Probabilistic Operator Learning for Climate Model Parameterisation

This repository contains the code required to reproduce the results presented in the MRes Thesis "Probabilistic Operator Learning for Climate Model Parameterisation", submitted to The University of Cambridge.

Setup

We recommend first cloning this repository locally. Navigate to your chosen directory and use

git clone https://github.com/tomcjackc/Probabilistic-Operator-Learning-for-Climate-Model-Parameterisation.git
cd Probabilistic-Operator-Learning-for-Climate-Model-Parameterisation

Edit the bottom line of environment.yml according to the local filepath to your Conda environments. The conda environment used to conduct this work can then be created by using

conda env create -f environment.yml

Notebooks in the notebooks/ diectory should now be runnable. If the import statements raise exceptions, please contact tc656@cam.ac.uk and we'll be happy to help.

Data download

Data for the benchmark problems --- Burgers', Darcy Flow, Helmholtz, Navier-Stokes --- can be obtained from https://doi.org/10.5281/zenodo.12529654. Data for each of these problems (details found on the Zenodo repository) should be placed into the burgers_data/, darcy_flow_data/rect_cont_PWC/, helmholtz_data/, and navier_stokes_data/ directories, respectively.

Code used for reproduction

The Jupyter notebooks found in the notebooks/ directory can be used to reproduce the results presented in the thesis. See notebooks/README.md for more details.

Model architectures for our operator learning framework are found in models.py, including a base class for our overall model and a gaussian process regressor class (written using GPJax) used for our latent representation of the operator in question.

Code used to load data for each benchmark problem and the PV parameterisation experiment is found in dataloaders.py. This ensures that data used in the notebooks is in a standardised format.

Lastly, tools.py includes some helper functions required to load data for the PV parameterisation experiment. This code is reused from https://github.com/m2lines/pyqg_generative.git.

The submodule pyqg_parameterisation_benchmarks is used to generate new data for PV parameterisation in experiments, but is not used in the final results of this work.