File size: 8,381 Bytes
45590a8 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 | {
"name": "Taylor2D-Impact",
"version": "0.1",
"description": "Autoregressive next-step surrogate of a 2D SPH copper bar under Taylor impact against a rigid wall (ADR-0019).",
"provenance": "LS-DYNA parametric sweep (3 bar lengths x 11 impact velocities) produced by Curtin collaborators; benchmark protocol per ADR-0019. One extra Convergence run is held aside for a mesh-resolution check.",
"data_license": "CC BY 4.0",
"solver": "LS-DYNA",
"discretisation": "SPH",
"materials": [
"*MAT_ELASTIC_PLASTIC_HYDRO",
"*EOS_GRUNEISEN"
],
"erosion": false,
"loading": "rigid-wall impact; initial velocity 100-200 m/s",
"source_units": "g-mm-ms",
"geometry": "2D bar, 20 mm x {60, 80, 100} mm",
"n_cases": 33,
"splits": {
"train": 21,
"val": 3,
"test_interp": 6,
"test_extrap": 3
},
"task": "autoregressive transition (ADR-0019)",
"aux_field": "von_mises_stress",
"aux_unit": "MPa",
"qois": [
"final_length",
"mushroom_width",
"peak_von_mises",
"t_peak_von_mises"
],
"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"
],
"particles_per_case": "4800-8000",
"n_frames": 152,
"output_dt_ms": 0.002,
"input_frames": 6,
"protocol_rationale": "input_frames = 6 gives the model C = 5 input velocities (input_frames - 1), the GNS reference history length (Sanchez-Gonzalez et al. 2023); under ADR-0035 the model observes exactly these 6 ground-truth frames to seed the rollout, with no constant-velocity history backfill. GT timeline analysis over all 33 cases (2026-07-05, python -m structbench.benchmarks.timeline; evidence table in docs/timelines/taylor_impact_2d.md): the rod is in free flight until first wall contact near frame 7, so a 6-frame observed prefix takes in 0.0% of the impact in every case, while the historical init = 11 handed models the shock onset (up to 10.6% of total KE already dissipated; nonzero in every case above 100 m/s). 99% displacement settlement lands as late as 296 us of the 300 us record and the last fifth of the horizon retains 1.6-8.1% of peak mean acceleration (elastic ringing), so the full horizon is dynamically active. n_frames = 152 counts stored frames; the working trajectory drops the terminal solver-output artifact frame (ADR-0028), giving a 151-frame / 300 us protocol horizon and a scored span of frames [6, 151) -- 145 predicted frames. Predictions are scored at the native 2 us output times; peak_von_mises/t_peak_von_mises (peak of the particle-mean field, e.g. 191 MPa at 44 us in T-20-80-150 ground truth) penalize temporally coarse surrogates.",
"size_gb": 2.4,
"horizon": "full",
"eval_times": "native",
"overview": "## The problem\n\nA copper bar strikes a rigid wall head-on and *mushrooms*: the impact face\nspreads outward while a plastic wave runs back up the bar. The **Taylor impact\ntest** is a classic high-strain-rate experiment for calibrating elasto-plastic\nmaterial models, and it makes a demanding learned-surrogate target \u2014 large\nplastic deformation, a travelling stress front, strain-hardening flow with a\npressure\u2013volume equation of state, and a moving contact boundary, all inside\n~300 microseconds.\n\nStructBench ships the 2D SPH version: an LS-DYNA `*MAT_ELASTIC_PLASTIC_HYDRO`\n+ `*EOS_GRUNEISEN` copper bar, 20 mm wide and 60 / 80 / 100 mm long, fired at\na rigid wall at 100\u2013200 m/s. The task is an **autoregressive next-step\nsurrogate** \u2014 from a short ground-truth prefix the model advances the particle\nstate one output step at a time to the end of the trajectory, predicting both\nposition and the per-particle von Mises stress.\n\n\n\n*Problem setup: a copper bar (D = 20 mm, L = 60 / 80 / 100 mm) fired head-on\nat a rigid wall at 100\u2013200 m/s.*\n\n## Interpolation vs. extrapolation\n\nThe split varies **only the impact velocity** across the three fixed\ngeometries, cleanly separating the two regimes a surrogate should be judged\non: `test_interp` (130 / 170 m/s) sits inside the training band, while\n`test_extrap` (200 m/s) sits beyond it; `val` (150 m/s) only picks each run's\ncheckpoint. Everything is scored in physical units \u2014 position RMSE in mm, the\nvon Mises field in MPa \u2014 and four quantities of interest read the engineering\noutcome directly: final bar length, mushroom width, and the peak mean von\nMises stress with its timing. The reference CGN baseline is strong in\ninterpolation and degrades honestly at 200 m/s; the numbers are below.\n\n## The von Mises stress floor\n\nThe per-particle von Mises field carries an irreducible, model-independent\nnoise floor, and the aux numbers should be read with it in mind. At the\nparticle scale, the raw von Mises field differs from its own 12-neighbour\nsmoothed version by ~0.12 relative L2 \u2014 sub-particle jitter from the SPH\ndiscretisation (tensile instability, particle disorder) that no surrogate\nshould be expected to reproduce. The field is also not fully converged at the\nbenchmark resolution: the same case re-run at 4x finer resolution differs by\n~0.27 relative L2 (an over-estimate, since the finer field carries its own\nparticle noise). The strongest baselines report von Mises relative L2 ~0.17,\nonly ~0.06 above the particle-scale floor \u2014 so the achievable headroom on this\nfield is small, and a von Mises relative L2 much below ~0.12 is fitting\ndiscretisation noise, not signal. This is a property of the DATA, not any one\nmodel: it is why displacement (relative L2 ~0.009) is ~20x more predictable\nthan stress here, and it is why a spatially-smoothed von Mises companion metric\n(which removes the sub-particle noise) is the fairer reading of stress accuracy\non this benchmark.",
"figures": [
{
"path": "assets/taylor_rollout_methods.gif",
"caption": "Ground truth vs the three baselines (CGN, Transolver, MGN) on T-20-80-170 (test_interp, 170 m/s), a 20x80 mm copper bar coloured by von Mises stress over the scored rollout. Transolver (time-conditioned) tracks the mushroom-head stress band most closely; CGN is slightly diffuse in the interior; MGN smears and fragments the field and its particles penetrate the rigid wall. Transolver and MGN are provisional native baselines (ADR-0044/0045).",
"alt": "Four-panel animation: ground truth and CGN, Transolver, MGN predictions of copper-bar von Mises stress."
},
{
"path": "assets/taylor_vms_methods_170.png",
"caption": "In-distribution von Mises stress (T-20-80-170, test_interp, 170 m/s): ground truth vs the three baselines at 12 / 84 / 156 / 228 / 300 us. Transolver tracks the mushroom-head stress band most closely (rollout von Mises RMSE 30 MPa); CGN is slightly diffuse (49.5 MPa); MGN smears and fragments the field (62.4 MPa) - the ranking the leaderboard reports. Transolver and MGN are provisional (ADR-0044/0045).",
"alt": "Grid of von Mises snapshots: ground truth and CGN, Transolver, MGN predictions at five times."
}
],
"data_access": "the canonical archive is public on Hugging Face \u2014 [StructBench/taylor-impact-2d](https://huggingface.co/datasets/StructBench/taylor-impact-2d) (CC BY 4.0): one `.h5` per case (the held-aside Convergence run included, `split=held_aside` in the manifest), `cases.csv` (split, loading/geometry parameters, SHA-256 manifest) and the LS-DYNA input decks under `decks/`. Fetch one case with `hf_hub_download` or the whole archive with `snapshot_download` and point `--data-root` at it; pin the dataset repo's `v0.1.0` tag (`revision=\"v0.1.0\"` \u2014 a data release, independent of the code version) for reproducible pipelines. The maintainer's OneDrive copy remains the master (ADR-0040, amended 2026-08-28).",
"data_access_label": "public on Hugging Face \u2014 [StructBench/taylor-impact-2d](https://huggingface.co/datasets/StructBench/taylor-impact-2d)"
}
|