Non-bonding intermolecular interactions largely dominate
the selective
dissolution of trace species into physical solvents and, therefore,
are fundamentally important to solvent development for the capture
of environment-undesired compounds or purification of chemicals. However,
acquirement of the interaction energy requires costly quantum chemical
computation and still encounters a practical challenge to build a
chemically interpretable machine learning (ML) prediction model using
documented molecular descriptors. Herein, we report an ML model for
predicting the interaction energies (E
int) between dimethyl sulfide and potential absorbing solvents. Applying
the reduced density gradient and quantum theory of atoms in molecules
analyses, the non-bonding intermolecular interactions of dimethyl
sulfide with solvent compounds were elucidated through focusing on
the molecular fragments containing the main center (MC) and secondary
center (SC) rather than the whole molecule. The training data set
was obtained using a molecular generation strategy, and 21 molecular
descriptors were defined to describe the electronic states of the
central atoms in each solvent molecule and its nearby hydrogen-bond
donors. Model analysis reveals that E
int is mainly determined by the charge state of the crucial fragments
and the hydrogen-bond donors of the solvent molecule. The custom-defined
descriptors not only improve the regression and prediction performance
but also enable the interpretability of the ML model. Additionally,
the absorption equilibrium measurements of solubilities of dimethyl
sulfide in several commercial solvents verified the strong correlation
between the dissolving affinity and solute–solvent intermolecular
interaction energy. The present study provides an approach to building
practical and interpretable intelligent algorithms to aid the development
of sustainable chemicals or processes.