Physics-informed supervised residual learning for electromagnetic modeling

In this paper, we propose a non-stationary iterative physics-informed supervised residual learning scheme (NSIP-ISRL) as a general framework for modeling electromagnetic wave propagation in inhomogeneous medium. NSIPISRL is based on the residual neural network (ResNet) that maps the residuals of matrix equation to the update of solutions [1]. It incorporates the concept of non-stationary iterative method, in which physical principles are embedded in the solution process through matrix-vector multiplication. NSIPISRL is applied to solve 2D volume integral equations in order to model electromagnetic wave interaction with lossy scatterers. The results show that NSIPISRL has a good accuracy and a strong generalization ability. The trained network can be applied to various scenarios with different scatterers, different incident angles, and different frequencies and still maintain a good accuracy.