xnn.common.data.hub.spookynet_states#

Builder for the electronic-state datasets of the SpookyNet paper.

Unke et al. (Nat. Commun. 12, 7273, 2021, Fig. 4) published two GFN2-xTB datasets in which the same geometries occur in two electronic states, so a model can only tell them apart from the total charge or the spin:

  • ag3: 2200 Ag3+ and Ag3- clusters;

  • carbene: 2200 singlet and triplet CH2 molecules.

550 structures were sampled around the minimum of each state by normal-mode sampling at 1000 K and every structure was recomputed in the other state. Each entry has the total charge, the number of unpaired electrons (spin_multiplicity = S + 1), the energy (eV), the forces (eV/Angstrom) and the dipole moment about the origin (e Angstrom).

Notes

  • Data: Zenodo record 10.5281/zenodo.5115732 (CC BY 4.0), SQLite files.

  • The paper trained on 1000 random structures; the builder draws the splits from a seeded permutation (1000 / 100 / 1100 by default), which is reproducible but not the authors’ own.

Classes

SpookyNetStatesBuilder()

Builder for the Ag3+/Ag3- and singlet/triplet CH2 sets.
Show Source