Discovering Active Flow Control strategies in a 2D cylinder with Deep Reinforcement Learning
In recent years, machine learning methods have been introduced to simple fluid dynamics problems that envision a promising future in this investigation field. This MsC thesis presents the coupling between Deep Reinforcement Learning (DRL) scheme using Artificial Neural Networks (ANN) and a CFD solve...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis de maestría |
| Fecha de publicación: | 2022 |
| 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:2117/371307 |
| Acceso en línea: | https://hdl.handle.net/2117/371307 |
| Access Level: | acceso abierto |
| Palabra clave: | Deep reinforcement learning Computational fluid dynamics Active flow control Machine learning Proximal policy pptimization Deep neural network Synthetic jets Supercomputer |
| Sumario: | In recent years, machine learning methods have been introduced to simple fluid dynamics problems that envision a promising future in this investigation field. This MsC thesis presents the coupling between Deep Reinforcement Learning (DRL) scheme using Artificial Neural Networks (ANN) and a CFD solver, Alya. This is done in the framework of an active flow control (AFC) case, everything running for the first time in a supercomputer. The studied case is a 2D cylinder at Re = 100. The flow is controlled by two opposed synthetic jets in the top and bottom surface. The DRL agent is able to find a strategy which improves the drag by 9%, maintaining zero lift condition. This supposed an improvement with the current literature on this case. The learnt strategy increases the recirculation bubble area and removes the K´arm´an streets, tending to recover the symmetrical flow before the oscillations appeared. This validates the coupling of the DRL agent with the CFD solver. In addition, there has been an exploratory work on the main setup parameters. Seeing how high Reynolds are more chaotic flows and less predictable to control, needing to adjust much most of the DRL setup parameters and mesh refinement. Or narrow window for the optimal action duration, DRL could lose information or be overloaded. Also there are some exercises trying different rewards or the maximum intensity of the jets. The promising results of this work show the great potential that ML applied to CFD has. Work is already ongoing for two publications, one of a 2D cylinder at Re=1000 and another changing the geometry to an airfoil that is going to be presented a the ECCOMAS congress 2022 conference in Oslo, Norway. This cases will benefit benefit from the available resources of a HPC system such as Marenostrum IV, being able to run large simulations or making use of the multienvironment features of DRL |
|---|