Google DeepMind WeatherNext 3 Offers 5 km Global Forecasts

Google DeepMind and Google Research have released WeatherNext 3, an advanced AI weather model designed to overcome limitations in resolution and initialization time inherent in traditional physics-based forecasting. This new model addresses the issue of coarse resolution, which previously hindered accurate local terrain predictions, and the delayed initialization tied to numerical weather prediction (NWP) analyses that typically arrive about six hours late. WeatherNext 3 directly utilizes a live global geostationary satellite mosaic as a primary input, enabling it to re-initialize forecasts every hour. The model's output resolution reaches down to 0.05°, which approximates to 5 kilometers, offering significantly more granular global weather predictions. A key innovation in its training process is the direct incorporation of raw weather station measurements, rather than relying solely on reanalysis grids. This approach allows WeatherNext 3 to capture and represent local atmospheric variations more faithfully.

According to Google AI, independent live evaluations conducted by Brightband have ranked WeatherNext 3 as the most accurate global weather model available to date. The model's architecture is based on a Functional Generative Network (FGN) mesh transformer, a probabilistic family previously introduced with WeatherNext 2, but scaled to support multi-resolution outputs. Its inputs include a live geostationary satellite mosaic alongside ECMWF HRES analysis data. The training dataset for WeatherNext 3 is comprehensive, drawing from ERA5/HRES-fc0, NASA's IMERG precipitation data, extensive weather station observations, and satellite mosaics. Unlike many AI forecasters that learn from NWP reanalysis, which tends to smooth out localized variations caused by geographical features like coastlines, valleys, and mountains, WeatherNext 3 employs dedicated observational heads. These heads are trained directly on raw station measurements, ensuring that its 0.05° outputs for temperature and dew point are calibrated to actual instrument readings rather than a model's generalized representation of the atmosphere.

While WeatherNext 3 represents a significant leap in AI weather forecasting, its deployment is currently partial. Forecast data is accessible through Google's BigQuery, Earth Engine, and Cloud Storage platforms, but requires an allowlist request for access. The model's weights are not open-sourced, and on-demand custom inference still utilizes the previous generation, WeatherNext 2. The development of WeatherNext 3 signifies a continued effort by Google DeepMind to leverage AI for improving weather prediction accuracy and timeliness, aiming to provide more reliable forecasts for a range of applications, from local planning to broader climate studies. The model's ability to integrate real-time satellite data and ground-truth observations marks a substantial advancement in the field of meteorological AI.