Google DeepMind's WeatherNext Gives Forecasters a Full Extra Day to Warn People About Cyclones

Key points

• Google DeepMind published a paper in Nature showing its WeatherNext AI model predicts cyclone track, intensity and wind structure with roughly 24 extra hours of lead time versus the best prior models.
• During the 2025 hurricane season, WeatherNext helped the US National Hurricane Center forecast Hurricane Melissa's rapid intensification and landfall in Jamaica.
• The model runs on a TPU, a Google-made AI chip, and produces a 15-day forecast in under a minute, generating 1,000 possible storm scenarios at a time.
• DeepMind is open-sourcing the code and weights for WeatherNext Cyclones, WeatherNext 2, and a lightweight version called WeatherNext 2-mini that runs in a free public notebook.
• Tropical cyclones have killed more than 700,000 people and caused $1.4 trillion in damage worldwide over the past 50 years.

A hurricane forecast is, at heart, a race. Every hour a warning goes out earlier is another hour to board up windows, move boats, or get people off the coast.

Google DeepMind says its new AI weather model has just won back a full day of that race.

In a paper published in Nature, the lab reports that its WeatherNext system predicts where a cyclone will go, how strong it will get, and how its winds are shaped, more accurately than any previous model. On average, its three-day forecast is as sharp as the best two-day forecasts were before. Meteorologists estimate that is worth about a decade of normal progress in the field.

And DeepMind is giving it away. The team is open-sourcing the code and model weights so any national weather service, university lab or nonprofit can run them.

What did the model actually do during a real storm?

It helped forecasters see Hurricane Melissa coming. In 2025, WeatherNext predicted that Melissa would rapidly intensify and make landfall in Jamaica, and the US National Hurricane Center used that signal to issue an unusually early warning. Rapid intensification, when a storm suddenly jumps in strength over a few hours, is the scenario forecasters fear most, because it leaves the least time to react.

This year the system is running 1,000 different simulated versions of each storm at once, a technique called an ensemble forecast. That lets forecasters see not just the most likely path, but rare and dangerous outliers too.

Why is this a big deal technically?

Cyclone forecasting has always been split in two. Where a storm goes is driven by huge global wind patterns, best captured by big, coarse global models. How strong a storm gets is driven by tiny, churning physics near its eye, which usually needs a separate, high-resolution local model.

WeatherNext does both jobs with one AI. Odder still, it does the intensity job well while looking at the world in fairly chunky 28-by-28-kilometre squares, about a hundred times coarser than traditional storm models. A smaller version, WeatherNext 2-mini, works at 111-by-111-kilometre resolution and still performs strongly. DeepMind admits it does not fully understand why the model is this good at that resolution, and calls it an open research question.

The system was trained on nearly 20 terabytes of global atmospheric data plus IBTrACS, an expert-curated record of roughly 5,000 historical storms.

What models are being released, and where can people see them?

Forecasts are viewable on Google's Weather Lab site, which now shows temperature, rain, wind and cyclone tracks in one view.

What does this mean for people in a storm's path?

Better, earlier warnings from your national weather agency. You will not be downloading WeatherNext yourself. What changes is what the National Hurricane Center, the UK Met Office and other agencies can tell you, and how confidently they can tell you days out. For official warnings, keep listening to your local meteorological service. The AI is a tool for the forecasters, not a replacement for them.