Speaker
Description
The pulsed ejection of air into a boundary layer represents an effective method to prevent or delay its separation from the surface. Despite many successful demonstrations, this approach has not reached a breakthrough in the aviation industry, which is arguably due to an excessive energy consumption.
Recent progress in the field of Deep Learning suggests that artificial neural networks can be used to great effect for the optimization of forcing parameters. In particular, reinforcement learning (RL) appears to be suited for this task, revealing robust control strategies through a bio-inspired interaction between the algorithm and the flow under consideration.
The objective in the current study is to test the capabilities of RL subjected to experimental conditions online (i.e., during a wind tunnel experiment). In a closed-loop, low-speed wind tunnel, the flow past a fully-turbulent backward-facing ramp is manipulated by an array of five pulsed-jet actuators relying on magnetic valves to introduce momentum in a pulsatile fashion. Based on partial insight into the instantaneous flow conditions via wall shear-stress measurements, the RL agent is trained to suppress flow separation. Specifically, the action space involves the decision whether to open or close the magnetic valve, thereby either forcing the flow at this specific time instance (open valve) or saving mass flow (closed valve).
This study will promote the implementation of RL for the benefit of controlling real-world, turbulent flow conditions. Furthermore, we expect to obtain interesting insight regarding the mechanisms of flow control by analysing the RL strategy.
