Self-attention enhanced CVAE-transformer for end-to-end inverse design of terahertz metasurfaces with polarization-multiplexed holography

To address the low efficiency and heavy reliance on iterative optimization in conventional metasurface design, in this work, we propose an end-to-end inverse design framework based on a self-attention enhanced CVAE-transformer. By introducing a self-attention mechanism, we construct a neural network model capable of accurately predicting the electromagnetic response of metal–insulator–metal metasurface elements, simplifying the optimization process and improving design efficiency. On this basis, a conditional variational autoencoder-based inverse retrieval model is established to rapidly generate metasurface unit structures covering a full 0−2 π phase range at specific operating frequencies. As a proof of concept, a beam deflection metasurface is designed and simulated, and the same framework is further employed to realize a terahertz metasurface with polarization-multiplexed holographic imaging functionality. Simulation results demonstrate that the proposed method enables the end-to-end generation of structural parameters without the need for multiple iterations and successfully reconstructs distinct holographic images under y- and x-polarization states. Compared with conventional parameter-sweep-based approaches, the proposed framework significantly improves design efficiency and shortens the development cycle under given target phase distributions, providing a new approach, to our knowledge, for efficient and automated metasurface design.