minimal_data_QM9_QMUGS dataset

This dataset is used to enable simple experiments with our orbital-free density functional theory code from https://github.com/sciai-lab/structures25.

Dataset 1: QM9_perturbed_fock

This dataset contains a small subset of QM9 molecules.

minimal_data_QM9_QMugs/
├── QM9_perturbed_fock/
│   ├── dataset_statistics/
│   │   ├── dataset_statistics_labels_local_frames_global_symmetric_natrep_e_kin_plus_xc.zarr
│   │   └── dataset_statistics_labels_no_basis_transforms_e_kin_plus_xc.zarr
│   ├── labels/
│   │   ├── 0000825.zarr.zip
│   │   └── ...
│   ├── labels_local_frames_global_symmetric_natrep/
│   │   ├── 0000825.zarr.zip
│   │   └── ...
│   ├── split.pkl
│   └── split.yaml    
└── QMUGS_perturbed_fock/
    ├── dataset_statistics/
    │   ├── dataset_statistics_labels_local_frames_global_symmetric_natrep_e_kin_plus_xc.zarr
    │   └── dataset_statistics_labels_no_basis_transforms_e_kin_plus_xc.zarr
    ├── labels/
    │   ├── 0342979.zarr.zip
    │   └── ...
    ├── labels_local_frames_global_symmetric_natrep/
    │   ├── 0342979.zarr.zip
    │   └── ...
    ├── split.pkl
    └── split.yaml  

Description folder:

• dataset_statistics: The dataset statistics contain information about the statistics of the energies, coefficients and gradients of the data. These are used for normalization techniques inside the model, and intial guesses of the density. They are loaded once at model or SADGuesser initialization. For each basis transform, the statistics will vary which is why each basis transformations gets its own directory. The dataset_statistics_labels_local_frames_global_symmetric_natrep_e_kin_plus_xc.zarr statistics used for the model, while the dataset_statistics_labels_no_basis_transforms_e_kin_plus_xc.zarr statistics are used for the SADGuesser.
• labels: The untransformed label zarr files. These contain all data necessary for model training as well as OFDFT evaluations.
• labels_local_frames_global_symmetric_natrep: The transformed label zarr files. The transformation is applied to all values in the spatial group according to their representation.
• split.pkl: The split file, defining which label files belong to train, val and test split.
• split.yaml: The same split file in yaml format, not used by the code but for humans to parse quickly.

Internal structure .zarr file (e.g., 0000825.zarr.zip)

geometry and of_labels are used during training and density optimization, while ks_labels is there to easily compare against Kohn-Sham energies but is not used in the usual workflows.

>>> print(zarr_store.tree())
/
├── geometry
│   ├── atom_pos (13, 3) float64
│   ├── atomic_numbers (13,) uint8
│   └── mol_id () <U11
├── ks_labels
│   ├── basis () <U12
│   └── energies
│       ├── e_electron (35,) float64
│       ├── e_ext (35,) float64
│       ├── e_hartree (35,) float64
│       ├── e_kin (35,) float64
│       ├── e_nuc_nuc (35,) float64
│       ├── e_tot (35,) float64
│       ├── e_xc (35,) float64
│       └── has_energy_label (35,) bool
└── of_labels
    ├── basis () <U96
    ├── energies
    │   ├── e_electron (35,) float64
    │   ├── e_ext (35,) float64
    │   ├── e_ext_mod (35,) float64
    │   ├── e_hartree (35,) float64
    │   ├── e_kin (35,) float64
    │   ├── e_kin_minus_apbe (35,) float64
    │   ├── e_kin_plus_xc (35,) float64
    │   ├── e_kinapbe (35,) float64
    │   ├── e_tot (35,) float64
    │   └── e_xc (35,) float64
    ├── n_scf_steps () int64
    └── spatial
        ├── basis_integrals (35, 904) float64
        ├── coeffs (35, 904) float64
        ├── dual_basis_integrals (35, 904) float64
        ├── grad_ext (35, 904) float64
        ├── grad_hartree (35, 904) float64
        ├── grad_kin (35, 904) float64
        ├── grad_kin_minus_apbe (35, 904) float64
        ├── grad_kin_plus_xc (35, 904) float64
        ├── grad_kinapbe (35, 904) float64
        ├── grad_tot (35, 904) float64
        └── grad_xc (35, 904) float64

Internal structure dataset_statistics_labels_local_frames_global_symmetric_natrep_e_kin_plus_xc.zarr

The dataset statistcs are split into multiple weigher_keys like constant or ground_state_only. This is because for some applications we want to filter, mainly which SCF iterations of our data we want to include. For example, for a good ground state prediction, only the ground state coefficients matter.

>>> print(dataset.tree())
/
├── constant
│   └── initial_guess_delta
│       ├── abs_max (477,) float64
│       ├── mean (477,) float64
│       └── std (477,) float64
├── ground_state_only
│   ├── atom_ref_atom_type_bias (5,) float64
│   ├── atom_ref_energy_only_atom_type_bias (5,) float64
│   ├── atom_ref_energy_only_global_bias () float64
│   ├── atom_ref_global_bias () float64
│   ├── coeffs
│   │   ├── abs_max (477,) float64
│   │   ├── mean (477,) float64
│   │   └── std (477,) float64
│   ├── energy_minus_atom_ref
│   │   ├── abs_max () float64
│   │   ├── mean () float64
│   │   └── std () float64
│   ├── energy_minus_scalar_atom_ref
│   │   ├── abs_max () float64
│   │   ├── mean () float64
│   │   └── std () float64
│   ├── gradient_label
│   │   ├── abs_max (477,) float64
│   │   ├── mean (477,) float64
│   │   └── std (477,) float64
│   ├── gradient_max_after_atom_ref (477,) float64
│   ├── gradient_max_after_scalar_atom_ref (477,) float64
│   ├── scalar_atom_ref_atom_type_bias (5,) float64
│   └── scalar_atom_ref_global_bias () float64
├── has_energy_label
│   ├── atom_ref_atom_type_bias (5,) float64
│   ├── atom_ref_energy_only_atom_type_bias (5,) float64
│   ├── atom_ref_energy_only_global_bias () float64
│   ├── atom_ref_global_bias () float64
│   ├── coeffs
│   │   ├── abs_max (477,) float64
│   │   ├── mean (477,) float64
│   │   └── std (477,) float64
│   ├── energy_minus_atom_ref
│   │   ├── abs_max () float64
│   │   ├── mean () float64
│   │   └── std () float64
│   ├── energy_minus_scalar_atom_ref
│   │   ├── abs_max () float64
│   │   ├── mean () float64
│   │   └── std () float64
│   ├── gradient_label
│   │   ├── abs_max (477,) float64
│   │   ├── mean (477,) float64
│   │   └── std (477,) float64
│   ├── gradient_max_after_atom_ref (477,) float64
│   ├── gradient_max_after_scalar_atom_ref (477,) float64
│   ├── scalar_atom_ref_atom_type_bias (5,) float64
│   └── scalar_atom_ref_global_bias () float64
├── initial_guess_only
│   └── initial_guess_delta
│       ├── abs_max (477,) float64
│       ├── mean (477,) float64
│       └── std (477,) float64
├── min_scf_iteration_3
│   ├── atom_ref_atom_type_bias (5,) float64
│   ├── atom_ref_energy_only_atom_type_bias (5,) float64
│   ├── atom_ref_energy_only_global_bias () float64
│   ├── atom_ref_global_bias () float64
│   ├── coeffs
│   │   ├── abs_max (477,) float64
│   │   ├── mean (477,) float64
│   │   └── std (477,) float64
│   ├── energy_minus_atom_ref
│   │   ├── abs_max () float64
│   │   ├── mean () float64
│   │   └── std () float64
│   ├── energy_minus_scalar_atom_ref
│   │   ├── abs_max () float64
│   │   ├── mean () float64
│   │   └── std () float64
│   ├── gradient_label
│   │   ├── abs_max (477,) float64
│   │   ├── mean (477,) float64
│   │   └── std (477,) float64
│   ├── gradient_max_after_atom_ref (477,) float64
│   ├── gradient_max_after_scalar_atom_ref (477,) float64
│   ├── initial_guess_delta
│   │   ├── abs_max (477,) float64
│   │   ├── mean (477,) float64
│   │   └── std (477,) float64
│   ├── scalar_atom_ref_atom_type_bias (5,) float64
│   └── scalar_atom_ref_global_bias () float64
└── min_scf_iteration_5
    ├── atom_ref_atom_type_bias (5,) float64
    ├── atom_ref_energy_only_atom_type_bias (5,) float64
    ├── atom_ref_energy_only_global_bias () float64
    ├── atom_ref_global_bias () float64
    ├── coeffs
    │   ├── abs_max (477,) float64
    │   ├── mean (477,) float64
    │   └── std (477,) float64
    ├── energy_minus_atom_ref
    │   ├── abs_max () float64
    │   ├── mean () float64
    │   └── std () float64
    ├── energy_minus_scalar_atom_ref
    │   ├── abs_max () float64
    │   ├── mean () float64
    │   └── std () float64
    ├── gradient_label
    │   ├── abs_max (477,) float64
    │   ├── mean (477,) float64
    │   └── std (477,) float64
    ├── gradient_max_after_atom_ref (477,) float64
    ├── gradient_max_after_scalar_atom_ref (477,) float64
    ├── scalar_atom_ref_atom_type_bias (5,) float64
    └── scalar_atom_ref_global_bias () float64

Dataset 2: QMUGS_perturbed_fock

This dataset contains a small subset of QMugs molecules, which we use to test generalization to larger molecules. The structure of the data is exactly the same as for QM9.