CrashSolver: Hierarchical Neural Crash Surrogate

• CrashSolver is a hierarchical neural surrogate for full-vehicle crash prediction that maps undeformed mesh data to predicted nodal displacements.
• It employs a finite-element part hierarchy with local encoders, a global transformer, and interface message passing to accurately model structural load transfer.
• It outperforms state-of-the-art baselines on the CarCrashNet benchmark, achieving lower RMSE and MAE on complex vehicles like the Chevrolet Silverado.

Searching arXiv for the relevant CrashSolver paper and closely related crash-simulation surrogates. CrashSolver is a hierarchical neural surrogate for full-vehicle structural crash prediction introduced alongside the CarCrashNet benchmark. Its purpose is to learn the mapping from an undeformed finite-element vehicle mesh and crash setup to the future nodal displacement field over the crash event, so that deformed geometry can be recovered as the initial coordinates plus predicted displacement. The method is explicitly framed as full-field, mesh-resolved crash prediction on complete vehicles rather than scalar crashworthiness regression, and it is designed to replace expensive explicit crash simulation with a fast learned model while still respecting vehicle part structure and load paths (Elrefaie et al., 8 May 2026).

1. Problem definition and mathematical formulation

CrashSolver addresses structural crash simulation in a regime dominated by nonlinear contact, large deformation, plasticity and energy dissipation, possible failure or erosion, and long-range coupling across many vehicle substructures. The target is not a reduced response quantity but the full transient deformation field on the vehicle mesh, which places it in the class of operator-learning surrogates for high-dimensional structural dynamics (Elrefaie et al., 8 May 2026).

The learning interface is described as

$$f_{\theta}\left(\mathbf{X}^{(0)}, \mathcal{E}, \mathbf{p}, \boldsymbol{\xi}\right)=\hat{\mathbf{U}}^{(1:n)} \in \mathbb{R}^{n \times N \times 3},$$

where $\mathbf{X}^{(0)}$ denotes undeformed nodal coordinates, $\mathcal{E}$ retained mesh or surface connectivity, $\mathbf{p}$ FE part identifiers or semantic labels, and $\boldsymbol{\xi}$ the design vector. The model predicts nodal displacements rather than stresses or forces directly. Deformed positions are then recovered by

$$\hat{\mathbf{X}}^{(t)}=\mathbf{X}^{(0)}+\hat{\mathbf{U}}^{(t)}.$$

The surrounding physical problem is expressed in semi-discrete form as

$$\mathbf{M}\,\ddot{\mathbf{u}}(t)+\mathbf{f}_{\mathrm{int}}\!\left(\mathbf{u}(t), \dot{\mathbf{u}}(t); \boldsymbol{\theta}_{\mathrm{mat}}\right)+\mathbf{f}_{\mathrm{cont}}\!\left(\mathbf{u}(t), \dot{\mathbf{u}}(t)\right)=\mathbf{f}_{\mathrm{ext}}\!\left(t; \boldsymbol{\xi}\right),$$

with $\mathbf{M}$ the mass matrix, $\mathbf{u}(t)$ the nodal displacement field, $\mathbf{f}_{\mathrm{int}}$ the internal force, $\mathbf{X}^{(0)}$0 the contact force, $\mathbf{X}^{(0)}$1 the external loading, $\mathbf{X}^{(0)}$2 the material parameters, and $\mathbf{X}^{(0)}$3 the design vector. This formulation makes clear that CrashSolver is intended as a surrogate over a transient, nonlinear, contact-dominated dynamical system rather than as a static regressor (Elrefaie et al., 8 May 2026).

2. Hierarchical architecture and inductive bias

The central design principle of CrashSolver is the use of the finite-element part hierarchy as an inductive bias. Rather than treating the complete vehicle as a single point cloud, the model groups nodes by semantic vehicle components, encodes each component with a shared local encoder, mixes component summaries with a global transformer, passes messages across component interfaces, and decodes a future displacement sequence (Elrefaie et al., 8 May 2026).

This hierarchy separates local deformation within parts from global crash load transfer between parts. The local component encoders operate within semantic structural groups such as bumper, rails, radiator support, subframe, cabin floor, rocker, and pillars. A global component transformer exchanges information across those component summaries, while interface message passing handles cross-component boundaries and load transfer. A temporal decoder then predicts the sequence of future deformation. The intended effect is preservation of crash load-path structure more effectively than a flat transformer or point-cloud model (Elrefaie et al., 8 May 2026).

The architecture is therefore multi-scale in a specifically structural sense. Local encoding captures part-wise geometric and deformation behavior; global mixing captures inter-part coupling; interface message passing handles the regions where force transfer is most consequential. This suggests that CrashSolver is less a generic mesh transformer than a solver whose representation is aligned with how automotive structures are engineered and analyzed.

3. Data model and benchmark setting

CrashSolver is benchmarked primarily on the vehicle-scale CarCrashNet dataset, a public high-fidelity benchmark for data-driven structural crash simulation. CarCrashNet combines component-scale and full-vehicle simulations in a multi-modal format, including more than 14,000 bumper-beam pole-impact simulations together with 825 full-vehicle crash simulations built from three industry-standard vehicle models of increasing structural complexity: Dodge Neon, Toyota Yaris, and Chevrolet Silverado. The paper also states that the open-source finite-element workflow is based on OpenRadioss and is validated against both experimental crash data and the commercial solver Ansys LS-DYNA (Elrefaie et al., 8 May 2026).

For the full-vehicle campaigns, the design vector is

$\mathbf{X}^{(0)}$4

where $\mathbf{X}^{(0)}$5 is impact velocity, $\mathbf{X}^{(0)}$6 scales front-support shell thicknesses, and $\mathbf{X}^{(0)}$7 scales lower-rail or subframe thicknesses. Velocity ranges from 50 to 64 km/h for all three vehicle campaigns, and thicknesses vary within $\mathbf{X}^{(0)}$8 of nominal.

The released full-vehicle data provide time-resolved VTKHDF fields including undeformed and deformed coordinates, displacement and velocity, von Mises stress, equivalent plastic strain, specific internal energy, erosion or failure flags, part and node IDs, and force or energy histories. These fields define a machine-learning-ready representation of crash dynamics at mesh resolution.

Vehicle campaign	Simulations and hidden test runs	Split
Toyota Yaris	500 simulations, 50 hidden test runs	80/10/10
Dodge Neon	250 simulations, 25 hidden test runs	80/10/10
Chevrolet Silverado	75 simulations, 15 hidden test runs	56/4/15

The benchmark protocol uses the same train, validation, and test split across compared models, trains on the retained node subset, and reports performance on hidden test sets. The reported metrics are MAE, RMSE, relative $\mathbf{X}^{(0)}$9 position error, relative displacement error, and final-frame or time-local RMSE (Elrefaie et al., 8 May 2026).

4. Quantitative performance and comparative evaluation

CrashSolver is compared against three state-of-the-art baselines: Transolver, GeoTransolver, and FIGConvUNet. On the main hidden test sets, it achieves the best mean RMSE on all three full-vehicle datasets, although the Toyota Yaris comparison with GeoTransolver is statistically tied (Elrefaie et al., 8 May 2026).

Dataset	CrashSolver	Selected comparison
Dodge Neon	RMSE 32.763 mm, MAE 18.036 mm	Transolver 33.947 mm; FIGConvUNet 34.044 mm; GeoTransolver 34.403 mm
Toyota Yaris	RMSE 21.769 mm, MAE 13.507 mm	GeoTransolver 21.773 mm; FIGConvUNet 21.910 mm; Transolver 22.583 mm
Chevrolet Silverado	RMSE 61.536 mm, MAE 37.753 mm	GeoTransolver 79.230 mm; Transolver 83.971 mm; FIGConvUNet 102.747 mm

The Chevrolet Silverado benchmark is the clearest separation point. There, CrashSolver substantially outperforms every listed baseline on the most structurally complex vehicle. On Dodge Neon, it is best on all reported metrics. On Toyota Yaris, the difference relative to GeoTransolver is marginal, with CrashSolver slightly better in RMSE and relative displacement error.

The appendix reports paired bootstrap confidence intervals and significance tests. The main findings are that CrashSolver and GeoTransolver are statistically tied on Toyota Yaris; CrashSolver is significantly better than Transolver and FIGConvUNet on Dodge Neon, with the GeoTransolver comparison described as mixed; and CrashSolver is significantly better than every listed baseline on Chevrolet Silverado. The paper also reports that, on an external concurrent dataset, SHIFT-Crash, a lower-capacity CrashSolver variant outperforms Transolver and GeoTransolver (Elrefaie et al., 8 May 2026).

5. Ablations and architectural interpretation

The ablation results support a specific reading of the architecture. The paper’s main findings are that hierarchy matters, component-aware local encoding improves accuracy especially on harder vehicles, global mixing helps but less than local structure, more capacity is not always better, part and contact conditioning help marginally, and CrashSolver scales best to complex vehicles (Elrefaie et al., 8 May 2026).

On Dodge Neon ablations, increasing local encoder slices from 16 to 32 improves RMSE from 37.135 mm to 36.707 mm. Deeper local encoders do not help. More global layers help slightly, changing attention heads matters little, reducing latent width to 96 is better than widening it, and part or contact conditioning gives small but consistent gains. The paper therefore argues for a conservative hierarchical design rather than a larger monolithic transformer.

These results locate the model’s principal gain in the local geometry encoder and in the explicit use of part structure. A plausible implication is that the hierarchy is functioning as a structural regularizer: it constrains the learned representation to follow the decomposition by which crash loads are physically transmitted through an assembled vehicle. This interpretation is reinforced by the fact that the advantage of CrashSolver grows with geometry and load-path heterogeneity.

6. Limitations, scope, and relation to adjacent surrogate-solver work

The paper explicitly limits the present benchmark to frontal rigid-wall impacts for the vehicle-scale campaigns. It does not include offset, side, rear, oblique, rollover, or pedestrian scenarios. The validation emphasizes global response more than local acceleration or intrusion; RMSE and MAE can underweight localized crash-critical regions; relative position error can look artificially small because of normalization by vehicle scale; and only a subset of future tasks is benchmarked (Elrefaie et al., 8 May 2026).

Within those bounds, CrashSolver is presented as a practical step toward AI-driven virtual crash testing. Its significance lies in the combination of a reproducible open benchmark with a structurally informed neural solver. The broader surrogate-modeling context includes geometry-aware operator-learning approaches such as GeoTransolver, later extended with FLARE-based low-rank attention for industrial-scale crash dynamics prediction (Akhare et al., 26 May 2026). This suggests a broader methodological convergence around mesh-resolved surrogates that must simultaneously capture local deformation, global load transfer, and long-range transient dependencies.

CrashSolver’s specific contribution within that landscape is the introduction of a full-vehicle surrogate whose organizing principle is the finite-element part hierarchy itself. It learns full-vehicle deformation fields rather than only scalar crashworthiness outputs, is benchmarked on public validated data, and is shown to outperform or match strong geometric-transformer baselines on hidden test sets. As a result, it occupies a distinct position between traditional explicit solvers and more generic neural operators: not a replacement for certification-grade simulation in all settings, but a structured surrogate for rapid full-field crash prediction under a clearly defined benchmark regime.