Reverse Design of Three-Band Terahertz Metamaterial Sensor

Terahertz metamaterial devices (TMDs) have demonstrated promising applications in biomass detection, wireless communications, and security inspection. Nevertheless, conventional design methodologies for such devices suffer from extensive iterative optimizations and significant dependence on empirical expertise, substantially prolonging the development cycle. This study proposes a reverse design framework leveraging a deep neural network (DNN) to enable rapid and efficient TMD synthesis, exemplified through a three-band terahertz metamaterial sensor. The developed DNN model achieves high-fidelity predictions (mean squared error = 0.03) and enables rapid inference for structural parameter generation. Experimental validation across four distinct target absorption spectra confirms high consistency between simulated and target responses, with near-identical triple-band resonance characteristics. Benchmarking against traditional CST-based optimization reveals a 36-fold acceleration in design throughput (200-device parameterization reduced from 36 h to 1 h). This work demonstrates a promising strategy for data-driven reverse design of multi-peak terahertz metamaterials, combining computational efficiency with rigorous electromagnetic performance.