GraphCast: Learning skillful medium-range global weather forecasting: A Milestone in AI Weather Forecasting
In December 2022, Google DeepMind released GraphCast, a medium-range global weather forecasting method based on graph neural networks and machine learning. It is trained directly on reanalysis data and predicts hundreds of weather variables for 10 days at 0.25-degree resolution in under one minute. It outperforms the most accurate operational deterministic system on 90% of 1380 verification targets and performs better on extreme events such as tropical cyclones and atmospheric rivers.
GraphCast is the first to demonstrate that a purely data-driven ML method can significantly surpass traditional numerical weather prediction (NWP) in medium-range global weather forecasting, with extremely low computational cost. It changes the paradigm of weather forecasting relying on physical simulation and supercomputers, bringing revolutionary efficiency gains to meteorology, agriculture, energy, disaster warning, and other fields.
GraphCast adopts an encoder-processor-decoder architecture, gridding the Earth's surface into a multi-scale graph structure and learning spatiotemporal dynamics through message passing. Training data is ERA5 reanalysis (1979-2018), inputting the state of the past two time steps and autoregressively predicting the next 10 days. Evaluation uses anomaly correlation coefficient (ACC) and root mean square error (RMSE), outperforming ECMWF's high-resolution forecast (HRES) on 90% of verification targets. Key limitation: the model may generalize poorly to extreme events outside the training data distribution, and lack of physical constraints may lead to non-physical solutions.
GraphCast has a disruptive impact on the meteorological services industry: traditional NWP requires hours on supercomputers, while GraphCast completes in one minute on a single GPU, enabling high-frequency updates and personalized forecasts. Energy companies can optimize renewable energy scheduling, agriculture can plan precisely, and insurance companies can improve risk assessment. Meanwhile, institutions like ECMWF may accelerate research into hybrid methods combining ML and NWP.
It is recommended that weather service providers immediately evaluate GraphCast's substitution or complementary value in their own businesses, especially for small and medium enterprises with limited computing resources. Energy and agricultural companies can explore customized forecasting services based on GraphCast. Investment should focus on startups productizing GraphCast, as well as companies providing platforms for ML weather model training and deployment.
Subsequent work needs to verify GraphCast's stability in real-time operations and how to integrate with physical models to improve interpretability. In terms of safety, the model's reliability under rare extreme events needs assessment. Commercially, cloud service providers may launch GraphCast APIs, and weather data companies need to adjust their business models.