Deep reinforcement learning for active flow control on a wing: A numerical study

This work presents a numerical study on the application of deep reinforcement learning (DRL) for active flow control (AFC) on a three-dimensional NACA0012 wing at a high angle of attack (AoA = 20◦ ) and low Reynolds number (Re = 1000). The aim of this work is to achieve improvements in the aerodynam...

Descripción completa

Detalles Bibliográficos
Autor: Molina Garcia, Pau
Tipo de recurso: tesis de maestría
Fecha de publicación:2025
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/455188
Acceso en línea:https://hdl.handle.net/2117/455188
Access Level:acceso embargado
Palabra clave:Deep learning (Machine learning)
Computational fluid dynamics
Aerodynamics
Deep Reinforcement Learning (DRL)
Active Flow Control (AFC)
CFD
Airfoil
NACA0012
Aprenentatge profund (Aprenentatge automàtic)
Dinàmica de fluids computacional
Aerodinàmica
Àrees temàtiques de la UPC::Aeronàutica i espai::Aerodinàmica
Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial::Aprenentatge automàtic
Descripción
Sumario:This work presents a numerical study on the application of deep reinforcement learning (DRL) for active flow control (AFC) on a three-dimensional NACA0012 wing at a high angle of attack (AoA = 20◦ ) and low Reynolds number (Re = 1000). The aim of this work is to achieve improvements in the aerodynamic performance through different actuation strategies. The DRL framework used is formed by TensorFlow, being the artificial neural network, and SOD2D, the CFD solver. This two entities are integrated using an in-memory database for efficient communication. The domain consists of three pseudo-environments, each with two jets on the top surface: one at the front and one at the middle. The rear jet operates with opposite sign to the front jet to ensure mass conservation. Several reward functions are explored, prioritizing drag, lift, or lift fluctuations performance, in order to evaluate the agent’s ability to optimize different flow control strategies depending on the objective. The baseline simulation exhibits similar results with literature, showing a three-dimensional flow with vortex shedding. The DRL-controlled cases demonstrated significant improvements over the baseline, achieving up to 22.85% drag reduction and 59.78% decrease in lift fluctuations. The most efficient case obtained a 60% lift increase with only a 4% drag penalty and a 44.5% reduction in fluctuations, resulting in a 150% overall efficiency improvement. Classic active flow control strategies are also performed, showing good results, specially when configuring them to mirror the optimal strategies found in the DRL cases. However, the DRL cases still outperform them. These findings underline DRL’s potential to control complex flows while discovering advanced AFC strategies beyond conventional periodic forcing methods.