Neural Residual Framework for Long-Term PDE Extrapolation

A physics-constrained method improves long-time stability in PDE simulations by approximating a one-step evolution operator using error-bound information. The framework uses a low-cost numerical prior and weak-form PDE residuals to enable training without ground-truth trajectory supervision.