Scandium-Dataset / docs /limitations.md
Scandiumlabs's picture
Upload docs/limitations.md with huggingface_hub
6af04d8 verified
|
Raw
History Blame Contribute Delete
2.47 kB

Limitations

Known Issues

1. 39,276 Gold entries below Strict Gold threshold

39,276 base Gold entries have quality scores 70-79. They pass the base Gold gate (≥70) but not Strict Gold (≥80). These entries are:

  • Mostly OQMD (recently promoted to Gold by symmetry pass)
  • Have valid structures and complete metadata
  • But lack the metadata completeness needed for higher scores

2. JARVIS missing energy_above_hull (100%)

All 25,673 JARVIS entries lack energy_above_hull. This is a fundamental limitation — JARVIS does not compute convex hull energies. These entries:

  • Cannot contribute to EaH prediction tasks
  • Are still valid for FE and BG prediction
  • Can still achieve Gold tier (Gate 7 passes with null EaH)

3. Garnet family: 23 entries

Only 23 garnet-type solid electrolytes (LLZO, etc.) in the dataset. This is insufficient for ML training. Garnet-specific models are not feasible with current data.

4. 16 OQMD entries without space group

Spglib failed to find a space group for 16 OQMD entries. These remain in Validated/Raw tier — cannot enter Gold without space group.

5. No experimentally validated entries

All entries are DFT-computed. No experimental validation is included. The quality score measures DFT self-consistency, not experimental accuracy.

6. No battery-specific properties

The dataset does not include:

  • Ionic conductivity
  • Migration barriers
  • Electrochemical stability windows
  • Interface compatibility metrics

These are planned for future releases.

7. Source imbalance

OQMD dominates at 64.4%. Models trained on the full dataset may be biased toward OQMD's distribution of structures and properties.

Known Non-issues

These were flagged during development and are now resolved:

  • OQMD space group missing → computed (171,764/171,780)
  • OQMD volume = 0 → all fixed (47,807 entries)
  • OQMD density = 0 → all computed (47,807 entries)
  • JARVIS volume = 0 → all extracted (25,673 entries)
  • OQMD coordinate artifact → repaired (137,405 entries)

Intended Use

✅ Materials property prediction (FE, EaH, BG) ✅ Battery materials screening ✅ Cross-source DFT comparison studies ✅ Representation learning pre-training ✅ Transfer learning to experimental data

Not Intended Use

❌ Predicting properties not in the dataset ❌ Modelling without acknowledging known limitations ❌ Claiming experimental accuracy for DFT-predicted properties ❌ Using without citation