Representing Subgrid-Scale Cloud Effects in a Radiation Parameterization using Machine Learning: MLe-radiation v1.0

This repository contains the code for the ML-enhanced radiation scheme which is based on RTE+RRTMGP used in the ICON Model. The corresponding paper is submitted to the Journal Geoscientific Model Development.

The corresponding paper is available on arxiv as preprint

Hafner, K., Shamekh, S., Bertoli, G., Lauer, A., Pincus, R., Savre, J., & Eyring, V., 2025, Representing Subgrid-Scale Cloud Effects in a Radiation Parameterization using Machine Learning: MLe-radiation v1.0 https://doi.org/10.48550/arXiv.2510.05963

If you want to use this repository, you can start by executing

conda env create -f environment_ml.yml
conda activate hafner1_ml_rad

for training and evaluation the ML-based part.

If you want to use pyrte-rrtmgp, activate the following environment

conda env create -f environment_pyrte.yml
conda activate hafner1_pyrte_rrtmgp

Repository content

• evaluation contains some functions for prediction and evaluation
• models contains the NN architecture including preprocessing layer
• nn_config contains the configuration of all NNs
• plotter contains plotting functions
• preprocessing contains the normalization file and data loader
• utils contains some helper functions
• train_jsc_cloudy.py contains the training script
• eval_jsc_cloudy.py contains the evaluation script
• pyrte_on_coarse_grained_data.py script to tun pyrte+rrtmgp on coarse grained data for reference
• config.py contains some general routines that are used for the training and evaluation script such as reading config files, loading data, creating an instance of the NN
• data_distribution.ipynb used to calculate and plot the distributions of variables in the simulations (Figure 2, 3, A1, B1)
• combined_eval_plots.ipynb used to calculate statisitcs and plot the results (Figure 4, 5, C1)

Previous Work

The code is partialy based on previous work on an ML-based radiaiton emulator which hase been published:

Hafner, K., Iglesias-Suarez, F., Shamekh, S., Gentine, P., Giorgetta, M. A., Pincus, R., & Eyring, V. (2025). Interpretable machine learning-based radiation emulation for ICON. Journal of Geophysical Research: Machine Learning and Computation, 2, e2024JH000501. https://doi.org/10.1029/2024JH000501