This repository includes codes for numerical cylinder flow control cases.
The folder rotaryRe100DRL contains a DRL agent based on the original OpenAI baselines implementation of the PPO algorithm.
The graphical abstract is available at https://youtu.be/9X8XtHk0R84
-
In order to run training before you have to download and build TFlows CFD solver here from the development_branch.
-
The solver Processor can be built with the follolwing command:
make MPI=yes FCOMP=mpif90.openmpi CGNS=no CGNS_MPI=openmpi DIR_CASE=<folder_with_the_agent>- Change the text inside the brackets with the folder name containing User_Mod folder coupling the Fortran code with the DLR agent. After successfull build create a symbolic link to the Processor binaries inside the DIR_CASE.
- Run python module to start the agent
cd OpenAI-baselines-CustomEnvs
python -m baselines.run-
Training log with be written to a train_data subfolder with correspondong current models of the controllers
-
Use tensorboard for monitoring the training progress
- Use the follwing code to test a trained model
python -m baselines.run --load_path=<absolute_path_to_the_trained_model>- Use pretrained model to see the control in action
python -m baselines.run --load_path=<absolute_path>/cylindercontrol/rotaryRe100DRL/OpenAI-baselines-CustomEnvs/trained_models/VirtualEnv/20200622-190846_dns_re100_entropy0.01_3-cd-0.2cl_50long/checkpoints/00050-
Original python conda environment used for the project can be restored using environment.yml file, see directions here
-
Most of the training parametres can be modified in /CustomEnvs/CustomEnvs.py and /CustomEnvs/VirtualEnv/VirtualEnv.py files.