Intelligent design of terahertz metamaterial sensors based on a transformer encoder model

Terahertz (THz) microfluidic sensors enable ultra-sensitive trace analyte detection through integrated metamaterial-enhanced sensitivity and microfluidic precision; however, the structural design of such sensors remains a complex and time-consuming challenge. To address this issue, a deep learning-based approach is proposed to achieve a rapid and intelligent design of THz microfluidic sensors. The proposed method employs a transformer encoder architecture to perform forward prediction of the absorption spectrum of THz metamaterial sensors based on structural parameters while enabling the inverse design of unit cell parameters corresponding to a given absorption spectrum. The simulation results show that the mean square errors (MSE) of the model for forward prediction and inverse design are 0.000 167 and 0.009 26, respectively, which are highly accurate for forward prediction and inverse design. Finally, the performance of the model is validated by performing a sensing analysis based on the inverse design results of the network model. The inverse-designed sensor showed two absorption peaks, both with absorption rates exceeding 99%, at 0.76 and 1.125 THz, achieving near-perfect absorption. The quality factors (Q) at these peaks are 15.6 and 22, respectively. The performance of this sensor is evaluated by metrics such as absorption characteristics and sensitivity, with refractive index sensitivities (S) of 260.6 GHz/RIU (Refractive Index Unit) and 400 GHz/RIU, respectively. Additionally, the sensor demonstrated insensitivity to both wide incident angles and polarization. This study offers new strategies for designing multi-band high-sensitivity THz metamaterial sensors and advances THz sensing applications, addressing the growing demand for THz detection.