Deep reinforcement learning for turbulent drag reduction in channel flows

Area of Science:

• Fluid Dynamics and Control Theory
• Artificial Intelligence and Machine Learning

Background:

• Turbulent fluid flows present a long-standing challenge in physics and engineering.
• Traditional drag reduction strategies for turbulent flows are often physically grounded but limited in complexity.
• Deep reinforcement learning (DRL) offers a novel approach to handle high-dimensional data for complex system control.

Purpose of the Study:

• To introduce a novel reinforcement learning (RL) environment for designing and benchmarking drag reduction control strategies.
• To enable computationally efficient, high-fidelity fluid simulations for RL agent integration.
• To advance the understanding of turbulent physical systems using advanced data-driven control methods.

Main Methods:

• Development of a parallelized, high-fidelity fluid simulation environment compatible with RL agents.
• Configurable state observation (velocity, pressure) and control actuation (blowing/suction) at the wall.
• Comparison of a Deep Deterministic Policy Gradient (DRL) algorithm against classical opposition control.

Main Results:

• DRL achieved 43% and 30% drag reduction in minimal and larger channel simulations, respectively (friction Reynolds number of 180).
• DRL outperformed classical opposition control by approximately 20 and 10 percentage points in the respective channel sizes.
• The RL environment successfully facilitated the testing and comparison of advanced turbulence control strategies.

Conclusions:

• DRL represents a powerful, data-driven approach for designing effective turbulence control strategies.
• The developed RL environment serves as a valuable benchmark for future research in flow control.
• This study demonstrates the potential of AI to significantly enhance drag reduction in turbulent flows.