Solving Combined Field Integral Equations with Physics-informed Residual Learning

This study applies physics-informed residual learning to compute electromagnetic scattering by perfect electric conductors (PECs). The formulation is based on the combined field integral equation. The graph neural network is used for its advantage of computing based on unstructured data. This allows machine learning schemes to handle triangular meshes. The physics-informed residual learning framework has the advantage of integrating information from both physics and data. Therefore, it provides a good generalization ability. After carefully preparing the dataset and finely training the network parameters, we can model electromagnetic scattering by various PEC targets with a relative error lower than 0.035. This study verifies the feasibility of further enhancing computation speed in electromagnetic modeling by combining data and physics.