Noé, F., Olsson, S., Köhler, J., and Wu, H. (2019), ‘Boltzmann generators: Sampling equilib-
rium states of many-body systems with deep learning’, Science 365(6457), eaaw1147.
Noether, E. (1918), ‘Invariante Variationsprobleme’, Nachrichten von der Gesellschaft der
Wissenschaften zu Göttingen, Mathematisch-Physikalische Klasse 1918, 235–257. English
Reprint: physics/0503066, http://dx. doi. org/10.1080/00411457108231446, p. 57.
Novati, G., Mahadevan, L., and Koumoutsakos, P. (2019), ‘Controlled gliding and perching
through deep-reinforcement-learning’, Physical Review Fluids 4(9), 093902.
Novati, G., Verma, S., Alexeev, D., Rossinelli, D., Van Rees, W. M., and Koumoutsakos, P.
(2017), ‘Synchronisation through learning for two self-propelled swimmers’, Bioinspiration
& Biomimetics 12(3), aa6311.
Nüske, F., Schneider, R., Vitalini, F., and Noé, F. (2016), ‘Variational tensor approach for
approximating the rare-event kinetics of macromolecular systems’, The Journal of Chemical
Physics 144(5), 054105.
Obukhov, A. M. (1941), ‘On the distribution of energy in the spectrum of turbulent flow’,
Doklady Akademii Nauk SSSR 32, 22–24.
Ogata, K. (2009), Modern Control Engineering, 5th ed., Pearson, Upper Saddle River, NJ.
Onsager, L. (1949), ‘Statistical hydrodynamics’, Nuovo Cimento Suppl. 6, 279–286.
Oppenheim, A. V. (2015), Signals, Systems and Inference, Pearson Education, Harlow.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 436 Data-Driven Fluid Mechanics

Oppenheim, A. V. and Schafer, R. W. (2009), Discrete-Time Signal Processing, 3rd ed.

(Prentice-Hall Signal Processing Series), Pearson, Harlow.
Oppenheim, A. V., Willsky, A. S., and with S. Hamid (1996), Signals and Systems, 2nd ed.,
Prentice-Hall, Upper Saddle River, NJ.
Orr, W. (1907a), ‘The stability or instability of the steady motions of a perfect liquid and of a
viscous liquid’, Proceedings of the Royal Irish Academy 27, 9–138.
Orr, W. M. (1907b), ‘The stability or instability of the steady motions of a perfect liquid, and of
a viscous liquid. Part I: A perfect liquid’, Proceedings of the Royal Irish Academy 27, 9–68.
Orszag, S. (1971), ‘Accurate solution of the Orr-Sommerfeld stability equation’, Journal of
Fluid Mechanics 50, 689–703.
Orszag, S. A. and Patera, A. T. (1983), ‘Secondary instability of wall-bounded shear flows’,
Journal of Fluid Mechanics 128, 347–385.
Östh, J., Noack, B. R., and Krajnović, S. (2014), ‘On the need for a nonlinear subscale
turbulence term in POD models as exemplified for a high Reynolds number flow over an
Ahmed body’, Journal of Fluid Mechanics 747, 518–544.
Otto, S. E. and Rowley, C. W. (2019), ‘Linearly-recurrent autoencoder networks for learning
dynamics’, SIAM Journal on Applied Dynamical Systems 18(1), 558–593.
Ouellette, N. T., Xu, H., and Bodenschatz, E. (2006), ‘A quantitative study of three-dimensional
Lagrangian particle tracking algorithms’, Experiments in Fluids 40(2), 301–313.
Paatero, P. and Tapper, U. (1994), ‘Positive matrix factorization: A non-negative factor model
with optimal utilization of error estimates of data values’, Environmetrics 5(2), 111–126.
Pan, S., Arnold-Medabalimi, N., and Duraisamy, K. (2020), ‘Sparsity-promoting algo-
rithms for the discovery of informative Koopman invariant subspaces’, arXiv preprint
arXiv:2002.10637.
Pao, H.-T. (2008), ‘A comparison of neural network and multiple regression analysis in
modeling capital structure’, Expert Systems with Applications 35, 720–727.
Parente, A. and Sutherland, J. C. (2013), ‘Principal component analysis of turbulent combustion
data: Data pre-processing and manifold sensitivity’, Combustion and Flame 160(2), 340–
350.
Parente, A., Sutherland, J., Dally, B., Tognotti, L., and Smith, P. (2011), ‘Investigation of
the MILD combustion regime via Principal Component Analysis’, Proceedings of the
Combustion Institute 33(2), 3333–3341.
Parente, A., Sutherland, J., Tognotti, L., and Smith, P. (2009), ‘Identification of low-dimensional
manifolds in turbulent flames’, Proceedings of the Combustion Institute 32(1), 1579–1586.
Parezanovic, V., Cordier, L., Spohn, A., Duriez, T., Noack, B. R., Bonnet, J.-P., Segond, M.,
Abel, M., and Brunton, S. L. (2016), ‘Frequency selection by feedback control in a turbulent
shear flow’, Journal of Fluid Mechanics 797, 247–283.
Parish, E. J. and Duraisamy, K. (2016), ‘A paradigm for data-driven predictive modeling using
field inversion and machine learning’, Journal of Computational Physics 305, 758–774.
Pastoor, M. (2008), Niederdimensionale Wirbelmodelle zur Kontrolle von Scher- und Nach-
laufströmungen, PhD thesis, Berlin Institute of Technology, Germany.
Pastoor, M., Henning, L., Noack, B. R., King, R., and Tadmor, G. (2008), ‘Feedback shear layer
control for bluff body drag reduction’, Journal of Fluid Mechanics 608, 161–196.
Pathak, D., Agrawal, P., Efros, A. A. and Darrell, T. (2017), Curiosity-driven exploration by
self-supervised prediction, in D. Precup and Y. W. Teh, eds., ‘Proceedings of the 34th
International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6–11

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 References 437

August 2017’, Vol. 70 of Proceedings of Machine Learning Research, PMLR, pp. 2778–
2787.
Peerenboom, K., Parente, A., Kozák, T., Bogaerts, A., and Degrez, G. (2015), ‘Dimension
reduction of non-equilibrium plasma kinetic models using principal component analysis’,
Plasma Sources Science and Technology 24(2), 025004.
Penland, C. (1996), ‘A stochastic model of IndoPacific sea surface temperature anomalies’,
Physica D: Nonlinear Phenomena 98(2–4), 534–558.
Penland, C. and Magorian, T. (1993), ‘Prediction of Niño 3 sea surface temperatures using
linear inverse modeling’, Journal of Climate 6(6), 1067–1076.
Pepiot-Desjardins, P. and Pitsch, H. (2008), ‘An efficient error-propagation-based reduction
method for large chemical kinetic mechanisms’, Combustion and Flame 154(1–2), 67–81.
Perko, L. (2013), Differential Equations and Dynamical Systems, Vol. 7 of Texts in Applied
Mathematics, Springer Science & Business Media, New York.
Perlman, E., Burns, R., Li, Y., and Meneveau, C. (2007), Data exploration of turbulence
simulations using a database cluster, in ‘Proceedings of the SC07’, ACM, New York,
pp. 23.1–23.11.
Perrin, R., Braza, M., Cid, E., Cazin, S., Barthet, A., Sevrain, A., Mockett, C., and Thiele,
F. (2007), ‘Obtaining phase averaged turbulence properties in the near wake of a circular
cylinder at high Reynolds number using POD’, Experiments in Fluids 43(2–3), 341–355.
Peters, N. (1984), ‘Laminar diffusion flamelet models in non-premixed turbulent combustion’,
Progress in Energy and Combustion Science 10(3), 319–339.
Phan, M. Q., Juang, J.-N., and Hyland, D. C. (1995), ‘On neural networks in identification and
control of dynamic systems,’ in Ardéshir Guran and Daniel J Inman, eds., ‘Wave Motion,
Intelligent Structures and Nonlinear Mechanics’, World Scientific, Singapore, pp. 194–225.
Picard, C. and Delville, J. (2000), ‘Pressure velocity coupling in a subsonic round jet’,
International Journal of Heat and Fluid Flow 21(3), 359–364.
Poincaré, H. (1920), Science et méthode, Flammarion, Paris. English translation in Dover books,
1952.
Pope, S. B. (2000), Turbulent Flows, Cambridge University Press, Cambridge.
Pope, S. B. (2013), ‘Small scales, many species and the manifold challenges of turbulent
combustion’, Proceedings of the Combustion Institute 34(1), 1–31.
Proctor, J. L., Brunton, S. L., and Kutz, J. N. (2016), ‘Dynamic mode decomposition with
control’, SIAM Journal on Applied Dynamical Systems 15(1), 142–161.
Protas, B. (2004), ‘Linear feedback stabilization of laminar vortex shedding based on a point
vortex model’, Physics of Fluids 16(12), 4473–4488.
Quarteroni, A. and Rozza, G. (2013), Reduced Order Methods for Modeling and Computational
Reduction, Vol. 9 of MS&A – Modeling, Simulation & Appplications, Springer, New York.
Rabault, J. (2019), ‘Deep reinforcement learning applied to fluid mechanics: Materials
from the 2019 Flow/Interface School on Machine Learning and Data Driven Methods’,
DOI:10.13140/RG.2.2.11533.90086.
Rabault, J., Belus, V., Viquerat, J., Che, Z., Hachem, E., Reglade, U., and Jensen, A. (2019),
Exploiting locality and physical invariants to design effective deep reinforcement learning
control of the unstable falling liquid film, in ‘The 1st Graduate Forum of CSAA and the
7th International Academic Conference for Graduates’, NUAA, November 21–22, 2019,
Nanjing, China.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 438 Data-Driven Fluid Mechanics

Rabault, J., Kuchta, M., Jensen, A., Reglade, U., and Cerardi, N. (2019), ‘Artificial neural
networks trained through deep reinforcement learning discover control strategies for active
flow control’, Journal of Fluid Mechanics 865, 281–302.
Rabault, J. and Kuhnle, A. (2019), ‘Accelerating deep reinforcement learning strategies of flow
control through a multi-environment approach’, Physics of Fluids 31(9), 094105.
Rabault, J. and Kuhnle, A. (2020), ‘Deep reinforcement learning applied to active flow control’,
ResearchGate Preprint https://doi.org/10.13140/RG 2.
Rabault, J., Ren, F., Zhang, W., Tang, H., and Xu, H. (2020), ‘Deep reinforcement learning in
fluid mechanics: A promising method for both active flow control and shape optimization’,
Journal of Hydrodynamics 32(2), 234–246.
Raibaudo, C., Zhong, P., Noack, B. R., and Martinuzzi, R. J. (2020), ‘Machine learning
strategies applied to the control of a fluidic pinball’, Physics of Fluids 32(1), 015108.
Raiola, M., Discetti, S., and Ianiro, A. (2015), ‘On PIV random error minimization with optimal
POD-based low-order reconstruction’, Experiments in Fluids 56(4), 75.
Raiola, M., Ianiro, A., and Discetti, S. (2016), ‘Wake of tandem cylinders near a wall’,
Experimental Thermal and Fluid Science 78, 354–369.
Raissi, M. and Karniadakis, G. E. (2018), ‘Hidden physics models: Machine learning of
nonlinear partial differential equations’, Journal of Computational Physics 357, 125–141.
Raissi, M., Perdikaris, P., and Karniadakis, G. (2019), ‘Physics-informed neural networks:
A deep learning framework for solving forward and inverse problems involving nonlinear
partial differential equations’, Journal of Computational Physics 378, 686–707.
Raissi, M., Yazdani, A., and Karniadakis, G. E. (2020), ‘Hidden fluid mechanics: Learning
velocity and pressure fields from flow visualizations’, Science 367(6481), 1026–1030.
Ranzi, E., Frassoldati, A., Stagni, A., Pelucchi, M., Cuoci, A., and Faravelli, T. (2014), ‘Reduced
kinetic schemes of complex reaction systems: fossil and biomass-derived transportation
fuels’, International Journal of Chemical Kinetics 46(9), 512–542.
Rasmussen, C. (2006), Gaussian Processes for Machine Learning, MIT Press, Cambridge, MA.
Rayleigh, L. (1887), ‘On the stability of certain fluid motions’, The Proceedings of the London
Mathematical Society 19, 67–74.
Recht, B. (2019), ‘A tour of reinforcement learning: The view from continuous control’, Annual
Review of Control, Robotics, and Autonomous Systems 2, 253–279.
Reddy, G., Celani, A., Sejnowski, T. J., and Vergassola, M. (2016), ‘Learning to soar in turbulent
environments’, Proceedings of the National Academy of Sciences of the United States of
America 113(33), E4877–E4884.
Reddy, G., Wong-Ng, J., Celani, A., Sejnowski, T. J., and Vergassola, M. (2018), ‘Glider soaring
via reinforcement learning in the field’, Nature 562(7726), 236–239.
Reddy, S. and Henningson, D. (1993), ‘Energy growth in viscous channel flows’, Journal of
Fluid Mechanics 252, 209–238.
Reimer, L., Heinrich, R., and Ritter, M. (2019), Towards higher-precision maneuver and gust
loads computations of aircraft: Status of related features in the CFD-based multidisciplinary
simulation environment FlowSimulator, in ‘Notes on Numerical Fluid Mechanics and
Multidisciplinary Design’, Springer, Cham, pp. 597–607.
Reiss, J., Schulze, P., Sesterhenn, J., and Mehrmann, V. (2018), ‘The shifted proper orthogonal
decomposition: A mode decomposition for multiple transport phenomena’, SIAM Journal
on Scientific Computing 40(3), A1322–A1344.
Rempfer, D. (1995), Empirische Eigenfunktionen und Galerkin-Projektionen zur Beschreibung
des laminar-turbulenten Grenzschichtumschlags (transl.: Empirical eigenfunctions and

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 References 439

Galerkin projection for the description of the laminar-turbulent boundary-layer transition),

Habilitationsschrift, Fakultät für Luft- und Raumfahrttechnik, Universität Stuttgart.
Rempfer, D. (2000), ‘On low-dimensional Galerkin models for fluid flow’, Theoretical and
Computational Fluid Dynamics 14(2), 75–88.
Rempfer, D. and Fasel, F. H. (1994a), ‘Dynamics of three-dimensional coherent structures in a
flat-plate boundary-layer’, Journal of Fluid Mechanics 275, 257–283.
Rempfer, D. and Fasel, F. H. (1994b), ‘Evolution of three-dimensional coherent structures in a
flat-plate boundary-layer’, Journal of Fluid Mechanics 260, 351–375.
Ren, F., Hu, H.-b., and Tang, H. (2020), ‘Active flow control using machine learning: A brief
review’, Journal of Hydrodynamics 32(2), 247–253.
Ren, F., Rabault, J., and Tang, H. (2020), ‘Applying deep reinforcement learning to active flow
control in turbulent conditions’, arxiv preprint arXiv:2006.10683.
Ren, Z. and Pope, S. (2008), ‘Second-order splitting schemes for a class of reactive systems’,
Journal of Computational Physics 227(17), 8165–8176.
Reynolds, O. (1883), ‘An experimental investigation of the circumstances which determine
whether the motion of water shall be direct or sinuous, and of the law of resistance in parallel
channels’, Proceedings of the Royal Society of London 35, 84–99.
Reynolds, O. (1894), ‘On the dynamical theory of incompressible viscous fluids and the
determination of the criterion’, Proceedings of the Royal Society of London 56, 40–45.
Reynolds, W. C. and Tiederman, W. G. (1967), ‘Stability of turbulent channel flow, with
application to Malkus’ theory’, Journal of Fluid Mechanics 27, 253–272.
Richards, J. I. and Youn, H. K. (1990), The Theory of Distributions, Cambridge University
Press, Cambridge.
Richardson, L. F. (1920), ‘The supply of energy from and to atmospheric eddies’, Proceedings
of the Royal Society A 97, 354–373.
Ripepi, M., Verveld, M. J., Karcher, N. W., Franz, T., Abu-Zurayk, M., Görtz, S., and Kier,
T. M. (2018), ‘Reduced-order models for aerodynamic applications, loads and MDO’, CEAS
Aeronautical Journal 9(1), 171–193.
Robinson, S. K. (1991), ‘Coherent motions in the turbulent boundary layer’, Annual Review
Fluid Mechanics 23, 601–639.
Rokhlin, V., Szlam, A., and Tygert, M. (2009), ‘A randomized algorithm for principal
component analysis’, SIAM Journal on Matrix Analysis and Applications 31, 1100–1124.
Rosenblatt, F. (1958), ‘The perceptron: A probabilistic model for information storage and
organization in the brain’, Psychological Review 65(6), 386.
Roussopoulos, K. (1993), ‘Feedback control of vortex shedding at low Reynolds numbers’,
Journal of Fluid Mechanics 248, 267–296.
Roweis, S. and Lawrence, S. (2000), ‘Nonlinear dimensionality reduction by locally linear
embedding’, Science 290(5500), 2323–2326.
Rowley, C. W. (2005), ‘Model reduction for fluids using balanced proper orthogonal decompo-
sition.’, International Journal of Bifurcation and Chaos 15(3), 997–1013.
Rowley, C. W., Colonius, T., and Murray, R. M. (2004), ‘Model reduction for compressible
flows using POD and Galerkin projection’, Physica D: Nonlinear Phenomena 189(1–
2), 115–129.
Rowley, C. W. and Dawson, S. T. M. (2017), ‘Model reduction for flow analysis and control’,
Annual Review of Fluid Mechanics 49(1), 387–417.
Rowley, C. W., Kevrekidis, I. G., Marsden, J. E., and Lust, K. (2003), ‘Reduction and
reconstruction for self-similar dynamical systems’, Nonlinearity 16(4), 1257.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 440 Data-Driven Fluid Mechanics

Rowley, C. W. and Marsden, J. E. (2000), ‘Reconstruction equations and the Karhunen–Loève

expansion for systems with symmetry’, Physica D: Nonlinear Phenomena 142(1), 1–19.
Rowley, C. W., Mezić, I., Bagheri, S., Schlatter, P., and Henningson, D. (2009), ‘Spectral
analysis of nonlinear flows’, Journal of Fluid Mechanics 645, 115–127.
Rowley, C. W. and Williams, D. R. (2006), ‘Dynamics and control of high-Reynolds-number
flow over open cavities’, Annual Review of Fluid Mechanics 38, 251–276.
Rudy, S. H., Brunton, S. L., Proctor, J. L., and Kutz, J. N. (2017), ‘Data-driven discovery of
partial differential equations’, Science Advances 3, e1602614.
Rumelhart, D. E., Hinton, G. E., Williams, R. J. (1986), ‘Learning representations by back-
propagating errors’, Nature 323(6088), 533–536.
Russell, B. (1912), ‘On the notion of cause’, The Proceedings of the Aristotelian Society 13, 1–
26.
Sabater, C., Bekemeyer, P., and Görtz, S. (2020), ‘Efficient bilevel surrogate approach for
optimization under uncertainty of shock control bumps’, AIAA Journal 58(12), 5228–5242.
Salimans, T. and Chen, R. (2018), ‘Learning Montezuma’s Revenge from a single demonstra-
tion’, arXiv preprint arXiv:1812.03381.
Salimans, T., Ho, J., Chen, X., and Sutskever, I. (2017), ‘Evolution strategies as a scalable
alternative to reinforcement learning’, CoRR, arXiv:1703.03864.
Saltzman, B. (1962), ‘Finite amplitude free convection as an initial value problem – I’, Journal
of Atmospheric Sciences 19(4), 329–341.
Sarlos, T. (2006), Improved approximation algorithms for large matrices via random projec-
tions, in ‘47th Annual IEEE Symposium on Foundations of Computer Science’, October
21–24, 2006, Washington, DC, pp. 143–152.
Savinov, N., Raichuk, A., Marinier, R., Vincent, D., Pollefeys, M., Lillicrap, T., and Gelly, S.
(2018), ‘Episodic curiosity through reachability’, arXiv preprint 1810.02274.
Schaeffer, H. (2017), Learning partial differential equations via data discovery and sparse
optimization, ‘Proceedings of the Royal Society A’, 473, 20160446.
Schaeffer, H. and McCalla, S. G. (2017), ‘Sparse model selection via integral terms’, Physical
Review E 96(2), 023302.
Schäfer, M., Turek, S., Durst, F., Krause, E., and Rannacher, R. (1996), Benchmark com-
putations of laminar flow around a cylinder, in ‘Flow Simulation with High-Performance
Computers II’, Springer, Berlin, pp. 547–566.
Schaffer, S. (1994), Making up discovery, in M. A. Boden, ed., ‘Dimensions of creativity’, MIT
Press, Cambridge, pp. 13–51.
Scherl, I., Strom, B., Shang, J. K., Williams, O., Polagye, B. L., and Brunton, S. L. (2020),
‘Robust principal component analysis for particle image velocimetry’, Physical Review
Fluids 5, 054401.
Schlegel, M. and Noack, B. R. (2015), ‘On long-term boundedness of Galerkin models’, Journal
of Fluid Mechanics 765, 325–352.
Schlegel, M., Noack, B. R., and Tadmor, G. (2004), Low-dimensional Galerkin models and
control of transitional channel flow, Technical Report 01/2004, Hermann-Föttinger-Institut
für Strömungsmechanik, Technische Universität Berlin, Germany.
Schmelzer, M., Dwight, R. P., and Cinnella, P. (2020), ‘Discovery of algebraic Reynolds-stress
models using sparse symbolic regression’, Flow, Turbulence and Combustion 104(2), 579–
603.
Schmid, P. (2007), ‘Nonmodal stability theory’, Annual Review of Fluid Mechanics 39, 129–
162.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 References 441

Schmid, P. and Henningson, D. (2001), Stability and Transition in Shear Flows, Springer Verlag,
New York.
Schmid, P. J. (2010), ‘Dynamic mode decomposition of numerical and experimental data’,
Journal of Fluid Mechanics 656, 5–28.
Schmid, P., Li, L., Juniper, M., and Pust, O. (2011), ‘Applications of the dynamic mode
decomposition’, Theoretical and Computational Fluid Dynamics 25(1–4), 249–259.
Schmid, P., Violato, D., and Scarano, F. (2012), ‘Decomposition of time-resolved tomographic
PIV’, Experiments in Fluids 52(6), 1567–1579.
Schmidt, M. and Lipson, H. (2009), ‘Distilling free-form natural laws from experimental data’,
Science 324(5923), 81–85.
Schmidt, O. T. and Towne, A. (2019), ‘An efficient streaming algorithm for spectral proper
orthogonal decomposition’, Computer Physics Communications 237, 98–109.
Schneider, K. and Vasilyev, O. V. (2010), ‘Wavelet methods in computational fluid dynamics’,
Annual Review of Fluid Mechanics 42, 473–503.
Schölkopf, B. and Smola, A. J. (2002), Learning With Kernels: Support Vector Machines,
Regularization, Optimization, and Beyond, MIT Press, Cambridge.
Schoppa, W. and Hussain, F. (2002), ‘Coherent structure generation in near-wall turbulence’,
Journal of Fluid Mechanics 453, 57–108.
Schulman, J., Levine, S., Abbeel, P., Jordan, M., and Moritz, P. (2015), Trust region policy
optimization, in F. Bach and D. Blei, eds., ‘Proceedings of the 32nd International Conference
on Machine Learning’, Vol. 37 of Proceedings of Machine Learning Research, PMLR, Lille,
France, pp. 1889–1897.
Schulman, J., Wolski, F., Dhariwal, P., Radford, A., and Klimov, O. (2017), ‘Proximal policy
optimization algorithms’, CoRR, arXiv:1707.06347.
Schulze, J., Schmid, P., and Sesterhenn, J. (2009), ‘Exponential time integration using Krylov
subspaces’, The International Journal for Numerical Methods in Fluids 60(6), 591–609.
Schumm, M., Berger, E., and Monkewitz, P. A. (1994), ‘Self-excited oscillations in the wake of
two-dimensional bluff bodies and their control’, Journal of Fluid Mechanics 271, 17–53.
Schwer, D., Lu, P. and Green Jr., W. (2003), ‘An adaptive chemistry approach to modeling
complex kinetics in reacting flows’, Combustion and Flame 133(4), 451–465.
Seber, G. A. (2009), Multivariate Observations, John Wiley & Sons, Hoboken, NJ.
Semaan, R., Fernex, D., Weiner, A., and Noack, B. R. (2020), xROM – A Toolkit for Reduced-
Order Modeling of Fluid Flows, Vol. 1 of Machine Learning Tools for Fluid Mechanics,
Technische Universität Braunschweig, Braunschweig.
Semaan, R., Kumar, P., Burnazzi, M., Tissot, G., Cordier, L. and Noack, B. R. (2015), ‘Reduced-
order modelling of the flow around a high-lift configuration with unsteady Coanda blowing’,
Journal of Fluid Mechanics 800, 72–110.
Semeraro, O., Bagheri, S., Brandt, L., and Henningson, D. S. (2011), ‘Feedback control of three-
dimensional optimal disturbances using reduced-order models’, Journal of Fluid Mechanics
677, 63–102.
Semeraro, O., Bagheri, S., Brandt, L., and Henningson, D. S. (2013), ‘Transition delay in a
boundary layer flow using active control’, Journal of Fluid Mechanics 731, 288–311.
Semeraro, O., Lusseyran, F., Pastur, L. and Jordan, P. (2017), ‘Qualitative dynamics of wave
packets in turbulent jets’, Physical Review Fluids 2(9), 094605.
Semeraro, O. and Mathelin, L. (2016), ‘An open-source toolbox for data-driven linear system
identification’, Technical report, LIMSI-CNRS, Orsay, France.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 442 Data-Driven Fluid Mechanics

Sharma, A. S., Mezić, I., and McKeon, B. J. (2016), ‘Correspondence between Koopman mode
decomposition, resolvent mode decomposition, and invariant solutions of the Navier-Stokes
equations’, Physical Review Fluids 1(3), 032402.
Sieber, M., Paschereit, C. O., and Oberleithner, K. (2016), ‘Spectral proper orthogonal
decomposition’, Journal of Fluid Mechanics 792, 798–828.
Siegel, S. G., Seidel, J., Fagley, C., Luchtenburg, D. M., Cohen, K., and McLaughlin, T. (2008),
‘Low dimensional modelling of a transient cylinder wake using double proper orthogonal
decomposition’, Journal of Fluid Mechanics 610, 1–42.
Sigurd Skogestad, I. P. (2005), Multivariable Feedback Control, John Wiley & Sons, Hoboken,
NJ.
Sillero, J. A. (2014), High Reynolds numbers turbulent boundary layers, PhD thesis, Aeronau-
tics, Universidad Politécnica, Madrid. https://tesis.biblioteca.upm.es/tesis/7538.
Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., Van Den Driessche, G., Schrittwieser,
J., Antonoglou, I., Panneershelvam, V., Lanctot, M. Sander, D., Dominik, G., John, N.,
Nal, K., Ilya, S., Timothy, L., Madeleine, L., Koray, K., Thore, G., and Demis H.
(2016), ‘Mastering the game of Go with deep neural networks and tree search’, Nature
529(7587), 484–489.
Silver, D., Hubert, T., Schrittwieser, J., Antonoglou, I., Lai, M., Guez, A., Lanctot, M., Sifre, L.,
Kumaran, D., Graepel, T., Lillicrap, T., Simonyan, K., and Hassabis, D. (2018), ‘A general
reinforcement learning algorithm that masters Chess, Shogi, and Go through self-play’,
Science 362(6419), 1140–1144.
Silver, D., Lever, G., Heess, N., Degris, T., Wierstra, D., and Riedmiller, M. (2014), Deter-
ministic policy gradient algorithms, in E. P. Xing and T. Jebara, eds., ‘Proceedings of the
31st International Conference on Machine Learning’, Vol. 32 of Proceedings of Machine
Learning Research, PMLR, Bejing, China, pp. 387–395.
Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, I., Huang, A., Guez, A., Hubert, T.,
Baker, L., Lai, M., and Bolton, A. (2017), ‘Mastering the game of Go without human
knowledge’, Nature 550(7676), 354.
Singh, A. P., Medida, S., and Duraisamy, K. (2017), ‘Machine-learning-augmented predictive
modeling of turbulent separated flows over airfoils’, AIAA Journal 55, 2215–2227.
Sirovich, L. (1987), ‘Turbulence and the dynamics of coherent structures. I–Coherent structures.
II–Symmetries and transformations. III–Dynamics and scaling’, Quarterly of Applied
Mathematics 45, 561–571.
Sirovich, L. and Kirby, M. (1987), ‘A low-dimensional procedure for the characterization of
human faces’, Journal of the Optical Society of America A 4(3), 519–524.
Skene, C., Eggl, M., and Schmid, P. (2020), ‘A parallel-in-time approach for accelerating direct-
adjoint studies’, Journal of Computational Physics 429(2), 110033.
Skogestad, S. and Postlethwaite, I. (2005), Multivariable Feedback Control: Analysis and
Design, 2 ed., John Wiley & Sons, Hoboken, NJ.
Skogestad, S. and Postlethwaite, I. (2007), Multivariable Feedback Control: Analysis and
Design, Vol. 2, Wiley, New York.
Smith, J. O. (2007a), Introduction to Digital Filters: With Audio Applications, W3K Publishing,
http://books.w3k.org/.
Smith, J. O. (2007b), Mathematics of the Discrete Fourier Transform (DFT): With Audio
Applications, W3K Publishing, http://books.w3k.org/.
Smith, S. W. (1997), The Scientist & Engineer’s Guide to Digital Signal Processing, California
Technical Publishing, San Diego, www.dspguide.com/.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 References 443

Sommerfeld, A. (1908), ‘Ein Beitrag zur hydrodynamischen Erklärung der turbulenten

Flüssigkeitsbewegungen’, Atti. del 4 Congr. Int. dei Mat. III, pp. 116–124.
Sorokina, M., Sygletos, S., and Turitsyn, S. (2016), ‘Sparse identification for nonlinear optical
communication systems: SINO method’, Optics Express 24(26), 30433–30443.
Spalart, P. R. (1988), ‘Direct simulation of a turbulent boundary layer up to Reθ = 1410’,
Journal of Fluid Mechanics 187, 61–98.
Stengel, R. F. (1994), Optimal Control and Estimation, Dover, New York.
Strang, G. (1968), ‘On the construction and comparison of difference schemes’, SIAM Journal
on Numerical Analysis 5(3), 506–517.
Strang, G. (1988), Linear Algebra and Its Applications, Harcourt Brace Jovanovich College
Publishers, San Diego, CA.
Strang, G. (1989), ‘Wavelets and dilation equations: A brief introduction’, SIAM Review
31(4), 614–627.
Strang, G. (1996), Wavelets and Filter Banks, Wellesley-Cambridge Press, Wellesley, MA.
Strang, G. (2007), Computational Science and Engineering, Wellesley-Cambridge Press,
Wellesley, MA.
Strang, G. and Nguyen, T. (1996), Wavelets and Filter Banks, SIAM, Philadelphia.
Strom, B., Brunton, S. L., and Polagye, B. (2017), ‘Intracycle angular velocity control of cross-
flow turbines’, Nature Energy 2(17103), 1–9.
Stuart, J. (1958), ‘On the non-linear mechanics of hydrodynamic stability’, Journal of Fluid
Mechanics 4, 1–21.
Stuart, J. T. (1971), ‘Nonlinear stability theory’, Annual Review of Fluid Mechanics 3, 347–370.
Succi, S. and Coveney, P. V. (2018), ‘Big data: the end of the scientific method’, Philosophical
Transactions of the Royal Society A 377, 20180145.
Suh, Y. K. (1993), ‘Periodic motion of a point vortex in a corner subject to a potential flow’,
Journal of the Physical Society of Japan 62(10), 3441–3445.
Sundqvist, B. (1991), ‘Thermal diffusivity measurements by Ångström method in a fluid
environment’, International Journal of Thermophysics 12(1), 191–206.
Sutherland, J. C. and Parente, A. (2009), ‘Combustion modeling using principal component
analysis’, Proceedings of the Combustion Institute 32(1), 1563–1570.
Sutherland, J. C., Smith, P. J., and Chen, J. H. (2007), ‘A quantitative method for a priori
evaluation of combustion reaction models’, Combustion Theory and Modelling 11(2), 287–
303.
Sutton, R. S. and Barto, A. G. (2018), Reinforcement Learning: An Introduction, 2nd ed., MIT
Press, Cambridge, MA.
Szegedy, C., Ioffe, S., Vanhoucke, V., and Alemi, A. A. (2017), Inception-v4, inception-resnet
and the impact of residual connections on learning, in ‘Thirty-First AAAI Conference on
Artificial Intelligence’, February 4–9, 2017, San Francisco, CA.
Tadmor, G., Lehmann, O., Noack, B. R., Cordier, L., Delville, J., Bonnet, J.-P., and Morzyński,
M. (2011), ‘Reduced order models for closed-loop wake control’, Philosophical Transac-
tions of the Royal Society A 369(1940), 1513–1524.
Tadmor, G. and Noack, B. R. (2004), Dynamic estimation for reduced Galerkin models of fluid
flows, in ‘The 2004 American Control Conference’, Boston, MA, June 30–July 2, 2004,
pp. 0001–0006. Paper WeM18.1.
Taira, K., Brunton, S., Dawson, S., Rowley, C., Colonius, T., McKeon, B., Schmidt, O.,
Gordeyev, S., Theofilis, V., and Ukeiley, S. (2017), ‘Modal analysis of fluid flows: An
overview’, AIAA Journal 55(12), 4013–4041.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 444 Data-Driven Fluid Mechanics

Taira, K., Hemati, M. S., Brunton, S. L., Sun, Y., Duraisamy, K., Bagheri, S., Dawson, S., and
Yeh, C.-A. (2019), ‘Modal analysis of fluid flows: Applications and outlook’, arXiv preprint
arXiv:1903.05750 .
Takeishi, N., Kawahara, Y., and Yairi, T. (2017), Learning Koopman invariant subspaces for
dynamic mode decomposition, in ‘Advances in Neural Information Processing Systems’,
December 4–9, 2017, Long Beach, CA, pp. 1130–1140.
Takens, F. (1981), Detecting strange attractors in turbulence, in ‘Dynamical systems and
turbulence, Warwick 1980’, Vol. 898 in Lecture Notes in Math., Springer, Heidelberg and
New York, pp. 366–381.
Talamelli, A., Persiani, F., Fransson, J. H., Alfredsson, P. H., Johansson, A. V., Nagib, H. M.,
Rüedi, J.-D., Sreenivasan, K. R., and Monkewitz, P. A. (2009), ‘CICLoPE —- A response to
the need for high Reynolds number experiments’, Fluid Dynamics Research 41(2), 021407.
Tanahashi, M., Kang, S., Miyamoto, T., and Shiokawa, S. (2004), ‘Scaling law of fine scale
eddies in turbulent channel flows up to Reτ = 800’, International Journal of Heat and Fluid
Flow 25, 331–341.
Tang, H., Rabault, J., Kuhnle, A., Wang, Y., and Wang, T. (2020), ‘Robust active flow control
over a range of reynolds numbers using an artificial neural network trained through deep
reinforcement learning’, Physics of Fluids 32(5), 053605.
Tedrake, R., Jackowski, Z., Cory, R., Roberts, J. W., and Hoburg, W. (2009), Learning to fly
like a bird, in ‘14th International Symposium on Robotics Research’, Lucerne, Switzerland.
Tenenbaum, J. B. (2000), ‘A global geometric framework for nonlinear dimensionality reduc-
tion’, Science 290(5500), 2319–2323.
Tennekes, H. and Lumley, J. L. (1972), A First Course in Turbulence, MIT Press, Cambridge,
MA.
Thaler, S., Paehler, L., and Adams, N. A. (2019), ‘Sparse identification of truncation errors’,
Journal of Computational Physics 397, 108851.
Theofilis, V. (2011), ‘Global linear instability’, Annual Review of Fluid Mechanics 43, 319–352.
Thiria, B., Goujon-Durand, S., and Wesfreid, J. E. (2006), ‘The wake of a cylinder performing
rotary oscillations’, Journal of Fluid Mechanics 560, 123–147.
Thormann, R. and Widhalm, M. (2013), ‘Linear-frequency-domain predictions of dynamic-
response data for viscous transonic flows’, AIAA Journal 51(11), 2540–2557.
Tibshirani, R. (1996), ‘Regression shrinkage and selection via the lasso’, Journal of the Royal
Statistical Society. Series B (Methodological) 58(1), 267–288.
Tinney, C. E., Ukeiley, L., and Glauser, M. N. (2008), ‘Low-dimensional characteristics of a
transonic jet. Part 2. Estimate and far-field prediction’, Journal of Fluid Mechanics 615, 53–
92.
Toedtli, S. S., Luhar, M., and McKeon, B. J. (2019), ‘Predicting the response of turbulent
channel flow to varying-phase opposition control: Resolvent analysis as a tool for flow
control design’, Physical Review Fluids 4(7), 073905.
Torrence, C. and Compo, G. P. (1998), ‘A practical guide to wavelet analysis’, Bulletin of the
American Meteorological Society 79(1), 61–78.
Towne, A., Schmidt, O. T., and Colonius, T. (2018), ‘Spectral proper orthogonal decomposition
and its relationship to dynamic mode decomposition and resolvent analysis’, Journal of
Fluid Mechanics 847, 821–867.
Townsend, A. A. (1976), The Structure of Turbulent Shear Flow, Cambridge University Press,
Cambridge.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 References 445

Tran, G. and Ward, R. (2016), ‘Exact recovery of chaotic systems from highly corrupted data’,
arXiv preprint arXiv:1607.01067.
Trefethen, L. and Bau, D. (1997), Numerical Linear Algebra, Society for Industrial and Applied
Mathematics, Philadelphia.
Trefethen, L., Trefethen, A., Reddy, S., and Driscoll, T. (1993), ‘Hydrodynamic stability without
eigenvalues’, Science 261, 578–584.
Tropp, J. A. and Gilbert, A. C. (2007), ‘Signal recovery from random measurements via
orthogonal matching pursuit’, IEEE Transactions on Information Theory 53(12), 4655–
4666.
Tu, J. H. and Rowley, C. W. (2012), ‘An improved algorithm for balanced POD through an ana-
lytic treatment of impulse response tails’, Journal of Computational Physics 231(16), 5317–
5333.
Tu, J. H., Rowley, C. W., Kutz, J. N., and Shang, J. K. (2014), ‘Spectral analysis of fluid flows
using sub-Nyquist-rate PIV data’, Experiments in Fluids 55(9), 1–13.
Tu, J., Rowley, C., Luchtenburg, D., Brunton, S., and Kutz, J. (2014), ‘On dynamic mode
decomposition: theory and applications’, Journal of Computational Dynamics 1(2), 391–
421.
Turns, S. R. (1996), Introduction to Combustion, Vol. 287, McGraw-Hill Companies, New York.
Uruba, V. (2012), ‘Decomposition methods in turbulence research’, EPJ Web of Conferences
25, 01095.
Van den Berg, J. (2004), Wavelets in Physics, Cambridge University Press, Cambridge.
Van Loan, C. F. and Golub, G. H. (1983), Matrix Computations, Johns Hopkins University
Press, Baltimore.
Vassberg, J., Dehaan, M., Rivers, M., and Wahls, R. (2008), Development of a common
research model for applied CFD validation studies, in ‘26th AIAA Applied Aerodynamics
Conference’, August 18-21, 2008, Honolulu, HI, American Institute of Aeronautics and
Astronautics, Reston, VA.
Venturi, D. (2006), ‘On proper orthogonal decomposition of randomly perturbed fields with
applications to flow past a cylinder and natural convection over a horizontal plate’, Journal
of Fluid Mechanics 559, 215–254.
Verma, S., Novati, G., and Koumoutsakos, P. (2018), ‘Efficient collective swimming by
harnessing vortices through deep reinforcement learning’, Proceedings of the National
Academy of Sciences 115(23), 5849–5854.
Vincent, A. and Meneguzzi, M. (1991), ‘The spatial structure and statistical properties of
homogeneous turbulence’, Journal of Fluid Mechanics 225, 1–20.
Viquerat, J., Rabault, J., Kuhnle, A., Ghraieb, H., and Hachem, E. (2019), ‘Direct shape
optimization through deep reinforcement learning’, arXiv preprint arXiv:1908.09885.
Vlachas, P. R., Byeon, W., Wan, Z. Y., Sapsis, T. P., and Koumoutsakos, P. (2018), ‘Data-driven
forecasting of high-dimensional chaotic systems with long short-term memory networks’,
Proceedings of the Royal Society A 474(2213), 20170844.
Vladimir Cherkassky, F. M. M. (2008), Learning from Data, John Wiley & Sons, New York.
Voit, E. O. (2019), ‘Perspective: dimensions of the scientific method’, PLoS: Computational
Biology 15(9), e1007279.
Voltaire, F. (1764), Dictionnaire Philosophique: Atomes, Oxford University Press, 1994.
von Helmholtz, H. (1858), ‘Über integrale der hydrodynamischen gleichungen, welche den
wirbelbewegungen entsprechen’, J. für die reine und angewandte Mathematik 55, 25–55.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 446 Data-Driven Fluid Mechanics

von Kármán, T. (1911), ‘über den Mechanismus des Widerstandes, den ein bewegter Körper in
einer Flüssigkeit erfährt’, Nachrichten der Kaiserlichen Gesellschaft der Wissenschaften zu
Göttingen, pp. 509–517.
von Storch, H. and Xu, J. (1990), ‘Principal oscillation pattern analysis of the 30- to 60-day
oscillation in the tropical troposphere’, Climate Dynamics 4(3), 175–190.
Voropayev, S. I., Afanasyev, Y. D., and Filippov, I. A. (1991), ‘Horizontal jets and vortex dipoles
in a stratified fluid’, Journal of Fluid Mechanics 227, 543–566.
Voss, R., Tichy, L., and R.Thormann (2011), A ROM based flutter prediction process and
its validation with a new reference model, in ‘International Forum on Aeroelasticity and
Structural Dynamics (IFASD-2011)’, June 26–30, 2011, Paris, France.
Vukasonivic, B., Rusak, Z., and Glezer, A. (2010), ‘Dissipative small-scale actuation of a
turbulent shear layer’, Journal of Fluid Mechanics 656, 51–81.
Wahde, M. (2008), Biologically Inspired Optimization Methods: An Introduction, WIT Press,
Southampton.
Wan, Z. Y., Vlachas, P., Koumoutsakos, P., and Sapsis, T. (2018), ‘Data-assisted reduced-order
modeling of extreme events in complex dynamical systems’, PloS One 13(5), e0197704.
Wang, J.-X., Wu, J.-L., and Xiao, H. (2017), ‘Physics-informed machine learning approach for
reconstructing reynolds stress modeling discrepancies based on DNS data’, Physical Review
Fluids 2(3), 034603.
Wang, M. and Hemati, M. S. (2017), ‘Detecting exotic wakes with hydrodynamic sensors’,
arXiv preprint arXiv:1711.10576 .
Wang, Q., Moin, P., and Iaccarino, G. (2009), ‘Minimal repetition dynamic checkpointing algo-
rithm for unsteady adjoint calculation’, SIAM Journal of Scientific Computing 31(4), 2549–
2567.
Wang, R. (2009), Introduction to Orthogonal Transforms, Cambridge University Press, New
York.
Wang, W. X., Yang, R., Lai, Y. C., Kovanis, V., and Grebogi, C. (2011), ‘Predicting catas-
trophes in nonlinear dynamical systems by compressive sensing’, Physical Review Letters
106, 154101-1–154101-4.
Wang, Y., Yao, H., and Zhao, S. (2016), ‘Auto-encoder based dimensionality reduction’, Neuro-
Computing 184, 232–242.
Wang, Z., Akhtar, I., Borggaard, J., and Iliescu, T. (2012), ‘Proper orthogonal decomposition
closure models for turbulent flows: a numerical comparison’, Computer Methods in Applied
Mechanics and Engineering 237, 10–26.
Wang, Z., Bapst, V., Heess, N., Mnih, V., Munos, R., Kavukcuoglu, K., and de Freitas, N.
(2017), Sample efficient actor-critic with experience replay, in ‘5th International Conference
on Learning Representations, ICLR 2017, Toulon, France, April 24–26, 2017, Conference
Track Proceedings’, OpenReview.net.
Wang, Z., Schaul, T., Hessel, M., Van Hasselt, H., Lanctot, M., and De Freitas, N. (2016),
Dueling network architectures for deep reinforcement learning, in ‘Proceedings of the 33rd
International Conference on International Conference on Machine Learning – Volume 48’,
ICML 16, JMLR.org, pp. 1995–2003.
Watkins, C. J. C. H. (1989), Learning from Delayed Rewards, PhD thesis, King’s College,
Cambridge, UK.
Wehmeyer, C. and Noé, F. (2018), ‘Time-lagged autoencoders: Deep learning of slow collective
variables for molecular kinetics’, The Journal of Chemical Physics 148(241703), 1–9.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 References 447

Welch, P. (1967), ‘The use of fast Fourier transform for the estimation of power spectra:
A method based on time averaging over short, modified periodograms’, IEEE Transactions
on Audio and Electroacoustics 15(2), 70–73.
Weller, J., Camarri, S., and Iollo, A. (2009), ‘Feedback control by low-order modelling of the
laminar flow past a bluff body’, Journal of Fluid Mechanics 634, 405.
White, A. P., Zhu, G., and Choi, J. (2013), Linear Parameter-Varying Control for Engineering
Applications, Springer, London.
Whittle, P. (1951), Hypothesis Testing in Time Series Analysis, Almqvist & Wiksells, Uppsala.
Wiener, N. (1948), Cybernetics: Control and Communication in the Animal and the Machine,
Wiley, New York.
Wigner, E. P. (1960), ‘The unreasonable effectiveness of mathematics in the natural sciences’,
Communications on Pure and Applied Mathematics 13, 1–14.
Willcox, K. and Peraire, J. (2002), ‘Balanced model reduction via the proper orthogonal
decomposition’, AIAA Journal 40(11), 2323–2330.
Willert, C. E. and Gharib, M. (1991), ‘Digital particle image velocimetry’, Experiments in
Fluids 10(4), 181–193.
Williams, M., Kevrekidis, I., and Rowley, C. (2015), ‘A data-driven approximation of the
Koopman operator: extending dynamic mode decomposition’, Journal of Nonlinear Science
6, 1307–1346.
Williams, M., Rowley, C., and Kevrekidis, I. (2015), ‘A kernel approach to data-driven
Koopman spectral analysis’, Journal of Computational Dynamics 2(2), 247–265.
Williams, R. J. (1992), ‘Simple statistical gradient-following algorithms for connectionist
reinforcement learning’, Machine Learning 8(3–4), 229–256.
Williamson, D. (1999), Discrete-Time Signal Processing, Springer, London.
Wright, J., Yang, A., Ganesh, A., Sastry, S., and Ma, Y. (2009), ‘Robust face recognition via
sparse representation’, IEEE Transactions on Pattern Analysis and Machine Intelligence
(PAMI) 31(2), 210–227.
Wünning, J. and Wünning, J. (1997), ‘Flameless oxidation to reduce thermal no-formation’,
Progress in Energy and Combustion Science 23(1), 81–94.
Xiao, H., Wu, J.-L., Wang, J.-X., Sun, R., and Roy, C. (2016), ‘Quantifying and reducing model-
form uncertainties in Reynolds-averaged Navier–Stokes simulations: A data-driven, physics-
informed Bayesian approach’, Journal of Computational Physics 324, 115–136.
Xu, H., Zhang, W., Deng, J., and Rabault, J. (2020), ‘Active flow control with rotating
cylinders by an artificial neural network trained by deep reinforcement learning’, Journal
of Hydrodynamics 32(2), 254–258.
Yan, X., Zhu, J., Kuang, M., and Wang, X. (2019), ‘Aerodynamic shape optimization using a
novel optimizer based on machine learning techniques’, Aerospace Science and Technology
86, 826–835.
Yang, J., Wright, J., Huang, T. S., and Ma, Y. (2010), ‘Image super-resolution via sparse
representation’, IEEE Transactions on Image Processing 19(11), 2861–2873.
Yang, Y., Pope, S. B., and Chen, J. H. (2013), ‘Empirical low-dimensional manifolds in
composition space’, Combustion and Flame 160(10), 1967–1980.
Yeh, C.-A. and Taira, K. (2019), ‘Resolvent-analysis-based design of airfoil separation control’,
Journal of Fluid Mechanics 867, 572–610.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i

 448 Data-Driven Fluid Mechanics

Yetter, R. A., Dryer, F., and Rabitz, H. (1991), ‘A comprehensive reaction mechanism for carbon
monoxide/hydrogen/oxygen kinetics’, Combustion Science and Technology 79(1–3), 97–
128.
Yeung, E., Kundu, S. and Hodas, N. (2017), ‘Learning deep neural network representations for
Koopman operators of nonlinear dynamical systems’, arXiv preprint arXiv:1708.06850.
Zdravkovich, M. (1987), ‘The effects of interference between circular cylinders in cross flow’,
Journal of Fluids and Structures 1(2), 239–261.
Zdybał, K., Armstrong, E., Parente, A., and Sutherland, J. C. (2020), ‘PCAfold: Python software
to generate, analyze and improve PCA-derived low-dimensional manifolds’, SoftwareX
12, 100630.
Zdybał, K., Sutherland, J. C., Armstrong, E., and Parente, A. (2021), ‘State-space informed data
sampling on combustion manifolds’, Combustion and Flame (manuscript in preparation).
Zdybał, K., Sutherland, J. C. & Parente, A. (2022), ‘Manifold-informed state vector subset for
reduced-order modeling’, Proceedings of the Combustion Institute .
Zdybał, K., Armstrong, E., Sutherland, J. C. & Parente, A. (2022), ‘Cost function for low-
dimensional manifold topology assessment’, Scientific Reports 12(1), 1–19.
Zebib, A. (1987), ‘Stability of viscous flow past a circular cylinder’, Journal of Engineering
Mathematics 21, 155–165.
Zhang, H.-Q., Fey, U., Noack, B. R., König, M., and Eckelmann, H. (1995), ‘On the transition
of the cylinder wake’, Physics of Fluids 7(4), 779–795.
Zhang, W., Wang, B., Ye, Z., and Quan, J. (2012), ‘Efficient method for limit cycle flutter
analysis based on nonlinear aerodynamic reduced-order models’, AIAA Journal 50(5), 1019–
1028.
Zheng, P., Askham, T., Brunton, S. L., Kutz, J. N., and Aravkin, A. Y. (2019), ‘Sparse relaxed
regularized regression: SR3’, IEEE Access 7(1), 1404–1423.
Zhong, Y. D. and Leonard, N. (2020), ‘Unsupervised learning of lagrangian dynamics from
images for prediction and control’, Advances in Neural Information Processing Systems 33.
Zhou, K. and Doyle, J. C. (1998), Essentials of Robust Control, Prentice Hall, Upper Saddle
River, NJ.
Zhou, Y., Fan, D., Zhang, B., Li, R., and Noack, B. R. (2020), ‘Artificial intelligence control of
a turbulent jet’, Journal of Fluid Mechanics 897, 1–46.
Zimmermann, R. and Görtz, S. (2010), ‘Non-linear reduced order models for steady aerody-
namics’, Procedia Computer Science 1(1), 165–174.
Zimmermann, R. and Görtz, S. (2012), ‘Improved extrapolation of steady turbulent aerody-
namics using a non-linear POD-based reduced order model’, The Aeronautical Journal
116(1184), 1079–1100.
Zimmermann, R., Vendl, A., and Görtz, S. (2014), ‘Reduced-order modeling of steady flows
subject to aerodynamic constraints’, AIAA Journal 52(2), 255–266.
Zou, H. and Hastie, T. (2005), ‘Regularization and variable selection via the elastic net’, Journal
of the Royal Statistical Society: Series B (Statistical Methodology) 67(2), 301–320.

h d i 0 0 9 8 088962 4 028 bli h d li b C b id i i