FE-PIRBN: Feature-enhanced physics-informed radial basis neural networks for solving high-frequency electromagnetic scattering problems

Huajian Zhang, Chao Li, Rui Xia, Xinhai Chen, Tiaojie Xiao, Xiaowei Guo, Jie Liu

Abstract: The solution of electromagnetic scattering problems holds significant application prospects in fields such as target detection and electromagnetic metamaterial design. However, efficient prediction of electromagnetic fields under high-frequency and intricate geometric conditions remains one of the most challenging problems. In this study, we proposed a feature-enhanced physics-informed radial basis neural network (FE-PIRBN) for solving high-frequency electromagnetic scattering problems. As a physics-driven neural network model, FE-PIRBN incorporates a PIRBN core network to preserve the local approximating property during the electromagnetic wave propagation process. It combines multi-resolution hash encoding techniques to optimize the capture of high-frequency electromagnetic field features. Under data-free training, the FE-PIRBN enables the accurate prediction of high-frequency electromagnetic scattering at sub-wavelength scales. Based on the testing results of typical cases including the electromagnetic scattering from single and double cylinder, it is evident that FE-PIRBN achieves computational accuracy within a 1.40%-5.82% relative error compared to numerical simulation results and the reference solutions. In increasing frequency scenarios, FE-PIRBN shows excellent generalization capabilities, achieving significant accuracy improvements and more robust convergence in high-frequency electromagnetic scattering tasks, ensuring future reliability and accuracy in wider band applications. Of particular importance is that as the electromagnetic frequency increases, the computational time for numerical methods grows exponentially. In contrast, FE-PIRBN, as a neural network model, maintains a relatively stable parameter size, resulting in significantly faster computation. This presents a promising and potential solution for addressing large-scale electromagnetic problems in the field.