Wall modeled computational fluid dynamics using machine learning. Using CFD coupled machine learning to create wall models for increased LES knowledge for future industrial use

Abstract

Machine learning is becoming a useful tool in many parts of engineering and science and computational fluid dynamics is no exception. At the same time the need for more accurate simulations with resolved turbulence is also increasing. But to make turbulence resolving methods such as Large Eddie Simulations (LES) accessible for industrial use computational speed up is still required. One method is to employ wall modeled LES. Classical wall models works well for simple flows but they get inaccurate for flows containing adverse pressure gradient. A solution to this problem is to use a machine learning based wall model.

In this project a machine learning based method for wall models is investigated and a new proposed way of coupling a CFD solver to the training process is tested. The method combines a supervised learning method with an unsupervised learning approach to take numerical inaccuracies of the CFD solver into the wall model. The models in this project were trained on channel flow and boundary layer flow. The results shown that machine learning based methods do work and give accurate results. These models are compared to classical wall functions and they can be more accurate. Some problems in the training process was found and especially the training time could become unreasonable if coupled with a LES solver. Solutions to these problems are discussed and in the future more complicated flows with adverse pressure gradients and LES simulations should be investigated.