Deep reinforcement learning based control of the wake of a bluff body

We experimentally investigate the manipulation of the near wake of a bluff body using pulsed jets located along the periphery of the model’s base and injected through rectangular slits. The mass flow rate, the duty cycle, and the actuation frequency were the forcing parameters. The actuation is determined using a deep reinforcement learning (DRL) based approach. The neural network is trained using 60 pressure taps that populate the model’s base and the lateral surfaces for a set of freestream conditions while the reward is based on the direct measure of the model’s drag provided by a load cell. We investigate a range of agents, and we show that it learns an optimal forcing leading to drag reduction of the order of 9% within tens of episodes, corresponding to about 3 hours of experiments.

We investigate the capability of the network to adapt to changing conditions, such as different freestream speeds and switching from the case of isolated wake to a platoon marching configuration and the trained agent shows a very good capability to keep satisfactory values of drag reduction.

The comparison of the most energetic POD modes of the microphones located across the model’s base measured in the baseline case (no forcing) and the most favorable forcing condition, reveals that the wake would tend to a symmetric state, suppressing the large-scale oscillations associated with the shedding mechanism, as evidenced looking at the spectra of the temporal coefficients of the POD modes.