Google’s latest AI weather model gives you no excuse to forget your umbrella

Google DeepMind and Google Research have unveiled a new AI-driven weather forecasting model that offers clearer insights into atmospheric changes and more frequent, reliable predictions.

WeatherNext 3 represents a significant leap in meteorology, driven by deep learning. Google plans to integrate this model into Search, Google Maps, and Gemini, while making it accessible to researchers and users via its cloud platforms.

“This marks the first time core variables will directly power a wide range of Google products,” said Samier Merchant, a senior staff engineer at Google, in an interview with TechCrunch.

Tested on Operational WeatherBench—a benchmark created by startup Brightband to compare AI forecasts—WeatherNext 3 has demonstrated superior accuracy across key metrics such as temperature, wind speed, and humidity.

It outperforms other deep-learning models from Google, Microsoft, Nvidia, and the European Centre for Medium-Range Weather Forecasts (ECMWF), as well as traditional forecasts from the US National Weather Service and ECMWF.

Image Credits:Brightband / Brightband

Traditional weather forecasts rely on government supercomputers solving complex physical equations. While accurate, these systems are costly and slow. Since 2018, when the ECMWF released decades of historical data, researchers have used deep learning to create faster models with comparable accuracy.

“Weather is chaotic; small differences can lead to massive perturbations,” explained Ferran Alet, a staff research scientist manager at DeepMind. “Machine learning addresses the core challenge: approximating noisy physics using incomplete information and limited compute, allowing it to learn patterns from vast datasets.”

Previous AI forecasting models faced key limitations: they covered broad areas (15–25 square km) with limited practical use, struggled with rain prediction, and relied on pre-processed government data.

WeatherNext 3 addresses these issues. It offers a resolution of 5km, improves rain prediction accuracy by 60% compared to WeatherNext 2, and generates hourly forecasts instead of the standard six-hour intervals.

Image Credits:Google / Google

These improvements stem from specific design choices. WeatherNext 3 has 2.4 times more parameters than its predecessor, with decoder heads tailored to provide more actionable insights. Unlike standard forecasts that average data across 3D grids, DeepMind’s model can visualize specific events like cyclone paths.

Additionally, the model is trained to predict data for specific weather stations. This enhances prediction granularity and allows for evaluation against ground-truth data.

“The goal with many AI applications is to run tasks end-to-end,” said Daniel Rothenberg, an atmospheric scientist at Brightband. “By predicting specific station data, such as hourly conditions at Denver’s airport, we connect forecasting more closely to real-world utility.”

The model’s increased frequency is enabled by ingesting real-time satellite data hourly. Using raw empirical observations rather than supercomputer-generated analysis improves accuracy, though handling unformatted data remains technically challenging.

Google claims WeatherNext 3 is the first AI model to incorporate raw observations for high-resolution global forecasts. However, AI startup WindBorne states its WeatherMesh 6 model has used raw data from weather balloons since late 2025. Google notes its forecasts offer higher global resolution. Both models still rely on national datasets, indicating further work is needed for true direct data assimilation.

While large language models dominate headlines, the transformer revolution in meteorology is equally transformative. Weather agencies in Europe and the US are already using AI in their forecasts, offering speed and cost benefits that could make accurate predictions accessible to poorer regions.

Bill Gates recently highlighted AI weather forecasting as a key benefit, noting its potential to improve crop yields in developing nations. Ferran Alet added that higher-resolution forecasts of wind, rain, and cloud cover will help make renewable energy projects more reliable.

“Ultimately, Google aims to provide useful information to users, and weather is a central part of that,” Alet concluded.