Home Dataset Research Capabilities Early Access
v1.0.0 · CC BY 4.0 · 267,230 entries

Open infrastructure for

MATERIALS AI

A harmonized, quality-scored foundation of DFT-computed structural and thermodynamic properties across 267,230 materials from Materials Project, OQMD, and JARVIS-DFT. Designed for AI-driven materials discovery.

267,230
Total Entries
3
Databases
47
Columns
178 MB
Parquet Size

WHY THIS DATASET

AI Needs Better Data,
Not Just Better Models

The biggest bottleneck in AI-driven materials discovery is no longer architecture — it's data. Most materials datasets are source-specific, lack quality scoring, and don't track provenance. Scandium Dataset solves this by harmonizing three major DFT databases into a single, quality-scored, provenance-tracked resource.

Cross-Source Harmony

Materials from MP, OQMD, and JARVIS are standardized to shared schema, units, and property definitions. No more manual mapping between database APIs.

Quality You Can Trust

Every entry is scored across 5 sub-scores. Calibrated against cross-source agreement. Score 80+ correlates with research-grade label quality.

Provenance Built In

Every entry tracks its source, source_id, checksum, and repair history. Know exactly where each value came from and how it was processed.

DATASET STATISTICS

By the Numbers

A comprehensive view of the dataset composition, quality distribution, and coverage across sources.

267,230
Total Entries
148,141
Unique Formulas
10
Material Families
8.42 GB
Total Dataset Size

TIER DISTRIBUTION

Strict Gold: 56,966
Gold: 96,242
Validated: 140,382
Raw: 30,108
Experimental: 498

PROPERTY COVERAGE

PropertyCoverageNotes
Formation energy (eV/atom)267,230 (100%)From MP, OQMD, JARVIS-DFT
Energy above hull (eV/atom)257,799 (96.5%)14,669 JARVIS entries computed via internal convex hull
Band gap (eV)267,079 (99.9%)151 OQMD entries with non-converged band gap
Space group267,214 (100%)spglib symmetry analysis
Volume (ų)267,230 (100%)From structure
Density (g/cm³)267,230 (100%)From elements + volume
Structure JSON267,230 (100%)pymatgen Structure serialization
Quality score267,230 (100%)Composite score (0–88)
Provenance267,230 (100%)Source, source_id, checksum
SSE family267,230 (100%)Composition-based heuristic

DATA PIPELINE

From Raw APIs to
Ready-to-Use Dataset

Every entry in the Scandium Dataset passes through a rigorous pipeline: ingestion, repair, deduplication, quality scoring, and tier classification.

I.

Source Ingestion

Data is pulled from Materials Project, OQMD, and JARVIS-DFT APIs with checksum verification. Raw archives are frozen and never modified.

II.

Coordinate Repair

OQMD coordinate artifacts detected and repaired. 137,405 entries with known coordinate issues undergo Type A repair with 100% success rate.

III.

Deduplication

Two-pass deduplication (intra-source + cross-source). 31,997 duplicate entries removed across 21,140 duplicate groups. Cross-source pairs identified.

IV.

Quality Scoring

Hierarchical scoring with 5 sub-scores: geometry, DFT convergence, metadata completeness, novelty, and chemical plausibility. Calibrated against cross-source agreement.

V.

Tier Classification

11-gate Strict Gold system. Entries classified as Strict Gold, Gold, Validated, Raw, or Experimental Gold based on validation gates passed.

VI.

Export & Release

Typed Parquet export with indexed lookup. Partitioned by source for selective loading. SHA256-checksummed releases with full MANIFEST.

DATA SOURCES

Three Foundations,
One Dataset

Each source contributes unique strengths. Together they provide comprehensive coverage of the inorganic crystal landscape.

Materials Project

69,279

CC BY 4.0

Highest data quality, complete metadata, reliable provenance. Serves as the gold standard reference.

OQMD

171,780

Non-Commercial

Largest source by volume. Required coordinate repair (137k entries) and symmetry computation (171k space groups).

JARVIS-DFT

25,673

CC0

Commercially safe (CC0). Missing EaH for all entries — computed via internal convex hull in v1.0.0.

OBELiX

498

Per-Article

Experimental Li-ion conductivity data. Integrated as experimental_gold tier with measured transport properties.

AVAILABLE PROPERTIES

Every Property
You Need

47 columns of structural, thermodynamic, electronic, and provenance data. Standardized units and null semantics documented in SCHEMA.md.

Thermodynamic

  • Formation energy (eV/atom)
  • Energy above hull (eV/atom)
  • Band gap (eV)
  • Magnetic ordering

Structural

  • Space group (spglib)
  • Crystal system
  • Volume (ų)
  • Density (g/cm³)
  • Nsites
  • Structure JSON

Metadata & Provenance

  • Source (MP/OQMD/JARVIS)
  • Source ID
  • Checksum
  • Quality score (0–88)
  • Quality tier
  • SSE family classification

QUICKSTART

Load the Dataset
in One Line

The dataset is distributed as typed Parquet files for fast, memory-efficient loading.

Inspect schema — no download needed Python
import pyarrow.parquet as pq
schema = pq.read_schema("dataset/entries_v4_typed_sample.parquet")
for f in schema:
    print(f"{f.name}: {f.type}")
Load full dataset with Pandas Python
import pandas as pd
df = pd.read_parquet("dataset/entries_v4_typed.parquet")
print(df.shape, list(df.columns))
Stream in batches — low memory mode Python
pf = pq.ParquetFile("dataset/entries_v4_typed.parquet")
for batch in pf.iter_batches(batch_size=1000, columns=["formula","band_gap"]):
    df_chunk = batch.to_pandas()
    print(f"Loaded {len(df_chunk)} rows")
Filter by source before loading Python
df_mp = pd.read_parquet("dataset/entries_v4_typed.parquet",
    filters=[("source","=","mp")])

DOWNLOAD

Get the Data

Choose the format that works best for your workflow. All files are checksum-verified.

Full Dataset (Parquet)

47 columns, typed, indexed

178 MB · entries_v4_typed.parquet

Download

Sample (1,000 rows)

Stratified sample for schema inspection

1.25 MB · entries_v4_typed_sample.parquet

Download

Battery Subset

82,925 battery-relevant entries

830 MB · battery_subset_v1.json

Download

Electrolyte Subset

41,665 strict Gold electrolyte candidates

492 MB · solid_electrolyte_candidate_subset_v1.json

Download
View on Hugging Face Dataset Hub

LICENSE & TERMS

Not Uniformly Licensed

Each entry carries its own license based on its source. Understand the terms before using the dataset commercially.

MP Entries

69,279

CC BY 4.0

Commercial safe. 26.1% of dataset.

JARVIS Entries

25,673

CC0 1.0

Commercial safe. 9.6% of dataset.

OQMD Entries

171,780

Non-Commercial

Non-commercial only. 64.3% of dataset.

Commercial-Safe Edition

94,952

MP + JARVIS

Extract via included script.

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

CITATION

How to Cite

If you use the Scandium Dataset in your research, please cite it.

BibTeX citation
@misc{scandium-dataset-v1,
    title  = {Scandium Dataset v1.0: A Harmonized Quality-Scored Foundation of DFT-Computed Properties for AI-Driven Materials Discovery},
    author = {Shamique Khan and Scandium Labs},
    year   = {2025},
    howpublished = {\url{https://huggingface.co/datasets/Scandium-Labs/Scandium-Dataset}},
    note   = {CC BY 4.0 (MP entries), CC0 1.0 (JARVIS entries), Non-commercial (OQMD entries)}
}

DOI pending — Zenodo archival in progress.

FAQ

Common Questions

What can this dataset be used for? ×
Phase stability screening, structural family classification, band gap prediction, materials-informatics benchmarking, pretraining property predictors on inorganic crystal structures, and cross-source DFT property harmonization studies.
What should I NOT use this dataset for? +
Direct ionic conductivity prediction, SSE performance ranking, or electrochemical stability assessment. The dataset contains migration barrier proxies for only 23% of Li/Na entries. It provides thermodynamic screening data, not transport property labels.
Can I use this dataset commercially? +
It depends on the entries you use. MP entries (CC BY 4.0) and JARVIS entries (CC0) are commercial-safe. OQMD entries (non-commercial + attribution) are not. A Commercial-Safe edition script is included to extract only MP + JARVIS entries.
How was quality scoring calibrated? +
Quality scores were calibrated against cross-source agreement. For example, MP-JARVIS overlap (11,741 formulas) shows FE MAE 0.20 eV and BG MAE 0.29 eV across sources. Score 80+ entries demonstrate monotonic improvements across 6 validation metrics — establishing research-grade reliability thresholds.
What is the difference between Gold and Strict Gold? +
Gold (96,242 entries) passes 8 validation gates including valid structure, unique composition, stable targets, and complete metadata. Strict Gold (56,966) adds 3 additional gates: quality score ≥ 80 (calibrated), complete provenance with checksum, and no known defects (volume > 0, density > 0, space group present).
Why are OQMD space groups now available? +
OQMD originally did not store symmetry data. A parallel spglib pass was run against all 171,780 OQMD structures, finding space groups for 171,764 entries (16 failed). This unblocked Gold tier classification for 29,068 OQMD entries that were previously stuck in Raw tier.
How do I load only MP data? +
Use Parquet filters: pd.read_parquet("dataset/entries_v4_typed.parquet", filters=[("source","=","mp")]). The dataset is also partitioned by source in the partitioned/ directory.

VERSION HISTORY

Changelog

Track the evolution of the Scandium Dataset across releases.

v1.0.0

Current Release

Harmonized MP + OQMD + JARVIS + OBELiX. 267,230 entries. BVSE migration barriers for 23% of Li/Na. CC BY 4.0 / CC0 / Non-commercial licensing. Parquet format with typed columns. Frozen benchmark splits.

v3.0-rc

Release Candidate

OQMD fixes applied (171,764 space groups, 47,807 volume/density). Strict Gold definition finalized. Quality score calibration completed. Cross-source agreement study published.

v2.0

Quality Scoring Introduced

Hierarchical quality scoring with 5 sub-scores. Tier classification (Gold/Validated/Raw). Cross-source deduplication pipeline established.

v1.0

Initial Release

First public release. Combined MP + OQMD + JARVIS. Basic validation and deduplication. JSON format.