Deep-Learning-Assisted Design of Polarization Conversion Metasurface With On-Demand Frequency Response and Ultra-Broadband Electromagnetic Scattering Reduction

Abstract

Designing compacted electromagnetic (EM) polarization conversion (PC) devices with high efficiency and various frequency response has become crucial due to their irreplaceable role in many applications such as satellite communications, imaging and radar detection. Here, we propose a method that combines prior-knowledge with deep-learning intelligent algorithm to enable fast customization of reflective metasurface polarization converter with on-demand frequency responses. The PC meta-atoms are designed through a combination of forward and inverse convolutional neural networks (FCNN and ICNN). Instead of time-consuming full-wave simulations, the FCNN can accurately predict the PC spectral response, enabling rapid generation of large datasets. While the ICNN, in conjunction with these datasets, facilitates efficient design of the PC meta-atoms. The proposed methodology is demonstrated through the generation of various PC meta-atoms with on-demand specified frequency bands, such as broadband, dual-band or tri-band responses. As an application, a reflective metasurface composed of the ultra-broadband PC atom and its mirror atom obtained with ICNN is designed and optimized with genetic algorithm which achieves a measured ultra-broadband radar cross-section reduction from 8–37 GHz. Our approach offers a quick and intelligent design solution for reflective PC devices, and may be potential in radar, antenna and communication fields.