Differentiating Between Tornadic and Nontornadic Synoptic Environments Using Self-Organizing Maps

The origins of tornado forecasting date back to March 25, 1948, after Fawbush and Miller successfully forecasted a tornado that damaged Tinker Air Force Base using synoptic charts of surface and upper-air data. This approach to forecasting severe weather was the standard for many decades, with NWP guidance gradually playing a more prominent role, and the emphasis also shifting (at least in the t–12 hour period) from synoptic analysis to mesoscale analysis of the severe storm threat. In the present day, forecasters routinely access predictions from ever-improving convection-allowing models. The goal of the present study is to take a fresh look at this subject using modern methods (e.g., machine learning) to determine whether any new synoptic-scale traits can aid in differentiating between tornadic and nontornadic storms. To perform this analysis, our study employs Self-Organizing Maps (SOMs), a type of machine learning method, to organize input data into different patterns. We train the SOM using 500 mb height anomalies for 2000 x 2000 km domains of ERA5 data centered on SPC proximity soundings for 8460 supercell thunderstorms at the time of a reported storm type (nontornadic, weakly tornadic, or strongly tornadic). This SOM resulted in the identification of 12 characteristic 500 mb patterns (nodes), each of which has a varying number of reported storm types. Composites of soundings and other tropospheric dynamic and thermodynamic fields were then created for each of the 12 nodes to identify features that set apart nodes with high percentages of strongly tornadic supercells from nodes with more weakly tornadic or nontornadic supercells. This analysis reveals that patterns associated with a higher percentage of strongly tornadic supercells in a node were characterized by a higher-amplitude 500 mb trough, stronger 500 mb jet streak, stronger quasi-geostrophic forcing, lower LCL heights, and stronger 0–1 and 0–6 km shear. These findings are consistent with our current understanding of the environments favoring tornadoes; however, the use of SOMs highlights the differences between tornadic and nontornadic storms in a way that is more readily usable in forecast scenarios than in previous studies.