Thermodynamic Properties of Actinides Systems at High Temperature with Machine Learning

At high temperatures, accurately describing the thermodynamic properties of materials requires considering anharmonic effects. Ab-initio molecular dynamics offers a natural approach to address this challenge by capturing atomic vibrations and obtaining phonon spectra. However, the computational expense, particularly for systems containing heavy elements, often necessitates reducing the size of the supercell to obtain timely results. To overcome this limitation, we introduce a method for accelerating the computation of finite temperature properties using Machine-Learning Assisted Canonical Sampling (MLACS)[1]. This approach involves sampling the canonical distribution using a machine learning potential adjusted through a self-consistent scheme that incorporates single-point ab-initio calculations on selected configurations generated with the potential from the previous step. With MLACS, we achieve results comparable to those obtained with ab-initio molecular dynamics but with a computational time reduced by one or two orders of magnitude. We illustrate the effectiveness of this method through examples involving complex systems containing actinide elements such UO2, U3Si2, and pure Pu[2].

[1] A. Castellano, F. Bottin, J. Bouchet, A. Levitt, and G. Stoltz, Phys. Rev. B 106, L161110 (2022)
[2] F. Bottin, R. Béjaud, B. Amadon, A. Castellano, and J. Bouchet, Phys. Rev. B , 109(6), L060304 (2024)