WB96MV-ORGANIC / README.md
janwein's picture
Update README.md
76d82f0 verified
|
Raw
History Blame Contribute Delete
19.5 kB
---
license: mit
task_categories:
- tabular-regression
tags:
- chemistry
- quantum-chemistry
- computational-chemistry
- density-functional-theory
- molecular-geometry
- molecular-machine-learning
- interatomic-potentials
- pyscf
pretty_name: Chembricks ωB97M-V Organic
size_categories:
- 100K<n<1M
---
# Chembricks ωB97M-V Organic
Chembricks ωB97M-V Organic is a curated computational quantum-chemistry dataset containing **801,781 molecular structures**. Each structure provides a gas-phase ωB97M-V single-point energy, HOMO and LUMO energies, an atom-resolved nuclear gradient, formal charge and spin, Cartesian coordinates, and traceable source provenance.
The DFT labels were evaluated on converged gas-phase GFN2-xTB geometries. They are protocol-defined computational reference values—not experimental measurements and not universal ground truth.
## Dataset summary
| Property | Value |
|---|---:|
| Retained molecular identities | 801,781 |
| Retained conformers | 801,781 |
| Raw source records curated | 805,046 |
| Duplicate source occurrences merged | 1,615 |
| Hard-invalid source records rejected | 1,650 |
| Quarantined source occurrences | 0 |
| Retained records with an explicit connectivity flag | 2,791 |
| Formal charge −1 / 0 / +1 | 3,669 / 790,722 / 7,390 |
| PySCF spin `2S = 0` | 801,781 |
| Electronic-structure driver | RKS for every retained structure |
| Elements | H, B, C, N, O, F, Si, P, S, Cl, Br, I |
| Phase | Gas |
The serialized format supports multiple conformers per molecular identity. In this release, the number of retained conformers equals the number of canonical identities, so each `vacuum_conformers` array contains one retained representative.
## Intended uses
This dataset is suitable for research on:
- molecular energy and orbital-energy regression;
- atomistic energy/gradient and force-model development;
- representation learning on molecular geometries;
- uncertainty estimation and out-of-distribution detection;
- quantum-chemistry surrogate models; and
- analysis of SCF behavior across organic and main-group chemical space.
It is not intended to provide experimental observables, solvent-phase properties, reaction barriers, DFT-optimized stationary points, or thermochemical quantities.
## Release scope
The release scope was fixed by a deterministic source-chunk rule:
| Source campaign | Included chunk IDs | Source files | Raw records |
|---|---:|---:|---:|
| Campaign 1 | 0–38 | 39 | 397,482 |
| Campaign 2 | 0–81 | 82 | 407,564 |
| **Total** | | **121** | **805,046** |
The chunk-prefix selection is reproducible, but it was not a randomized or distribution-preserving sample. Keep the supplied source references when constructing evaluations, and do not assume that records are independently and identically distributed solely because canonical identities are unique.
## From SMILES to the stored DFT result
```text
SMILES + formal charge
10 candidate 3D structures generated with g-xTB
gas-phase GFN2-xTB very-tight geometry optimization
identity, composition, convergence and geometry validation
gas-phase PySCF ωB97M-V single-point calculation
energy + HOMO/LUMO + nuclear gradient + xTB geometry
deduplication, conflict checks and robust outlier curation
```
The workflow starts from a molecular SMILES and explicit formal charge. g-xTB generates ten candidate three-dimensional structures. A selected candidate is optimized in the gas phase with GFN2-xTB using the very-tight convergence setting. PySCF then evaluates one ωB97M-V single point on that xTB geometry.
The DFT stage does **not** optimize the geometry (`perform_opt: false`) and does not request thermochemistry (`thermo: false`). The dataset therefore contains no DFT stationary geometries, Hessians, harmonic frequencies, zero-point energies, enthalpies, or finite-temperature Gibbs free energies. A nonzero stored gradient is expected because it is evaluated at an xTB-optimized rather than ωB97M-V-optimized geometry.
## Quantum-chemistry protocol
| Setting | Value |
|---|---|
| Functional | ωB97M-V (`wb97m-v`) |
| Functional class | Range-separated hybrid meta-GGA with VV10 nonlocal correlation |
| Program | PySCF 2.9.0 |
| Phase | Gas; `solvent: null` |
| Geometry | Gas-phase GFN2-xTB very-tight optimization |
| DFT calculation | Single point only |
| Numerical grid | PySCF grid level 5 plus the VV10 nonlocal grid |
| Density fitting | RI-JK through PySCF `density_fit()` |
| Orbital basis/ECP scheme | `qmcbasis`: NWChem-format, commit-declared aug-cc-pVTZ orbital-basis and matching ccECP map for observed elements |
| Auxiliary basis | `cc-pvdz-jkfit`; iodine override `def2-universal-jkfit` |
| SCF level shift | 0.5 Hartree |
| Maximum SCF cycles | 250 |
| Driver and spin | RKS; every retained entry has PySCF `spin = Nα − Nβ = 2S = 0` |
| Direct SCF | PySCF default behavior; no explicit screening tolerance was serialized |
| Coordinates | Å |
| Energies | Hartree (`Eh`) |
| Nuclear gradient | Hartree/bohr (`Eh/a₀`) |
ωB97M-V includes exact Hartree–Fock exchange. RI-JK therefore fits both the Coulomb (J) and exchange (K) contributions rather than using a Coulomb-only J-fit.
CDS means **cavitation–dispersion–solvent-structure** correction. Because these calculations are gas phase and do not use SMD, `solvent_energy` and `cds_correction` are exactly zero for every retained structure.
## Curation and quality control
The curation procedure canonicalized identity with RDKit and applied the following non-negotiable checks:
- converged GFN2-xTB geometry optimization and converged DFT SCF;
- finite SCF, solvent, CDS, HOMO, LUMO, coordinate, and gradient values;
- matching element composition, formal charge, electron-count parity, spin, and RKS driver;
- one coordinate and one gradient vector per atom, with three finite Cartesian components;
- `HOMO <= LUMO` and an SCF-cycle count between 1 and 250;
- the expected gas-phase ωB97M-V method, basis scheme, grid, level shift, and calculation mode;
- no severe atomic collisions; and
- robust median/MAD energy screens within canonical-identity and molecular-formula groups.
Within one canonical identity, symmetry-aware exact duplicates were identified after atom mapping and rigid alignment. Both the heavy-atom and all-atom RMSD had to be at most `10^-6 Å`. Duplicate labels were required to agree within `10^-5 Hartree` for SCF energy, solvent energy, CDS correction, HOMO, and LUMO; aligned gradient components were required to agree within `10^-6 Hartree/bohr`. One deterministic representative was retained and redundant occurrences were attached to its provenance rather than emitted as extra training examples.
The released curation policy allows a composition-consistent geometry to differ in connectivity from its canonical SMILES. The diagnostic proximity graph uses `1.25 ×` the sum of RDKit covalent radii when checking expected SMILES bonds. The **2,791 retained structures** for which an expected bond was not present carry:
```json
"curation_flags": ["connectivity_mismatch_allowed_by_mild"]
```
These structures are intentionally visible rather than silently relabeled. Users can retain them, exclude them, or evaluate them as a separate slice. Geometry-derived identity inference was permitted when source SMILES was absent and uniquely resolvable, but no retained record in this release required it.
### Curation accounting
| Outcome | Count | Share of 805,046 raw records |
|---|---:|---:|
| Retained distinct representative | 801,781 | 99.594% |
| Duplicate occurrence merged into provenance | 1,615 | 0.201% |
| Quarantined | 0 | 0.000% |
| Hard rejected | 1,650 | 0.205% |
| **Total** | **805,046** | **100.000%** |
The final artifact passed streaming standard-JSON parsing, non-finite-value checks, ID uniqueness checks, exact accounting closure, source-reference validation, geometry/gradient shape checks, method-profile validation, and a complete SHA-256 rehash.
## Distributed file
| File | Size | SHA-256 |
|---|---:|---|
| `wb97m_v_organic_public_2026-08-16.json` | 3,266,960,267 bytes | `e1abc9a68ca2a782d9bf0c9605325a9be67607c68e394f7b365df529599cffeb` |
The artifact is standard JSON, not JSON Lines. It uses an identity-grouped envelope rather than storing one independent object per line:
```text
root
├── schema_version
├── metadata
├── source_files[]
└── molecules[]
├── molecular identity
└── vacuum_conformers[]
├── geometry
├── log_data
└── provenance and curation fields
```
## JSON field reference
### Top-level fields
| Field | Type | Meaning |
|---|---|---|
| `schema_version` | string | Serialization schema version; currently `1.0.0`. |
| `metadata` | object | Dataset identity, version, counts, curation settings, units, calculation protocol, and manifest fingerprints. |
| `source_files` | array of objects | Table of source-file provenance. `source_ref.file_id` values resolve against this table. |
| `molecules` | array of objects | The 801,781 retained canonical molecular identities. |
### `metadata`
| Field | Meaning |
|---|---|
| `dataset_name` | Human-readable dataset name. |
| `dataset_version` | Version of the curated release. |
| `description` | Short description of the retained labels and structures. |
| `counts` | Exact `molecules`, `conformers`, and `source_records` counts. |
| `state_mapping` | Declares that `vacuum_conformers` stores validated gas-phase DFT states. |
| `identity_policy` | Canonical molecular-identity rule. |
| `curation_policy` | Duplicate, label-tolerance, identity-inference, and connectivity-policy settings. |
| `units` | Canonical units for coordinates, energies, gradients, and timing data. |
| `calculation_protocols` | Full protocol object referenced by each conformer's `protocol_id`. |
| `source_manifest_logical_sha256` | Logical fingerprint of the source manifest used for the build. |
| `dependency_manifest_logical_sha256` | Logical fingerprint of the curation implementation and runtime dependencies. |
| `curation_config_logical_sha256` | Logical fingerprint of the effective curation configuration. |
### `source_files[]`
| Field | Type | Meaning |
|---|---|---|
| `id` | integer | Dataset-local file identifier used by conformer provenance. |
| `relative_path` | string | Stable logical source path; it is not a local machine path. |
| `collection` | string | Source campaign identifier. |
| `protocol_id` | string | Calculation protocol applied to records in the source file. |
| `record_count` | integer | Number of raw records in that source file. |
| `size_bytes` | integer | Raw source-file size in bytes. |
| `sha256` | string | SHA-256 digest of the raw source file. |
### `molecules[]`
| Field | Type | Meaning |
|---|---|---|
| `molecule_id` | string | Deterministic ID of the form `mol-` plus 24 hexadecimal characters, derived from the canonical SMILES with a versioned SHA-256 namespace. It is a dataset ID, not an InChIKey. |
| `canonical_smiles` | string | RDKit canonical isomeric SMILES used as the molecular identity. |
| `original_smiles` | array of strings | Unique original SMILES carried by retained representative records grouped under this identity. |
| `formal_charge` | integer | Molecular formal charge; validated against both identity and `log_data.charge`. |
| `vacuum_conformers` | array of objects | Retained gas-phase conformer records. The schema supports multiple conformers; this release contains one per identity. |
### `vacuum_conformers[]`
| Field | Type | Meaning |
|---|---|---|
| `conformer_id` | string | Deterministic ID of the form `conf-` plus 24 hexadecimal characters, derived from canonical identity and the normalized conformer payload. |
| `geometry` | object | Ordered atom symbols and Cartesian coordinates of the converged xTB geometry. |
| `log_data` | object | Normalized DFT labels, protocol fields, gradient, and optional runtime diagnostics. |
| `source_occurrence_count` | integer | Number of raw occurrences represented by this retained conformer. A value greater than one means redundant occurrences were merged. |
| `source_ref` | object | Provenance of the deterministic representative. |
| `merged_source_refs` | array of objects, optional | Provenance of redundant source occurrences merged into the representative. These are not extra training examples. |
| `curation_flags` | array of strings, optional | Explicit retained-policy flags. Absence means that no such flag applies. |
### `geometry`
| Field | Type | Meaning |
|---|---|---|
| `elements` | array of strings, length `N` | Atomic symbols in the stored geometry atom order. |
| `coordinates` | array of `[x, y, z]`, shape `[N, 3]` | Finite Cartesian coordinates aligned row-for-row with `elements`; unit is Å. |
| `unit` | string | Always `angstrom`. |
| `method` | string | `xTB`, indicating that the geometry came from the GFN2-xTB optimization. |
`geometry.elements[i]`, `geometry.coordinates[i]`, and `log_data.gradient[i]` always refer to the same atom.
### `log_data`
| Field | Type | Unit | Meaning |
|---|---|---|---|
| `converged` | boolean | — | Always `true` for retained SCF calculations. |
| `scf_energy` | number | Hartree | Total gas-phase ωB97M-V SCF energy. |
| `homo` | number | Hartree | Highest occupied molecular-orbital energy. |
| `lumo` | number | Hartree | Lowest unoccupied molecular-orbital energy. |
| `gradient` | array of `[gx, gy, gz]`, shape `[N, 3]` | Hartree/bohr | Nuclear derivative `∂E/∂R` in geometry atom order. A force target is the negative gradient. |
| `charge` | integer | elementary charge | Total molecular charge; equals the enclosing `formal_charge`. |
| `spin` | integer | — | PySCF spin `Nα − Nβ = 2S`; every retained value is `0`. This is not the multiplicity `2S + 1`. |
| `driver` | string | — | SCF driver; always `RKS` in this release. |
| `cycles` | integer | — | Number of SCF cycles, from 1 through 250. |
| `xc` | string | — | Exchange-correlation functional identifier; `wb97m-v`. |
| `basis` | string | — | Basis/ECP scheme identifier; `qmcbasis`. Resolve details through the matching protocol object. |
| `grid_level` | integer | — | PySCF numerical grid level; `5`. |
| `level_shift` | number | Hartree | SCF level shift; `0.5`. |
| `solvent` | null | — | Always `null`; calculations are gas phase. |
| `solvent_energy` | number | Hartree | Always `0.0` in this gas-phase release. |
| `cds_correction` | number | Hartree | Always `0.0`; CDS is the cavitation–dispersion–solvent-structure correction. |
| `perform_opt` | boolean | — | Always `false`; no DFT geometry optimization was performed. |
| `thermo` | boolean | — | Always `false`; no Hessian/frequency thermochemistry was requested. |
| `protocol_id` | string | — | Foreign key into `metadata.calculation_protocols`; here `dev_crux_0153337`. |
| `time` | number, optional | seconds | Reported total calculation timing diagnostic. |
| `pyscf_time` | number, optional | seconds | Reported PySCF-stage timing diagnostic. |
| `memory_usage_mb` | number, optional | MB | Reported memory-use diagnostic. |
| `ncores` | integer, optional | cores | Reported CPU-core count. |
Runtime and resource fields are environment-dependent diagnostics, not molecular labels.
### Provenance objects
`source_ref` and entries in optional `merged_source_refs` contain:
| Field | Meaning |
|---|---|
| `file_id` | Foreign key into top-level `source_files`. |
| `result_index` | Zero-based record index inside that source file. |
| `uuid` | Original source UUID when supplied. |
| `relative_path` | Stable logical source path. |
| `collection` | Source campaign identifier. |
For every conformer, `source_occurrence_count` is the representative occurrence plus the number of entries in `merged_source_refs`.
## Memory-safe loading
The JSON file is approximately 3.27 GB, so avoid `json.load()` unless sufficient memory is available. The top-level `molecules` array can be streamed with [`ijson`](https://pypi.org/project/ijson/):
```bash
pip install huggingface_hub ijson
```
```python
from huggingface_hub import hf_hub_download
import ijson
path = hf_hub_download(
repo_id="chembricks/WB96MV-ORGANIC",
filename="wb97m_v_organic_public_2026-08-16.json",
repo_type="dataset",
)
with open(path, "rb") as handle:
molecules = ijson.items(handle, "molecules.item")
for molecule in molecules:
conformer = molecule["vacuum_conformers"][0]
elements = conformer["geometry"]["elements"]
coordinates_angstrom = conformer["geometry"]["coordinates"]
energy_hartree = conformer["log_data"]["scf_energy"]
gradient_hartree_per_bohr = conformer["log_data"]["gradient"]
# Train, transform, or yield the record here.
```
## Recommended ML practice
- Split on `molecule_id`, canonical SMILES, or a molecular scaffold—not on provenance occurrences.
- Do not expand `merged_source_refs` into repeated examples; those objects document redundancy.
- Keep the coordinate, element, and gradient arrays in exactly the same atom order.
- If training forces, use `force = -gradient` and convert both energy and length units consistently.
- Preserve `protocol_id`, even though the release currently contains one protocol.
- Decide explicitly how to handle the 2,791 `connectivity_mismatch_allowed_by_mild` structures and report that decision.
- Do not compare raw total energies across different elemental compositions as though they were directly normalized molecular properties.
- Prefer scaffold-aware or source-aware evaluation when estimating chemical generalization.
## Limitations
- These are computed labels and inherit the approximations of GFN2-xTB geometries, ωB97M-V, the basis/ECP scheme, RI-JK density fitting, and the numerical settings.
- The structures are not DFT stationary points; gradients generally do not vanish.
- The release contains closed-shell RKS calculations only and does not represent open-shell chemistry.
- It contains gas-phase values only. Solvation, condensed-phase effects, and finite-temperature thermochemistry are absent.
- Canonical SMILES is the identity key, but 2,791 retained geometries carry an explicit connectivity-change flag.
- The deterministic source-chunk selection is not equivalent to a random sample of chemical space.
- Timing and memory fields depend on the calculation environment and should not be treated as intrinsic molecular properties.
## License
The dataset is released under the MIT License.
## Citation
If you use the dataset, cite the Hugging Face repository and the ωB97M-V method paper:
```bibtex
@misc{chembricks_wb97mv_organic_2026,
author = {Chembricks},
title = {Chembricks ωB97M-V Organic},
year = {2026},
howpublished = {Hugging Face dataset},
url = {https://huggingface.co/datasets/chembricks/WB96MV-ORGANIC},
note = {Version 2026-08-16-v1-mild}
}
```
- N. Mardirossian and M. Head-Gordon, “ωB97M-V: A combinatorially optimized, range-separated hybrid, meta-GGA density functional with VV10 nonlocal correlation,” *Journal of Chemical Physics* **144**, 214110 (2016). [DOI: 10.1063/1.4952647](https://doi.org/10.1063/1.4952647)
- [PySCF density-fitting documentation](https://pyscf.org/user/df.html)