Machine-Learning Database for paper "Defect-Driven Oxide Phase Transformation: Deep-Learning Molecular Dynamics of the Magnetite-to-Wüstite Transition"

The first file (01.train_data.tar.gz) contains the initial training dataset used to construct the Deep Potential model. The second file (run.tar.gz) provides the expanded dataset generated through the iterative sampling strategy described in this work, which incorporates additional Fe₃O₄ bulk, defect (Fe and O vacancies), and surface configurations. This dataset also includes the previously developed FeO dataset from our prior work (npj computational materials. 2025, 11, 32; Acta Materialia. 2026: 121832), ensuring comprehensive coverage of the Fe–O compositional and structural space. The third file (graph.pb) corresponds to the final trained Deep Potential model, which encapsulates the learned interatomic interactions and is used for all subsequent large-scale simulations reported in this study.