RF Metasurface Array Design Using Deep Convolutional Generative Adversarial Networks

Abstract

In recent years, metasurface (MTS) arrays have shown promising abilities to control and manipulate electromagnetic (EM) waves through modified surface boundary conditions. The constituent unit elements of a MTS have become increasingly complex with rise of anisotropic radio-frequency (RF) applications such as beam scanning through anomalous reflection/refraction, beam focusing, and polarization conversion in an extremely low-profile. Designing these meta-atoms or metagratings is a challenging and time-consuming procedure. Each new MTS design typically requires numerous iterations of manual tuning and full-wave simulations. In this paper, we employ deep convolutional generative adversarial networks (DC-GANs) to generate anisotropic RF metamaterial unit cell designs for MTS arrays. Using a small set of simulated meta-atom spectra, these networks learn the relationship between the physical structure of meta-atoms and their reflection spectra for vertical and horizontal polarizations. Our numerical experiments demonstrate that DC-GANs are able to generate meta-atom structures that resemble design features in the training data. Numerical experiments with design test case showed 90% accurate reflection responses with errors within 0.2 (1.3) dB in magnitude and 3.0° (4.4°) in phase for the co-polar (cross- polar) component.