structmeshdata/underbody-impact-data

Automotive Underbody Panel Impact Dataset

Version: 1.1.1
Data type: finite-element simulation trajectories
Task: impact-conditioned displacement and shell von Mises effective-stress field prediction

The Automotive Underbody Panel Impact Dataset contains independent impact simulations on three automotive structural geometries. Each geometry has 500 Latin-hypercube-sampled impact conditions. Every case stores 17 aligned states of the full three-dimensional nodal displacement field and shell-element von Mises effective stress.

The dataset supports research on graph neural operators, mesh-based surrogate models, spatiotemporal field prediction, peak-event prediction, and simulation-based design screening.

Detailed generation documentation is provided in:

• metadata/SIMULATION_PROTOCOL.md for units, the impactor, boundary/contact conditions, and solver/output settings;
• metadata/LHS_DESIGN.md for parameter definitions, design bounds, and the geometry-specific constrained-LHS construction;
• metadata/TEMPORAL_SAMPLING.md for the 17-state reduction and the discrete peak-time definition.

Dataset summary

Geometry	Cases	Nodes	Directed graph edges	Shell elements	Displacement	von Mises effective stress
floorfrontdriver	500	7,408	29,572	7,374	[7408,17,3]	[7374,17]
floorfrontR	500	12,011	48,138	12,055	[12011,17,3]	[12055,17]
trunkfloor	500	14,440	58,074	14,589	[14440,17,3]	[14589,17]

The three geometries are independent datasets. Equal case identifiers across geometries do not denote paired physical simulations.

Conditions and units

The simulations use the tonne--mm--s--N consistent unit system. Coordinates and displacements are in mm, time is in s, velocity is in mm/s, mass is in tonne, density is in tonne/mm^3, and stress and Young's modulus are in MPa.

Symbol	Released field	Definition	Design domain	Unit
p	impact_xyz	centroid of the selected eligible panel shell	geometry-dependent discrete candidate set	mm
v	velocity_xyz	initial rigid-impactor translational velocity	derived from speed and angles	mm/s
s	impact_speed	velocity magnitude	[1732.05, 5196.15]	mm/s
mu	mass_ratio	common scale factor for reference impactor mass and density	[0.75, 1.25]	dimensionless
theta	theta_deg	polar angle measured from global +Z	[0, 15]	degree
phi	phi_deg	azimuth in global XY, from +X toward +Y	[0, 360]	degree
E	material_young_mpa	rigid-impactor Young's modulus	{70000, 110000, 210000}	MPa
nu	material_poisson	rigid-impactor Poisson ratio	{0.33, 0.34, 0.30}	dimensionless

The velocity is constructed as v = s [sin(theta) cos(phi), sin(theta) sin(phi), cos(theta)]. The mass-ratio convention is impactor_mass = 0.01 tonne * mu and impactor_density = 5.205e-5 tonne/mm^3 * mu.

Impact positions are not sampled as three independent continuous coordinates. Eligible positions are shell-element centroids at least 80 mm from the topological outer boundary. Two LHS coordinates are mapped in normalized XY to the nearest unused eligible centroid; the released Z coordinate is the actual centroid height.

Constrained-LHS design

Each design uses seven normalized coordinates: position X, position Y, speed, mass ratio, theta, phi, and material class. Candidate designs are generated by Latin hypercube sampling, and the design with the largest normalized minimum pairwise distance among 128 trials is retained. Material is a balanced three-level categorical coordinate.

floorfrontR and trunkfloor use independent single-batch designs with seeds 20260728 and 20260723, respectively. floorfrontdriver is a staged design: cases 001--100 use seed 20260721, cases 101--200 are a complementary nested extension using seed 20260722, and cases 201--500 form an independent augmentation using seed 20260722. The geometries share parameter bounds and simulation rules but do not share paired physical conditions.

Impactor and simulation setup

The impactor is a rigid spherical shell with radius 12.5 mm and an initial 5.0-mm gap from the target centroid along the direction opposite to travel. It uses shell ELFORM 2, shear factor 0.833333, three through-thickness integration points, thickness 0.1 mm, and *MAT_RIGID. The material class changes only the rigid impactor's E and nu; mass ratio changes only its density and nominal mass.

All nodes on the panel's topological outer boundary are constrained in all six degrees of freedom. Impactor--panel interaction uses *CONTACT_AUTOMATIC_SURFACE_TO_SURFACE_ID with static and dynamic friction coefficients of 0.15. A body acceleration of 9810 mm/s^2 is applied in global +Z. Simulations end at 0.03 s. They were run with LS-DYNA SMP single precision R12 through ANSYS v221 lsdyna_sp.exe using ncpu=8 and memory=400m.

The panel material definitions and shell sections remain those of the cited upstream Version 3 model; their full keyword cards are not redistributed in this compact release.

Stress definition

The effective_stress field is the shell-element von Mises equivalent stress exported from LS-PrePost. The LS-PrePost etime 9 component corresponds to Effective Stress (v-m), ip#max: for each shell element and retained state, the stored scalar is the maximum von Mises stress over all through-thickness integration points. The maximizing integration-point index is not retained. Stress values are in MPa, and the tensor shape is [Ne, 17].

Repository structure

underbody-impact-data/
├── data/
│   ├── floorfrontdriver/cases_001_100.zip ... cases_401_500.zip
│   ├── floorfrontR/cases_001_100.zip ... cases_401_500.zip
│   └── trunkfloor/cases_001_100.zip ... cases_401_500.zip
├── meshes/
├── metadata/
├── generation_evidence/
├── scripts/
├── manifest.csv
├── checksums.sha256
├── DATASHEET.md
└── schema.json

Each ZIP member is stored as cases/caseNNN.pt. The files are PyTorch-serialized plain dictionaries. manifest.csv records the byte size and SHA-256 digest of every case.

Download

from huggingface_hub import snapshot_download

dataset_root = snapshot_download(
    repo_id="structmeshdata/underbody-impact-data",
    repo_type="dataset",
    revision="v1.1.1",
)

Loading a case

PyTorch 2.6 or newer is recommended. The loader uses weights_only=True and reads cases directly from ZIP shards:

python scripts/load_case.py \
  --dataset-root . \
  --geometry floorfrontdriver \
  --case case001

from pathlib import Path
import sys

sys.path.insert(0, str(Path("scripts").resolve()))
from load_case import load_case, load_mesh

case = load_case(Path("."), "floorfrontdriver", "case001")
mesh = load_mesh(Path("."), "floorfrontdriver")
print(case["disp"].shape)
print(case["effective_stress"].shape)

Validation

python scripts/validate_dataset.py --dataset-root . --verify-checksums

The validator checks the case schema, tensor shapes, finite values, aligned time arrays, split coverage, mesh connectivity, archive membership, and SHA-256 digests.

Fixed split and peak-event task

The fixed split is 400 train / 50 validation / 50 test cases per geometry with seed 12345. Normalization statistics must be computed from the training cases only.

For peak-event prediction, the supplied script selects the state containing the global maximum valid nodal displacement magnitude and uses the von Mises effective-stress field from that same state.

Temporal sampling

The solver writes D3PLOT output at a nominal interval of 0.0002 s. Compact conversion retains every tenth raw state and appends the final state. Every released case therefore uses indices [0, 10, 20, ..., 150, 151]; the first 16 retained states have a nominal 0.002-s spacing, while the appended terminal state can be very close to index 150. Exact floating-point times are stored in each case and should be used instead of reconstructing them from the nominal interval.

The peak time t* is discrete: it is the argmax of nodal displacement magnitude over valid nodes and the 17 retained states only. It is not a continuous-time solver maximum. Stress at the same selected retained state is used as the paired peak-event stress target.

Data-version note

The included floorfrontR data passed the release quality audit. Files from earlier internal builds must not be mixed with this release; see metadata/floorfrontR_DATA_NOTE.md.

Limitations

• The fields are numerical simulation results, not physical crash-test measurements.
• The dataset covers three fixed meshes and their documented sampled conditions.
• It does not establish generalization to arbitrary vehicle geometries or real tests.
• Public test labels reproduce the fixed paper protocol but are not a hidden benchmark.
• The exact LS-DYNA R12 sub-build is not retained for every case.
• The recorded impactor_mass is the generator-defined nominal mass; users requiring an independently recomputed solver mass should inspect an original solver MATSUM output.
• Peak-event time is quantized to the released 17-state temporal grid.

License

This repository is released under the MIT License. Third-party names and source model provenance are documented in THIRD_PARTY_NOTICES.md.

Citation

Please cite the versioned Hugging Face repository for release v1.1.1: https://huggingface.co/datasets/structmeshdata/underbody-impact-data/tree/v1.1.1. Citation metadata is also provided in CITATION.cff.