Enhancing High-Degree-of-Freedom Meta-Atom Design Precision and Speed with a Tandem Generative Network

Abstract

Traditional metasurface design approaches largely rely on the prior knowledge of researchers and iterative trial-and-error methods using full-wave simulations, resulting in lengthy and inefficient processes. Deep-learning techniques, such as tandem neural networks (TNNs) and generative networks, show considerable promise in addressing the inverse-design problem. However, TNN faces challenges in creating high-freedom structures and neglects learning one-to-many mappings in inverse problems. The denoising diffusion probabilistic model (DDPM), while superior to other generative networks in generation precision and quality, is hindered by slow structure generation. This paper proposes a novel metasurface design method called the tandem generative network (TGN) to realize accurate and efficient high-degree-of-freedom meta-atom design. TGN constructs an original probabilistic generative model and generates free-form meta-atoms by sampling from the probability space. TGN-generated patterns are validated to produce matching transmittance with an average mean absolute error of 0.0356, achieving decreases of 38% and 86% compared to DDPM and TNN, respectively. Furthermore, the generation speed of TGN is 2990 times faster than that of DDPM. By employing the first probabilistic generative model for metasurface design, TGN paves new avenues in deep learning for inverse design, providing a swift and accurate means to design complex meta-atom structures with desired electromagnetic properties.