Physics-informed dynamic mode decomposition

Peter J. Baddoo, Benjamín Herrmann, Beverley McKeon, J. Nathan Kutz, Steven L. BruntonProceedings of the Royal Society A Mathematical Physical and Engineering Sciences 2023 · DOI

TL;DR

This paper introduces "physics-informed dynamic mode decomposition" (piDMD), a novel method that embeds known physical principles (like conservation laws and symmetries) directly into the dynamic mode decomposition (DMD) framework. By constraining the learned model to adhere to these principles, piDMD produces more accurate, robust, and generalizable models compared to standard DMD, especially in the presence of noise or limited data.

What Problem Does This Paper Address?

Dynamic Mode Decomposition (DMD) is a powerful tool for extracting coherent structures and dynamics from high-dimensional time-series data. However, standard DMD has several limitations: it's highly sensitive to noise, struggles with traveling wave phenomena, and can produce models that violate fundamental physical laws, leading to poor generalization outside the training data. This paper addresses these shortcomings by developing a method that explicitly incorporates physical knowledge into the DMD process.

Methods

The core innovation is "physics-informed DMD" (piDMD). Instead of simply finding the best low-rank linear approximation of the system's dynamics (as in standard DMD), piDMD reformulates the problem as a Procrustes problem. This means the learned model (represented by a matrix $A$) is constrained to lie on a specific "matrix manifold" that embodies known physical principles. The authors derive closed-form solutions and efficient algorithms for piDMD under various physical constraints, including:

  • Shift-invariance: The system behaves the same regardless of spatial translation.
  • Conservation laws: Quantities like energy or mass are preserved.
  • Self-adjointness: Often related to energy preservation and real eigenvalues.
  • Locality: Interactions are primarily between nearby states.
  • Causality: Effects follow causes in a specific direction.

These principles are translated into specific matrix structures (e.g., circulant, symmetric, tridiagonal, upper triangular), and the optimization problem is solved to find a matrix $A$ that best fits the data while adhering to these structural constraints.

Datasets

The paper demonstrates piDMD on several challenging problems, including:

  • Advection equation: Used to illustrate shift-invariance and compare standard DMD with piDMD.
  • Energy-preserving fluid flow: Specifically, flow past a cylinder at Re=100, using vorticity measurements.
  • Schrödinger equation: Modeling a quantum Hamiltonian with a hexagonal well.
  • Solute advection-diffusion: Used to demonstrate data-driven resolvent analysis.
  • Three-dimensional transitional channel flow: A complex fluid dynamics problem.

The paper does not introduce new datasets but rather applies piDMD to existing benchmark problems and simulation data.

Key Insights and Ideas

  • Physics as a Constraint: The most significant insight is that physical principles can be elegantly incorporated as constraints on the learned model's structure (matrix manifold), rather than just as penalties in a loss function. This leads to more principled and robust models.
  • Procrustes Problem Connection: Reframing the DMD regression as a Procrustes problem provides a powerful theoretical foundation and allows leveraging existing mathematical tools for solving constrained optimization problems.
  • Improved Generalization and Robustness: By restricting the model space to physically plausible structures, piDMD inherently reduces overfitting, making models less sensitive to noise and better at generalizing to unseen data.
  • Efficient Algorithms: For many physical structures, the authors derive efficient, closed-form solutions (often leveraging Fast Fourier Transforms or other linear algebra techniques) that are computationally competitive with or superior to standard DMD.
  • Spectral Accuracy: piDMD consistently achieves higher accuracy in identifying the true spectral properties (eigenvalues and eigenvectors) of the underlying dynamical system compared to standard DMD.

Key Results

  • Superior Spectral Identification: In all tested examples, piDMD significantly outperforms standard DMD in accurately identifying the system's eigenvalues and eigenvectors. For instance, in the channel flow example, piDMD captures the true spectrum of the linearized Navier-Stokes equations, while standard DMD produces spurious eigenvalues.
  • Reduced Noise Sensitivity: piDMD models exhibit lower variance and are less affected by noise in the training data, as demonstrated in the flow past a cylinder and Schrödinger equation examples.
  • Accurate Predictions: piDMD models provide more faithful predictions of future system dynamics, as illustrated by the advection equation example where standard DMD fails catastrophically.
  • Efficient Computation: For shift-invariant systems, piDMD leverages FFTs to achieve efficient computation, often outperforming standard DMD in terms of speed and memory requirements.
  • Physical Structure Discovery: The method can implicitly discover or enforce physical structures like locality (tridiagonal matrices) from data, as shown in the convection-diffusion example.

Limitations and Future Work

  • Physics Understanding: The effectiveness of piDMD relies on having some prior knowledge of the physical principles governing the system. For systems where physics is poorly understood, identifying the correct manifold can be challenging.
  • Manifold Complexity: For intricate geometries or multi-dimensional systems, translating physical principles into a tractable matrix manifold can be difficult.
  • Optimization Challenges: While many exact solutions are provided, optimization on complex manifolds can still be computationally demanding.
  • Noise Characterization: A full characterization of piDMD's sensitivity to noise, especially when both input (X) and output (Y) data are noisy, requires further investigation.
  • Extension to Non-Linear Systems: The current framework primarily focuses on linear dynamical systems. Extending piDMD to non-linear systems remains an open area.
  • Data Requirements: While piDMD generally requires less data than standard DMD, the exact data requirements for different manifold constraints are still an area for exploration.

Summary updated

Abstract

In this work, we demonstrate how physical principles—such as symmetries, invariances and conservation laws—can be integrated into the dynamic mode decomposition (DMD). DMD is a widely used data analysis technique that extracts low-rank modal structures and dynamics from high-dimensional measurements. However, DMD can produce models that are sensitive to noise, fail to generalize outside the training data and violate basic physical laws. Our physics-informed DMD (piDMD) optimization, which may be formulated as a Procrustes problem, restricts the family of admissible models to a matrix manifold that respects the physical structure of the system. We focus on five fundamental physical principles—conservation, self-adjointness, localization, causality and shift-equivariance—and derive several closed-form solutions and efficient algorithms for the corresponding piDMD optimizations. With fewer degrees of freedom, piDMD models are less prone to overfitting, require less training data, and are often less computationally expensive to build than standard DMD models. We demonstrate piDMD on a range of problems, including energy-preserving fluid flow, the Schrödinger equation, solute advection-diffusion and three-dimensional transitional channel flow. In each case, piDMD outperforms standard DMD algorithms in metrics such as spectral identification, state prediction and estimation of optimal forcings and responses.
BibTeX
@inproceedings{baddoo2023physics,
  author = {Peter J. Baddoo and Benjamín Herrmann and Beverley McKeon and J. Nathan Kutz and Steven L. Brunton},
  title = {{Physics-informed dynamic mode decomposition}},
  booktitle = {Proceedings of the Royal Society A Mathematical Physical and Engineering Sciences},
  year = {2023},
  doi = {10.1098/rspa.2022.0576},
}
Figure