The short version
- The new AI model provides forecasters with approximately one additional day of accurate prediction compared to traditional methods.
- It successfully predicted the rapid intensification of Hurricane Melissa, allowing for earlier warnings in Jamaica.
- Researchers remain uncertain how the system achieves high accuracy using lower-resolution atmospheric data than previously thought necessary.
A significant advancement in meteorological forecasting has emerged from Google’s DeepMind and Google Research, offering emergency planners a crucial extension of lead time for hurricane response. The new artificial intelligence model, known as WeatherNext, demonstrates the ability to predict cyclone trajectories and intensities with greater accuracy than existing systems. On average, this technology provides forecasters an additional day of reliable prediction. This means that forecasts generated three days in advance by the AI are now comparable in precision to two-day forecasts produced by traditional models.
The practical value of this extra time was evident during Hurricane Melissa in October 2025. As the storm developed over the Caribbean Sea, conventional weather models offered conflicting projections regarding its path and strength. Some suggested it would remain a weak system affecting Haiti, while others hinted at a more dangerous trajectory toward Jamaica. WeatherNext predicted with eighty percent confidence that the storm would strike Jamaica as a Category 5 hurricane five days before landfall. This early warning allowed communities to prepare for the catastrophic flooding and landslides that followed.
Mike Brennan, director of the US National Hurricane Center, emphasized the operational importance of these additional hours. Evacuations, supply staging, and resource deployment are highly time-sensitive activities where incorrect decisions can have severe consequences. Historically, achieving a one-day improvement in forecast accuracy required a decade of scientific effort. The ability to push predictive capabilities forward by even a few hours is considered highly valuable for mitigating risk and saving lives.
Predicting hurricanes presents unique challenges because these storms operate across multiple spatial scales. Determining a storm’s track requires global data, including information on cold fronts and prevailing winds. However, predicting intensity demands much finer-scale data focused on local atmospheric and oceanic conditions. Traditional global models often fail to capture these smaller-scale details, which is why earlier AI systems struggled with intensity predictions despite performing well on trajectory tracking.
Ferran Alet, a research scientist at Google DeepMind and lead author of the study published in Nature, explained that extreme weather events are rare, limiting the amount of specific cyclone data available for training machine learning models. To overcome this, the team trained WeatherNext to excel at general weather prediction as well as cyclone-specific forecasting. This approach leveraged the vast amounts of available weather data to improve performance on rarer, more dangerous events.
Kate Musgrave, tropical cyclone group lead at the Cooperative Institute for Research in the Atmosphere and a co-author of the paper, noted that Hurricane Melissa marked a historic milestone. It was the first time the National Hurricane Center predicted a Category 5 hurricane when the system was still only a Category 1 storm. Such rapid intensification can transform a manageable situation into an emergency overnight, making early detection critical.
Despite its success, the model’s internal mechanics remain partially opaque to its creators. WeatherNext achieves high accuracy using lower-resolution atmospheric data than traditional models require for intensity forecasting. This has surprised the scientific community, as it suggests that coarser inputs contain more predictive signal than previously understood. Alet described the system as a black box, noting that while researchers do not fully understand how it extracts these insights, the results provide physicists with new signals about atmospheric dynamics.
Rather than producing a single deterministic forecast, WeatherNext generates a wide range of potential scenarios for developing storms. This approach helps account for the butterfly effect, where small deviations can lead to significant changes in outcomes. The model now generates one thousand scenarios per storm, up from fifty last year. Forecasters use these probabilistic outputs alongside other models to inform their final predictions, improving their ability to anticipate how a storm system will likely unfold.
Sources behind this briefing
Go to the original reporting
- Ars Technica↗DeepMind’s hurricane breakthrough has surprised weather scientists
