Research map / Vision-language and multimodal / Human and embodied modeling
AutoHood3D: A Multi‑Modal Benchmark for Automotive Hood Design and Fluid–Structure Interaction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Releases AutoHood3D, a 16,000+ variant automotive hood dataset with one-way LES-FEA fluid-structure interaction simulations, STL meshes, point clouds, and text prompts, plus baselines for five neural architectures.
Problem
Lack of large-scale, reproducible, multimodal 3D datasets for multiphysics (fluid-structure interaction) problems, such as automotive hood deformation during rotary dip painting, to support ML surrogates and generative design.
Approach
Generates 16,000+ hood geometry variants by combining convex-hull base shells with SAM-2-segmented, K-Means-clustered engineered cutouts, and simulates each with one-way coupled OpenFOAM LES fluid and linear-elastic solid solvers via preCICE, releasing the outputs in multiple modalities including text prompts for LLM fine-tuning.
Key results
- MLP and PointNet achieve the lowest in-distribution errors (e.g., MLP Ux 0.25e-2, PointNet Dy 0.40e-2) with sub-12 ms inference.
- Graph U-Net attains the best OOD pressure MSE (4.05e-2) and PointGNNConv the lowest OOD D_y and D_z errors, at 200-250 ms inference for Graph U-Net.
- The dataset comprises 16,000+ hood geometries and over 108,000 time-resolved LES-FEA snapshots, with roughly 1.2 TB of storage and about 5000 CPU-core hours of simulation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MLP rather than the graph-based architectures | surrogate models · ID test, pressure p | MSE: 0.37 vs 3.09–7.70 | Helped |
| Use PointNet rather than the other architectures | surrogate models · ID test, displacement D_x | MSE: 0.29 vs 2.80–14.88 | Helped |
| Use Graph U-Net rather than the point- and MLP-based models | surrogate models · OOD test, pressure p | MSE: 4.05 vs 4.96–15.45 | Helped |
| Use PointNet rather than the other architectures | surrogate models · OOD test, displacement D_x | MSE: 73.79 vs 84.34–181.02 | Helped |
| Increase GraphSAGE max_neighbors from 4 to 8 (ablation) | GraphSAGE · ID test, D_y | MSE: 3.78 -> 1.76 | Helped |
| Increase GraphSAGE max_neighbors from 4 to 16 (ablation) | GraphSAGE · ID test, D_y | MSE: 3.78 -> 4.28 | Hurt |
| Increase GraphSAGE max_neighbors from 4 to 32 (ablation) | GraphSAGE · ID test, D_y | MSE: 3.78 -> 1.51 | Helped |
| Increase GraphSAGE max_neighbors from 4 to 32 (ablation) | GraphSAGE · training runtime per epoch | time per epoch: 147 ± 2 -> 2361 ± 245 | Hurt |
| Reduce MLP hidden width from 256 to 128 (ablation) | MLP · ID test, U_x | MSE: 0.25 -> 0.32 | Hurt |
| Reduce MLP hidden width from 256 to 64 (ablation) | MLP · ID test, U_x | MSE: 0.25 -> 0.39 | Hurt |
| Reduce MLP hidden width from 256 to 128 (ablation) | MLP · OOD test, D_x | MSE: 181.02 -> 175.28 | Helped |
| Reduce MLP hidden width from 256 to 64 (ablation) | MLP · OOD test, D_x | MSE: 181.02 -> 124.11 | Helped |
Surprising findings
- Simple MLP and PointNet outperform graph-based models in-distribution, while graph models generalize better OOD.
- The MLP's ID advantage is attributed to point-wise memorization/overfitting rather than neighbor aggregation, since homophily is high (assortativity above 0.96).
Limitations
- Cutouts are defined from 2D projections, not fully 3D contours.
- Only one-way coupling with a linear-elastic small-strain solver and a single impulsive loading condition is modeled.
- The MSE loss does not enforce momentum or force conservation at the fluid-structure interface.
Experimental setup
Datasets: AutoHood3D 4.5k hood subset Base inner-hood designs (100) and cutout designs (~250) from prior work
Benchmarks: AutoHood3D in-distribution test AutoHood3D out-of-distribution test
Metrics: mean squared error inference time per sample parameter count time per epoch
Baselines: MLP PointNet GraphSAGE Graph U-Net PointGNNConv
Builds on: OpenFOAM preCICE pimpleFoam solidDisplacementFoam PointNet GraphSAGE
Closest papers by meaning
- DrivAerStar: An Industrial-Grade CFD Dataset for Vehicle Aerodynamic Optimization (NeurIPS 2025 · similarity 0.86)
- WindsorML: High-Fidelity Computational Fluid Dynamics Dataset For Automotive Aerodynamics (NeurIPS 2024 · similarity 0.86)
- DrivAerNet++: A Large-Scale Multimodal Car Dataset with Computational Fluid Dynamics Simulations and Deep Learning Benchmarks (NeurIPS 2024 · similarity 0.85)
- DragSolver: A Multi-Scale Transformer for Real-World Automotive Drag Coefficient Estimation (ICML 2025 · similarity 0.85)
- AeroGTO: An Efficient Graph-Transformer Operator for Learning Large-Scale Aerodynamics of 3D Vehicle Geometries (AAAI 2025 · similarity 0.85)
- 3DID: Direct 3D Inverse Design for Aerodynamics with Physics-Aware Optimization (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.