Training data for the 2nd generation of AI-enhanced QM method (AIQM2)

This repository contains the training data for the 2nd generation of AI-enhanced QM method (AIQM2) [10.26434/chemrxiv-2024-j8pxp-v2]. There are two datasets. wb97xtz_minus_gfn2xtbstar.20250620.h5 is used for training AIQM2@DFT and ccsdtstarcbs_minus_gfn2xtbstar_minus_d4wb97x_chno_mlatom.20240828.h5 is used for transfer learning to AIQM2 at coupled-cluster level. They are generated by subtracting the baseline and dispersion energies from the reference levels (wb97x/def2-TZVPP and CCSD(T)*/CBS), respectively. The reference data comes from ANI-1x and ANI-1ccx datasets (https://github.com/aiqm/ANI1x_datasets). Details can be checked in the preprint. These two datasets are in h5 format. The structure follows: /2 dict # number of atoms/2/coordinates array (1325, 2, 3) [float64] # in angstroms/2/energies array (1325,) [float64] # total energy in Hartree/2/forces array (1325, 2, 3) [float64] # the coupled-cluster level data doesn't contain forces/2/species array (1325, 2) [int8] # atomic numbers of the molecule They can be easily loaded with MLatom (https://github.com/dralgroup/mlatom) and transformed into the standard molecular database format. The example codes are provided in load_dataset.py. The MLatom used is in version 3.18.0