Source code for "Deep Dynamical Modeling and Control of Unsteady Fluid Flows" from NIPS 2018. The paper can be found here.
A description of the individual files is given below.
koopman_model.py- script for defining architecture of and constructing Deep Koopman models for training.train_koopman.py- training script for Deep Koopman models.bayes_filter.py- script for defining architecture of and constructing Deep Variational Bayes Filter models.train_bayes_filter.py- training script for Deep Variational Bayes Filter models.dataloader.py- script for loading and processing data prior to training.utils.py- contains functions for evaluating trained models.find_matrices.py- script to load a trained neural network model and determine the B-matrix, action normalization parameters, and goal state encoding.find_dynamics.py- script to load a trained neural network model and output the current state encoding and the A-matrix based on the previous sequence of observed states and actions.
config.ini- example config file that defines parameters for a PyFR simulation of the 2D cylinder system. Relevant parameters that may need to be modified can be found in thesoln-plugin-controllersection.controller.patch- patch that can be applied to PyFR to allow for prescribing an angular velocity on the surface of the cylinder and performing model predictive control.mesh.pyfrm- mesh file required to run simulation of 2D cylinder system.cyl-2d-p2-1530.pyfrs- solution file that can be used to initialize simulations of the 2D cylinder system.base.h5- snapshot of steady base flow that defines the goal state in model predictive control.loc_to_idx.json- JSON file containing a map from spatial locations in full CFD solutions to indices in the arrays used as neural network inputs.
Below we detail the steps required to install the necessary software, generate training data, train a Deep Koopman model, and perform model predictive control.
Deep Flow Control depends on both TensorFlow and PyFR which require python 3.5 to work properly. The training code depends on python 2.7 so it is easiest to manage different python environments using conda. Once conda is installed, you can create the two needed python environments by
conda create -n py27 python=2.7 anaconda
conda create -n py35 python=3.5 anaconda
When installing PyFR or tensorflow start by activating the python 3.5 environment
source activate py35
Deep Flow Control requires PyFR v1.7.6. Once downloaded, copy controller.patch into the top-level directory of PyFR and run git apply controller.patch to modify the PyFR code in order to enable simulation with control inputs. This patch will create a file named controller.py in the pyfr/plugins directory that contains the necessary code for defining control laws and performing model predictive control.
After applying the patch, PyFR can be installed by navigating to the PyFR directory and running
python setup.py build
python setup.py install
Tensorflow can be installed via pip install tensorflow or pip install tensorflow-gpu
Once PyFR has been successfully installed, simulations can be run in order to generate training data. Move the files config.ini, mesh.pyfrm, cyl-2d-p2-1530.pyfrs, and loc_to_idx.json to the same directory. Modify config.ini and controller.py as desired. Direct the python path to include the training files by adding the following to your .bashrc or .bash_profile file
export PYTHONPATH=$PYTHONPATH:/path/to/deep_flow_control/training
To generate the training data run the command
pyfr restart -bcuda -p mesh.pyfrm cyl-2d-p2-1530.pyfrs config.ini
to begin a simulation using the CUDA backend (OpenCL and OpenMP are also available as backends). Training data will be saved to the directory save_dir, as defined in config.ini but make sure the directory already exists.
The scripts in the training directory can be used to train a Deep Koopman model. Navigate to the directory containing the scripts
cd path/to/deep_flow_control/training
Before executing the scripts, make sure to switch to python 2.7 with
source activate py27
Examine train_koopman.py to get a sense for the arguments that can be used to define the model architecture. An example command to train a Deep Koopman model is:
python train_koopman.py --num_filters 256 128 64 32 16 --control_input True --data_dir (data_dir)
where (data_dir) is the directory where training data has been stored. By default checkpoints will be written to a directory named ./checkpoints.
Once you have a trained model, modify the arguments in config.ini to perform model predictive control. In particular, you will need to set mpc = 1, change training_path to be the path to the directory where the training scripts are located, modify checkpoint to correspond to the desired model checkpoint, and change base_flow to contain the correct path to the file base.h5. From this point, simulations can be run with the same command used to generate the training data. Make sure to activate python 2.7 before running PyFR with model predictive control.