In this video, created by Pol Suárez, we summarize some of the results of our recent work:

[ Ссылка ]

We show the visualization of the (top) velocity and (bottom) pressure field of an actuated two-dimensional cylinder at Reynolds number Re=2,000. Deep neural networks trained through deep reinforcement learning (DRL), coupled with the numerical solver Alya, are used to perform active flow control (AFC). The control consists of two jets located at the highest and lowest points of the cylinder surface. The agent finds a strategy yielding 17% drag reduction. At this Reynolds number the control is very different from the ones obtained for Re=100 or 1,000. The jet actuation starts at t=50, and it can be observed how the jets clearly break the stabilized uncontrolled vortices into smaller and less-energetic ones thanks to high-frequency actuations. This control mechanism, obtained through the DRL agent, is similar to the drag-crisis phenomenon.

свернуть