Neural partial differential equation (PDE) surrogates trained with mean squared error (MSE) often reproduce the dominant dynamics of a system while losing high-frequency information and fine-scale structures. This limitation can make them unreliable for problems where small-scale behavior is important.

The new preprint introduces ReAl, a refinement method that aligns the surrogate's dynamical flow with the true system's evolution. Rather than only matching outputs at individual time steps, ReAl aims to keep the learned trajectory consistent with the actual flow, preserving details that MSE training tends to discard.

The work connects to generative modeling, specifically Flow Matching, which provides a way to model complex distributions over trajectories. The abstract suggests that this alignment can improve the fidelity of neural surrogates without sacrificing their computational efficiency.

Because the source is a brief arXiv abstract, details on benchmarks, implementation, and comparisons are not yet available. The preprint is listed as new under identifier 2610.20369.