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

TL;DR

AutoHood3D tackles the lack of large-scale, multi-modal 3D datasets for automotive hood design under fluid-structure interaction. By integrating a three-stage workflow—base geometry generation, design variation, and high-fidelity LES-FFD co-simulation—the authors produce 16k+ hood variants with time-resolved FSI data. They provide multiple data modalities (STL, CFD/FEA fields, point clouds, and text prompts) and an end-to-end reproducible pipeline, including OpenFOAM-based solvers and preCICE collaboration. Benchmarks across five neural architectures reveal a speed-accuracy trade-off and highlight the potential of graph-based models for out-of-distribution generalization, enabling physics-aware ML and generative design in automotive contexts.

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.

Figures (16)

• Figure 1: Dataset modalities (left): base 3D STL hood shell, surface point‐cloud sampling, raw fluid and solid meshes, and sample surface fields (normalized displacement and pressure‐gradient) extracted on the hood surface. Example variants (right): selected hood geometries with interpolated deformation fields mapped onto the STL, demonstrating how different cutout topologies yield distinct deflection patterns on both the front and rear faces.
• Figure 2: Workflow for generating multiple hood geometries. Starting from the base inner‐hood CAD (left), the surface is projected onto a 2D plane and segmented into individual cutout masks. Each mask boundary is extracted as an ordered point-cloud curve. These curves are then re-embedded into the full 3D hood shell to produce the final geometries with engineered openings (right).
• Figure 3: Co-simulation workflow.
• Figure 4: One-way coupled FSI co-simulation overview. An impulsive inflow acceleration is applied to the fluid domain, which uses free-slip exterior walls and a no-slip hood surface. The computed pressure field on the fluid mesh is transferred via preCICE’s nearest-projection adapter to the solid mesh, where the hood shell is fully clamped for structural deformation analysis.
• Figure 5: Supervised fine‑tuning (SFT) example data point: an LLM is trained to map structured text prompts, specifying cutout parameters such as curve count and geometric spacing, to the corresponding 3D point cloud hood geometry.