xnn.common.data.hub.rmd17#

Revised MD17 (rMD17) dataset builder.

rMD17 recomputes 100,000 conformations of ten small molecules from the original MD17 at the PBE/def2-SVP level with tight SCF and dense grids, so it is practically free of numerical noise. Each molecule ships as a .npz and the authors provide five fixed train/test index splits of 1000 structures each.

Reference#

Christensen & von Lilienfeld, “On the role of gradients for machine learning of molecular energies and forces”, Mach. Learn.: Sci. Technol. 1 045018 (2020). Data: https://figshare.com/articles/dataset/12672038

Notes

  • Upstream units are kcal/mol (energies) and kcal/mol/angstrom (forces); positions are in angstrom. By default these are converted to eV and eV/A – the convention used elsewhere in xnn and by MACE/NequIP – controlled by the units argument.

  • The authors warn: do not train on more than 1000 samples, as the conformations are consecutive MD frames and thus correlated. The five official splits respect this; split="all" (100k frames) is offered only for custom splitting.

Classes

RMD17Builder()

Builder for the revised MD17 dataset.
On this page
Show Source