Severe convective weather (SCW) forecasting has emerged as a cutting-edge research focus due to its complex dynamical characteristics and significant socioeconomic impacts. However, the advent of deep learning (DL), with its capacity for extracting massive information and modeling non-linear relationships, offers a promising alternative for enhancing the accuracy of SCW forecasts. This study reviews advances in DL applications to SCW forecasting. It begins by summarizing current mainstream DL techniques and analyzes the similarities between SCW tasks and typical DL problems. Subsequently, it surveys the utilization in forecasting four types of SCW events: rainstorms, hail, thunderstorm winds and tornadoes, while identifying existing challenges and outlines future directions. DL effectively improved SCW forecasting accuracy and forecast lead times for integrating multi-source observational data and analyzing both observational and numerical weather prediction (NWP) datasets. Current challenges include lack of physical constraints, weak interpretability and insufficient high-quality samples. Future efforts will leverage higher-resolution and fully-integrated multi-source observation data to deepen the understanding of the small-scale structure and stage characteristics of SCW. Physical-based mechanism understanding, numerical prediction, and DL technologies are being continuously advanced in an integrated manner, aiming to build cognitively capable forecasting foundation models that can better support forecasters’ operational expertise and drive SCW forecasting toward greater intelligence and interpretability.
Liu et al. (Wed,) studied this question.