On-the-fly Machine-learned Force Fields for MD Simulations of Ferroelectric Phase Transitions and Domain Wall Pinning

On-the-fly machine-learned force fields (MLFFs) have recently opened a new avenue for finite temperature calculations with first-principles accuracy. The force fields are trained using a combination of molecular dynamics (MD), density functional theory (DFT) and machine learning, with the latter deciding when to perform DFT steps and update the force fields.

Here, we apply machine-learned interatomic potentials to predict structural phase transitions in well-studied ferroelectric perovskite oxides, BaTiO₃, PbTiO₃, LiNbO₃ and BiFeO₃. Lattice parameters, thermal expansion and predicted phase transitions from ab initio MD simulations and experiments match well, with an expected underestimation of the structural phase transition temperatures of about 20 %. Furthermore, in BaTiO₃, our simulations accurately predict the order-disorder transition of Ti displacements that give rise to the multiple phases observed upon heating. Despite the ability to predict the drastic ferroelectric transition observed experimentally in BiFeO₃, the more elusive gamma-phase is however not observed in our simulations. Furthermore, we use the generated MLFFs together with nudged elastic band (NEB) calculations to study the mobility of ferroelectric domains walls and the pinning of these by point defects and compare with DFT-predicted values. Finally, we discuss the potential for further use of machine-learned potentials for investigating ferroelectric materials.