Yuto Koroyasu, "Exploring 3D Point Cloud Autoencoders as Generative Models in Similarity-based Topology Optimization", Technical University of Munich, 2023.

Topology Optimization (TO) is a powerful computational technique utilized to optimize the material distribution in a design space and create 3D structures that maximize their mechanical performance under a set of constraints. When applied during early stages of product development, TO enables rapid creation of design concepts and reduces the number of design modifications. However, in industrial settings, engineers often neglect the manufacturing constraints (e.g.,costs) in the TO, which leads to organic geometrical components that increase the manufacturing costs and require adaptations before being manufactured. Here we propose TO framework driven by a similarity metric based on features learned by a 3D point cloud autoencoder (AE). In this framework, we generate cost-optimized designs using a 3D point cloud AE, which are transferred to the TO domain as reference shapes and utilized to guide the density updates during the topology optimization. Hence, our approach allows engineers to generate cost-optimized reference shapes based on existing sets of designs, which can be exploited in similarity-based topology optimization. Furthermore, this framework utilizes the distance in the latent space learned by the AE to adaptively scale the strain energy density and balance design similarity with mechanical performance of the structure. We evaluate our method on a set of benchmark jet bracket structural optimization problems with respect to both cost and mechanical performance. Our results demonstrate how the proposed framework guides TO the reference shape generated by the AE, and consequently produces a structure with a lower cost while maintaining mechanical performance.