The next generation of predictive models arrives courtesy of Google DeepMind and Google Research with WeatherNext 2, an AI architecture designed to deliver more useful and accurate forecasts. In a context of increasingly extreme weather events, the ability to anticipate unlikely but high-impact scenarios It becomes key to planning.
Among its most notable advances, this system can produce forecasts 8 times faster and with a resolution of up to one hourIt also generates hundreds of possible atmospheric evolutions from a single starting point. According to internal tests, it outperforms the previous model in the 99,9% of variables and time horizons from 0 to 15 days.
What is WeatherNext 2 and why does it matter?
WeatherNext 2 is an AI-powered prediction model that is incorporated into core of Google's forecasting systemThis is what powers the weather functions in their products. This translates into visible improvements for end users and for multiple sectors that depend on the atmosphere.
Google has already begun using this technology to enhance forecasts in Search, Gemini, and Pixel Weather, with plans to I also got to Google Maps Its Weather API is already integrated into the map platform. Adoption is gradual and focuses on providing more consistent and detailed information.
How it works: the Functional Generative Network approach
WeatherNext 2's performance relies on an approach called Functional Generative Network (FGN)This method introduces noise directly into the model architecture, so that the forecasts remain realistic and physically consistent, avoiding inconsistencies between variables.
The FGN trains the system on what meteorologists call "marginals," that is, isolated elements such as a specific temperature, wind speed, or humidity. From there, the model learns to forecast "joints." complex and interconnected systems, such as large areas of heat or the expected production of a wind farm.
Starting from a single initial atmospheric state, WeatherNext 2 generates hundreds of possible pathsThese simulations are fed back into the model itself to consolidate the coherence between variables and regions, capturing both the most probable evolution and the extremes that are of most interest to risk management.
Under the hood, the system calculates four 6-hour shifts a dayalways starting from the most recent global situation. This cadence allows the range of scenarios to be updated as the atmosphere changes and new signals emerge.
Speed, resolution, and accuracy
In computational terms, the leap is significant: Each prediction runs in less than a minute on a single TPU, whereas approaches based on traditional physical models would require hours on supercomputers for comparable tasks.
Besides speed, it provides temporary resolution of up to one hour and notable improvements in variables such as temperature, wind, humidity, precipitation, and pressure. Overall, it outperforms the previous model in 99,9% of variables and in most timeframes from 0 to 15 daysopening the door to more informed decisions.
Another advantage is its ability to incorporate the full range of possibilities, including low probability but high impact eventsThis probabilistic approach is essential for planning operations and emergencies without underestimating extreme scenarios.
Where you'll see it and how to access it
Google has integrated WeatherNext 2 into Search, Gemini and the Pixel Weather appand is expected to arrive on Google Maps and the platform's Weather API. Users will notice more detailed and consistent forecasts on a daily basis.
For businesses, scientists, and developers, access is available through Google Cloud Vertex AI, BigQuery and Earth EngineThis approach makes it easier to explore datasets, build industry-specific applications, and assess risks with greater granularity.
The roadmap includes integrate new data sources and expand access, with the aim of accelerating research and enabling solutions that rely on high-quality predictions.
Potential impact in Spain and Europe
In contexts like Spain and the rest of Europe, a faster and more accurate prediction can translate into improved management of heat waves, windstorms and heavy rainfallSectors such as energy, mobility, or agriculture tend to benefit from more consistent 0-15 day timeframes.
By generating multiple scenarios, WeatherNext 2 allows for the evaluation of worst cases and the dispersion of resultsThis is especially useful for civil protection, infrastructure operators, and water resource planning in the face of extreme events.
In the renewable energy sector, improvements in variables such as wind and cloud cover help to estimate wind and solar power production with greater time detail, optimizing the operation of the electrical system and reducing uncertainty in supply.
The new model combines speed, resolution, and physical consistency to deliver more useful forecasts for both users and organizationsIts integration into Google products and access via the cloud open up its use to multiple cases, from daily weather checks to planning weather-sensitive activities.
