Equivalent-circuit-intervened deep learning metasurface

Abstract

Recently, some advanced computer algorithms, including the inverse design method based on deep learning, have been widely applied to design versatile electromagnetic devices, featuring fast design and high efficiency especially in the global optimization problem. However, enormous amount of data generated from blind preliminary and computationally expensive simulations to predict the optical responses is required. Herein, to overcome this shortcoming of traditional deep-learning-based inverse design method, a fast, highly efficient, and sufficiently accurate synergetic strategy for devising photonic applications is presented, which integrates an explicit equivalent-circuit-intervened model as the guide in the implicit deep learning method to avoid blind calculations. As a case study of the proposed synergetic paradigm, the absorbers with customized and challenging spectral patterns are explored. For verification, three proof-of-concept designs were characterized: single-layer notched-band absorption, dual-layer trap absorption, and triple-layer ultra-wideband absorption. Our approach solves the knotty issues of batch designs, as well as the complex and time-consuming processes, suffered in traditional approaches. Moreover, our demonstrated equivalent-circuit-intervened intelligent strategy can be directly extended to various other photonic architectures, advancing a significant step toward real-world applications.

Keywords

• Metasurfaces
• Adaptable absorber
• Arbitrary spectral pattern
• Indium tin oxide
• Machine learning