GoldDIGR — DFT Reaction Calculation Dataset

A large open dataset of ~1,000,000 xTB/DFT reaction calculations mined from the supporting information of published chemistry papers and recomputed with a uniform pipeline (geometry/TS search, IRC, single-point DFT, and an xTB spin/structure scan). Each calculation is a self-contained .zip; the corpus is distributed as 425 .tar bundles (~800 GB total).

Status: continuously updated — new DFT results are injected over time and the affected bundles are re-uploaded (see Updates below).

Quick start

Download and unpack (all bundles extract into one unified tree):

# one bundle
huggingface-cli download <user>/GoldDIGR bundles/bundle_0000.tar --repo-type dataset --local-dir .
tar -xf bundles/bundle_0000.tar          # -> 10.<prefix>/<doi>/<stem>/<reaction>.zip

# everything
huggingface-cli download <user>/GoldDIGR --repo-type dataset --local-dir .
for t in bundles/*.tar; do tar -xf "$t"; done

Verify integrity:

cd bundles && sha256sum -c SHA256SUMS

Layout

bundles/
  bundle_0000.tar … bundle_0424.tar     # the data (≤2500 zips each, ~62 MB–4.3 GB)
  bundle_XXXX.tar.sha256                 # per-bundle checksum
  SHA256SUMS                             # all checksums combined
bundle_manifest.tsv                      # key <TAB> bundle_id  (every zip -> its bundle)
bundle_lists/bundle_XXXX.paths           # the zip list inside each bundle

Inside each bundle, zips are stored at their DOI path:

<DOI_prefix>/<DOI_suffix>/<SI_stem>/<reaction>_<charge>_<multiplicity>.zip
e.g.  10.1039/C8SC02758G/c8sc02758g2/03_1_1.zip

Inside a reaction .zip

<reaction>/
  DFT-SinglePoint/        ORCA single-point outputs per stationary point
                          (.out, .xyz, .inp, CHELPG charges, fuzzy/Mayer/Wiberg
                          bond-order matrices), and status.json
  xTB-scan/               GFN2-xTB spin/structure scan: per-frame geometries,
                          spin_scan_summary.csv (energies per spin/frame),
                          bem_snapshots/step_*.json (energy + charges + bond orders)
  IRC_Analysis/           IRC frame analysis (json + csv) and reaction summaries
  *.trj / *.xyz / *.yaml  TS-opt / IRC trajectories and metadata

status.json (in DFT-SinglePoint/) records which stationary points (input, finished_first, finished_last, *_opt, ts_final_geometry) completed.

How it was processed

The raw outputs were cleaned losslessly/redundancy-only before release: ORCA/xTB log banners and citation/grid boilerplate trimmed; uncompressed members recompressed; intermediate scratch removed (xTB restart files, crashed-run Hessian .h5); and per-step xTB .out files removed where their energy is already preserved in spin_scan_summary.csv (energy-verified, never lossy). All scientific quantities (geometries, energies, charges, bond orders, orbital energies) are retained.

Updates

The dataset grows as new DFT results are computed. Updates are incremental: only the bundles containing changed/added reactions are rebuilt and re-uploaded (tracked via bundle_manifest.tsv; the Hub's Xet backend dedups unchanged data across versions). Check the commit history for the latest additions.

Citation

@misc{golddigr,
  title  = {GoldDIGR: A DFT Reaction Calculation Dataset},
  author = {<authors>},
  year   = {<year>},
  howpublished = {Hugging Face Hub, <user>/GoldDIGR},
  doi    = {<Zenodo DOI if minted>}
}

License

Released under <CC-BY-4.0 / your choice>. Underlying structures derive from published SI; please also cite the original papers (DOIs are encoded in each reaction's path).