Ice storms are relatively rare yet high-impact winter phenomena, which can cause significant disruptions to the transportation and energy sectors, with wide-ranging socioeconomic impacts. Evaluating the ability of climate models to adequately simulate frozen extremes is essential for improving model performance and increasing confidence in future projections used by stakeholders and policymakers. However, the vast majority of models do not output precipitation types other than snow and rain, making assessment of icing impacts challenging.

Here, we leverage three-dimensional thermodynamic fields produced by global models to define four classes of precipitation (rain, snow, sleet, and freezing rain) by implementing an ensemble precipitation typing approach using the techniques typically applied to numerical weather prediction output. We validate this approach with ERA5 and station observations from the Automated Surface Observing System (ASOS) and NOAA’s Integrated Surface Database (ISD) across the eastern U.S. during 1980-2019. Whilst the mean annual climatology of the different precipitation types is reasonable in ERA5, there is varied accuracy at the station and event level. Self-organizing maps (SOMs) are then applied as an automated machine-learning approach to characterize the large-scale meteorological patterns (LSMPs) associated with freezing rain events over the eastern U.S. Trained on ERA5 dynamic and thermodynamic fields, we define atmospheric setups critical for a model to capture for it to accurately simulate frozen extremes. We then apply the reanalysis-derived SOM as a reference to evaluate the skill of CMIP6 historical experiments in simulating the synoptic environments that can support freezing rain events over the eastern U.S.