Clustering of Right‐Moving Tornadic Supercell Proximity Sounding Profiles Using an Explainable Framework

Abstract

Abstract This study demonstrates an explainable machine learning framework for explicitly selecting important environmental features that determine clusters of right‐moving tornadic supercells. Although previous studies have clustered and analyzed proximity sounding profiles associated with tornadic right‐moving supercells, few have identified key features driving cluster assignments. Using self‐organizing maps, we group soundings based on virtual buoyancy and wind profiles. Recursive feature elimination reveals that virtual buoyancy at heights of approximately 3 and 10.6 km (near the equilibrium level) are the most significant features in this clustering, with both acting as a proxy for the potential magnitude of Convective Available Potential Energy (CAPE) and Convective Inhibition (CIN). Mid‐level wind vectors (around 5 km) and low‐level (below 3 km) meridional winds also play major roles in determining cluster assignments. Using these features at their identified key levels achieves skill comparable to traditional composite parameters in distinguishing between significant and non‐significant tornado environments associated with right‐moving supercells. We interpret these features as drivers of SOM node classification and as useful low‐dimensional descriptors of already tornadic environments, not as independent evidence that each feature controls tornado intensity.