Real-Time Model-Based Reinforcement Learning for Active Flow Control on the NASA Hump Model

Active flow control (AFC) offers promising approaches for enhancing aerodynamic performance by manipulating flow structures. The NASA wall-mounted hump model, a benchmark for separated flow control, provides an ideal testbed for investigating AFC strategies. However, optimizing control parameters in complex flow regimes remains challenging due to the high-dimensional, nonlinear nature of fluid dynamics.

Reinforcement learning (RL) has emerged as a powerful tool for solving complex control problems, but its application to AFC has been limited by partial observation and prohibitively long training times in computational fluid dynamics simulations. Model-based reinforcement learning (MBRL) addresses this challenge by learning a dynamics model concurrently with policy optimization, potentially reducing the number of required environment interactions.

We present a novel interface that enables RL agents to directly actuate synthetic jets on a wind tunnel model of the NASA hump. Our setup integrates with the 20” x 28” Shear Wind Tunnel at NASA Langley Research Center, allowing for real-time control and data acquisition. The system comprises pressure sensors, electronically-controlled valves, and an extensible environment that allows for the training of RL agents.

This framework facilitates the application of state-of-the-art MBRL algorithms to AFC, potentially accelerating the discovery of optimal control policies for flow separation mitigation. Our approach paves the way for data-efficient, adaptive flow control strategies that can be developed and validated in physical experiments, bridging the gap between simulation-based and experimental AFC research.