See the preprint Synthetic Data Enable Experiments in Atomistic Machine Learning for more details.

The dataset contains 546 uncorrelated carbon trajectories at a variety of densities (ranging from 1.0 gcm-3 to 3.5 gcm-3) and temperatures.

As can be seen from the examples above (generated using Ovito), this dataset captures a wide variety of chemical environments and features, including carbon nano-tubes, graphitic films, buckyball-esque clusters, cubic and hexagonal diamond and tetrahedral amorphous carbon.

Each atomic environment has been labelled with a "local-energy" (together with a force) by the C-GAP-17 potential, and these are included in the .extxyx files as per atom quantities. These can be accessed using, for instance, the Atomic Simulation Environment package (ase):

from ase.io import read


trajectory = read("results/density-1.0-T-2000.extxyz", index=":")
structure = trajectory[0]
local_energies = structure.get_array("gap17_energy")

The density, anneal temperature, trajectory id and timestamp for each structure is given as a per-structure quantity in the header of each .extxyz entry.

density = structure.info["density"]          # in gcm-3
temperature = structure.info["temperature"]  # in K
trajectory_id = structure.info["run_id"]     # integer
timestamp = structure.info["time"]           # in ps

Each trajecotry was seeded using a structure generated using the ./generate_structure.py script, which generates random structures for a given density using a hard-sphere constraint.

The LAMMPS molecular dynamics package, together with the C-GAP-17 potential was then used to perform a melt-quench-anneal simulation, with temperature profile as depicted above. Snapshots were taken at 1ps intervals, for a total of 210 snapshots per trajectory. (546 trajectories _ 210 snapshots _ 200 atoms = 22.9 million atomic enviroments).

If you use this dataset in your research, please cite the following:

@misc{Gardner-22,
  title = {Synthetic Data Enable Experiments in Atomistic Machine Learning},
  author = {Gardner, John L. A. and Beaulieu, Zo{\'e} Faure and Deringer, Volker L.},
  year = {2022},
  number = {arXiv:2211.16443},
  eprint = {2211.16443},
  eprinttype = {arxiv},
  primaryclass = {physics},
  doi = {10.48550/arXiv.2211.16443},
  archiveprefix = {arXiv}
}