Interpretable Deep Learning of Molecule-Solvent Interactions Using a Solvatochromic Subgroup Contribution Framework

Solvent effects critically influence the photophysical and optical properties of molecules in solution, yet a quantitative and interpretable description of solvent effects remains a long-standing challenge. In this study, to quantitatively understand and chemically interpret solvatochromism, deep learning (DL) models based on the solvatochromic subgroup contribution (SSC) framework were introduced, which describes solvatochromic shifts through localized interactions between molecular subgroups and the surrounding solvent. In this framework, the electronic transition energies of molecules in solution are successfully modeled based on three chemically interpretable contributions: the reference property (RP), subgroup contribution (SC), and proximity effect factor (PEF), providing a chemistry-informed representation of solvent-dependent electronic transition energies. Trained on an experimental database of solvent-dependent absorption and emission energies, SSC-based deep learning models achieve both high prediction accuracy and subgroup-level interpretability of solvatochromic shifts. More importantly, the learned SC values show strong correlations with classical solvent polarity parameters (i.e., E_T(30) and Kamlet-Taft parameters) and reproduce Bilot-Kawski relationships, confirming that the SSC framework captures explicit and physically meaningful molecule-solvent interactions. The SSC framework is a scalable and generalizable tool for modeling molecule-solvent interactions in chemistry and materials science, providing improved predictive accuracy and chemical interpretability.