Identifying topological prototypes using deep point cloud autoencoder networks

Data mining of engineering designs generated by topology optimization methods is a challenging task. A topology optimization method maximizes one or more performance objectives by redistributing the material in a design space for a given set of boundary conditions and constraints. The performance objective could be stiffness and the constraint could be the allowed mass fraction in the design space. If the constraints are not too restrictive or unknown at the initial stages of the design process, multiple feasible designs are possible and are generally generated to inspire new design ideas. Since a designer cannot manually review all the designs, one needs to select a few representative and interesting topologies based on performance or geometric features. In this paper, we propose a method to group designs with similar geometric features which are extracted from a point cloud representation of the geometry using a deep autoencoder network. The point cloud representation is generated by sampling points on the surface of geometry. The extracted features - called latent code - can be used to cluster topologies and identify prototypes in each cluster. The proposed method could be used on designs generated by topology optimization and other design generation methods. To evaluate the method, we use it on complex truss-like topology datasets with prespecified design types which have been recognized by the proposed method with high precision and recall. Also, the prototypes of the different categories are identified.