Physics-Guided Transformer Enables Efficient Inverse Design of Terahertz Metasurfaces for Glucose Sensing

Inverse design of terahertz (THz) metasurfaces via algorithmic models has become a mainstream trend and finds wide application in THz biosensing devices. However, traditional inverse design methods without optimization often rely solely on mathematical correlations, lacking physical interpretability and resulting in outcomes that deviate from practical requirements. This work proposes a Physics-Guided Transformer (PGT) model tailored for the inverse design of THz biosensing metasurfaces. By incorporating an additional physics-guided spectral attention (PGS-Attention) module, the proposed model effectively captures global electromagnetic coupling effects and long-range dependencies between THz spectral sequences and metasurface units, significantly enhancing physical interpretability. As a result, the computation time of the PGT model is 6.38 × 10⁶ times faster than that of full-wave simulations. Furthermore, label-free biosensing experiments are conducted on the inversely designed metasurface using glucose solutions with 0.1-0.5 g/mL. Experimental results demonstrate that the electromagnetically induced transparency (EIT) window generated at 0.96 THz achieves a sensitivity of 280 GHz/RIU. This study provides a comprehensive technical route for physically interpretable inverse design models of THz metasurfaces and offers a solution to the black-box dilemma in artificial intelligence (AI)-driven micro/nano-optics design.