Atomic Neural Network for Calculation of Solvation Free Energies in Organic Solvents

Abstract

This paper introduces AtomicESE, an artificial neural network for calculating solvation-free energies ΔG°_solv of molecules in organic solvents. AtomicESE calculates ΔG°_solv by summing atomic contributions, each evaluated by a dense neural network. This atomic network uses 13 physically relevant input features, comprising six local atomic features, two global charge-related molecular properties, and five solvent-specific properties. For neutral solutes, AtomicESE achieves an average RMSE below 0.6 kcal/mol, demonstrating strong performance across all solvent classes, with particularly high accuracy for aromatic, haloaromatic, alkane, and ketone solvents. AtomicESE also works reliably for ionic solutes.