Hybrid Neural Network-Driven Inverse Design of Metasurface for Ultrabroadband RCS Reduction

An inverse design approach based on hybrid artificial intelligence (AI) network is proposed for designing a metasurface with ultra-broadband radar cross section (RCS) reduction. The metasurface comprises 1-bit coded unit cells designed by a tandem neural network (TNN), which enables efficient inverse prediction of unit cell parameters with absorption and polarization conversion as input. A deep Q-network (DQN) optimizes the array arrangement to maximize RCS reduction bandwidth. This hybrid artificial network enables an efficient end-to-end inverse design from target RCS performance to unit cells and array arrangement, significantly improving the design efficiency and scalability. As an example, we design and fabricate an optically transparent metasurface with polyethylene terephthalate and indium tin oxide, achieving over 10 dB RCS reduction in an ultra-wide frequency band from 3.7 to 18.7 GHz (133.9% relative bandwidth) and maintains angular stability up to 45° under dual polarizations. The proposed metasurface demonstrates strong potential for electromagnetic stealth applications that require optical transparency. Moreover, the hybrid AI network provides a scalable approach for developing advanced multifunctional metasurfaces.