Part-Aware Product Design Agent Using Deep Generative Network and Local Linear Embedding

Abstract

In this study, we present a data-driven generative design approach that can augment human creativity in product shape design with the objective of improving system performance. The approach 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) a low-fidelity evaluation module that can rapidly assess the engineering performance of 3D objects based on locally linear embedding (LLE). This approach has two unique features. First, it generates 3D meshes that can better capture surface details (e.g., smoothness and curvature) given individual parts' interconnection and constraints (i.e., part-aware), as opposed to generating holistic 3D shapes. Second, the LLE-based solver can assess the engineering performance of the generated 3D shapes to realize real-time evaluation. Our approach is applied to car design to reduce air drag for optimal aerodynamic performance.

Mapping artificial intelligence-based methods to engineering design stages: a focused literature review

“…Deep learning algorithms such as generative adversarial networks (GANs) (Goodfellow et al 2020) and autoencoders (Bank et al 2020) can effectively be used to encode the data in which they are trained and generate new data, and hence are applied to generate designs in 2D (Quan et al 2018;, 3D (Khan and Awan 2018;Shu et al 2020), and point cloud (Achlioptas et al 2018;Krahe et al 2020) formats. With such capabilities, GANs have also been leveraged for data-driven generative design models (Li et al 2021), synthesizing designs with (Krahe et al 2020). Such effective utilization of GANs has also been applied to generating designs for topological optimization of designs conducted in the subsequent detailed stage (Oh et al 2019;Kallioras and Lagaros 2020).…”

Design representation for performance evaluation of 3D shapes in structure-aware generative design

“…In the design literature, these deep generative models are often referred to as data-driven generative design (DDGD) methods. DDGD methods have been increasingly used to improve design creativity and facilitate conceptual design, such as airfoil design (Chen et al 2020; Chen & Ahmed 2021), car wheel design (Oh et al 2019; Yoo et al 2021) and car shape design (Li, Xie & Sha 2021, 2022). DDGD methods can learn to synthesize designs from data without explicit human configuration by training a deep neural network model and learning a latent vector space with a predefined (often reduced) dimensionality.…”

“…In this paper, we refer to them as traditional DDGD methods. Recently, there have been emerging interests in developing structure-aware DDGD methods (Chen & Fuge 2019; Mo et al 2019; Gao et al 2019 b ; Li et al 2021). Compared to traditional DDGD methods, structure-aware DDGD methods can handle complex geometries consisting of interconnected components and learn interdependencies between components (i.e.…”