Reinforcement learning for bluff body active flow control in experiments and simulations.

Area of Science:

• Fluid mechanics
• Control theory
• Artificial intelligence

Background:

• Active flow control aims to manipulate fluid dynamics for desired outcomes.
• Turbulent flow control presents significant challenges due to its complexity and unpredictability.
• Reinforcement learning (RL) has shown promise in optimizing control strategies.

Purpose of the Study:

• To demonstrate the effectiveness of reinforcement learning (RL) in discovering active flow control strategies for drag reduction.
• To maximize power gain efficiency by optimizing the rotational speed of cylinders in a turbulent flow.
• To validate RL-discovered strategies through experimental and simulation-based approaches.

Main Methods:

• Utilized reinforcement learning (RL) agents to autonomously explore and identify control strategies.
• Conducted tens of towing tank experiments with automated control sequences.
• Employed noise reduction techniques and carefully defined reward functions for the RL agent.
• Verified experimental findings using computational fluid dynamics (CFD) simulations.

Main Results:

• RL agent successfully discovered a drag reduction strategy comparable to optimal strategies found through traditional methods.
• The discovered strategy significantly improved power gain efficiency in bluff body flow control.
• Simulations provided insights into the physical mechanisms underlying the drag reduction process.

Conclusions:

• Reinforcement learning is effective for discovering active flow control strategies in experimental settings, not just simulations.
• This study validates RL's potential for efficient exploration of advanced flow control solutions.
• The findings pave the way for applying RL to a broader range of complex fluid mechanics problems.