EC: Provable Extrapolation for Neural PDE Surrogates
Proceedings of the AAAI Symposium Series · 06.11.2026 · 10 (1) · 655-656
DOI: 10.1609/aaaiss.v10i1.43469 ↗Kurz gefasst
Fachartikel in der Zeitschrift Proceedings of the AAAI Symposium Series, online erschienen am 06.11.2026. Verfasst von Ignacy Nieweglowski.
Abstract (Originalsprache)
The traditional limitations of neural network architectures in reliably generalizing beyond the convex hulls of their training data have posed a significant problem for computational physics, in which one frequently wishes to predict the dynamics of physical systems beyond the surrogate model’s training domain. The recently introduced BEACONS framework offers a principled circumvention of these limitations for hyperbolic PDEs; using the method of characteristics we can determine the analytical properties of a solution a priori for regions arbitrarily far from the training domain and thus prove extrapolatory bounds for worst case L-infinity errors of these surrogate approximations. We can then construct deep architectures whose composite error remains bounded. The empirical advantage of BEACONS under strict protocol, however, remains largely unexplored. This work evaluates and validates BEACONS against Physics-Informed Neural Networks (PINNs), Fourier Neural Operators (FNOs), and standard Multi-Layer Perceptrons (MLPs) under a matched and preregistered protocol. The talk will present the theoretical promise of bound-carrying architectures for hyperbolic problems and the empirical advantage BEACONS yields over conventional surrogates.
Zitieren
APA
Nieweglowski, I. (2026). EC: Provable Extrapolation for Neural PDE Surrogates. Proceedings of the AAAI Symposium Series, 10(1), 655-656. https://doi.org/10.1609/aaaiss.v10i1.43469
Vancouver
Nieweglowski I. EC: Provable Extrapolation for Neural PDE Surrogates. Proceedings of the AAAI Symposium Series. 2026;10(1):655-656. doi:10.1609/aaaiss.v10i1.43469
BibTeX
@article{Nieweglowski2026,
author = {Nieweglowski, Ignacy},
title = {EC: Provable Extrapolation for Neural PDE Surrogates},
journal = {Proceedings of the AAAI Symposium Series},
year = {2026},
volume = {10},
number = {1},
pages = {655--656},
doi = {10.1609/aaaiss.v10i1.43469}
}
EUVIMED