Rapid on-demand design of a tri-band electromagnetically induced transparency terahertz metamaterial antibiotic sensor assisted by deep learning

In this paper, we present a tri-band electromagnetically induced transparency-like (EIT-like) terahertz metamaterial antibiotic sensor based on multi-mode coupling, incorporating a <i>π</i>-shaped resonator and a rectangular resonant ring to generate three distinct transparency windows centered at 0.76, 0.79, and 0.85 THz. These frequencies correspond to the characteristic fingerprint spectra of chlortetracycline, tetracycline, and sodium penicillin, respectively, enabling the specific detection of each antibiotic. The sensor demonstrates excellent performance with sensitivities of 163, 93, and 155 GHz/RIU, quality factors (Q) of 14.8, 18.5, and 26.5, and figures of merit (FOM) of 3.2, 2.2, and 4.8 for the three resonant peaks. Furthermore, we developed an efficient inverse design method using a one-dimensional parallel convolutional neural network (CNN) architecture, where each branch employs differently sized convolutional kernels to capture multi-scale local features. The optimized network achieves rapid convergence within 300 iterations and attains a mean squared error (MSE) of 0.026 on the validation set, demonstrating both fast and accurate design capabilities for metamaterial sensors. This work significantly advances the development of multi-frequency EIT sensors and their practical applications.