Physics-guided machine-learning parameterizations of subgrid processes for climate modeling

Due to computational resource limits, small-scale processes, such as clouds and convection, are not explicitly included in state-of-the-art global climate models. Yet, these processes are crucial for accurate predictions of climate, and therefore the effect of these processes on the climate system needs to be estimated using parameterizations. Traditional parameterizations usually rely on simplified models and suffer from inaccuracies that lead to large uncertainties in climate projections.

Reducing these uncertainties is crucial since it will assist policymakers to plan proper adaptation and mitigation policies for climate change. Therefore, novel and computationally efficient approaches to subgrid parameterization development are urgently needed and are at the forefront of climate research. One alternative to traditional parameterizations is to use machine learning to learn new parameterizations which are data driven. However, machine-learning parameterizations might violate physical principles and often lead to instabilities when coupled to an atmospheric model. Furthermore, machine-learning parameterizations can make large generalization errors when evaluated in conditions they were not trained on. I will show how machine learning algorithms can be used to learn new parameterizations from the output of a three-dimensional high-resolution atmospheric model, while obeying physical constraints such as energy conservation. Implementing these parameterizations in the atmospheric model at coarse resolution leads to stable simulations that replicate the climate of the high-resolution simulation, and capture important statistics such as precipitation extremes. Moreover, I will discuss how we can rescale the inputs of machine learning algorithms to help them generalize to unseen climates. I will also discuss how machine-learning parameterizations can give further physical insights into the parameterization problem. Specifically, I will show how machine-learning algorithms combined with explainable artificial intelligence tools can be used to better understand the relationship between large-scale fields and subgrid processes.