The in silico modeling of molten salts is of crucial importance to emerging "carbon free" energy applications, but is inhibited by the computational cost of quantum mechanically treating the high polarizabilities characteristic of molten salts. Here, we integrate configurational sampling using classical force-fields with active learning to automate the generation of near-DFT accurate machine learning Gaussian Approximation Potentials (GAP) for molten LiCl using fewer than 600 atomic configurations. Relative to conventional ab initio molecular dynamics, the molten LiCl GAP model exhibits a 19,000x speedup and improved experimental agreement as gauged by calculated R-factors. The accuracy of the GAP parametrization workflow is validated by its ability to reproduce experimental structure factors, densities, self-diffusion coefficients, and ionic conductivities for molten LiCl. This hybrid simulation strategy significantly accelerates the generation of machine learning potentials for molten salts by reducing the expensive ab initio calculations required for parameterization to O(100) evaluations, enabling the facile generation of first-principles quality predictions of structural and dynamical properties of molten salts.
Automated Development of Molten Salt Machine Learning Potentials: Application to LiCl
18 March 2021, Version 1
This content is a preprint and has not undergone peer review at the time of posting.