Deep-learning surrogate modeling for rapid design of high-sensitivity terahertz metasurface sensors

The design of terahertz metasurface sensors commonly relies on extensive full-wave electromagnetic simulations, making systematic exploration of high-dimensional design spaces computationally demanding. In this work, a deep-learning-assisted surrogate modeling framework is proposed to accelerate the design and optimization of refractive-index-sensitive metasurfaces. A dataset comprising 2048 distinct CST-simulated metasurface configurations was generated over the frequency range of 1.0–1.4 THz. With 1001 frequency points evaluated for each configuration, the pointwise regression dataset contains 2,050,048 input–output samples of the complex transmission response. A multilayer neural network was trained to predict the real and imaginary components of the transmission coefficient as a function of geometric parameters, operating frequency, and analyte refractive index. The trained surrogate achieved strong predictive performance on an independent test dataset, yielding a mean squared error of \(4.12\times 10^{-3}\) , a mean absolute error of \(4.27\times 10^{-2}\) , and a coefficient of determination of \(R^{2}=0.9792\) . The model accurately reproduced the spectral features governing resonance-based sensing and enabled rapid evaluation of candidate metasurface designs without repeated full-wave simulations. The surrogate model was subsequently employed to screen thousands of candidate geometries and identify promising sensing configurations. Selected candidates were validated through full-wave CST simulations, confirming the effectiveness of the proposed screening strategy. The best-performing design exhibited a refractive-index sensitivity of approximately \(0.2~\mathrm {THz/RIU}\) ( \(232~\mathrm {GHz/RIU}\) ), which is roughly 2.5 times higher than the baseline value of \(80~\mathrm {GHz/RIU}\) reported for closely related metasurface geometries. The observed resonance shifts and the agreement between surrogate predictions and CST results demonstrate the capability of the framework to guide the discovery of high-performance sensor designs. The results highlight the potential of surrogate-assisted electromagnetic design as an efficient and scalable alternative to conventional parameter-sweep-based optimization. The proposed framework substantially reduces computational cost while enabling systematic exploration of complex metasurface design spaces, providing a practical route toward data-driven development of terahertz sensing devices.