Research on multi-stage topology optimization method based on latent diffusion model

Abstract

Topology optimization aims to identify the optimal material distribution to enhance structural performance. Traditional methods such as Solid Isotropic Material with Penalization (SIMP) involve extensive finite element iterative calculations, limiting their ability to address complex or large-scale problems. This paper introduces a novel topology optimization method based on latent diffusion models (TOLDM). It is a multi-stage topology optimization approach that incorporates latent diffusion models with the SIMP method. TOLDM reduces the dimensionality and memory usage of data processing by shifting the diffusion process from the image space to a lower-dimensional latent space, decreasing the reliance on high-performance computing resources and accelerating the training and inference processes. Moreover, by incorporating a cross-attention module, the proposed method introduces physical conditions of various modalities into the diffusion process, enhancing the feasibility and manufacturability of the generated topological structures. This integrated optimization strategy not only improves the diversity and efficiency of structural designs but also achieves an optimized balance between resource consumption and performance. We have also compared TOLDM with other methods based on deep generative models, demonstrating its superiority. Furthermore, by incorporating transfer learning, TOLDM was successfully applied to the structural optimization of automobile engine hoods, enabling the generation of efficient and manufacturable topological structures. This indicates that TOLDM has high applicability in resource-constrained engineering fields, effectively advancing the practical implementation of topology optimization technology and expanding its technical boundaries.