Learning in two dimensions and controlling in three: Generalizable drag reduction strategies for flows past circular cylinders through deep reinforcement learning

Abstract

We present the automated discovery of control strategies for drag reduction in cylinder flows. Reinforcement Learning algorithms discover control strategies for two-dimensional configurations that generalize to three dimensional flows. We discuss the physical processes involved in the drag reduction mechanisms along with their generalization capabilities. This work demonstrates a practical approach to handling the computationally intensive task of deploying Reinforcement Leaning for bluff body flow control problems: namely train in 2D and control in 3D.

Topics & Concepts