Terahertz (THz) technology is finding increasingly widespread applications in biosensing, high-speed communication, and stealth materials. Meanwhile, graphene, as a quintessential two-dimensional material, has emerged as a core component of THz devices due to its unique optoelectronic properties. However, the precise and non-destructive characterization of the complex conductivity of graphene at the microscopic scale remains a formidable challenge. Conventional measurement methods often suffer from limitations associated with contact resistance or intricate sample preparation processes. In this paper, we propose a non-invasive parameter inversion method based on deep learning. We design a tri-layer graphene-silica-copper metasurface structure featuring a central cavity and establish a high-fidelity scattering model that incorporates physical effects such as edge diffraction and multi-mode resonance. Utilizing the Radar Cross Section (RCS) data generated by this model, we train a Deep Enhanced Conductivity Predictor (DECP) network integrated with a Convolutional Block Attention Module (CBAM). Experimental results demonstrate that the proposed network can accurately reconstruct the complex conductivity of graphene from far-field RCS data. The coefficients of determination (R2) for the prediction of both real and imaginary parts exceed 0.99, with a Root Mean Square Error (RMSE) as low as the order of 10−5. This study not only validates the effectiveness of data-driven approaches in material characterization but also provides a novel paradigm for the real-time monitoring and intelligent design of terahertz metasurfaces.
Huang et al. (Fri,) studied this question.