Part-Aware Product Design Using Deep Generative Network

We propose a data-driven generative design framework that can facilitate the design exploration of engineering products in support of creative conceptual design and optimization. The framework consists of two modules: 1) a 3D mesh generative design module that can generate part-aware 3D objects using variational auto-encoder (VAE), and 2) an evaluation module that can assess the engineering performance of 3D objects in real-time based on locally linear embedding (LLE) [1]. The novelty of our framework lies in three aspects. First, it generates 3D shapes with the consideration of individual parts’ interconnection and constraints (i.e., part-aware) as opposed to generating a holistic 3D shape. Second, it transfers 3D shapes to watertight meshes with the same connectivity (i.e., the same number of vertices and the same way how vertices are connected) so that the surface details (e.g., smoothness, curvature) can be captured by mesh representation and learned by neural networks. Third, the LLE-based solver can quickly assess the engineering performance of the generated 3D shapes to realize real-time evaluation. In this study, we apply our framework to the aerodynamic car design.