Dynamically training machine learning based force-fields for strongly anharmonic materials

When it comes to computing materials properties at finite temperatures, machine learning (ML) based force-field methods are a promising alternative in particular for cases where commonly used phonon-based perturbation theories break down or cannot easily be extended beyond the harmonic approximation. While sharing the same advantages with ab initio molecular dynamics (MD) these methods require only a fraction of the computational cost. The quality of the force-field and thus the success of the method however crucially depends on choosing adequate training data. In particular with statically obtained training data the force-field method will fail once it encounters a not well represented structure during the MD.

In this work we are going to showcase our dynamically trained ML force-field method applied to materials with varying degree of anharmonicity (c-BAs, Si, SnSe). The force-field itself is based on the usual lattice dynamics expansion of the total energy, which allows us to utilize Bayesian error estimation for the decision about updating the training data. We find that evaluating the Bayesian error as trajectory average leads to an efficient exploration of the configuration space.