Judicious Curation of DFT Machine Learning Datasets for Accurate, Flexible, and Transferrable Atomistic Potentials for Elemental Systems and Metal Oxides

Machine learning techniques have accelerated material discovery and expanded the impact of atomistic potentials by leveraging the predictive accuracy of density functional theory (DFT) to simulations with numbers of atoms and timescales unsuitable for DFT. We focus on the initial curation and subsequent systematic expansion of these training data for the machine-learning deep neural-network potentials (MLPs) training. Compact datasets were generated for element systems across the periodic table and transition metal oxides using clear and concise criteria applicable to most elemental systems or transition metal oxides. MLPs are validated after each iteration to gauge the impact of the new data and MPL accuracy by comparing calculated material properties with DFT reference values. Remarkably, we observe good transferability for material properties not included in the training (e.g., melting points). Furthermore, we explore the systematic dataset expansion to describe oxide surface energies and thermal expansion accurately. These MLPs effectively represent materials trained with small configurational datasets; they are transferrable and flexible, providing a launch point for the broad community to adapt and expand these data and their material discovery and optimization applications.