Google DeepMind and Google Research have introduced a new version of the WeatherNext 3 model. According to an independent ongoing evaluation by Brightband, this system predicts global weather more accurately than any other available model. The main innovation lies in where the system obtains its input data. Instead of learning exclusively from the outputs of conventional physics-based simulations, it uses satellite imagery and measurements from ground stations directly. It can issue a new forecast every hour.
A sharper picture of the atmosphere
The usefulness of a forecast depends on how precisely it can describe a specific place and time. WeatherNext 3 calculates near-surface temperature and humidity on a grid with a resolution of five kilometers. It processes other surface values at a resolution of ten kilometers and maps variables in the upper layers of the atmosphere, such as wind speed, at twenty-five-kilometer intervals. Overall, this picture of the weather is approximately five times more detailed than that produced by the previous version, WeatherNext 2, which used only a twenty-five-kilometer grid and issued forecasts at six-hour intervals.
This difference is clearly demonstrated by a comparison of temperatures measured two meters above the ground over the British Isles. While the older model blurred the landscape into large, coarse areas, the newer version can handle even highly varied terrain. Mountains and valleys therefore no longer disappear beneath a single average value on the map.
Satellite data
Most artificial intelligence weather models learn from data generated by numerical weather prediction systems. These systems run on supercomputers, simulate atmospheric physics, and have a six-hour delay. As a result, deviations arise for variables that change quickly, particularly precipitation and near-surface temperatures.
The new model therefore continuously loads a mosaic of images from geostationary satellites, giving it a constantly updated overview of what is actually happening in the atmosphere. Each new hourly forecast is thus based on the latest observations. If a storm, frontal system, or precipitation band forms unexpectedly quickly, emergency services, dispatchers, and air traffic control receive information sooner and in greater detail.
Learning directly from weather stations
Temperature and humidity can change over distances of just a few kilometers, as anyone living on the coast, in enclosed valleys, or in foothills will know. Conventional models fail in these areas because they work with a simplified picture of the atmosphere in which local variations are lost. WeatherNext 3 is instead trained directly on data from ground-based weather stations, meaning that its five-kilometer grid accurately reflects the shape of the landscape.
Google expects this approach to deliver the greatest benefits in Latin America, Africa, and the Asia-Pacific region. Detailed local models are generally unavailable in these regions because operating them on supercomputers is extremely expensive. Detailed forecasts will therefore become available to billions of people and local businesses that previously had to rely on rough estimates.
Forecasting rain and snowfall is among the most complex tasks in meteorology. Precipitation forms in clouds through processes that take place on a very small scale and change rapidly, making them difficult for conventional physics-based simulations to capture. The result is often an inaccurate estimate or a forecast that completely misses the boundary of a severe storm.
The developers therefore trained the model using two exceptionally high-quality sources. The first is NASA's IMERG satellite product, while the second is its own global precipitation reconstruction based on satellite radar. Measurements show that the accuracy of medium-range precipitation forecasts has improved significantly. Compared with older methods, accuracy against IMERG data increased by as much as sixty percent, by thirty percent against the US MRMS network, and by ten percent against rain gauges for short-term forecasts. On maps, this improvement means that instead of blurred areas, sharply delineated precipitation bands appear that closely match actual satellite observations.
The model also calculates variables that are crucial for the renewable energy sector. It predicts wind speeds at a height of one hundred meters above the ground, corresponding to the level at which wind turbine blades rotate. It also provides detailed information about cloud cover and the intensity of solar radiation. This allows transmission system operators and energy companies to estimate the volume of electricity generated more accurately and better align it with current demand.
Where the model is already running and what this means for Europe
Since its announcement, the WeatherNext 3 model has been gradually integrated into Google's services, specifically Search, the Gemini app, Maps, the Google Maps Platform Weather API, and Google Earth Engine. Google plans to deploy it worldwide, with no regional exceptions for the European Union. People in the Czech Republic will therefore encounter the results of this model most often in ordinary weather forecasts in Search or Maps.
The greatest technological advance is expected in longer-range outlooks. When planning activities several days or a weekend ahead, users will receive precipitation forecasts that are up to fifty percent more accurate. The greatest improvement will be felt in areas where forecasts have previously lagged behind in reliability.
Those interested in working with weather data independently do not need to configure or train anything complicated. Globally available forecasts updated every hour can be queried using BigQuery and Earth Engine, or downloaded in bulk from Google Cloud Storage.
