Uncertainty-informed transferable deep learning potentials for simulating BeF₂-LiF system

BeF₂-LiF multicomponent molten salts are promising prototypes for salts in reactor systems due to their favorable thermophysical and transport properties. However, complex structures across different salt concentrations have not yet been explored because of the limitations associated with experimental measurements. Herein, predictive molecular dynamics plays a crucial role in assisting, enhancing, and validating the interpretation of the experimental data. However, developing accurate models for simulating the structure properties of ionic melts is challenging. While ab initio simulations are limited by the computational cost, the classical models are accuracy-limited. This significantly inhibits our ability to design new compositions and materials. Here, an accurate, efficient, and transferable machine learning potential is developed to predict structure-properties relationships for a wide concentration range for BeF₂-LiF system. Optimal training compositions were chosen using uncertainty quantification based on latent-space (principal component analysis) exploration. As such, developed neural network potentials (NNIP) trained on 33% BeF₂ (FLiBe) along with pure LiF, BeF₂, and 75% BeF₂ were shown to simulate structures of liquid phases for different compositions of LiF-BeF₂ accurately. Further theoretical neutron-weighted structure factors (S(Q) reveals the source of pre-peak seen in the experiment. As such, the transferability of NNIP across different phases and compositions opens up the possibility of advanced screening for unseen compositions with different structural complexity.