The Future of Severe Thunderstorms in the United States—Insights from Combining Deep Learning and High-Resolution Modeling

Climate change is altering the large-scale thermodynamic and kinematic landscape over North America. Understanding the resultant influence of modified environments on severe thunderstorm hazards (e.g., tornadoes, hail) remains difficult. Past studies have used environmental proxies and dynamical downscaling to analyze trends because they bypass spatiotemporal biases associated with tornado and hail reporting practices. However, their ability to discern the interplay between thermodynamics and kinematics on severe storm hazards is limited. Here we addressed this problem using deep learning, a subset of machine learning that uses neural networks as a vehicle to approximate nonlinear relationships within input data, which can be multidimensional in nature. In this study, convolutional neural networks (CNNs) were trained to estimate the probability of tornado and hail occurrence from environments and simulated radar parameters of convective-permitting model simulations covering most of North America. These simulations were created using the Weather Research and Forecasting model (WRF) and were used to derive an estimate of severe thunderstorm frequency and intensity under current climate conditions. The trained CNN model was subsequently applied to a convection-permitting WRF simulation generated under a business-as-usual, end of the century climate scenario. A synthesis of regularization techniques, inductive biases, and hyper-parameter optimization results is provided. Challenges in deciphering CNN model layers and co-dependencies between atmospheric variables are also discussed. Our results focus on the intercomparison between severe thunderstorm frequencies and intensities during current and future climate conditions, which indicate how changing environments will impact convection over North America.