AeroGP: Machine Learning How Aerosols Impact Regional Climate

Abstract Aerosol particles from both natural and anthropogenic sources play a critical role in the Earth's climate by interacting with solar radiation and clouds. Anthropogenic aerosol and precursor emissions have historically exerted a global cooling effect, which has partially offset the warming from concurrent greenhouse gas emissions. Recent reductions and shifts in aerosol and precursor emission patterns may reduce this offset and introduce spatially and temporarily varying climate impacts. Investigating aerosol‐climate effects is typically done with computationally expensive Earth System Models, which include complex representations of physical, chemical, biological, and geological processes and their coupled interactions for the entire global climate system. In this study, we develop a machine‐learning climate emulator using Gaussian processes, called AeroGP , that can be used to quickly assess, for example, the impact of different policy decisions on future climate mitigation strategies. The emulator is trained on a unique data set from the Norwegian Earth System Model (NorESM), analyzed as an ensemble here for the first time. AeroGP accounts for the joint spatial covariance of the output variables and captures the complex, heterogeneous impacts of aerosols on surface temperature using coregionalization. We believe this is the first time this method has been used to account for the spatial correlation of such climate data. We show that AeroGP retains the spatial complexity of NorESM at a fraction of the computational cost and demonstrate its usefulness to assess the sensitivity of temperature to idealized future aerosol emission scenarios.