Scandium-Dataset / DATASET_CARD.md
shamique's picture
Upload DATASET_CARD.md with huggingface_hub
8e83be6 verified
|
Raw
History Blame Contribute Delete
12 kB
metadata
license: other
license_details: >-
  Per-entry licenses vary — see LICENSE_BREAKDOWN.md. MP entries are CC BY 4.0,
  JARVIS entries are CC0, OQMD entries are non-commercial only.
language:
  - en
size_categories:
  - 100K<n<1M
task_categories:
  - other
tags:
  - dft
  - materials-project
  - oqmd
  - jarvis
  - materials-screening
  - dft-stability
  - parquet
  - materials-science
  - battery-materials
  - inorganic-materials
pretty_name: Scandium-Dataset
configs:
  - config_name: default
    data_files:
      - split: train
        path: dataset/entries_v3.parquet

Dataset Card — Scandium-Dataset v1.0.0

Summary

Scandium-Dataset provides a harmonized, quality-scored foundation of DFT-computed structural and thermodynamic properties across 267,230 materials from Materials Project, OQMD, and JARVIS-DFT. It supports the early screening stage of battery materials discovery — filtering by phase stability, electronic structure, and structural family — before downstream property prediction (ionic conductivity, mechanical stability, electrochemical window) via AIMD, NEB, or targeted DFT.

Previous versions positioned this as a solid-state battery discovery dataset. v1.0.0 corrects that framing: the data covers thermodynamic screening, not ionic transport. See CHANGELOG for details.

  • Version: v1.0.0

  • Total entries: 267,230 (Experimental: 498; Raw: 30,108; Validated: 140,382; Gold: 96,242)

  • Strict Gold: 56,966 entries (subset of Gold; the three exclusive tiers sum to 266,732)

  • Tiers are hierarchical (nested): Strict Gold ⊆ Gold. Validated, Gold, and Raw are exclusive tiers that sum to 266,732. Exclusive breakdown: Raw = 30,108 | Validated = 140,382 | Gold (non-Strict) = 39,276 | Strict Gold = 56,966

  • Battery-relevant subset: 82,925 entries | Electrolyte candidate subset: 41,665 entries (strict Gold)

    File names on disk (battery_candidate_subset_v1.json, solid_electrolyte_candidate_subset_v1.json) — the v3 suffix was retired in v1.0.0-rc2 when the scope was clarified.

  • Storage: Parquet (dataset/entries_v3.parquet, 47 columns) with indexed lookup — 184 MB instead of 1.6 GB JSON.

  • Experimental data: 498 OBELiX entries integrated (Therrien et al. 2025) as experimental_gold tier with measured Li-ion conductivity.

  • Transport proxy draft: 24,873 BVSE migration barriers computed (bvlain engine) at 23% Li/Na coverage — see "Roadmap" below for the gap this addresses.

Included vs. Not Included

Included Not Included (requires augmentation)
Formation energy (eV/atom) — 100% Ionic conductivity (S/cm)
Energy above hull (eV/atom) — 90.4% Migration/activation energy (NEB)
Band gap (eV) — 99.9% Electrochemical stability window
Space group, volume, density — 99.9%+ Elastic/shear moduli (full DFT)
Crystal structure (pymatgen Structure JSON) — 100% Experimental validation beyond pilot OBELiX integration
Quality tier (Gold/Validated/Raw) — 100%
Provenance tracking (source, source_id, checksum) — 99.8%
SSE family classification (composition-based) — 100%
BVSE migration barrier proxy — 23% of Li/Na entries

Suitable For / Not Suitable For

Suitable for: phase stability screening, structural family classification, band gap prediction, materials-informatics benchmarking, pretraining general property predictors on inorganic crystal structures, cross-source DFT property harmonization studies.

Not suitable for (without augmentation): direct ionic conductivity prediction, SSE performance ranking, electrochemical stability assessment. The dataset contains migration barrier proxies for 23% of Li/Na entries, but at ~75% skip rate due to bond-valence parameter coverage, these do not constitute a complete transport-property layer.

What the Data Actually Supports

Property coverage (verified against Parquet store)

Property Coverage Notes
Formation energy 267,230 (100%) From MP, OQMD, JARVIS-DFT
Energy above hull 241,557 (90.4%) 0% for JARVIS (25,673 entries); JARVIS EaH script has not been run against the Parquet store
Band gap 267,079 (99.9%) 151 OQMD entries with non-converged band gap
Space group 267,214 (100%) spglib symmetry analysis
Volume 267,230 (100%) From structure
Density 267,230 (100%) From elements + volume
Structure JSON 267,230 (100%) pymatgen Structure serialization
Quality score 267,230 (100%) Composite score (0–88)
Provenance 267,230 (100%) Source, source_id, checksum
Duplicate group 21,140 (7.9%) Entries in dedup groups (31,997 total removed upstream)

SSE screening fields (stored in ssb_screening block)

Field Coverage Notes
SSE family 267,230 (100%) Composition-based heuristic (not structure-based)
Mobile ion 267,230 (100%) Li/Na/Mg presence-based
CAVD channel dimensionality 0% Algorithm was not re-run against Parquet store
SSE candidate score 100% 5-gate system (thermo + electronic + mobility + window + mechanical)
Thermo stability flag 99.8% E_hull < 0.025 eV/atom
Bulk/shear modulus 100% Geometric density-based proxy (not DFT elastic tensors)
Stability window 1,814 (0.7%) Grand-potential phase diagrams — computed for subset with low EaH
Interfacial reaction energy 39,706 (14.9%) Decomposition energy vs Li
BVSE migration barrier 24,873 (23% of Li/Na) bvlain v0.25.1, softBV percolation. 74.8% skip rate on attempted entries (98,773) due to bond-valence parameter coverage gaps

Sources

Source Entries License Download Date
Materials Project 69,279 CC BY 4.0 2026-07-20
OQMD 171,780 Non-commercial + attribution 2026-07-20
JARVIS-DFT 25,673 CC0 2026-07-20
OBELiX (experimental) 498 Per-article terms 2026-07-24

License Warning

⚠️ This dataset is NOT uniformly licensed. Each entry carries its own license.

  • license: "CC-BY-4.0" → MP entries (commercial safe, 26.1%)
  • license: "CC0-1.0" → JARVIS entries (commercial safe, 9.6%)
  • license: "OQMD-noncommercial" → OQMD entries (non-commercial only, 64.3%)

See LICENSE_BREAKDOWN.md. A Commercial-Safe edition (MP+JARVIS, ~94,952 entries) is extractable via scripts/extract_commercial_safe_edition.py.

Tier System

Tier Count Criteria
Strict Gold 56,966 11 gates: base Gold + quality ≥ 80 + no defects + provenance
Gold 96,242 8 gates: validated + unique + stable + complete metadata
Validated 140,382 5 gates: valid structure + targets + no critical issues
Raw 30,108 Source + formula present (may have quality issues)
Experimental Gold 498 OBELiX entries with measured conductivity

Family Imbalance

The dataset skews heavily toward intermetallics (62.5%) and layered oxides (15.7%). Solid-electrolyte-relevant composition families are a small fraction:

Family Total Gold Validated Raw
Intermetallic 166,930 36,520 109,056 21,354
Layered oxide 42,015 26,295 12,756 2,964
Halide SSE 18,803 12,560 5,211 1,032
Sulfide SSE 16,359 9,458 5,591 1,310
NASICON 560 488 70 2
Garnet 23 + 113 experimental 19 4 0

Source × Family Interaction

SSE-relevant families (halide + sulfide) are not uniformly distributed across sources — they skew toward the higher-quality sources:

Source Halide SSE Sulfide SSE SSE % of source Mean quality score
MP (69,279) 9,082 (13.1%) 7,666 (11.1%) 24.2% 87.4
JARVIS (25,673) 3,887 (15.1%) 2,251 (8.8%) 23.9% 77.8
OQMD (171,780) 5,834 (3.4%) 6,442 (3.8%) 7.1% 73.2

MP contributes more SSE-relevant entries (16,748) than OQMD (12,276) despite being 40% of its size, and MP's mean quality score is 14 points higher. This means quality-score thresholds do not disproportionately filter SSE-relevant chemistries — the interaction runs in the opposite direction. However, within-source analyses should still account for source effects.

If your target chemistry is garnets or sulfides specifically, usable Gold-tier entries number in the tens to low thousands — consider targeted acquisition (ICSD, structured literature extraction) before expecting ML models to generalize within these families.

BVSE Migration Barriers (Draft Quality)

Metric Value
Total Li/Na entries 108,015
Excluded (>60 sites) 8,744
No structure (experimental) 498
Attempted 98,773
Barriers computed 24,873
Superionic (≤0.25 eV) 1,501
Good (0.25–0.40 eV) 4,705
Moderate (0.40–0.55 eV) 5,009
Poor (>0.55 eV) 13,658
Skipped (no BV params) 73,900
Errors 0
Coverage of attempted 25.2%
Coverage of all Li/Na 23.0%

Caveat: 73,900 skipped entries (~75% of attempted) reflect bond-valence parameter coverage, not a sampling gap. Entries are flagged with bvse_skip_reason. Engine: bvlain v0.25.1, validated against 7 known SSEs (5/7 pass within literature, 2 known-marginal outliers documented in KNOWN_ISSUES.md).

Roadmap to True SSE-Property Coverage

v1.0.0 provides thermodynamic and structural screening data. Planned extensions:

  1. CAVD re-computation against the Parquet store (currently 0% coverage; algorithm exists in scripts/compute_cavd_channel_dimensionality.py)
  2. JARVIS EaH re-computation against the Parquet store (currently 0% for 25,673 JARVIS entries; script exists in scripts/compute_jarvis_hull_energy.py)
  3. BVSE barrier expansion — re-run with lower --max-sites threshold and extended parameter table to reduce the 75% skip rate
  4. MLIP-NEB migration barriers for top-tier stable candidates (active development)
  5. DFT NEB validation for a curated set of halide/sulfide/garnet candidates
  6. Elastic tensor data from MP API for mechanical property validation (scaffold exists in scripts/compute_mechanical_properties.py)
  7. Experimental conductivity cross-references beyond OBELiX pilot

Benchmark

Frozen train/val/test splits at dataset/splits/. Four split types with RF+Ridge baselines in MODEL_LEADERBOARD.md:

Split Train Val Test Purpose
Random 80/10/10 200,122 26,670 39,940 Basic generalization
Composition held-out 213,383 27,011 26,338 No formula overlap
Family held-out 260,984 5,165 583 Cross-family
Chemistry held-out 227,384 26,673 12,675 OOD (halides)

Caveat on family held-out split: 583 test entries (0.2% of corpus) is small — benchmark numbers quoted against this split will have wide error bars. It is useful as a smoke test for cross-family generalization but not sufficient to support strong SOTA claims. Consider supplementing with per-family stratified evaluation.

GNN baselines (CGCNN, MEGNet, ALIGNN) are in progress.

Intended Use

  • Primary: Upstream materials screening — filtering by phase stability, electronic structure, and structural family
  • Secondary: Cross-source DFT property harmonization, materials-informatics benchmarking, pretraining structure-based property predictors
  • Not recommended for: Quantitative phase diagram construction (use MP/OQMD directly), SSE conductivity ranking (requires transport-property labels not in this dataset)

Maintenance

  • Version: v1.0.0
  • DOI: pending (Zenodo archival in progress)
  • Issue tracking: GitHub Issues
  • Contact: Scandium Labs
Free AI Image Generator No sign-up. Instant results. Open Now