A Fourier Neural Operator Enhanced Physics-Embedded Iterative Learning Solver for Electromagnetic Scattering Analysis

In this letter, an efficient and accurate physics-embedded Fletcher–Reeves conjugate gradient (FRCG) projection method, termed conjugate gradient-based Fourier neural operator (CGFNO), is proposed to solve the electromagnetic scattering problems iteratively. As a neuralnetwork, the Fourier neural operator (FNO) in the proposed method is deployed to facilitate the update of the unknown total electric field directly. In addition, the weight-sharing mechanism inherent in FNO substantially reduces the neural network complexity. The proposed CGFNO, which incorporates neural networks, demonstrates superior performance over the traditional FRCG method, offering improved accuracy and efficiency while significantly reducing both iteration counts and computational time in scattering field calculations. The accuracy and generalization ability of the CGFNO are verified with two representative numerical tests. Comparative analyses further reveal that CGFNO exhibits enhanced computational efficiency compared to the conventional numerical methods.