Artificial intelligence (AI), particularly deep learning (DL), is revolutionizing optical metamaterial design by overcoming the fundamental challenges of multidimensional parameter spaces, nonlinear structure–property relationships, and the intrinsic non-uniqueness of inverse problems. By learning complex mappings between geometric structures and electromagnetic responses, DL enables rapid forward prediction and on-demand inverse design without computationally intensive full-wave simulations. This review provides a comprehensive survey of AI-driven design methodologies across four key metamaterial platforms: localized resonant nanostructures, metasurfaces, periodic and guided-wave photonic structures, and complex scattering systems. For each platform, we systematically examine the neural network architectures employed, the specific design challenges addressed, and the representative achievements attained. These data-driven approaches not only significantly accelerate the discovery of high-performance structures but also offer new opportunities for extracting physical insights into light–matter interactions. We assess the critical challenges of data efficiency, model interpretability, and experimental feasibility, and outline emerging research directions that may address these barriers. This review aims to provide both a comprehensive summary of the current state of the art and forward-looking perspectives for this rapidly evolving interdisciplinary field.
Xu et al. (Thu,) studied this question.