An advanced deep learning-driven Terahertz metamaterial sensor for distinguishing different red wines

Abstract

Terahertz time-domain spectroscopy (THz-TDS) encounters two issues in the detection field, which refer to the detection target for solid samples caused by the strong absorption of water and the large content target substance led by the low sensitivity. Fortunately, terahertz metamaterial (THz-MM) that can carry liquid samples and amplify signals solves the above problems well. In addition, the THz-MM can achieve the detection of trace substances through the resonance peak shift generated by designing suitable structures. However, since most researchers focus on designing complex structures rather than analyzing data, deep learning (DL) that can mine new features from original features and construct decision models is used to research the rich information in THz-MM sensor data. In the current research, a flexible transmissive THz-MM in the shape of a circle (‘O’ shape) was designed by depositing the gold on the polyimide substrate. Firstly, the structures referring to substrate thickness (ST), metal thickness (MT) and ring width (RW) were optimized, and the performances referring to principle, stability and sensitivity were evaluated. Next, the best THz-MM (ST: 16 µm, MT: 0.2 µm, RW: 6 µm) was prepared and characterized from morphology, thickness and consistency. Then, different concentrations of anthocyanins (R²: 0.9982) and tannic acid (R²: 0.9736) were successfully predicted by combining the resonance peak shifts. Finally, resonance peak descriptors were constructed and combined with DL referring to a fully connected neural network (FCNN) model to successfully identify different varieties of red wines (Precision: 91.11 %; Recall: 90.74 %, F1-score: 90.83 %; Accuracy: 90.74 %). Overall, the current research presents an advanced DL-driven THz-MM sensor, which promotes the process of THz-TDS technology in the food detection field