Generative adversarial networks for high degree of freedom metasurface designs

Abstract

Due to the continuous development of microwave technology and the gradually expanding demand, attention has turned to free-form metasurfaces capable of realizing complex electromagnetic responses. Recent studies have shown that metasurface design can be accelerated and improved with the aid of deep learning methods. Here, we propose a generative adversarial network with raw network framework (RGAN) for realizing inverse design from a given response to a metasurface pattern. With the proposed approach, a metasurface design meeting requirements can be obtained immediately without the need for complex, repetitive iterative processes. Moreover, guided by the agent model within the network, the network is able to maximize exploration of the parameter space, ultimately generating novel designs completely distinct from those in the training set. Simulations demonstrate good spectral response matching. The feasibility of the proposed method is verified through experiments. The RGAN-based inverse prediction framework shows potential in the field of metasurface engineering and can be easily extended to other metasurface application areas, such as optical metamaterials and nanophotonic devices.