공학해석용 격자 생성을 위한 기계학습 기반 방법론

Abstract

CAE has been widely used to reduce time and cost for product design and development in various industries. A typical CAE process, starting after CAD model creation, comprises of preprocessing, solving, and postprocessing steps. In the preprocessing step, mesh models are generated considering CAD geometry and material properties as well as boundary conditions (external forces and constraints). In the solving step, analysis solvers are executed with the mesh models and the boundary conditions. In the postprocessing step, the results are presented and visualized for review. In order to obtain the desired quality of results, mesh models need to be carefully generated by considering mesh grid size, geometric shape, and boundary conditions. However, the generation of such mesh models requires an iterative process of mesh generation, solving, result review, mesh modification steps, which is not only time and effort consuming, but also demanding some CAE related expertise. In this dissertation, we propose a novel machine learning-based methodology for making good use of conventional mesh generation modules to generate mesh models appropriate for CAE analysis, without any iterative process and deep knowledge of CAE. The methodology comprises of heuristic and machine learning methods. The heuristic method provides an overall frame for mesh generation and determines proper values of global and local parameters required for running a mesh generation module. The global parameters are used for mesh generation based on CAD geometry, and the local parameters for mesh refinement considering both on CAD geometry and boundary conditions. In this study, MeshGems is used as the module. After investigating the characteristics of the parameters of the MeshGems module and their major effects on mesh quality, we have selected global and local parameters, and developed the heuristic method to determine their proper values. The machine learning method, consisting of two-step models, is used to determine local refinement regions, more specifically, geometric entities around which a mesh is refined with smaller and denser elements. The first step model is a machine learning model used to classify CAD models into part families with boundary conditions, and the second step models, separately developed for each pair of a part family and a boundary condition, are machine learning models used to determine local refinement regions of the CAD model. Based on the local refinement regions, we make the heuristic method determine local parameters used for mesh refinement. After establishing how to represent the input and output data of the machine learning models, we have developed a testbed to manage the input/output data. From five part families with similar shape and three different boundary conditions for each family, we have generated more than 300 similar CAD models for each family, from which we have obtained a set of learning data. Then, we have developed multi-layer perceptron (MLP) and convolutional neural network (CNN) models both for part family classification and for estimating local refinement regions. Using TensorFlow, we have implemented the machine learning models and conducted machine learning with the learning data. We have found from experimental comparison that the CNN models outperform the MLP models both for part family classification and for estimating local refinement regions. Finally, we have applied the proposed methodology to generate mesh models from simple CAD models with boundary conditions, performed CAE analysis with the mesh models, and verified results, from which we have found the usefulness and potential possibilities of the methodology.