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

AutoHood3D is a high-fidelity data-generation framework and multimodal dataset of 16k+ automotive-hood variants targeting a practical multiphysics challenge: deformation from fluid entrapment and inertial loading during rotary-dip painting. Each sample couples LES–FEA (1.2M cells) and provides time-resolved flow/structure fields, STL meshes, and structured language prompts. Unlike previous benchmarks (often 2D, low diversity, or missing multiphysics), AutoHood3D offers scalable 3D FSI, broad geometric diversity, and a fully reproducible pipeline. We benchmark point- and graph-based surrogates, establish in-distribution and out-of-distribution baselines, and reveal systematic errors in displacement and force; point-based models excel in-distribution, while graph-based models generalize better to novel geometries. The results motivate consistent multiphysics-aware losses that enforce fluid–solid coupling. AutoHood3D provides a reproducible foundation for physics-aware surrogates, rapid generative design, and new 3D FSI benchmarks.