Google DeepMind’s WeatherNext 3 Trains on Weather Station Observations to Deliver 5 km Global Forecasts, Refreshed Every Hour
AI climate fashions have spent three years closing the hole with physics-based forecasting, however two issues stayed open: decision too coarse for native terrain, and initialization tied to numerical climate prediction (NWP) evaluation that arrives about six hours late. WeatherNext 3, launched by Google DeepMind and Google Research, assaults each. It takes a stay world geostationary satellite tv for pc mosaic as a direct mannequin enter, re-initializes each hour, and emits forecasts down to 0.05° (~5 km) whereas coaching in opposition to uncooked climate station measurements quite than reanalysis grids alone. According to Google AI, unbiased stay evaluations from Brightband rank it as essentially the most correct world climate mannequin to date.
Is it deployable? Partially. Forecast knowledge is offered now by way of BigQuery, Earth Engine and Cloud Storage after an allowlist request, however WeatherSubsequent 3 weights aren’t open supply and on-demand customized inference nonetheless runs WeatherSubsequent 2.
Architecture and inputs
WeatherSubsequent 3 is a Functional Generative Network (FGN) mesh transformer, the identical probabilistic household launched with WeatherNext 2, scaled to multi-resolution output. Inputs are a stay geostationary satellite tv for pc mosaic plus ECMWF HRES evaluation. Training attracts on ERA5/HRES-fc0, NASA’s IMERG, station observations and satellite tv for pc mosaics.
Most AI forecasters study from NWP reanalysis, which smooths away the native variation that coastlines, valleys and mountains truly produce. WeatherSubsequent 3 trains devoted observational heads straight on uncooked station measurements, so its 0.05° temperature and dew level outputs are calibrated to what devices document quite than to a mannequin’s illustration of the environment.