Google DeepMind and Google Research have introduced WeatherNext 3, an AI weather forecasting model that provides global predictions at 0.05-degree (~5 km) resolution refreshed every hour. Unlike traditional AI weather forecasters that rely primarily on delayed numerical weather prediction analysis, WeatherNext 3 incorporates live geostationary satellite mosaics and trains directly on raw weather station measurements. Independent evaluations from Brightband rank it as the most accurate global weather model to date.
The model utilizes a Functional Generative Network mesh transformer architecture to output multi-resolution forecasting tiers, including surface temperature, wind, solar radiation, and precipitation. While forecast data is accessible via BigQuery, Earth Engine, and Cloud Storage, the model’s weights remain proprietary. The high-resolution outputs for wind and solar irradiance are targeted directly at operational buyers such as energy grid operators.
Why it matters
Enables operational buyers like energy grid operators to better forecast renewable output using hourly 5 km resolution data.
Demonstrates the shift in AI weather modeling from numerical reanalysis gridded data to direct training on live station and satellite observations.
Source: marktechpost.com
