Quantum-Accurate Modeling of Ferroelectric Phase Transition in Perovskites from Message-Passing Neural Networks

Graph-based message-passing neural networks (MPNNs) have been proposed to facilitate computational research on materials at the atomic scale, which represent the chemical structure as an indirect graph and incorporate the message-passing scheme to learn the interaction between atoms. Here, we employ the MPNN framework to investigate the temperature-dependent structural phase transitions of perovskites. We take two prototypical ferroelectric perovskites, BaTiO3 and PbTiO3, as examples to demonstrate the application of this approach. Our results show that well-trained MPNN models achieve a similar level of accuracy as density functional theory (DFT) calculations in terms of both energy and force, with an accuracy of a few meV per atom. This level of accuracy fulfills the requirement for investigating structural changes. By integrating MPNN models as calculators in molecular dynamics simulations, we investigated the structural phase transitions of the two compounds and reproduced their phase transition sequences. The simulated phase transition temperatures and lattice parameters are comparable to the experimental or DFT results. Moreover, we examined the influence of exchange–correlation functionals on the trained MPNN models. This study demonstrates that MPNN, which can serve as a universal energy model, presents an appealing approach to treating the structural properties of perovskites.