Scientists at Google Deepmind and Google Research today released a new artificial intelligence model for weather forecasting that understands our changing atmosphere more clearly and predicts its behavior more frequently.
WeatherNext 3 is the latest wave of big changes in meteorology brought about by deep learning technology, which Google says will begin feeding into the weather information users see in Search, Google Maps, and Gemini, and will be available to users and researchers on Google’s cloud platform.
“This is the first time that some of the core variables impact and power many Google products,” Google senior staff engineer Sameer Merchant told TechCrunch.
The new model has already proven to be the most accurate of the leading candidates tested on Operational WeatherBench, a utility for comparing AI forecasts built by startup Brightband. Look at indicators such as temperature, wind speed, and humidity.
It outperforms other deep learning models built by Google, Microsoft, Nvidia, and the European Center for Medium-Range Weather Forecasts, as well as traditional forecasts from the US National Weather Service and ECMWF.
Most weather forecasts are generated by government-owned supercomputers that painstakingly calculate mathematical formulas written to describe the physics of the weather. Although these systems have become significantly more accurate, they are expensive and relatively slow. After ECMWF published more than half a century of weather data generated by these systems in 2018, deep learning researchers began training models that could make predictions as accurate as government tools and much faster.
“Weather is chaotic, so small differences start to really mess things up…Machine learning targets the problems we actually solve, the approximate noisy physics that comes from incomplete information and finite calculations, so it learns patterns from large amounts of data,” says Ferran Allais, Staff Research Scientist Manager at DeepMind.
Since then, modelers have identified key weaknesses in AI predictive models. This means that AI forecasting models tend to predict over a wider area (15-25 square kilometers) than is actually useful, are not always suitable for rain, and still rely on formatted datasets created by government agencies.
WeatherNext 3 addresses all three challenges. Researchers told TechCrunch that they can predict key variables down to 5km resolution. Rain ratings are 60% better than WeatherNext 2 and can now generate hourly forecasts instead of the standard 6-hour forecast.
These improvements are the result of specific choices made by the designers. WeatherNext 3 is a larger model, with 2.4 times more parameters than its predecessor, and adjusts the targets of the decoder head to give more useful answers. Most weather forecasts are output as metrics averaged across a 3D grid, but DeepMind researchers have already achieved great success by adjusting their model to also visualize cyclone paths.
This time, the designers trained the model to make predictions for specific weather data stations. This is important not only to provide more detailed predictions, but also to be able to evaluate the work against specific ground truth data.
“The idea in many AI applications is to try to perform tasks as end-to-end as possible,” said Daniel Rotenberg, an atmospheric scientist at Brightband. “Adding to this model the ability to predict what, for example, the Denver airport weather station measures on an hourly basis only brings that predictive task closer to the core.”
The model can incorporate weather satellite data collected in real time on an hourly basis, so it can forecast more frequently. Feeding AI models with raw empirical observations rather than analysis produced by weather supercomputers promises more accurate predictions, but it remains technically difficult to make models work with unformatted data.
Google says WeatherNext 3 is the “first” AI model to incorporate raw observations directly into high-resolution global forecasts, while AI weather startup WindBorne said its model, WeatherMesh 6, will incorporate raw observations from its weather balloon fleet and other sources starting in late 2025. Asked about that, Google pointed out that its forecasts are of higher resolution around the world. In any case, both models still rely on national weather datasets to perform their predictions, so true direct data assimilation requires more work.
While LLMs are getting a lot of attention, the transformer revolution in meteorology is just as important. Weather agencies in Europe and the United States are already using AI models in their forecasting products, and their speed and low cost are expected to have an economic impact on poorer regions where the cost of high-quality sensors and supercomputers makes accurate predictions impossible.
Bill Gates recently cited AI-powered weather forecasting as a key benefit of the technology, saying better forecasts have improved crop yields in developing countries. Alet, the DeepMind researcher, said higher-resolution forecasts of wind, rain and cloud cover can help increase the reliability of renewable energy projects.
“At the end of the day, Google is about providing useful information to users, and I think a lot of what users are looking for is related to weather in some way,” Areto said.
If you make a purchase through links in our articles, we may earn a small commission. This does not affect editorial independence.
