CausalNNCAM - Causally-informed deep learning to improve climate models and projections

Authors:

• Fernando Iglesias-Suarez - fernando.iglesias-suarez@dlr.de
• Breixo Soliño Fernández - breixo.solinofernandez@dlr.de
• Original CBRAIN-CAM code by Stephan Rasp - raspstephan@gmail.com - https://raspstephan.github.io

This repository provides the source code used on the paper Causally-informed deep learning to improve climate models and projections, by Iglesias-Suarez et al.

Corresponding DOI:

Installation

To install the dependencies, it is recomended to use Anaconda or Mamba. An environment file is provided in dependencies.yml.

How to reproduce results

The results described in the paper where obtained by execute the following steps:

1. pipeline.ipynb: Run PC1 within the PCMCI framework. This yields a set of causal drivers for each output at every grid column of SPCAM.
2. aggregate_results.ipynb: Collect, aggregate and evaluate causal links. Produces causal (correlation) matrix plots.
3. Find the appropriate threshold to filter spurious links, using SHERPA.
   1. SHERPA_threshold_GridSearch.ipynb: Best general threshold.
   2. notebooks_SHERPA_thrs_optimization_per_output/Create_optimized_numparents_dict_mse.ipynb: Best threshold for each output.
4. Creation and training of neural networks (NN).
   1. NN_Creation.ipynb: Can create both Causally-informed NN that use the best general threshold (from 2.1) and Non-causal NN that use all inputs.
   2. NN_Creation_optimized_threshold.ipynb: Causally-informed NN that use the best threshold for each output (from 2.2).
   3. NN_Creation_random_links.ipynb: NN using random links
5. Evaluation
   1. notebooks_evaluate_CausalNNs_r2/evaluate_nonlinearities_in_SPCAM.ipynb: Comparation between the different types of NN
   2. notebooks_online_evaluation: Comparation with SPCAM
      1. cross_section_online_evaluation.ipynb
      2. latitudinal_2Dfields_online_evaluation.ipynb
6. notebooks_xai/shap_xai.ipynb: Use explainable AI to evaluate the importance of the inputs in each NN