Bayesian Active Learning for Accelerated Design of Broadband Polarization-Insensitive Metasurfaces

Electromagnetic metamaterials are suitable for a wide variety of civil and military applications because they are thin and lightweight and have strong absorption, tunable frequencies, and adjustable electromagnetic parameters. However, due to the large number of degrees of freedom in design and complex objectives, it is difficult to develop metamaterials quickly and efficiently by empirical trial and error. Here, we propose an active learning strategy combining Bayesian optimization and electromagnetic simulation to accelerate the design of lightweight broadband metasurfaces. With only 36 initial data points, the proposed method takes only 5 iterations to determine the optimal structure from over 11 million candidates. Broadband absorption is achieved with an absorption rate greater than 90% from 6.37 to 18 GHz at normal incidence (while the relative bandwidth reaches 95%), a rate 16% higher than that of the best built-in optimization. Both electromagnetic simulation and experimental validation confirm broadband absorption, polarization insensitivity, and wide-incident-angle absorption. Importantly, by using a resistive indium tin oxide film and a dielectric polymethacrylimide foam sheet, the entire structure is ultralight, with a weight of only 60 g per square meter. This study opens up a new avenue for the rational design of high-performance microwave metasurfaces and can be easily extended to promote the discovery of terahertz, infrared, and visible-light metamaterials.