Unsupervised deep learning of spatial organizations of coherent structures in a turbulent channel flow

Abstract

We examined the capability of an unsupervised deep learning network to capture the spatial organizations of large-scale structures in a cross-stream plane of a fully developed turbulent channel flow at Reτ≈180. For this purpose, a generative adversarial network (GAN) is trained using the instantaneous flow fields in the cross-stream plane obtained by a direct numerical simulation (DNS) to generate similar flow fields. Then, these flow fields are examined by focusing on the turbulent statistics and the spatial organizations of coherent structures. We extracted the intense regions of the streamwise velocity fluctuations (u) and the vortical structures in the cross-stream plane. Comparing the DNS and GAN flow fields, it is revealed that the network not only presents the one-point and two-point statistics quite accurately but also successfully predicts the structural characteristics hidden in the training dataset. We further explored the meandering motions of large-scale u structures by measuring their waviness in the cross-stream plane. It is shown that as the size of the u structures increases, they exhibit more aggressive waviness behavior which in turn increases the average number of vortical structures surrounding the low-momentum structures. The success of GAN in this study suggests its potential to predict similar information at a high Reynolds number and, thus, be utilized as an inflow turbulence generator to provide instantaneous boundary conditions for more complicated problems, such as turbulent boundary layers. This has the potential to greatly reduce the computational costs of DNS related to a required large computational domain at high Reynolds numbers.