Non-intrusive Learning of Physics-Informed Spatio-temporal Surrogate for Accelerating Design

Sudeepta Mondal, Soumalya Sarkar · 2026

Read Non-intrusive Learning of Physics-Informed Spatio-temporal Surrogate for Accelerating Design by Sudeepta Mondal, Soumalya Sarkar (2026) 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-2604-14424.txt.

Introduction

Mondal and Sarkar build a physics-based spatio-temporal surrogate for design, where costly detailed simulations bottleneck the process. Purely data-driven stand-ins run far faster than simulations but falter outside their training range. The team constrains a Koopman autoencoder model with the system's own physics, learning dynamics without intruding into the simulator. A second surrogate predicts how the Koopman operator behaves in a set time window under new conditions. The team evaluates the framework on two-dimensional incompressible flow around a cylinder.

Catalog record

CREATORSudeepta Mondal, Soumalya Sarkar
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE19567
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.14424

Browse the full library · About the Tower of Babel

Full text

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

Non-intrusive Learning of Physics-Informed Spatio-temporal Surrogate for Accelerating Design Sudeepta Mondal, Soumalya Sarkar Abstract Most practical engineering design problems involve nonlinear spatio-temporal dynamical systems. Multi-physics simulations are often performed to capture the fine spatio-temporal scales which govern the evolution of these systems. However, these simulations are often high-fidelity in nature, and can be computationally very expensive. Hence, generating data from these expensive simulations becomes a bottleneck in an end-to-end engineering design process. Spatio-temporal surrogate modeling of these dynamical systems has been a popular data-driven solution to tackle this computational bottleneck. This is because accurate machine learning models emulating the dynamical systems can be orders of magnitude faster than the actual simulations. However, one key limitation of purely data-driven approaches is their lack of generalizability to inputs outside the training distribution. In this paper, we propose a physics-informed spatio-temporal surrogate modeling (PISTM) framework constrained by the physics of the underlying dynamical system. The framework leverages state-of-the-art advancements in the field of Koopman autoencoders to learn the underlying spatio-temporal dynamics in a non-intrusive manner, coupled with a spatio-temporal surrogate model which predicts the behavior of the Koopman operator in a specified time window for unknown operating conditions. We evaluate our framework on a prototypical fluid flow problem of interest: two-dimensional incompressible flow around a cylinder. 1 Introduction Figure 1: Schematic overview of the proposed physics-informed spatio-temporal modeling (PISTM) framework of dynamical systems. In the recent years there has been a lot of work involving machine learning (ML) frameworks for emulating spatio-temporally varying fields encountered in nonlinear dynamical systems ( 4 ) . These systems are frequently encountered in engineering design, and data from such systems are often generated by high fidelity (HF) simulations that are multi-physics and multi-scale in nature. Such simulations are often computationally very expensive to perform, and become a bottleneck in the overall engineering design process. There has been a lot of work in the domain of data-driven modeling to formulate surrogate models for spatio-temporal dynamical systems. Specifically, deep learning models such as convolutional neural networks (CNN) ( 5 ) , long short-term memory networks (LSTM) ( 13 ) , fully connected neural networks (FCNN) ( 10 ; 14 ) have been extensively used along with more traditional approaches such as proper orthogonal decomposition (POD) and dynamic mode decomposition (DMD) ( 19 ) for reduced order modeling and emulation of various dynamical systems. However, most of these approaches do not incorporate any physical constraints in their formulation, and hence have limitations with respect to their generalizability for input conditions that are outside the training domain. To this end, there has been an ever growing body of work in the field of physics-informed machine learning, whereby ML models are constrained with physical laws in their training phase. 16 developed a physics-informed neural network (PINN) architecture that is constrained by the Navier Stokes equation to reconstruct the flowfields which have guarantees of satisfying physical laws and boundary constraints. Recently, PINNs have been popularly used in various predictive learning problems involving partial differential equations (PDEs) ( 12 ; 2 ; 7 ) that govern the dynamics of the systems. However, most PINN formulations ( 7 ) require the knowledge of the underlying PDEs governing a physical process, either partially or fully. For any generic dynamical system, where the user is restricted to merely observing the spatio-temporal outputs from the system, the knowledge of the governing equations is often not there. PINN frameworks that rely on the physical laws are not completely suited for such scenarios, and this makes the application of such frameworks challenging for practical engineering applications. To this end, researchers have looked into physics-informed approaches which are informed by the physics of dynamical systems in general, and not the underlying physical laws. Some of the notable works in this regard are the ones by 3 which incorporate constraints inspired by Lyapunov stability in their framework, and 1 ; 18 which model the underlying dynamical systems using Koopman operators ( 8 ) . Koopman operator theory has been recently receiving considerable attention in the modeling and analysis of several nonlinear systems, with applications in various fields such as fluid mechanics. Koopman’s theory assumes that a nonlinear dynamical system can be fully encoded using an infinite dimensional linear operator such that the system evolves linearly in time in the transformed space. Based on this assumption on both forward and backward dynamics of dynamical systems, 1 ; 18 showed that consistent Koopman Autoencoders (KAE) are able to perform time series forecasting tasks, along with incorporating high-level constraints in the model. This is a significant advantage when compared to traditional time series prediction techniques in deep learning using Recurrent Neural Network (RNN) architectures such as LSTMs, whereby it is difficult to incorporate such constraints in the model learning phase. However, till date, the research based on these approaches have considered only the temporal learning problem, i.e. forecasting how the dynamical system evolves in time for a given operating and initial condition. The state-of-the-art cannot be directly applicable in practical problems of spatio-temporal surrogate modeling in which it is required to predict how a dynamical system evolves in time for unknown operating and boundary conditions. In this work, we aim to address this gap using a novel physics-informed spatio-temporal modeling framework (PISTM) which is based on: (a) learning a reduced order model (ROM) of the spatio-temporal evolution predicted by Koopman convolutional autoencoders on available training data (b) employing Gaussian process regression ( 17 ) models to predict the latent space coefficients of the ROM for the Koopman autoencoder for unknown operating/boundary conditions and (c) using the pre-trained decoder to predict the Koopman evolution for unknown operating/boundary conditions. A schematic overview of the approach is shown in Figure 1 . The Koopman autoencoder framework involves fully connected layers for encoding the dynamics of the underlying system, and assume that the dynamics in the latent space evolve linearly both forward and backward in time. This is achieved through the enforcement of the forward mapping \Phi and the backward mapping \Psi to be approximate inverse of each other ( 18 ) (c.f. Figure 1 ). The ROM of the Koopman autoencoder is a convolutional autoencoder framework, similar in architecture to the one proposed by the authors in a previous work ( 15 ) . Our framework has been validated on the problem of predicting the two-dimensional incompressible viscous flow around a circular cylinder for unknown Reynolds numbers. 2 Problem Statement The problem of interest in this work is to predict the spatio-temporal signature f(t) of a nonlinear dynamical system for a discrete time window of t=T to t=T+k for an unknown operating condition ( \textbf{X}_{test}\in\textbf{X}_{in} ), given that we know how the system behaves for a specified time history of t=T-h to t=T-1 for some operating conditions \textbf{X}_{train}\in\textbf{X}_{in} , with h>1 . Problems in this setup are very common in many practical engineering design problems, where the challenge is to predict how a complex system would behave in a specified time-frame for an unknown operating condition. It is to be noted that this problem is different and more challenging than classical forecasting problems, whereby given a temporal evolution of the system dynamics for a particular condition, it is

Read the complete text