TY - JOUR
T1 - A posteriori study on wall modeling in large eddy simulation using a nonlocal data-driven approach
AU - Tabe Jamaat, Golsa
AU - Hattori, Yuji
AU - Kawai, Soshi
N1 - Publisher Copyright:
© 2024 Author(s).
PY - 2024/6/1
Y1 - 2024/6/1
N2 - The feasibility of wall modeling in large eddy simulation (LES) using convolutional neural network (CNN) is investigated by embedding a data-driven wall model developed using CNN into the actual simulation. The training dataset for the data-driven wall model is provided by the direct numerical simulation of turbulent channel flow at R e τ = 400 . The data in the inner layer, excluding y + ≤ 10 , are used in the training process. The inputs of the CNN wall model are the velocity components, and the outputs of the wall model are the streamwise and spanwise components of the wall shear stress. An a priori test has already been carried out in our previous study to assess the potential of CNN in establishing a wall model, and the results have shown the reasonable accuracy of the CNN model in predicting the wall shear stress. In this study, the focus is on the a posteriori test, and the performance of the CNN wall model is investigated in the actual LES under various conditions. Initially, the model is used in a simulation with the same specifications as those used for obtaining the training dataset, and the effect of the wall-normal distance of the CNN model inputs is investigated. Then, the model is tested for coarser grid sizes and higher Reynolds number flows to check its generalizability. The performance of the model is also compared with one of the commonly used existing wall models, called ordinary differential equation (ODE)-based wall model. The results show that the CNN wall model has better accuracy in predicting the wall shear stress in the a posteriori test compared to the ODE-based wall model. Moreover, it is able to predict the flow statistics with reasonable accuracy for the wall-modeled LES under various conditions different from those of the training dataset.
AB - The feasibility of wall modeling in large eddy simulation (LES) using convolutional neural network (CNN) is investigated by embedding a data-driven wall model developed using CNN into the actual simulation. The training dataset for the data-driven wall model is provided by the direct numerical simulation of turbulent channel flow at R e τ = 400 . The data in the inner layer, excluding y + ≤ 10 , are used in the training process. The inputs of the CNN wall model are the velocity components, and the outputs of the wall model are the streamwise and spanwise components of the wall shear stress. An a priori test has already been carried out in our previous study to assess the potential of CNN in establishing a wall model, and the results have shown the reasonable accuracy of the CNN model in predicting the wall shear stress. In this study, the focus is on the a posteriori test, and the performance of the CNN wall model is investigated in the actual LES under various conditions. Initially, the model is used in a simulation with the same specifications as those used for obtaining the training dataset, and the effect of the wall-normal distance of the CNN model inputs is investigated. Then, the model is tested for coarser grid sizes and higher Reynolds number flows to check its generalizability. The performance of the model is also compared with one of the commonly used existing wall models, called ordinary differential equation (ODE)-based wall model. The results show that the CNN wall model has better accuracy in predicting the wall shear stress in the a posteriori test compared to the ODE-based wall model. Moreover, it is able to predict the flow statistics with reasonable accuracy for the wall-modeled LES under various conditions different from those of the training dataset.
UR - http://www.scopus.com/inward/record.url?scp=85197209186&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85197209186&partnerID=8YFLogxK
U2 - 10.1063/5.0210851
DO - 10.1063/5.0210851
M3 - Article
AN - SCOPUS:85197209186
SN - 1070-6631
VL - 36
JO - Physics of Fluids
JF - Physics of Fluids
IS - 6
M1 - 065164
ER -