This has been tested with

• Ubuntu 18.04.2 LTS (GNU/Linux 4.15.0-55-generic x86_64)
• Python 3.6.8
• CUDA 10.2

Clone this repo using git clone --recurse-submodules which also pulls content from submodules. Initialize a virtual environment and install necessary dependencies with

virtualenv -p python3 env
. env/bin/activate
pip install -r requirements.txt

Compile the spconv package according to these instructions. No need to clone the repo itself since it is already included as a submodule. If you run into problems during the installation please refer to their issues section. We only use the spconv package as provided.

Install OpenFOAM 7 with sudo apt install openfoam7. We implemented a custom interFoam solver which dumps L and d from the discrete pressure Poisson equation Lp=d to disk. Change the directory cd foam/newInterFoam/ and compile it by running wmake. Check out https://openfoam.org/download/7-ubuntu/ if you have questions regarding the installation process of OpenFOAM 7 on Ubuntu.

In the preconditioner/ folder you can find PyTorch code for the machine learning part. Adjust the settings to your liking in config.py. Generate a data set of system matrices L representing the discretized Laplacian and start the train/test loop in the background with

python3 gen_data.py
nohup python3 train.py &

We use TensorBoard to log hyperparameters, train/validation loss, and the performance of the model on the test data set. Run

tensorboard --logdir runs/ &

if you want to monitor the loss during training and check the test results in your browser.