Machine learning interatomic potentials for sulfur vacancy dynamics in single sheet MoS2

This dataset contains two types of machine learning interatomic potentials (MLIPs) that have been trained to accurately describe the dynamics of sulfur vacancies in single sheet 2H-molybdenum disulfide (MoS2): (i) A graph neural network (GNN) that was fine-tuned for MoS2 starting with the MACE MP-0 foundation model using the MACE python package and (ii) a gaussian approximation potential that was trained on defective MoS2 structures using the Vienna ab-initio Simulation Package (VASP, version 6.4.0, https://vasp.at/). Both MLIPs are able to accurately describe the potential energy curves of the atomic jump that governs the sulfur vacancy dynamics. Benchmarking the MLIPs by running density functional theory calculations on equidistantly selected snapshots of MLIP MD trajectories yields force test errors of 42 meV/A and 74 meV/A for the GNN and GAP potentials, respectively. The test and training data set for the MACE model was obtained using the CP2K software package (version 2022.2, https://www.cp2k.org/). In addition to the trained MLIP models, this dataset also contains the corresponding training data sets.