Development and Validation of Versatile Deep Atomistic Potentials for Metal Oxides

Metal oxide systems are of great interest for various technologies such as coatings and thin film applications. Simulating these materials at larger timescales and system sizes remain a challenge for materials modeling. Calculations based on standard density functional theory while accurate are computationally expensive. On the other hand, methods based on atomistic potentials that can be extended to large system sizes and time scales often fail to describe the different oxidation states of metal oxides. Herein we focus on creating machine learning deep neural-network potentials for different metal oxide systems in a systematic and replicable manner using recently introduced method for convergence acceleration of the training set. Each of these potentials are trained on experimentally verified structures without limitations on the selection of their oxidation states. We validate these interatomic potentials by matching various computed material properties with density functional theory. We demonstrate how these interatomic potentials and their associated training data can be used for further studies to reduce the gap between atomistic simulation and experiments, such as polymorphism, interfaces between different phases, crystalline growth, and other transformations that might include changes in composition. Our study, in addition to our previous studies of metals and binary metallic alloys, demonstrates that machine learning potentials are effective in describing systems with mixed levels of ionic/covalent bonding.