Using Machine Learning Methods to Predict and Understand Severe Weather Over the United States

The ability to predict severe convective weather at longer lead times over the United States has remained a persistent challenge despite recent insights and technological advancements. However, increasing lead times beyond a few days can be vital for those such as emergency managers and stakeholders, as it enables them to prepare for emergencies. Previous research has demonstrated the potential for improved predictability of extended severe weather events through the application of machine learning techniques and the analysis of various tropical and extratropical teleconnections. In this study, we employ deep learning within a scientific discovery framework to enhance our understanding and expand the predictability of severe weather events across different regions of the United States. Practically perfect hindcasts, Climate Data Record tropical outgoing longwave radiation anomalies, and ERA-5 synoptic and mesoscale reanalysis data are used to assess the likelihood of severe weather hazards, including tornadoes, hail, and wind, over timescales ranging from one to three weeks. A fully convolutional neural network, designed with both encoder and decoder components, is trained using these data sets for predictive purposes. Additionally, key features and locations are identified within the input data that significantly influence the model's predictions using explainable artificial intelligence frameworks. The integration of deep learning techniques in a scientific discovery framework provides valuable insights into the dynamical processes and teleconnections that contribute to enhanced predictability of severe convective weather.