AeroGP: machine learning how aerosols impact regional climate

Anthropogenic aerosol emissions have historically exerted a net cooling effect which has masked some of the simultaneous warming from greenhouse gases (roughly -0.5°C since pre-industrial times). This mean effect is the result of heterogenous climate forcing through aerosol-radiation and aerosol-cloud interactions both locally close to emission sources and remotely via teleconnections. Future reductions and shifts in aerosol emission patterns due to regional clean air policies and shifting industrial production could therefore unmask additional warming and induce spatially complex climate impacts. Therefore, there is a need for computationally efficient tools to assess the climate impacts of possible future aerosol policy decisions.

We have developed a machine-learning emulator using Gaussian Processes (GP), trained on output from the Norwegian Earth System Model (NorESM), to predict the global spatially resolved surface temperature response to regional aerosol emission perturbations. We use a novel design for our GP model which considers the joint spatial covariance of the outputs. We show the efficacy of the emulator is comparable to that of the parent model NorESM for a fraction of the computational cost, and then use it to assess potential future aerosol emission scenarios that might be relevant to European policy decisions.