Parameterization of Stochastically Entraining Convection Using Machine Learning Technique

Abstract

A stochastic mixing model with a machine learning technique is proposed for mass flux convection schemes. The model consists of the stochastic differential equations (SDEs) for the fractional entrainment rate, fractional detrainment rate, fractional dilution rate, and vertical acceleration. Unknowns in SDEs are parameterized using a deep neural network with the inputs of cloud and environment properties. The deep neural network is found to predict entrainment and detrainment rates better than previously proposed parameterizations. The new mixing model is implemented in a unified convection scheme (UNICON) and tested in a single-column mode for two marine shallow convection cases. It is shown that the simulations with the new mixing model produce realistic mean and variance of various convective updraft properties and that the appropriate amount of stochasticity is generated. Consistently accurate simulations of updraft mass fluxes and moist conserved variables reduce model errors in the original UNICON. Additional sensitivity simulations enabling or disabling the stochasticity in mixing and initialization suggest that most of the cloud variabilities are generated from the mixing process.