Towards a priori mesh quality estimation using Machine Learning Techniques

Abstract

Since the quality of FE meshes strongly affects the quality of the FE simulations, it is known to be very important to generate good quality meshes. Thus, it is crucial to be able to estimate very early what can be the expected quality of a mesh without having to play in loop with several control parameters. This paper addresses the way the quality of FE meshes can be estimated a priori, i.e. before meshing the CAD models. In this way, designers can generate good quality meshes at first glance. Our approach is based on the use of a set of rules which allow estimating what will be the mesh quality according to the shape characteristics of the CAD model to be meshed. Those rules are built using Machine Learning Techniques, notably classification ones, which analyse a huge amount of configurations for which the shape characteristics of both the CAD models and meshes are known. For an unknown configuration, i.e. for a CAD model not yet meshed, the learnt rules help understanding what can be the expected classes of quality, or in another way what are the control parameters to be set up to reach a given mesh quality. The proposed approach has been implemented and tested on academic and industrial examples.