Bridging the Theory–Experiment Gap in Vibrational Spectra for Molecular Structure Elucidation

Structure elucidation from experimental vibrational spectra is a fundamental task in chemistry, yet data-driven approaches remain limited by scarce paired experimental data and systematic deviations between computed and measured spectra. Here, we bridge this theory–experiment gap through two coupled advances. First, we introduce Spec2Spec, a U-Net translator that fuses theoretical vibrational spectra with three-dimensional molecular embeddings from a pretrained Uni-Mol encoder via cross-attention, learning a structure-conditioned, nonlinear mapping from computed to experimental infrared spectra. Applying Spec2Spec to the 238,869-molecule USTC-FG26 database yields FG26AUG, a large-scale pseudo-experimental dataset. Second, we pretrain the spectrum-based molecular generative model Spec2Mol on FG26AUG and fine-tune it on limited experimental data, markedly reducing the reliance on large paired experimental datasets. On two independent benchmarks (NIST gas-phase spectra and OCR-digitized literature spectra), this pipeline consistently outperforms counterparts pretrained on purely theoretical corpora. Gradient-weighted class activation mapping (Grad-CAM) analyses indicate that the fine-tuned model attends to chemically meaningful spectral regions. Together, these results establish a practical pathway for molecular structure generation directly from experimental vibrational spectra.