Interpretable Multiscale Machine Learning‐Based Parameterizations of Convection for ICON.

Machine learning (ML)‐based parameterizations have been developed for Earth System Models (ESMs) with the goal to better represent subgrid‐scale processes or to accelerate computations. ML‐based parameterizations within hybrid ESMs have successfully learned subgrid‐scale processes from short high‐resolution simulations. However, most studies used a particular ML method to parameterize the subgrid tendencies or fluxes originating from the compound effect of various small‐scale processes (e.g., radiation, convection, gravity waves) in mostly idealized settings or from superparameterizations. Here, we use a filtering technique to explicitly separate convection from these processes in simulations with the Icosahedral Non‐hydrostatic modeling framework (ICON) in a realistic setting and benchmark various ML algorithms against each other offline. We discover that an unablated U‐Net, while showing the best offline performance, learns reverse causal relations between convective precipitation and subgrid fluxes. While we were able to connect the learned relations of the U‐Net to physical processes this was not possible for the non‐deep learning‐based Gradient Boosted Trees. The ML algorithms are then coupled online to the host ICON model. Our best online performing model, an ablated U‐Net excluding precipitating tracer species, indicates higher agreement for simulated precipitation extremes and mean with the high‐resolution simulation compared to the traditional scheme. However, a smoothing bias is introduced both in water vapor path and mean precipitation. Online, the ablated U‐Net significantly improves stability compared to the non‐ablated U‐Net and runs stable for the full simulation period of 180 days. Our results hint to the potential to significantly reduce systematic errors with hybrid ESMs. Plain Language Summary: Due to their computational costs, it is currently not feasible to run more accurate high‐resolution climate models on a global domain on climate (century) time‐scales. However, high‐accuracy climate simulations are needed for more robust and detailed projections of our future climate. Here, we develop and evaluate various machine learning‐based convection parameterizations learned on reconstructed and coarse‐grained high‐resolution subgrid fluxes to solve this problem, and benchmark their performance. The data set is chosen from simulations of the Icosahedral Non‐hydrostatic modeling framework (ICON) in a realistic setting of the tropical Atlantic and at storm‐resolving resolutions. We focus only on convective subgrid fluxes that are isolated from other components. We improve the best ML algorithms further by excluding variables that cause unphysical correlations. Finally, we explain the learned relations of the best data‐driven schemes based on physical process understanding, test their performance when coupled to the ICON model, and achieve stable coupled simulations for 180 days as well as improved precipitation predictions. Key Points: We train/benchmark machine learning models on convective fluxes derived from realistic coarse‐grained data of storm‐resolving simulationsShapley values reveal that the best offline model, a U‐Net, learns non‐causal links to precipitation and shows poor online performanceA model, without non‐causal precipitation connections, runs more stable coupled to ICON and indicates better precipitation predictions [ABSTRACT FROM AUTHOR]