When Known Physics Helps Neural PDE Models: Residual Constraints Out-Regularize Generic Priors for Nonlinear Dynamics
My rating
This paper studies whether the benefits of structural priors in neural PDE surrogates come from physics-specific information or from regularization and training choices. Using a common evaluation protocol, the authors compare several priors with a matched from-scratch neural operator baseline. Their main finding is that a known-equation residual consistently outperforms the best generic regularizer at equal tuning budgets, with the benefit appearing across nonlinear dynamics.
Tags: