Machine-learning potentials based on Meta-GGA functionals for solution enthalpy calculations in liquid metals

Gil, J.; 小林 恵太 ; 板倉 充洋  

Gil, J.; Kobayashi, Keita; Itakura, Mitsuhiro

This study aims to calculate the solution enthalpies of solutes in liquid metals with experimental accuracy at reasonable cost. For this purpose, we train machine-learning potentials on first-principles data generated with meta-GGA exchange-correlation functionals. Previous GGA-based calculations required empirical post-calculation corrections, while direct meta-GGA simulations - though more accurate - are generally too costly to achieve statistically meaningful results for liquid metals. Machine-learning potentials trained on meta-GGA data via active learning enable much faster calculations with improved accuracy. Applying our approach to oxygen in liquid sodium, the calculated solution enthalpy closely matches experimental values without any corrections. This workflow can be extended to other solutes in liquid metals, particularly where experimental data are lacking or unreliable, to build a comprehensive database.