Molecular dynamics simulations of lanthanum chloride by deep learning potential

Abstract It has been known for many years that mixtures of lanthanum chloride (LaCl3) with alkali chlorides in spent nuclear fuel show severe departures from ideality. To go deep into this non-ideality, the local structure of molten LaCl3 was investigated by experiments and classic molecular dynamics simulations. However, there appears to be some controversy concerning conclusions of these research methodology such as coordination numbers of the first coordination shell. In this work, interatomic potential driven by machine learning is developed based on data sets generated by ab initio calculations in order to research the local structure and property of molten LaCl3 at 1200 K, 1300 K, 1400 K and 1500 K. The machine learning potential enables higher efficiency and similar accuracy relative to DFT and yields precise descriptions of microstructures and properties. Microstructural evolution with target temperatures is analyzed through partial radial distribution functions, coordination numbers, angular distribution functions and total neutron structural factors. We observe short- and intermediate-range order and the latter disappears at high temperature. The sevenfold and eightfold coordinated structures are dominant in LaCl3 melt and the network structure is composed of corner-sharing, edge-sharing and face-sharing configurations. Evolution of properties including density and self-diffusion coefficient over the entire operating temperature range are documented. This work exhibits a thorough understanding of the local structure and property of LaCl3 melt and reveals the accuracy of machine learning potential on molten trivalent metal chlorides for first time.