DeepMind achieves breakthrough in weather forecasting

Ronni Holmvig Strøm · 2026-08-10

In October 2025 a storm was forming over the Caribbean, and the models could not agree on what it would become. One future had it staying weak and drifting toward Haiti. Another had it intensifying and turning for Jamaica. Five days before landfall, DeepMind's WeatherNext model committed to the second future with 80 percent confidence: Jamaica, Category 5. Hurricane Melissa arrived as forecast, and the extra warning gave communities in its path time they would not otherwise have had. Last Thursday, in a paper in Nature, the team showed the Melissa call was not luck. On average the model buys forecasters about a full day of lead time, and its builders open-sourced it the same week.

A day sounds modest until you sit with what it costs. Bringing cyclone forecasts forward by twenty-four hours has historically taken about a decade of meteorological work. WeatherNext delivers that step change in a single model, its three-day forecasts now as accurate as the previous generation's two-day forecasts. Mike Brennan, who directs the US National Hurricane Center, put the value plainly: "Time is really golden when it comes to those types of decisions."

The strange part is how the model does it. Intensity forecasting has always been treated as a resolution problem. To know how strong a storm will get, you model the fine-scale thermodynamics churning around its core, which demands high-resolution data and specialized local models. WeatherNext runs on inputs at 28-by-28 kilometers, about 100 times coarser than traditional systems, and still beats them on intensity. Ferran Alet, one of the lead authors, described the reaction: when the team told the community how coarse the inputs were, "they were shocked, because that means that the lower-resolution inputs capture more signal about what's going to happen than previously believed."

Cyclone prediction has long forced a division of labor. Track, meaning where the storm goes, is steered by continent-scale atmospheric currents that coarse global models handle well. Intensity, meaning how strong it gets, lives in the small-scale physics near the eye. Earlier AI models inherited the split. They forecast track respectably and, as Kate Musgrave of the Cooperative Institute for Research in the Atmosphere put it, "intensity they could not do well at all."

WeatherNext collapses the two problems into one network. It was co-trained on two kinds of data at once: roughly 20 terabytes of global atmospheric dynamics and the IBTrACS archive of nearly 5,000 historical storms. Scarce cyclone records taught it the rare behavior, abundant weather data taught it the physics underneath. From that base it produces not one forecast but an ensemble of possible futures, and the ensemble has grown fast. Last year the system generated 50 scenarios per storm. This year it runs 1,000, enough to catch the low-probability, high-consequence tail events like the rapid intensification that turned Melissa from a Category 1 into a Category 5, the first time the National Hurricane Center predicted a Category 5 while the storm was still only a Category 1. A full 15-day forecast takes under a minute on a single TPU.

The obvious worry writes itself: a system nobody fully understands, steering evacuation decisions. Alet does not dodge it. "It's a black box at the end of the day," he said, "but that gives physicists a signal that something is happening that was not previously understood." The model points to a piece of atmospheric physics the field had written off as unreachable at coarse resolution, and now has reason to go looking for.

Which is why open-sourcing it matters more than the benchmark. DeepMind released the code and the weights, including a compact version that runs on one TPU in a free Colab notebook. A meteorological agency with a modest budget can now run forecasts that beat what the frontier could produce two years ago. The decade of progress did not stay in the lab. It shipped, with the door left open behind it.

Sources: Ars Technica, 2026, Google DeepMind, 2026.