Aerosol-cloud interactions are mediated in global climate models with the help of aerosol activation parameterizations. These parameterizations translate how changes in the background, ambient aerosol simulated by the model influence the nucleation of cloud droplets - specifically the number concentration of how many are nucleated. Altering aerosol emissions in the model thus influences cloud droplet number concentrations, giving rise to an “aerosol indirect effect” on climate as the simulated clouds exhibit different radiative properties. Better understanding the magnitude of these potential effects is critical towards understanding and anticipating contemporary climate change.
Estimates of aerosol activation under different conditions are based on a mixture of both observations and detailed theoretical calculations. But neither alone is suitable for directly implementing inside a climate model. Thus, parameterizations have been developed over the past two decades as climate models have incorporated increasingly sophisticated representations of both aerosol and cloud microphysics. Prior work has shown that model-derived estimates of the aerosol indirect effect are critically sensitive to the subjective choice of which activation parameterization is run online with the model (Rothenberg et al, 2018). Therefore, objective and general parameterizations with limited assumptions are desirable in order to faithfully represent aerosol-cloud interactions in these models.
In this work, we extend prior work which leveraged techniques from uncertainty quantification to built robust emulators of the detailed cloud parcel models traditionally used as “ground truth” for evaluating activation parameterizations. We explore a bevy of machine learning and data driven modeling techniques, exploiting two large ensembles (50,000 members each) of parcel model simulations run on both (a) aerosol-meteorology samples culled from online CESM simulations, and (b) similar samples from a simplified and artificially constructed phase space of aerosol and meteorology parameters. We further highlight how auto-machine learning tools and pipelines developed on Google Cloud Platform can aid researchers in quickly and efficiently developing new, machine-learning based parameterizations of physical processes for inclusion in global climate models.
Dr. Rothenberg is the Chief Scientist at ClimaCell, a weather technology company which leverages novel atmospheric measurements to better understand the weather as it impacts people and businesses. Previously, he completed his doctoral and post-doctoral studies at MIT where his dissertation focused on better understanding how aerosols influence climate. In this work he leveraged data-driven modeling and machine learning to build process emulators which could be embedded in climate models to more faithfully represent the physics underlying aerosol-cloud interactions.
