Proefschrift
Deep Generative Models for Engineering Design
In this thesis, we investigate how deep generative models (AI models that can create new designs) can be better utilized for technical and industrial design. Our goal is to generate designs that are not only realistic (plausible), but also efficient, reliable, and usable in complex 2D and 3D applications.
- Auteur
- J. Fan
- Datum
- 24 maart 2026
- Links
- Thesis in Leiden Repository
We first demonstrate that the way noise is used in so-called diffusion models has a major impact on the quality of generated designs. We develop an analysis method to determine the optimal noise range for a dataset. We then show that many common evaluation methods are poorly aligned with human judgment. Therefore, we introduce a new assessment method that better considers the structural quality of designs.
Building on these methodological developments, we illustrate their practical relevance through a design autocompletion scenario in automotive engineering. We implement our tool in the BMW A-pillar design process, where engineers provide incomplete geometry and the model generates plausible initial design proposals. This significantly reduces the time required to complete a starting design and allows engineers to move more quickly toward downstream optimization and validation tasks, demonstrating how such generative tools can directly support early-stage engineering workflows.
We then shift the focus from 2D to 3D designs. We develop a method to greatly simplify complex 3D mesh shapes without losing important information, making them suitable for generative models. We also show how these models can generate three-dimensional design variations that can be explored in practice. As an example, we apply the approach to BMW car rim design, where the generative system produces diverse candidate geometries that designers can use for inspiration and exploration. Rather than replacing the designer, the system helps broaden the explored design space and supports overcoming creative bottlenecks during concept development.
Finally, we show how these models can directly generate CAD geometry (NURBS surfaces and B-Rep solids), which is an important step towards practical applications in industry. This representation enables integration with downstream engineering tools, and we theoretically demonstrate how generative modelling can be coupled with simulation workflows. In such a setup, simulation feedback can guide the model toward structurally improved designs, enabling iterative refinement driven by performance evaluation rather than purely geometric criteria.
Overall, this research bridges the gap between fundamental AI research and industrial design practice. The results contribute to more reliable and usable AI tools for designers and engineers, and demonstrate how generative AI can help develop technical products faster and smarter.