EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Physics-informed radial basis function neural networks for two-dimensional magnetostatic field modeling

Huseyin Yildiz, Yasemin Poyraz Kocak, Erol Uzal

Physica Scripta · 06.11.2026 · 101 (44) · 446005

DOI: 10.1088/1402-4896/aeb8cf ↗

Kurz gefasst

Fachartikel in der Zeitschrift Physica Scripta, online erschienen am 03.11.2026. Verfasst von Huseyin Yildiz und 2 weiteren Autor:innen. Die Arbeit nennt 68 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Turkish Health Institutes Presidency.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Magnetic field modeling is essential for a wide range of applications, from electromagnetic device design to the analysis of geomagnetic and material-based systems. This study proposes two distinct neural network frameworks based on physics-informed radial basis function neural networks (PI-RBFNNs) to solve Maxwell’s equations on four different domains. The proposed unified and domain-specific network architectures use region-aware radial basis function kernels and are initialized by collocation governing equations in boundary and jump conditions. The models capture magnetic field behavior in both homogeneous and heterogeneous regions, with the level of agreement depending on the problem complexity and network architecture. All models are implemented using the TensorFlow framework, which enables efficient automatic differentiation and scalable optimization. Unlike conventional mesh-based numerical methods, the proposed PI-RBFNN framework provides a mesh-free formulation for magnetostatic field modeling. To evaluate the effectiveness of the models, results obtained from both configurations are compared with benchmark data from the literature and solutions generated by the finite difference method (FDM). The results show good agreement for simpler configurations, while discrepancies increase for more heterogeneous problems. For intermediate cases, the domain-specific architecture reduces the relative L 2 error with respect to FDM compared with the unified architecture. These findings demonstrate the potential of PI-RBFNN architectures for mesh-free modeling of magnetostatic field problems involving heterogeneous material regions and interface conditions, while further validation against independently converged reference solutions is required for broader assessment of their accuracy.

Förderung

  • Turkish Health Institutes Presidency

Zitieren

APA

Yildiz, H., Kocak, Y. P., & Uzal, E. (2026). Physics-informed radial basis function neural networks for two-dimensional magnetostatic field modeling. Physica Scripta, 101(44), 446005. https://doi.org/10.1088/1402-4896/aeb8cf

Vancouver

Yildiz H, Kocak YP, Uzal E. Physics-informed radial basis function neural networks for two-dimensional magnetostatic field modeling. Physica Scripta. 2026;101(44):446005. doi:10.1088/1402-4896/aeb8cf

BibTeX
@article{Yildiz2026,
  author = {Yildiz, Huseyin and Kocak, Yasemin Poyraz and Uzal, Erol},
  title = {Physics-informed radial basis function neural networks for two-dimensional magnetostatic field modeling},
  journal = {Physica Scripta},
  year = {2026},
  volume = {101},
  number = {44},
  pages = {446005},
  doi = {10.1088/1402-4896/aeb8cf}
}
Kostenlos starten →