Towards Skillful Tropical Cyclone Forecasting by AI-Model-Driven High-Resolution Regional Coupled Model

Abstract

With the recent rise of artificial intelligence (AI), data-driven global weather forecasting models have demonstrated superior performance compared to state-of-the-art physics-based global models across various weather elements. This work reports on tropical cyclone (TC) simulations using a hybrid weather modeling system that harnesses the advantages of both AI-based and physics-based models. The system utilizes AI-based global models, Pangu-Weather and AIFS, to drive the atmospheric model within a regional atmosphere–ocean-wave coupled model (abbreviated as UWIN-CM). It preserves skillful TC track forecasting from the global AI models while gaining the benefits of predicting fine-scale details contributed by the high-resolution UWIN-CM model. The performances in forecasting seven TCs that necessitated the issuance of TC warning signals in Hong Kong in 2024 are studied. Results show that the AI-model-driven UWIN-CM can achieve a reduction in track error by 34% compared to the UWIN-CM driven by IFS. The track error is reduced to a level comparable to that of the AI models themselves. In terms of intensity, the AI-model-driven UWIN-CM also gives a reduction in intensity error by 20% compared to the UWIN-CM driven by IFS, and very significantly improves the intensity forecast provided by the AI global models. Other forecasting aspects, such as genesis, rapid intensification, and wind structure of TCs, are also investigated. The AI-model-driven results generally outperform those driven by IFS in these aspects. This work demonstrates that AI-based global models and high-resolution physics-based regional models can complement each other to achieve more accurate TC forecasts.