The company has released the code and model weights for WeatherNext 2 alongside WeatherNext Cyclones, allowing meteorological agencies, scientists and developers to adapt the technology for research, operational forecasting and specialised regional applications. The move follows publication of results in Nature showing a substantial improvement over leading operational forecasting systems.
WeatherNext Cyclones was evaluated on tropical cyclones covering the 2023-2025 period. Across predictions of storm tracks, intensity and wind radii, the system produced an average lead-time advantage of about 24 hours or more compared with leading models. Researchers said the improvement was comparable to roughly a decade of progress in conventional operational cyclone forecasting.
The gain does not mean every cyclone will now be predicted exactly one day earlier. Rather, a WeatherNext prediction made several days ahead can achieve an accuracy comparable with conventional models issued roughly a day later. That distinction is significant for emergency planners because additional reliable lead time can allow authorities to move equipment, prepare shelters and organise evacuations before conditions deteriorate.
The technology combines global atmospheric forecasting with specialised information about tropical cyclones. Its training used almost 20 terabytes of atmospheric data as well as the International Best Track Archive for Climate Stewardship, or IBTrACS, which contains records covering nearly 5,000 historical storms. The system learns both the large-scale atmospheric flows controlling cyclone movement and the smaller-scale characteristics associated with storm strength.
Cyclone forecasting has traditionally posed a difficult computational problem. Global models are effective at representing broad atmospheric circulation that determines where a storm travels, while high-resolution regional models have generally been needed to estimate intensity changes around a cyclone’s core.
WeatherNext Cyclones produced strong results despite operating with atmospheric information at a resolution of about 28 kilometres by 28 kilometres. That is substantially coarser than the resolution normally associated with specialised cyclone models, raising scientific questions about how much information about storm intensification can be extracted from large-scale atmospheric patterns.
The model can generate as many as 1,000 possible forecast scenarios, rather than relying on a single deterministic prediction. Such ensembles allow forecasters to assess probabilities surrounding uncertain outcomes, including low-probability but potentially catastrophic rapid-intensification events.
A 15-day forecast can be generated in less than a minute using a Google tensor processing unit. WeatherNext 2 uses a technology called a Functional Generative Network, which introduces controlled variation into predictions so that hundreds of physically coherent possible weather outcomes can be produced from the same starting conditions.
The broader WeatherNext 2 system, introduced in November 2025, provides forecasts extending as far as 15 days with resolution down to one-hour intervals. Google says it outperformed its previous WeatherNext generation across 99.9 per cent of tested combinations of atmospheric variables and forecast lead times, including temperature, wind and humidity.
The cyclone technology also underwent operational testing during the 2025 Atlantic hurricane season. During Hurricane Melissa, WeatherNext projected a Category 5 landfall in Jamaica five days ahead with an estimated 80 per cent probability while other model guidance still showed considerable uncertainty about the storm’s development and destination. The probability subsequently rose to almost 100 per cent three days before landfall.
Google has worked with the US National Hurricane Center on evaluating the system and is expanding research partnerships with forecasting agencies in the Philippines, Taiwan, Indonesia and Vietnam. Cooperation with agencies in Japan, Australia and India is also planned. Official warnings remain the responsibility of national meteorological authorities rather than Google’s AI systems.
WeatherNext 2 forecast information has already been made available through Earth Engine and BigQuery, while customised inference is being offered through Google’s Vertex AI platform. WeatherNext technology is also being incorporated into Search, Gemini, Pixel Weather and Google Maps services.
The open-source release could broaden the technology’s significance beyond major national forecasting centres. Universities and weather agencies with limited access to expensive supercomputing infrastructure may be able to experiment with advanced ensemble forecasting at substantially lower computational cost. Google has additionally released WeatherNext 2-mini, a compact model designed to operate on a single TPU through a publicly accessible Colab environment.
Follow Arabian Post
Select Arabian Post as your preferred source on Google and MSN News for trusted business news and Arab politics and updates.
