Google has invested for many years in the forefront of machine learning weather research, demonstrated by our high-performing models in nowcasting, medium-range, and subseasonal predictions. These SOTA models often outperform traditional models in accuracy and speed, while utilizing a fraction of the computational resources.
WeatherNext publications
FGN
- New SOTA generative model for probabilistic weather forecasting, designed to generate ensembles with learned model-perturbation and by modeling both epistemic and aleatoric uncertainty.
- Preprint on arXiv in 2025: "Skillful joint probabilistic weather forecasting from marginals" by Ferran Alet et al.
- Published in Nature in 2026: "Operational tropical cyclone forecasting with AI" by Ferran Alet et al.
GenCast
- GNN/Transformer-based probabilistic (diffusion) model for ensemble forecasting. It is designed to model the conditional distribution of future weather states, generating an ensemble of possible future weather trajectories. Outperforms ECMWF ENS in 96% of targets.
- Published in Nature in late 2024: "Probabilistic weather forecasting with machine learning" by Ilan Price et al.
GraphCast
- Graph Neural Network (GNN)-based learned weather simulator that produces deterministic forecasts. Outperforms ECMWF HRES over 90% of tested cases and variables.
- Published in Science in late 2023: "Learning skillful medium-range global weather forecasting" by Remi Lam et al..
- Named a runner-up to Science's Breakthrough of the Year in 2023.
- Won the MacRobert Award in 2024.
Other research
While WeatherNext focuses on mid-range forecasting, Google also invests in other models. For short-range, high-resolution precipitation, MetNet predicts rain and snow up to 12 hours ahead, vital for Google's Nowcasting efforts. The latest model expands nowcasting globally using geostationary satellite data enabling support in data-sparse regions. For longer-range global simulations, NeuralGCM, a machine learning-powered General Circulation Model, simulates global weather faster than traditional models, enabling large forecast ensembles for long-range uncertainty.
Blogs
WeatherNext: AI model achieves breakthrough in forecasting cyclones
WeatherNext 2: Our most advanced weather forecasting model
GraphCast: AI model for faster and more accurate global weather forecasting
GenCast predicts weather and the risks of extreme conditions with state-of-the-art accuracy
How we're supporting better tropical cyclone prediction with AI
MetNet-3: A state-of-the-art neural weather model available in Google products
Demos
Weather Lab
Weather Lab is an interactive website developed by Google DeepMind and Google Research to demonstrate state-of-the-art AI weather forecasting models such as WeatherNext. Originally launched as an AI-powered tropical cyclone tracker, Weather Lab provides an accessible, map-based interface you can visualize high-resolution global forecast layers (including temperature, precipitation, wind, and pressure), analyze active and historical cyclone tracks, evaluate model performance against traditional numerical weather prediction (NWP) baselines like ECMWF, and inspect 15-day point forecasts with probabilistic uncertainty bands.
