Learning to Predict the Unpredictable: Deep Models for Severe Convective Weather

Abstract

Breakthroughs in deep learning have ushered in transformative opportunities for interdisciplinary research in the emerging field of "AI + Meteorology." Among the most challenging and societally impactful problems in this domain is the prediction of severe convective weather, which is characterized by highly dynamic and complex atmospheric processes. This paper provides a comprehensive overview of recent theoretical advancements and methodological innovations in applying deep neural networks to severe convective weather forecasting. It systematically reviews the limitations of traditional numerical and statistical methods, discusses representative datasets and evaluation metrics, and emphasizes the integration of physical and data-driven modeling principles. The application and performance of various deep learning models—including recurrent and non-recurrent architectures, generative approaches, and large-scale meteorological models—are thoroughly analyzed. In addition, the paper highlights critical challenges such as long-tailed data distributions, model interpretability, and lack of physical consistency. Finally, it outlines prospective research directions and open questions, aiming to offer both theoretical insights and practical guidance for developing next-generation intelligent weather prediction systems.