Author(s): Seongeun Choi; Jin Hwan Hwang

Linked Author(s): Seongeun Choi, Jin Hwan Hwang

Keywords: Deep learning; Convolutional Neural Network; Jet; Sub-grid scale

Abstract: The jet is a common phenomenon such as the disposal of treated wastewater through outfall, and it is a complicated flow. This phenomenon usually focuses on improving the mixing. Therefore, the study related to the jet has been conducted with lots of methods such as hydraulic models and field experiments. The numerical methods are used in this study since it relatively consumes less space, time, and labor costs. The representative numerical methods include Direct Numerical Simulation (DNS), Large Eddy Simulation (LES), and Reynolds Averaged Navier Stokes (RANS). Reynolds Averaged Navier Stokes (RANS) yields only time-averaged flow properties, and it has a limitation in reproducing the complexity of the jet. On the contrast, DNS does not require turbulence modeling because it resolves all scales of eddies directly. However, it requires a large number of grids and takes much time. LES resolves the larger-scale eddies directly and requires the model for the smaller-scale eddies. But it has slightly different results depending on the sub-grid model. So, DNS and LES models are appropriate methods to realize the instantaneous flow. . But it has somewhat different results depending on the sub-grid model with DNS results. This study proposes a method to use the results of DNS learned by deep learning without using the existing sub-grid models of LES. There are some models in deep learning, such as Convolutional Neural Network (CNN), Recurrent Neural Networks (RNN), and Artificial Neural Network (ANN). Among them, CNN has filters named kernels, and they are used to extract the features from the input using the convolution operation. It means that CNN can capture the spatial features from an image. Also, CNN has attractive properties such as local connectivity, nonlinear embedded mapping, and parameter sharing. Therefore, the study will use CNN to learn the turbulent properties of flows and be applied flexibly to the fluid flow with similar dominant features. This study is aimed to pursue two objectives; (1) Comparing the LES and DNS energy spectrum and finding the wavenumber of sub-grid scale in LES; (2) Optimizing the algorithm for jet flow and combining the resolved area in LES and algorithm.

DOI: https://doi.org/10.3850/IAHR-39WC2521711920221074

Year: 2022