High-accuracy thermodynamic properties to the melting point from ab initio calculations aided by machine-learning potentials

Accurate prediction of thermodynamic properties requires an extremely accurate representation of the free-energy surface. Requirements are twofold-first, the inclusion of the relevant finite-temperature mechanisms, and second, a dense volume-temperature grid on which the calculations are performed. A systematic workflow for such calculations requires computational efficiency and reliability, and has not been available within an ab initio framework so far. Here, we elucidate such a framework involving direct upsampling, thermodynamic integration and machine-learning potentials, allowing us to incorporate, in particular, the full effect of anharmonic vibrations. The improved methodology has a five-times speed-up compared to state-of-the-art methods. We calculate equilibrium thermodynamic properties up to the melting point for bcc Nb, magnetic fcc Ni, fcc Al, and hcp Mg, and find remarkable agreement with experimental data. A strong impact of anharmonicity is observed specifically for Nb. The introduced procedure paves the way for the development of ab initio thermodynamic databases.

Automated summary

Accurate prediction of thermodynamic properties requires a precise representation of the free-energy surface, including finite-temperature mechanisms and dense volume-temperature grid calculations. A new framework involving direct upsampling, thermodynamic integration, and machine-learning potentials is introduced, which greatly improves computational efficiency and reliability. The method has been applied to calculate equilibrium thermodynamic properties for several materials and shows remarkable agreement with experimental data, particularly emphasizing the impact of anharmonicity for Nb. This procedure paves the way for developing ab initio thermodynamic databases.