Upload folder using huggingface_hub

README.md

@@ -0,0 +1,292 @@
+---
+license: other
+license_name: mixed
+license_link: LICENSE
+language:
+- en
+tags:
+- materials-science
+- fusion-energy
+- irradiation
+- mechanical-properties
+- nuclear-materials
+- scientific-data
+pretty_name: FusionMatDB
+size_categories:
+- 10K<n<100K
+task_categories:
+- tabular-regression
+- other
+---
+
+# FusionMatDB — Fusion Irradiation Materials Database
+
+**The first publicly accessible, ML-ready database of fusion materials irradiation effects.**
+
+Extracted from 65 ORNL Fusion Materials Program semiannual progress reports (1990–2024) using Gemini 3 Flash vision on Vertex AI, plus the SDC-IC ITER Structural Design Criteria material library.
+
+No equivalent open-access database exists. EUROfusion EDDI (~3,000 records) is restricted to EU consortium members. This dataset fills that gap.
+
+---
+
+## Dataset Visualisations
+
+![Irradiation coverage map](fig1_coverage_map.png)
+*Figure 1: Every record plotted in dose–temperature space. The database covers fission reactor conditions (1–100 dpa, 200–750°C) across 19 material classes — the regime relevant to ITER, DEMO, and private fusion machines.*
+
+![Radiation hardening curves](fig2_hardening_curves.png)
+*Figure 2: Yield strength vs dose for RAFM steels and vanadium alloys — the core scientific signal. Higher doses and lower temperatures produce more hardening, consistent with dispersed barrier hardening (DBH) theory.*
+
+![Coverage bars](fig3_coverage_bars.png)
+*Figure 3: Records by material class (left) and property type (right). RAFM steels dominate — reflecting 40 years of ORNL focus on ferritic/martensitic steels for fusion first-wall applications.*
+
+![Analysis panels](fig4_analysis_panels.png)
+*Figure 4 (left to right): Void swelling vs dose shows the expected increasing trend — a physics-consistency check validating extraction accuracy. Records per ORNL report volume spans 1990–2024. Confidence score distribution shows 85% of records scoring ≥ 0.7.*
+
+---
+
+## Dataset Summary
+
+| | |
+|---|---|
+| **Total records** | 22,269 |
+| **Train / Val / Test** | 17,800 / 2,225 / 2,225 (80/10/10, stratified by material class) |
+| **Features per record** | 54 |
+| **Source documents** | 65 ORNL semiannual reports + SDC-IC ITER Material Library |
+| **Material classes** | 19 (RAFM steel, vanadium alloy, copper alloy, tungsten, SiC, ceramics, ODS, austenitic SS, nano-laminates, and more) |
+| **Date extracted** | April 2026 |
+| **Extraction model** | Gemini 3 Flash (Vertex AI Express, temperature=0) |
+
+---
+
+## Data Sources and Licence
+
+| Source | Records | Licence |
+|---|---|---|
+| ORNL Fusion Materials Program semiannual progress reports (vols. 10–75) | 20,318 | Public domain (U.S. DOE) |
+| SDC-IC ITER Structural Design Criteria Material Library | 1,951 | EUPL-1.2 |
+
+**Licence note:** ORNL data is U.S. federal government work (public domain). SDC-IC data is EUPL-1.2 (copyleft, allows commercial use with attribution). When using only ORNL-sourced records (`source == "gemini_vision"`), the dataset is effectively public domain. When including SDC-IC records (`source == "sdc_ic_parse"`), attribution under EUPL-1.2 applies.
+
+Filter by `source` column to use the licence appropriate for your use case.
+
+---
+
+## Splits
+
+```python
+from datasets import load_dataset
+
+ds = load_dataset("khalizo/fusionmatdb")  # all splits
+train = load_dataset("khalizo/fusionmatdb", split="train")
+val   = load_dataset("khalizo/fusionmatdb", split="validation")
+test  = load_dataset("khalizo/fusionmatdb", split="test")
+```
+
+Splits are stratified by `material_class`. Rare classes (<30 records) are pooled for stratification purposes.
+
+---
+
+## Features
+
+### Material identification
+| Column | Type | Description |
+|---|---|---|
+| `material_name` | string | Canonical name (e.g. `"EUROFER97"`, `"V-4Cr-4Ti"`, `"F82H"`) |
+| `material_class` | string | Class (see Material Classes below) |
+| `source` | string | `gemini_vision` or `sdc_ic_parse` |
+| `paper_id` | string | Source document ID (e.g. `"ornl_70"`) |
+
+### Elemental composition (weight %)
+`W_wt_pct`, `Cr_wt_pct`, `V_wt_pct`, `Ta_wt_pct`, `Fe_wt_pct`, `C_wt_pct`, `Mn_wt_pct`, `Mo_wt_pct`, `Ni_wt_pct`, `Si_wt_pct`, `Ti_wt_pct`, `Al_wt_pct`
+
+### Processing
+| Column | Type | Description |
+|---|---|---|
+| `manufacturer` | string | Manufacturer name |
+| `product_shape` | string | Form (e.g. `"rolled plate"`, `"rod"`) |
+| `temper_temp_C` | float | Tempering temperature (°C) |
+| `grain_size_um` | float | Grain size (µm) |
+| `layer_spacing_nm` | float | Bilayer thickness for nano-laminates (nm) |
+
+### Irradiation conditions
+| Column | Type | Description |
+|---|---|---|
+| `irradiation_state` | string | `"irradiated"` or `"unirradiated"` |
+| `dose_dpa` | float | Displacement per atom (0–500 dpa validated) |
+| `irradiation_temp_C` | float | Irradiation temperature (°C; cryogenic values are physically correct) |
+| `reactor` | string | Facility (e.g. `"HFIR"`, `"BOR-60"`, `"EBR-II"`, `"ion_beam"`) |
+| `neutron_spectrum` | string | `"fission"`, `"fast"`, `"mixed"`, `"ion"` |
+| `helium_appm` | float | Transmutation helium (appm) |
+
+### Mechanical properties
+| Column | Type | Description |
+|---|---|---|
+| `yield_strength_mpa_unirradiated` | float | Yield strength before irradiation (MPa) |
+| `yield_strength_mpa_irradiated` | float | Yield strength after irradiation (MPa) |
+| `yield_strength_mpa_std` | float | Measurement uncertainty (MPa) |
+| `uts_mpa_unirradiated` | float | Ultimate tensile strength, unirradiated (MPa) |
+| `uts_mpa_irradiated` | float | Ultimate tensile strength, irradiated (MPa) |
+| `elongation_pct_irradiated` | float | Elongation after irradiation (%) |
+| `dbtt_k_irradiated` | float | Ductile-to-brittle transition temperature after irradiation (K) |
+| `fracture_toughness_mpa_sqrt_m` | float | Fracture toughness (MPa√m) |
+| `charpy_energy_j` | float | Charpy impact energy (J) |
+| `hardness_value` | float | Hardness (HV or as noted in `hardness_type`) |
+| `volumetric_swelling_pct` | float | Void swelling (%) |
+| `void_diameter_nm` | float | Average void diameter (nm) |
+| `void_density_per_m3` | float | Void density (m⁻³) |
+| `dislocation_loop_diameter_nm` | float | Dislocation loop diameter (nm) |
+| `creep_rate_per_s` | float | Steady-state creep rate (s⁻¹) |
+| `electrical_resistivity_uohm_cm_irradiated` | float | Electrical resistivity post-irradiation (µΩ·cm) |
+| `dielectric_breakdown_kv_per_mm_irradiated` | float | Dielectric breakdown strength post-irradiation (kV/mm) |
+
+### ML metadata
+| Column | Type | Description |
+|---|---|---|
+| `confidence_score` | float | Extraction quality (0.0–1.0); based on field completeness |
+| `reviewed_by_human` | bool | True for SDC-IC records (human-curated) |
+| `split` | string | `train`, `validation`, or `test` |
+
+---
+
+## Material Classes
+
+| Class | Example materials | Records |
+|---|---|---|
+| `RAFM_steel` | F82H, EUROFER97, HT-9, 9Cr, T91, Grade 91 | 5,512 |
+| `vanadium_alloy` | V-4Cr-4Ti, V-5Cr-5Ti, V-2.5Ti-1Si | 2,956 |
+| `copper_alloy` | CuCrZr, GlidCop, MARZ copper, OFHC Cu | 2,415 |
+| `austenitic_steel` | 316 SS, 304 SS, JPCA, PCA, Fe-Cr-Ni alloys | 2,118 |
+| `other` | Multi-material, ambiguous, or LWR-specific | 2,391 |
+| `ceramic_insulator` | Al₂O₃, MgAl₂O₄, AlN, SiC, BN, BeO | 1,094 |
+| `SiC_composite` | SiC/SiC, Hi-Nicalon composites | 887 |
+| `ODS_steel` | MA957, PM2000, 14YWT | 784 |
+| `tungsten_alloy` | W-Re, K-doped W, La-doped W, W-NiFe | 712 |
+| `tungsten` | Pure W, W single crystal | 638 |
+| `ferritic_model_alloy` | Fe-3Cr, Fe-12Cr, Fe-18Cr, alpha-Fe | 391 |
+| `nickel_alloy` | Ni, Inconel, BAM-11, Alloy 718 | 182 |
+| `beryllium` | Be, BeO | 181 |
+| `refractory_metal` | Mo, Mo-Re, Cr, Nb-1Zr | 127 |
+| `carbon_graphite` | H451, IG-110, graphite | 114 |
+| `nanolaminate` | Cu-Fe, Cu-Nb (Helion magnet candidates) | 101 |
+| `titanium_alloy` | Ti-6Al-4V | 86 |
+| `HTS_tape` | REBCO, YBCO | 77 |
+| `max_phase` | Ti₂AlC, Ti₃SiC₂ | 30 |
+| `zirconium_alloy` | Zircaloy | 18 |
+
+---
+
+## Property Coverage
+
+| Property | Total records | Irradiated | Unirradiated |
+|---|---|---|---|
+| Yield strength (MPa) | 4,868 | 2,267 | 2,601 |
+| UTS (MPa) | 3,681 | 1,757 | 1,924 |
+| Elongation (%) | 1,920 | 1,920 | — |
+| Volumetric swelling (%) | 2,066 | 2,066 | — |
+| Hardness | 1,496 | — | — |
+| Fracture toughness (MPa√m) | 1,262 | 1,262 | — |
+| DBTT (K) | 510 | 510 | — |
+| Void diameter (nm) | 796 | 796 | — |
+| Creep rate (s⁻¹) | 216 | 216 | — |
+| Electrical resistivity (µΩ·cm) | 254 | 254 | — |
+
+---
+
+## Intended Uses
+
+### ✅ Gaussian Process property predictors
+
+Best-supported GP training targets (complete: dose + temp + property all present):
+
+| Material class | GP rows | Input → Target |
+|---|---|---|
+| RAFM steels | 457 | dose_dpa, irradiation_temp_C → yield_strength_mpa_irradiated |
+| Vanadium alloys | 371 | dose_dpa, irradiation_temp_C → yield_strength_mpa_irradiated |
+| Copper alloys | 156 | dose_dpa, irradiation_temp_C → yield_strength_mpa_irradiated |
+| Tungsten | 109 | dose_dpa, irradiation_temp_C → yield_strength_mpa_irradiated |
+
+```python
+import pandas as pd
+df = pd.read_parquet("train.parquet")
+
+# RAFM steel GP dataset
+rafm = df[
+    (df["material_class"] == "RAFM_steel") &
+    df["yield_strength_mpa_irradiated"].notna() &
+    df["dose_dpa"].notna() &
+    df["irradiation_temp_C"].notna()
+][["dose_dpa", "irradiation_temp_C", "yield_strength_mpa_irradiated"]]
+```
+
+### ✅ Radiation damage world model
+
+142 records with both `yield_strength_mpa_unirradiated` and `yield_strength_mpa_irradiated` in the same row. Paired format: (state_before, action) → state_after.
+
+### ✅ Bayesian active learning
+
+Load as GP prior data in [FusionBAL](https://github.com/khalizo/fusionbal) to recommend which irradiation experiments to run next. Replicates the Digilab Bristol result reducing V-4Cr-4Ti experiments from 30 to ~9.
+
+### ✅ Materials NLP and information extraction
+
+Each record traces to a specific ORNL report page. Useful for training materials NER models or evaluating LLM extraction accuracy.
+
+### ❌ Deep learning / neural networks
+
+Not enough data per material class (would need 5,000+ per class). Use GPs.
+
+---
+
+## Known Limitations
+
+1. **Extraction accuracy is estimated, not fully verified.** All 20,318 ORNL records are LLM-extracted. Spot-checked against source PDFs (EUROFER97 RT yield = 580 MPa ✓, W yield range ✓) but not systematically validated. `reviewed_by_human = True` only for SDC-IC records.
+
+2. **Sparse paired data.** Only 142 records have both irradiated and unirradiated yield strength in the same row. Most ORNL papers report one or the other, not both.
+
+3. **Fission proxy, not fusion neutrons.** All ORNL data uses fission reactor spectra (HFIR, BOR-60, EBR-II). No DT fusion neutron irradiation data exists — IFMIF is not yet operational. Fission data is the standard proxy for fusion conditions.
+
+4. **202 high-dose records flagged.** Records with `dose_dpa > 150` are flagged in the source data for expert review. Doses 150–500 dpa are achievable in fast reactors (EBR-II, FFTF); values previously >500 dpa have been nulled.
+
+5. **Cryogenic temperatures are correct.** Records with `irradiation_temp_C < -50°C` represent real cryogenic irradiation experiments (10–196 K = liquid helium to liquid nitrogen). These are not unit errors.
+
+6. **Material name fragmentation.** 68 distinct RAFM steel variants are stored separately. For class-level GP training, group by `material_class` rather than `material_name`.
+
+7. **`other` class (20%).** 2,391 records have ambiguous or multi-material names that couldn't be classified. Filter with `material_class != "other"` to work with the classified 80%.
+
+---
+
+## Comparison to Existing Databases
+
+| Database | Records | Irradiation data | Access | ML-ready |
+|---|---|---|---|---|
+| **FusionMatDB** (this dataset) | **22,269** | **Yes — core focus** | **Open** | **Yes** |
+| EUROfusion EDDI | ~3,000 | Yes | EU consortium only | No |
+| MatDB4Fusion (KIT) | 353 | No (baseline only) | Public CSV | Partial |
+| JRC ODIN | >20,000 | Some | Tiered | No |
+| ITER MPH | Unknown | Yes | Closed | No |
+
+---
+
+## Related Projects
+
+- **FusionBAL** — Bayesian active learning for fusion experiment planning (loads this dataset as GP prior)
+- **FusionUQ** — Uncertainty quantification for MACE-MP-0 ML interatomic potentials (calibrates against this dataset)
+
+---
+
+## Citation
+
+```bibtex
+@dataset{fusionmatdb2026,
+  title  = {FusionMatDB: An Open-Source Fusion Irradiation Materials Database},
+  author = {Khalizo},
+  year   = {2026},
+  url    = {https://huggingface.co/datasets/khalizo/fusionmatdb},
+  note   = {Extracted from ORNL Fusion Materials Program semiannual progress
+            reports (1990--2024) using Gemini 3 Flash vision on Vertex AI.}
+}
+```
+
+Source data attribution:
+- **ORNL Fusion Materials Program reports**: U.S. Department of Energy, public domain. Available at https://fmp.ornl.gov/semiannual-progress-reports/
+- **SDC-IC Material Library**: ITER Structural Design Criteria, EUPL-1.2. Available at https://github.com/Structural-Mechanics/SDC-IC-Material-Library

fusionmatdb.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4b8651f4629b2f9b84c0f25d27a641fba9cc5d3a84d899b758d65f2cf0838455
+size 400938

test.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:11cfb075385470d70f0e72fe2e0c2b3c70b053e584eb0a04a2c31a1330ecbd87
+size 92113

train.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d9f0cac64c5d31cc441fd846a7c4399683e0c8472f0ecddd4ff39543c9c1c4bd
+size 418102

validation.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:98bcfe3212972eab4c34b6ffe3f897573318bdfbe20d3e7c6d40cf4b14fc224f
+size 91771