Google’s WeatherNext 3 sharpens hourly forecasts to 5 km globally
Google DeepMind and Google Research have released WeatherNext 3, an AI-driven weather model that pushes global forecasts to 5 km resolution and refreshes every hour—closing two long-standing gaps in numerical weather prediction.
The model ingests a live geostationary satellite mosaic and re-initializes 24 times per day, sidestepping the six-hour latency of traditional NWP analyses. Outputs span three tiers: 5 km station-trained 2 m temperature and dew point, 10 km surface wind and pressure, and 25 km atmospheric fields across 13 pressure levels. Independent live evaluations by Brightband rank it as the most accurate global model to date.
From smoothed grids to sharp observations
Most AI forecasters learn from NWP reanalysis grids that smooth away local terrain effects. WeatherNext 3 instead trains dedicated observational heads on raw station measurements, so its 5 km temperature and dew point outputs align directly with what instruments record. That approach shows clear gains: up to 60% improvement in CRPS against NASA’s IMERG satellite precipitation data, 30% against MRMS, and 10% against rain gauges at early lead times.
Built for operations, not just benchmarks
Forecast data is already accessible via BigQuery, Earth Engine and Cloud Storage after an allowlist request. However, the model weights remain closed and on-demand custom inference still uses WeatherNext 2. The package also delivers wind at 100 m, cloud layers and both solar irradiance components—variables grid operators need to forecast renewable generation against demand.
Why it matters
WeatherNext 3 demonstrates that AI can outpace physics-based models on resolution and cadence while staying grounded in real observations. For utilities and grid managers, hourly 5 km forecasts mean tighter integration of wind and solar into daily operations. The closed weights and limited deployment options, though, suggest the tech is still evolving toward full operational readiness rather than an immediate open alternative.
Source: MarkTechPost. AI-assisted editorial synthesis — TechnoExpress.