Overview
What is the Scandium Dataset?
The Scandium Dataset is a curated, validated, versioned compilation of DFT-calculated material properties from three major open databases — Materials Project (MP), Open Quantum Materials Database (OQMD), and JARVIS-DFT — unified under a single schema, deduplicated, repaired, quality-scored, and organized for thermodynamic screening and benchmarking of battery-relevant inorganic materials.
⚠️ Important: The dataset contains formation energy, band gap, and energy
above hull — not ionic conductivity, migration barriers, or electrochemical
stability windows. See sse_readiness.md for a detailed
assessment of what it can and cannot do for solid-state electrolyte discovery.
Why does it exist?
Materials science has a data problem:
- Fragmented: 3+ major databases, 5+ formats, incompatible APIs
- Untrustworthy: Raw database dumps contain artifacts, duplicates, missing metadata
- Not battery-focused: General-purpose data is 62% intermetallics
- Not reproducible: No standard splits, no frozen versions, no provenance
- Not validated: Quality scores without calibration, claims without evidence
The Scandium Dataset solves all five problems.
Design Principles
- Trust through transparency — every claim has evidence, every defect is documented
- Battery-first — three editions (General, Battery, Electrolyte) serve different use cases
- Research-grade quality — strict Gold gates ensure only the best entries pass
- Versioned immutability — once frozen, data never changes
- Reproducible benchmarks — fixed splits, deterministic evaluation
The Pipeline
MP ─┐
├── Merge → Validate → Repair → Dedup → Score → Tier → Release
OQMD ┘
JARVIS ─┘
Key Facts
| Metric | Value |
|---|---|
| Total entries | 266,732 |
| Sources | 3 (MP, OQMD, JARVIS) |
| Families | 10 (7 battery-relevant) |
| Properties | FE, EaH, BG (per entry) |
| Quality sub-scores | 5 (geometry, DFT, metadata, novelty, chemical) |
| Gold tier | 96,242 (36.1%) |
| Strict Gold | 56,966 (21.4%) |
| Structures repaired | 137,405 (OQMD coordinates) |
| Space groups computed | 171,764 (OQMD via spglib) |
| Duplicates removed | 31,997 |