Repairing Noise-Contaminated\nLow-Frequency Vibrational\nSpectra with an Attention U‑Net

Low-frequency\nvibrational modes in infrared (IR) and\nRaman spectra,\noften termed molecular fingerprints, are sensitive probes of subtle\nstructural changes and chemical interactions. However, their inherent\nweakness and susceptibility to environmental interference make them\nchallenging to detect and analyze. To tackle this issue, we developed\na deep learning denoising protocol based on an attention-enhanced\nU-net architecture. This model leverages the inherent correlations\nbetween high- and low-frequency vibrational modes within a molecule,\neffectively reconstructing low-frequency spectral features from their\nhigh-frequency counterparts. We demonstrate the effectiveness of this\nmethod by recovering low-frequency signals of <i>trans</i>-1,2-bis(4-pyridyl)ethylene (BPE) adsorbed on an Ag surface, a representative\nsystem for surface enhancement Raman spectroscopy (SERS). Notably,\nthe trained model exhibits promising transferability to SERS spectra\nacquired under different surface and external field conditions. Furthermore,\nwe applied this method to experimental IR and Raman spectra of BPE,\nachieving high-quality, low-frequency spectral recovery.