Deep reinforcement learning (DRL) has recently emerged as a novel approach to discover efficient control strategies for active flow control (AFC).
In a nutshell, AFC has the potential to minimise the drag of a vehicle by modifying the flow around it, ultimately reducing its fuel consumption and thus its carbon footprint – a major goal in the transport industry.
In this project, the team proposes to use DRL to control boundary-layer separation in a case of interest for the aeronautical industry, namely the high-lift 30P30N wing at chord-based Reynolds number 400,000. Separation control can also yield to increased aerodynamic efficiency, a major goal in aeronautics.
While most of the DRL-based AFC studies in the literature still rely on training the DRL agent in a simplified (typically two dimensional) system, the study's approach leverages the use of high-performance computing (HPC) to train the model on a realistic high-fidelity environment.
This is performed through our state-of-the-art computational-fluid-dynamics (CFD)–DRL framework, which performs distributed training using multiple GPU-based parallel CFD simulations.
With this, the team will demonstrate the potential of discovering DRL-based AFC strategies for an industrially relevant turbulent flow, with critical implications for the aerospace, maritime, and automotive sectors.
Bernat Font, Technische Universiteit Delft - Netherlands