Google DeepMind's WeatherNext 3 Trains on Weather Station Observations to Deliver 5 km Global Forecasts, Refreshed Every Hour
WeatherNext 3 replaces the traditional 6-hour NWP analysis lag with hourly initialization from live geostationary satellite mosaics, enabling faster adaptation to rapidly evolving weather systems The model delivers multi-resolution outputs in a single forward pass: 0.05° (~5 km) station-calibrated temperature/dew point, 0.1° gridded surface variables, and 0.25° atmospheric fields across 13 pressure levels Training directly on raw weather station observations rather than smoothed reanalysis grids
Analysis
TL;DR
- WeatherNext 3 replaces the traditional 6-hour NWP analysis lag with hourly initialization from live geostationary satellite mosaics, enabling faster adaptation to rapidly evolving weather systems
- The model delivers multi-resolution outputs in a single forward pass: 0.05° (~5 km) station-calibrated temperature/dew point, 0.1° gridded surface variables, and 0.25° atmospheric fields across 13 pressure levels
- Training directly on raw weather station observations rather than smoothed reanalysis grids allows the model to capture local terrain effects (coastlines, valleys, mountains) that conventional AI forecasters miss
- Precipitation forecasting shows CRPS improvements of up to 60% against IMERG and 30% against MRMS at early lead times, addressing a historic weakness of global AI weather models
- The model is accessible via BigQuery, Earth Engine, and Cloud Storage on an allowlist basis, but weights remain proprietary and custom inference still defaults to WeatherNext 2
Why It Matters
WeatherNext 3 represents a significant step toward operational deployment of AI weather models by solving two persistent bottlenecks: coarse resolution and stale initialization. For AI practitioners and energy sector professionals, the direct inclusion of turbine-height wind, solar irradiance, and cloud layer outputs signals that Google is targeting real-world commercial buyers—not just academic benchmarks. The hourly cadence and 5 km resolution make it competitive with regional numerical weather prediction systems, potentially reshaping how grid operators and disaster response teams access global forecasts.
Technical Details
- Architecture: Functional Generative Network (FGN) mesh transformer, the same probabilistic family as WeatherNext 2, scaled to multi-resolution output. A single forward pass generates all three resolution tiers simultaneously.
- Inputs: Live global geostationary satellite mosaic combined with ECMWF HRES analysis, replacing the traditional reliance on lagged NWP reanalysis for initialization.
- Training data: ERA5/HRES-fc0 reanalysis, NASA's IMERG satellite precipitation retrievals, Google's own satellite-radar precipitation reanalysis, raw weather station observations, and satellite mosaics. Dedicated observational heads are trained directly on station measurements for temperature and dew point calibration.
- Resolution tiers: 0.05° (~5 km) for 2 m temperature and dew point (station-trained); 0.1° (~10 km) for 10 m/100 m wind, pressure, sea surface temperature, cloud layers, solar radiation, and 1-hour precipitation; 0.25° (~25 km) for atmospheric fields across 13 pressure levels.
- Forecast cadence and ensembles: 24 initializations per day. Synoptic cycles at 00/06/12/18 UTC run out to 15 days (360 hours) with 64 ensemble members; interim hourly runs cover 48 hours.
- Evaluation: Independent live evaluations from Brightband rank it as the most accurate global weather model to date. Google reports up to 50% reduction in Brier score and CRPS versus NWP baselines against IMERG.
Industry Insight
- The shift from reanalysis-only training to raw station observation calibration sets a new standard for AI weather models—future systems that ignore local observational data will struggle to match terrain-aware accuracy, particularly in complex topography.
- Google's focus on clean energy variables (100 m wind, solar irradiance, cloud distributions) indicates a strategic push into the renewable energy forecasting market; grid operators should evaluate WeatherNext 3 access through BigQuery or Earth Engine as a potential replacement for legacy NWP-dependent scheduling tools.
- The proprietary access model (allowlist-only data, no open weights) reinforces the trend of major AI weather labs treating forecasting as a differentiated cloud service rather than an open research artifact—practitioners should plan for API-based or query-based integration rather than self-hosted deployment.
Disclaimer: The above content is generated by AI and is for reference only.