Unveiling ionic redox potentials: Advancing prediction through large-scale MLMD and FEP integration

Abstract

The simulation of redox potentials in molten salts faces challenges in achieving large-scale automation, limiting the accuracy and flexibility of molecular dynamics predictions. This paper proposes a novel and accurate workflow for predicting redox potentials using deep neural networks, free energy perturbation theory, and large-scale machine learning molecular dynamics. Feasibility and accuracy of the workflow are validated by using industrial magnesium electrolysis, comparing results with cyclic voltammetry curves. Maximum simulated error is just 0.3 V. Training the potential energy surface models and exploring solvation structure and energy distribution are integral to our research process. Our trained models exhibit DFT-level accuracy, with energy and force errors of 10-4 and 10-2 orders of magnitude, respectively. Solvation structure investigation reveals that introducing La3+ can release Mg2+ from MgO. A linear correlation between the free energy difference and temperature is observed. This workflow addresses the mismatch between scale and accuracy in predicting redox potentials.