A Foundational Machine-Learned Interatomic Potential for Perovskites

Abstract

Perovskite materials cover a wide range of chemical compositions, all of which crystallize into a common structure. This structure admits a range of distortions and derivatives, from which a multitude of physical properties and applications arise. This project presents a foundational machine-learned model trained on 59 perovskite compositions. The ambition is that this model will facilitate perovskite research by allowing accurate atomistic simulations multiple orders of magnitude faster than density functional theory (DFT). The model uses the MACE architecture and the training data is generated via DFT computations. The training data includes relaxed structures at 0K as well as finite temperature structures generated by running molecular dynamics (MD) simulations. For the total energy, the model achieves a root mean square error on the order of 0.1 meV to 1 meV/atom, which is small compared to the room temperature thermal energy scale of order 10 meV/atom. The model is tested by running MD simulations and determining the structural phase transitions of BaTiO3, BaZrO3, BaZrS3, CsGeBr3, CsPbI3, NaNbO3 and SrTiO3. The model is generally able to determine the structural phases within about 100K of previously presented experimental and computational results. One exception is NaNbO3, where the model is unable to recognize the multitude of fast structural transitions. It is, however, able to recognize that complex dynamics take place, directing further study to this material. Possible improvements to the model include ensuring maximally diverse training sets, training on more chemical compositions and fine-tuning more universal models with the perovskite data.