Predicting practical reduction potential of electrolyte solvents via computational hydrogen electrode and interpretable machine-learning models

Abstract

Accurate prediction of practical reduction electrode potentials (E_red) of electrolyte solvents of electrochemical energy storage devices relies on calculating the Gibbs free energy in their reduction reaction. However, the emergence of new electrolyte solvents and additives leaves most of the reaction mechanisms unveiled. Here, we provide a machine-learning-assisted workflow of thermodynamically quantified E_red prediction for electrolyte solvents. A computational hydrogen electrode model based on density functional theory calculation is generalized for calculating the reaction free energy of electrochemical elementary steps. Machine-learning models are trained based on the organic and inorganic electrolyte solvents that possess experimentally identified reduction mechanisms. Validation of the best-scoring model is conducted by experimental validation of 6 additional solvents. Multiple thermodynamics features are found impactful on E_red through different chemical bonding with reaction intermediates. This workflow enables accurate E_red prediction for electrolyte solvents without identified reduction mechanisms, and is widely applicable in the electrochemical energy storage area.