Physics-informed neural networks for the heat equation with dynamic (Wentzell) boundary conditions
Open Journal of Mathematical Sciences · 09.05.2026 · 10 · 526-536
DOI: 10.30538/oms2026.0303 ↗Kurz gefasst
Fachartikel in der Zeitschrift Open Journal of Mathematical Sciences, online erschienen am 09.05.2026. Verfasst von Muhammad Awais.
Abstract (Originalsprache)
Dynamic boundary conditions provide a stringent verification setting for physics-informed neural networks (PINNs), because the approximation must satisfy an interior parabolic equation together with an evolution law posed on the boundary. This paper develops a transparent benchmark study for the one-dimensional heat equation with Wentzell-type dynamic boundary conditions. Two manufactured solutions are employed. Benchmark A is a cosine solution for which homogeneous boundary forcing is admissible only under the compatibility condition \(k=\alpha\pi^2\); it is intentionally retained as a reduced Wentzell test because the endpoint derivatives vanish and the boundary-gradient term is inactive. Benchmark B is a fully active test with nonzero endpoint derivatives at both boundaries, so that the \(\beta u_x\) contribution is explicitly exercised. The numerical evidence is based on directly computed norms, residuals, five independent random-feature seeds, an implemented finite-difference reference solver, and a targeted ablation study. The reported neural solver is a least-squares physics-informed random-feature network with hard initial-condition enforcement, \(250\) hidden features, \(3980\) interior collocation points, and \(401\) boundary times per endpoint. Over five seeds, the mean relative errors are \(E_{L^2}=(8.65\pm1.27)\times10^{-6}\) and \(E_{L^{\infty}}=(1.12\pm0.20)\times10^{-5}\) for Benchmark A, and \(E_{L^2}=(9.45\pm2.42)\times10^{-8}\) and \(E_{L^{\infty}}=(2.77\pm0.99)\times10^{-7}\) for Benchmark B. The finite-difference baseline displays approximately second-order convergence under the parabolic scaling \(\Delta t=2h^2\). These results show that the two-case manufactured suite provides a useful verification framework for smooth one-dimensional Wentzell-type dynamics, while the scope of the claims remains restricted to this benchmark setting rather than general neural-solver superiority.
Zitieren
APA
Awais, M. (2026). Physics-informed neural networks for the heat equation with dynamic (Wentzell) boundary conditions. Open Journal of Mathematical Sciences, 10, 526-536. https://doi.org/10.30538/oms2026.0303
Vancouver
Awais M. Physics-informed neural networks for the heat equation with dynamic (Wentzell) boundary conditions. Open Journal of Mathematical Sciences. 2026;10:526-536. doi:10.30538/oms2026.0303
BibTeX
@article{Awais2026,
author = {Awais, Muhammad},
title = {Physics-informed neural networks for the heat equation with dynamic (Wentzell) boundary conditions},
journal = {Open Journal of Mathematical Sciences},
year = {2026},
volume = {10},
pages = {526--536},
doi = {10.30538/oms2026.0303}
}
EUVIMED