A two-moment machine learning parameterization of the autoconversion process

Abstract

Autoconversion is the mass transfer from cloud to precipitation water in an early stage of cloud development, and is the dominant process in the formation of embryonic droplets that trigger precipitation formation. The accurate parameterization of this process is key, in order to improve the interaction between cloud microphysics and cloud dynamics for models from cloud scale to the global climate scale. For model based parameterizations of the auto-conversion process, the usual approach to develop an autoconversion parameterization is by curve fitting the autoconversion rates obtained from simulations or numerical solutions of the kinetic collection equation under a wide range of initial conditions. However, in this case, the autoconversion is modeled by a function that is a nonlinear product of liquid water content and droplet concentration and depends on a small number of parameters. As a result, a large amount of scatter around the actual values can be obtained, indicating a weak relationship between actual and fitted autoconversion rates.The purpose of this paper is to analyze whether neural networks are better than traditional curve fitting or regression to obtain parameterizations of autoconversion. Then, a deep neural network was trained from an autconversion rates dataset generated by solving the kinetic collection equation for a wide range of droplet concentrations and liquid water contents. The obtained machine learned parameterization shows a very good match with actual rates calculated from the kinetic collection equation.