Building Trust in PINNs: Error Estimation through Finite Difference Methods

Aleksander Krasowski, René P. Klausen, Aycan Celik, Sebastian Lapuschkin, Wojciech Samek, Jonas Naujoks

Late-breaking Work · 4th World Conference on eXplainable AI (xAI 2026)

We propose a post-hoc method for estimating pointwise errors of Physics-Informed Neural Networks by exploiting the fact that, for linear PDEs, the error satisfies the same differential operator driven by the PINN residual. This yields spatially resolved error maps — interpretable diagnostics analogous to attribution maps in XAI.

Interactive demo coming soon. This page will host predictions, error-over-time curves, and time-sliced error fields for the five PDE configurations from the paper.

Code on GitHub · Paper (preprint)