Cross-type generalization of extreme-event learning

Determine the extent to which training AI weather/climate models on one class of extreme weather event (for example, heat waves) enables generalization to different classes of extreme events (for example, cold snaps or atmospheric rivers) that may share some physical processes but differ in dynamics.

Background

While the paper focuses on tropical cyclones, the authors point out that extreme events have differing dynamics, which may hinder cross-type generalization, but also share common physical processes that could facilitate transfer. They explicitly state that the degree of such transfer remains unresolved.

Establishing the scope of cross-type generalization would inform training set design and evaluation protocols for AI models tasked with multi-hazard forecasting and emulation.