Finite-Temperature Cation Disorder and Dynamics in ZnTiN2 with Equivariant Neural Network Potentials

Zn-based ternary nitride semiconductors have been the subject of avid research due to their potential photoelectrochemical applications. Recent density functional theory (DFT) studies of ZnTiN2 have demonstrated a strong correlation between cation disorder and absorption onsets in these structures, but left open questions about the nature of disorder at non-zero temperatures and the disorder dynamics at interfaces. Equivariant neural networks interatomic potentials, trained on DFT calculations, have emerged as an accurate means for accelerating finite-temperature molecular dynamics simulations with limited training data. Here, we train an equivariant neural network interatomic potential (NNP) on DFT data of different cation-disordered ZnTiN2 configurations. We use an active learning approach, which incorporates data from NNP-driven Monte Carlo and molecular dynamics simulations, to an interatomic potential that is accurate across a broad range of thermodynamic phase space. Our model achieves a root mean square error of <1 meV/atom. With this potential, we generate a wide range of disordered atomic configurations beyond what was possible in previous studies and predict quantitatively how the cation disorder evolves with temperature.