Data-driven cloud cover parameterizations for the ICON earth system model using deep learning and symbolic regression

This thesis delves into the improvement of cloud parameterizations in climate models through machine learning trained on coarse-grained output from high-resolution simulations. Utilizing the ICOsahedral Non-hydrostatic (ICON) modeling framework, it specifically targets the enhancement of cloud cover parameterization within the ICON Earth System Model. Three types of neural networks (NNs) differing in vertical locality are developed to estimate cloud cover, with globally trained NNs even applicable to distinct regional simulations. Interpretability analysis exposes model-specific biases and local relationships with the thermodynamic environment. Despite achieving high predictive performance, NNs necessitate post-hoc interpretation tools. To tackle this issue, a combined hierarchical modeling framework incorporating symbolic regression, feature selection, and physical constraints is proposed. The resulting equations, characterized by simplicity and physical consistency, attain performance comparable to NNs while demonstrating superior transferability to other realistic datasets. Our best equation adeptly captures cloud cover distributions across various regimes, notably excelling in representing marine stratocumulus clouds by learning to utilize the vertical relative humidity gradient. This research underscores the potential of deep learning in achieving accurate cloud parameterizations and emphasizes the effective role of symbolic regression in deriving interpretable, consistent equations for cloud cover.