Our most modern model provides 10-day weather forecasts with unprecedented accuracy in a smaller minute
The weather affects us all, in a large and small way. He can dictate how we dress in the morning, provide us with green energy, and in the worst cases they create storms that can destroy communities. In the world of more and more extreme weather, fast and accurate forecasts have never been more important.
In paper Published in scienceWe present Graphcast, the most modern model of artificial intelligence, which is able to make medium -range weather forecasts with unprecedented accuracy. Graphcast provides weather conditions to 10-day in advance more precisely and much faster than the high-resolution-resolution sector (Hres), produced by the European Center for Medium Survey Forecasts (ECMWF).
Graphcast can also offer earlier warnings about extreme weather events. It can predict the traces of cyclones with high accuracy into the future, identify atmospheric rivers associated with flood risk and predicts the beginning of extreme temperatures. This skill can save lives thanks to greater readiness.
Graphcast takes a significant step forward in artificial intelligence to predict the weather, offering more accurate and efficient forecasts and opening paths to support decisions for the needs of our industries and societies. And through Open the model of the Graphcast model, We enable scientists and forecasts around the world to the benefit of billions of people in their daily lives. Graphcast is already used by weather agencies, including ECMWF, which conducts live experiment Forecasts of our model on your website.
The choice of Graphcast forecasts takes place for 10 days showing specific humidity in 700 hectopascals (about 3 km above the surface), surface temperature and surface wind speed.
The challenge of the global weather forecast
Weather forecasting is one of the oldest and most difficult – zones. Medium range forecasts are important for supporting key decision -making in various sectors, from renewable energy to the logistics of events, but they are difficult to do carefully and efficiently.
Forecasts are usually based on numerical weather forecasts (NWP), which begins with carefully defined physics equations, which are then translated into computer algorithms activated on supercomputers. Although this traditional approach was a triumph of science and engineering, the design of equations and algorithms is time consuming and requires deep specialist knowledge, as well as expensive calculation resources to make accurate forecasts.
Deep learning offers a different approach: the use of data instead of physical equations to create a weather forecast system. Graphcast is trained in decades of historical weather data to learn about the model of reasons and effects that govern the evolution of earth weather, from the present to the future.
Most importantly, GraphCast and traditional approaches go hand in hand: we trained Graphcast for four decades of weather reanalysis, from the Era5 ERAMWF data set. This change is based on historical weather observations, such as satellite images, radar and weather stations using traditional NWP to “fill empty places”, in which observations are incomplete to reconstruct the rich record of global historical weather.
Graphcast: AI model for weather forecasting
Graphcast is a weather forecasting system based on machine learning and neural networks (GNN), which are particularly useful architecture for spatially structural data processing.
Graphcast makes forecasts with a high resolution of 0.25 degrees of latitude/latitude (28 km x 28 km on the equator). These are over a million mesh points covering the entire surface of the earth. At each mesh point, the model provides five variables of the earthly surfaces-in this temperature, speed and direction of wind as well as medium pressure at sea level and six atmospheric variables at each of the 37 levels of height, including specific humidity, wind speed and direction and temperature.
While the GraphCast training was intensively computing, the resulting forecasting model is highly efficient. Making 10-day forecasts from Graphcast lasts less than a minute on one Google TPU V4 computer. For comparison, a 10-day forecast using a conventional approach, such as Hres, can take the time of calculations in supercomputers with hundreds of machines.
In a comprehensive assessment of performance in relation to the determinist system of gold, Hres, Graphcast, provided more accurate forecasts for over 90% of 1380 test variables and forecasting time (see ours Scientific paper Detailed information). When we limited the rating to the troposphere, 6-20 kilometers of the atmosphere of the closest surface of the Earth, in which the most important is accurate forecasting, our model exceeded HRE to 99.7% of test variables for future weather.
In the case of input data, Graphcast requires only two sets of data: weather state 6 hours ago and the current weather. The model then predicts the weather of 6 hours in the future. This process can then be transfused in 6-hour increases to provide the latest forecasts up to 10 days earlier.
Better warnings about extreme weather events
Our analyzes have revealed that Graphcast can also identify difficult weather events earlier than traditional forecasting models, although they have not been trained to look for them. This is a perfect example of how Graphcast can help in readiness to save life and reduce the impact of storms and extreme weather on the community.
Using simple cyclone tracking directly to Graphcast forecasts, we can predict the cyclone movement more precisely than the Hres model. In September, the live version of our publicly available GRAPHCAST model, implemented on the ECMWF website, thoroughly expected about nine days earlier that Hurricane Lee made a landing in Nova Scotland. On the other hand, traditional forecasts had greater variability, where and when the land would occur, and closed only in Nova Scotland from about six days earlier.
Graphcast can also characterize atmospheric rivers – narrow regions of the atmosphere that transmit most water vapor outside the tropics. The intensity of the atmospheric river may indicate whether it will bring a favorable rain or a flood that causes a flood. Graphcast forecasts can help characterize atmospheric rivers, which can help plan to plan your emergency reactions along with AI models for flood forecasting.
Finally, predicting extreme temperatures is becoming more and more important in our warming world. Graphcast can characterize when the heat is set to historical highest temperatures for anywhere on earth. This is especially useful in predicting the heat waves, destructive and dangerous events that are becoming more and more common.
Forecasting about heavy events – as Graphcast and Hres compare.
On the left: Cyclone tracking performances. As the movement of the movement of the cyclone, Graphcast increases more accuracy than Hres.
On the right: atmospheric river forecast. Graphcast forecast errors are much lower than Hres for all their 10-day forecasts
The future of artificial intelligence for the weather
Graphcast is currently the most accurate 10-day global weather forecasting system in the world and can predict extreme weather events in the future than possible. When weather patterns evolve in a changing atmosphere, Graphcast evolves and improves along with available higher quality data.
To forecast the weather powered by artificial intelligence, we are more available, we have Open our model code. ECMWF is already there Experimenting with 10-day forecasts Graphcast And we are glad that we see opportunities that he unlocks for scientists – from adapting the model for individual weather phenomena to optimization for different parts of the world.
Graphcast joins other most modern weather forecasting systems with Google Deepmind and Google Research, including the regional NovCasting model, which creates forecasts up to 90 minutes ahead and Metnet-3The regional weather forecasting model already operating in the USA and Europe, which creates a more accurate 24-hour forecasts than any other system.
The pioneering of the use of artificial intelligence in weather forecasting will bring benefits to billions of people in their daily lives. But our wider research is not only predict the weather – it is about understanding the wider patterns of our climate. By developing new tools and accelerating research, we hope that AI can enable the global community to face our greatest environmental challenges.
We are grateful to Matthew Chantry, Peter Dueben and Linus Magnusson from ECMWF, for their help and feedback. We also want to thank Svetlan Grant, Jon Small for providing legal support. This work was done thanks to the contribution of co-authors: Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Ferran Alet, Suman Ravuri, Timo Ewalds, Zach Eaton-Rosen, Weihua Hu, Alexander Merose, Stephan Hoyer, George Holland, Jacka, Jacko, Jacko, Jacko, Jacko, Jacko, Jacko, Jacko, Jacko, Jacko, Jacko, Jacko. Stott, Alexander Pritzel, Shakir Mohamed and Peter Battaglia.
*This is the version of the author's work. This is published here with the consent of AAA for personal use, not for redistribution. The final version was published in Science DOI: 10.1126/science.adi2336.
