Google DeepMind and Google Research Launch WeatherNext 3 with Hourly Multi-Resolution Forecasts

Google DeepMind and Google Research released WeatherNext 3 on September 3, 2026, introducing an artificial intelligence weather forecasting model that produces global predictions updated every hour at spatial resolutions down to five kilometers. The system generates global atmospheric forecasts on hourly cycles by combining live geostationary satellite imagery mosaics with numerical weather analysis, improving upon the update frequency and spatial granularity of previous data-driven and physical models.

Hourly Initialization and Rapid Forecast Cycles

Traditional numerical weather prediction systems solve fluid dynamic and thermodynamic differential equations on supercomputers. Because these physical simulations demand significant computational overhead, operational meteorological services typically limit their primary global forecast initialization cycles to intervals of six or twelve hours.

WeatherNext 3 relies on a neural network architecture based on Functional Generative Networks and mesh transformers. Once trained on decades of atmospheric records, running inference through the model requires a fraction of the computational time and energy demanded by raw physical simulations.

This efficiency allows the system to initialize new forecast runs 24 times a day:

• Primary cycles: Every six hours (00, 06, 12, and 18 UTC), the system produces extended global forecasts extending across a 15-day (360-hour) horizon using a 64-member ensemble.
• Interim hourly runs: During the intervening hours (such as 01–05 and 07–11 UTC), the model generates rapid 48-hour forward projections to track near-term atmospheric developments.

These hourly refreshes allow the model to adjust rapidly to moving storm systems, convective precipitation, and localized fronts without waiting for the next multi-hour numerical modeling cycle.

Multi-Resolution Output Architecture

Rather than computing all atmospheric metrics on a single uniform grid, WeatherNext 3 produces a multi-resolution output from a single forward pass. This tiered approach concentrates high-density computation on parameters heavily modified by surface friction and topography, while modeling upper-air dynamics on broader grids:

• 0.05-degree resolution (~5 kilometers): Applied to select surface variables, specifically 2-meter air temperature and dew point moisture fields.
• 0.1-degree resolution (~10 kilometers): Applied to standard gridded surface variables, including 10-meter surface winds, 100-meter turbine-height winds, surface pressure, and solar irradiance.
• 0.25-degree resolution (~25 kilometers): Applied to 3D atmospheric pressure-level variables aloft, such as upper-air geopotential height, temperature, specific humidity, and wind vectors across multiple vertical levels.

In atmospheric modeling, resolving surface parameters at five kilometers allows the system to reflect microclimatic variations caused by localized terrain features. Coarser global models operating at 25-kilometer intervals average terrain elevations across wide areas, often obscuring narrow river valleys, ridgelines, coastlines, and urban heat islands. By generating key surface moisture and thermal outputs at five kilometers, WeatherNext 3 captures local temperature inversions, sea breeze boundaries, and elevation-driven thermal shifts.

Live Inputs and Observational Training

WeatherNext 3 establishes its initial global atmospheric state through real-time telemetry and numerical analysis, while using ground-truth station records during model training to refine its output accuracy:

• Real-time initialization feeds: For each hourly run, the model directly ingests live global geostationary satellite imagery mosaics paired with operational numerical weather analysis from the European Centre for Medium-Range Weather Forecasts (ECMWF HRES). The live satellite mosaics provide immediate, un-lagged radiometric observations of cloud formations, moisture tracks, and convective development worldwide.
• Ground-truth station training: Rather than ingesting ground weather stations as real-time hourly feeds, WeatherNext 3 uses historical terrestrial weather station measurements as training data for dedicated observational decoder heads. This trains the model to correct for local elevation biases and output 5-kilometer surface metrics that closely reflect ground sensor readings.
• Multi-source precipitation training: To improve precipitation modeling—historically a challenging domain for artificial intelligence weather models—the architecture is trained against three distinct precipitation datasets: ECMWF reanalysis, NASA’s Integrated Multi-satellitE Retrievals for GPM (IMERG), and Google’s internal satellite-radar precipitation reanalysis.

Clean Energy Variables and Platform Integration

Alongside standard meteorological metrics, the model outputs specialized atmospheric layers designed for industrial and clean energy forecasting:

• 100-meter wind vectors: Directly predicts wind speeds at wind-turbine hub heights to assist grid operators with wind farm generation estimates.
• Solar radiation metrics: Predicts surface solar radiation downwards (SSRD) and direct solar irradiance (FDIR) alongside total and multi-layer cloud cover distributions to model photovoltaic output.

Google is deploying WeatherNext 3 data across consumer applications—including Google Search, Google Maps, and Gemini—as well as enterprise platforms such as Google Earth Engine, Google Cloud Storage, BigQuery, and the Google Maps Platform Weather API.