Characterizing Extreme Rainfall Events Associated with Large-Scale Circulation by LLM

Jui-Chung Hung

Abstract: We integrated retrieval-augmented generation (RAG) with a multimodal large language model (LLM) to cluster large-scale circulation patterns associated with extreme rainfall events (>80 mm day−1) in Taiwan. Employing an event-perspective approach on 62 years (1960–2022) of ERA5 reanalysis data, we identified discrete rainfall events by season. The LLM-based classification effectively captured intra-seasonal variability and mixed-regime events, outperforming conventional seasonal categorizations. The results of this analysis indicate that 56.7% of extreme rainfall events occurred between July and October. During this period, the LLM–RAG framework performed event-based clustering and identified four representative synoptic patterns: typhoon systems, subtropical high perturbations, southwesterly monsoonal flow, and southeasterly flow regimes.