Hybrid ROM-PINN Framework for Closure Modeling in Convection-Dominated Systems

TL;DR

This paper introduces a novel hybrid ROM-PINN framework that combines physics-based modeling and machine learning to improve closure modeling in convection-dominated fluid flow simulations, enhancing accuracy and robustness.

Contribution

It develops a new ROM closure method integrating VMS-based physics principles with PINNs, advancing data-driven closure modeling in reduced-order models.

Findings

Improved accuracy of ROMs in convection-dominated regimes.

Enhanced robustness through physics-informed machine learning.

Bridging classical multiscale closure with modern data-driven methods.

Abstract

Reduced-order models (ROMs) have become an essential tool for reducing the computational cost of fluid flow simulations. While standard ROMs can efficiently approximate laminar flows, their accuracy often suffers in convection-dominated regimes due to the truncation of dynamically important modes. To account for the influence of unresolved scales, ROM closure models are commonly introduced. Classical closure strategies are typically based on phenomenological arguments or analogies with large eddy simulation (LES), often formulated within a variational multiscale (VMS) framework, in which the resolved and unresolved scales are explicitly separated and their interactions are systematically modeled. More recently, advances in data-driven modeling and machine learning have opened new opportunities to construct ROM closures that are both more accurate and more consistent with the underlying physics. In this work, we develop a new ROM closure that combines machine learning with physics-based modeling principles. The closure term is derived within a VMS framework, where the reduced solution space is decomposed into resolved and unresolved components. This VMS-derived closure term is then modeled using PhysicsInformed Neural Networks (PINNs) and incorporated into a newly constructed C-PINN-ROM. The resulting closure leverages high-fidelity data while enforcing physical constraints imposed by the reduced-order equations, thereby ensuring consistency with the underlying dynamics and enhanced robustness in convection-dominated regimes. Through this PINN-based framework, we demonstrate how physics-informed machine learning can substantially improve the accuracy and robustness of ROMs, effectively bridging classical multiscale closure modeling with state-of-the-art data-driven methodologies.