Extend physics‑informed learning to complex multi‑physics with legacy codes

Develop physics‑informed learning methods and surrogate models that are applicable to complex, coupled multi‑physics scenarios in contexts where mature legacy codes cannot easily be replaced, ensuring the resulting approaches remain effective and reliable in high‑dimensional stochastic settings.

Background

The paper surveys four strategies for tackling high‑dimensional Bayesian inverse problems. Within the third strategy—learning directly from governing equations and physics‑informed residues—the authors note that such approaches often struggle in high‑dimensional stochastic settings and are mainly applied to simpler physical models. They emphasize that bringing physics‑informed learning and surrogates to complex coupled multi‑physics problems, especially when legacy codes cannot be replaced, is a challenging frontier.

This open problem is central to broadening the applicability of physics‑informed machine learning beyond toy or single‑physics cases to realistic engineering and biomechanics scenarios where multiphysics couplings, nonlinearities, and mature software ecosystems complicate derivative access and model integration.