Prediction of hail size from environmental point sounding data is of high interest to forecasters, particularly when issuing forecasts for likelihood of severe hail on a Day 1-3 time scale. Physics-based approaches that use a combination of updraft parameterization and hail modeling are frequently used. In one such approach, a limited number of environmental parameters inform the updraft parameterization, which provides updraft characteristics to the improved HAILCAST hail model. Recent research has found this approach is sensitive to how updraft duration is determined, presenting a potential avenue of improvement.

Machine learning (ML) has been shown to be a powerful tool in synthesizing numerical model data, particularly environmental data, to improve forecasts of high impact weather events including severe hail. This study applies a number of machine learning techniques to information obtained from environmental point soundings to predict hail size as determined by the Maximum Estimated Size of Hail (MESH) product. Results from two main experiments are emphasized: prediction of hail size from sounding parameters alone; and prediction of updraft duration to use as input to the HAILCAST hail model, which then produces a hail size forecast. Results from these experiments are compared with those obtained from operational and updated HAILCAST versions alone, as well as results from the operational Sounding Analog Retrieval System (SARS).