Machine learning-assisted flow control based on an Iterative Linear System Identification

High dimensionality and strong nonlinearity of fluid flow phenomena pose challenges for designing control laws. One promising approach to control such phenomena is low-dimensional modeling. However, flow control remains challenging when the low-dimensionalized dynamics is a nonlinear system. We propose an effective control strategy based on the Iterative Linear System Identification (ILSI). The ILSI extracts linear systems through iterative application of a linear system extraction autoencoder (LEAE), which consists of a convolutional neural network-based autoencoder (CNN-AE) and a linear ODE layer extracting the governing equation of the latent dynamics in the form of a linear ODE. Based on the obtained linear governing equation for the uncontrolled flow, we construct an initial control law and perform the desired control. Then, to obtain more effective control law, we repeat the process of obtaining new linear equations and control laws by applying LEAE to the controlled flow fields. This method is applied to a two-dimensional flow around a circular cylinder with a blowing and suction mechanism at Re_D=100, and the control law is constructed to suppress the drag on the cylinder. In the talk, we will introduce the control effect of the application of ILSI and the influence of the location of the blowing and suction on the control performance.