Reconstruction of molecular vibrational spectra from light-molecular vibration coupling spectra using deep learning

Light-molecular vibration coupling on metasurfaces induces complex optical phenomena, such as Fano resonance and Rabi splitting, which complicate molecular identification via conventional spectroscopic analysis. In this study, we developed an analytical system that uses deep feature learning to directly extract molecule-specific absorption information from these complex spectra. For a molecular vibration model with a single peak, we evaluated a DenseNet-169 convolutional neural network (CNN) model using 1,496 spectra generated via finite-difference time-domain (FDTD) simulations; however, the results were suboptimal. In contrast, for a two-peak molecular vibration model, we trained multiple CNN models, including DenseNet-169, on a dataset of 80,267 spectra. Consequently, we successfully reconstructed absorption coefficients with an extremely high accuracy, achieving a mean coefficient of determination ([Formula: see text]) of 0.9209, even in complex systems with overlapping vibrational peaks. This approach demonstrates significant potential as a foundational technology for next-generation molecular sensing.