Free-form Feature Classification for Finite Element Meshing based on Shape Descriptors and Machine Learning

Finite element analysis (FEA) has helped modern manufacturers create efficient and reliable product developments. Finite element mesh generation (FE meshing from 3D-CAD) models is generally the most important process in the FEA pipeline, therefore fully automated meshing that can secure analysis accuracy is strongly required to streamline the pipeline. Many manufacturers strictly prescribe FE meshing patterns for specific classes of free-form features on CAD models established company-specific FE meshing rules of where and how many node points of elements should be placed over and inside a form feature, to secure the analysis accuracy. Meshing rules for “boss” or “rib” features are often specially specified, as these play critical roles in securing strength for a part or transmitting forces between parts. As such, when an FE mesh is to be generated for a cylindrical boss feature, the node points of elements must be placed concentrically around a medial axis of the boss at an angle interval of 15 degrees. In the case of a rib feature, the node points must be arranged along a ridge curve on top of the rib at a maximum interval of 3.0mm. Therefore, it is crucial for manufacturers to develop software where features such as bosses or ribs with complex free-form surfaces can be extracted from CAD models and categorized under classes where meshing rules are prescribed and where an FE mesh for the feature region can be automatically generated according to rules realizing a high-quality and reliable FEA pipeline. To date, some feature recognition methods aimed for FE meshing have been studied. Lai et al. proposed a method that recognizes rib features from a B-rep CAD model by finding specific topological and geometrical patterns of virtual loops around these features and then decomposes them into regions that can be meshed with hexahedral or prismatic FE meshes. Lu et al. introduced a feature-based hexahedral meshing method, which decomposes a B-rep CAD model into a set of hex meshable volumes by extracting protrusion features bounded by concave zones through the identification of three loop types in a CAD model to serve as the feature boundaries. Moreover, Boussuge et al. presented a method for recognizing protrusion features on a CAD model whose shape can be partitioned into plate and shell elements. Unfortunately, these feature recognition methods cannot be directly applied to our case for the following reasons. First, the feature geometries discussed in the previous studies were basically 2.5-dimensional, consisted only of simple planes and cylinders, and were bounded by sharply concaved loops on a B-rep CAD model. In our study, however, the feature (i.e., boss) that needs recognition is designed as a portion of a casted or forged part’s surface whose geometry is generally defined by 3D free-form surfaces. Moreover, the feature is usually bounded by smooth free-form filet surfaces that are comparatively not discernible as the ones mentioned above. Second, feature classes for FE meshing are normally defined subjectively based on the knowledge of skilled FEA engineers, and they often differ from one company to another. On the contrary, the recognition algorithms of the previous studies were designed for the elaborate procedural search of loops on a B-rep CAD model and coded in an ad hoc manner to fit the recognition of specific feature classes. This way, the algorithm is not easily expanded when a new feature class is added or a current feature class is to be modified. Third, it was assumed in previous studies that an input B-rep CAD model is provided without any topological or geometric defect. However, it is well known that the data quality of CAD models may possibly degrade because of loss of information during the translation process and that some quality issues on the B-rep data (i.e., small cracks between faces) may be inducted. Therefore, a recognition algorithm relying mainly on the topological and geometrical search on the B-rep CAD model is more likely to fail. To solve the abovementioned issues, we propose an algorithm of the free-form feature classification for feature-based FE meshing, which we regard to consist of three steps: feature extraction from the CAD model, feature classification, and feature-compliant mesh generation. Our study focuses on the feature classification step. In principle, our algorithm accepts a triangular mesh of a free-form feature easily converted from a B-rep CAD model. Moreover, it identifies a feature class label, such as boss and rib, of the input mesh model via 3D shape descriptors, Bag-of-Features (BoF), and machine learning. The advantages of the proposed approach are summarized as follows: ? The local and global shape descriptors allow us to encode both the local and the global features’ geometry as a single multidimensional vector even when a feature has complex free-form shapes bounded by smooth filet surfaces. Moreover, the BoF technique facilitates the application of the shape descriptor representation to the machine learning scheme, thereby solving the first problem. ? The machine learning technique makes the design of the feature classification algorithm uniform and portable regardless of the classes. As such, the algorithm can be easily expanded by the addition of newly labeled training feature samples, thereby addressing the second problem. ? Instead of B-rep representation, it uses free-form features based on shape descriptors at the vertices on a triangular mesh, thus avoiding unstable feature extraction and classification processes caused by product data quality issues, which solves the third problem.