Physics-informed Neural-operator Predictive Control for Drag Reduction in Turbulent Flows

Turbulence control for wall friction reduction poses a significant challenge due to computational costs associated with modeling turbulent dynamics. In this work, we present a learning and control scheme termed physics-informed neural-operator-based predictive control (PINO-PC). Our method uses a Neural Operator framework that enables accurate learning and control of turbulent flows. It carries out predictive control where both the policy and the observer model for turbulence control are learned jointly. We show that PINO-PC outperforms prior model-free reinforcement learning methods with a stable policy learning procedure. We experiment with various challenging generalized scenarios where flows are of unseen high Reynolds numbers, and we find that our method achieves a drag reduction of 39.0% under a bulk-velocity Reynolds number of 15k, outperforming previous fluid control methods by more than 32%.