Scandium-Labs/Scandium-Dataset

1# Limitations2 3## Known Issues4 5### 1. 39,276 Gold entries below Strict Gold threshold639,276 base Gold entries have quality scores 70-79. They pass the base Gold7gate (≥70) but not Strict Gold (≥80). These entries are:8- Mostly OQMD (recently promoted to Gold by symmetry pass)9- Have valid structures and complete metadata10- But lack the metadata completeness needed for higher scores11 12### 2. JARVIS missing energy_above_hull (100%)13All 25,673 JARVIS entries lack energy_above_hull. This is a fundamental14limitation — JARVIS does not compute convex hull energies. These entries:15- Cannot contribute to EaH prediction tasks16- Are still valid for FE and BG prediction17- Can still achieve Gold tier (Gate 7 passes with null EaH)18 19### 3. Garnet family: 23 entries20Only 23 garnet-type solid electrolytes (LLZO, etc.) in the dataset. This is21insufficient for ML training. Garnet-specific models are not feasible with22current data.23 24### 4. 16 OQMD entries without space group25Spglib failed to find a space group for 16 OQMD entries. These remain in26Validated/Raw tier — cannot enter Gold without space group.27 28### 5. No experimentally validated entries29All entries are DFT-computed. No experimental validation is included. The30quality score measures DFT self-consistency, not experimental accuracy.31 32### 6. No battery-specific properties33The dataset does not include:34- Ionic conductivity35- Migration barriers36- Electrochemical stability windows37- Interface compatibility metrics38 39These are planned for future releases.40 41### 7. Source imbalance42OQMD dominates at 64.4%. Models trained on the full dataset may be biased43toward OQMD's distribution of structures and properties.44 45## Known Non-issues46 47These were flagged during development and are now resolved:48- OQMD space group missing → computed (171,764/171,780)49- OQMD volume = 0 → all fixed (47,807 entries)50- OQMD density = 0 → all computed (47,807 entries)51- JARVIS volume = 0 → all extracted (25,673 entries)52- OQMD coordinate artifact → repaired (137,405 entries)53 54## Intended Use55 56✅ Materials property prediction (FE, EaH, BG)57✅ Battery materials screening58✅ Cross-source DFT comparison studies59✅ Representation learning pre-training60✅ Transfer learning to experimental data61 62## Not Intended Use63 64❌ Predicting properties not in the dataset65❌ Modelling without acknowledging known limitations66❌ Claiming experimental accuracy for DFT-predicted properties67❌ Using without citation68