DATA - Exploring Lithium Diffusion in LiF with Machine Learning Potentials: From Point Defects to Collective Ring Diffusion

This dataset accompanies the article Exploring Lithium Diffusion in LiF with Machine Learning Potentials: From Point Defects to Collective Ring Diffusion (DOI: https://doi.org/10.1038/s41524-026-02132-8). It contains the heavy data needed to reproduce the LiF machine-learning interatomic potential (MLIP) training workflow: Quantum ESPRESSO DFT single-point outputs used as reference labels, DeepMD training and validation datasets derived from those outputs, and the final four-member DeepMD model committee. The DFT labels were computed with Quantum ESPRESSO using the PBE exchange-correlation functional, PAW pseudopotentials for Li and F from the QE pseudopotential library, scalar-relativistic flavor, Gamma-point sampling, wavefunction and charge-density cutoffs of 110 Ry and 440 Ry, Gaussian smearing of 0.005 Ry, SCF convergence threshold of 5e-7 Ry, TF-local charge mixing, and mixing parameter beta = 0.2. Initial candidate geometries were generated from CP2K AIMD trajectories, but CP2K energies and forces were not used as MLIP labels. The archive is organized into dft/, dataset/, and mlip/. The dft/ folder stores renamed QE .pwo files grouped by physical system and source (aimd or active_learning), with full provenance in dft/metadata_copy_source.tsv. The dataset/ folder contains the DeepMD train/validation databases and conversion scripts. The mlip/04-final_20240408/ folder contains the final DeepMD model inputs, outputs, frozen graphs, compressed frozen graphs, and SLURM retraining templates. The data are distributed as mlip_data.zip. Unzip this archive before use; extraction creates the dft/, dataset/, mlip/, and metadata files described in the README. The companion reproducibility repository is available at https://github.com/paolodeangelis/lif-defect-and-ring-diffusion.