A New Framework for Systematically Characterizing and Improving Extreme Weather Phenomena in Climate Models

Extreme weather phenomena remain a significant challenge for climate models due in part to the relatively small space and time scales at which such events occur. Accordingly, robust simulation of extreme events requires models with high fidelity at these relatively small scales. However, numerous recent studies have shown evidence that current climate models exhibit non-convergent changes in extreme weather statistics as spatial and temporal resolution increase. These studies also provide evidence that such non-convergence originates in the subgrid parameterization suites (e.g., micro/macrophysics and convection). In order to provide a framework for identifying parameterization characteristics that cause non-convergent behavior and for testing parameterization improvements, we have developed a hindcast-based system characterizing the fidelity of extremes as a function of spatial and temporal resolution. The use of hindcasts as a model evaluation tool allows us to identify modes of failure (e.g., false-hits and misses) that systematically vary as a function of resolution. We have implemented this framework for the Community Earth System Model, and we have created a dataset of hindcast ensembles at multiple horizontal resolutions. Preliminary analysis of this multi-resolution set of hindcasts shows that in some regions, (1) the tail of the precipitation probability density (PDF) function grows as resolution increases (in accord with recent studies), and that (2) a large portion of this increase in the PDF tail comes from increases in Type I model errors—simulated extreme events that do not occur in observations. We explore possible causes of this inconsistent model behavior.