j-IR-vis: Vision model for Infrared spectroscopy embeddings

Infrared (IR) spectroscopy provides rich structural information but interpreting spectra at scale remains challenging. Here we introduce j-IR-vis, a vision-based neural model that learns chemically interpretable representations directly from IR spectra for functional-group prediction and downstream molecular characterization. Trained separately on simulated ($\D_{sim}$) and experimental ($\D_{exp}$) datasets, j-IR-vis achieves strong functional group classification accuracy, with Exact-Match Ratios of 0.77–0.81 on $\Dsim$ and 0.67–0.76 on $\D_{exp}$, and Macro-F$_1$ scores up to 0.93. Grad-CAM attention maps reveal that the model consistently focuses on the fingerprint region (1500–400 cm$^{-1}$), confirming that it leverages physically meaningful vibrational features. Embedding-space analyses show that inclusion of this region improves all geometric metrics—RankMe, Silhouette, Separation, and Lift—indicating a richer and more structured latent manifold. Fixed embeddings further enable accurate prediction of aromatic ring counts (weighted F$_1$ = 0.69) and octanol–water partition coefficients ($\mathrm{RMSE}=1.18$, Spearman’s $\rho=0.59$), demonstrating transferability to structural and physicochemical tasks without retraining the vision backbone. Finally, cosine-similarity analyses reveal that j-IR-vis captures complementary chemical relationships compared to traditional fingerprint-based representations. Finally, cosine-similarity analyses reveal that j-IR-vis embeds molecules in a spectroscopically informed similarity space that captures relationships not present in standard fingerprint representations. Together, these results establish j-IR-vis as a general-purpose spectral encoder that bridges experimental spectroscopy and molecular machine learning, offering a physically grounded route to multi-modal chemical representations. j-IR-vis is openly accessible through the GitHub repository \url{https://github.com/ChemAI-Lab/jirvis}.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.