What question did this study set out to answer?

Investigate active control strategies for separated flows around airfoils using deep reinforcement learning.

March 26, 2026The Proceedings of Mechanical Engineering Congress Japan0 citationsOpen Access

Analysis of Separation Flow Control Strategies acquired by Deep Reinforcement Learning of Wide-area Flow Field Inputs around High Angle-of-attack Airfoils

NTNaoki Takada AWAyano Watanabe SSSatoshi Shimomura

Key Points

Investigate active control strategies for separated flows around airfoils using deep reinforcement learning.
Utilized a dielectric barrier discharge plasma actuator for flow control.
Trained a controller using an Attention Branch Network in a 3D CFD environment.
Employed Ape-X DQN for dynamic selection of actuation frequencies.
The trained controller achieved a higher lift-to-drag ratio than predefined strategies.
Attention maps highlighted key flow features influencing control decisions.
Findings provide insights into separation control mechanisms and sensor placement.

Abstract

In this study, we investigate active feedback control of separated flows around NACA0015 airfoil using a Dielectric Barrier Discharge Plasma Actuator (DBDPA) trained via deep reinforcement learning (DRL) in a 3D CFD environment. A deep neural network-based controller with an Attention Branch Network (ABN) structure was trained using Ape-X DQN, enabling dynamic selection of burst frequencies for DBDPA actuation. Compared to predetermined control strategies, the trained controller achieved a higher lift-to-drag ratio by actively adapting to unsteady separation dynamics. Attention maps generated by ABN revealed that the controller focused on key flow features such as separation zones, leading edge vortex shedding, and low-pressure regions near the leading edge, providing interpretable insights into its decision-making process. These findings contribute to the understanding of separation control mechanisms and offer design guidelines for effective sensor placement in practical feedback control systems.

AI से पूछें

Bookmark

View Full Paper