Dataset and Deep Potential Model for Vacancy-driven atomistic reconstruction and phase stability across the Fe2O3–Fe3O4–FeO–Fe reduction sequence

This repository contains the datasets, input files, and trained Deep Potential model supporting the study “Vacancy-driven atomistic reconstruction and phase stability across the Fe2O3–Fe3O4–FeO–Fe reduction sequence” The deposited files provide the core data and simulation inputs used to construct and apply the Fe–O machine-learning interatomic potential, with particular emphasis on the α-Fe₂O₃ configurations introduced in the present work and on the stochastic oxygen-removal molecular-dynamics simulations used to analyse vacancy-driven hematite reconstruction. The repository includes: developed_dataset.tar.gz — the developed Fe–O dataset used for model training and validation, including the α-Fe₂O₃ configurations generated for the present study. init_data.rar — the initial reference dataset used in the Deep Potential workflow. cp2k.input — representative CP2K input settings used for first-principles calculations of reference energies, atomic forces, and stresses. param_cp2k.json — parameters used for CP2K-based first-principles labelling within the active-learning workflow. graph.pb — the trained Fe–O Deep Potential model used for the molecular-dynamics and structural analyses reported in the manuscript. Fe2O3_delO_MD.in — LAMMPS input for stochastic oxygen-removal Deep Potential molecular dynamics of α-Fe₂O₃. The protocol progressively removes oxygen from hematite and relaxes the structure between deletion events, enabling analysis of Fe–O coordination, oxygen-sublattice reconstruction, and apparent atomic diffusivity during imposed oxygen depletion. The Fe–O potential used in this work spans metallic Fe, FeO/Fe₁₋ₓO, Fe₃O₄, and α-Fe₂O₃ environments. In the present study, the model was used to investigate the structural response of hematite to progressive oxygen deficiency and to connect vacancy-driven oxygen-sublattice reconstruction with finite-temperature Fe–O phase stability. The oxygen-removal simulations are designed to probe the solid-state structural response to imposed oxygen deficiency. They do not explicitly model the elementary gas–solid reaction sequence involving H₂ adsorption, hydrogen transfer, hydroxyl formation, or H₂O desorption. These files are provided to facilitate reproducibility of the first-principles labelling, Deep Potential training workflow, and oxygen-removal DPMD simulations reported in the associated manuscript.