GraphCast
DeepMind's ML weather-forecasting model that rivals numerical simulations for 10-day global weather prediction at a fraction of the compute cost.
What it does
GraphCast is a machine learning weather forecasting model from DeepMind that predicts global weather up to 10 days ahead by learning patterns from 40 years of historical weather data (ERA5 reanalysis). It matches or outperforms traditional numerical weather prediction (NWP) systems like ECMWF’s HRES on many metrics, while running in minutes rather than hours on a single GPU.
Access is via open research release (code and model weights available). Results are published in Science.
Best for
Climate scientists and atmospheric researchers who want a fast, competitive baseline for weather prediction — either for direct use in research or as a comparison benchmark against physical models. Also relevant for researchers studying predictability, model uncertainty, or data-driven vs. physics-based approaches to atmospheric science.
Pricing
Open research release. Code and model weights are publicly available. No commercial pricing — this is a research tool.
Strengths
- Matches ECMWF HRES performance on many standard metrics — the benchmark that professional meteorology has used for decades
- Dramatically faster than numerical simulation — runs in minutes on a single GPU vs. hours of supercomputer time
- Open source: weights, code, and training methodology published
- Demonstrates that purely data-driven approaches can rival physics-based simulation on a complex dynamical system — scientifically significant beyond the practical forecasting result
Limitations
- Optimized for global medium-range forecasting (10 days); less validated for regional or sub-seasonal prediction
- As a data-driven model, it may not generalize well to rare events not well-represented in the training period
- Does not incorporate the interpretable physical constraints that make numerical models useful for mechanistic understanding
- Requires familiarity with ML infrastructure to deploy — not a point-and-click interface
How it compares
| vs. | Key difference |
|---|---|
| ECMWF HRES (traditional NWP) | GraphCast matches NWP accuracy with orders-of-magnitude less compute; NWP provides physical interpretability |
| Pangu-Weather / FourCastNet | Competing ML weather models with similar capabilities; GraphCast is currently among the best-validated |
| Regional climate models | GraphCast is global and medium-range; regional models provide finer spatial resolution for specific areas |
Related content
Field Guide: Climate & Earth Science (coming soon)