Fully AI-driven regional system for skillful tropical cyclone forecasting

Forecasting tropical cyclones (TCs) remains one of the most enduring challenges in numerical weather prediction, particularly in terms of storm intensity and track at medium to extended lead times. Recent advances in artificial intelligence (AI)-based global models have demonstrated strong skill in capturing large-scale atmospheric dynamics; however, their coarse spatial resolution and limited variable suites constrain their ability to produce storm-resolving forecasts critical for high-impact events. Here, we introduce the first fully AI-driven regional TC forecasting system, in which the European Centre for Medium-Range Weather Forecasts’ AI Forecast System (AIFS) is run locally to provide all initial and lateral boundary conditions for a convection-permitting regional model. This framework operates using a single atmospheric analysis and incorporates dynamic vortex initialization and spectral nudging to enhance storm realism and environmental consistency. Applied to Northwest Pacific TCs, the system demonstrated substantial improvement in intensity forecasts relative to those of the AIFS alone, whilst also delivering more accurate track guidance compared with that of conventional regional systems driven by physics-based global models. These gains highlight the system’s ability to simultaneously deliver accurate forecasts of both track and intensity—a capability traditionally considered difficult to achieve within a single framework. The proposed approach establishes a scalable, efficient, and operationally viable paradigm for tropical cyclone prediction based entirely on AI-driven regional modeling.