Deep Reinforcement Learning-Guided Design of Broadband Electromagnetic Wave Absorbing Coatings

Abstract

Artificial intelligence-assisted design of electromagnetic wave absorbing coatings is often restricted to geometry or topology optimization within fixed materials. Here, we present a modular particle swarm optimization−proximal policy optimization (PSO−PPO) framework for radar-absorbing metastructures that combines a progressive feature fusion surrogate for 8−18 GHz reflection-loss prediction, a ResNet-based empirical filter for low-performance patterns, and reinforcement learning optimization in a mixed discrete-continuous design space. For broadband single-layer optimization, the framework identifies a generated-material M2/Pt metasurface absorber with a 1.30 mm thickness and a CST-validated effective bandwidth of 6.32 GHz, with field simulations indicating absorption from multiple localized resonances and dielectric loss. The same strategy is extended to multilayer inverse design for prescribed single-peak Gaussian spectra, yielding target responses at 10, 12, and 16 GHz with mean absolute errors (MAEs) of 1.64−2.19 dB. This work demonstrates an efficient route for automated absorber optimization and customized spectral regulation.