In the Climate Informatics 2026, we will together explore how deep learning weather prediction (DLWP) models see clouds. Clouds are a critical component of the Earth's climate system, influencing energy balance, precipitation, and atmospheric dynamics. However, the cloud cover often remains unpredicted in DLWP models, which limits their evaluation and interpretability. In this hackathon, we will work together to develop and apply machine learning techniques to analyze the cloud representation in DLWP models, with the goal of improving our understanding of these models and ultimately enhancing their performance.
The goal is to learn an adapter that maps typical DLWP model output to cloud cover, and to use this adapter to analyze the cloud representation in the model. As most DLWP models are trained on the ERA5 reanalysis dataset, we will take a perfect model approach: we will train neural networks to map ERA5 reanalysis outputs to ERA5 cloud cover. This will allow us to analyze the raw cloud representation from DLWP models. Testing their generalisation and robustness, we will apply the trained adapter to hold-out ERA5 data, as well as model output from the AIMIP project, which includes a variety of DLWP model output on climate-like timescales. In the end, we hope to identify key problems and opportunities for improving the cloud representation in DLWP models, which can ultimately contribute to advancing our understanding of DLWP models for climate projections and to improve their weather prediction capabilities.

Key Problems We Aim to Address