A physics-informed machine learning parameterization for cloud microphysics in ICON

This repository contains the code for the physics-informed machine learning parameterization for cloud microphysics in ICON. The simulation data used to train and evaluate the machine learning algorithms was generated with the ICON model. The corresponding paper is currently under Review in Environmental Data Science

Sarauer, Ellen, et al. "A physics-informed machine learning parameterization for cloud microphysics in ICON."

If you want to use this repository, start by executing

conda env create -f environment.yml
conda activate sarauer_ml_mig

Repository content

• models: contains the trained ML models "Microphysics Trigger Classifier" and "Microphysics Regression".
• notebooks: contains notebooks for data exploration.
• scripts: contains batch scripts on how to submit scripts in src folger to DKRZ with slurm + coarse-graining script coarse-graining.sh.
• src: contains all important scripts for the pipeline.
  • preprocessing: preprocess_classifier.py and preprocess_regression.py
  • build models: build_classifier_model.py and build_regression_model.py
  • training: train_classifier.py and train_regression.py
  • postprocessing and explainability: postprocess_classifier_and_explain.py and postprocess_regression_and_explain.py

Data

To fully reproduce the results it is first necessary to have access to accounts on DKRZ/Levante. The source code is available on the GitLab of the DKRZ (https://gitlab.dkrz.de/icon/icon-mpim) under a BSD 3-clause license (https://gitlab.dkrz.de/icon/icon-mpim/-/tree/master/LICENSES). The simulations were performed with the branch feature-nextgems-aerosol-microphysics at commit 260364f1.

Important packages

• all models are trained with Pytorch (https://github.com/pytorch/pytorch) and using Sklearn (https://github.com/scikit-learn/scikit-learn)
• code related to the explainability of model predictions uses the Shap library (https://github.com/shap/shap)