dtsdts/exciting-multifidelity-dft-deltas

exciting multi-fidelity DFT deltas

Version: 30 October 2026.

This dataset contains all-electron density-functional calculations performed with exciting, with numerical settings varied through the APWprecision parameter and k-point density. It provides energies, band gaps, DOS comparison measures, available relaxed geometries and calculation provenance. It is a companion to the FHI-aims multi-fidelity dataset, with a similar tabular layout and Hugging Face loading interface.

The archive has 3,920 calculation records: 3,885 historical NHR@GWDG records and 35 HU@HPC records. It combines the existing 3,899-row table with 21 newly ingested records; 12 already-present HU labels were checked without duplication. HU timings are retained in their own timing domain and have not been converted into equivalent NHR CPU times.

Configurations

Configuration	Split	Rows	Use
all_calculations (default)	data	3,920	Full calculation archive, including reference settings and available sensitivity data
locked_density8	test	335	Observed labels for the frozen density8 basis-selection benchmark
locked_slate	test	342	All 57 materials × six APW settings, with explicit missing-label masks
hu_supplement	data	35	HU records with their separate provenance and timing fields

The locked slate uses APW settings 0.4, 0.5, 0.6, 0.7, 0.8 and 0.9, RMT scaling 0.95 and the original density8 meshes. It has 335/342 observed energy outcomes, 50/57 APW0.7 labels and 50 materials with all six settings. All other APW settings in the slate are complete. Seven APW0.7 labels remain missing; they are null in locked_slate, never imputed or treated as zero.

The density labels 2, 4 and 8 are the historical mesh-density convention, not universal 2×2×2, 4×4×4 and 8×8×8 grids. Actual meshes vary with the unit cell; where retained, ngridk gives the mesh. Density2/4 rows are a separate sensitivity study and do not fill density8 benchmark cells.

Loading

from datasets import load_dataset

archive = load_dataset("dtsdts/exciting-multifidelity-dft-deltas")
calculations = archive["data"]

locked = load_dataset(
    "dtsdts/exciting-multifidelity-dft-deltas",
    "locked_density8",
    split="test",
)
slate = load_dataset(
    "dtsdts/exciting-multifidelity-dft-deltas",
    "locked_slate",
    split="test",
)
assert len(locked) == 335
assert sum(slate["energy_observed"]) == 335

Direct pandas access is also available:

import pandas as pd

url = "https://huggingface.co/datasets/dtsdts/exciting-multifidelity-dft-deltas/resolve/main/data.parquet"
df = pd.read_parquet(url)

data.csv.gz is a CSV mirror of the full archive. Structure arrays and nested provenance objects are JSON-encoded strings; use json.loads where needed. Numeric scalars and missing values are typed in Parquet. No executable dataset loader or pickled graph object is required. This release does not invent graph features or starting geometries that are absent from the retained tables.

Settings, structures and references

The locked HU calculations retain the original species files, initial structures, PBE settings, atomic-position relaxation with a fixed cell, original meshes and numerical tolerances. Archived nominal APW1/density8 calculations provide matched references; APW1 is not asserted to be an exact complete-basis limit. Anchor calculations on HU check compatibility with the archived references before the corresponding lower-precision target is accepted.

formula, compound_name, ICSD_number, atomic numbers and available relaxed lattice/Cartesian coordinates describe the material. An ICSD identifier alone should not replace the original structure, species and protocol hashes when matching calculations. Missing structural/provenance fields remain null.

Targets and sign conventions

Signed delta columns use calculation minus matched reference:

Column	Meaning	Unit
delta_total_energy_per_atom	Signed total-energy delta	eV/atom
absolute_energy_error_eV_per_atom	Absolute value of the signed energy delta, used by the energy recommender	eV/atom
delta_band_gap_scf_eV	Signed gap delta from the SCF occupation/eigenvalue rule	eV
absolute_band_gap_error_eV	Absolute SCF gap delta	eV
dos_tanimoto	Historical DOS fingerprint similarity	dimensionless
dos_tanimoto_error	1 - dos_tanimoto	dimensionless
dos_wasserstein1_eV	Historical DOS W1 target, including the retained double-Fermi-subtraction convention	eV
dos_wasserstein1_fermi_relative_eV	Separately tagged corrected Fermi-relative DOS W1 where available	eV

The two Wasserstein columns are different protocols and must not be silently pooled. The historical fingerprint uses dg_cut:56:-2:7:(-10, 5) on the raw DOS. The W1 comparison interpolates and normalizes on −10 to 5 eV at 0.05 eV spacing. The corrected W1 column uses the already Fermi-relative DOS axes without a second absolute-Fermi shift. Protocol/status columns retain the distinction.

The historical single-spin SCF gap rule classifies occupation >0.5 as occupied, then uses max(0, minimum empty eigenvalue - maximum occupied eigenvalue). Band-structure-derived gaps are retained separately where available.

Legacy total_energy and energy components retain their eV conventions; columns explicitly ending in Ha_cell retain Hartree per cell. Relaxed lattice vectors and Cartesian positions are in Å, angles in degrees and volumes in ų, except unit_cell_volume_Bohr3. column_schema.json describes the exported types and units. No numerical target has been recomputed to manufacture missing labels.

Splits and use in model evaluation

The default data split is an archive, not a recommended random train/test partition. source_split preserves historical labels; split, where present, also retains its source value. These historical labels are not advertised as a new leakage-safe split of all 3,920 rows.

frozen_partition identifies train, tune, calibration or locked-test membership where the existing frozen artifact maps the material group. outside_frozen_artifact means that no such mapping is asserted. All rows for an identified locked material, including different APW values or density2/4 sensitivity calculations, must be kept out of training/tuning/calibration when reproducing the paper evaluation. The authoritative group lists are in provenance/frozen_group_splits.json.

locked_density8 is evaluation-only. Its 335 unique group/APW cells reproduce the original 310 observed outcomes plus 25 validated HU additions. The 310 historical rows were joined to their original material metadata only after unique matching of precision/density/RMT and all four frozen targets within 1e−12; group/ICSD identity was then checked for consistency. No predictions, fitting or calibration were changed by this export.

locked_slate contains the saved 100-component composition features plus the three numerical settings as explicit columns, group IDs and observed masks. It is useful for reproducing candidate coverage and missing-label analyses. It does not contain predictions fitted to the new test labels.

Provenance and quality checks

The release includes source-file hashes, row-source hashes, input/reference/species hashes where available, accepted-job/proof identifiers, a completed-batch index and the exact remaining-cell dispositions. All 12 jobs in the latest batch (nine APW0.7 targets plus three anchors) completed with scheduler exit 0:0 and passed maintained scientific, allocation, reference and stageout proofs.

One HU row is a previously approved, source-specific legacy adoption: ZrO₂ ICSD66782, APW0.5, source job1631615. Its application finished normally, but the original wrapper is recorded as scheduler FAILED 1:0. That native status is preserved explicitly; the row is not represented as a new successful Slurm job. Its existing v22 adoption and exact original reference comparison were reverified, with independent checks of energy, gap and DOS metrics. Its unproven elapsed timing is null. This exception does not certify other failed jobs.

Historical rows have less detailed runtime evidence than recent HU rows. Null proof/timing fields are not evidence of failure or success. status="ok" in the archive describes label/parser acceptance; inspect the explicit native scheduler and adoption fields for the legacy case.

Private storage paths, raw module/log text and path-bearing error diagnostics were removed; hashes retain source linkage. Credentials, socket paths and private keys are excluded. The release is checksum-bound by manifest.json and audited in audit.json. Source code and discussion: errorbar_modelling PR262, paper status issue218, and remaining-gap issue258.

Limitations

• Seven locked APW0.7 labels still need specific accounting, recovery or legacy validation work; no complete 342-outcome claim is made.
• The locked cohort is retrospective/post-development, not a claim of a pristine prospective benchmark. Complete-case material subsets can be nonrepresentative.
• Basis error and k-point sensitivity are distinct effects. Nominal reference convergence and matching do not establish an exact physical ground truth.
• No common CPU-time scale, controlled speedup, joint energy/gap/DOS safety guarantee or independent new uncertainty-calibration result is supplied here.
• Rows in different configurations overlap by design. Do not concatenate the full archive, HU subset and locked subset as independent observations.

Files

• data.parquet, data.csv.gz: full archive.
• locked_density8.parquet: 335 unique observed locked cells.
• locked_slate.parquet: 342 candidates with masks and saved composition features.
• hu_supplement.parquet: 35 HU records, a subset of the archive.
• column_schema.json, audit.json, manifest.json: schema, audit and checksums.
• provenance/: coverage, frozen group partitions, latest completed jobs and remaining-cell dispositions.

Citation and license

@misc{speckhard2026excitingdeltas,
  author = {Speckhard, Daniel},
  title = {exciting multi-fidelity DFT deltas and locked basis-selection benchmark},
  year = {2026},
  url = {https://huggingface.co/datasets/dtsdts/exciting-multifidelity-dft-deltas},
  note = {Dataset release, 30 October 2026}
}

The derived dataset is released under CC BY 4.0, matching the companion dataset's license. Please acknowledge the exciting code and the underlying structure sources when using the data. The associated paper is in preparation; no journal publication or DOI is implied by this dataset citation.