binghuang2018/qmc4aqml

A small dataset for testing multi-level quantum machine learning models aiming for QMC accuracy

The QMC4AQML dataset

Authors:

• Bing Huang (University of Vienna, [email protected])
• O. Anatole von Lilienfeld (University of Toronto, [email protected])
• Jaron T. Krogel (Oak Ridge National Laboratory, [email protected])
• Anouar Benali (Oak Ridge National Laboratory, [email protected])

General

• Purpose: used for testing multi-level quantum machine learning model aiming for an accuracy at the quantum Monte Carlo (QMC) level
• Molecular data are provided in the extxyz format (https://wiki.fysik.dtu.dk/ase/ase/io/formatoptions.html#extxyz).
• Geometries were optimzed at the level of B3LYP/cc-pVTZ
• All FN-QMC energies were obtained using the B3LYP/cc-pVTZ nodal surfaces.
  • retrievable through the key qmc_b3lyp in the extxyz file
• Quantum chemical (QC) levels of theory include
  • HF, DFT (PBE, B3LYP)
  • post-HF levels of theory: MP2, CCSD(T)
  • and basis set is fixed to cc-pVTZ
  • with SPE retrievable as the key of format theory+vtz
    • e.g., B3LYP/cc-pVTZ energy is assigned the key b3lypvtz (which is identical to key b5vlypvtz)

Test molecules

• located in the folder targets/ (all energies in Hartree)
• totalling 50 molecules, drawn randomly from the QM9 dataset
• calculated at FN-QMC level and various other cheaper levels

Training set

Two sets of amons are available:

• graph amons (with QMC energy) of the QM9 dataset (in the folder amons-ni5-qmc/) (all energies in kcal/mol)
  • made up of at most 5 heavy atoms
  • totalling 1175
  • amounts to force-field global minima
• conformer amons (without QMC energy) for the 50 test molecules (in amons-ni7/) (all energies in Hartree)
  • the maximal number of heavy atoms is now 7
  • include all conformational degrees of freedom
    • that is, multiple conformer amons may share the same SMILES string
  • the conformer amons for, say the first test molecule (targets/frag_0001.xyz), are amons-ni7/frag_0001_a*xyz