Datasets:

Scandiumlabs
/

Scandium-Dataset

Transparency note: Every known limitation of this dataset is documented below. We believe honest documentation is the foundation of trust. No issue has been hidden.

Critical Issues (Affect Specific Tiers)

1. Dataset Does Not Include Ionic Transport Properties (By Design, v1.0.0 Scope)

Status: By design, v1.0.0 scope — flagged prominently for all users
Scope: Entire dataset (100%)
Impact: The dataset supports thermodynamic and structural screening only. It cannot be used to rank materials by ionic conductivity or SSE performance without augmentation (see ROADMAP.md for planned transport-property layers).
Mitigation: Use for upstream screening (stability, band gap, structural family), then pair with AIMD/NEB or experimental conductivity data for downstream SSE evaluation. The BVSE migration barrier proxy covers 23% of Li/Na entries at draft quality — use with caution, and consult the 74.8% skip rate documentation before filtering.

2. JARVIS Energy Above Hull (EaH) — ✅ Resolved

Status: ✅ Computed via internal convex hull (v1.0.0-rc2)
Scope: 25,673 JARVIS entries (100%)
Coverage: 96.5% of all entries now have EaH; 14,669 JARVIS entries got new EaH values via internal hull
Unresolved: 8,933 JARVIS entries lack EaH because their chemical system has <3 entries (too few for a convex hull). This is a fundamental sparsity limitation that cannot be resolved without cross-method normalization.
Note: This uses JARVIS-relative hull, not the MP/OQMD reference. Cross-method normalization remains future work.

3. Extreme Formation Energy Outliers (FE > 5 eV/atom)

Status: By design, tier-isolated
Scope: 42 OQMD entries (max = 45.17 eV/atom)
Impact: Zero entries in Gold tier
Root cause: OQMD multi-element compounds with unphysical stoichiometries (e.g., Be₄H₁₆K₈O₃₈S₈, Ba₆Mg₁₂Pt₆). These are likely computational artifacts from high-throughput enumeration.
Resolution: Removed from Gold tier by gate G10 (quality score ≥ 80). Documented for raw/validated tiers.

4. Extreme Energy Above Hull Outliers (EaH > 5 eV/atom)

Status: By design, tier-isolated
Scope: 87 entries (44 OQMD + 43 MP, max = 47.03 eV/atom)
Impact: Zero entries in Gold tier
Root cause: Highly metastable phases computed far from convex hull. These are legitimate DFT results for unphysical structures.
Resolution: Removed from Gold tier by gate G10.

High-Impact Issues

5. BVSE Coverage Gap (~75% of Li/Na Entries Skipped)

Status: Documented, structural limitation
Scope: 73,900 of 98,773 attempted Li/Na entries (74.8%)
Impact: Only 1 in 4 Li/Na entries with crystal structures receives a BVSE migration barrier estimate
Root cause: bvlain's bond-valence parameter table only covers a limited set of element/cation combinations. Entries with unparameterized compositions (e.g., uncommon transition metals, complex anion chemistries) cannot produce a barrier — this is a fundamental limitation of the empirical bond-valence method, not a sampling gap.
Manifestation: Entries have bvse_skip_reason: "no BV params for composition" in their ssb_screening block.
Mitigation: Filter with has_bvse_barrier: true to work exclusively with the 24,873 entries that have computed barriers. The 1,501 superionic + 4,705 good entries within this subset are valid candidates for further screening.
Additional exclusions: 8,744 entries excluded by --max-sites 60 (unit cells >60 atoms) and 498 experimental OBELiX entries lacking crystal structures. Combined, these reduce the effective BVSE coverage to 23.0% of all 108,015 Li/Na entries.

6. BVSE Sanity Validation — 2 Known Marginal Outliers

Status: Documented, pre-production
Scope: 5/7 known SSE structures pass within literature ranges; 2 fail marginally
Details:
- Li7P3S11 (mp-641703) → 0.420 eV (lit. 0.18–0.30) — bvlain systematically overestimates barriers for this structure
- Li6PS5Cl (JVASP-38005) → 0.176 eV (lit. 0.24–0.38) — JVASP-38005 is a microporous variant with artificially low barrier
Impact: These are known, characterized limitations of the bvlain parameter set for specific structure types. All 5/7 non-outlier structures reproduce within ±0.05 eV of literature.

7. Single-Atom Entries (38 entries)

Scope: 38 entries with nsites = 1
Impact: Minimal (0.01% of dataset)
Detail: Primarily OQMD single-atom calculations. Included but flagged.
Recommendation: Filter using nsites > 1 if single atoms are inappropriate for your task.

8. MP vs OQMD Quality Score Bias

Scope: All entries
Impact: MP entries have mean score 87.4, OQMD have 73.2
Root cause: MP uses finer computational parameters (ENCUT, KSPACING, semi-core pseudopotentials). This is a genuine quality difference, not a systematic bias.
Mitigation: Score calibration is monotonic and consistent per-source. Use tier-based filtering to isolate quality levels.

9. Quality Score × SSE Chemistry Interaction

Scope: SSE-relevant families (halide, sulfide)
Detail: SSE-relevant entries skew toward higher-quality sources. MP provides 24.2% of its entries as halide/sulfide (mean quality 87.4), while OQMD provides only 7.1% (mean quality 73.2). A quality-score threshold applied uniformly does not disproportionately filter SSE chemistries — the interaction runs protectively. However, within-source analyses should still account for source effects, and the interaction is documented here so downstream users are not surprised when SSE-family entries consistently rank higher on quality than intermetallics.
Recommendation: If filtering by source or quality threshold, verify that SSE-relevant coverage remains sufficient for your task. See reports/source_reports/source_reports.json for per-source family breakdowns.

10. Missing Structured Formula (Resolved)

Status: ✅ Fixed in v0.0.0
Scope: Was 266,732 entries
Resolution: All entries now have structured_formula via formula reduction.

11. Stale Quality Flags (Resolved)

Status: ✅ Fixed in v0.0.0
Scope: Was 171,764 OQMD entries
Resolution: missing_spacegroup flag removed from all OQMD entries that now have space_group.

Medium-Impact Issues

12. Conservative Quality Scoring

Scope: Maximum score = 88 (no entry ≥ 90)
Impact: The scoring system is intentionally conservative. Any missing sub-score information caps the total.
Recommendation: Use scores for relative ranking, not absolute quality assessment.

13. Family Imbalance

Scope: Entire dataset
Detail: ~63% intermetallics, ~16% layered oxides, ~7% halide SSEs, ~6% sulfide SSEs
Impact: Models trained on full dataset may be biased toward intermetallics.
Mitigation: Use stratified sampling by family, or use battery/electrolyte subsets. Garnet enrichment script (scripts/enrich_garnet_family.py) available for structure-based reclassification.

14. Missing Band Gap (151 OQMD entries)

Scope: 0.1% of OQMD entries
Root cause: OQMD band gap calculation did not converge for these entries.
Impact: Minimal; entries remain usable with band_gap: null.

15. Space Group Determination Failures (16 OQMD entries)

Scope: 0.01% of OQMD entries
Root cause: Structures with too few atoms or disordered configurations where spglib cannot determine symmetry.
Impact: Minimal; entries remain usable with space_group: null.

16. Cross-Source Formula Duplicates (195 formulas)

Scope: 195 formulas appear in all three sources
Impact: Deduplication preserved best entry per group; all resolutions logged.

17. Family Held-Out Benchmark Split Is Small

Scope: Family held-out test split
Detail: 583 test entries (0.2% of corpus) in the cross-family split, compared to 39,940 (random) and 12,675 (chemistry held-out). This is a consequence of the family imbalance — rare families necessarily produce small test sets under held-out sampling.
Impact: Benchmark numbers quoted against this split will have wide error bars. Suitable as a smoke test for cross-family generalization but not sufficient to support strong SOTA claims.
Recommendation: Supplement with per-family stratified evaluation or bootstrapped confidence intervals when reporting cross-family performance.

Future-Priority Issues

Raw source downloads not included — due to source redistribution policies, raw download scripts must be run by the user
Garnet undercount — 136 entries (improved from 23 via broader composition heuristic + 95 OBELiX experimental garnets, but still far below the ~1,000+ needed for ML training on LLZO and variants). Only 19 garnets are in Gold tier. The enrichment script (scripts/enrich_garnet_family.py) identifies ~200 new candidates from structure-based verification, but these need manual validation before reclassification. True garnet-family expansion requires targeted acquisition from ICSD and MP for LLZO-type structures.
SSE family tags are composition-based — sse_family is classified by elemental heuristic, not crystal structure. A compound with Li+S+P but no percolating channels will still be tagged as argyrodite/LGPS. Cross-check with CAVD dimensionality when available.
Experimental data coverage is thin — OBELiX integrated (498 experimental_gold entries with measured conductivity), but this is a pilot-level integration. Full experimental coverage requires ICSD, Inorganic Crystal Structure Database, and additional curated experimental conductivity data.
No DOI assigned yet — Zenodo archival in progress for v1.0.0
License complexity — the dataset has 3 different licenses; the license field per entry was added in v0.1.0 to enable programmatic filtering. A Commercial-Safe edition (MP+JARVIS only, ~95k entries) is now extractable via scripts/extract_commercial_safe_edition.py.
CAVD channel dimensionality is 0% coverage — the algorithm exists in scripts/compute_cavd_channel_dimensionality.py but was never re-run against the Parquet store. When computed, the field uses Voronoi-based connectivity analysis (a geometric proxy), not full NEB migration barriers. 0D/1D classifications reliably rule out good conductors, but 2D/3D classifications overestimate true percolation. Use as pre-filter only.
Mechanical properties are geometric proxies — the bulk_modulus_GPa and shear_modulus_GPa fields use density-based estimation, not DFT elastic tensors. They provide order-of-magnitude estimates only. True DFT elastic data for MP entries can be queried via scripts/compute_mechanical_properties.py --api-key YOUR_KEY.
SSE candidate scores are uncalibrated — the 5-gate sse_candidate_score is a heuristic composite (30+25+20+15+10 pts). Gates 1-2 are fully populated; gates 3-5 are populated only when underlying data exists. Scores from entries with missing gates are underestimates.

Reporting New Issues

If you discover an issue not listed here, please open a GitHub Issue with:

Affected entry (source_id or formula)
Description of the issue
Expected vs observed value
Any relevant context (pipeline version, date)