METHODOLOGY TO ACHIEVE ROBUST CLOSED-LOOP TURBULENT FLOW CONTROL USING MACHINE LEARNING

A methodology to achieve robust closed-loop feedback control of a turbulent flow using machine learning is outlined. The chosen candidate system is a square cross-sectional cylinder with two moving surface actuators embedded in the windward face at the leading corners. A Long Short-Term Memory (LSTM) Neural Network is trained using motor actuation and pressure sensor data to forecast future system states. This LSTM model is then implemented with Model Predictive Control (MPC) in order to achieve closed-loop flow control. The derived controller performance was tested experimentally using three objective functions: recovery of mean-base pressure set-point from perturbation, and drag or wake fluctuation intensity optimizations. An adaptive learning algorithm, which adjusts the model to new Reynolds number (Re) conditions without user intervention, is implemented to extend controller performance and achieve robust control.