we_are_coded.by CODE · The world, decoded
БГ
DeepMind

A cyclone model gained a full day's lead at once - and shipped open source

NatureScience

WeatherNext predicts a storm's path, strength and wind extent up to 15 days ahead. Per the Nature paper, it gives an average of a day or more lead over the leading operational models. One forecast computes in under a minute on a single chip.

In short
  • Published in Nature on 6 August. The evaluation covers the 2023-2025 seasons; at a three-day forecast the position error is about 100 kilometers.
  • What's new is that the model gives both strength and wind extent, not just the track. It computes up to 1000 scenarios versus about 50 for classic ensembles.
  • Its role is advisory. The repository itself states it isn't an officially supported product and doesn't replace warnings from national agencies.
Checked on7 August 2026Responsible editorTsvetelin IvanovHow we workMethod · Corrections

The number here is good. What matters more is what it buys. One extra day's warning on a hurricane is the difference between an orderly evacuation and a last-minute run. And that day was won all at once - not over a decade of slow progress.

The facts: the paper came out in Nature on 6 August, accepted 24 July, submitted in December. The WeatherNext Cyclones model predicts track, strength and wind radii up to 15 days ahead. Per the abstract, it gives an average of a day or more lead over the leading operational models. The numbers shown in the charts for the 2023-2025 period: about 100 kilometers of position error and about 11 knots of strength error at a three-day forecast, compared respectively to the European centre's ensemble and to the US hurricane model. This year the ensemble reaches 1000 storm scenarios versus 50 last year. One 15-day forecast computes in under a minute on a single chip, and that's on data with a grid step of about 28 kilometers - a hundred times coarser than classic models. The US National Hurricane Center used it during the 2025 season, including for Hurricane Melissa. The code is released under an open license.

I want to flag something usually skipped over. The fair criticism of machine forecasts so far was that they know where a storm is going, but not how strong it will get. This paper hits exactly there. And it pulls out an uncomfortable conclusion: coarser data carries more signal about strength than was thought so far. Meaning the high resolution that supercomputers get built for may not be a required condition.

The machine gives the scenarios. The warning is still issued by a human.

And here's where you need to read carefully. The announcement talks about joint work with forecasters from the national center and the UK's met office. The code repository itself, though, carries a legal note: the model wasn't made jointly with, nor endorsed by, any government weather service. The two things don't contradict each other - people took part, no institution signed off - but this exact distinction is what journalists routinely erase.

The best result in the paper comes when the model is blended with the others. A boring but true way for the new thing to enter service - it doesn't displace the old, it sits next to it. And since the code is open and one forecast computes on a single chip, agencies without a supercomputer get access too. That's the bigger news than any kilometer of error.

The visual is generated code art. No third-party images.
Follow usFacebookLinkedIn
Official primary sources
→Nature - Operational Tropical Cyclone Forecasting with AI→DeepMind - WeatherNext: forecasting cyclones→GitHub - google-deepmind/weathernext
Original: https://wearecoded.com/en/articles/weathernext-cikloni-den-avans.html
ShareFacebookXLinkedInTelegramWhatsApp
← Back to all news