Physics-Infused Reduced-Order Modeling

for high-speed thermal protection systems

Published: Jun 1, 2025 by Daning Huang

The tackling of the Impossible Trinity of Modeling has been one of the research themes of our group, that is central to our efforts in the multi-disciplinary optimization of dynamical systems.

Our recent work is a milestone of such effort. We developed a systematic reduced-order modeling methodology that integrates the coarse-graining theory from statistical mechanics and numerical methods from machine learning community. The end product is the physics-infused reduced-order modeling (PIROM), that is simultaneously accurate, computationally efficient, and extrapolates to unseen scenarios. In conventional modeling, achieving any of the two sacrifices the third.

We demonstrated the methodology on a high-speed thermal protection system, governed by a parameterized nonlinear parabolic PDE. While the PIROM is trained using only one set of material properties, it achieves zero-shot learning and extrapolates with ease to completely new material properties without any new samples.

The work is done jointly with collaborators from the Sandia National Lab, Drs. Patrick Blonigan and John Tencer.