Particle.news
Download on the App Store

DeepMind’s WeatherNext Gives Forecasters an Extra Day of Hurricane Lead Time

DeepMind open-sourced the WeatherNext models so researchers and weather agencies can test and adapt them in real-world forecasting.

Overview

  • The WeatherNext Cyclones model and related WeatherNext 2 weights and code were published and released to the public on Thursday, August 6, 2026, alongside an unedited Nature manuscript and a DeepMind blog post.
  • Across retrospective tests and operational trials from 2023–2025, WeatherNext on average extends accurate lead time by about one day for cyclone track, intensity, and wind radii versus leading operational models.
  • Operational partners including the U.S. National Hurricane Center, CIRA, and the UK Met Office corroborate the model’s strong performance and say it helped forecast Hurricane Melissa’s rapid intensification and Jamaica landfall in 2025.
  • The model produces large probabilistic ensembles—now up to 1,000 members—and can generate a 15-day forecast in under a minute on TPUs, giving forecasters faster access to rare but high-impact scenarios.
  • Researchers note a key scientific puzzle and an operational caveat: WeatherNext achieves strong intensity skill using coarse inputs (about 28×28 km), and agencies must still address validation, integration, oracle and governance work before wide downstream adoption.