| { |
| "name": "Wave1D-Propagation", |
| "version": "0.1", |
| "description": "Autoregressive next-step surrogate of an elastic stress wave in a 2D SPH bar strip under initial-velocity excitation (ADR-0025). Entry tier: onboarding, tutorial, and fast CI.", |
| "provenance": "LS-DYNA parametric sweep (4 bar lengths x 4 initial velocities) produced by Curtin collaborators; benchmark protocol per ADR-0025.", |
| "data_license": "CC BY 4.0", |
| "solver": "LS-DYNA", |
| "discretisation": "SPH", |
| "materials": [ |
| "*MAT_ELASTIC (scaled toy constants: E=0.01 GPa, rho=2e-6 kg/mm3)" |
| ], |
| "erosion": false, |
| "loading": "initial velocity 1-8 m/s; elastic wave propagation; wave speed ~70.7 m/s (4-11 traversals per trajectory, by bar length)", |
| "source_units": "kg-mm-ms", |
| "geometry": "2D strip, 5 particle rows, {200, 300, 400, 500} mm x 8 mm", |
| "n_cases": 16, |
| "splits": { |
| "train": 12, |
| "val": 2, |
| "test_interp": 2 |
| }, |
| "task": "autoregressive transition (ADR-0025)", |
| "aux_field": "axial_stress", |
| "aux_unit": "MPa", |
| "qois": [ |
| "arrival_time_25", |
| "arrival_time_50", |
| "arrival_time_75", |
| "peak_stress" |
| ], |
| "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": "500-1250", |
| "n_frames": 302, |
| "output_dt_ms": 0.1, |
| "input_frames": 6, |
| "protocol_rationale": "input_frames = 6 (ADR-0035): C = 5 input velocities (input_frames - 1), the GNS reference history length; the model observes exactly these 6 ground-truth frames (indices 0-5) to seed the rollout, with no constant-velocity backfill. GT timeline analysis run 2026-07-06 (docs/timelines/wave_propagation_1d.md): a 6-frame observed prefix takes in 14.8% of initial KE worst-case (3.7% at 3 frames), and at the measured front speed ~70.7 m/s the wave reaches the first (25%) gauge about 7 frames in -- after the observed prefix -- so the arrival_time QoI is predicted, not observed. 6 is near the ceiling for this benchmark: a larger input_frames would risk seeding past first arrival.", |
| "size_gb": 0.23, |
| "horizon": "full", |
| "eval_times": "native", |
| "overview": "## The problem\n\nSet a slender elastic bar moving along its axis and stop it dead at one end:\na compression front leaps from the impact end, sweeps back along the bar at\nthe material wave speed, reflects off the free end, and keeps traversing \u2014\nthe textbook one-dimensional stress-wave problem. The particles barely move;\nthe travelling stress wave *is* the physics. That leaves a learned surrogate\nnothing to hide behind: it must carry a sharp front through hundreds of\nautoregressive steps and keep the wave speed, the reflection timing, and the\nstress amplitude honest over 4\u201311 bar traversals, depending on length.\n\nStructBench ships the 2D SPH version: an LS-DYNA `*MAT_ELASTIC` strip\n(scaled toy constants) five particle rows deep, 8 mm wide and 200 / 300 /\n400 / 500 mm long, arrested from an initial axial velocity of 1\u20138 m/s\n(wave speed ~70.7 m/s). The task is an **autoregressive next-step\nsurrogate** \u2014 from a short ground-truth prefix the model advances the\nparticle state one output step at a time over the 30 ms record, predicting\nposition and the per-particle axial stress. Axial stress is the headline\ntarget and position the sanity check \u2014 the reverse of Taylor 2D's emphasis.\n\n\n\n*Problem setup: an elastic bar strip (length L = 200\u2013500 mm) moving axially\nat 1\u20138 m/s is arrested dead at one end.*\n\n## The entry tier\n\nThis is the platform's entry benchmark \u2014 16 linear-elastic cases, 0.23 GB \u2014\nsized for onboarding, the docs tutorial, and CI-scale runs rather than for\nseparating strong methods. The split is an interior holdout on the 4-length\n\u00d7 4-velocity grid, interpolation only: every length and every velocity\nappears in training, `val` (300 mm at 2, 400 mm at 4 m/s) only picks each\nrun's checkpoint, and `test_interp` (300 mm at 4, 400 mm at 2 m/s) is\nscored. Everything is reported in physical units \u2014 axial-stress RMSE in MPa,\nposition RMSE in mm \u2014 and the quantities of interest read the wave directly:\narrival time at the 25 / 50 / 75 % gauge stations and the peak stress. The\nreference CGN baseline nails the arrival times and overshoots the peak.\nDespite its entry-tier size, the benchmark separates prediction schemes\nsharply (2026-08-28 multi-method fleet): the time-conditioned operators\n(Transolver, GeoFLARE) beat CGN ~3x on the fields and ~10x on peak stress,\nwhile every mesh-native autoregressive family (MGN, Transolver, GeoFLARE)\nfails outright (relative L2 > 1) - the sustained 30 ms reverberation\nrecirculates rollout errors instead of letting them decay, and only CGN's\nrelative-displacement particle-graph formulation survives it. The\nautoregressive rows are kept as deliberate negative results; the numbers\nare below.", |
| "figures": [ |
| { |
| "path": "assets/wave_rollout.gif", |
| "caption": "Ground truth (top) vs CGN prediction (bottom) on held-out W1D-300-4 (test_interp): a 300 mm bar at 4 m/s initial velocity, coloured by axial stress, y-axis exaggerated x4. The surrogate tracks the compression front, the free-end reflections, and the cycle timing over the 30 ms rollout; degradation concentrates in the final ~5 ms.", |
| "alt": "Stacked animation of ground-truth and CGN-predicted axial stress waves in a slender bar." |
| }, |
| { |
| "path": "assets/wave_axial_interp_400_2.png", |
| "caption": "In-distribution (test_interp, 400 mm bar at 2 m/s): ground truth (top) vs CGN prediction (bottom), axial stress at t = 0.6 / 10.4 / 20.2 / 30.0 ms (y x4). The prediction reproduces the wavefront position and reflection cycles; late-horizon fields roughen and overshoot near the impact end (rollout position RMSE 0.95 mm).", |
| "alt": "Prediction-vs-truth axial-stress snapshots for the 400 mm bar, in-distribution." |
| } |
| ], |
| "data_access": "the canonical archive is public on Hugging Face \u2014 [StructBench/wave-propagation-1d](https://huggingface.co/datasets/StructBench/wave-propagation-1d) (CC BY 4.0): one `.h5` per case, `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/wave-propagation-1d](https://huggingface.co/datasets/StructBench/wave-propagation-1d)" |
| } |
|
|
