Machine-Learning Molecular Dynamics and Descriptor-Based Analysis of Solution Enthalpies and Diffusion Coefficients of Solutes in Liquid Sodium

Liquid metals have been actively investigated for applications in advanced nuclear systems. However, fundamental thermodynamic and transport properties for impurities remain sparse and are sometimes inconsistent across datasets, reducing their reliability for safety assessments. To address this gap, we conduct machine-learning molecular dynamics (MLMD) to obtain solution enthalpies and diffusion coefficients of solutes in liquid metals. This approach achieves accuracy close to that of density functional theory (DFT) while largely reducing computational cost.

In this study, we calculate solution enthalpies and diffusion coefficients for all elements from hydrogen (Z = 1) to curium (Z = 96) in liquid sodium using MLMD. For each solute, machine-learning moment tensor potentials are constructed from DFT-based first-principles molecular dynamics data and used in the MLMD. To interpret the results, we implement a descriptor-driven regression framework that systematically identifies the physicochemical factors governing dissolution and diffusion. Our analysis reveals that solution enthalpies are primarily determined by bond energy, electronegativity, and valence electron count, whereas diffusion coefficients are largely controlled by atomic mass, atomic radius, and electronegativity.