Google DeepMind releases WeatherNext 3, an AI model that bests traditional forecasts

Scientists at Google DeepMind and Google Research have released a new artificial intelligence model for weather forecasting that the company says sees the atmosphere more clearly and predicts its behavior more frequently.

WeatherNext 3 is the latest advance in a shift toward deep learning in meteorology. It has already posted the highest scores on Operational WeatherBench, a benchmark for comparing AI forecasts built by the startup Brightband, across metrics like temperature, windspeed and humidity. According to Google, it beats models from Microsoft, Nvidia and the ECMWF itself, as well as traditional forecasts from the U.S. National Weather Service and the ECMWF.

“This is going to be the first time that some of the core variables feed and power a lot of the Google products,” Samier Merchant, a Google senior staff engineer, told TechCrunch. The company plans to integrate the model into Google Search, Google Maps and Gemini.

Traditional weather forecasts rely on supercomputers running physics-based equations. The ECMWF released more than 50 years of historical data in 2018, which allowed deep learning researchers to train models that predict weather faster and with comparable accuracy.

“Weather is chaotic, and so small differences really start to perturb massively… Machine learning targets the problem we are really solving, which is approximate noisy physics from incomplete information and finite compute, and so it learns patterns from a lot of data,” said Ferran Alet, a staff research scientist manager at DeepMind.

AI models still have known limitations: they tend to forecast over wide areas (15 to 25 square kilometres) that may be less useful for precise local predictions, they struggle with precipitation, and they continue to depend on formatted datasets from traditional models.

Why This Matters

For the first time, an AI weather model will directly power core Google products, giving billions of users potentially more accurate and frequent forecasts.
The model's superior performance on a standardized benchmark suggests AI-based forecasting is closing the gap with—and in some measures surpassing—traditional physics-based methods.
If AI models can address their remaining weaknesses (resolution, rain prediction), they could become the standard for operational weather forecasting, reducing reliance on expensive government supercomputers.

Background

For decades, weather forecasting has depended on numerical weather prediction (NWP) — solving complex physics equations on supercomputers. These systems are accurate but computationally expensive and slow. In 2018, the ECMWF released decades of reanalysis data, which researchers used to train deep-learning models that could make predictions in seconds. Google has been a leader in this space, with earlier models like GraphCast. WeatherNext 3 is its latest attempt to improve on both AI and traditional methods.

Key Perspectives

Google DeepMind: The model is a breakthrough that will deliver better weather information directly to users through everyday Google products. The company emphasises speed and accuracy gains from machine learning.

Traditional meteorological agencies (U.S. NWS, ECMWF): While their models have been surpassed on some metrics, they remain the gold standard for certain variables and spatial resolution. AI models still depend on their data and lack the physical interpretability that forecasters trust.

Critics/Skeptics: AI weather models still forecast over coarse grid cells (15–25 km²), often miss fine-scale phenomena like localized rain or thunderstorms, and rely on data from the very systems they aim to replace. They may also struggle in extreme or out-of-distribution weather events not well represented in training data.

What to Watch

How quickly the model is integrated into Google products and whether users notice a meaningful improvement in forecast accuracy.
Whether Google releases further benchmarks or peer-reviewed results on precipitation and high-resolution forecasting.
If other AI labs (Microsoft, Nvidia) respond with models that close the gap or leapfrog WeatherNext 3 on OperationalWeatherBench.

ZOTPAPER