Graph neural network based isomorphic learning of finite element computation for optical scattering problems
Optics Express · 16.11.2026 · 34 (23) · 44555
DOI: 10.1364/oe.610464 ↗Kurz gefasst
Fachartikel in der Zeitschrift Optics Express, online erschienen am 10.11.2026. Verfasst von Jingwei Wang und 4 weiteren Autor:innen. Die Arbeit nennt 51 Literaturangaben. Gefördert von: National Natural Science Foundation of China, China Postdoctoral Science Foundation, Natural Science Foundation of Hubei Province.
Abstract (Originalsprache)
Accurate and efficient optical simulations are essential for the design of nanophotonic devices. Although deep learning presents a promising path to accelerate these simulations, many data-driven models suffer from limited generalization and poor physical interpretability. To address this issue, we propose a FEM-based graph isomorphic learning (FemGIL) framework for solving optical scattering problems. FemGIL relies on the isomorphism between finite element data and graph structure. In this model, nodes represent unknown coefficients, and their features encode key physical information, such as the incident field and material parameters. The network’s aggregation operation has the same mathematical form as the weak form of FEM and is used to simulate local electric field interactions. The accuracy and efficiency of FemGIL are validated in three 2-D scenarios involving basic scatterers, single metallic scatterers, and multiple metallic scatterers. FemGIL demonstrates good predictive performance for geometric sizes outside the training range. Compared to the conventional BiCGSTAB solver in FEM, FemGIL achieves over a 70% reduction in solution-stage computation time while maintaining comparable accuracy. By employing transfer learning, FemGIL can efficiently adapt a pretrained model to predict the scattering field of complex structures. By deeply integrating numerical methods with deep learning, FemGIL provides a fast, generalizable, and physics-consistent simulator, paving the way for rapid optical device design.
Förderung
- National Natural Science Foundation of China
- China Postdoctoral Science Foundation
- Natural Science Foundation of Hubei Province
- Optics Valley Laboratory
Zitieren
APA
Wang, J., Zhang, Z., Zhang, Z., Liang, K., & Chen, Y. (2026). Graph neural network based isomorphic learning of finite element computation for optical scattering problems. Optics Express, 34(23), 44555. https://doi.org/10.1364/oe.610464
Vancouver
Wang J, Zhang Z, Zhang Z, Liang K, Chen Y. Graph neural network based isomorphic learning of finite element computation for optical scattering problems. Optics Express. 2026;34(23):44555. doi:10.1364/oe.610464
BibTeX
@article{Wang2026,
author = {Wang, Jingwei and Zhang, Zicen and Zhang, Zexuan and Liang, Kun and Chen, Yuntian},
title = {Graph neural network based isomorphic learning of finite element computation for optical scattering problems},
journal = {Optics Express},
year = {2026},
volume = {34},
number = {23},
pages = {44555},
doi = {10.1364/oe.610464}
}
EUVIMED