Recognition of Free-form Features for Finite Element Meshing using Deep Learning

Large-scale finite element models (FE models), having several million elements, require high-quality FE meshes in compliance with some company-specific meshing rules to ensure simulation accuracy. In these meshing rules, the meshing patterns, location and resolution are strictly defined for specific free-form features such as ribs and bosses. For example, as shown in Fig. 1, when an FE mesh is to be generated for a cylindrical boss feature, the node points of elements must be placed concentrically and evenly around a medial axis of the boss at specified number. However, the automatic feature compliant finite element meshing for CAD models is not yet fully supported in commercial CAE software. And it still requires many manual operations resulting in a high person-hour ratio to the whole CAE process. The feature-compliant finite element meshing mainly consists of a series of operations; 1) extracting the finite-element free-form feature shapes such as “ribs” from a given CAD model, 2) performing the segmentation of each feature shape into local feature areas such as “top”, “side”, and “fillet” areas, and 3) generating a finite-element mesh such that it complies with the specific meshing rules defined on the recognized feature shapes and areas. However, if the geometry of a given CAD model is large-scale and has high complexity, the operations relying on only the engineer’s decisions become incredibly stressful, time-consuming, and error-prone. Therefore, an automated feature shape extraction and feature area recognition technique targeting a feature-compliant finite element meshing is strongly required. There has been some research on the feature extraction techniques from CAD models to generate meshes of FE models. However, three main problems remain in them. First, the feature extraction algorithm does not work robustly when the CAD models include Product Data Quality (PDQ) issues such as cracked or degenerated geometries. Second, the free-form features surrounded by complicated and smooth boundaries, commonly found in casted or molded parts, remain challenging to detect by these techniques. Finally, the extraction algorithm must be designed in an ad-hoc way for different feature types and features with similar shapes. Thus, it is challenging to apply these previous feature extraction techniques when developing a feature-compliant finite element meshing. As a solution to these problems, the 3D shape-descriptor-based finite-element feature classification and feature extraction technique using the dense point cloud representation have been proposed by our research group. However, they provide the solutions only to the first operations in the feature-compliant finite element meshing process. Unlike these previous studies, this paper proposes a deep-learning approach to extract free-form feature shapes and perform the segmentation of each feature shape into local feature areas for finite element meshing from CAD models of a product. PointNet++ that is one of the popular convolutional neural network for 3D point cloud classification and segmentation is used to recognize our free-form feature shapes and local feature areas. The performance of the free-form feature recognition is experimentally verified in this paper.