Machine-learned super-resolution analysis of three-dimensional turbulent channel flow

Abstract

We use machine-learning-based super-resolution analysis to reconstruct high-resolution flow field data from grossly coarse low-resolution data, for three-dimensional fully developed turbulent channel flow at Re_τ = 180. The training data is obtained by three-dimensional direct numerical simulation (DNS). We use an average pooling operation used commonly in image tasks, to prepare the coarse input data set. As a machine learning model, the hybrid downsampled skip-connection multi-scale (DSC/MS) model based on convolutional neural network is utilized in this study. Remarkable about this model are its robustness against rotation/translation of the flow images and its ability to consider multi-scale property of turbulence. The super-resolved flow fields recovered through the proposed machine learning model are in agreement with the reference DNS data in terms of velocity color distributions, root mean squared values of velocity fluctuations and L₂ error norm defined as the difference between the reference DNS data and super-resolved flow field. The maximum wavenumbers of streamwise and spanwise energy spectrum recovered by machine learning are increased by the super-resolution reconstruction. The proposed method holds great potential for various applications in experimental and numerical situations to handle the fluid big data efficiently, e.g., PIV measurements and subgrid-scale modeling of large-eddy simulation.