CarCrashNet

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

Mohamed Elrefaie¹, Dule Shu², Matt Klenk², Faez Ahmed¹ ¹ Massachusetts Institute of Technology ² Toyota Research Institute

---

Overview of the CarCrashNet framework: a large bumper-beam pole-impact dataset and three full-vehicle crash datasets (Toyota Yaris, Dodge Neon, Chevrolet Silverado), benchmarked across state-of-the-art neural solvers and validated against Ansys LS-DYNA and physical crash testing.

---

Overview

CarCrashNet is the first public, high-fidelity, open-source benchmark for data-driven structural crash simulation. It combines validated component-scale and full-vehicle finite-element crash data in a multi-modal, machine-learning-ready format, together with CrashSolver, a new hierarchical neural solver that achieves state-of-the-art performance across every released benchmark.

Structural crash simulation governs vehicle safety, but until now the field has lacked an open, validated, large-scale benchmark of the kind that has fueled progress in fluid dynamics, weather, and atomistic modeling. CarCrashNet closes that gap.

---

Key Contributions

1. Validated open-source crash workflow. OpenRadioss benchmarked against both the industry-standard commercial solver Ansys LS-DYNA and physical crash-test references, agreeing within 7.2% on peak wall force, 2.6% on wall-force duration, and 0.5% on peak internal energy.

2. Two large-scale public datasets at two levels of fidelity and complexity (see Datasets).

3. CrashSolver, a hierarchical, part-aware neural solver for full-vehicle crash field prediction that outperforms all state-of-the-art baselines on every released benchmark (see Benchmarks).

---

Solver Validation — OpenRadioss vs. Ansys LS-DYNA

To establish dataset credibility we validate the open-source OpenRadioss workflow against the commercial industry-standard solver Ansys LS-DYNA on the detailed Toyota Yaris model, and against independent physical crash-test references.

Side-by-side, time-synchronised crash response of the same Toyota Yaris model simulated in OpenRadioss (open-source) and Ansys LS-DYNA (commercial).

Quantitative agreement against the LS-DYNA reference:

Quantity	OpenRadioss vs. LS-DYNA
CFC60 peak wall force	7.2 %
Wall-force duration	2.6 %
Peak internal energy	0.5 %

Relative to the published physical crash test, OpenRadioss matches the impact speed within 0.2 %, overpredicts peak wall force by ~14.6 %, and underpredicts wall-force duration by ~19.6 % — comparisons that support global-response agreement while transparently identifying contact- and pulse-shape quantities as solver-sensitive.

---

Datasets

CarCrashNet releases 15,567 validated finite-element crash simulations — 6.65 TB of mesh-resolved fields, time histories, and reduced crashworthiness metrics.

1. Bumper-Beam Pole-Impact Dataset — 14,742 simulations

A large component-scale corpus of frontal pole impacts on a DP1000 bumper beam + DP600 crash-box assembly. Seven engineering design variables are sampled with a scrambled Sobol design of experiments:

Variable	Range
Impact velocity	7.2 – 54.0 km/h
Crash-box thickness	1.0 – 3.0 mm
Bumper-beam thickness	1.0 – 3.0 mm
DP600 yield strength	0.150 – 0.600 GPa
DP1000 yield strength	0.250 – 1.000 GPa
Pole diameter	100 – 500 mm
Lateral pole offset	0 – 800 mm

Each case ships with full-field VTKHDF trajectories, global/local time histories, and five scalar crashworthiness targets: peak pole contact force, peak deceleration, peak internal energy, peak plastic work, and kinetic-energy absorption fraction.

Multi-physics field outputs across the design space — displacement, von Mises stress, and equivalent plastic strain — shown from three viewing angles:

View	Displacement	von Mises stress	Plastic strain
Isometric
Side
Top

2. Full-Vehicle Crash Dataset — 825 simulations

Vehicle-scale explicit FE crash simulations across three industry-standard models of increasing structural and geometric complexity:

Vehicle	Class	Simulations	Edited structural groups
Toyota Yaris	Passenger sedan	500	37
Dodge Neon	Passenger sedan	250	27
Chevrolet Silverado	Pickup truck	75	78

Low- (top row) and high-velocity (bottom row) crash cases for the three vehicle-scale models.

Each campaign varies impact velocity (50 – 64 km/h) and physically meaningful front-support and lower-rail / subframe shell thicknesses (±10%), producing realistic deformation modes across the frontal crash load path.

Every case is released with:

• Field trajectories (VTKHDF): reference + deformed mesh, displacement, velocity, von Mises stress, plastic strain, internal energy, erosion flags, part IDs.
• Global & local time histories: rigid-wall force, kinetic / internal / contact / hourglass energies, local accelerations.
• Reduced scalar targets: peak wall force, peak internal energy, KE absorption, peak deceleration, impulse duration.

---

CrashSolver

CrashSolver is a hierarchical machine-learning solver designed for full-vehicle crash field prediction. It exploits the finite-element part hierarchy as an inductive bias.

1. Semantic decomposition of the vehicle into structural groups (bumper, rails, radiator support, shock housings, subframe, engine bay, cabin floor, rocker, pillars, exterior panels).
2. Shared local component encoders that learn deformation on smaller structural token sets — not over a single monolithic mesh.
3. Global component transformer that mixes component summaries.
4. Mesh-derived interface message passing that exchanges latent features across component boundaries.
5. Temporal decoder / nodal readout that emits the full crash displacement trajectory.

This design preserves the high-resolution nodal deformation field while giving the network an explicit representation of the crash load paths.

---

Benchmarks

Full-vehicle crash field prediction — CrashSolver achieves the lowest error on every dataset

Mean error on the unseen hidden test set of each CarCrashNet vehicle benchmark. Lower is better; best in bold.

Dataset	Model	RMSE (mm) ↓	MAE (mm) ↓	Relative L₂ (pos) ↓	Relative L₂ (disp) ↓
Dodge Neon	CrashSolver	32.763	18.036	0.02499	0.08837
	Transolver	33.947	18.678	0.02589	0.09148
	FIGConvUNet	34.044	18.850	0.02597	0.09196
	GeoTransolver	34.403	18.973	0.02628	0.09349
Toyota Yaris	CrashSolver	21.769	13.507	0.01537	0.09043
	GeoTransolver	21.773	13.359	0.01537	0.09059
	FIGConvUNet	21.910	13.576	0.01547	0.09105
	Transolver	22.583	14.049	0.01594	0.09391
Chevrolet Silverado	CrashSolver	61.536	37.753	0.03143	0.17069
	GeoTransolver	79.230	45.366	0.04049	0.21844
	Transolver	83.971	47.510	0.04291	0.23184
	FIGConvUNet	102.747	62.405	0.05248	0.28432

CrashSolver wins on every metric, on every vehicle. The advantage grows with structural complexity: on the Silverado pickup truck, CrashSolver reduces RMSE by ~22% relative to the strongest competing baseline (GeoTransolver).

Bumper-beam tabular surrogate benchmark

We additionally release a compact tabular benchmark on the 14,742-simulation bumper-beam corpus that compares 12 model families — linear models, kernel methods, MLP, tree ensembles, gradient boosting (XGBoost, LightGBM, CatBoost), a tabular foundation model (TabPFN v2), and AutoGluon AutoML — across five crashworthiness targets. Full per-target results are reported in the paper.

---

Dataset Availability

The CarCrashNet datasets will be publicly released upon acceptance / completion of peer review.

This repository currently hosts the high-level description of the work. We will update this page with download links and loading utilities for the bumper-beam and full-vehicle datasets once the peer-review process is complete.

---

Citation

If you use CarCrashNet, CrashSolver, or any of the released datasets, please cite:

@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}
}

---

Contact

For questions or collaboration: mohamed.elrefaie@mit.edu