Dual-band terahertz metamaterial sensor integrated with deep learning for synergistic identification of red wine varieties

Terahertz time-domain spectroscopy (THz-TDS) technology is confronted with a significant challenge: the low sensitivity in the detection of food products containing water and complex matrices. Accordingly, a dual-band terahertz metamaterial (THz-MM), comprising a dielectric layer of polyimide and a resonance layer (shaped as a circle and a cross) of gold, was devised to address the aforementioned limitation. Subsequently, the structural parameters including the substrate thickness (ST), metal thickness (MT), ring width (RW), cross length (CL) and cross width (CW) were optimized. Furthermore, the sensor principle, stability and sensitivity were evaluated. Ultimately, the dual-band THz-MM exhibiting the optimal performance (ST: 18 μm, MT: 0.2 μm, RW: 6 μm, CL: 61 μm, CW: 1.8 μm) was prepared and characterised. To ascertain the viability of the dual-band THz-MM for the detection of trace substances, the weight factor method was used to successfully predict anthocyanin (R 2 : 0.9989) and tannin (R 2 : 0.9916) concentrations. At present, the majority of researchers are engaged in the design of intricate structures, rather than the analysis of data. In order to facilitate the extraction of valuable insights from the vast amount of information available, ten structural descriptors and one non-structural descriptor were devised. The results demonstrated that the fusion model with Add mode (A-Add-R), based on the AEFCNN and ResNet7 models, was effective in identifying different red wine varieties (precision: 0.9848, recall: 0.9833, F1-score: 0.9833, accuracy: 0.9833). In conclusion, a dual-band THz-MM sensor was constructed and combined with a deep learning (DL) model that fuses 1D and 2D descriptors for the purpose of identifying red wine varieties in a synergistic manner, which not only verify the feasibility of using THz-MM combined with DL for quality detection of liquid foods with complex matrices but also provide a new detection technology for the problem of counterfeit and shoddy red wine. • Dual THz-MM was designed, optimized and characterised. • Dual THz-MM can predict anthocyanin and tannic acid contents. • Ten structural descriptors and one unstructured descriptor were constructed. • Dual THz-MM combined with fusion models can distinguish different red wines.