Deep-Learning-Based Metasurface Design Method Considering Near-Field Couplings

Abstract

Planar metasurfaces have been applied in several fields. Near-field coupling is typically neglected in traditional metasurface designs. A numerical modeling method for macrocells that considers near-field couplings between meta-atoms is proposed. A deep neural network (DNN) is constructed to accurately predict the electromagnetic response from different macrocells. Transfer learning is employed to reduce the number of the training datasets. The designed neural network is embedded in the optimization algorithm as an effective surrogate model. Both the deflector and high numerical aperture (NA) metalens are simulated and optimized with our design framework, approximately 30% improvements of efficiencies are achieved.