Thermodynamic properties at high temperature with machine learning

It is of crucial importance at high temperature to take into account anharmonic effects to correctly describe the thermodynamics properties of materials.

Ab-initio molecular dynamics is a natural way to tackle this issue, capture the atomic vibrations and obtain the phonon spectra. Unfortunately, its computational cost, especially for systems containing heavy elements, can impose to reduce the size of the supercell to obtain results in a correct time limit. Here we present a method to accelerate the computation of finite temperature properties with Machine-Learning Assisted Canonical Sampling (MLACS). The idea is to sample the canonical distribution with a machine learning  potential adjusted with a self-consistent scheme involving single points ab-initio calculations on selected configurations generated with the potential at the previous step. With this method we reproduce the results obtained with ab-initio molecular dynamics in a time divided by one or two order of magnitude. We will give here examples on complex systems involving actinide elements like actinide oxides (UO₂, PuO₂) and U₃Si₂.