Molten Salt Crystallization studied by Machine-Learned Potentials

Molecular dynamics (MD) simulations based on accurate interatomic force-fields are employed to calculate solid-liquid phase diagrams and investigate mechanisms of crystallization at the atomic level. The work is motivated by the potential relevance of molten salts in the context of lithium extraction. We report on the development of efficient and robust machine learning (ML) potentials for three unary salts (LiF, NaF and KF), three binary mixtures (LiF-NaF, LiF-KF and NaF-KF) as well as the ternary mixture of LiF-NaF-KF in the framework of Atomic Cluster Expansion and with the training datasets generated using density functional theory (DFT) with the strongly constrained and appropriately normed (SCAN) functional. We demonstrate that simulations based on these potentials produce properties consistent with DFT and available experimental data, including lattice constant, liquid density, melting point and latent heat of melting. The phase diagrams for the three binary mixtures and the ternary mixture calculated based on our ML potentials will be presented. In addition, we discuss results of simulations for homogeneous crystal nucleation, with focus on the crystallization pathways and role of metastable polymorphs. These studies are enabled by the attention paid in the potential development to accurate description of competing crystal phases, such as those with wurtzite, zinc blende and CsCl-prototype structures. The atomic-level mechanism of homogeneous melting of the rocksalt structure of LiF, NaF and KF will also be illustrated.