Recently, Google DeepMind and Google Research jointly released the new global weather AI model WeatherNext3. This model breaks through the inherent bottlenecks of traditional meteorological AI in local terrain resolution and initialization delay by directly accessing real-time geostationary satellite mosaics for hourly initialization and using raw meteorological station observation data for training, achieving a comprehensive upgrade from data architecture to forecast accuracy.

In terms of core architecture and input mechanism, WeatherNext3 adopts the same probabilistic functional generation network (FGN) grid Transformer architecture as WeatherNext2 and further expands to multi-resolution output. Its input directly connects to global real-time geostationary satellite mosaics and the HRES analysis data from the European Centre for Medium-Range Weather Forecasts (ECMWF). During the training phase, it extensively absorbs multi-source information including ERA5/HRES-fc0, NASA's IMERG, meteorological station measured data, and satellite mosaics. Traditional meteorological AI models usually rely on reanalysis grids from numerical weather prediction (NWP), which smooths out local climate changes caused by coastlines, valleys, and mountains. WeatherNext3, however, trains a dedicated output head using raw meteorological station observation data, making its 0.05-degree temperature and dew point forecasts more consistent with actual instrument records rather than relying solely on the model's own physical atmospheric representation.

In terms of resolution and forecast frequency (Cadence), WeatherNext3 can output three levels of forecast results with a single forward pass. The first level is 2-meter air temperature and dew point at a resolution of 0.05 degrees (approximately 5 kilometers) based on meteorological station training. The second level is grid surface data at a resolution of 0.1 degrees (approximately 10 kilometers), including 10-meter and 100-meter wind speed, pressure, sea surface temperature, cloud distribution, solar radiation, and one-hour precipitation. The third level is an atmospheric field covering 13 pressure levels at a resolution of 0.25 degrees (approximately 25 kilometers). Compared to the previous model's output of 0.25-degree fields every six hours, its accuracy has improved by about five times. In terms of update frequency, the model is initialized 24 times a day. The weather system cycles at UTC 00, 06, 12, and 18 can extend the forecast up to 15 days (360 hours) and include 64 ensemble members, while temporary hourly runs cover a 48-hour window. This hourly rolling refresh based on current satellite observations completely breaks the previous cycle limitations that relied on lagging six-hour traditional numerical forecast analysis, offering significant value for capturing rapidly developing convective weather.

In terms of precipitation simulation and support for clean energy variables, WeatherNext3 demonstrates strong business application potential. Precipitation forecasting has always been a historical pain point for global models, often producing unclear fields that miss storm boundaries. This model trains using multi-source data including ECMWF reanalysis, NASA's IMERG satellite inversion, and Google's own satellite radar precipitation reanalysis. In the early forecast stage, its Continuous Ranked Probability Score (CRPS) can improve by up to 60% compared to the IMERG benchmark and shows significant improvement compared to meteorological station rain gauges. For the new energy sector, the model can also accurately output 100-meter wind speed at the hub height of wind turbines, complete distributions of low, medium, and high clouds, and all components of solar irradiance. These key data combinations can directly empower power grid operators to accurately predict wind and solar energy generation output.

Currently, the forecast data of WeatherNext3 is open for application, and users can access it through BigQuery, Earth Engine, and Cloud Storage. However, the model weights are not open-sourced, and custom inference is still running the previous version of WeatherNext2. As this model is fully implemented, global weather forecasting is accelerating towards a new intelligent era of high precision and high timeliness.