How DeepMind’s WeatherNext Gives AI a One‑Day Head‑Start on Superstorm Intensity Forecasts

WeatherNext AI Model for Global Tropical Cyclone Forecasting

WeatherNext combines two data streams: (1) global atmospheric analysis data that captures large‑scale weather patterns, and (2) a curated historical tropical‑cyclone database (IBTrACS) containing nearly 5,000 past storms. The model is trained end‑to‑end on roughly 20 TB of global weather data, enabling it to learn a form of “meteorological background knowledge”.

The system generates probabilistic forecasts rather than a single deterministic trajectory. It can produce ensemble scenarios covering up to 15 days ahead and supports ensembles of up to 1,000 members.

Evaluation against Operational Models

Five‑day average track error: 230 km, which is 140 km lower than the leading operational ensemble system ECMWF‑ENS.

At the same accuracy level, WeatherNext provides about a 30‑hour lead‑time advantage over ECMWF‑ENS.

Intensity forecasts surpass those of the high‑resolution regional model HAFS, marking the first time a global AI‑based system outperforms a dedicated hurricane‑intensity model.

Resolution and Model Variants

WeatherNext operates at a 28 km grid spacing—approximately 100 times coarser than traditional high‑resolution models—yet matches or exceeds their performance on track and intensity metrics. A smaller variant, WeatherNext 2‑mini, runs at 111 km resolution and still delivers strong results.

Operational Deployment

Since June 2025 the model has been running in real time, providing forecasts to the U.S. National Hurricane Center and other World Meteorological Organization agencies. Forecast outputs have been adopted by operational forecasters.

Open‑Source Release

The model code and trained weights have been open‑sourced to enable collaboration between machine‑learning researchers and operational forecasters.

Key references: Nature article “Operational Tropical Cyclone Forecasting with AI” (2026‑08‑06) https://www.nature.com/articles/s41586-026-10953-2; DeepMind blog post https://deepmind.google/blog/weathernext-ai-model-achieves-breakthrough-in-forecasting-cyclones/.

Code example

来源：ScienceAI
本文
约2000字
，建议阅读
5
分钟
面向全球热带气旋的AI预报系统。