Temperature dependence of O solubility in liquid Na by atomistic simulation of Na(l)-Na2O(s) interfaces using corrected machine learning potential: a step towards simulating Na combustion

Liquid Na combustion is a significant safety concern in sodium-cooled fast reactors. Atomistic simulations are an alternative to experiments for studying detailed mechanisms of complex combustion processes. However, accurate simulations of the interfaces involved in combustion are challenging even for density functional theory (DFT), because the systematic error between different chemical systems cannot be fully cancelled. Herein, we report the achievement of a key milestone in atomistic simulation of liquid Na combustion, which involves the development of a machine learning (ML) moment tensor potential that allows accurate simulation of interface systems between liquid Na and solid Na2O. The ML potential is trained by using supervised and active learning to ensure DFT-level accuracy. An empirical correction is then applied to achieve experimental accuracy by reducing systematic error. Consequently, the basic properties of liquid Na and solid Na2O are accurately simulated. In addition, with empirical correction, experimental O solubility data for liquid Na at 350-900 K are reproduced by using interface molecular dynamics simulations and a thermodynamic model. The temperature dependence of the enthalpy and entropy of the Na2O solution and their effect on O solubility are evaluated. The results show that, despite the increase in solution enthalpy with temperature, O solubility increases more rapidly than the linear Arrhenius plot due to the effect of solution entropy. The results of this study indicate that, with appropriate correction, ML potentials can achieve near-experimental accuracy, beyond the accuracy of DFT, in interface simulations and material properties calculations, paving the way for sodium combustion simulations in the future.

Automated summary

In this study, a machine learning moment tensor potential that accurately simulates the interface systems between liquid Na and solid Na2O was developed. The potential was trained using supervised and active learning to ensure DFT-level accuracy, and an empirical correction was applied to achieve experimental accuracy. As a result, the basic properties of liquid Na and solid Na2O were accurately simulated.