14-storm study finds AI-driven typhoon forecasts cut day 4–5 track errors 40%–50%

At 96 to 120 hours ahead, the conventional regional systems (a WRF run driven by the ECMWF physics-based forecast, and NOAA's HAFS) had track errors above 500–600 km. The AI-driven system held errors to about 320–340 km, a 40%–50% reduction. Accurate long-range tracks matter for early warning and disaster response.

The system still struggles with rapid intensification and weakening, and intensity errors of 10–15 m/s persisted across several cases. The test covered one season and 14 storms in one basin, so how it performs in other years and regions is not shown.

The team targeted a known trade-off: AI global models such as AIFS predict storm tracks well but at roughly 31 km resolution they cannot resolve eyewalls and inner-core structure, so intensity suffers. Running AIFS locally and nesting a 2-km WRF model inside it, with a dynamically balanced synthetic vortex built from best-track data and spectral nudging of large-scale fields, was meant to keep AIFS's track skill while recovering storm-scale detail.

Spectral analysis of Typhoon Yagi forecasts showed the nested model kept substantially more wind energy at 50–200 km scales than AIFS, which lost fine-scale variance even relative to its starting conditions. Verification used mean absolute errors against China Meteorological Administration best-track data at 12, 24, 48, 72, 96 and 120 hours, with all WRF runs sharing identical physics to isolate the effect of the boundary forcing.

In case studies, the AI-driven system avoided the premature northward deflection EC_WRF showed for Typhoon Gaemi and the late-stage divergence in HAFS. For Kong-Rey it reproduced the observed recurvature and avoided the excessive intensity bias of the regional baselines. The authors say the setup could suit forecasting centres with limited access to conventional global model output.

The authors list remaining work: higher vertical resolution in AI models, ensemble-based probabilistic forecasts and better storm-core observations for training and validation. Rapid intensification and weakening remain a major limitation of the system.