Datasets:

Scandium-Labs
/

Scandium-Dataset

Size: ~154,000 inorganic compounds
Properties: Formation energy, band gap, elastic constants, piezoelectric, dielectric, magnetic
Method: DFT (PAW-PBE/PBE+U/HSE06)
License: CC BY 4.0
Limitation: Single-source, no cross-validation. Some properties (e.g., ion mobility) not available.
Relationship to this work: MP is one of three constituent sources (69,279 entries).

Size: ~1,000,000 structures
Properties: Formation energy, stability
Method: DFT (PAW-PBE)
License: Non-commercial + attribution
Limitation: Less curated than MP. Known issues with coordinate artifacts and missing symmetry metadata.
Relationship to this work: OQMD is the largest constituent source (171,780 entries). The Scandium Dataset repairs known OQMD defects (coordinate artifacts, space group determination, volume extraction).

Size: ~76,000 structures (DFT 3D subset: 25,673)
Properties: Formation energy, band gap (TBmBJ), elastic, optical, magnetic, phonons, 2D
Method: DFT (optPBE + TBmBJ)
License: CC0 (Public Domain)
Limitation: No convex hull distance (energy above hull). Uses different functionals than MP/OQMD.
Relationship to this work: JARVIS contributes 25,673 entries. Its TBmBJ band gaps provide complementary information. The EaH gap is documented as a known limitation.

Size: ~3,500,000 entries
Properties: Formation energy, band gap, elastic
Method: DFT (PAW-PBE)
License: Varies
Relationship to this work: Not included as a source (planned for v4.0). AFLOW is larger but less curated per-entry.

Size: 12,000,000+ calculations
Properties: Raw DFT outputs
Method: Various (community-contributed)
License: CC BY 4.0
Relationship to this work: Not included (planned for v4.0). NOMAD's heterogeneity makes quality scoring more difficult.

2. Aggregated/Harmonized Datasets

Source: MP + Alexandria + OQMD
Key innovation: Harmonized schema across multiple DFT functionals; reusable fetching/curation library; trajectory data for MD simulations
Differences from this work:
- LeMat-Traj includes trajectory/relaxation data; Scandium focuses on static properties
- LeMat-Traj does not implement a quality tiering system with calibrated gates
- Scandium adds battery-specific family classification and frozen benchmark splits
- Scandium provides cross-source agreement quantification (FE MAE 0.20 eV/atom across 11,741 overlapping formulas)
- Scandium adds provenance tracking and explicit defect documentation per entry
Relationship: LeMat-Traj is the closest prior art. The tiering methodology and battery-specific curation are the primary differentiators. The paper must explicitly acknowledge LeMat-Traj and narrow the novelty claim to these differentiators.

Format: Benchmark suite with 13 tasks across formation energy, band gap, elastic constants, phonons, perovskites
Key innovation: Standardized train/val/test splits and evaluation protocol
Differences from this work:
- Matbench tasks are smaller (1.3K–132K entries)
- Matbench uses frozen single-source splits (primarily MP)
- Scandium provides multi-source splits with multiple held-out strategies
- Scandium includes battery-specific task definitions and tier-based evaluation
- Matbench has broad property coverage; Scandium focuses on battery-relevant properties
Relationship: Matbench is the standard benchmark for crystal property prediction. The Scandium benchmark adds multi-source and OOD dimensions not covered by Matbench.

Format: Benchmark across DFT, force fields, ML, and 2D materials
Key innovation: Multi-property benchmark with leaderboard
Differences from this work:
- JARVIS-Leaderboard includes force fields and molecular dynamics; Scandium is property-prediction focused
- Scandium provides per-entry quality scoring, tiering, and provenance
- Scandium's battery-specific subset and cross-source validation are unique

Focus: Liquid and solid electrolyte properties
Properties: Ionic conductivity, electrochemical stability, diffusion barriers
Relationship: The Electrolyte Genome focuses on computed transport properties; Scandium provides foundational structural and thermodynamic data that can be used to train models for such properties.

Focus: Battery materials property database
Properties: Formation energy, band gap, ionic conductivity, experimental data
Relationship: BatteryHub includes experimental data not present in Scandium. Future versions plan to integrate experimental validation.

Dataset	Size	Sources	Quality Scoring	Benchmark Splits	Battery Focus	License
Scandium (this work)	266,732	MP+OQMD+JARVIS	✓ (4 tiers, 11 gates)	✓ (4 types, OOD)	✓	Multi
Materials Project	~154K	Single	✗	Partial (Matbench)	Partial	CC BY 4.0
OQMD	~1M	Single	✗	✗	Partial	Non-commercial
JARVIS-DFT	~76K	Single	Partial	Leaderboard	✗	CC0
LeMat-Traj	~500K	MP+Alexandria+OQMD	✗	✗	✗	Multi
Matbench	1.3K–132K	MP (+others)	✗	✓ (frozen)	✗	CC BY 4.0
AFLOW	~3.5M	Single	✗	✗	✗	Varies
Electrolyte Genome	~50K	MP	✗	✗	✓	CC BY 4.0

The Scandium Dataset's contributions are:

Multi-source unification with quality tiering — not just schema harmonization (done by LeMat-Traj), but a calibrated 4-tier quality system with provenance tracking per entry
Cross-source validation — quantifying agreement across sources (FE MAE 0.20 eV/atom) and documenting systematic biases (MP vs OQMD quality score gap: 87.4 vs 73.2)
Battery-specific curation — family classification, battery (82,925) and electrolyte (41,665) subsets with relevance methodology
Frozen OOD benchmark splits — including chemistry held-out (halides) and family held-out splits that test generalization beyond training distribution
Repair documentation — 137,405 coordinate repairs, 47,807 volume repairs, with full provenance chain per entry