Deep convolutional generative adversarial networks assisted inverse design of quad-channel full-space metasurface

This is the authors' abstract. We don't add key points for this paper.

In recent years, deep learning has emerged as a powerful data-driven approach to transform the way we discover, design, and utilize metasurfaces (MSs). In this paper, we propose an inverse method for designing a multifunctional metasurface based on deep convolutional generative adversarial networks (DCGANs). The meta-atom patterns are designed in topological configurations and encoded as binary matrices. By comprehensively analyzing the varying metal proportions within the meta-atoms, the design space is exhaustively explored, resulting in a high-quality training dataset with a full phase span. In addition, Gramian angular difference fields (GADFs) are introduced as a data representation method that encodes one-dimensional (1D) electromagnetic (EM) response data into two-dimensional (2D) images. This enables the DCGAN model to flexibly design MSs using the computer vision methodology. As a novel inverse design paradigm for functional MSs, we develop an asymmetric reflection-transmission-integrated quad-channel full-space metadevice. Upon illumination with different frequencies and linearly polarized excitations, the proposed design functions as both an orbital angular momentum (OAM) beam generator and a holographic imager, with both simulation and measurement confirming the design's accuracy.