Machine Learning Enhancement for Climate Change Projection of Hazardous Convective Weather

Hazardous convective weather (HCW; i.e. tornadoes, large hail, and severe wind) is a significant, damaging phenomenon of extreme weather. However, the degree to which HCW distributions will change spatially and temporally during early-season months over the course of the 21st century is unknown, in part associated with the high computational cost of performing global convection-permitting climate model simulations. We used the Weather Research and Forecasting (WRF) Model to conduct convection-permitting (4 km) simulations in order to understand changes in early-season (i.e. January - April) HCW over the continental United States due to anthropogenic climate change. Simulations were performed for two ten-year climate epochs: a historical period (1991-2000) and a future period forced by Representative Concentration Pathway 8.5 (2091-2100) to analyze climate change effects on HCW. Changes in HCW were quantified in two ways: through changes in HCW environmental favorability, and through proxies for HCW occurrence (namely, updraft velocity > 18 m/s). HCW environmental favorability, defined by large-scale environmental variables, is found to increase substantially across all months simulated between the historical and future climates, with the greatest absolute increases in April. The increase in environmental favorability is primarily driven by a changing thermodynamic environment, with substantial projected increases in CAPE and CIN (102% increase in CAPE > 500 J/kg; 173% increase in CIN > 100 J/kg), while kinematic profiles (SRH, 0-6 km wind shear) were not projected to change substantially between climate states. Additionally, HCW proxies are also projected to increase between the historical and future climates, both in terms of absolute probability of occurrence and in terms of probability conditional on the presence of a favorable environment. Finally, we utilize a random forest model to test the ability of machine learning to enhance projections from coarser resolution (27 km) simulations of the same climate epochs. If the machine learning model can successfully predict storm-scale variables from coarser resolution data, future computational costs could be reduced by performing coarse resolution global climate simulations and applying machine learning as post-processing to infer information at finer spatial scales.