Transfer Learning Lithium and Electrolyte Potential Energy Surfaces from Pure and Hybrid DFT

Training and test datasets of labeled data (atomic coordinates, system net charges, system energies, atomic forces, and atomic partial charges) associated with the NeurIPS 2022 workshop paper "Transfer Learning Lithium and Electrolyte Potential Energy Surfaces from Pure and Hybrid DFT".

atomic_energies_PBE-D3.json

Per-atom energy offsets for PBE, fitted to minimize totalEnergy - sum(offsets)

atomic_energies_wB97X-D3BJ__def2-TZVPD.json

Per-atom energy offsets for wB97X-D3BJ/def2-TZVPD, fitted to minimize totalEnergy - sum(offsets)

atom_types.json

Mapping of elements to consecutive integer atom types

layers_uniform.json

Hidden layer sizes by element (in this case, the same for all elements)

symmetry_function_parameters_v2.json

Symmetry function parameters for AEV featurization

training_set.json

Training dataset, a superset of that from Dajnowicz et al, containing 600,390 total datapoints, of which 362,382 are from Dajnowicz et al and 238,008 are new calculations in a format compatible with the old. Of the newly added points, 6902 are periodic systems. GFN2-xTB charges have also been added to all non-periodic systems for which GFN2-xTB labeling converged (581,954 points), including the points from the Dajnowicz et al dataset.

The keys are arbitrary integers (stored as strings)

Each value is a dictionary with the following properties:

{

  'atomicNumbers': [integers representing atomic number for each atom],

  'elements': [strings representing the element for each atom (redundant with atomic numbers)],

  'positions': [[x, y, z] coords of each atom, in Angstroms],

  'charge': integer representing system total charge,

  'multiplicity': integer representing spin multiplicity,

  'lattice': [[lattice vectors, in Angstroms, only exists if periodic]],

  'labels': dict of QM reference data for this point (energy, gradient, etc)

}

Every system needs at least one entry under 'labels': 'PBE-D3' or 'wB97X-D3BJ__def2-TZVPD'.

Some label types will be missing from some points, for example xtbCharges for periodic systems.

Labels not present for a datapoint are simply excluded from the loss function for that datapoint.

Example 'labels' entry:

{

  'PBE-D3': {

    'totalEnergy': system energy in Hartree,

    'atomizationEnergy': system energy, in Hartree, minus per-atom energy offsets (see atomic_energies_PBE-D3.json or atomic_energies_wB97X-D3BJ__def2-TZVPD.json),

    'gradient': [[dE/dx, dE/dy, dE/dz] for each atom, in Hartree/Angstrom]

  }

  'wB97X-D3BJ__def2-TZVPD': {

     'totalEnergy': system energy in Hartree,

     'gradient': [[dE/dx, dE/dy, dE/dz] for each atom, in Hartree/Angstrom],

     'dipoleMoment': [system dipole moment vector x, y, z in electron-Angstroms],

     'atomizationEnergy': system energy, in Hartree, minus per-atom energy offsets (see atomic_energies_PBE-D3.json or atomic_energies_wB97X-D3BJ__def2-TZVPD.json)

     'xtbCharges': [atomic partial charges from GFN2-xTB, in e-]

  }

}

barriers_set.json

Carbonate decomposition barrier test set of 20 systems, in the same format as the dataset psei0.json, where the key for each point is a string of form NAME_start (for starting systems) or NAME_transition (for transition states). Test set RMSEs are calculated for the relative energy between each system pair (NAME_transition - NAME_start).