Data-driven multiscale modeling of subgrid parameterizations in climate models

Climate models, which simulate the long-term evolution of the Earth's atmosphere, oceans, and terrestrial weather, are critical tools for projecting the impacts of climate change around the globe. Due to limits on available computational resources, these models must be run at a coarsened spatial resolution which cannot resolve all physical processes relevant to the climate system [4]. To reflect the contribution of these subgrid-scale processes, closure models are added to climate models to provide the needed subgrid-scale forcing. These parameterizations model the contribution of these fine-scale dynamics and are critical to high quality and accurate long term predictions [14, 5]. A variety of approaches to designing these parameterizations have been tested, ranging from hand-designed formulations [16], to modern machine learning with genetic algorithms [14], or neural networks trained on collected snapshots [17, 7, 12, 13], or in an online fashion through the target simulation [6].