A general purposed machine learning interatomic potential for Mg-Al-Si-O system suitable for Earth materials at high pressure and temperature conditions

Abstract Accurate phase diagrams and thermodynamic properties of Earth materials are essential for advancing geophysical, geodynamical and geological studies. Apart from experiment, atomistic simulations, particularly molecular dynamics, can be used to obtain thermodynamic data, but they often fail to reproduce correct phase relations. In this study, we develop a machine learning interatomic potential for the Mg–Al–Si–O system, optimized for accuracy and computational speed. Among several tested functionals, the r2SCAN exchange-correlation functional proves most suitable for generating training data encompassing over 20 minerals and melts. To enhance accuracy, a pairwise Gaussian correction is applied, reducing the energy error from 5.2 kJ/mol to 1.2 kJ/mol. Predicted isochemical phase diagrams show good agreement with experiments. Beyond phase diagrams, we calculate solid-melt interfacial free energy for periclase and forsterite and find that the anisotropy of solid-melt interfacial free energy is low (6%) for periclase and moderate (12%) for forsterite. The influence of nonhydrostatic stress on the α-β quartz transition is systematically examined, demonstrating that mean stress serves as a reliable proxy with about 17% error. This work illustrates that molecular dynamics simulations powered by machine learning interatomic potentials offer a powerful approach to investigating the physical properties of deep Earth materials.