This repository contains a parallel fluid solver written from scratch using navier stokes equations. The solver is parallelized using MPI and supports FSI simulations through preCICE. CalculiX can be configured using a simple yaml, the same applies for openFOAM which can be used to test and validate the solver. This Repository also contains a simple yet configurable CNN to predict the flow of training data generated from the parallel solver.
In addition to a compiler and cmake you need libvtk to output simulation data which can be viewed with paraview.
To build the solver run the following commands:
mkdir build
cd build
cmake ..
make installThen the program can be run with mpirun ./numsim_parallel {path to scenario file}
The output will be in the out folder next to the executable.
It contains vti files for each timestep which can be opened with paraview.
To run FSI simulations, you need to have preCICE installed.
Then you can build the adapter which is located in the adapter subfolder.
You can couple against any structure solver which accepts forces and provides displacements through preCICE.
There is an example setup for CalculiX in the calculix subfolder.
This setup can also be configured through a yaml file.
To validate the solver, the configurations of the solver can be run though openFOAM with the Python entrypoint located in the openfoam subfolder.
Below is an example FSI simulatino for a 2d elastic tube:
More detail about the FSI setup can be found in the final presentation slides.
Example prediction:
For the CNN you need to have pytorch and numpy installed.
To generate training data, the solver can be run with a scenario file that specifies the number of samples and ranges to interpolate data. This could look something like this:
generateTrainingData = 1
nSamples = 101
reMin = 500.0
reMax = 1500.0
uMin = 0.5
uMax = 1.5
This will generate 101 samples with Reynolds numbers between 500 and 1500 and inflow velocities between 0.5 and 1.5.
The generated data will be stored in the train folder next to the executable.
The model is specified by a config.json file which contains parameters for the model architecture and training process.
To invoke training, the command python fluid_ml/main.py --config {path to config file} can be used.
There are a few default config files in cfg/ml.
After training, the model will be saved in the models folder and a pdf with summary plots will be saved alongside.
The pdf is automatically generated from a jupyter notebook. For this you need to have a LaTeX installation.
The whole Data Generation and Training process can be run with run_pipeline [-g generation_config] [-t training_config].
For the C++ code run_tidy.sh and run_format.sh can be used to run clang-tidy and clang-format respectively.
For the python code flake8 and pylint are used for linting. Linting can be invoked with make.
For linting, you need to have flake8, wemake-python-styleguide and pylint installed.
This can be done in a virtual environment with pip.