Deep learning-enhanced graphene metasurface terahertz absorber for real-time multi-gas identification and quantification

We present an intelligent, chip-scale electronic nose based on a tunable dual-layer graphene terahertz (THz) metasurface absorber integrated with a multi-task deep neural network. The sensor simultaneously identifies and quantifies four hazardous gases: carbon monoxide (CO), carbon dioxide (CO \(\:{\text{}}_{2}\) ), nitric oxide (NO), and nitrogen dioxide (NO \(\:{\text{}}_{2}\) ). The absorber consists of patterned graphene disks, rings, and ribbons separated by KAPTON and TOPAS dielectric spacers, backed by a gold ground plane. Using the Kubo formalism, we derive the surface conductivity of graphene and incorporate it into a transfer-matrix model that accurately predicts the absorption spectrum as a function of bias voltage and gas concentration. The molecular fingerprints of the target gases are modelled via Voigt line profiles based on HITRAN data, and the gas–graphene interaction is described by a Langmuir adsorption isotherm that leads to concentration-dependent frequency shifts and amplitude modulations. A multi-task convolutional neural network is trained on 20,000 synthetic mixture spectra (10–500 ppm) to simultaneously perform multi-label classification (F1 \(\:>\) 0.97) and multi-output regression (mean absolute error 8.9 ppm, R \(\:{\text{}}^{2}\) = 0.986). The architecture allows rapid bias switching (four chemical potentials: 0.35, 0.45, 0.65, 0.90 eV) to generate a multi-dimensional data cube. Transfer learning enables extension to new gases with minimal retraining. The proposed sensor achieves sub-second response, high selectivity, and excellent scalability. The limitations of synthetic training data and the need for experimental validation are explicitly discussed.