Are Deep-Learning-Based Hybrid Partial Differential Equation Solvers Reliable? Why Training Paradigms and Update Strategies Matter
Computing in Science & Engineering · 2026-07 · 28 (3) · 56-67
DOI: 10.1109/mcse.2026.3696587 ↗Kurz gefasst
Fachartikel in der Zeitschrift Computing in Science & Engineering, 2026-07. Verfasst von Yuhan Wu und 3 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 16 Literaturangaben.
Abstract (Originalsprache)
Deep-learning-based hybrid iterative methods (DL-HIMs) integrate stationary iterative solvers with neural operators, utilizing their complementary spectral biases to accelerate convergence. However, many DL-HIMs stagnate at false fixed points where updates vanish while significant physical residuals persist. We demonstrate that reliability depends on training paradigms and update strategies rather than neural architectures alone. Through a detailed study of a nonlinear deep operator network-based and a linear fast Fourier transform-based DL-HIM, we show that convergence degrades when training objectives are misaligned with solver dynamics and problem physics. We examine Anderson acceleration (AA) and find its classical form ill-suited for nonlinear neural operators. We introduce physics-aware AA (PA-AA), which minimizes the physical residual rather than the fixed-point update. Numerical experiments confirm that PA-AA restores reliable convergence in substantially fewer iterations. These findings suggest that addressing the controversies that surround artificial intelligence-based partial differential equation solvers requires attention not only to architectures but also physically informed training and iteration design.
Zitieren
APA
Wu, Y., Beek, J. W. V., Dolean, V., & Heinlein, A. (2026). Are Deep-Learning-Based Hybrid Partial Differential Equation Solvers Reliable? Why Training Paradigms and Update Strategies Matter. Computing in Science & Engineering, 28(3), 56-67. https://doi.org/10.1109/mcse.2026.3696587
Vancouver
Wu Y, Beek JWV, Dolean V, Heinlein A. Are Deep-Learning-Based Hybrid Partial Differential Equation Solvers Reliable? Why Training Paradigms and Update Strategies Matter. Computing in Science & Engineering. 2026;28(3):56-67. doi:10.1109/mcse.2026.3696587
BibTeX
@article{Wu2026,
author = {Wu, Yuhan and Beek, Jan Willem van and Dolean, Victorita and Heinlein, Alexander},
title = {Are Deep-Learning-Based Hybrid Partial Differential Equation Solvers Reliable? Why Training Paradigms and Update Strategies Matter},
journal = {Computing in Science & Engineering},
year = {2026},
volume = {28},
number = {3},
pages = {56--67},
doi = {10.1109/mcse.2026.3696587}
}
EUVIMED