A Dynamic Mode Decomposition Extension for the Forecasting of Parametric Dynamical Systems

Francesco Andreuzzi, Nicola Demo, Gianluigi Rozza

TL;DR

This paper introduces a novel extension of Dynamic Mode Decomposition (DMD) to handle parametric dynamical systems. By combining Proper Orthogonal Decomposition (POD) for dimensionality reduction with regression techniques, the method can accurately forecast the behavior of systems with varying parameters and time, even for unseen configurations. This opens doors for efficient real-time analysis and prediction in complex scientific and engineering applications.

What Problem Does This Paper Address?

Many real-world dynamical systems are not only time-dependent but also influenced by various parameters. For instance, the flow around an object can change drastically with its shape or the fluid's velocity. Predicting the system's future behavior in such scenarios, especially for parameter values not encountered during training, is crucial for design, control, and understanding. Traditional methods often struggle with the high dimensionality and computational cost associated with exploring this parameter space over time. While DMD is excellent for forecasting time-dependent systems, extending it to handle parameter variations efficiently and accurately has been a significant challenge. This paper aims to bridge this gap by developing a data-driven approach that can predict the system's evolution for both new parameters and future time instances.

Methods

The core of this work lies in extending Dynamic Mode Decomposition (DMD) to parametric systems. The proposed methodology involves a two-stage process:

  • Offline Stage:
    • Dimensionality Reduction (POD): High-fidelity snapshots of the system for various parameter values and time instances are first reduced in dimensionality using Proper Orthogonal Decomposition (POD). This captures the dominant spatial structures and modal coefficients.
  • DMD Application: DMD is then applied to these reduced-order snapshots. Two approaches are presented:
  • Monolithic: A single DMD operator is learned to represent the dynamics across all parameter configurations. This is beneficial when parameters lead to similar system behaviors.
  • Partitioned: Separate DMD operators are learned for each parameter configuration. This is more robust when parameter variations lead to significantly different dynamics.
  • Regression for Parametric Mapping: A regression technique is employed to learn the mapping between the parameter space and the reduced-order snapshots. This allows for the prediction of reduced snapshots for unseen parameter values.
  • Online Stage:
    • Prediction of Reduced Snapshots: The learned DMD operators are used to predict the reduced snapshots at future time instants for the training parameters.
  • Interpolation for Untested Parameters: The regression model is used to interpolate the predicted reduced snapshots for untested parameter values.
  • Full-Dimensional Reconstruction: The interpolated reduced snapshots are then mapped back to the full-dimensional space using the POD basis, providing the final prediction.

Additionally, a knowledge-based DMD operator stabilization technique is introduced. This method addresses the issue of DMD producing non-physical modes by stabilizing the DMD eigenvalues, ensuring more robust forecasting, especially for stable systems.

Datasets

The paper does not introduce new benchmark datasets. Instead, it utilizes synthetic data generated from three distinct numerical experiments:

  1. A simple dynamical system with linear parameter dependency.
  2. A heat conductivity problem with nonlinear parameter dependency.
  3. A Navier-Stokes flow problem around a cylinder, exhibiting complex nonlinear parameter dependency.

These examples are chosen to demonstrate the method's effectiveness across varying levels of complexity and parameter influence.

Key Insights and Ideas

  • POD + DMD + Regression Synergy: The paper convincingly demonstrates that combining POD for dimensionality reduction, DMD for temporal forecasting, and regression for parameter mapping creates a powerful framework for parametric system prediction. This multi-stage approach effectively tackles the curse of dimensionality inherent in parametric problems.
  • Monolithic vs. Partitioned DMD: The distinction between monolithic and partitioned DMD approaches is crucial. The choice depends on the nature of the parameter dependency, with monolithic being efficient for similar dynamics and partitioned offering better robustness for disparate behaviors.
  • Stabilization for Robustness: The proposed eigenvalue stabilization technique for DMD operators is a key innovation. It significantly improves forecasting accuracy by filtering out non-physical or exponentially growing modes, which are common issues in DMD, especially when dealing with noisy data or limited training sets.
  • Equation-Agnostic Nature: A significant advantage of this method is its data-driven nature. It does not require explicit knowledge of the underlying physical equations, making it applicable to a wide range of problems where high-fidelity simulations or experimental data are available.

Key Results

The paper presents numerical results from three test cases, showcasing the method's effectiveness:

  • Numerical Study 1 (Linear System): The method accurately predicts the system's behavior for untested parameters and future time instants, with a small error attributed to the limited number of POD modes used.
  • Numerical Study 2 (Heat Conductivity): The approach demonstrates good performance, although it requires a higher number of POD modes to capture the system's complexity. The results show reasonable accuracy in predicting the system's state for unseen parameters and time.
  • Numerical Study 3 (Navier-Stokes): This complex case highlights the method's ability to capture dominant structures and predict wave-like behaviors. While the error increases in regions with complex flow dynamics or due to a limited number of POD modes, the method successfully captures key phenomena and shows improved accuracy with stabilization. The stabilization technique significantly reduces prediction errors for future time instants by mitigating the exponential growth of spurious modes.

Limitations and Future Work

  • POD Mode Selection: The accuracy of the method is sensitive to the number of POD modes chosen. Selecting an optimal number that balances fidelity and computational cost remains a challenge.
  • Regression Model Choice: The performance of the online stage depends on the chosen regression technique. The paper explores linear, cubic, nearest neighbor, and Gaussian Process Regression (GPR), suggesting that further investigation into optimal interpolators for specific problems is warranted.
  • Clustering Parameters: For systems with highly diverse parameter behaviors, the paper suggests that clustering parameters into groups with similar dynamics could further improve accuracy. This would involve developing techniques to automatically identify and group these parameters.
  • Noise-tolerant DMD Variants: The authors acknowledge that noise in the data can affect DMD results and suggest exploring noise-tolerant DMD variants as a future direction to enhance robustness and accuracy.
  • Autoencoder-based Reduction: The paper mentions the potential of using autoencoders as an alternative to POD for dimensionality reduction, which could offer advantages in capturing complex nonlinear structures.

Summary updated

Abstract

Dynamic mode decomposition (DMD) has recently become a popular tool for the non-intrusive analysis of dynamical systems. Exploiting Proper Orthogonal Decomposition (POD) as a dimensionality reduction technique, DMD is able to approximate a dynamical system as a sum of spatial basis evolving linearly in time, thus enabling a better understanding of the physical phenomena and forecasting of future time instants. In this work we propose an extension of DMD to parameterized dynamical systems, focusing on the future forecasting of the output of interest in a parametric context. Initially all the snapshots -- for different parameters and different time instants -- are projected to a reduced space; then DMD, or one of its variants, is employed to approximate reduced snapshots for future time instants. Exploiting the low dimension of the reduced space the predicted reduced snapshots are then combined using regression techniques, thus enabling the possibility to approximate any untested parametric configuration in future. This paper depicts in detail the algorithmic core of this method; we also present and discuss three test cases for our algorithm: a simple dynamical system with a linear parameter dependency, a heat problem with nonlinear parameter dependency and a fluid dynamics problem with nonlinear parameter dependency.
BibTeX
@article{andreuzzi2023dynamic,
  author = {Francesco Andreuzzi and Nicola Demo and Gianluigi Rozza},
  title = {{A Dynamic Mode Decomposition Extension for the Forecasting of Parametric Dynamical Systems}},
  journal = {SIAM J. Appl. Dyn. Syst},
  year = {2023},
  doi = {10.1137/22m1481658},
}
Figure