Do AI-Based Ensembles Improve Typhoon Inland Extreme Rainfall Prediction through Better Representation of Physical Processes? A Case Study of Typhoon GAEMI (2024)

Abstract AI-based ensemble weather prediction is emerging as a new paradigm for probabilistic forecasting, yet its ability to represent the key physical processes that govern high-impact extreme rainfall remains unclear. Here, we compare an AI-based ensemble system (AIFS-CRPS) with a conventional physics-based ensemble (IFS-ENS) in relation to the July 2024 extreme rainfall event in Hunan, China, associated with the remnants of Typhoon GAEMI. Both systems captured the large-scale process chain linking the western Pacific subtropical high (WPSH), remnant-vortex evolution, moisture transport, and inland convergence, but differed in their representation of the steering environment. In this event, AIFS-CRPS members were more concentrated around the rainfall-producing regime, whereas IFS-ENS sampled a broader range of circulation states. These case-specific results illustrate how the organization of large-scale ensemble uncertainty can influence rainfall placement and provide a process-oriented framework for interpreting ensemble forecasts. Broader multi-event evaluation is needed to determine whether this relationship is systematic across high-impact rainfall events.