Scientists at Google DeepMind and Google Research today launched a new AI model for weather forecasting that sees our changing atmosphere more clearly and predicts its behavior more frequently. WeatherNext 3 is the latest wave of a sea change in meteorology brought about by deep learning techniques, and Google says it will begin feeding the weather information that users see in search, Google Maps and Gemini, as well as making it available to users and researchers on Google’s cloud platforms. “This will be the first time that some of the core variables feed and power many of Google’s products,” Samier Merchant, a senior engineer at Google, told TechCrunch. The new model has already proven to be the most accurate among the main competitors tested in Operational WeatherBench, a utility for comparing AI forecasts created by the startup Brightband. Analyze metrics such as temperature, wind speed and humidity. In addition to outperforming other deep learning models created by Google, Microsoft, Nvidia, and the European Center for Medium-Range Weather Forecasts (ECMWF), it also outperforms traditional forecasts from the US National Weather Service and ECMWF. Image credits: Brightband Most weather forecasts come from government-owned supercomputers that laboriously process mathematical equations written to describe the physics of climate; While these systems have become remarkably accurate, they are expensive and comparatively slow. After ECMWF released more than half a century of weather data produced by these systems in 2018, deep learning researchers began training models that could make predictions much more quickly and with accuracy comparable to government tools. “The climate is chaotic, so small differences really start to be hugely disruptive…Machine learning targets the problem we’re really solving, which is noisy, approximate physics from incomplete information and finite computation, so it learns patterns from a huge amount of data,” said Ferran Alet, research scientist manager at DeepMind. Since then, model makers have exploited key weaknesses of AI forecast models: they tend to forecast over a wider area (15 to 25 square kilometers) than is really useful, they are not always good with rain, and they still rely on formatted data sets produced by government agencies. WeatherNext 3 takes on all three challenges. On key variables, the researchers told TechCrunch, it can predict up to a resolution of 5 km. Its rainfall assessments have improved by 60% over WeatherNext 2 and it can now produce forecasts every hour, instead of the standard six-hour forecast. Image credits: Google Those improvements are the result of specific decisions made by the designers. WeatherNext 3 is a larger model, with 2.4 times more parameters than its predecessor, adapting the targets of the decoder heads to provide more useful responses. While most weather forecasts are generated as metrics averaged on a 3D grid, DeepMind researchers have already won plaudits by adjusting their model to also visualize cyclone tracks. This time, the designers also trained the model to target its forecasts to specific weather data stations. This is important not only to provide more granular predictions, but also to be able to evaluate your work with specific, real-world data. “The idea, with many AI applications, is to try to execute tasks as end-to-end as possible,” said Daniel Rothenberg, an atmospheric scientist at Brightband. “Adding a capability where this model now also predicts, say, what the Denver airport weather station is going to measure every hour, just plugs that forecasting task closer to the core.” The model can forecast more frequently because it can ingest data from weather satellites collected in real time every hour. Feeding AI models with raw empirical observations, rather than analyzes produced by weather supercomputers, promises more accurate forecasting, but it remains technically challenging to get the models to work on raw data. Google says WeatherNext 3 is the “first” AI model to directly incorporate raw observations for a high-resolution global forecast, but AI weather startup WindBorne says its model, WeatherMesh 6, has been incorporating raw observations from its fleet of weather balloons and other sources since late 2025. When asked about it, Google noted that its forecasts have higher resolution around the world. Regardless, both models still rely on national weather data sets for forecasting, so more work will be required for true direct data assimilation. While LLMs receive most of the attention, the transformer revolution in meteorology has been equally important. European and US weather agencies are already using AI models in their forecasting products, and their speed and low cost promise to deliver economic impact in poorer regions where spending on high-quality sensors and supercomputers has put accurate forecasts out of reach. Bill Gates recently cited AI-based weather forecasting as a crucial benefit of the technology, as better forecasts improve crop yields in developing countries. Alet, the DeepMind researcher, said higher-resolution forecasts of wind, rain and cloud cover will be useful in making renewable energy projects more reliable. “At the end of the day, I think Google is about providing useful information to the user, and a lot of what users search for has to do with weather in one way or another,” Alet said. When you buy through links in our articles, we may earn a small commission. This does not affect our editorial independence.
