Machine-Learning Database for paper "Atomic-scale mechanism of iron crystallization during hydrogen reduction revealed by an accurate deep-learning force field"

This dataset provides the complete training database and Deep Potential (DP) model parameters developed for the study:“Atomic-scale mechanism of iron crystallization during hydrogen reduction revealed by an accurate deep-learning force field.”It contains the atomic configurations, first-principles reference data, and the trained machine-learning potential used to model the Fe–O system (Fe₁₋ₓO and metallic Fe phases) with DFT-level accuracy. 1. init.rar — Initial Dataset and Structural Inputs This archive contains the foundational datasets and structural files used for developing the Deep Potential (DP) force field for Fe₁₋ₓO. 1.1 01.train_data Subfolder dpmd_dataset/ — Processed dataset for DeepMD training, including atomic coordinates, forces, and energies derived from DFT and AIMD simulations. dpmd_rawfiles/ — Raw outputs from ab initio molecular dynamics simulations (CP2K), used to construct the DeepMD training dataset. 1.2 02.develop_data Subfolder Contains .vasp structure files used to expand and refine the DP training space, covering multiple physical states of the Fe–O system: bulk — Bulk Fe₁₋ₓO structures with different lattice constants. defect — Fe and O vacancy configurations (single and double vacancies). surface — Fe₁₋ₓO surfaces in (100), (110), and (111) orientations. bcc / fcc — Metallic Fe phases with body-centered and face-centered cubic structures, including lattice-scaled and defect/surface-modified configurations. These datasets ensure that the DP model accurately describes both oxide and metallic phases of iron across different thermodynamic and structural environments. 2. iterxx.rar — Iterative DP-GEN Refinement Data (Iterations 00–60) Each folder corresponds to one iteration in the DP-GEN automated training workflow. Every iteration includes three subfolders: 00.train/ — Training data and model outputs generated during the current iteration. 01.model_devi/ — Model deviation analysis comparing DP predictions with DFT results, guiding data selection for the next iteration. 02.fp/ — First-principles (FP) reference results from CP2K used for model correction and refinement. These iterative cycles systematically improve the DP model’s accuracy and transferability across bulk, defect, and surface environments. 3. Additional Files cp2k.input — Input file for CP2K ab initio simulations. param_cp2k.json — Parameter configuration file defining CP2K calculation settings. dpgen.log — Log file documenting DPGEN operations during data generation and training. record.dpgen — Comprehensive record of iterative DPGEN processes, including parameters, task assignments, and training results. 4. frozen_model.pb — Trained Deep Potential Model This is the final frozen DP model obtained after 60 iterative refinement cycles.It represents the optimized DeepMD interatomic potential for the Fe–O system, capable of performing large-scale molecular dynamics simulations with near-DFT accuracy.The model successfully reproduces key energetic, structural, and dynamic properties of Fe₁₋ₓO, as well as the atomic-scale crystallization process of disordered Fe into the bcc Fe phase.