Performance of a Random Forest Parameterization in Predicting the Diurnal Cycle of Precipitation

The diurnal cycle of precipitation is a useful measure by which we may evaluate a convective parameterization's skill in capturing the unresolved physical processes. With this in mind, we construct training datasets from a combination of high resolution SHiELD and coarse resolution FV3GFS atmospheric simulation data and train a random forest parameterization to predict vertical profiles of air temperature and specific humidity tendencies in the FV3GFS model. Here, we present an evaluation of our machine-learned parameterization's performance in predicting the diurnal cycle of precipitation. Using the high resolution SHiELD simulation as a benchmark, we find that the diurnal cycle of precipitation over land predicted by the random forest parameterization is closer in phase and amplitude to the SHiELD benchmark than the FV3GFS physics parameterization. We will also present results of the ML parameterization's diurnal cycle in localized climate regions, such as the Sahel.