DATA-DRIVEN MODEL DISCOVERY AND
COORDINATE EMBEDDINGS FOR
PHYSICAL SYSTEMS
ICERM 2019
J. Nathan Kutz
Department of Applied Mathematics
University of Washington
Email: [email protected]
Web: faculty.washington.edu/kutz
Model Discovery
Steven Brunton
Mechanical Engineering
University of Washington
Joshua Proctor
Institute for Disease Modeling
Mathematical Framework
Dynamics State-space Parameters
x
Dynamics Stochastic effects
Measurement
Measurement model Measurement noise
What Could the Right Side Be?
Limited by your imagination
2nd degree polynomials
Sparse Identification of Nonlinear Dynamics (SINDy)
Identifying ROMs
Discovery Innovations
Schaeffer et al -- corrupt data, PDEs, integral formulation, convergence
Dongbin Xiu & co-workers (2018) – Sampling strategies
Guang Lin & co-workers (2018) -- Uncertainty Metrics
Hod Lipson and co-workers (2006) — Symbolic/genetic regression
Karniadakis, Raissai, Perdikaris …. — Neural Nets
Zheng, Askham, Brunton, Kutz & Aravkin (2018) – SR3 sparse relaxed
regularized regression (for SINDy, LASSO, CS, TV, Matrix Completion …)
Manifolds and Embeddings
Observables & Coordinates
Bernard Koopman 1931
Mezic (2004)
Coifman, Kevrekidis, co-workers
Koopman Invariant Subspaces
Brunton, Proctor & Kutz, PLOS ONE (2018)
Burgers’ Equation
Cole-Hopf
Kutz, Proctor & Brunton, Complexity (2018)
Dynamic Mode Decomposition
Travis Askham
Askham & Kutz, SIADS (2018)
Approximate Dynamical Systems
Linear dynamics
(equation-free)
Eigenfunction
expansion
Least-square fit
DMD with Control
Input
Input
Snapshots
DMD
generalization
Proctor, Brunton & Kutz, SIADS (2016)
Koopman vs DMD: All about Observables!
Neural Nets
The Zoo
NN
Zoo
NNs for Koopman Embedding
Bethany Lusch Lusch et al. Nat. Comm (2018)
Handling the Continuous
NNs for Koopman Spectra
Embedding
Bethany Lusch Lusch, Kutz
Lusch& Brunton, arxiv
et al. Nat. (2017)
Comm (2018)
The Pendulum
Flow Around a Cylinder
Autoencoder + SINDy
Kathleen Champion
Fast Learning
Charles Delahunt
Moth Olfactory System
Learning New Odors
Sparsity for Learning
Rapid Learning in NNs
Comparisons
Decoder Networks
Structure of Mapping
Linear Maps: SVD (left singular vector) defines layer
Erichson, Mathelin, Brunton, Mahoney & Kutz (2019)
Mathematical Framework
State space
Measurements
Mapping
Approximate the full state space from limited measurements
Optimization
Linear Maps
Singular value decomposition Data
Linear measurements H
Optimize (least-squares)
Optimal Placement
Point measurements
Optimal Sensors via QR pivots
Manohar, Kutz & Brunton (2018) IEEE Control Systems Magazine
Clark, Askham, Brunton & Kutz (2019) IEEE Sensors
Nonlinear Mapping
General Form: Compositional Layers
Universal Approximators: Hornik 1990
Shallow Layer Mapping
Two Layers
Composition
Erichson, Mathelin, Brunton, Mahoney & Kutz (2019) SIAM
Activation Functions
Linear vs Nonlinear Maps
Improved Interpretability of Modes
Improved Performance
Robustness to Noise
Improved Performance
Super Resolution Analysis
Conclusion
Model Discovery: Sparse regression provides parsimonious dynamical models
Coordinates: Learning Koopman embeddings can provide optimal basis for dynamics
Neural Networks: Structure and function matter
- Connect discovery and coordinates
- Structure can lead to fast (one-shot) learning with limited data
- Discovery of improved coordinate embeddings through decoding
COMING SOON: A multi scale physics challenge set
