Enhanced Low-resolution Contrast Operator Using Neural Networks for E-polarized EM Scattering Problems

Coarse discretization introduces significant errors in the solution of scattering problems, in part due to discretization errors in the contrast operator. We present a procedure for the automatic construction of a modified contrast operator for electromagnetic scattering problems by using trainable neural networks to represent a modified contrast operator. We achieve a higher accuracy on a coarse discretization while still keeping computation time down compared to a fine discretization. By using synthetic data from a full-wave Maxwell solver to train the network for one-dimensional slab scatterers and two-dimensional polygonal scatterers, we are able to use the techniques found in deep learning to improve accuracy in coarse-grid forward scattering problems.