When Hurricane Melissa barrelled toward Jamaica in October 2025, it was still a relatively modest Category 1 storm. Most forecasting models were uncertain about its trajectory and strength. But one system was not: Google DeepMind’s WeatherNext AI flagged an 80% probability five days in advance that Melissa would intensify to Category 5 and make direct landfall on the island. It proved exactly right. Forecasters at the US National Hurricane Center had never before predicted a Category 5 hurricane from Category 1 strength. The extra warning time allowed evacuation efforts to begin earlier, saving lives in what became the strongest storm ever to hit Jamaica.
It was the real-world debut of a system that, in August 2026, was published in Nature and officially open-sourced to the world.
The Problem With Traditional Forecasting
Hurricane prediction has always involved an uncomfortable trade-off. Global physics-based models — which simulate the entire atmosphere using complex mathematical equations requiring hours of supercomputer time — are good at predicting where a storm will go, but too coarse to capture how strong it will get. Specialised regional models handle intensity better but lose track accuracy. Forecasters have long had to stitch together two imperfect tools to get a full picture.
Rapid intensification — when a hurricane’s wind speed jumps by at least 34 mph within 24 hours — is the most dangerous and historically the hardest scenario to predict. Melissa was exactly that kind of storm.
What WeatherNext Does Differently
DeepMind’s WeatherNext Cyclones model was built to solve both problems simultaneously in a single system. Trained on nearly 20 terabytes of atmospheric data and nearly 5,000 historical storms going back to 1940, it learns the patterns that precede intensification rather than simulating atmospheric physics from first principles — allowing it to spot signals that traditional models miss.
The headline result, now validated in Nature, is striking: WeatherNext’s three-day forecasts are as accurate as what the best existing systems could deliver at two days. One extra day of warning may sound modest, but for coastal communities facing a major hurricane, it can mean the difference between an ordered evacuation and a desperate scramble. As the DeepMind team put it, that single-day gain “corresponds roughly to a decade’s worth of meteorological progress.”
The speed advantage is equally dramatic. A full 15-day ensemble forecast — running 1,000 possible storm scenarios to capture the full range of uncertainty — now completes in under a minute on a single specialised chip. Traditional models running the same calculations would require hours on supercomputers costing vastly more energy.
Better in Almost Every Way
On track prediction, WeatherNext’s five-day estimate places the storm centre an average of 230 kilometres off target — compared to 370 kilometres for the leading European ensemble model. On three-day intensity forecasts, it outperforms the best operational intensity model by nearly five miles per hour in wind speed accuracy.
Perhaps most surprisingly, it achieves all of this using weather data at roughly 100 times lower resolution than conventional high-resolution hurricane models. The AI extracts more insight from less raw detail — a finding that genuinely surprised the researchers.
For the 2026 hurricane season, the system has been scaled up to generate 1,000 ensemble scenarios per cyclone, up from 50 in 2025, making it powerful enough to surface rare but catastrophic tail-risk events — the sudden intensifications that have historically caught forecasters and communities off guard.
Open for Everyone
In a significant move, DeepMind has now open-sourced WeatherNext Cyclones alongside its companion general weather model, WeatherNext 2, on GitHub under a permissive licence. A lightweight version can run in a free Google Colab notebook, putting meaningful hurricane forecasting capability within reach of researchers and agencies in countries that lack the infrastructure for traditional supercomputer-based modelling.
The collaboration behind the model — involving the US National Hurricane Center, the UK Met Office, and academic research institutions — was designed from the outset to keep human forecasters at the centre. WeatherNext doesn’t replace meteorologists; it gives them better information, faster, to guide decisions that remain ultimately human.
The Bigger Picture
Hurricane forecasting has improved steadily for decades, but progress has been measured in small steps — a few hours here, a marginal accuracy gain there. WeatherNext represents something qualitatively different: a step change driven not by better physics simulations but by a model that learned what severe storms look like across thousands of historical examples.
For the hundreds of millions of people living in hurricane-prone coastlines around the world, an extra day of reliable warning is not an incremental improvement. It is, in the most literal sense, a matter of life and death.