Mesh Adaptation Using Machine Learning

Abstract

The quality of the mesh plays an important factor in the accuracy and efficiency of computational fluid dynamics (CFD) simulations. Mesh adaptation techniques are typically computationally expensive or rely on an experienced professional to identify the locations for refinement. The focus of this research is to show that a deep neural network can learn the mesh adaptation patterns needed to accurately imitate complex mesh adaptation indicators, therefore saving computational resources by avoiding costly mathematical functions such as taking derivates and replacing them with a combination of simple matrix multiplications. For this study, two different machine learning models will be trained. The first is trained through a modification of the Larsson and Toosi error indicator by focusing only on the averaged flow field. The second model is trained using the true Larsson and Toosi error indicator based on the unsteady flow fields. In both cases, the averaged flow field serve as the input data. The necessary model parameter choices will be explained, and the final model architecture will be described. The results show that machine learning models can produce a pattern that accurately represent the original adaptation indicators using either unsteady flow fields or the averaged flow field.