What if the weather forecast could update every hour and better detect the differences between a coast, a valley, and a mountain? Google DeepMind and Google Research are introducing WeatherNext 3, an artificial intelligence model that promises more accurate, faster, and more localized global forecasts.
The tool uses real-time weather observations, including data from geostationary satellites, to generate higher-resolution predictions. According to Google, independent evaluations by Brightband rank it as the company’s most advanced global weather model to date.
More detailed and frequent forecasts
WeatherNext 3 produces forecasts every hour and works with different spatial resolutions depending on the variable being analyzed:
- Surface temperature and humidity at a resolution of up to 5 kilometers.
- Other surface variables at a resolution of 10 kilometers.
- Atmospheric variables, such as wind speed, at a resolution of 25 kilometers.
Taken together, this represents a picture of the weather that is approximately five times more detailed than WeatherNext 2, which generated forecasts every six hours on a 25-kilometer grid.
Can you notice the difference in everyday life? It could be especially important in areas where the weather changes quickly or the terrain alters local conditions, such as coastal regions, valleys, and mountainous areas.
AI learns from real-time weather data
Many artificial intelligence models for weather are trained using information produced by traditional numerical forecasting systems. These systems simulate atmospheric physics using supercomputers, but their data can be several hours old.
WeatherNext 3 incorporates continuously updated image mosaics from geostationary satellites. This allows the model to observe how clouds, storms, and other phenomena evolve before generating a new forecast.
It also uses data from weather stations distributed across different parts of the world. This helps its forecasts represent regional details more accurately, instead of offering an overly smoothed-out view of the weather.
Resolution is not just a technical figure. A more detailed forecast can help you know whether rain will affect an entire city or only a nearby area.
Improvements in rain and snow
Precipitation remains one of the most complex challenges for weather models. Rain and snow depend on processes that occur inside small clouds and change quickly, so forecasts can end up showing areas that are too broad or miss the boundaries of an intense storm.
To improve this aspect, WeatherNext 3 was trained with high-quality precipitation data, including NASA’s IMERG satellite records and a global analysis system based on radar and satellite data.
Google reports that, in medium-range forecast evaluations, the model achieved improvements of up to:
- 60% compared with IMERG.
- 30% compared with MRMS.
- 10% compared with rain-gauge measurements during the first forecast periods.
These figures use the CRPS metric, or Continuous Ranked Probability Score, which evaluates the quality of probabilistic forecasts. In simple terms, it measures how closely the predicted probability distribution matches what ultimately happens.
A tool for renewable energy and agriculture
WeatherNext 3 is not designed only to tell you whether you should bring an umbrella. It also generates forecasts designed for sectors that depend directly on the weather.
The model estimates wind speeds at a height of 100 meters, roughly the height of many wind turbines. This can help calculate how much electricity wind farms will produce.
It also provides information about cloud cover and solar radiation. With these data, solar plants can better anticipate how much sunlight they will receive, while power grid operators can balance renewable generation with consumer demand.
For farmers, transportation companies, infrastructure operators, and emergency teams, more frequent updates can make it easier to decide on irrigation, routes, harvests, logistics, and responses to extreme weather events.
More useful for historically underserved regions
High-resolution regional weather models usually require enormous computing capacity. As a result, some regions of Latin America, Africa, and Asia-Pacific have not had access to the same level of detail as other markets.
Google says WeatherNext 3 can bring localized forecasts to billions of people without requiring every organization to build its own model from scratch. The data will be available to researchers, developers, and businesses through BigQuery, Earth Engine, and Google Cloud Storage.
The model will also begin integrating with products such as Google Search, Gemini, Google Maps, Google Maps Platform Weather API, and Google Earth Engine.
According to the company, precipitation forecasts for one or more days ahead could be up to 50% more accurate, with the greatest improvements in regions where forecasts have been less reliable.
Forecasting does not eliminate uncertainty
Artificial intelligence can analyze huge volumes of data and update a forecast quickly, but the atmosphere remains a difficult system to predict. No model can guarantee that a storm will behave exactly as calculated.
That is why forecasts generated by WeatherNext 3 should be understood as a support tool. For severe weather alerts, emergencies, and safety decisions, Google always recommends consulting each country’s official weather services.
The importance of this advance lies in bringing forecasting closer to what is actually happening on the ground. If AI can combine recent observations, greater geographic detail, and more frequent forecasts, weather stops being just information for planning your day and becomes a tool for protecting communities, managing energy, and making better decisions.
Original source
https://blog.google/innovation-and-ai/models-and-research/google-deepmind/introducing-weathernext-3