Machine Learning Control for Experimental Turbulent Flow Targeting the Reduction of a Recirculation Bubble

We experimentally control the turbulent flow over backward-facing step (Re_H = 31500). The goal is to modify the internal (Xr) and external (Lr) recirculation points and consequently the recirculation zone (Ar). A model-free machine learning control (MLC) is used as control logic. As benchmark, an optimized periodic forcing is employed. MLC generalizes periodic forcing by a multi-frequency actuation. In addition, a sensor-based control and a non-autonomous feedback, open- and closed-loop laws, were use to optimize the control. The MLC multi-frequency forcing outperforms, as expected, periodic forcing. The non-autonomous feedback brings a further improvement. The unforced and actuated flows have been investigated in real-time with a TSI particle image velocimetry (PIV) system. The current study shows that a generalization of multi-frequency forcing and sensor feedback significantly reduces the turbulent recirculation zone, far beyond optimized periodic forcing. The study suggests that MLC can effectively explore and optimize new feedback actuation mechanisms and we anticipate MLC to be a game changer in turbulence control.