GraphCast: AI models for faster, more accurate global weather forecasting

Our state-of-the-art model provides 10-day weather forecasts with unprecedented accuracy in less than a minute

Weather affects all of us in big or small ways. It influences how we dress in the morning, provides green energy, and in the worst cases can cause storms that destroy communities. In a world of increasingly extreme weather events, fast and accurate forecasts are more important than ever.

In a paper published in Science, we introduce GraphCast, a cutting-edge AI model that can make medium-range weather forecasts with unprecedented accuracy. GraphCast predicts weather conditions for up to 10 days more accurately and faster than the industry’s leading weather simulation system, the High Resolution Forecast (HRES) produced by the European Center for Medium-Range Weather Forecasts (ECMWF).

GraphCast can also provide early warning of extreme weather events. It can predict cyclone paths far into the future with high accuracy, identify atmospheric rivers associated with flood risk, and predict the onset of extreme temperatures. This ability has the potential to save lives by increasing preparedness.

GraphCast is a major advance in AI-powered weather forecasting, paving the way to providing more accurate and efficient forecasts and supporting decision-making critical to industry and societal needs. Open sourcing GraphCast’s model code will also enable scientists and forecasters around the world to benefit billions of people in their daily lives. GraphCast is already used by weather agencies, including ECMWF, which runs live experiments of model predictions on its website.

Challenge to global weather forecasting

Weather forecasting is one of the oldest and most challenging scientific endeavors. Mid-range forecasting is critical to supporting critical decisions across sectors, from renewable energy to event logistics, but is difficult to do accurately and efficiently.

Forecasting typically relies on numerical weather prediction (NWP). NWPs start with carefully defined physical equations that are translated into computer algorithms that run on supercomputers. Although this traditional approach is a triumph of science and engineering, designing the equations and algorithms is time-consuming, requires deep expertise, and expensive computational resources to make accurate predictions.

Deep learning offers a different approach, using data instead of physical equations to create weather forecasting systems. GraphCast is trained on decades of historical weather data to learn a model of the cause-and-effect relationships that determine how Earth’s weather evolves from now into the future.

Importantly, GraphCast and traditional approaches work together. We trained GraphCast on 40 years of weather reanalysis data from ECMWF’s ERA5 dataset. This treasure trove is based on historical weather observations such as satellite imagery, radar, and weather stations, and uses traditional NWP to “fill in the gaps” in observations to provide a comprehensive overview of global historical weather. We are reconstructing a wealth of records.

GraphCast: AI models for weather prediction

GraphCast is a weather forecasting system based on graph neural networks (GNNs), an architecture particularly useful for machine learning and processing spatially structured data.

GraphCast performs forecasts at a high resolution of 0.25 degrees longitude/latitude (equator 28 km x 28 km). This equates to more than 1 million grid points covering the entire surface of the Earth. At each grid point, the model predicts five land surface variables, including temperature, wind speed and direction, and mean sea level pressure, and six atmospheric variables, including specific humidity, wind speed and direction, at each of 37 levels of altitude. temperature.

Although training GraphCast was computationally intensive, the resulting predictive model is highly efficient. Making a 10-day forecast using GraphCast takes less than a minute on a single Google TPU v4 machine. For comparison, a 10-day forecast using traditional approaches such as HRES can take several hours of calculation on a supercomputer with hundreds of machines.

In a comprehensive performance evaluation against the gold standard deterministic system, HRES, GraphCast provided more accurate predictions for more than 90% of 1,380 test variables and predicted lead times (for more information, see Science ). When we limit our evaluation to the troposphere, the region of the atmosphere closest to the Earth’s surface between 6 and 20 kilometers in height, where accurate predictions are most important, our model yields HRES for 99.7% of the test variables for future weather. I surpassed it.

Better warning for extreme weather

Our analysis reveals that GraphCast can identify severe weather faster than traditional predictive models, even though it is not trained to look for severe weather. This is a great example of how GraphCast can help save lives and prepare communities to reduce the impact of storms and extreme weather events.

A simple cyclone tracker applied directly to GraphCast forecasts can predict cyclone movement more accurately than the HRES model. In September, a live version of the publicly available GraphCast model deployed on the ECMWF website accurately predicted Hurricane Lee to make landfall in Nova Scotia about nine days in advance. In contrast, traditional forecasts were more variable about when and where landfall would occur, limiting it to Nova Scotia only about six days in advance.

GraphCast can also characterize atmospheric rivers, the small regions of the atmosphere that move most of the water vapor out of the tropics. The strength of a river in the atmosphere can indicate whether it brings beneficial rain or a deluge that causes flooding. GraphCast’s predictions can help characterize atmospheric rivers and, when combined with AI models to predict floods, could help plan emergency responses.

Finally, in a warming world, predicting temperature extremes is becoming increasingly important. GraphCast can characterize times at any location on Earth when heat exceeds historic high temperatures. This is especially useful when predicting heat waves and other destructive and dangerous events that are becoming increasingly common.

The future of AI in weather

GraphCast is currently the world’s most accurate 10-day global weather forecasting system, able to predict extreme weather events further out than previously possible. As weather patterns evolve in a changing climate, GraphCast will evolve and improve as higher quality data becomes available.

To make AI-powered weather forecasts more accessible, we have open sourced the model code. ECMWF is already experimenting with GraphCast’s 10-day forecast to see the possibilities it unlocks for researchers, from tuning models to specific weather phenomena to optimizing them for different regions of the world. I’m excited about it.

GraphCast builds on other cutting-edge weather prediction systems from Google DeepMind and Google Research, including regional nowcasting models that generate forecasts up to 90 minutes in advance, and MetNet-3, a regional weather prediction model already in operation across the United States. will join. And Europe produces more accurate 24-hour forecasts than any other system.

Pioneering the use of AI in weather forecasting will benefit billions of people in their daily lives. But our extensive research is not just about predicting the weather, but about understanding broader patterns in climate. By developing new tools and accelerating research, we hope that AI will empower the global community to tackle our biggest environmental challenges.