WeatherNext: How Google DeepMind's AI Models Rewrote Global Weather Forecasting

Key points

• Resolution and scope: WeatherNext 2 runs at 0.25° (~30 km) global resolution for medium-range forecasting (10–15 days), with additional 100 m wind and tropical cyclone track prediction. The cyclone variant FNV3 was deployed operationally during the 2025 Atlantic hurricane season.
• Training data: ERA5 reanalysis combined with ECMWF HRES operational data, with a training cutoff in 2024.
• Model family evolution: WeatherNext Graph (GraphCast, deterministic graph neural network) → WeatherNext Gen (GenCast, diffusion-based ensemble) → WeatherNext 2 (unified deterministic + probabilistic, atmospheric + cyclone). The trajectory moves from single deterministic answers to probabilistic ensembles and from single-task to multi-task unification.
• Probabilistic design: WN2's technical report (arXiv:2506.10772) is titled "Skillful joint probabilistic weather forecasting from marginals." Rather than sampling the joint distribution directly, the model estimates per-variable marginals and combines them, reframing forecast mathematics around classical probability principles.
• Data access infrastructure: Weights total multiple gigabytes across four files, so DeepMind provides multiple alternative channels for forecast data:
• Google Cloud (Earth Engine, BigQuery, Vertex AI)
• WeatherLab (cyclone tracks)
• OpenMeteo (API and interactive builder)
This makes WeatherNext a full data-distribution platform, not just a model repository.
• Operational impact: AI forecasting collapses the inference stage from hours (traditional NWP on supercomputers) to seconds on a single GPU. Ensemble forecasting becomes cheap (one diffusion pass vs. 50 model runs), extreme-event warnings speed up from hours to minutes, and access expands to regions without HPC infrastructure.
• Lessons for AI research:
1. Domain expertise is non-substitutable: the DeepMind team included meteorologists who understood ERA5, HRES, and data assimilation. 2. Infrastructure matters as much as the model: Google Cloud and OpenMeteo channels are core to WN2's real-world value. 3. Determinism-to-probability is a recurring AI trajectory: GraphCast → GenCast → WN2 mirrors patterns seen across other ML domains.
• Open-source limitation: Model weights and code are released, but training data and compute pipelines are not, so WN2 can be used but not reproduced. DeepMind's mitigation is open data-access channels, though this differs from full reproducibility, raising long-term questions about concentration of AI-for-science capability inside a few large labs.

Source

• Repository: github.com/google-deepmind/weathernext
• Technical report: arXiv:2506.10772