EUVIMEDOpdag EUVIMED
Open access · CC BY 4.0

Curriculum-Enhanced Adaptive Sampling for Physics-Informed Neural Networks: A Robust Framework for Stiff PDEs

Hasan Cetinkaya, Fahrettin Ay, Mehmet Tunçel, Hazem Nounou, Mohamed Numan Nounou, Hasan Kurban, Erchin Serpedin

Mathematics · 2025-12-15 · 13 (24) · 3996

DOI: 10.3390/math13243996 ↗

At a glance

Journal article in Mathematics, published online on 2025-12-15. Written by Hasan Cetinkaya and 6 co-authors. Cited 5 times so far (Crossref); 36 references. The full text is freely available (open access, CC BY 4.0). Funded by: Texas A&M University at Qatar.

AbstractBibliografiske oplysningerCiteRelated studies

Abstract (original language)

Physics-Informed Neural Networks (PINNs) often struggle with stiff partial differential equations (PDEs) exhibiting sharp gradients and extreme nonlinearities. We propose a Curriculum-Enhanced (CE) Adaptive Sampling framework that integrates curriculum learning with adaptive refinement to improve PINN training. Our framework introduces four methods: CE-RARG (greedy sampling), CE-RARD (probabilistic sampling), and their novel difficulty-aware dynamic counterparts, CED-RARG and CED-RARD, which adjust refinement effort based on task difficulty. We test these methods on five challenging stiff PDEs: the Allen–Cahn, Burgers’ (I and II), Korteweg–de Vries (KdV), and Reaction equations. Our methods consistently outperform both Vanilla PINNs and curriculum-only baselines. In the most difficult regimes, CED-RARD achieves errors up to 100 times lower for the Burgers’ and KdV equations. For the Allen–Cahn and Reaction equations, CED-RARG proves most effective, reducing errors by over 40% compared to its non-dynamic counterpart and by over two orders of magnitude relative to Vanilla PINN. Visualizations confirm that our methods effectively allocate collocation points to high-gradient regions. By demonstrating success across a wide range of stiffness parameters, we provide a robust and reproducible framework for solving stiff PDEs, with all code and datasets publicly available.

Funding

  • Texas A&M University at Qatar

Cite

APA

Cetinkaya, H., Ay, F., Tunçel, M., Nounou, H., Nounou, M. N., Kurban, H., & Serpedin, E. (2025). Curriculum-Enhanced Adaptive Sampling for Physics-Informed Neural Networks: A Robust Framework for Stiff PDEs. Mathematics, 13(24), 3996. https://doi.org/10.3390/math13243996

Vancouver

Cetinkaya H, Ay F, Tunçel M, Nounou H, Nounou MN, Kurban H, et al. Curriculum-Enhanced Adaptive Sampling for Physics-Informed Neural Networks: A Robust Framework for Stiff PDEs. Mathematics. 2025;13(24):3996. doi:10.3390/math13243996

BibTeX
@article{Cetinkaya2025,
  author = {Cetinkaya, Hasan and Ay, Fahrettin and Tunçel, Mehmet and Nounou, Hazem and Nounou, Mohamed Numan and Kurban, Hasan and Serpedin, Erchin},
  title = {Curriculum-Enhanced Adaptive Sampling for Physics-Informed Neural Networks: A Robust Framework for Stiff PDEs},
  journal = {Mathematics},
  year = {2025},
  volume = {13},
  number = {24},
  pages = {3996},
  doi = {10.3390/math13243996}
}
Start for free →