Build with Your Diode: Inverse Design of Pre-Defined Multi-Layer Reconfigurable Transmissive Metasurfaces

This paper proposes a novel deep learning framework based on Wasserstein Generative Adversarial Networks with Gradient Penalty (WGAN-GP) for the inverse design of sandwiched-structure transmissive metasurfaces. Addressing the challenges of multi-layer structure design and the non-uniqueness problem in inverse mapping, we introduce a hybrid approach combining multi-medium transmission theory with data-driven models. The framework consists of two key components: a forward proxy predictor utilizing a Dual-Head ResNet-Variational AutoEncoder (VAE) to accurately map geometries to electromagnetic responses, and a physics-aware inverse model employing Feature-wise Linear Modulation (FiLM) to handle highdimensional design spaces. The proposed method is validated through the design of magnitude-reconfigurable and 1-bit phasereconfigurable unit cells working at millimeter-wave frequencies. Results demonstrate that the inverse model achieves a low Mean Squared Error (MSE) of 0.0013, effectively generating structures that meet target transmission properties.