Prediction and control of turbulent channel flow with deep learning

We apply deep learning to predict, and then control turbulent channel flow for drag reduction. A well-known turbulence control problem, opposition control (Choi et al., 1994, JFM), is adopted for applying deep learning. We consider several deep learning techniques based on neural networks. Deep learning models are trained to predict wall-normal velocity at y⁺=10 from wall variables such as wall pressure and shear stresses with database of uncontrolled turbulent channel flow. Among various deep learning techniques, convolutional neural network trained with generative adversarial framework has the best prediction capability. Simple rescaling is applied to predict the flow being controlled, because deep learning models are trained with uncontrolled flow. With the wall-normal velocity predicted by the deep learning based on the wall variable measurements, we conduct active control and obtain a significant amount of skin friction reduction.