National Aeronautics and Space Administration

Abstract:
Cloud microphysics parameterizations, simplified representations of cloud particle populations and their evolutions, are a crucial but highly uncertain part of many climate models. Because of the multi-scale nature of clouds, a lack of governing microphysical equations, process nonlinearity and stochasticity, and limited direct observations, microphysics parameterization design and quantitative evaluation have historically been challenging. Here, I present work to identify, characterize, and reduce parameterization errors that will enable building better cloud microphysics parameterizations. Specifically I consider (a) parametric error driven by poorly constrained and unphysical parameters and (b) structural error driven by inadequate representation of cloud particle properties and processes via parameterization variables and functional forms. First, I show machine learning can make Bayesian parameter inference computationally tractable for computationally expensive 3D climate models. This methodology enables characterizing parametric error and distinguishing it from structural error. By varying the formulation of the parameter inference, we can further pinpoint origins of structural error, which allows for a clearer path toward parameterization improvement. The talk focuses on applications to a microphysics parameterization in a cloud resolving model, but these approaches are of interest for climate model parameterization development more generally. Second, I pair idealized modeling with in situ drop size distributions to interrogate ubiquitous structural assumptions governing warm rain initiation in CMIP6 models. The results suggest a need to reformulate the structure of parametrized warm rain initiation congruent with the Bayesian parameter inference approach.