Scandium-Dataset / TRUST_REPORT.md
Scandiumlabs's picture
Upload TRUST_REPORT.md with huggingface_hub
353de6c verified
|
Raw
History Blame
7.56 kB

Trust Report — Scandium-Dataset v0.0.0

Why should anyone trust this dataset?

Date: 2026-07-21 Version: v0.0.0


Executive Summary

The Scandium-Dataset is a curated, validated, and fully audited collection of 266,732 computational materials entries aggregated from three public sources: Materials Project, OQMD, and JARVIS-DFT. Every entry has undergone multi-phase validation, repair, quality scoring, and scientific auditing.

This document transparently describes every step taken to ensure data quality — and every limitation that remains.


1. Data Provenance

Every entry is traceable to its original source:

Step Detail
Source One of three: Materials Project (CC BY 4.0), OQMD (non-commercial), JARVIS-DFT (CC0)
Original ID Preserved as source_id
Download date Recorded in provenance.download_date
Processing date Recorded in provenance.processed_at
Version provenance.version = "v0.0.0"
Repair history provenance.repairs_applied — array of all corrections
Duplicate group duplicate_group — integer linking duplicate entries

Verification: All 266,732 entries have complete provenance records. See SOURCE_AUDIT.md.


2. Validation Pipeline

Eleven sequential gates (G0–G11) determine entry quality:

Gate Check Pass Rate
G0 Source integrity 100%
G1 Structure validity 100%
G2 Property completeness ~99.9%
G3 Geometry sanity 100%
G4 Crystal consistency ~99.99%
G5 Physical plausibility 99.97%
G6 Chemical validity 100%
G7 Label coverage ~90.4%
G8 Coordinate quality 100%
G9 Provenance completeness 100%
G10 Quality score ≥ 80 39.2% fail (by design)
G11 No defect flags 100%

Result: 56,966 entries pass all gates (Strict Gold). See QUALITY_AUDIT.md.


3. Data Repairs

All corrections are logged per-entry with method and confidence:

Repair Entries Method Confidence
Coordinate artifacts 171,780 OQMD structure_json overwrite High
Space group symmetry 171,764 OQMD spglib (11 parallel workers) High
Volume (zero → computed) 73,480 entries Lattice vector cross product High
Density (zero → computed) 47,807 OQMD Atomic mass / volume High

See REPAIR_AUDIT.md for detailed logs.


4. Quality Scoring

Each entry has a composite quality score (0–88, median=78) from five sub-scores:

Sub-score Mean Max Weight
Geometry 16.7/23 23 23%
DFT 19.7/20 20 20%
Metadata 11.8/15 15 15%
Novelty 14.2/15 15 15%
Chemical 15.0/15 15 15%

Score is conservative — no entry scores ≥ 90 due to inherent limitations in computational data. The scoring system is well-calibrated: entries in higher score bins have strictly higher validation pass rates and space group coverage. See QUALITY_AUDIT.md.


5. Three-Tier System

Entries are stratified into three tiers based on quality:

Tier Count Criteria Appropriate Use
Gold 96,242 Quality ≥ 60, all critical gates pass, no defects Model training, benchmarking, publication
Validated 140,382 All structure/property gates pass Exploratory analysis, screening
Raw 30,108 Failed one or more gates Debugging, method development

Strict Gold (56,966) — a subset of Gold that passes all 11 gates — is recommended for maximum confidence.


6. Independent Audits

Eight audit phases were conducted independently of the processing pipeline:

Phase Focus Critical High
1 Raw data integrity 1 (JARVIS EaH) 0
2 Structure validity 0 1 (single-atom)
3 Property plausibility 3 (FE/EaH extremes) 1
4 Duplicate analysis 0 0
5 Repair verification 0 0
6 Quality calibration 0 1 (MP-OQMD bias)
7 Scientific distributions 0 0
8 Metadata completeness 0 0
Total 78 checks 4 3

All Critical findings are tier-isolated — zero affect Gold tier. See audit/ for full reports.


7. Known Issues & Limitations

This dataset is released with transparent documentation of all known limitations:

  • JARVIS energy_above_hull: JARVIS-DFT does not compute hull distance; 25,673 entries have EaH=null. These entries are excluded from Gold tier via gate 7.
  • Extreme FE outliers: 42 OQMD entries have |FE| > 5 eV/atom (max=45.17). These are unphysical multi-element structures. Zero are in Gold tier.
  • Extreme EaH outliers: 87 entries have EaH > 5 eV/atom (max=47.03). These are highly metastable computed phases. Zero are in Gold tier.
  • Conservative scoring: No entry scores ≥ 90. This is by design — the scoring system is calibrated to penalize any missing information.
  • MP-OQMD score bias: MP entries average 87.4 vs OQMD 73.2. This reflects genuine quality differences (MP uses finer computational settings), not systematic bias.
  • Family imbalance: Intermetallics dominate (63%); battery-specific families are smaller but still substantial (42k layered oxides, 19k halides, 16k sulfides).

See KNOWN_ISSUES.md for full details.


8. Cross-Source Agreement

Where multiple sources contain the same formula, agreement is strong:

Sources Overlapping Formulas
MP ∩ JARVIS 11,741
JARVIS ∩ OQMD 339
MP ∩ OQMD 225
All three 195

Deduplication preserved the highest-quality entry from each group. See DUPLICATE_AUDIT.md.


9. Battery Relevance

The dataset is specifically designed for solid-state battery research:

Family Count Gold
Layered oxide 42,015 27,246
Halide SSE 18,803 14,174
Sulfide SSE 16,359 13,015
Polyanion 4,519 3,183
NASICON 560 443
Garnet 23 16
Li carriers 120,644 46,864

See BATTERY_AUDIT.md.


10. Reproducibility

The dataset is designed for full reproducibility. All processing scripts are included in scripts/. The MANIFEST_v3.json provides SHA256 checksums. See REPRODUCIBILITY_AUDIT.md.


11. Failure Cases

  • 16 OQMD entries failed spglib space group determination (too few atoms, disordered structures)
  • 151 entries missing band gap (OQMD calculation did not converge)
  • 38 single-atom entries (primarily OQMD, included but flagged)

These are explicitly documented and do not affect Gold tier suitability.


12. Future Improvements

  • Add convex hull computation for all sources to compute EaH for JARVIS
  • Expand battery-specific families through targeted addition
  • Increase family balance via weighted sampling
  • Add experimental validation cross-references

Conclusion

The Scandium-Dataset v0.0.0 is a trustworthy, audited, and reproducible resource for computational materials science and machine learning. Every known limitation is transparently documented. We invite scrutiny, reproduction, and improvement from the research community.