Machine Learning Based Closure Optimization for the Unified Convection Parametrization

The EMC Unified Convection (UC) parameterization combines the Simplified Arakawa Schubert (SAS) and Grell-Freitas (GF) convective parameterizations in order to improve performance on regional and global scales. The UC parameterization uses the average of an ensemble of closures to determine the strength and location of convection. Deficiencies in optimizing the selection of these closures, used in deep convection parameterizations in General Circulation Models (GCMs), at different scales and in changing environmental conditions have critical impacts on climate simulations. Some closures may produce more accurate output in particular environmental conditions but currently the GF parametrization takes a uniform average over all closures. This work uses Machine learning (ML) methods combined with satellite and global model datasets in order to weight the closure average based on location and meteorological conditions. First dimensionality reduction techniques are applied in order to define a set of conditions where certain groups of closures tend to perform better. From these groups a weight vector is generated from the relative error each closure demonstrates compared with observations. A decision tree is then responsible for deciding which weight vectors are best in particular environmental situations. One advantage of this approach is that it is explainable; a human expert familiar with the behaviors of the closures (the conditions where they perform best/worst, etc.) can determine why the decision tree chose the particular weight vector.