WheelSDF: Implicit Representation-Based 3D Shape Generation and Interpolation for Wheel Design

Hayata Morita, Kohei Shintani, Chenyang Yuan, Frank Permenter2025 · DOI

Abstract

Abstract Surrogate models have enabled the rapid evaluation of the engineering performance of multiple design candidates at the conceptual design stage. However, training these data-driven surrogate models require a large amount of data paired with computational results for engineering performance metrics, making data shortage to be a significant challenge. A promising approach to overcome this limitation is to use generative models to augment and extend existing datasets. In this study, we explore the practical applicability of generative AI in the automotive industry in the context of 3D wheel design. This is a particularly challenging problem as wheels can be topologically different, with large variations in the number and size of openings. To overcome this problem and to handle such variations, we propose WheelSDF, an implicit representation-based generative AI model, with novel refinements to the sampling method and loss function during training. WheelSDF is able to generate high-fidelity shapes with sufficient accuracy for performance evaluation. In addition, we demonstrate how to use a pre-trained WheelSDF model to generate novel wheel shapes, interpolating between two topologically different geometries. By extracting engineering performance metrics from shapes, we show that this interpolation is smooth, illustrating its potential to generate new shapes along with their corresponding metrics. This demonstrates the potential of our proposed method to address the issue of data shortage when training surrogate models.
BibTeX
@misc{morita2025wheelsdf,
  author = {Hayata Morita and Kohei Shintani and Chenyang Yuan and Frank Permenter},
  title = {{WheelSDF: Implicit Representation-Based 3D Shape Generation and Interpolation for Wheel Design}},
  year = {2025},
  doi = {10.1115/detc2025-165362},
}

Loading related content...

Figure