ConStruct-LDM: Conditional Latent Diffusion Model for Structured Design

Abstract

Engineering design is a complex process that involves balancing various performance trade-offs, requiring both the expertise of skilled designers and significant costs. In recent years, the introduction of machine learning techniques has raised expectations for automating the design process and reducing associated costs. However, the high dimensionality of design spaces and the scarcity of training data remain major obstacles to practical deployment. This paper proposes a novel method for generating diverse and realistic design candidates in a conditional manner by leveraging image generation techniques, based on the observation that three-dimensional (3D) shapes can be represented as two-dimensional (2D) images without significant loss of information. The proposed model, named ConStruct-LDM, effectively extracts global features of shapes using VQ-VAE-2 and performs efficient generation in the latent space via a Latent Diffusion Model. Additionally, we introduce zero-initialization for conditional inputs to enhance training stability. We apply the proposed method to hull form design and demonstrate substantial improvements in all evaluation metrics compared to conventional approaches. Furthermore, we show that the generated shapes can serve as effective data augmentation for downstream tasks, highlighting the practicality of the method in real-world design workflows. Our approach enables both the acceleration of the design process and the expansion of the design space, indicating the potential applicability of generative AI not only in ship design but also in other engineering shape design domains.