Many applications rely on scanned data, which can come from a variety of sources: optical scanners, coordinate measuring machines, or medical imaging. We assume that the data input to these applications is an unorganized point cloud or mesh of vertices. The objective may be to find particular features (medical diagnostics or reverse engineering) or comparison to some reference geometry (e.g. dimensional metrology). This paper focuses on the feature fitting of a segmented point cloud, specifically for branched, organic structures or structural frames, and targets non-monolithic geometries. In this paper, a methodology is presented for the automated recovery of cross-sectional shapes using centrally located curves. We assume a triangulated surface mesh is generated from the scanned point cloud. This surface mesh is the starting point for our methodology. We then find the curve skeleton of the part which abstractly describes the global geometry and topology. Next after segmenting the curve skeleton into non-branching segments, orthogonal planes to the curve skeleton segments, at preset or adaptive intervals, make slices through the surface mesh edges. The intersection points are extracted creating a 2D point cloud of the cross section. A number of application specific post-processing operations can be performed after obtaining the 2D point cloud cross sections and the curve skeleton paths including: calculations such as area or area moments of inertia, feature fitting or recognition, and digital shape reconstruction. Case studies are presented to demonstrate capabilities and limitations, and to provide insight into appropriate uses and adaptations for the presented methodology. Highlights Automated cross-sectional extraction for branching structures is presented. Methodology utilized skeletonization of object as reference for sampling planes. Surface mesh is sliced to extract a 2D point cloud. Filter algorithm for exclusion of peripheral slicing is presented. Several case studies demonstrate capabilities and limitations of the method.

Topology Optimization (TO) is used in the initial design phase to optimize certain objective functions under given boundary conditions by finding suitable material distributions in a specified design domain. Currently available methods in industry work very efficiently to get topologically-optimized design concepts under static and dynamic load cases. However, conventional methods do not address the designer's preferences about the final material layout in the optimized design. In practice, the final design might be required to have a certain degree of local or global structural similarity with an already present good reference design because of economic, manufacturing and assembly limitations or the desire to re-use parts in different systems. In this article, a heuristic Energy Scaling Method (ESM) for similarity-driven TO under static as well as dynamic loading conditions is presented and thoroughly evaluated. A 2D cantilever beam under static point load is used to show that the proposed method can be coupled with gradient-based and also heuristic, non-gradient methods to get designs of varying similarity w.r.t. a reference design. Further testing of the proposed method for similarity-driven TO on a 2D crash test case and a large-scale 3D hood model of a car body indicates the effectiveness of the method for a wide range of problems in the industry. Finally, the application of similarity-driven TO is further extended to show that ESM also has the potential for sensitivity analysis of performance w.r.t. the extension of design domain.

Machine learning is opening up new ways of optimizing designs but it requires large data sets for training and verification. While such data sets already exist for financial, sales and business applications, this is not the case for engineering product design data. This paper discusses our efforts in curating a large Computer Aided Design (CAD) data set with desired variety and validity for automotive body structural compositions. Manual creation of 60,000 CAD variants is obviously not viable so we examine several approaches that can be automated with commercial CAD systems such as Parametric Design, Feature Based Design, Design Tables/Catalogs of Variants and Macros. We discuss pros and cons of each method and how we devised a combination of these approaches. This hybrid approach was used in association with DOE tables. Since the geometric configurations and characteristics need to be correlated to performance (structural integrity), the paper also demonstrates automated workflows to perform FEA on CAD models generated. Key simulation results can then be associated with CAD geometry and, for example, processes using machine learning algorithms for both supervised and unsupervised learning. The information obtained from the application of such methods to historical CAD models may help to understand the reasoning behind experiential design decisions. With the increase in computing power and network speed, such datasets together with novel machine learning methods, could assist in generating better designs, which could potentially be obtained by a combination of existing ones, or might provide insights into completely new design concepts meeting or exceeding the performance requirements.

No abstract

Large-scale, high-quality data sets are central to the development of advanced machine learning techniques that increase the effectiveness of existing optimization methods or even inspire novel ones. Especially in the engineering domain such high quality data sets are rare due to confidentiality concerns and generation costs, be it computational or manual efforts. We therefore introduce the OSU-Honda Automobile Hood Dataset (CarHoods10k), an industry-grade 3D vehicle hood data set of over 10000 shapes along with mechanical performance data that were validated against real-world hood designs by industry experts. CarHoods10k offers researchers and practitioners the unique opportunity to develop novel methods on realistic data with relevance to real-world vehicle design. To illustrating central use cases, we first apply methods from geometric deep learning to learn a compact latent representation for design space exploration. Second, we use machine learning models to predict mechanical hood performance from the learned latent representation. We thus demonstrate the effectiveness of machine learning for building metamodels, which are used in design optimization whenever possible to replace costly engineering simulations. Third, we integrate CarHoods10k in a topology optimization approach based on evolutionary algorithms to demonstrate its capability to search for high-performing structures, while maintaining manufacturability constraints.