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

Area of Science:

• Computational chemistry
• Machine learning in chemistry

Background:

• Accurate atomic charges are crucial for predicting molecular properties.
• Existing methods for calculating atomic charges may have limitations in accuracy or applicability.

Purpose of the Study:

• To introduce BoostCha, a novel gradient-boosting based atomic-charge scheme.
• To evaluate the performance of BoostCha-derived atomic charges in predicting solvation free energies using machine learning models.

Main Methods:

• BoostCha predicts atomic charges in three steps: local pseudo-charge prediction, global refinement, and charge conservation restoration.
• Three-dimensional Kocer-Mason-Erturk descriptors represent atomic local environments.
• Two machine learning models, ESE-Boost (gradient-boosting) and ESE-ANN (artificial neural network), were used to predict solvation free energies.

Main Results:

• Both ESE-Boost and ESE-ANN models achieved strong predictive performance for solvation free energies.
• The average root-mean-square errors were 0.49 kcal/mol for ESE-Boost and 0.52 kcal/mol for ESE-ANN.
• The models demonstrated consistent accuracy across diverse organic solvent classes, including alkanes, alcohols, ethers, esters, ketones, and aromatic solvents.

Conclusions:

• The BoostCha atomic-charge scheme provides accurate input features for machine learning models predicting solvation free energies.
• The developed machine learning approaches show robust and reliable performance in diverse chemical environments.
• This work advances the application of machine learning in computational chemistry for predicting solvation properties.