CarCrashNet: A Large-Scale Dataset and Hierarchical Neural Solver for Data-Driven Structural Crash Simulation

The gap

Crash simulation finally gets its benchmark

Structural crash simulation governs vehicle safety, yet the field has lacked the open, validated, large-scale benchmarks that fueled progress in fluid dynamics, weather, and atomistic modeling. CarCrashNet closes that gap: the first public, high-fidelity benchmark for data-driven crash simulation, pairing component-scale and full-vehicle finite-element data with CrashSolver, a hierarchical neural solver that sets state of the art on every released benchmark.

Validation

Open-source physics you can trust

Dataset credibility comes first: the OpenRadioss workflow is validated head-to-head against the industry-standard commercial solver Ansys LS-DYNA on the detailed Toyota Yaris model — agreeing within 7.2% on peak wall force, 2.6% on wall-force duration, and 0.5% on peak internal energy — and benchmarked against published physical crash-test references.

Time-synchronized crash of the same Toyota Yaris model: OpenRadioss (open-source) vs. Ansys LS-DYNA (commercial).

Datasets

Two corpora, two scales of complexity

Bumper-beam pole impacts

14,742 component-scale simulations of a DP1000 bumper beam + DP600 crash-box assembly. Seven design variables — velocity, thicknesses, yield strengths, pole diameter and offset — sampled with a scrambled Sobol DOE.

14,742 simulations

Full-vehicle crashes

825 explicit FE simulations across three industry-standard models of increasing complexity: Toyota Yaris (500), Dodge Neon (250), Chevrolet Silverado (75), varying impact velocity and structural shell thicknesses.

825 simulations · 3 vehicles

ML-ready everything

Full-field VTKHDF trajectories (displacement, velocity, stress, plastic strain, erosion), global and local time histories, and reduced crashworthiness scalars for every single case.

Multi-modal · mesh-resolved

Displacement fields across the bumper-beam design space (isometric view).

von Mises stress fields across the design space (isometric view).

Equivalent plastic strain across the design space (isometric view).

Low-velocity (top) and high-velocity (bottom) crash cases for Yaris, Neon, and Silverado.

CrashSolver

A neural solver that thinks in parts, not points

Vehicles crash along structural load paths, not as monolithic point clouds. CrashSolver exploits the finite-element part hierarchy as an inductive bias: semantic decomposition into structural groups, shared local component encoders, a global component transformer, mesh-derived interface message passing, and a temporal decoder that emits the full crash displacement trajectory.

CrashSolver architecture: part-aware encoding, global component mixing, interface message passing, and temporal readout.

Results

Lowest error on every metric, on every vehicle

On the unseen hidden test set of each vehicle benchmark, CrashSolver beats Transolver, GeoTransolver, and FIGConvUNet across the board — and its advantage grows with structural complexity: on the Silverado pickup, RMSE drops by ~22% versus the strongest baseline.

Dataset	Model	RMSE (mm) ↓	MAE (mm) ↓	Rel. L₂ (pos) ↓	Rel. L₂ (disp) ↓
Dodge Neon	CrashSolver	32.763	18.036	0.02499	0.08837
Dodge Neon	Transolver	33.947	18.678	0.02589	0.09148
Toyota Yaris	CrashSolver	21.769	13.507	0.01537	0.09043
Toyota Yaris	GeoTransolver	21.773	13.359	0.01537	0.09059
Chevrolet Silverado	CrashSolver	61.536	37.753	0.03143	0.17069
Chevrolet Silverado	GeoTransolver	79.230	45.366	0.04049	0.21844

Benchmark overview: CrashSolver achieves the lowest error on every released dataset.

Reference

Citation

@article{elrefaie2026carcrashnet,
  title   = {CarCrashNet: A Large-Scale Dataset and Hierarchical Neural Solver
             for Data-Driven Structural Crash Simulation},
  author  = {Elrefaie, Mohamed and Shu, Dule and Klenk, Matt and Ahmed, Faez},
  journal = {arXiv preprint arXiv:2605.07098},
  year    = {2026}
}

Keep exploring