Evaluation of Hail Size Forecasting Models during the 2016 Hazardous Weather Testbed Spring Experiment

Convection-allowing models (CAMs) can produce coarse representations of individual thunderstorms but do not explicitly represent severe hazards, such as tornadoes and hail. The probability of these hazards occurring and their expected intensity must be diagnosed from the available CAM output. During the 2016 Hazardous Weather Testbed Spring Experiment, three hail size diagnostic methods were evaluated in real-time both subjectively by participants and by object-based and grid-based verification methods. The Thompson hail size method estimates the largest hail size possible at a given point from the hail or graupel size distribution predicted by the CAM microphysics scheme. WRF-HAILCAST runs a 1D hail growth model at each grid point to estimate the hail size distribution from an ensemble of hail embryos. The Gagne machine learning hail size method estimates the probability of hail occurring and a spatial radar-estimated hail size distribution using storm and environmental information. The three methods are all evaluated on the Center for Analysis and Prediction of Storms Mixed-Physics and Single-Physics Ensembles and the National Center for Atmospheric Research CAM Ensemble. Hail forecasts from each method are grouped by location, time, and severe weather parameter values, such as CAPE and wind shear, and are compared with radar-estimated hail swaths and National Weather Service hail reports. The presentation will highlight where each method performs well and struggles along with insights about CAM model configurations and their effects on hail forecasts.