SCHEMA.md

# SCHEMA — Scandium-Dataset v1.0.0

## File: `entries_v4_typed.parquet`

**Rows:** 267,230  
**Columns:** 47  
**Storage:** Typed Parquet (PyArrow) with proper int/float/bool types  
**Index:** `entries_v4_typed.index.json` — O(1) lookup by `source_id`

---

## Column Dictionary

| # | Column | Type | Unit | Nullable | Description | Source Field Mapping |
|---|--------|------|------|----------|-------------|---------------------|
| 1 | `source` | string | — | No | Source database: `"mp"`, `"oqmd"`, `"jarvis"`, or `"experimental"` | All |
| 2 | `source_id` | string | — | No | Original ID in the source database (e.g. `mp-12345`, `oqmd-67890`, `JVASP-123`) | MP: `material_id`; OQMD: `id`; JARVIS: `jid` |
| 3 | `formula` | string | — | No | Pretty chemical formula (e.g. `Li6PS5Cl`) | All |
| 4 | `structured_formula` | string | — | No | Reduced/structured formula via pymatgen `reduced_formula` | Computed |
| 5 | `band_gap` | double | eV | Yes (151 OQMD entries) | Electronic band gap. `null` when calculation did not converge. | MP: `band_gap`; OQMD: `band_gap`; JARVIS: `bandgap` |
| 6 | `formation_energy_per_atom` | double | eV/atom | No | Formation energy per atom. Negative = stable relative to elements. | MP: `formation_energy_per_atom`; OQMD: `delta_e`; JARVIS: `formation_energy` |
| 7 | `energy_above_hull` | double | eV/atom | Yes (8,933 JARVIS + 5,238 others) | Energy above convex hull. `null` when chemical system has <3 entries (hull construction impossible) or JARVIS subset lacks hull coverage. | MP: `energy_above_hull`; OQMD: `delta_e_hull`; JARVIS: computed via internal hull |
| 8 | `nsites` | int64 | atoms | No | Number of atoms in the conventional unit cell | All (from structure) |
| 9 | `nelements` | string | — | No | JSON array of element symbols (e.g. `["Li","P","S","Cl"]`) | Computed |
| 10 | `volume` | double | ų | No | Unit cell volume | MP: `volume`; OQMD/JARVIS: computed from lattice vectors |
| 11 | `density` | double | g/cm³ | No | Computed density from atomic masses and volume | Computed |
| 12 | `mobile_ion` | string | — | No | Mobile ion carrier: `"Li"`, `"Na"`, `"Mg"`, or `"None"` | Computed (composition heuristic) |
| 13 | `tier` | string | — | No | Quality tier: `"gold"`, `"validated"`, `"raw"`, `"experimental_gold"` | Computed (11-gate system) |
| 14 | `quality_score` | int64 | score (0–88) | No | Composite quality score from 5 sub-scores. Range 48–88 (conservative; no entry ≥90) | Computed |
| 15 | `is_experimental` | bool | — | No | `true` if entry is from experimental data (OBELiX), `false` otherwise | Computed |
| 16 | `is_stable` | string | — | No | Stability flag: `"True"`, `"False"`, or `"null"` (lacking hull data) | MP: `is_stable`; OQMD: `is_stable`; JARVIS: computed |
| 17 | `is_metal` | bool | — | No | `true` if band gap ≤ 0.1 eV | Computed from `band_gap` |
| 18 | `space_group` | int64 | number | Yes (16 OQMD entries) | International space group number (1–230). `null` when spglib fails (disordered or too few atoms). | Computed via spglib |
| 19 | `space_group_number` | string | — | Yes | Space group number as string (for compatibility) | Computed |
| 20 | `crystal_system` | string | — | Yes | Crystal system: `"triclinic"`, `"monoclinic"`, etc. | Computed |
| 21 | `sse_family` | string | — | No | SSE structural family classification: `"garnet"`, `"NASICON"`, `"argyrodite"`, `"LGPS_type"`, `"LISICON"`, `"anti_perovskite"`, `"perovskite"`, `"halide"`, `"borohydride"`, `"sulfide"`, `"oxide"`, `"none"` | Computed (composition-based heuristic) |
| 22 | `license` | string | — | No | Entry-level license: `"CC-BY-4.0"`, `"CC0-1.0"`, or `"OQMD-noncommercial"` | Per-source |
| 23 | `duplicate_group` | int64 | — | Yes (245,490 are null) | Group ID for entries with identical formula from multiple sources. Non-null only for members of duplicate groups. | Computed |
| 24 | `source_weight` | double | — | No | Relative source weight for ensemble modeling (0.0–1.0). Reflects source-level confidence. | Computed |
| 25 | `total_magnetization` | double | μB | Yes (non-magnetic entries) | Total magnetic moment. 0.0 for non-magnetic systems; `null` if unavailable. | MP: `total_magnetization`; OQMD: `total_magnetization`; JARVIS: `magmom` |
| 26 | `fe_per_carrier` | string | eV/carrier | Yes | Formation energy per mobile ion carrier for battery-relevant entries. | Computed |
| 27 | `volume_per_carrier` | string | ų/carrier | Yes | Volume per mobile ion carrier. | Computed |
| 28 | `carrier_fraction` | double | — | Yes | Fraction of sites occupied by mobile ions | Computed |
| 29 | `electronic_insulation` | bool | — | No | `true` if band gap ≥ 1.0 eV (rough electronic insulation criterion) | Computed |
| 30 | `space_group_symbol` | string | — | Yes | Hermann-Mauguin space group symbol (e.g. `"P-1"`, `"Fm-3m"`) | Computed via spglib |
| 31 | `magnetic_ordering` | string | — | Yes | Magnetic ordering classification: `"NM"` (non-magnetic), `"FM"` (ferromagnetic), `"AFM"` (antiferromagnetic), `"FiM"` (ferrimagnetic), or `null` | Computed |
| 32 | `structure_json` | string | — | No | JSON-serialized pymatgen `Structure` object. Contains lattice parameters, site positions, species. | All sources |
| 33 | `ssb_screening` | string | — | No | JSON block of SSE screening fields (see SSB Screening sub-schema below). | Computed |
| 34 | `provenance` | string | — | No | JSON block of provenance tracking fields (see Provenance sub-schema below). | Computed |
| 35 | `elements` | string | — | No | JSON array of all elements present (space-separated for some sources) | All sources |
| 36 | `families` | string | — | No | JSON array of battery-relevant family classifications | Computed |
| 37 | `carrier_elements` | string | — | Yes | JSON array of mobile carrier elements (Li, Na, Mg) if present | Computed |
| 38 | `quality_flags` | string | — | Yes | JSON array of quality defect flags (empty if none) | Computed |
| 39 | `tags` | string | — | Yes | JSON array of user-facing tags (e.g. `["single_atom"]`, `["extreme_fe"]`) | Computed |
| 40 | `notes` | string | — | Yes | Free-text notes about the entry | Computed |
| 41 | `anions` | string | — | Yes | JSON array of anion elements present | Computed |
| 42 | `strict_gold` | string | — | No | Strict Gold eligibility: `"True"` or `"False"` | Computed |
| 43 | `tier_detail` | string | — | No | JSON dictionary of tier classification details | Computed |
| 44 | `tier_gates` | string | — | No | JSON dictionary of gate pass/fail flags (G0–G11) | Computed |
| 45 | `quality_sub_scores` | string | — | No | JSON dictionary of 5 quality sub-scores: geometry, dft, metadata, novelty, chemical | Computed |
| 46 | `references` | string | — | Yes | JSON array of literature references for the entry | All sources |
| 47 | `merged_sources` | string | — | Yes | JSON array of source IDs merged into this entry during deduplication | Computed |

---

## Sub-schemas

### `provenance` (JSON block)

```json
{
  "source": "mp",
  "source_id": "mp-12345",
  "download_date": "2026-07-20",
  "processed_at": "2026-07-21T18:30:00Z",
  "version": "v1.0.0",
  "repairs_applied": ["repair_coordinate_artifact", "repair_volume"],
  "checksum": "sha256:abc123..."
}
```

| Field | Type | Description |
|-------|------|-------------|
| `source` | string | Source database name |
| `source_id` | string | Original ID in source |
| `download_date` | string | Date when raw data was extracted from source API (ISO 8601) |
| `processed_at` | string | When the entry was processed through the pipeline (ISO 8601) |
| `version` | string | Dataset version that created this entry |
| `repairs_applied` | array of strings | List of all repairs applied to this entry (empty if none) |
| `checksum` | string | SHA-256 checksum of the entry's structure + properties JSON blob |

### `ssb_screening` (JSON block)

```json
{
  "sse_candidate_score": 12,
  "gates_passed": ["thermo_stability", "electronic_insulation"],
  "thermo_stability_flag": true,
  "stability_window_low_V": 0.0,
  "stability_window_high_V": 5.2,
  "window_width_V": 5.2,
  "interfacial_reaction_energy_vs_Li_eV_atom": 0.15,
  "passivating_interphase": true,
  "bulk_modulus_GPa": 35.2,
  "shear_modulus_GPa": 18.1,
  "dendrite_suppression_flag": true,
  "bvse_migration_barrier_eV": 0.42,
  "has_bvse_barrier": true,
  "bvse_skip_reason": null,
  "cavd_channel_dimensionality": null
}
```

| Field | Type | Coverate | Description |
|-------|------|----------|-------------|
| `sse_candidate_score` | int | 100% | 5-gate composite score (10 base + 1 per gate passed). Range 10–15. |
| `gates_passed` | array | 100% | List of passed gate names |
| `thermo_stability_flag` | bool | 99.8% | True if E_hull < 0.025 eV/atom |
| `stability_window_low_V` | float | 0.7% | Lower bound of electrochemical stability window vs Li/Li+ |
| `stability_window_high_V` | float | 0.7% | Upper bound of electrochemical stability window |
| `window_width_V` | float | 0.7% | Width of stability window |
| `interfacial_reaction_energy_vs_Li_eV_atom` | float | 14.9% | Decomposition energy vs Li metal |
| `passivating_interphase` | bool | 14.9% | True if reaction products are electronically insulating |
| `bulk_modulus_GPa` | float | 100% | Geometric density-based proxy (not DFT elastic tensor) |
| `shear_modulus_GPa` | float | 100% | Geometric density-based proxy (not DFT elastic tensor) |
| `dendrite_suppression_flag` | bool | 100% | True if shear_modulus > 6 GPa |
| `bvse_migration_barrier_eV` | float | 23% Li/Na | BVSE migration barrier estimate via bvlain v0.25.1 |
| `has_bvse_barrier` | bool | 100% | True if BVSE barrier was successfully computed |
| `bvse_skip_reason` | string | 77% Li/Na | Reason BVSE was skipped (e.g. "no BV params for composition", "max_sites_exceeded") |
| `cavd_channel_dimensionality` | string | 0% | CAVD channel dimensionality (0D/1D/2D/3D). Not yet computed — see CHANGELOG. |

---

## Null Semantics

| Null Pattern | Meaning | Examples |
|-------------|---------|----------|
| `null` in `band_gap` | Calculation did not converge | 151 OQMD entries |
| `null` in `energy_above_hull` | Chemical system has <3 entries (sparse) or source does not compute hull distance | 8,933 JARVIS entries |
| `null` in `space_group` | spglib failed (disordered, too few atoms) | 16 OQMD entries |
| `null` in `duplicate_group` | Entry has no duplicates across sources | 245,490 entries |
| `null` in `total_magnetization` | Non-magnetic system or field not computed | Varies |
| `null` in `ssb_screening.*` | Screening field not computed for this entry | CAVD (0%), stability windows (99.3%) |

**Never null:** `source`, `source_id`, `formula`, `structured_formula`, `formation_energy_per_atom`, `nsites`, `nelements`, `volume`, `density`, `mobile_ion`, `tier`, `quality_score`, `is_experimental`, `is_metal`, `sse_family`, `license`, `source_weight`, `electronic_insulation`, `structure_json`, `ssb_screening`, `provenance`, `elements`, `families`, `strict_gold`, `tier_detail`, `tier_gates`, `quality_sub_scores`.

---

## Units Reference

| Property | Unit | Notes |
|----------|------|-------|
| Formation energy | eV/atom | Per-atom, not per-formula-unit |
| Energy above hull | eV/atom | Per-atom |
| Band gap | eV | Electronic band gap (GGA-level) |
| Volume | ų | Conventional cell |
| Density | g/cm³ | Computed from elements + volume |
| Total magnetization | μB | Bohr magnetons |
| BVSE migration barrier | eV | SoftBV percolation method |
| Stability window | V vs Li/Li+ | Grand potential phase diagram |
| Interfacial reaction energy | eV/atom | Decomposition energy vs Li |
| Bulk/shear modulus | GPa | Geometric proxy (not DFT) |

---

## Source DFT Parameter Heterogeneity

| Parameter | MP | OQMD | JARVIS-DFT |
|-----------|-----|------|------------|
| Functional | PBE (Perdew-Burke-Ernzerhof) | PBE | PBE (OPT variants for some) |
| U correction | Hubbard U for transition metals | Custom U values (differs from MP) | No U correction (default PBE) |
| Pseudopotentials | PAW (VASP-supplied) | PAW (VASP-supplied, older set) | USPP / PAW (mixed) |
| Encodings | ENCUT = 520 eV (default +30% for some) | ENCUT = 520 eV (standard) | ENCUT varies (USPP-compatible) |
| k-point density | KSPACING ~0.22 Å⁻¹ | Variable (denser for small cells) | Variable |
| Band gap method | PBE (indirect) | PBE | PBE + OPT / TB-mBJ (for some) |

**Important:** Direct comparison of `formation_energy_per_atom`, `energy_above_hull`, or `band_gap` across sources is scientifically questionable due to differing DFT parameters. Cross-source agreement statistics are published in `dataset/statistics/cross_source_agreement.json`. See `docs/limitations.md` for guidance on source-aware analysis.

---

## Version History

| Version | File | Notes |
|---------|------|-------|
| v1.0.0 | `entries_v4_typed.parquet` | Current. Typed Parquet with proper int/float/bool columns. 47 columns. |
| v0.3.0 | `entries_v3.parquet` | Previous. Un-typed Parquet, same 47 columns. |
| v0.0.0 | `entries_final_v3.json` | Original JSON store (1.6 GB). Deprecated. |

**Migration v3 → v4:** Column names and semantics are identical. v4 adds explicit PyArrow schema typing (int64, float64, bool) to enable HF Dataset viewer compatibility and streaming. No data values changed.