Deep Learning-Driven Design of Freeform Metasurfaces for Holographic Imaging

Deep learning (DL), with its strong nonlinear fitting capabilities, has found extensive application in the inverse design of metasurfaces (MSs). This paper presents a deep learning-driven approach to design MSs for holographic imaging. We first pre-analyzed the metallic structures within the meta-atoms in our training dataset to achieve a full <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$2\pi$</tex> phase range. A forward network is then designed to predict the electromagnetic (EM) responses given a meta-atom pattern. Furthermore, an inverse network capable of determining the necessary structure for a desired EM response is constructed. To mitigate the issue of non-uniqueness, we cascade these two networks, forming a tandem network to achieve holographic imaging. The simulation results demonstrate that under plane wave illumination at 10 GHz and 15 GHz, the MS array can project high-resolution holograms of the letters “T” and “F”, respectively.