NequIP-OAM-M-ft-periclase-brucite-water-njord-2026-v1: an r²SCAN machine-learning interatomic potential and active-learning dataset for the Mg–O–H system (periclase, brucite, water)

A machine-learning interatomic potential (MLIP) and its training dataset for the Mg-O-H system: periclase (MgO), brucite (Mg(OH)2), liquid water, and their reactive interfaces (dissolution, surface hydroxylation, water dissociation, and proton transport). The model is a fine-tune of the NequIP-OAM-M-0.1 foundation on r2SCAN DFT labels. The training set was grown over three rounds of committee-based active learning. Model and reference data. Elements and type map [H, O, Mg]; units eV, Angstrom, and eV/Angstrom for forces; DFT reference VASP 6.4.3, r2SCAN, ENCUT 1200 eV, EDIFF 1e-7, identical across all iterations. The final production model was trained on a 96/4 train/validation split with a constant learning rate and EMA, shipping the last checkpoint (no early stopping). Its validation force MAE is 6.16 meV/Angstrom The dataset (final_dataset.xyz) is a cumulative 6054-frame extended-XYZ corpus (iter_00 3600, iter_01 1000, iter_02 1454). Every frame is self-describing through its comment-line info keys: seed_id, system, source_class, iter_added, selected_by, T_K and P_GPa (where the sampling run id defines them), cluster and rescued_on (which HPC produced the DFT label), and the DFT settings. Composition: mineral+water_surface 70.6%, bulk_water 11.9%, bulk_brucite 8.7%, bulk_mgo 8.7%. How it was grown. Each active-learning round sampled configurations with LAMMPS MD and well-tempered metadynamics (PLUMED reactive collective variables) driven by the previous round's committee, selected frames by farthest-point sampling on SOAP descriptors plus committee disagreement, collective-variable coverage, and quotas, labeled them with VASP r2SCAN under a locked INCAR, and trained a six-member NequIP-OAM-M committee. The terminal round trained the single production model released here. Quality evidence. The committee's per-frame disagreement tracks the realised force error (log-log r about 0.89), committee-versus-DFT force parity is R^2 about 0.9996, and the labels are consistent across the HPC backends used. Files in this record: nequip-oam-m-ft-periclase-brucite-water-njord-2026-v1.nequip.zip – the model, a self-contained NequIP package. final_dataset.xyz.gz – the 6054-frame training dataset with per-frame provenance. final_dataset.e0.json – the isolated-atom reference energies (E0 shifts). active_learning_history.zip – the full provenance (the three iterations, step manifests, a hash-verified MANIFEST, the committees, and quality evidence). active_learning_raw.zip – the optional raw trajectories and VASP outputs, verifiable against the history bundle's hashes. README.md – usage, units, quality evidence, and attribution. License and attribution. Released under CC-BY-4.0. This model is a fine-tune of NequIP-OAM-M-0.1 (mir-group, https://www.nequip.net, CC-BY-4.0, pre-trained on OMat24, sAlex, and MPTrj); please attribute it and cite the NequIP and Allegro papers (Batzner et al., Nature Communications 13, 2453 (2022); Musaelian et al., Nature Communications 14, 579 (2023)). Workflow orchestration, dataset assembly, and release packaging were carried out with assistance from Claude Code (Anthropic); the authors are responsible for the released model and data.