DeepMind’s Weather AI Predicts Hurricanes a Day Earlier Than Traditional Forecasting

When Hurricane Melissa made landfall in Jamaica in 2025, it was the strongest storm ever to hit the island. The hurricane’s rapid intensification left forecasters stunned.

But thanks to WeatherNext, an AI model developed by Google DeepMind, the island had an early warning. Working with the National Hurricane Center, the model predicted Melissa’s sudden jump in strength with nearly 100 percent confidence three days in advance. That gave experts more time to help people prepare and evacuate. It was the first time a storm that began with relatively low wind speeds was successfully predicted to reach Category 5.

When it comes to cyclones—including hurricanes and typhoons—every extra hour counts. These storms are among nature’s most destructive weather events and notoriously hard to anticipate. A cyclone’s path and strength can change rapidly. Seemingly tame storms can explode into monsters; those expected to skirt populated areas can suddenly veer towards a city. Longer forecasts gives communities time to mobilize resources and get out of harm’s way.

But cyclones are chaotic systems. Tiny differences can dramatically alter their behavior, making them harder to predict the further out we look. Existing forecasts rely on physics-based simulations that extrapolate two days ahead. But DeepMind says their algorithm extends the warning period to three days without sacrificing accuracy.

An extra day may seem trivial. But “this scale of improvement corresponds roughly to a decade’s worth of meteorological progress,” the team wrote in a blog post.

Beyond cyclones, WeatherNext also generates 15-day weather forecasts faster and using less energy than conventional models. That’s not to say it’ll replace them though. Instead, the two complement each other, giving human forecasters better information to guide critical decisions.

“By combining advanced machine learning with the indispensable real-world expertise of human forecasters, we aim to create a collaborative weather forecasting ecosystem that can save lives and help communities adapt to a changing climate,” the team wrote.

Crystal Ball

Predicting weather has always been challenging. Standard forecasting software uses physical models of the Earth’s atmosphere, incorporating temperature, air pressure, wind, humidity, and many other variables. It then calculates how these factors will evolve. Given current pressure and temperature gradients and moisture levels, for example, how will air move, and how likely is it that moisture will condense into clouds and rain?

Supercomputers crunch the numbers and churn out predictions. Though relatively accurate, the process is slow—often taking hours—costly, and rigid. Weather is one of the most complex physical systems on Earth, and even small changes in conditions can throw these models off.

So DeepMind turned to AI. Five years ago, they developed an AI modeI that outperformed physics-based models at 90-minute forecasts. In 2023, the AI lab’s GraphCast algorithm nailed 10-day predictions from historical data, beating leading systems roughly 90 percent of the time across thousands of scenarios. GenCast soon followed, cutting the time and energy required to generate predictions. Broadly speaking, these systems divide the globe into small geographical chunks called pixels and learn how weather conditions in one area influence neighboring areas.

But extreme weather presents an additional challenge. Massive databases exist to train AI on everyday weather patterns. Cyclones, on the other hand, are relatively rare and highly unpredictable.

One way to tackle this problem it to generate many slightly different versions of what might happen by adding random noise after training. But because the noise affects each pixel differently, it can disrupt their relationships and produce unrealistic weather patterns.

For WeatherNext, DeepMind instead built uncertainty into the AI itself.

Bridging the Gap

 There’s traditionally been a tradeoff between accuracy and scale in cyclone prediction.