{"success":true,"data":{"author":"Wolverton group, Northwestern University (OQMD); preprocessed by materials-toolkits","bucketAccessType":"public","bucketName":"materials-toolkits/oqmd","bucketPrefix":"","category":"Chemistry & Materials","cloudProvider":"huggingface","createdAt":"2026-11-16T02:38:31.137Z","description":"Structures and energies from the Open Quantum Materials Database (OQMD, Wolverton group, Northwestern University), extracted from the v1.8 SQL dump of February 2026 and preprocessed for machine learning. all.hdf5 has 1,407,269 structures (14.3 million atoms, 1 to 312 per cell), one per OQMD entry, choosing the converged static or standard calculation with the lowest energy per atom; unique.hdf5 drops the 147,716 entries OQMD marks as duplicates, leaving 1,259,553. Labels are the VASP PBE (GGA+U) total energy, formation energy per atom, stability relative to the convex hull (negative below the hull, as OQMD stores it) and band gap. Labels are not cleaned: about 1,200 entries have unphysical energies or stabilities from failed calculations and should be filtered. Parquet index files map each row to OQMD entry, calculation and structure ids; an older 2024 export is kept as oqmd.tar.gz. The repository is about 1.2 GB (measured 2026-11-15). Used for formation energy and stability prediction, phase diagrams and crystal generative models. OQMD data are licensed under CC-BY-4.0; cite Saal et al., JOM 2013, and Kirklin et al., npj Computational Materials 2015.","displayTitle":"OQMD v1.8: 1.4 Million DFT Crystal Structures and Formation Energies (HDF5)","id":"5509616e-d05b-5703-8fa7-3fca617ed523","license":"CC-BY-4.0","licenseEvidenceUrl":"https://oqmd.org/download/","objectCountEstimate":9,"probeStatus":"ok","referenceUrl":"https://doi.org/10.1038/npjcompumats.2015.10","region":"us","schemaCard":{"accessHints":[{"language":"python-datasets","snippet":"from datasets import load_dataset\nds = load_dataset(\"materials-toolkits/oqmd\")"},{"language":"python-hub","snippet":"from huggingface_hub import snapshot_download\nsnapshot_download(repo_id=\"materials-toolkits/oqmd\", repo_type=\"dataset\")"}],"caveats":["HuggingFace metadata: author=materials-toolkits, license:cc-by-4.0, size_categories:1M<n<10M, format:parquet, downloads=305"],"format":"mixed","generatedAt":"2026-11-16T02:38:36.162Z","layout":{"example":"all.hdf5","pattern":"{repoId}/{filename}"},"mountPath":"/mnt/archive","sampleHeaders":[{"bytes":"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","encoding":"hex","key":"all.hdf5","truncatedAt":512}],"sampleObjects":["all.hdf5","all_index.parquet","dump.py","oqmd.tar.gz","unique.hdf5"],"sidecars":[{"bytes":"---\nlicense: cc-by-4.0\npretty_name: OQMD\ntags:\n- chemistry\n- crystal\n- material\nsize_categories:\n- 1M<n<10M\n---\n\n# OQMD\n\nStructures and energies of the [Open Quantum Materials Database](https://oqmd.org),\npreprocessed for machine learning.\n\n| File | Structures | Atoms | Content |\n|---|---|---|---|\n| `all.hdf5` | 1,407,269 | 14,266,404 | one structure per entry of OQMD v1.8 |\n| `unique.hdf5` | 1,259,553 | 13,089,068 | same, without the entries marked as duplicates in OQMD |\n| `all_index.parquet`, `unique_index.parquet` | | | OQMD `entry_id`, `calculation_id` and `structure_id` of each row of the file above |\n| `oqmd.tar.gz`, `dump.py` | 1,021,320 | 7,117,087 | older export (2024) in another layout, kept unchanged, see the end of this page |\n\nStructures have 1 to 312 atoms (median 6).\n\n## How structures are selected\n\nSource: the SQL dump of OQMD v1.8 (`qmdb__v1_8__022026.sql.gz`, February 2026), 1,407,285 entries.\n\nFor each entry, one calculation is selected:\n\n- its output structure must be consistent: the number of atoms declared for the structure\n  equals its number of rows in the `atoms` table (47,859 of 8,374,305 structures are not);\n- the calculation is converged, its label starts with `static` or `standard`, and does not\n  end with `lda`;\n- among those, the one with the lowest energy per atom is kept (lowest calculation id when\n  equal).\n\n1,407,269 entries have such a calculation. Nothing else is filtered in `all.hdf5`.\n\n`unique.hdf5` groups each entry with the entry it duplicates according to OQMD\n(`entries.duplicate_of_id`) and keeps the entry with the lowest energy per atom of each group;\nwhen energies are exactly equal the original entry is kept. 147,716 entries are removed this\nway. Duplicates that OQMD does not mark are not removed.\n\n## Labels\n\n| Field | Source | Meaning |\n|---|---|---|\n| `total_energy` | `calculations.energy_pa` × number of atoms | DFT total energy of the cell in eV |\n| `formation_energy_per_atom` | `formation_energies.delta_e` (fit `standard`) | formation energy in eV/atom |\n| `energy_above_hull` | `formation_energies.stability` (fit `standard`) | distance to the convex hull in eV/atom, **stored as in OQMD: negative for phases below the hull** |\n| `band_gap` | `calculations.band_gap` | band gap in eV, NaN for 3,155 structures of `all.hdf5` |\n\nLabels are not cleaned. `all.hdf5` contains unphysical values from failed calculations: 559\nstructures with a positive total energy (up to +704 eV/atom), 668 with a stability above\n5 eV/atom, and one negative band gap. Filter them for training.\n\n| | `all.hdf5` | `unique.hdf5` |\n|---|---|---|\n| stability ≤ 0 | 97,228 | 96,246 |\n| stability ≤ 0.1 eV/atom | 458,906 | 402,102 |\n\n## Format\n\nFlat concatenated arrays:\n\n| Dataset | Shape | Type | Content |\n|---|---|---|---|\n| `lattice` | (n_struct, 3, 3) | float32 | lattice vectors as rows, in angstrom |\n| `num_atoms` | (n_struct,) | int64 | number of atoms of each structure |\n| `ptr` | (n_struct,) | int64 | index of the first atom of each structure in `x` and `z` |\n| `x` | (n_atoms, 3) | float32 | fractional coordinates in [0, 1) |\n| `z` | (n_atoms,) | int64 | atomic numbers |\n| `total_energy`, `formation_energy_per_atom`, `energy_above_hull`, `band_gap` | (n_struct,) | float32 | labels, NaN when unknown |\n\nThe lattice is rebuilt from cell lengths and angles, so it is in the standard orientation\n(first vector along x). Fractional coordinates are wrapped into the cell. The `batch` array is\nan artefact of the writer and is not used when reading. Structures are ordered by OQMD entry\nid; row *i* of an index file describes structure *i*.\n\n## Usage\n\n```python\nfrom pycrystalgen.data import HDF5Dataset\n\ndataset = HDF5Dataset.from_hub(\"materials-toolkits/oqmd\", \"unique.hdf5\")\nstructures = dataset[:32]  # batched Structures\n```\n\n## Processing and checks\n\nBuilt with pycrystalgen (`scripts/build_oqmd.py`): the dump is read as a stream without a MySQL\nserver, the tables are extracted, structures are written to CIF then preprocessed to HDF5.\n\nEvery structure of both files was compared with the tables of the dump: number","encoding":"utf8","key":"README.md","name":"README.md","size":5062,"truncated":true}],"sourceObjectKey":"all.hdf5","starterCell":"from huggingface_hub import hf_hub_download\npath = hf_hub_download(repo_id=\"materials-toolkits/oqmd\", filename=\"all.hdf5\", repo_type=\"dataset\")\nprint(path)","starterSamplePath":"all.hdf5","totalObjects":9},"summary":"Structures and energies from the Open Quantum Materials Database (OQMD, Wolverton group, Northwestern University), extracted from the v1.8 SQL dump of February 2026 and…","title":"OQMD v1.8: 1.4 Million DFT Crystal Structures and Formation Energies (HDF5)","updatedAt":"2026-11-16T17:16:48.706Z","visibility":"public","websiteUrl":"https://oqmd.org"}}