Inverse design of dielectric metasurface transmitarray using physics-enhanced generative adversarial networks

This paper proposes a physics-enhanced wasserstein generative adversarial network with gradient penalty for the inverse design of dielectric metasurface transmitarrays. Unlike existing approaches, the proposed network incorporates a differentiable simulator as a carrier of physical prior knowledge, achieving high spectral accuracy while preserving structural diversity. Furthermore, a lightweight physics regularization module is integrated to enforce sparsity and connectivity constraints, ensuring that the generated structures satisfy practical fabrication requirements. As a proof of concept, a dielectric metasurface transmitarray designed by the proposed network is fabricated using 3D printing technology and experimentally demonstrated for first-order orbital angular momentum beam generation. The measured results show good agreement with full-wave simulations, validating the feasibility and effectiveness of the proposed inverse design methodology.