Convolutional neural network-assisted design and validation of terahertz metamaterial sensor

This paper proposes a convolutional neural network (CNN)-assisted method for both forward optimization and inverse design of terahertz metamaterial sensors (TMSs), addressing the limitations imposed by reliance on manual trial-and-error processes. A hollow n-shaped TMS based on copper foil was developed, exhibiting two distinct resonance peaks between 0.3 and 1.4 THz. The formation mechanisms of resonance peaks were analyzed based on electric field and current distribution, while the sensing performance of the TMS was investigated. In the forward optimization stage, the n-shaped unit of TMS was converted into a data matrix, and the CNN was developed to predict the resonance frequency. In the inverse design stage, a predictive model for estimating the size of the TMS was developed by applying one-dimensional convolution to the transmission coefficients. The training dataset employed for forward optimization and inverse design achieved coefficients of determination (R2) of 0.99 and 0.99, respectively, with corresponding mean absolute error (MAE) values of 3.90 and 1.04. The efficacy of the proposed method was validated through terahertz time-domain spectroscopy (THz-TDS) measurements of TMS. Experimental assessments were conducted on glucose solutions of varying concentrations to ascertain the sensing capabilities. The proposed method contributes to the efficient design and optimization of TMS.