Scandium-Dataset / ROADMAP.md
Scandiumlabs's picture
Upload ROADMAP.md with huggingface_hub
92b38dd verified
|
Raw
History Blame
3.18 kB

Scandium-Dataset Roadmap

Vision: Become the field's reference open dataset for battery-relevant materials screening — combining scale (267k+ entries) with harmonized quality scoring, cross-source provenance, and benchmark splits that no single-source dataset provides. Transport-property layers (ionic conductivity, migration barriers) are planned for future versions.

Strategic plan: See docs/ssb_strategic_roadmap.md.


v0.1.0 — Foundation Release (Done)

Focus: Clean, validated, audited general inorganic materials aggregation

  • Aggregate MP, OQMD, JARVIS-DFT (267,230 entries)
  • Coordinate repair (171,780 OQMD entries)
  • Symmetry pass (171,764 space groups)
  • Volume/density repair (73,480 entries)
  • 11-gate validation pipeline
  • Multi-source deduplication (31,997 duplicates removed)
  • Quality scoring with 5 sub-scores
  • Three-tier system (Gold / Validated / Raw)
  • Battery-relevant (82,925) and electrolyte candidate (41,665) subsets
  • Benchmark splits (4 types: random, composition-held-out, family-held-out, chemistry-held-out)
  • Complete 8-phase audit
  • Full documentation (audit reports, trust report, dataset card, known issues)
  • Licensing resolved (per-entry license field)

v1.0.0 — Scope Clarification + Parquet Store (Current)

Focus: Corrected scope (thermodynamic screening, not SSE discovery), Parquet storage, experimental integration, BVSE draft

  • Parquet dataset store (47 columns, 184 MB, indexed lookup)
  • OBELiX experimental data integration (498 entries with measured conductivity)
  • BVSE migration barrier proxy (24,873 barriers, 1,501 superionic)
  • KNOWN_ISSUES.md: scope clarified as thermodynamic screening
  • Dataset card: Included/Not Included and Suitable/Not Suitable sections added
  • Version bumped to v1.0.0, framing corrected from "SSB discovery" to "materials screening"

Phase 1 — Property Gaps to Close (Next)

Focus: Close documentation-vs-data gaps, expand SSE screening fields

  • Run CAVD channel-dimensionality analysis against Parquet store (currently 0% coverage; algorithm exists)
  • Compute JARVIS EaH against Parquet store (currently 0% for 25,673 entries; script exists)
  • Re-run garnet enrichment (target 500+) — currently 136 entries
  • Reduce BVSE skip rate by adjusting --max-sites threshold and parameter table
  • Expand electrochemical stability windows beyond 0.7% coverage

Phase 2 — MLIP-Driven Transport Properties

Focus: Migration barriers + elastic tensors

  • MLIP-NEB migration barriers on top SSE candidates
  • DFT NEB validation for halide/sulfide/garnet subset
  • Elastic tensor data from MP API (scaffold exists)
  • Full sse_candidate_score with all 5 gates populated
  • GNN baselines (CGCNN, MEGNet, ALIGNN) added to MODEL_LEADERBOARD

Phase 3 — Production Transport Layer

  • Benchmark-quality ionic conductivity predictions
  • Experimental conductivity cross-references (beyond OBELiX pilot)
  • DOI registration via Zenodo
  • Third-party audit of data pipeline