A dual-branch transformer architecture for inverse design of reconfigurable metasurface

Abstract Metasurfaces enable advanced multifunctional applications by controlling electromagnetic (EM) wavefronts. While deep learning has aided EM design, existing methods struggle with global spectral features and discontinuous phase transitions and require large datasets. To address these limitations, we propose a Transformer-based framework that captures long-range spectral correlations and integrates physics-informed adaptive weighting to perform well under data-scarce conditions. Our framework includes two parts: a hybrid Transformer-Gated Recurrent Unit (GRU) forward prediction network and a Mirror Dual-Branch Transformer (MDT) inverse design network. The forward network maps structural parameters to EM responses via adaptive weighting that emphasizes abrupt phase transitions. The inverse network uses an innovative MDT architecture with three modules: polarization decoupling, phase-sensitive attention, and a spectral mirror branch, which decouple polarization features, refine spectral representation, and improve modeling accuracy. We validate the approach with a vanadium dioxide (VO 2 ) reconfigurable holographic metasurface. Results show that MDT reduces design loss by 45.5% compared to conventional Transformer methods, notably improving accuracy under data-scarce conditions. The method moves beyond static design, offering a scalable framework for phase-change material control, supporting temperature-responsive modulation and secure information encoding.