Parameterization of Autoconversion Rates in a Climate Model Using Machine Learning with Training from Large Eddy Simulations

Abstract:

Phenomenologically, cloud water autoconversion is the rate at which cloud droplets grow into raindrops, predominantly through collision-coalescence. As autoconversion helps control cloud lifetime, by controlling the rate at which cloud water is depleted through precipitation, parameterization of autoconversion plays an important role in determining a climate model’s climate sensitivity. The rate at which cloud droplets collide and coalesce depends on many factors that are not explicitly treated in climate models. For this reason, many atmospheric models rely on empirical relations to represent autoconversion that are obtained through curve fits to data from a relatively limited number of idealized simulations that explicitly resolve autoconversion processes.

Here we discuss our experience leveraging a new large eddy simulation capability, called Predicting INteractions of Aerosol and Clouds in Large Eddy Simulation (PINACLES), that is equipped with spectral bin microphysics to explicitly simulate autoconversion processes to generate training data for deep neural networks (DNN) that we then use to parameterize autoconversion processes in the Department of Energy’s Energy Exascale Earth System Model (E3SM). In particular, we will describe our end-to-end workflow from the generation of training datasets spanning a range of boundary layer cloud types, to the training of deep neural networks, and finally to the implementation in a climate model. Further, we will assess the impact of these new DNN-based parametrizations on aerosol-induced change of cloud and precipitation properties, the effective radiative forcing (ERF) associated with aerosol-cloud interactions (ERFaci), and the simulated climate.