Data-Driven Science and Engineering: Machine Learning, Dynamical Systems, and Control (2nd ed.)

My reading notes

Why it matters

This is the reference text behind the SINDy/DMD/Koopman/POD-DEIM baselines that any physics-informed or spatiotemporal method (Pi-DPM, MIRROR, surrogate spatial nets) should beat, and Chapter 14 directly grounds PINNs, SINDy autoencoders, and neural-operator learning. It gives Arun the classical-dynamics vocabulary and reduced-order-model machinery to position generative/physics-aware work against strong, interpretable baselines.

Summary

Brunton and Kutz (both at the University of Washington) present an integrated graduate-level textbook (761 pages, 14 chapters across 4 parts) that fuses data-driven methods and machine learning with classical engineering mathematics, dynamical systems, and control theory. The recurring thesis is that high-dimensional complex systems (turbulence, climate, brain, epidemiology, robotics) hide dominant low-dimensional structure, and that finding the right coordinate transforms plus optimization unlocks sensing, prediction, estimation, and control. The authors frame data-driven discovery as a "fourth paradigm" that augments rather than replaces empirical, analytical, and computational science.

Part I builds the linear-algebra and signal foundations: the SVD (with PCA, randomized SVD, tensor decompositions), Fourier/wavelet/Laplace transforms and image processing, and sparsity / compressed sensing (sparse regression, robust PCA, sparse sensor placement). Part II covers core ML: regression and model selection (cross-validation, information criteria, the Pareto front), clustering and classification (k-means, EM/mixture models, SVM, random forests), and a neural-networks chapter spanning backprop, SGD, CNNs, RNNs, autoencoders, and GANs.

Part III turns to dynamics and control: data-driven dynamical systems (DMD, SINDy, Koopman operator theory), classical linear control (LQR, Kalman filter, LQG, robust/frequency-domain methods), and balanced model reduction plus system identification. Part IV reaches research-level material: data-driven and machine-learning control (MPC, extremum seeking), a new reinforcement-learning chapter (model-free Q-learning, deep RL), reduced-order models (POD-Galerkin, gappy POD, DEIM, neural-network time-steppers), and a new physics-informed ML chapter covering SINDy autoencoders, Koopman forecasting, operator learning, PINNs, and learned coarse-graining for PDEs.

The second edition adds the RL and physics-informed ML chapters, parallel Python and MATLAB code throughout (with R available online), per-chapter homework that ranges up to reproducing modern research papers, and companion YouTube video lectures. Each chapter is largely modular and designed to double as a stand-alone short course.

Key ideas

• Dominant low-dimensional structure: complex high-dimensional systems admit compact representations via SVD/PCA, POD, and learned coordinate transforms, which is the backbone for both modeling and control.
• Data-driven dynamics: DMD, SINDy (sparse identification of nonlinear dynamics), and Koopman operator theory extract interpretable governing equations and linear embeddings directly from time-series data.
• Sparsity and compressed sensing: l1/sparse-regression machinery enables reconstruction from few measurements, robust PCA, and optimal/sparse sensor placement.
• Control integration: classical control (LQR, Kalman, LQG, MPC, robust control) is unified with data-driven model reduction and a new reinforcement-learning chapter (Q-learning, deep RL).
• Reduced-order models: POD-Galerkin, gappy POD, and DEIM (plus randomized linear algebra and neural-net surrogates) give cheap, accurate surrogates for expensive PDE simulations.
• Physics-informed ML (2nd-ed chapter): PINNs, SINDy autoencoders, Koopman forecasting, and neural operator learning embed physical constraints and symmetries into trainable models.

Takeaways for my work

• Use the SINDy/DMD/Koopman and POD-DEIM chapters as the classical, interpretable baselines that Pi-DPM, MIRROR, and PINN-style work must outperform; cite them when justifying why a generative or learned approach is needed.
• The reduced-order-model material (gappy POD, DEIM, randomized SVD, neural-net time-steppers) is a direct template for building fast surrogate spatial networks and for the surrogate side of the MIRROR anomaly platform.
• Sparse sensor placement and compressed-sensing chapters map onto spatial-data settings with sparse/denied observations, relevant to trajectory-gap and deception-detection work.
• Companion Python/MATLAB code on github.com/dynamicslab and per-section YouTube video lectures make this a practical pull-from reference for runnable baselines, not just theory.