Deep Learning Design for Loss Optimization in Metamaterials.

Inherent material loss is a pivotal challenge that impedes the development of metamaterial properties, particularly in the context of 3D metamaterials operating at visible wavelengths. Traditional approaches, such as the design of periodic model structures and the selection of noble metals, have encountered a plateau. Coupled with the complexities of constructing 3D structures and achieving precise alignment, these factors have made the creation of low-loss metamaterials in the visible spectrum a formidable task. In this work, we harness the concept of deep learning, combined with the principle of weak interactions in metamaterials, to re-examine and optimize previously validated disordered discrete metamaterials. The paper presents an innovative strategy for loss optimization in metamaterials with disordered structural unit distributions, proving their robustness and ability to perform intended functions within a critical distribution ratio. This refined design strategy offers a theoretical framework for the development of single-frequency and broadband metamaterials within disordered discrete systems. It paves the way for the loss optimization of optical metamaterials and the facile fabrication of high-performance photonic devices.