Weather Forecasting AI Models

4 posts · Nov 2025 – Sep 2026

What this trail covers

WeatherNext 3 is introduced as the most advanced and accurate global weather model to date, according to independent live evaluations. It learns directly from real-time observations, enabling timely and more localized predictions. The model generates hourly forecasts at multiple spatial resolutions (down to 5km for surface variables), maintaining physical consistency. It ingests live, global geostationary satellite data and sparse weather station observation data, bypassing the six-hour data lag

All DeepMind trails

Sep 2026Introducing WeatherNext 3, our most advanced and accurate global weather AI modelWeatherNext 3 advances weather forecasting by ingesting raw satellite data and sparse weather station observations directly, enabling hourly forecasts at up to 5km resolution. This contrasts with WeatherNext 2's 6-hour increments and 25km grid, and traditional NWP models' data lag. The model achieves breakthrough accuracy in precipitation forecasting with up to 60% CRPS improvement against IMERG. It also introduces specific forecasts for renewable energy production (100m wind speed, cloud cover, sun radiation). The post details the technical shift from NWP-trained models to real-time observation ingestion and the resulting improvements in spatial resolution, temporal frequency, and accuracy, particularly for localized and rapidly changing weather events.
Aug 2026AI model achieves breakthrough in forecasting cyclonesThis post details the advancements in the WeatherNext AI model for cyclone forecasting, achieving state-of-the-art accuracy and providing an extra day of predictive accuracy. Key technical contributions include: the model's ability to bridge the gap between global atmospheric currents and localized thermodynamic processes with a single AI model; co-training on global weather dynamics and historical cyclone observations (IBTrACS database); the use of Functional Generative Networks (FGNs) for efficient ensemble prediction generation, enabling the capture of rare but consequential scenarios; and achieving high accuracy with a coarser input resolution (28x28km) than traditional models, which remains an open research question. The post also announces the open-sourcing of WeatherNext 2, WeatherNext Cyclones, and WeatherNext 2-mini models.
May 2026AI breakthrough: WeatherNext predicts Hurricane MelissaThis post details the successful application of the WeatherNext AI model in predicting Hurricane Melissa's rapid intensification and landfall five days in advance with high confidence. It highlights the historical trade-off between global and local models for track and intensity prediction, which WeatherNext bridges by training on decades of global weather patterns and specialized extreme tropical cyclone datasets. The model's ensemble approach of running 50 "what-if" scenarios is also emphasized as a key factor in providing experts with a broader range of possibilities. The post also mentions the integration of WeatherNext into forecasts on Search and ongoing collaborations with national meteorological agencies worldwide.
Nov 2025WeatherNext 2: Our most advanced weather forecasting modelWeatherNext 2 is introduced, an advanced and efficient forecasting model that generates forecasts 8x faster with resolution up to 1-hour. It utilizes a new Functional Generative Network (FGN) approach that injects noise into the model architecture to ensure physically realistic and interconnected forecasts. This model surpasses previous state-of-the-art on 99.9% of variables and lead times, enabling more useful and accurate predictions. The forecast data is now available in Earth Engine and BigQuery, with an early access program on Google Cloud's Vertex AI platform.
Advertisement