Exploring the Potentials and Challenges of Deep Generative Models in Product Design Conception

Deep Generative Models (DGMs) offer potential for automating and streamlining product design conception, enhancing innovation and efficiency. However, their adoption in this field is limited due to challenges in handling specific modalities like requirement tables and sketches, lack of robustness and reliability in outputs, complexity for non-expert users, and rapid evolution requiring substantial data and computational resources. This paper explores reasons for this limited application and outlines requirements for successful integration of DGMs into product design. DGM families, including Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), Diffusion Models, Transformers, and Radiance Fields, are analyzed for their strengths, weaknesses, and general applicability. The objective is to provide insights for engineers to determine the most effective method for their specific challenges. The paper aims to contribute to a fundamental understanding of DGMs in product design conception, proposing potential solutions and offering a roadmap for leveraging these technologies. Key challenges in integrating DGMs into existing workflows include the need for precise domain-specific knowledge and the rapid evolution of the field, which restricts accessibility and scalability. The study focuses on deriving fundamental requirements for effective DGM integration, which will help evaluate model families, assess their suitability for tasks within product design, and guide practitioners in selecting appropriate models for specific problems. These requirements form the basis for the technical analysis of DGM families and provide application recommendations for non-DGM experts. The text discusses the application of DGMs in product design conception, focusing on the early-phase process of defining and translating engineering requirements into functional representations. The concept development phase involves refining and comparing alternative solutions, with a focus on representations depicting the exterior design of product concepts. The study highlights the importance of 2D-shape, image, and 3D-object representations in the design process, each with varying levels of information richness and challenges in synthesis. VAEs and GANs are recommended for product design concept exploration, with VAEs excelling in generating novel design variations within a given design space and GANs offering higher detail compared to VAEs. The text also discusses the potential of diffusion models for generating high-quality, detailed visual content, with transformer-based models showing promise in enhancing real-world coherence in generated content. Radiance Field models are noted for enabling dynamic 3D scene rendering without explicit geometry, suitable for product concept visualization. The study concludes by emphasizing the transformative potential of DGMs in product design and construction, but also highlights the need for technological innovations and accessibility improvements to fully integrate these models into traditional design processes. Key factors for successful integration include data availability, computational resources, target representation adequacy, conditioning mechanisms, expected accuracy levels, and suitable representation. The text also discusses the importance of developing specific datasets for engineering-oriented domains, addressing the lack of appropriate metrics for evaluating model performance in PDC applications, and the need for benchmark challenges aligned with existing benchmarks in image and 3