Master GraphCast, GenCast, and FGN â the AI models that outperform traditional weather forecasting and are reshaping how we predict our planet's future.
Comprehensive resource for Research Engineers targeting DeepMind Sustainability
The evolution of ML-based weather prediction at DeepMind
First ML model to outperform ECMWF HRES on 90% of verification targets. Uses multi-scale icosahedral mesh with 16 message-passing layers.
Probabilistic forecasts via diffusion. Outperforms ECMWF ENS on 97.4% of targets. Superior tropical cyclone and extreme weather prediction.
Single-pass ensemble via 32D noise injection. Powers WeatherNext 2 in Google Search, Maps, and Pixel Weather.
The shared architecture pattern powering all three models
Lat-lon grid â Icosahedral mesh
721Ă1440 â 40,962 nodes
Message passing on multi-mesh
16-24 GNN/Transformer layers
Mesh â Output grid
Predicts Î from current state
Everything you need to master DeepMind Weather AI
Message passing, aggregation functions, and how GNNs capture spatial relationships on Earth's spherical surface.
Learn more âScore matching, denoising, and probabilistic generation. How GenCast creates ensemble forecasts.
Learn more âWhy spherical representation matters. Multi-scale hierarchies from M0 (12 nodes) to M6 (40,962 nodes).
Learn more âCRPS, RMSE, ACC, Brier scores. How to measure probabilistic forecast quality and ensemble calibration.
Learn more âTraining pipelines, TPU optimization, gradient checkpointing, and distributed systems for weather ML.
Learn more â12+ technical questions with detailed answer frameworks. System design and research scenarios.
Start prep âThe interactive study guide covers all three models, mathematical foundations, atmospheric science basics, and practical interview preparation.
Open Interactive Study Guide â