Chebyshev Fourier Neural Operators (CFNO)

📜 Extension of the Fourier Neural Operators to PDE problems solved using pseudo-spectral (Chebyshev) space discretization methods.

Content

• cfno : base Python module for CFNO
• dedalus : scripts to run RBC simulations with Dedalus and pySDC
• docs : some documentations about the FNO
• script : scripts for the full training pipeline (data generation, training, evaluation)
• utils : utility scripts for cluster run

Installation

In this folder, run this command to install cfno in your environment :

pip install -e .

The -e option installs in editable mode, which means any modification in the code won't need a re-installation to take the change into account.

You can also use the --user option with pip to install without admin rights.

Additional dependencies :

• dedalus : spectral discretization for RBC. Recommended installation approach: build from source.

• pySDC : base package for SDC, need to be installed using a development version available in the neuralpint branch of its main Github repo. To do that :

# Somewhere in a root folder ...
git clone https://github.com/Parallel-in-Time/pySDC.git
cd pySDC
git switch neuralpint
pip install -e .

Some changes may happen regularly on the development branch, to update your own version simply do

# In the pySDC repo
git pull

How to use the code

See the full pipeline description in scripts. In particular, the main code parts it uses are :

• cfno.models.cfno2d : implementation of the CFNO model
• cfno.losses : module for the different losses
• cfno.training.pySDC : base FourierNeuralOp class used for training and / or inference
• cfno.data.preprocessing : HDF5Dataset class and createDataset function used to create training datasets

Acknowledgements

This project has received funding from the European High-Performance Computing Joint Undertaking (JU) under grant agreement No 101118139 (Inno4Scale - NeuralPint). The JU receives support from the European Union's Horizon 2020 research and innovation programme.