dataset page: layout, loading, splits, manifest; protocol pointer

README.md

@@ -43,75 +43,94 @@ Code, benchmark protocol, and leaderboards:
 
 Autoregressive next-step surrogate of a 2D SPH notched concrete beam under drop-weight impact (ADR-0026). Covers 3 beam widths, 3 impactor shapes, 3 notch positions, and 4 velocities. Three bodies: the K&C concrete beam (part 1) is the predicted deformable; the steel impactor (part 2) and the two support blocks (part 3) are protocol-kinematic (ADR-0026) — driven by ground truth during rollout (both move: the impactor decelerates from its case velocity to ~10-20% on contact, the supports displace a few mm), excluded from the training loss and from position/strain metrics, with both QoIs restricted to concrete particles.
 
+## Dataset summary
+
 - Solver: LS-DYNA (SPH; erosion: no)
 - Loading: drop-weight impact, initial velocity 40-160 m/s, impactor cross-sections plate/disk/rod (case names Rectangular/Sphere/Bullet)
+- Geometry: 2D SPH notched beam, H 80 x W {320,480,640} mm
+- Materials: *MAT_CONCRETE_DAMAGE_REL3 (K&C; density 2.4e-6 kg/mm3); *MAT_PLASTIC_KINEMATIC
 - Source units: kg-mm-ms (files are strict SI, ADR-0012)
 - Cases: 110 (train 88, val 8, test_interp 12, probe 2)
-- Particles per case: 4264-12966; 502 frames at 0.001 ms
+- Particles per case: 4264-12966; 502 frames at 0.001 ms; 24.9 GB on disk
+- Fields: node/displacement, node/velocity, node/acceleration, sph/stress, sph/strain, sph/strain_rate, sph/effective_plastic_strain, sph/pressure, sph/density, sph/internal_energy, sph/mass, sph/radius, sph/n_neighbors, sph/deletion, global/kinetic_energy, global/internal_energy, global/total_energy
 - Provenance: LS-DYNA parametric sweep (3 widths x 3 shapes x 3 notches x 4 velocities) produced by Curtin collaborators — extends the published 81-specimen drop-weight study (plate/disk/rod impactors at 80/120/160 m/s) with a 40 m/s velocity level; benchmark protocol per ADR-0026.
 - License: CC BY 4.0
 
-## Task
-
-autoregressive transition (ADR-0026). Auxiliary target: `max_principal_strain` (-). Models advance the particle state autoregressively from a short ground-truth prefix and are scored on the full predicted rollout.
-
-## Evaluation criteria
+## Files
 
-- Protocol (benchmark-owned, ADR-0032, ADR-0035): 6 input frames, horizon frames [6, 250) of 502 scored (250 µs, ADR-0039); full-length diagnostic, scored at native output times.
-- Protocol rationale: Confirmed (maintainer, 2026-07-20): input_frames = 6 gives C = 5 input velocities (input_frames - 1), the GNS reference history length — the velocity budget is the criterion, not a rigid prefix. The timeline analysis (2026-07-20, on the DUG data copy) shows impact contact from frame 0, so the observed window takes in the first 6 us of contact; accepted. Scored horizon (ADR-0039): rollout metrics and QoIs are scored on frames [input_frames, 250) (250 µs). Internal energy reaches 99% of its final value by frame 77-213 (width-dependent); the remaining frames are ballistic separation and elastic ringing, which dominated full-horizon RMSE (half the final error accrued after frame 301 in baseline rollouts) while adding no fracture physics. The full 502-frame error curve remains a non-leaderboard long-horizon diagnostic. The cracked_fraction QoI threshold 0.01 is a declared protocol definition (ADR-0029, amended 2026-08-06): the SPH source model has no erosion or crack criterion; a 221-case sweep shows the GT fraction shifts ~0.05 mean per case across the factor-2 band [0.005, 0.02], and frame-249 vs frame-501 fractions are nearly identical (0.305 vs 0.317 mean), corroborating the 250 us horizon. Probe split (characterisation, 2026-08-15; height axis added 2026-08-27): the probe cases are out-of-distribution on up to FOUR axes at once — beam width (400/800 mm) and impactor velocity (140/60 m/s) both off the training grids ({320,480,640} mm; {40,80,120,160} m/s), beam height H=100 mm on the 800 mm case (every grid case is H=80; verified from canonical frame-0 extents, 2026-08-27), and, decisively, an OFF-CENTRE impact. All 108 train/val/test_interp cases are struck exactly at midspan (impact offset 0.0 mm, every notch a/b/c variant and width); the probe impacts land ~6% off-centre — a loading mode absent from training entirely. Probe scores therefore measure graceful failure on a genuinely new loading configuration, not ordinary interpolation: global-attention operators mis-localise the response to the learned midspan prior, while relative-position message-passing (MGN/CGN) degrades more gracefully.
-- Metrics: headline is the pooled space+time relative L2 (displacement + max_principal_strain); also reported are position/max_principal_strain RMSE (physical units) and one-step RMSE, plus the quantities of interest below.
-- Quantities of interest: midspan_deflection_peak, cracked_fraction.
+- `<case_id>.h5` — one HDF5 file per case; the file name is the case id (layout below).
+- `card.json` — machine-readable card metadata (ADR-0027): the facts above plus the split sizes.
+- `README.md` — this file; `LICENSE-*.txt` — the data licence (CC BY 4.0).
 
-## Leaderboard
+## Manifest and input decks (Hugging Face mirror)
 
-The headline metric is the pooled space+time relative L2 (↓ lower is better); the Scheme column is the prediction scheme. RMSE and quantities of interest are shown for the in-distribution `test_interp` split.
+- `cases.csv` — one row per `.h5`: `case_id`, `split` (`held_aside` for files shipped outside the protocol splits), the loading/geometry parameters parsed from the id, `n_nodes`, `n_frames`, `file_bytes`, `sha256` (integrity manifest; also what the Dataset Viewer shows).
+- `decks/<case_id>.k` — the LS-DYNA input deck of every case (also embedded verbatim in each file's `metadata/source_deck`); re-running a deck regenerates the raw output the adapter converts to canonical HDF5.
+- Case ids: `NB-I-<W>-<Shape>-<n>-<V>` — beam width W mm (height 80 mm), impactor `Rectangular` / `Sphere` / `Bullet` (plate / disk / rod cross-section), notch position code n ∈ {a, b, c}, impact speed V m/s; the two off-grid probes are `S_<H>_<W>_V<V>_<label>` (beam height H mm, width W mm, disk impactor at V m/s).
 
-_Headline — pooled relative L2 (↓ better)_
+## HDF5 layout
 
-| Method | Scheme | interp·disp | interp·aux | probe·disp | probe·aux |
-|---|---|---|---|---|---|
-| MGN | autoregressive | 0.2245 | 0.6988 | 0.7672 | 1.255 |
-| CGN | autoregressive | 0.2827 | 0.5876 | 0.5905 | 0.8535 |
-| Transolver | time-conditioned | 0.03517 | 0.2294 | 1.047 | 1.236 |
-| Transolver++ | time-conditioned | 0.03699 | 0.2396 | 1.01 | 1.263 |
+One HDF5 file per case, readable with `h5py` or any HDF5 tool. Every quantity is stored in strict SI (m, s, kg, Pa, J) regardless of the solver's `kg-mm-ms` source convention. Small scalars are HDF5 attributes; arrays are datasets (float64 geometry and time, float32 response, int64 ids); response arrays are gzip-compressed and chunked along the frame axis, so one frame or transition can be read without loading the whole trajectory. Shapes below use N nodes, P SPH particles, E elements, T frames and d = `metadata.dimension`; the exact schema version is the `schema_version` attribute (ADR-0013 — 0.2.0 readers read 0.1.0 files unchanged, ADR-0042).
 
-_Trajectory error — RMSE_
-
-| Method | Scheme | interp·pos (mm) | interp·strain |
+| Path | Shape | Dtype | Content |
 |---|---|---|---|
-| MGN | autoregressive | 0.1984 | 0.01997 |
-| CGN | autoregressive | 0.2497 | 0.01697 |
-| Transolver | time-conditioned | 0.03365 | 0.006467 |
-| Transolver++ | time-conditioned | 0.03509 | 0.006714 |
+| `metadata` (attrs) | — | — | `case_id`, `dataset_id`, `dimension`, `schema_version`, `source_units`, `units_convention` (= `SI`) |
+| `metadata/provenance` (attrs) | — | — | `solver_name`, `solver_version`, `generation_date` |
+| `metadata/source_deck` | scalar | str | the complete solver input deck, verbatim |
+| `nodes/coords` | (N, d) | f64 | initial node coordinates [m] |
+| `nodes/node_id` | (N,) | i64 | solver node ids |
+| `materials/{canonical_model, source_model, source_params, material_id}` | (M,) | str / i64 | material models; `source_params` is the solver's material card as JSON |
+| `response/time/t` | (T,) | f64 | output times [s]; frame 0 is the initial state |
+| `elements/sph/connectivity` | (P, 1) | i64 | particle → node index (0-based) |
+| `elements/sph/{element_id, part_id}` | (P,) | i64 | solver element id, part id |
+| `elements/<other>/…` | (E, n), (E,) | i64 | any further element group (e.g. a single rigid-wall / boundary `shell`) follows the same connectivity, element_id, part_id pattern |
+| `response/node/{displacement, velocity, acceleration}` | (T, N, d) | f32 | [m], [m/s], [m/s²] |
+| `response/element/sph/{stress, strain, strain_rate}` | (T, P, 6) | f32 | Voigt (xx, yy, zz, xy, yz, zx): [Pa], [–], [1/s] — six components even for 2D cases |
+| `response/element/sph/{pressure, density, mass, internal_energy}` | (T, P) | f32 | [Pa], [kg/m³], [kg], [J] |
+| `response/element/sph/{effective_plastic_strain, radius, n_neighbors, deletion}` | (T, P) | f32 | [–], smoothing length [m], neighbour count, 0/1 deletion flag |
+| `response/element/<other>/…` | (T, E, …) | f32 | per-element response of any further element group |
+| `response/global/{kinetic_energy, internal_energy, total_energy}` | (T,) | f32 | [J] |
 
-_Quantities of interest (MAE)_
+`sph/stress` and `sph/strain` are 6-component Voigt tensors; scalar targets are loader-derived (see the card's aux field).
 
-| Method | Scheme | interp·midspan_deflection_peak (mm) | interp·cracked_fraction |
-|---|---|---|---|
-| MGN | autoregressive | 0.5842 | 0.1081 |
-| CGN | autoregressive | 0.5843 | 0.1892 |
-| Transolver | time-conditioned | 0.04067 | 0.0212 |
-| Transolver++ | time-conditioned | 0.04696 | 0.02438 |
+## Loading
 
-## Baseline details
+Plain HDF5 — nothing beyond `h5py` is needed:
 
-**MGN** (mgn, 2026-08-17, commit `59d5786`)
+```python
+import h5py
+
+with h5py.File("<case_id>.h5") as f:
+    t = f["response/time/t"][:]                # (T,) s
+    x0 = f["nodes/coords"][:]                  # (N, d) m
+    u = f["response/node/displacement"]        # (T, N, d) m, chunked along T
+    u_last = u[-1]                             # one frame, no full read
+    sig = f["response/element/sph/stress"][:]  # (T, P, 6) Pa, Voigt
+```
 
-*Native MeshGraphNets (ADR-0047 baseline, ADR-0049 repair): autoregressive next-step on the SPH particle set as a mesh, val-selected model-best-212000.pt, seed 2 of the s1-s2 pair, run notch-mgn-base-s2. PROVISIONAL (ADR-0044/0045). On test_interp it roughly matches CGN (displacement relative L2 0.22 vs 0.28) and both trail Transolver ~6x; on the off-centre triple-OOD PROBE the relative-position message passing degrades more gracefully than the global-attention operator (disp 0.77 vs Transolver 1.05), though CGN is best there (0.59). Pooled relative L2 headline (ADR-0055) from the 2026-08-17 re-eval.*
+Or through StructBench's loader, which returns the ML working frame (positions in mm; `max_principal_strain` in dimensionless) with the auxiliary target derived on the fly:
 
-**CGN** (cgn, 2026-07-24, commit `5956d81`, checkpoint: `models/notch_beam_2d_impact/cgn-5956d81/model-best-186000.pt` — private archive; publication parked)
+```python
+from structbench.datasets import load_case_trajectory
 
-*Single-scale CGN (ADR-0034) on the ADR-0039 §4 truncated recipe with the ADR-0038 strain knobs (train_frames 250, aux_tail_weight 3, asinh aux transform at scale 0.01; hidden 192 / 15 MP steps / 2-layer node MLP, noise_std 0.01, batch 4) at 250k steps; seed 1 of the 2026-07-24 h250c pair (seeds 1-2), val-selected checkpoint model-best-186000.pt (186k), one A100-80GB, ~80 h. Extending the same recipe from 200k to 250k steps cut seed-mean test rollout position RMSE 21% and deflection MAE 30% while validation strain RMSE stayed flat (0.0173 -> 0.0163): the extra budget buys kinematics, not damage-field quality. Caveats: the model over-predicts cracked fraction on the reviewed cases (crack MAE 0.19 vs sibling seed s2's 0.13, the one metric s2 wins); the off-grid probe case S_80_400_V140 is this seed's worst rollout (0.59 mm scored vs 0.40 for s2); predictions break the mirror symmetry of centered-notch cases while the ground truth stays symmetric (2026-07-24 finding); full-horizon (502-frame) rollout position RMSE is 0.87 mm on test_interp - diagnostic only, not scored. Relative L2 (rollout_rel_l2_disp/aux) is the pooled space+time headline (ADR-0055), added 2026-08-16 from a re-eval on this checkpoint; RMSE reproduced to <1%, so the blessed RMSE/QoI values are unchanged.*
+traj = load_case_trajectory("<case_id>.h5", aux_field="max_principal_strain")
+traj.positions   # (T, P, d) float32, mm
+traj.aux         # (T, P) float32, dimensionless
+traj.time        # (T,) float64, s
+```
 
-**Transolver** (transolver, 2026-08-16, commit `59d5786`)
+## Splits
 
-*Native time-conditioned Transolver (ADR-0054): history-free independent-time-query, no rollout accumulation, so one-step is N/A. Seed 2 of the s1-s2 pair, val-selected model-best-230000.pt, run notch-transolver-tc-s2. PROVISIONAL (ADR-0044/0045). On test_interp it is the strongest baseline (~8x lower displacement relative L2 than CGN); on the PROBE it fails hard (relative L2 > 1) - the off-centre triple-OOD case (ADR-0026 amendment) where the global-attention operator mis-localises the response to the learned midspan prior, while CGN's relative-position message passing degrades more gracefully (probe disp 0.59 vs 1.05). Pooled relative L2 headline (ADR-0055) from the 2026-08-16 re-eval.*
+The benchmark's fixed split assignment, by case id (the file name without `.h5`):
 
-**Transolver++** (transolver_plus, 2026-08-18, commit `5b84119`)
+- **train** (88): `NB-I-320-Bullet-a-120`, `NB-I-320-Bullet-a-160`, `NB-I-320-Bullet-a-40`, `NB-I-320-Bullet-b-160`, `NB-I-320-Bullet-b-40`, `NB-I-320-Bullet-c-120`, `NB-I-320-Bullet-c-160`, `NB-I-320-Bullet-c-40`, `NB-I-320-Bullet-c-80`, `NB-I-320-Rectangular-a-120`, `NB-I-320-Rectangular-a-160`, `NB-I-320-Rectangular-a-40`, `NB-I-320-Rectangular-a-80`, `NB-I-320-Rectangular-b-160`, `NB-I-320-Rectangular-b-40`, `NB-I-320-Rectangular-b-80`, `NB-I-320-Rectangular-c-160`, `NB-I-320-Rectangular-c-40`, `NB-I-320-Rectangular-c-80`, `NB-I-320-Sphere-a-160`, `NB-I-320-Sphere-a-40`, `NB-I-320-Sphere-b-160`, `NB-I-320-Sphere-b-40`, `NB-I-320-Sphere-c-120`, `NB-I-320-Sphere-c-160`, `NB-I-320-Sphere-c-40`, `NB-I-320-Sphere-c-80`, `NB-I-480-Bullet-a-120`, `NB-I-480-Bullet-a-160`, `NB-I-480-Bullet-a-40`, `NB-I-480-Bullet-a-80`, `NB-I-480-Bullet-b-120`, `NB-I-480-Bullet-b-160`, `NB-I-480-Bullet-b-40`, `NB-I-480-Bullet-b-80`, `NB-I-480-Bullet-c-120`, `NB-I-480-Bullet-c-160`, `NB-I-480-Bullet-c-40`, `NB-I-480-Bullet-c-80`, `NB-I-480-Rectangular-a-120`, `NB-I-480-Rectangular-a-160`, `NB-I-480-Rectangular-a-40`, `NB-I-480-Rectangular-a-80`, `NB-I-480-Rectangular-b-160`, `NB-I-480-Rectangular-b-40`, `NB-I-480-Rectangular-c-120`, `NB-I-480-Rectangular-c-160`, `NB-I-480-Rectangular-c-40`, `NB-I-480-Rectangular-c-80`, `NB-I-480-Sphere-a-160`, `NB-I-480-Sphere-a-40`, `NB-I-480-Sphere-a-80`, `NB-I-480-Sphere-b-160`, `NB-I-480-Sphere-b-40`, `NB-I-480-Sphere-b-80`, `NB-I-480-Sphere-c-160`, `NB-I-480-Sphere-c-40`, `NB-I-480-Sphere-c-80`, `NB-I-640-Bullet-a-120`, `NB-I-640-Bullet-a-160`, `NB-I-640-Bullet-a-40`, `NB-I-640-Bullet-a-80`, `NB-I-640-Bullet-b-160`, `NB-I-640-Bullet-b-40`, `NB-I-640-Bullet-c-120`, `NB-I-640-Bullet-c-160`, `NB-I-640-Bullet-c-40`, `NB-I-640-Bullet-c-80`, `NB-I-640-Rectangular-a-160`, `NB-I-640-Rectangular-a-40`, `NB-I-640-Rectangular-a-80`, `NB-I-640-Rectangular-b-120`, `NB-I-640-Rectangular-b-160`, `NB-I-640-Rectangular-b-40`, `NB-I-640-Rectangular-b-80`, `NB-I-640-Rectangular-c-120`, `NB-I-640-Rectangular-c-160`, `NB-I-640-Rectangular-c-40`, `NB-I-640-Sphere-a-120`, `NB-I-640-Sphere-a-160`, `NB-I-640-Sphere-a-40`, `NB-I-640-Sphere-b-120`, `NB-I-640-Sphere-b-160`, `NB-I-640-Sphere-b-40`, `NB-I-640-Sphere-b-80`, `NB-I-640-Sphere-c-160`, `NB-I-640-Sphere-c-40`, `NB-I-640-Sphere-c-80`
+- **val** (8): `NB-I-320-Bullet-b-80`, `NB-I-320-Sphere-a-120`, `NB-I-320-Sphere-a-80`, `NB-I-480-Rectangular-b-80`, `NB-I-480-Sphere-c-120`, `NB-I-640-Bullet-b-120`, `NB-I-640-Rectangular-a-120`, `NB-I-640-Rectangular-c-80`
+- **test_interp** (12): `NB-I-320-Bullet-a-80`, `NB-I-320-Bullet-b-120`, `NB-I-320-Rectangular-b-120`, `NB-I-320-Rectangular-c-120`, `NB-I-320-Sphere-b-120`, `NB-I-320-Sphere-b-80`, `NB-I-480-Rectangular-b-120`, `NB-I-480-Sphere-a-120`, `NB-I-480-Sphere-b-120`, `NB-I-640-Bullet-b-80`, `NB-I-640-Sphere-a-80`, `NB-I-640-Sphere-c-120`
+- **probe** (2): `S_100_800_V60_extrapolation`, `S_80_400_V140_intrapolation`
 
-*Transolver++ (ADR-0057): the native time-conditioned Transolver with both eidetic-state knobs ON - per-point adaptive slice temperature + train-only Gumbel Rep-Slice reparameterisation. Seed 2 of the s1-s2 pair (val-selected model-best-250000.pt, run notch-transolver-tcpp-s2), seed-matched to the plain Transolver entry above (also seed 2). PROVISIONAL method comparison (ADR-0046), not a blessed baseline. It is neutral-to-worse than plain Transolver: test_interp displacement is ~5% worse (0.03699 vs 0.03517) and strain ~4% worse; on the off-centre triple-OOD PROBE displacement is marginally better (1.01 vs 1.05) but strain is worse (1.263 vs 1.236) and both still fail hard (relative L2 > 1). No robust generalisation gain. Consistent with Transolver++ being designed for million-scale geometries: its eidetic-state edits do not help these O(10^4)-particle impact meshes. Pooled relative L2 headline (ADR-0055).*
+## Benchmark protocol
 
-## Using this archive
+This archive backs the **NotchBeam2D-Impact** benchmark in StructBench. Task: autoregressive transition (ADR-0026); auxiliary target `max_principal_strain` (dimensionless); 6 input frames, horizon frames [6, 250) of 502 scored (250 µs, ADR-0039); full-length diagnostic, scored at native output times; quantities of interest: midspan_deflection_peak, cracked_fraction. The full evaluation protocol and its rationale, the baseline recipes and checkpoints, and the current leaderboard live on the benchmark page in the code repository — <https://github.com/qilinli/StructBench/blob/main/docs/benchmarks/notch_beam_2d_impact.md> — so the numbers have a single home. To train a baseline on this archive:
 
 ```bash
 pip install structbench  # or: pip install -e . from the repo
@@ -119,10 +138,6 @@ structbench-train --mode train --config configs/notch_beam_2d_impact/cgn.toml \
     --data-root /path/to/this/folder --out runs/notch_beam_2d_impact-cgn
 ```
 
-Machine-readable metadata: `card.json` alongside this file.
-`sph/stress` and `sph/strain` are 6-component Voigt tensors; scalar targets are loader-derived (see the card's aux field).
-Consume with `structbench.datasets.load_case_trajectory` or any HDF5 reader (layout per ADR-0013).
-
 ## References
 
 - **MGN** — Pfaff, T., Fortunato, M., Sanchez-Gonzalez, A., & Battaglia, P. W. (2021). Learning Mesh-Based Simulation with Graph Networks. *ICLR*. https://arxiv.org/abs/2010.03409