Atomic Charges via Gradient Boosting: Development and Application for Solvation Energies in Organic Solvents.

A gradient-boosting based atomic-charge scheme, BoostCha, is introduced. The BoostCha model operates in three steps: it first predicts pseudo-charges for individual atoms based on their local environments, represented by three-dimensional descriptors of Kocer-Mason-Erturk type, then refines these values using global molecular information, and finally restores the charge conservation. The BoostCha charges are employed as input features in two independent machine-learning models for predicting solvation free energies in organic solvents: ESE-Boost, a gradient-boosting model, and ESE-ANN, a dense artificial neural network. Both approaches yield strong predictive performance, with average root-mean-square errors of 0.49 and 0.52 kcal/mol, respectively. The methods demonstrate consistent performance across diverse solvent classes and are particularly accurate for alkanes, alcohols, ethers, esters, ketones, and aromatic and haloaromatic solvents.