Online Physics-Informed Dynamic Mode Decomposition: Theory and Applications

Biqi Chen, Ying Wang · 2024

Read Online Physics-Informed Dynamic Mode Decomposition: Theory and Applications by Biqi Chen, Ying Wang (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-03609.txt.

Introduction

Biqi Chen and Ying Wang rebuild a data-driven modeling method so it learns on the fly and obeys physics, in a 2024 paper. Dynamic mode decomposition extracts simple patterns from complex flows, but it is slow, noise-shy, and free to break physical law. Their version, OPIDMD, recasts the task as a convex optimization with one best answer.

A Bayesian reading of physics-informed decomposition shows how constraints bend the learned operator. Online gradient methods then update the model in real time under varied physical rules. In short-term forecasting tests, including a noisy Lorenz system scored at 0.991, the method beat its exact, online, and physics-informed rivals. The authors aim it at live simulation and control of complex systems.

Catalog record

CREATORBiqi Chen, Ying Wang
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE117 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-03609

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Online Physics-Informed Dynamic Mode Decomposition: Theory and Applications Biqi Chen, Ying Wang Abstract Dynamic Mode Decomposition (DMD) has received increasing research attention due to its capability to analyze and model complex dynamical systems. However, it faces challenges in computational efficiency, noise sensitivity, and difficulty adhering to physical laws, which negatively affect its performance. Addressing these issues, we present Online Physics-informed DMD (OPIDMD), a novel adaptation of DMD into a convex optimization framework. This approach not only ensures convergence to a unique global optimum, but also enhances the efficiency and accuracy of modeling dynamical systems in an online setting. Leveraging the Bayesian DMD framework, we propose a probabilistic interpretation of Physics-informed DMD (piDMD), examining the impact of physical constraints on the DMD linear operator. Further, we implement online proximal gradient descent and formulate specific algorithms to tackle problems with different physical constraints, enabling real-time solutions across various scenarios. Compared with existing algorithms such as Exact DMD, Online DMD, and piDMD, OPIDMD achieves the best prediction performance in short-term forecasting, e.g. an R 2 R^{2} value of 0.991 for noisy Lorenz system. The proposed method employs a time-varying linear operator, offering a promising solution for the real-time simulation and control of complex dynamical systems. keywords Dynamic mode decomposition, Dynamical systems, Physics-informed, Proximal gradient descent, Bias-variance trade-off, Bayesian DMD 1 Introduction In the realm of scientific and engineering challenges, the ability to understand and predict the behaviours of complex systems is highly valuable Brunton and Kutz (2022) ; Xu et al. (2020) ; Fu and Wang (2023) . These systems are often intricate, composed of numerous interacting components that exhibit different dynamic behaviors Abdullahi et al. (2022) ; Achillopoulou et al. (2020) . Gaining an understanding of these systems is vital for a variety of purposes, such as evaluating current conditions, enhancing performance efficiency, predicting future conditions, and formulating control strategies Wang and Hao (2015) ; Zhou et al. (2023) . However, traditional analysis methods, including linear stability analysis and modal analysis, often find it difficult to cope with the intricacies of these systems Brunton et al. () . This is particularly true in cases where the system displays nonlinear characteristics or the analysis involves handling extensive datasets. DMD emerges as a robust and flexible tool in this context, providing a novel approach to uncovering the dynamic essence of time series data from these complex systems Kutz et al. (2016) ; Schmid et al. (2011) ; Schmid () . Due to its capability to reveal intricate spatial and temporal patterns within high-dimensional datasets, DMD has been widely used to analyse nonlinear dynamical systems Askham and Kutz () ; Tu et al. () ; Proctor et al. () ; Korda and Mezić () ; Bai et al. () ; Grosek and Kutz () . Its applications span fluid dynamics Schmid () , neuroscience Brunton et al. (2016) , and epidemiology Proctor and Eckhoff (2015) , demonstrating its versatility and depth of insight into system behaviors Rowley and Dawson () . However, it encounters challenges in terms of computational efficiency, noise sensitivity, and difficulty in complying with physical laws, which adversely affect its performance in real-time applications Dawson et al. () . Integrating physical principles into DMD has led to substantial improvements in model accuracy and relevance. These advances are marked by the development of sparsity-promoting algorithms for Koopman-invariant subspaces, which enhance the identification of key physical processes in nonlinear systems Pan et al. (2021) . Innovations such as piDMD Baddoo et al. () , physics-fusion DMD Yina et al. (2023) , physics-Fusion DMD Yina et al. (2023) and geometry-informed DMD Li et al. (2023) address the integration of various types of physical information and the reduction of noise interference, achieving accurate dynamic modeling and improving prediction accuracy. Researchers have demonstrated its application in real-time and predictive settings by creating surrogate models with DMD in multiple real or simulated cases, significantly enhancing the model’s interpretability and reliability Sharma et al. () ; Karniadakis et al. () ; Baddoo et al. () . However, the influence of physical constraints on the DMD linear approximation of dynamics remains an area that requires further investigation, providing an open avenue for exploration despite these advancements Baddoo et al. () . Although piDMD successfully enhances the accuracy and noise resistance of DMD by extending the low-rank constraint to a manifold constraint, it retains certain limitations. Like standard DMD, piDMD was originally solved using a non-probabilistic, linear model-based approach. Consequently, this method lacks optimality guarantees, and its statistical performance remains undetermined without further assumptions about the probabilistic data structure. Bayesian DMD has improved the method, introducing a probabilistic approach to modal analysis through prior knowledge integration and uncertainty quantification Takeishi et al. () . Extensions such as Bayesian sparse DMD and probabilistic DMD mixtures enhance DMD by accounting for observation noise and enabling posterior analysis. Variational Matrix Factorization in Bayesian DMD aligns with traditional DMD while providing a fully Bayesian framework, offering notable benefits for incomplete data handling Kawashima and Hino (2023) . This synergy between Bayesian methods and optimized DMD algorithms produces robust and stable forecasts for spatio-temporal data, emphasizing the quantification of essential uncertainty in both spatial and temporal dimensions Sashidhar and Kutz () . As data volumes from simulations and experiments expand, the computational costs of modal extraction algorithms such as DMD increase significantly Erichson et al. () . This increase in computational complexity is directly related to the growing quantity and dimensions of the incoming data, resulting in substantial computational burdens Zhang et al. () . Handling large-scale measurement data thus becomes challenging, especially in scenarios requiring online or real-time analysis. This initiative is supported by various studies, especially compressed sensing-like approaches to data compression and algorithmic optimization Brunton et al. (2016) ; Bai et al. (2020) ; Erichson et al. (2019) ; Zhang et al. (2022) . The essence of these efforts is to tailor DMD for high fidelity in aerodynamics, fluid dynamics, and large-scale thermofluid systems, catering to both overconstrained and underconstrained data sets. Recent efforts to optimize DMD for online applications have primarily focused on a common scenario in which the system evolves slowly over time while data arrives continuously Zhang et al. () ; Hemati et al. () . Notable progress in wind tunnel testing, EEG data examination, and wind farm forecasting demonstrates the enhanced efficiency and real-world relevance of these methods Nedzhibov () ; Matsumoto et al. (2018) ; Alfatlawi and Srivastava (2019) ; Nedzhibov (2023) ; Liew et al. () . Despite these advancements, integrating physical constraints into DMD’s online learning algorithms remains an unaddressed challenge. In this study, we introduce OPIDMD, a novel method for the online analysis of time-varying systems. It represents a linear operator that best captures the data characteristics at the current time. By integrating the physical rigor of DMD with advancements in online analytics, OPIDMD offers an effective solution for accurately capturing the system state in real time while demonstrating significant advantages in noise robustness and generalization compared to existing online algorithms. It is particularly well-suited for low-dimensional, nonlinear dynamical systems with large amounts of sample data. We leverage convex optimization techniques to integrate physical constraints with the probabilistic modeling of the linear approximation of dynamics,

Read the complete text