(376b) A Machine Learning Guided Approach for Studying Local Solvation Environments

In this work we introduce a calculation scheme that uses state of the art computational tools to study ion solvation environments. We first generate microsolvated clusters with global optimization algorithm called ABCluster and then introduce a measure of convergence based on the structural similarities of solvent molecules near the solute using unsupervised machine learning algorithms called Smooth Overlap of Atomic Positions (SOAP) and sketchmaps. SOAP is a clustering technique that identifies structures that are closely related to each other, and the sketchmaps reduce the dimensionality of the feature vectors to represent the SOAP kernels. After studying different solvation environments, we calculate solvation energies using Quasi Chemical Theory (QCT) and thermodynamic cycles. We demonstrate that low energy molecular clusters produced by our procedure have structurally similar local solvation environments, and this suggests that this calculation scheme can be used to quantify the number of explicit solvent molecules needed to accurately model the relevant local solvation environment of a charged solute. We expect this approach will be broadly applicable for the atomistic modeling of ions in different solvents.