Application of Machine Learning for Dryline Climatologies and Extended-Range Severe Weather Forecasts

Abstract

The dryline is an important boundary for convection initiation in the central US. Althoughit has been studied for decades, there are no long-term climatological studies of both the dryline and its role in the severe thunderstorm climatology. In this dissertation, a machine learning (ML) model designed to identify fronts (FrontFinder AI) is applied to ERA5 reanalysis data to create an extended dryline climatology. The climatology for 16 years overlapping available NOAA Weather Prediction Center (WPC) surface analyses is compared for drylines across the entire contiguous US and for each month, expanding upon previous dryline climatologies which only considered a subset of months for specific regions. The ML dryline climatology corrects a non-meteorological change in annual dryline frequency observed in the WPC surface analyses. The ML model is then applied to ERA5 data back to 1940 to create a 26- year (1995–2020) climatology of dryline-associated severe thunderstorm reports in the Southern Great Plains, and to investigate dryline variability associated with drought and the El Ni˜no–Southern Oscillation (ENSO) over 81 years (1940–2020). Drylines are associated with over 50% of springtime severe days from west Texas through southwest Oklahoma, and drylines are more frequent and further east during months of Southern Great Plains drought. The final chapter creates a random forest (RF) daily severe thunderstorm forecast model to evaluate whether a nested convection-allowing model (CAM) ensemble and its associated simulated storm attributes provide any additional skill over a convection-parameterized ensemble for extended-range forecasts (days 3– 15). When both ensembles have daily synoptic-scale forecast skill (days 1–8), the CAM RF forecasts are significantly better. Preliminary results exploring forecasted dryline skill through day 15 are presented.