DeepWheel

Generating a 3D Synthetic Wheel Dataset for Design and Performance Evaluation

Published in ASME Journal of Mechanical Design (2026) 148(5): 051702  ·  KAIST SmartDesignLab / Narnia Labs

---

Contents

• Overview
• The data, qualitatively
• Dataset structure
• Usage
• Applications
• Citation
• License

---

Overview

DeepWheel is a synthetic automotive wheel dataset for design and performance evaluation, generated with a generative-AI framework: 2D renders are synthesized with Stable Diffusion, lifted to 3D via 2.5D depth estimation + reconstruction, and structurally simulated to extract engineering performance. It provides 6,000+ photo-realistic images and 900+ structurally-analyzed 3D models, each as multi-view images, depth maps, a reconstructed surface mesh, and a parametric CAD model, paired with performance labels (mass and modal natural frequencies) — coupling 3D geometry ↔ performance for data-driven design, surrogate modelling, and image-to-3D research.

_{A reconstructed wheel design, spinning about its axle.}

Domain	automotive wheels (rims)
Scale	6,000+ rendered images · 900+ structurally-analyzed 3D models
Modalities	rendered images · depth maps · 3D meshes · CAD models
Performance labels	mass · modal frequencies (Mode 7, Mode 11)
Paper	J. Mech. Des. (2026) 148(5):051702 · doi:10.1115/1.4069899 · arXiv:2504.11347

---

The data, qualitatively

_{Generated wheel-design variety — reconstructed 3D rim meshes spanning a range of spoke geometries.}

---

Dataset structure

DeepWheel/
├── 1_rendered_images.zip      # multi-view renders of each design
├── 2_predicted_depth_maps.zip # predicted depth maps
├── 3_3D_recon_meshes.zip      # reconstructed surface meshes (.stl)
├── 4_3D_cad_models.zip        # parametric CAD models
├── deepwheel_sim_results.csv  # performance labels (per design)
└── readme.docx                # official documentation (incl. file-matching instructions)

Modalities

Modality	File	Content
Rendered images	1_rendered_images/…	multi-view RGB renders
Depth maps	2_predicted_depth_maps/…	predicted per-view depth
3D meshes	3_3D_recon_meshes/*.stl	reconstructed surface meshes
CAD models	4_3D_cad_models/…	parametric CAD geometry
Performance labels	deepwheel_sim_results.csv	file_name, Mass, Mode7 Freq, Mode11 Freq

---

Usage

huggingface-cli download KAIST-SmartDesignLab/DeepWheel --repo-type dataset --local-dir DeepWheel
cd DeepWheel && for f in *.zip; do unzip -q "$f"; done

import pandas as pd, trimesh
labels = pd.read_csv("deepwheel_sim_results.csv").set_index("file_name")
case   = labels.index[0]
mesh   = trimesh.load(f"stl/{case}.stl")          # reconstructed wheel
mass, f7, f11 = labels.loc[case, ["Mass", "Mode7 Freq", "Mode11 Freq"]]

---

Applications

• Surrogate modelling — predict mass and modal frequencies from 3D geometry or rendered views.
• Image-to-3D — reconstruct wheel geometry from images / depth (paired renders + meshes + CAD).
• Generative & inverse design — generate manufacturable wheel rims with target performance.
• Multimodal learning — images ↔ depth ↔ mesh ↔ CAD ↔ performance.

---

Citation

@article{deepwheel2026,
  title   = {DeepWheel: Generating a 3D Synthetic Wheel Dataset for Design and Performance Evaluation},
  journal = {Journal of Mechanical Design},
  volume  = {148},
  number  = {5},
  pages   = {051702},
  year    = {2026},
  doi     = {10.1115/1.4069899}
}

---

License

Released under CC BY-NC 4.0 (Creative Commons Attribution-NonCommercial 4.0 International). Use, modification, and redistribution are permitted for non-commercial purposes with attribution; commercial use of the dataset or derivative models is prohibited. See readme.docx and the paper (doi:10.1115/1.4069899) for details. DeepWheel — KAIST Smart Design Lab / Narnia Labs.