Deep learning-assisted metasurface-enhanced near-infrared spectroscopy for alcohol detection

In this study, we introduce a deep learning-assisted metasurface (MTS) platform for near-infrared (NIR) spectroscopic analysis that enables the simultaneous qualitative identification and quantitative concentration determination of multiple alcohols. Surface lattice resonances (SLRs) supported by the MTS are harnessed to amplify and spectrally structure the intrinsically overlapping NIR absorption fingerprints of alcohols. Leveraging this resonance-enhanced spectral response, we constructed a representative spectral dataset by sampling at strategically selected concentration points. To analyze these data, we developed a hybrid alcohol detection (HA) model that integrates convolutional neural networks (CNNs) with a Transformer architecture. Within this framework, the CNN branch is employed to extract local, fine-grained spectral features, whereas the Transformer branch is designed to capture long-range, global spectral dependencies and contextual relationships. The optimized HA model exhibits high predictive accuracy, achieving a coefficient of determination (R²) of 0.9653, a root mean square error (RMSE) of 0.9245 vol%, and 100% identification accuracy. Comprehensive validation confirms the robustness and generalization capability of the model, maintaining a mean R² > 0.93 and an RMSE on the order of 1 vol%across datasets acquired on the same day, over one-month time intervals in aqueous environments, and under more complex conditions in a beer matrix. These results demonstrate that deep learning-assisted MTS-based NIR spectroscopy offers substantial potential for miniaturized integration and provides a versatile platform for advanced biosensing and Internet of Things (IoT)-enabled sensing applications.