This project provides a clean and extensible reference implementation of Finite Element Networks as proposed in our paper

Learning the Dynamics of Physical Systems from Sparse Observations with Finite Element Networks
Marten Lienen & Stephan Günnemann
Published at ICLR 2022

A learned flow field

First, install pytorch and torch-scatter according to your CPU/GPU setup. Then you can clone the repository and install the package locally.

# Clone the repository
git clone https://github.com/martenlienen/finite-element-networks

# Change into the repository
cd finite-element-networks

# Make sure that a recent version of pip is available that supports PEP 518 projects
# with a pyproject.toml
pip install --upgrade pip

# Install the code as a local package
pip install --editable .

After this you can import finite_element_networks in your own code, python shell, and notebooks, and easily integrate it into your existing code.

If you want to run our example training script, explore the notebooks or use our pytorch lightning data modules, you have to install the package with the lightning extra.

pip install --editable '.[lightning]'

If you are a weights & biases user, there is some additional code to log plots and animations there during training.

To get you started quickly, we have published a preprocessed version of the Black Sea and ScalarFlow datasets as well as pre-trained checkpoints of both FEN and T-FEN. To get them, run the following commands in the root of the repository.

curl -o data.zip https://zenodo.org/record/6366269/files/fen.zip
unzip data.zip

Now you can train a new model with examples/train.py black-sea or examples/train.py scalar-flow. Note, that you can also train models or use the project in any other way without downloading the data and checkpoints. However, if you use the data modules without having downloaded the preprocessed datasets, the will download and prepare the data for you, which is almost 500G for ScalarFlow and 13G for Black Sea.

If you end up using these datasets in your own work, please note that the raw data of both datasets comes with their own license as we have described in the appendix of our paper.

Additionally, we provide some notebooks for you to recreate results and figures similar to what we present in the paper.

• forecasting_evaluation.ipynb loads one of the checkpoints and computes the MAE over the test set
• flow_fields_and_disentanglement.ipynb renders animated versions of the inferred flow field and the disentanglement between free-form and transport term

The simplest way for you to apply this model to your own data will be to create your own data module, which you can model directly after the ones that we provide.

If you build upon this work, please cite our paper as follows.

@inproceedings{lienen_fen2022,
  title = {Learning the Dynamics of Physical Systems from Sparse Observations with Finite Element Networks},
  author = {Lienen, Marten and G\"unnemann, Stephan},
  booktitle={International Conference on Learning Representations (ICLR)},
  year = {2022},
}