AutoHood3D: A Multi-Modal Benchmark for Automotive Hood Design and Fluid-Structure Interaction

AI summary

Overview

Research area: Machine learning for engineering design and scientific computing (SciML) — specifically multiphysics surrogate modeling and generative design, built around fluid–structure interaction (FSI) during automotive manufacturing.

Technical level: Advanced. The paper combines large-eddy simulation, finite element analysis, co-simulation libraries, graph and point-cloud neural networks, and LLM-based text-to-geometry synthesis.

Scope: AutoHood3D is an open-source, multi-modal benchmark containing more than 16,000 3D automotive hood geometries with coupled LES–FEA solution fields, STL meshes, point clouds, and natural language annotations, plus baseline performance for five neural architectures.

What This Paper Is About

Automotive hoods must be evaluated against many competing constraints, but high-fidelity simulation of every design variant is computationally expensive, and publicly available datasets are typically two-dimensional, RANS-based, geometrically limited, or missing the annotations and reproducible generation pipelines that data-driven workflows need. The authors target a specific, under-modeled multiphysics problem — hood deformation caused by trapped coating fluid and inertial loading during the rotary dip paint process — and build a large, fully reproducible dataset around it. They then train and compare five neural network architectures to establish how well current models can predict the resulting flow and structural response.

Key Contributions

1. First open-source dataset of more than 16,000 3D automotive hood geometries with engineered cutouts, providing design diversity for fluid dynamics and structural studies (built from 100 base inner-hood geometries taken from prior work and extended with cutouts drawn from roughly 250 inner-hood designs).
2. A fully reproducible end-to-end workflow, spanning convex-shell reconstruction, cutout segmentation, automated mesh generation, and preCICE-mediated FSI co-simulation, designed to be extensible to other components and physical scenarios.
3. A multi-modal benchmark containing high-fidelity STL meshes, raw CFD solutions with time-resolved flow and surface fields, FEA displacement and stress outputs, surface point clouds, and tokenized natural language annotations — including a text-to-geometry corpus of 2,587 hood geometries generated with Gemma3 (gemma3-27b-it).
4. Quantitative baselines across five neural architectures (MLP, PointNet, GraphSAGE, Graph U-Net, PointGNNConv), reported separately for in-distribution and out-of-distribution testing with runtime and parameter counts.

Main Findings

• Simulation setup and cost: The LES mesh contains approximately 750,000 polyhedral cells (Δ = 0.003125 m), of which 170,000 to 180,000 are concentrated in the boundary layer around the hood; the structural surface mesh contains approximately 400,000 cells (Δ = 0.00220 m). The abstract describes this as 1.2M cells in total. The dataset occupies approximately 1.2 TB across 37,000 output files, and the campaign consumed roughly 5000 CPU-core hours, with each case parallelized over 12 cores and 72 GB of memory.

• Loading conditions: A constant acceleration pulse of α = 2.7 m/s² is applied over t = 0.07 s at 30° to the flow direction, with a sampling time of 0.01 s and a total of 8 solutions per case.

• Material and solver settings: Structural material properties are density ρs = 2700 kg/m³, Poisson ratio ν = 0.33, and Young's modulus E = 68.9e9 N/m. The fluid solver (UM_pimpleFoam) extends OpenFOAM's pimpleFoam with a Spalart–Allmaras Delayed Detached-Eddy-Simulation closure and an added transient body-force term; the structural solver (UM_solidDisplacementFoam) uses a small-strain linear-elastic model. Coupling is one-way, with the pressure field exchanged through preCICE using a nearest-projection scheme.

• In-distribution results favor simple models: On normalized fields (MSE, ×10⁻²), the MLP achieved the lowest errors on Ux (0.25), Uz (0.44), and p (0.37), while PointNet showed the best overall balance (Ux 0.31, Uy 0.26, Uz 0.48, p 0.55, Dx 0.29, Dy 0.40, Dz 0.43). GraphSAGE was substantially worse (Dx 14.88, Dz 6.86), and PointGNNConv also degraded (Dx 11.71, Dz 15.41). Graph U-Net produced intermediate in-distribution errors (Ux 1.91, p 3.09, Dx 2.89).

• Out-of-distribution results reverse the ranking: Graph-based models generalized better to a novel geometry. Graph U-Net attained the best OOD pressure MSE at 4.05 × 10⁻² with competitive velocity errors, and PointGNNConv achieved the lowest errors on the transverse displacement components Dy (4.78) and Dz (13.67). MLP degraded sharply, especially on Dx (181.02 × 10⁻²), and PointNet reached 15.45 × 10⁻² on pressure.

• Speed–accuracy trade-off: MLP (319,623 parameters) and PointNet (154,463) achieved inference times of 9.66 ms and 11.14 ms respectively, with GraphSAGE (93,703 parameters) at 9.74 ms. Graph U-Net required 248.86 ms per inference with 3,550,535 parameters, and PointGNNConv 34.34 ms with 168,598 parameters. Inference time was averaged for 200 samples. The conclusion summarizes this as roughly 7–11 ms and in-distribution MSE ~10⁻³ for MLP/PointNet, versus 34–249 ms and OOD MSE as low as 1.2 × 10⁻² for the graph models. All models were trained for 750 epochs; Graph U-Net results are reported at 320 epochs due to its longer per-epoch runtime.

• The hood surface is strongly homophilic: Reported mean assortativity values across fields were Ux +0.9756, Uy +0.9739, Uz +0.9711, p +0.9650, Dx +0.9910, Dy +0.9933, Dz +0.9877, and +0.9449 for the multi-dimensional cosine measure (with standard deviations ranging from 0.0037 to 0.0322). The paper states that all fields exhibit strong homophily (r̄ > 0.96), so neighbor aggregation should help graph networks if configured well.

• Neighborhood size controls a cost–accuracy trade-off: Sweeping GraphSAGE's max-neighbor setting (mNN = 4, 8, 16, 32) reduced in-distribution MSE by up to 40–50%, but per-epoch time rose from 147 ± 2 s at mNN = 4 to 2361 ± 245 s at mNN = 32, and the stored dataset grew from 201.1 GB to 826.6 GB. The mNN = 32 configuration was stopped at 130 epochs (other runs used 150 epochs), and mNN = 4 was judged the most favorable speed–accuracy–size balance.

• MLP advantage is attributed to capacity, not local aggregation: An ablation with hidden widths of 256, 128, and 64 showed in-distribution error rising steeply as width shrank (e.g., Dx from 2.80 to 4.59 to 7.63 × 10⁻²), while out-of-distribution error was marginally better for the smaller networks (Dx 124.11 for width 64 versus 181.02 for width 256; p 3.38 versus 4.96). The authors conclude the width-256 MLP memorizes point-wise patterns rather than aggregating neighbor information. Their hypothesis table marks the non-homophily explanation as rejected, the sensitivity to neighbor count as confirmed, and the overfitting explanation as supported.

• Dataset composition: A sample subset of approximately 4,500 hoods and an extended superset of 12,000+ hoods are organized in a curriculum learning sequence from fewest to most cutout curves, with an additional set of 10,000+ hoods having randomized cutout configurations. Pre-split training, validation, in-distribution test, and out-of-distribution test subsets are provided. The benchmark experiments used the 4,500-hood subset split into 75% training, 15% validation, 10% in-distribution testing, and a single isolated geometry for out-of-distribution testing.

• Solver validation details are not present in the provided text: The paper states that the numerical methodology is validated, but the corresponding material sits in Appendix C, whose content is truncated in the supplied document.

Methodology in Plain English

The authors started from 100 existing inner-hood CAD geometries and used a convex-hull algorithm to generate a matching outer envelope, then extruded a uniform 10 mm thickness to form a complete two-shell hood. To create design variety, they took roughly 250 inner-hood designs (each containing about four to eight cutouts), projected each 3D surface onto a reference plane, and used the SAM-2 segmentation model to isolate individual cutout shapes. Each cutout boundary was converted into an ordered point-cloud curve parameterized by smooth Bézier curves, yielding a database of about 1,750 unique cutout profiles. These profiles were clustered with K-Means in a two-dimensional perimeter-versus-area feature space, producing nine groups. New hood geometries were

Authors’ abstract

This study presents a new high-fidelity multi-modal dataset containing 16000+ geometric variants of automotive hoods useful for machine learning (ML) applications such as engineering component design and process optimization, and multiphysics system surrogates. The dataset is centered on a practical multiphysics problem-hood deformation from fluid entrapment and inertial loading during rotary-dip painting. Each hood is numerically modeled with a coupled Large-Eddy Simulation (LES)-Finite Element Analysis (FEA), using 1.2M cells in total to ensure spatial and temporal accuracy. The dataset provides time-resolved physical fields, along with STL meshes and structured natural language prompts for text-to-geometry synthesis. Existing datasets are either confined to 2D cases, exhibit limited geometric variations, or lack the multi-modal annotations and data structures - shortcomings we address with AutoHood3D. We validate our numerical methodology, establish quantitative baselines across five neural architectures, and demonstrate systematic surrogate errors in displacement and force predictions. These findings motivate the design of novel approaches and multiphysics loss functions that enforce fluid-solid coupling during model training. By providing fully reproducible workflows, AutoHood3D enables physics-aware ML development, accelerates generative-design iteration, and facilitates the creation of new FSI benchmarks. Dataset and code URLs in Appendix.