EUVIMEDEUVIMED entdecken
Crossref

A simple remedy for failure modes in physics informed neural networks

Ghazal Farhani, Nima Hosseini Dashtbayaz, Alexander Kazachek, Boyu Wang

Neural Networks · 2025-03 · 183 · 106963

DOI: 10.1016/j.neunet.2024.106963 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Neural Networks, 2025-03. Verfasst von Ghazal Farhani und 3 weiteren Autor:innen. Beteiligte Einrichtung: National Research Council Canada, Automotive and Surface Transportation, 800 Collip Cir, London, N6G 4X8, Canada. Hauptthemen laut MEDLINE: Neural Networks, Computer, Physics. Die Arbeit wurde bisher 22-mal zitiert (Crossref) und nennt 30 Literaturangaben. Gefördert von: Natural Sciences and Engineering Research Council of Canada.

AbstractSchlagwörterBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Physics-informed neural networks (PINNs) have shown promising results in solving a wide range of problems involving partial differential equations (PDEs). Nevertheless, there are several instances of the failure of PINNs when PDEs become more complex. Particularly, when PDE coefficients grow larger or PDEs become increasingly nonlinear, PINNs struggle to converge to the true solution. A noticeable discrepancy emerges in the convergence speed between the PDE loss and the initial/boundary conditions loss, leading to the inability of PINNs to effectively learn the true solutions to these PDEs. In the present work, leveraging the neural tangent kernels (NTKs), we investigate the training dynamics of PINNs. Our theoretical analysis reveals that when PINNs are trained using gradient descent with momentum (GDM), the gap in convergence rates between the two loss terms is significantly reduced, thereby enabling the learning of the exact solution. We also examine why training a model via the Adam optimizer can accelerate the convergence and reduce the effect of the mentioned discrepancy. Our numerical experiments validate that sufficiently wide networks trained with GDM and Adam yield desirable solutions for more complex PDEs.

Schlagwörter

MeSH-Schlagwörter (MEDLINE)

Neural Networks, ComputerPhysicsAlgorithmsNonlinear DynamicsComputer Simulation

Stichwörter der Autor:innen

Failure modes in PINNsOptimizationPhysics-informed neural networks

Autor:innen und Institutionen

  • National Research Council Canada, Automotive and Surface Transportation, 800 Collip Cir, London, N6G 4X8, Canada.
  • Western University, Department of Computer Science, 1151 Richmond St, Middlesex College, London, N6A 5B7, Canada.
  • Western University, Department of Mathematics, 1151 Richmond St, Middlesex College, London, N6A 5B7, Canada.
  • Western University, Department of Computer Science, 1151 Richmond St, Middlesex College, London, N6A 5B7, Canada; Vector Institute, Toronto, 661 University Ave Suite 710, M5G 1M1, Ontario, Canada. Electronic address: bw…

Förderung

  • Natural Sciences and Engineering Research Council of Canada

Zitieren

APA

Farhani, G., Dashtbayaz, N. H., Kazachek, A., & Wang, B. (2025). A simple remedy for failure modes in physics informed neural networks. Neural Networks, 183, 106963. https://doi.org/10.1016/j.neunet.2024.106963

Vancouver

Farhani G, Dashtbayaz NH, Kazachek A, Wang B. A simple remedy for failure modes in physics informed neural networks. Neural Networks. 2025;183:106963. doi:10.1016/j.neunet.2024.106963

BibTeX
@article{Farhani2025,
  author = {Farhani, Ghazal and Dashtbayaz, Nima Hosseini and Kazachek, Alexander and Wang, Boyu},
  title = {A simple remedy for failure modes in physics informed neural networks},
  journal = {Neural Networks},
  year = {2025},
  volume = {183},
  pages = {106963},
  doi = {10.1016/j.neunet.2024.106963}
}
Kostenlos starten →