xnn.common.data.hub.ani1ccx

ANI-1ccx dataset builder (coupled-cluster subset of the ANI-1x release).

The ANI-1ccx data set holds ~500 k conformations for H/C/N/O organic molecules with energies at an approximate CCSD(T)/CBS level: the CCSD(T)*/CBS composite extrapolation (DLPNO-CCSD(T) + MP2/HF basis-set extrapolation) of Smith et al. It is an intelligently selected ~10 % sub-sample of the ANI-1x set, recomputed with coupled cluster and used to train the ANI-1ccx model (xnn.dnn.models.ani.ANI.ani1ccx()) by transfer learning. Upstream ships it inside the same ani1x-release.h5 file as ANI-1x – an ANI-1ccx conformation is simply one whose ccsd(t)_cbs.energy is not NaN – so this builder is a thin restriction of ANI1xBuilder to that level, sharing its cached download. CCSD(T)*/CBS is energy-only (no coupled-cluster forces were computed).

Reference

Smith et al., “Approaching coupled cluster accuracy with a general-purpose neural network potential through transfer learning”, Nat. Commun. 10, 2903 (2019) (the ANI-1ccx model), and Smith et al., “The ANI-1ccx and ANI-1x data sets, coupled-cluster and density functional theory properties for molecules”, Sci. Data 7, 134 (2020) (the release used here). Data: https://doi.org/10.6084/m9.figshare.10047041 Format / reader spec: aiqm/ANI1x_datasets

Notes

• Energies are in Hartree upstream; by default they are converted to eV (units="hartree" keeps the raw values). Positions are in angstrom.

• The DFT properties of these same conformations (e.g. for the transfer-learning pre-training stage) are available through load_dataset("ani1x", ...).

• Splits, caps, and caching behave exactly as for ani1x; the one ~5.6 GB release file is shared with the ani1x cache directory.

Classes

ANI1ccxBuilder()

Builder for the ANI-1ccx data set (CCSD(T)*/CBS energies, no forces).