Predicting Coherent Turbulent Structures via Deep Learning

Abstract

Turbulent flow is widespread in many applications, such as airplane wings or turbine blades. Such flow is highly chaotic and impossible to predict far into the future. Some regions exhibit a coherent physical behavior in turbulent flow, satisfying specific properties; these regions are denoted as coherent structures. This work considers structures connected with the Reynolds stresses, which are essential quantities for modeling and understanding turbulent flows. Deep-learning techniques have recently had promising results for modeling turbulence, and here we investigate their capabilities for modeling coherent structures. We use data from a direct numerical simulation (DNS) of a turbulent channel flow to train a convolutional neural network (CNN) and predict the number and volume of the coherent structures in the channel over time. Overall, the performance of the CNN model is very good, with a satisfactory agreement between the predicted geometrical properties of the structures and those of the reference DNS data.

Identifying regions of importance in wall-bounded turbulence through explainable deep learning

“…A U-net architecture 24 is used for predicting the velocity field. This architecture efficiently exploits spatial correlations in the data, and further develops the work by Schmekel et al 30 . Note that U-nets and other computer-vision architectures have been successfully used in the context of turbulent-flow predictions 28,[64][65][66][67] .…”

“…To accomplish this objective, we will first show how U-nets can predict the evolution of turbulent channel flow, extending our earlier work 30 . We start with a database of 6000 instantaneous realizations obtained from turbulent channel flow simulations, see the “Methods” section for additional details on the data generation.…”

Explaining wall-bounded turbulence through deep learning

“…A convolutional neural network (CNN) is used for predicting the velocity field, as discussed by Schmekel et al [30]. Note that CNNs and other computer-vision architectures have been successfully used in the context of turbulent-flow predictions [28,[58][59][60][61].…”

“…The network employed in this work consists of 4 layers of 3D CNN blocks, which contain plain convolutional and residual blocks [62]. The architecture, similar to the one used by Schmekel et al [30], is shown in epochs, where all training data is used once in a single epoch.…”

“…To accomplish this objective, we will first show how CNNs can predict the evolution of turbulent channel flow, as documented in our earlier work [30].…”