Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,121 @@
+# WBM (Matbench Discovery test set)
+
+## Overview
+WBM (after authors **W**ang, **B**otti & **M**arques) is a set of **256,963 periodic
+inorganic crystals** generated by five successive rounds of chemical-similarity
+**elemental substitution** of Materials Project prototypes, each relaxed with DFT
+(VASP, PBE / PBE+U). It is the prospective **test set of
+[Matbench Discovery](https://matbench-discovery.materialsproject.org/)** (MBD): given the
+*unrelaxed* structure, a model must predict the *relaxed* formation energy and hence the
+energy above the MP convex hull, i.e. thermodynamic **stability**. Chemical space is broad
+and distinct from MP (85 elements incl. lanthanides/actinides, ternary-dominated ≈ 75 %).
+Every material is stored with **two geometries** — the initial (unrelaxed) and the
+DFT-relaxed endpoint (WBM ships no relaxation trajectory). It is a single-file, periodic,
+labelled crystal dataset — **not** a conformer ensemble.
+
+## Statistics
+<!-- from wbm_stats.json (atom range + element set are exact; distributions from an 8,000-structure sample) -->
+
+| property | value |
+|---|---|
+| structures (test) | 256,963 (215,488 unique prototypes; 42,825 stable) |
+| atoms per structure (min / mean / max) | 2 / 7.76 / 100 |
+| unit-cell edge, Å (sampled min / max) | 2.26 / 30.3 (bounding-box diagonal) |
+| distinct elements | 85 |
+| element set | H, Li, Be, B, C, N, O, F, Na, Mg, Al, Si, P, S, Cl, K, Ca, Sc–Zn, Ga, Ge, As, Se, Br, Rb, Sr, Y, Zr–Cd, In, Sn, Sb, Te, I, Xe, Cs, Ba, La–Lu, Hf–Hg, Tl, Pb, Bi, Ac, Th, Pa, U, Np, Pu |
+| `formation_energy` (`y`, eV/atom) | −4.56 … 4.74 |
+| `e_above_hull` (eV/atom) | −2.73 … 6.48 |
+| `band_gap` (eV) | 0 … 8.07 |
+| `volume` (ų) | 12.5 … 3740 |
+| substitution rounds (`wbm_step` 1–5) | 61,466 / 52,755 / 79,160 / 40,314 / 23,268 |
+
+Distributions in the figure are computed from a random sample of 8,000 structures; the
+atom-count range and element set are exact (from the collated tensors).
+
+## Usage / loading
+```python
+from StructureCloud.Datasets import WBM
+
+ds = WBM()                                  # split='test', mode='relaxed', label='formation_energy'
+d  = ds[0]                                   # torch_geometric.data.Data
+# d.pos [N,3] Å, d.frac_pos [N,3], d.cell [1,3,3] Å, d.z [N], d.pbc [1,3] bool,
+# d.id 'wbm-1-1', d.y = formation energy (eV/atom)
+
+# attach more labels, restrict to unique prototypes:
+ds = WBM(label='e_above_hull', extra_labels=['band_gap', 'wbm_step'], unique_only=True)
+
+# the Matbench Discovery benchmark pairing (predict relaxed target from unrelaxed input):
+from torch_geometric.loader import DataLoader
+ds = WBM(mode='both', label='formation_energy')
+for input_batch, target_batch in DataLoader(ds, batch_size=32):
+    ...   # input_batch = unrelaxed geometry (no label); target_batch = relaxed + y
+```
+
+Returned sample keys (single-structure modes `'relaxed'` / `'unrelaxed'`):
+- `pos` — `[N, 3]` float32 (Å); Cartesian, computed as `frac_pos @ cell`
+- `frac_pos` — `[N, 3]` float32; fractional coordinates
+- `cell` — `[1, 3, 3]` float32 (Å); lattice matrix. `pbc` — `[1, 3]` bool (all True)
+- `z` — `[N]` int64; atomic numbers
+- `id` — `str`; material id (e.g. `wbm-3-12345`). `y` — the selected label (omitted if `label=None`)
+
+**Modes** (`mode=`): `'relaxed'` (default) → only the relaxed geometry; `'unrelaxed'` → only
+the initial geometry (same key schema, mutually stackable with `'relaxed'`); `'both'` → a
+tuple `(input, target)` — unrelaxed input **without** label, relaxed target **with** labels —
+which a plain PyG `DataLoader` collates into `(input_batch, target_batch)`. Labels are always
+properties of the DFT-relaxed structure regardless of mode. `'both'` returns a tuple, so it is
+**not** compatible with `StackedDataset`. `unique_only=True` restricts to the 215,488
+`unique_prototype` rows. Loading is memory-mapped (`mmap=True` default; `mmap=False` loads into RAM).
+
+## Preprocessed (collated) format
+Ships as one memory-mappable `preprocessed/test.pt` holding tensors only:
+- `frac_pos` / `init_frac_pos` `[ΣN, 3]` float32 — relaxed & initial fractional coordinates, one
+  atom offset table `ptr` `[M+1]`; `cell` / `init_cell` `[M, 3, 3]` float32 (Å).
+- `z` / `init_z` `[ΣN]` int64 — atomic numbers for each geometry. **Relaxation regroups atoms by
+  element**, so the relaxed and initial atom orderings differ (same composition); each geometry is
+  stored self-consistently with its own `z`, and there is no atom-index correspondence between them.
+- per-structure label tensors `[M]` (all 18 summary columns) + derived `wbm_step`; packed strings
+  `material_id`, `formula`, `protostructure_spglib`, `protostructure_spglib_initial_structure`.
+
+## Labels
+All labels are properties of the **DFT-relaxed** structure, per-structure (regression), no NaNs.
+
+| loader name | collated column | units |
+|---|---|---|
+| `formation_energy` (default `y`) | `e_form_per_atom_mp2020_corrected` | eV/atom |
+| `e_above_hull` | `e_above_hull_mp2020_corrected_ppd_mp` | eV/atom (stability: stable if ≤ 0) |
+| `band_gap` | `bandgap_pbe` | eV |
+| `e_form_uncorrected`, `e_form_wbm` | `e_form_per_atom_uncorrected`, `e_form_per_atom_wbm` | eV/atom |
+| `e_above_hull_wbm` | `e_above_hull_wbm` | eV/atom |
+| `energy`, `energy_from_cse` | `uncorrected_energy`, `uncorrected_energy_from_cse` | eV (total) |
+| `e_correction`, `e_correction_legacy` | `e_correction_per_atom_mp2020` / `..._mp_legacy` | eV/atom |
+| `volume`, `n_sites`, `wbm_step` | `volume`, `n_sites`, `wbm_step` | ų, count, 1–5 |
+| `unique_prototype` | `unique_prototype` | bool |
+| `relax_displacement` | `site_stats_fingerprint_init_final_norm_diff` | fingerprint distance |
+| `formula`, `id`, `protostructure`, `protostructure_initial` | string fields | — |
+
+`e_form_per_atom_mp2020_corrected` applies the **MP2020 compatibility corrections** (anion/GGA+U
+corrections) so WBM energies are on the same scale as the MP convex hull used to score stability.
+
+## Splits
+Single split `'test'` — WBM is an evaluation set (MBD holds it out; the MBD training set is MPtrj).
+`unique_only=True` selects the 215,488 unique-prototype subset. `wbm_step` (1–5) records which of
+the five substitution rounds produced each structure.
+
+## Preprocessing & cleaning
+- Coordinates are **fractional** in the source (pymatgen `Structure` / `ComputedStructureEntry`);
+  stored as `frac_pos` + `cell`, with Cartesian `pos = frac_pos @ cell` (Å) computed on access.
+- **Both** geometries (relaxed + initial) are kept; the intermediate ionic relaxation steps are not
+  in the WBM release. Relaxed and initial atom orderings differ (see collated-format note).
+- All 256,963 structures are kept (no filtering); the summary CSV, relaxed CSE JSONL, and initial
+  structure JSONL are 1:1 aligned on `material_id`. No unit conversion (labels already eV / eV·atom⁻¹).
+
+## Data distribution
+![atom / cell / element / label distributions](./wbm_histogram.png)
+
+## Sources & citation
+- Data (figshare, CC-BY-4.0), via `matbench_discovery.data.DataFiles`:
+  `wbm/2023-12-13-wbm-summary.csv.gz`, `wbm/2022-10-19-wbm-computed-structure-entries.jsonl.gz`
+  (relaxed), `wbm/2022-10-19-wbm-init-structs.jsonl.gz` (initial).
+- Wang, Botti & Marques, *A high-throughput framework for determining synthesizability*,
+  npj Comput. Mater. 7, 12 (2021). DOI 10.1038/s41524-020-00481-6.
+- Riebesell et al., *Matbench Discovery*, Nat. Mach. Intell. (2025). arXiv:2308.14920.