Multi-Discriminator Distributed Generative Model for Multi-Layer RF Metasurface Discovery

Abstract

Metasurface-based antenna beam control is defining a new engineering paradigm in radio-frequency applications such as communications, radar, and analog spatial signal processing. Metasurfaces are composite electromagnetic material surfaces that are made of subwavelength scattering particles, or meta-atoms, with negligible thickness and optimized to control electromagnetic waves in unprecedented fashions through modified boundary conditions. Conventional metasurface design is a tedious process that requires iteratively solving Maxwell's equations. This becomes increasingly challenging as state-of-the-art metasurfaces require complex bi-anisotropic responses over multiple layers of meta- atoms and several frequency bands. In this paper, to reduce design time and optimization overhead, we employ a multi-discriminator distributed generative adversarial network for inverse design of multi- layer metasurfaces. Unlike conventional design approaches, our proposed approach is able to jointly design multiple layers, discover new meta- atom patterns, and avoid solving Maxwell's equations numerically or analytically. Results show that generated triple-layer meta-atoms can achieve frequency resonances within 7% of the input values.