Proceedings of the 54th Hawaii International Conference on System Sciences | 2021 Part-Aware Product Design Agent Using Deep Generative Network and Local Linear Embedding Xingang Li Charles Xie Zhenghui Sha* University of Arkansas Institute for Future Intelligence University of Arkansas [email protected] [email protected] [email protected] Abstract as intelligent design agents, particularly in product shape design. GD is a term for a class of tools that can In this study, we present a data-driven generative generate novel yet realistic designs by leveraging design approach that can augment human creativity in computational and manufacturing capabilities [5]. product shape design with the objective of improving Deep GD models can produce a large amount of new system performance. The approach consists of two 2D or 3D data [6–9] given a set of training data, which modules: 1) a 3D mesh generative design module that has shown promises in computer graphics and can generate part-aware 3D objects using variational computer vision. In the design field, deep GD models, auto-encoder (VAE), and 2) a low-fidelity evaluation like variational auto-encoder (VAE) [10] and module that can rapidly assess the engineering generative adversarial network (GAN) [11], have been performance of 3D objects based on locally linear used to assist human designers in generating novel and embedding (LLE). This approach has two unique realistic designs [12–14] for design conceptualization. features. First, it generates 3D meshes that can better Building upon existing models, we develop a data- capture surface details (e.g., smoothness and driven GD approach in this study for product shape curvature) given individual parts’ interconnection and generation based on deep neural networks. Our constraints (i.e., part-aware), as opposed to assumption is that existing product designs (e.g., cars, generating holistic 3D shapes. Second, the LLE-based chairs, tables, etc.) on the market must have gone solver can assess the engineering performance of the through a complete design cycle, so both their generated 3D shapes to realize real-time evaluation. appearances or functionalities are optimized. Using a Our approach is applied to car design to reduce air deep GD approach with existing designs as training drag for optimal aerodynamic performance. data, it is expected that the generated design candidates would be promising ones. Also, learning- 1. Introduction and Motivation based GD methods have the potentials to reduce the high dependencies on design expertise because With advances in Artificial intelligence (AI), AI machines learn design knowledge in advance, which has shown its capabilities in many “human” jobs, like will assist the designer in realizing design automation. speech translation, customer service, and even However, realizing this idea in engineering design decision-making. Dellermann et al. [1] argue that, in is challenging. In engineering design, a product is a the following decades, AI will not replace but rather system that consists of interconnected components. collaborate with humans in most domains. They treat Traditionally, the design of such systems starts from this human-AI collaboration as hybrid intelligence, the system requirements analysis and is driven by a which leverages the complementary strengths of top-down hierarchical decomposition, followed by the human intelligence and AI. In the design field, AI has design of subsystems and components. Each also greatly facilitated human designers’ decision- component in a system is first designed separately and making in different design processes. For example, finally integrated into a complete system and validated researchers have successfully embedded intelligent against system-level requirements. Most existing GD agents into traditional computer-aided design (CAD) [15–20] are focused on generating holistic 3D shapes software (e.g., [2]) and some custom research without considering the structural dependencies or platforms (e.g., [3,4]) in support of conceptual design relations between components (e.g., a car is treated as and design optimization. This could save a large a whole piece instead of dividing it into the car body, amount of human labor, thus significantly shorten the mirrors, etc.). Even if efforts have been taken to cycle and iteration of product design and development. generate part-aware 3D shapes [6, 21, 22], they ignore Among various AI techniques, generative design the evaluation of the engineering performance of the (GD) models using deep neural networks can be used generated 3D shapes. However, assessing the * Corresponding author URI: https://hdl.handle.net/10125/71258 978-0-9981331-4-0 Page 5250 (CC BY-NC-ND 4.0) engineering performance of a product is essential in Our approach applies a deep generative model as an engineering design. intelligent agent, which learns existing designs on the Nowadays, most industries use computer-aided market to generate a large number of designs for users. engineering (CAE) tools, such as finite element analysis (FEA) and computational fluid dynamics 2.2. Deep generative models (CFD) software, to evaluate the engineering Deep generative models aim at synthesizing new performance of a preliminary design before the samples using the distribution learned from the physical prototyping and testing. Nonetheless, the existing data. The strategy of deep generative models engineering evaluation is costly. For instance, the is trying to approximate a distribution as similar to the assessment of the aerodynamic performance of a 3D true data distribution as possible by using a multi-layer car model using CFD software could take hours. of neural networks. GAN [11] and VAE [10] are the Therefore, it is impractical to evaluate every single two most widely used deep generative models. design candidate, let alone the vast number of design GAN consists of two parts: a generator and a alternatives obtained from GD models. A fast discriminator. The discriminator tries to distinguish engineering evaluation method is needed in realizing the data generated by the generator from the training GD in engineering design. data. In contrast, the generator aims to fool the To address these challenges, this study develops a discriminator with data that are highly similar to the new GD approach for engineering systems design that training data. They compete with each other in the integrates fast engineering evaluation and deep training process, driving the generator to produce data generative models that allows the generation of part- as identical to the training data as possible. GANs have aware 3D meshes. We validate and demonstrate the achieved success in 2D and 3D visualizations and effectiveness of our approach through an aerodynamic reconstructions [7,8]. However, since the car design problem. The remainder of this paper is discriminator judges the generated data based on organized as below. Section 2 gives a literature review distance metrics, the generator can synthesize of the relevant research. The details of the proposed unrealistic data (e.g., a face with a displaced mouth). approach are introduced in Section 3. Section 4 The basic idea of VAE is to find a hidden presents the results and discussion, and the paper is representation of the training data using low- concluded in Section 5, in which we also summarize dimensional latent variables. Those latent variables the closing insights and future work. contain information like specific structural and semantic properties of the training data. Compared to 2. Literature Review GANs, the training of VAEs is faster and easier via The review presented in this section is relevant to backpropagation, thus gaining increasing popularity the literature in the fields of intelligent design agents, [23]. VAEs have also been successfully applied in deep generative models, 3D shape synthesis, and data- both 2D images [9] and 3D models [6]. driven CFD evaluation methods. 2.3. 3D shape synthesis 2.1. Intelligent design agents The increasing availability of large 3D shape Rules-driven parametric design tools have datasets, like ShapeNet [24], provides a large amount introduced the generative design module to enable of training. the deep generative methods have been automatic design exploration. Users usually set the applied in object detection [25,26], classification constraints and requirements for their designs, and [27,28], and semantic segmentation [29,30]. They can those tools can then run hundreds of simulations to also generate diverse 3D objects in various generate various designs for users to select. There are representations, such as point cloud [15,16], voxels also learning-based/data-driven design platforms. For [17–19], and meshes [20,31]. Compared to point cloud example, Hu and Taylor [3] developed a CAD and voxels, meshes can better capture the geometric platform with an intelligent tutoring system. The details (e.g., smoothness, curvature) of 3D objects system can first learn all possible ways to design a 3D without consuming large storage space. So, the mesh model and then instruct users to draw 3D models. representation is more suitable for engineering design In rules-driven parametric design tools, setting applications (e.g., the representation of automobiles). proper constraints and requirements for the design However, 3D meshes from open-source datasets requires a high degree of domain expertise, which is are usually non-manifold triangles that may contain not friendly to novice designers or fast design