Machine Learning with Physically Inspired Descriptors to Predict Solvation Free Energies of Neutral and Ionic Solutes in Aqueous and Nonaqueous Solvents

Solvation free energy is a key property for understanding various chemical processes such as ion solvation and phase transfer. The former corresponds to ionic solutes, while the latter is relevant to nonaqueous solvents. However, more attention has been paid to the prediction of the solvation free energies of neutral solutes in aqueous solvents. In the present work, we start from our published research (J. Phys. Chem. Lett. 2023, 14, 1877–1884), which was developed for predicting experimental hydration free energies of neutral solutes, and propose extensive machine learning models to predict solvation free energies of neutral and ionic solutes in aqueous and nonaqueous solvents. Two types of descriptors have been developed for solvents and ionic solutes. The former accounts for fundamental physical and chemical properties of solvents, and the latter is rationally designed based on thermodynamic cycles for the ion solvation process. Combined with our previously developed physically inspired descriptors, three machine learning predictors are built, achieving mean absolute errors of 0.44, 1.72, and 1.60 kcal/mol for neutral, anionic, and cationic solutes, respectively. Further analysis of the prediction performance and feature importance suggests the potential to improve prediction accuracy, especially for ionic solutes.