Cylinder2DFlowControlGeneral

This repository contains the code corresponding to our manuscript, "Robust active flow control over a range of Reynolds numbers using an artificial neural network trained through deep reinforcement learning", Tang et al., Physics of Fluids 32(5), 053605 (2020). This paper is availbale at https://aip.scitation.org/doi/10.1063/5.0006492, the preprint is also available at https://arxiv.org/abs/2004.12417

This work focuses on the active flow control of a computational fluid dynamics simulation over a range of Reynolds numbers using deep reinforcement learning (DRL). The DRL framework utilized in the present work for performing AFC is shown as follow:

The environment, i.e., a numerical simulation of the flow past a cylinder, is coupled in a closed-loop fashion with the learning agent. Iteratively, the mass flow rate of the jets is controlled by the agent according to the observed flow state. In response, the simulation produces the updated flow field as next state, and a reward signal is used to guide the control strategy towards controlling the flow so as to reduce the drag.

The learning environment supports four flow configurations with Reynolds numbers 100, 200, 300 and 400. It is shown that the DRL controller is able to significantly reduce the lift and drag fluctuations and to actively reduce the drag by approximately 5.7%, 21.6%, 32.7%, and 38.7%, at Re=100, 200, 300, and 400 respectively. More importantly, it can also effectively reduce drag for any previously unseen value of the Reynolds number between 60 and 400, as indicated by the following results:

Getting started

The main code is located in RobustDRLCylinder2DControl. You can use the code to obtain a model that is capable to take effective control over a range of Reynolds numbers. The model we obatined for the results presented in our paper is located in BestModel.

The recommended method of execution is with the docker container provided here.

For details about how to launch the simulation, please see this video:

This work is based on the multi-environment approach proposed by Rabault and Kuhnle, and the reader can also refer to the open source code.

If you use this work for your research, please consider citing our work:

@article{doi:10.1063/5.0006492,
author = {Tang,Hongwei  and Rabault,Jean  and Kuhnle,Alexander  and Wang,Yan  and Wang,Tongguang },
title = {Robust active flow control over a range of Reynolds numbers using an artificial neural network trained through deep reinforcement learning},
journal = {Physics of Fluids},
volume = {32},
number = {5},
pages = {053605},
year = {2020},
doi = {10.1063/5.0006492},
URL = {https://doi.org/10.1063/5.0006492},
eprint = {https://doi.org/10.1063/5.0006492}
}

Help

If you encounter problems, please:

• look for help in the readme file of this repo as well as the tutorial video.
• look for help on the repo for serial training and parallel taining.
• feel free to open an issue and ask for help.

Notes

• You can use script_launch_parallel.sh to launch the simulation or run launch_servers.py and launch_parallel_training.py respectively.
• For evaluating the obtained model/policy, run single_runner.py.
• You can refer to this tutorial for the details about running fenics in docker.

Typo

• There is a i missing in Eq. (20) in our paper. And the n in the words following Eq. (20) should be i.

  i.e., Eq. (20) shuld be c_i=a_{j−1}+ i *(a_j−a_{j−1}) / Ne where i=1, 2, ..., Ne. (See issues3. Thank @jviquerat for pointing this out.)

  This is purely a typo and the code is correctly implemented.

Contact and Support

[email protected] (Hongwei Tang)