Transferable Deep Learning Potential Reveals Intermediate-Range Ordering Effects in LiF-NaF-ZrF4 Molten Salt

Abstract

LiF-NaF-ZrF4 multicomponent molten salts are promising candidate coolants for advanced clean energy systems owing to their desirable thermophysical and transport properties. However, the complex structures enabling these properties, and their dependence on composition, is scarcely quantified due to limitations in simulating and interpreting experimental spectra of highly disordered, intermediate-ranged structures. Specifically, size-limited ab-initio simulation and accuracy-limited classical models used in the past, are unable to capture a wide range of fluctuating motifs found in the extended heterogeneous structures of liquid salt. This greatly inhibits our ability to design tailored compositions and materials. Here, accurate, efficient and transferable machine learning potentials are used to predict structures far beyond the first coordination shell in LiF-NaF-ZrF4. Neural networks trained at only eutectic compositions with 29% and 37% ZrF4 are shown to accurately simulate a wide range of compositions (11 to 40% ZrF4) with dramatically different coordination chemistries, while showing a remarkable agreement with theoretical and experimental Raman spectra. The theoretical Raman calculations further uncovered the previously unseen shift and flattening of bending band at ~250 cm-1 which validated the simulated extended-range structures as observed in compositions with higher than 29% ZrF4 content. In such cases, machine learning-based simulations capable of accessing larger time- and length-scales (beyond 17 {\AA}) were critical for accurately predicting both structure and ionic diffusivities.