DeepWheel

Overview

DeepWheel is a large-scale, multi-modal synthetic dataset for automotive wheel design and performance evaluation. It was created using a fully automated framework combining generative AI (Stable Diffusion), monocular depth prediction, 3D reconstruction, and finite element structural simulation.

The framework addresses the scarcity of publicly available, high-quality datasets for wheel design by generating diverse wheel geometries — from reference designs and topology-optimized structures — and pairing them with engineering performance data. The dataset supports surrogate model training, data-driven inverse design, and comprehensive design space exploration.

Generation Pipeline

The dataset is produced through a four-stage automated pipeline:

• Stage 1: Image-to-Image Translation — Binary wheel masks (from reference designs and topology optimization) are converted to photo-realistic 512×512 RGB renderings using Stable Diffusion with text-conditioned generation
• Stage 2: Depth Prediction — A fine-tuned Marigold depth estimation model predicts depth maps from the 2D renderings, generating 2.5D representations of each design
• Stage 3: 2D-to-3D Reconstruction — Representative designs are sampled via Latin Hypercube Sampling (LHS) on depth-based embeddings and reconstructed into watertight 3D meshes using depth-to-point-cloud conversion and marching cubes
• Stage 4: CAE Simulation — Reconstructed meshes are converted to STEP format, meshed with 10-node tetrahedral elements (Tet10), and analyzed via modal analysis using Altair OptiStruct

Applications

• Surrogate model training for structural performance prediction (mass, natural frequencies)
• Data-driven inverse design for automotive wheel components
• Design space exploration and diversity analysis
• Depth estimation model fine-tuning for engineering domains
• Benchmarking 2D-to-3D reconstruction methods
• Product visualization, 3D printing, and simulation

Structural Analysis Details

All 3D models are analyzed under free-free boundary conditions using modal analysis. The material properties reflect A356-T6 aluminum alloy, a standard cast aluminum used in automotive wheel production:

• Density: 2,680 kg/m³
• Elastic modulus: 72 GPa
• Poisson's ratio: 0.33
• Yield strength: 175 MPa
• Ultimate tensile strength: 250 MPa

Performance metrics include mass, 1st mode natural frequency (rim mode), and 5th mode natural frequency (spoke mode), which are key indicators for NVH performance and structural fatigue.

Dataset Contents

Category	Path	Description
Rendered Images	1_rendered_images/	6,249 photo-realistic RGB images (512×512) generated via Stable Diffusion
Predicted Depth Maps	2_predicted_depth_maps/	6,249 single-channel grayscale depth maps aligned with rendered images
3D Reconstruction Meshes	3_3D_recon_meshes/	904 STL mesh files reconstructed from predicted depth maps
3D CAD Models	4_3D_cad_models/	904 STEP files converted from reconstructed meshes for CAE analysis
Simulation Results	deepwheel_sim_results.csv	Modal analysis results including mass and natural frequencies

Design Diversity

The dataset includes designs from two sources — existing reference wheel designs found on the market and topology-optimized structures — resulting in a broad and diverse design space. Topology optimization contributed a 7.6% increase in design space diversity and a 21.0% increase in performance space diversity compared to reference-only designs. Depth-based embeddings are used for balanced sampling, outperforming RGB-based approaches for clustering and exploration tasks.

Citation

If you use this dataset in your research, please cite: