PREDICTION AND CONTROL OF TURBULENT FLOWS USING DEEP LEARNING

In this study, prediction and control of 2D decaying homogeneous isotropic turbulence (DHIT) using deep learning have been performed as an example of a fundamental study to improve understanding of the dynamic behavior of turbulence. PredictionNet, a neural network used for prediction, showed high accuracy in the prediction of flow up to one integral time scale, T_L, with a correlation coefficient of 0.855. Our model based on generative adversarial network (GAN) also showed much higher accuracy in the enstrophy spectrum than convolutional neural network (CNN) model. Predicability depending on the scale is also analyzed using scale decomposition. Another neural network used for control, ControlNet, was able to generate disturbances that allow the time-evolution of the flow field to be in a direction that fits the objective function. In addition, it was possible to bring some physical understanding of the input vorticity fields through the analysis of the disturbance fields.