Machine learning based force-fields for strongly anharmonic materials

Computing materials properties at finite temperatures, such as thermal or electric conductivity, poses a particular challenge for anharmonic materials because the commonly used phonon-based perturbation theories break down and going beyond the harmonic approximation is costly and tedious. This applies to some extent also to other materials, since at finite temperatures the structure and the symmetry are thermodynamic averages of energetically similar, lower symmetry structures and therefore anharmonic effects are more common than generally assumed. Alternatively, sufficiently long ab initio molecular dynamics trajectories would contain the required information to calculate the abovementioned properties, however the computational cost of is in most cases prohibitively high. Recently, machine learning based force-fields have made obtaining said the trajectories feasible. However, those methods come with the intrinsic problem of choosing appropriate training data, since encountering a not well-represented structure during the molecular dynamics will lead to failure of the calculation, which is an issue for anharmonic materials in particular.

In this talk we are going to introduce an “on-the-fly” machine learning method, that will automatically update the training data and refit the force-field once a new structure is encountered. The force-field in this approach is based on the usual lattice-dynamics expansion of the total energy [1], which is particularly suitable for taking into account anharmonic effects. As an example, the method will be applied to obtain force-fields for the strongly anharmonic SnSe.

 

[1] F. Zhou et al., Phys. Rev. Lett. 113, 185501 (2014)