Machine-Learned Modeling for Accelerating Organic Solvent Design in Metal-Ion Batteries

Abstract

Organic solvents offer a promising avenue for enhancing metal-ion battery performance, for instance, in suppressing dendritic formation. To expedite the discovery of optimal electrolyte formulations, this study integrates density functional theory calculations with machine learning to accurately predict binding energies between metal ions and organic solvents. Leveraging a vast dataset of over 300 organic molecules, an extra trees regressor model is developed and demonstrated to exhibit exceptional predictive capabilities. The model's performance is underscored by its high ${\mathrm{R}}^2$ values on both validation and test sets. Key descriptors contributing to the model's accuracy include the number of valence electrons in the metal ion, the atomic number of the metal ion, and features associated with the van der Waals surface. By applying the trained model to a dataset of up to 20 000 unseen organic molecules, potential high-performance electrolyte additives are identified. Notably, ${\mathrm{C}}_{13}{\mathrm{H}}_{24}{\mathrm{N}}_2{\mathrm{O}}_2$ and ${\mathrm{C}}_{11}{\mathrm{H}}_{18}{\mathrm{O}}_5$ emerge as promising candidates for Zn-ion and Mg-ion batteries, respectively, outperforming conventional additives. To gain deeper insights into the microscopic behaviour of these identified molecules, molecular dynamics simulations are conducted. This research establishes a robust in silico framework for accelerating the design of advanced metal-ion batteries through the rational selection of organic solvents.