Research
Our state-of-the-art model provides 10-day weather forecasts with unprecedented accuracy in less than a minute
The weather affects us all, in ways big and small. It can dictate how we get dressed in the morning, provide us with green energy and, at worst, create storms that can destroy communities. In a world of increasingly extreme weather, fast and accurate forecasts have never been more important.
On a piece of paper published in Science, we present GraphCast, a state-of-the-art artificial intelligence model capable of making medium-range weather forecasts with unprecedented accuracy. GraphCast predicts weather conditions up to 10 days in advance more accurately and much faster than the industry gold standard weather simulation system – the High Resolution Forecast (HRES), produced by the European Center for Medium-Range Weather Forecasts (ECMWF).
GraphCast can also provide advance warnings of extreme weather events. It can predict cyclone tracks with great accuracy in the future, identify atmospheric rivers associated with flood risk and predict the onset of extreme temperatures. This capability has the potential to save lives through greater preparedness.
GraphCast takes a major step forward in artificial intelligence for weather forecasting, delivering more accurate and efficient forecasts and paving the way to support decision-making critical to the needs of our industries and societies. And from open source the model code for GraphCast, we enable scientists and meteorologists around the world to benefit billions of people in their daily lives. GraphCast is already used by weather services, including ECMWF, which is running a live experiment predictions of our model on its website.
A selection of GraphCast forecasts rolling over 10 days showing specific humidity at 700 ectopascals (about 3 km above the surface), surface temperature, and surface wind speed.
The challenge of global weather forecasting
Weather forecasting is one of the oldest and most challenging scientific endeavors. Medium-range forecasts are important to support key decision-making in everything from renewable energy to event logistics, but are difficult to do accurately and efficiently.
Forecasts are typically based on numerical weather prediction (NWP), which starts with carefully defined physics equations, which are then translated into computer algorithms run on supercomputers. Although this traditional approach has been a triumph of science and engineering, designing the equations and algorithms is time-consuming and requires deep expertise, as well as expensive computing resources to make accurate predictions.
Deep learning offers a different approach: using data instead of physical equations to build a weather forecasting system. GraphCast trains on decades of historical weather data to learn a model of the cause-and-effect relationships that govern how Earth’s weather evolves, from the present to the future.
Mainly, GraphCast and traditional approaches go hand in hand: we trained GraphCast on four decades of weather analysis data, from ECMWF’s ERA5 dataset. This crowd relies on historical weather observations, such as satellite imagery, radar, and weather stations that use a traditional NWP to “fill in the gaps” where observations are incomplete, to reconstruct a rich record of global historical weather.
GraphCast: An AI model for weather forecasting
GraphCast is a weather forecasting system based on machine learning and graph neural networks (GNNs), which are a particularly useful architecture for processing spatially structured data.
GraphCast makes predictions at the high resolution of 0.25 degrees longitude/latitude (28km x 28km at the equator). That’s more than a 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 at each of the 37 elevation levels, including the special humidity, wind speed and direction and temperature.
While training GraphCast was computationally intensive, the resulting prediction model is extremely efficient. Making 10-day forecasts with GraphCast takes less than a minute on a single Google TPU v4 machine. In comparison, a 10-day forecast using a conventional approach such as HRES can take hours of computation on a supercomputer with hundreds of machines.
In a comprehensive performance evaluation against the gold standard deterministic system, HRES, GraphCast provided more accurate predictions on more than 90% of the 1380 test variables and predicted lead times (see Scientific paper for details). When we restricted the evaluation to the troposphere, the 6-20 km region of the atmosphere closest to the Earth’s surface where accurate prediction is most important, our model outperformed HRES on 99.7% of the test variables for future weather conditions.
For inputs, GraphCast requires only two sets of data: the weather condition 6 hours ago and the current weather condition. The model then predicts the weather 6 hours into the future. This process can then be advanced in 6-hour increments to provide up-to-date forecasts up to 10 days in advance.
Better extreme weather warnings
Our analyzes revealed that GraphCast can also identify severe weather events earlier than traditional forecast models, even though it hasn’t been trained to look for them. This is a great example of how GraphCast could help with preparedness to save lives and reduce the impact of storms and extreme weather on communities.
By applying a simple cyclone detector directly to GraphCast forecasts, we could predict cyclone movement more accurately than the HRES model. In September, a live version of the publicly available GraphCast model, developed on the ECMWF website, accurately predicted about nine days in advance that Hurricane Lee would make landfall in Nova Scotia. In contrast, traditional forecasts had more variance in where and when landfall would occur, locking in only Nova Scotia about six days in advance.
GraphCast can also characterize atmospheric rivers – narrow regions of the atmosphere that carry most of the water vapor out of the tropics. The intensity of an atmospheric river can indicate whether it will bring beneficial rain or a deluge that will cause flooding. GraphCast forecasts can help characterize atmospheric rivers, which could help emergency response planning along with AI models for flood forecasting.
Finally, forecasting extreme temperatures is of increasing importance in our warming world. GraphCast can characterize when heat is set to rise above historical peak temperatures for any given location on Earth. This is especially useful for predicting heat waves, disruptive and dangerous events that are becoming more common.
Predicting severe events – how GraphCast and HRES compare.
Left: Cyclone tracking performances. As the lead time for forecasting cyclone motions increases, GraphCast maintains greater accuracy than HRES.
Right: Atmospheric river forecast. The forecast errors of GraphCast are significantly lower than those of HRES for the entire 10-day forecast
The future of artificial intelligence for weather
GraphCast is now the world’s most accurate 10-day global weather forecasting system and can predict extreme weather events further into the future than previously possible. As weather patterns evolve into a changing climate, GraphCast will evolve and improve as higher quality data becomes available.
To make AI weather forecasting more accessible, we have open source our model code. ECMWF already is experiment with GraphCast’s 10-day forecasts and we’re excited to see the possibilities it unlocks for researchers – from tailoring the model for specific weather phenomena to optimizing it for different parts of the world.
GraphCast joins other state-of-the-art weather forecasting systems from Google DeepMind and Google Research, including a regional Nowcasting model that produces forecasts up to 90 minutes ahead and MetNet-3a regional weather forecast model that already operates across the US and Europe and produces more accurate 24-hour forecasts than any other system.
The ground-breaking use of artificial intelligence in weather forecasting will benefit billions of people in their daily lives. But our broader research isn’t just about predicting the weather – it’s about understanding the broader patterns of our climate. By developing new tools and accelerating research, we hope that artificial intelligence can empower the global community to tackle our greatest environmental challenges.