Efficient high-fidelity turbulence modeling using convolutional neural networks
| In the past few years, the field of Machine Learning (ML) in general, and Deep Learning (DL) in particular, has been getting a lot of attention, both due to the continued growth of computational power, and the generation of large datasets. The field of fluid mechanics is rich in large amount of datasets from experiments and simulations. Moreover, the fact that fluid flow often contains reoccurring patterns that can be studied and predicted makes fluid mechanics an excellent candidate for the use of ML. This talk discusses the ability of convolutional neural networks (CNN) to replace classic turbulence models for a specific set of problems efficiently. Although some attempts have been made to use neural networks to create a data-informed turbulence model, none have managed to do so efficiently for an unsteady case with walls. The unique approach of this work, necessitating fast predictions, required using a specific CNN architecture and the use of image processing algorithms. The model can be effectively used for problems that require a lot of high-fidelity simulations of flow fields that share similarities. Nevertheless, the chosen CNN would not necessarily improve cases substantially different from the ones used for the model dataset. The focus is on predicting the Reynolds stresses using CNNs. Specifically, U-Net . The chosen test case is a flow over a cylinder since it presents different patterns in a relatively low-Reynolds number range and has a lot of documentation for comparison. While most papers attempt to create a general model, this paper explores the possibility of a case-specific DL model, enabling the use of different methods from the ones used so far. Nevertheless, some generalization is still required from the model and will be tested. Specifically, its ability to generalize over various Reynolds numbers.
|
Light refreshments will be served before the lecture
