GraphCast: AI model for faster and more accurate global weather forecasting

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

The weather affects us all, in big and small ways. You can dictate how to wear in the morning, provide us with green energy, and in the worst cases, create storms that can destroy societies. In an increasingly increased weather world, rapid and accurate predictions were not more important.

In a paper published in Science, we offer Graphast, a modern Amnesty International model that is able to make weather forecasts unprecedented. Graphcast predicts weather conditions up to 10 days more accurate and much faster than the HRES-HRES expectations, produced by the European Center for Medium-range weather forecast (ECMWF).

GraphCast can also make previous warnings of harsh weather events. It can predict hurricane paths with great accuracy in the future, determine the weather rivers related to the risks of flooding, and expect maximum temperatures to appear. This ability has the ability to save life with greater preparation.

GraphCast takes a big step forward in artificial intelligence to predict weather, providing more accurate and effective predictions, and opening paths to support critical decision -making of the needs of our industries and our societies. Through the open sources, the model symbol of Graphast, we enable scientists and predictors all over the world to benefit from billions of people in their daily lives. GraphCast is already used by weather agencies, including ECMWF, which runs a live experience of our model expectations on its website.

The challenge of predicting global weather

Weather prediction is one of the oldest and most challenging endeavors. Medium -range predictions are important to support major decisions across sectors, from renewable energy to logistical services, but it is difficult to do it accurately and efficiently.

The predictions usually depend on the numerical prediction of the weather (NWP), which begins with carefully defined physics equations, which are then translated into computer algorithms that work on supercomputs. Although this traditional approach was a victory for science and engineering, the design of equations and algorithms takes a long time and requires deep experience, as well as expensive calculations for making accurate predictions.

Deep learning provides a different approach: using data instead of material equations to create a weather forecast system. GraphCast is trained in contracts of historical weather data to learn a form of cause and result that governs how ground weather is evolved, from present to the future.

Critically, the GraphCast and the traditional approach goes alongside: We have trained GraphCast on four decades of weather analysis data, from ECMWF data set. This TRVE depends on historical weather notes such as satellite images, radar and weather stations using the traditional NWP “to fill the voids” where the notes are incomplete, to rebuild a rich record of global historical weather.

GRAPHCAST: Amnesty International weather forecast model

GRAPHCAST is a weather forecast system that depends on machine learning and neural networks (GNNS), which is a particularly useful structure for spatial data processing.

GraphCast highlights high -precision expectations from the length line/latitude 0.25 degrees (28 km x 28 km in the equator). This is more than a million network points covering the entire surface of the Earth. At each mesh point, the model predicts five Earth’s surface variables-including temperature, wind speed and direction, pressure at sea level-curtain of variables in the atmosphere in each of 37 levels of height, including specified moisture, wind speed, direction and temperature.

While the GraphCast training was an accurate intense, the resulting prediction model is very effective. It takes 10 days to make a 10 -day expectation using one minute on the one Google Tpu V4. For comparison, expectations for 10 days can use a traditional approach, such as HRES, take hours of account on the super computer with hundreds of machines.

In a comprehensive evaluation of performance against the golden impossible system, HRES, GraphCast provided more accurate predictions of more than 90 % than 1380 test variables and expected lead times (see our science paper to get details). When we limit ourselves to the evaluation on troposphere, a high area of ​​6-20 km of the closer atmosphere to the surface of the Earth where the exact prediction is the most important, our model outperformed 99.7 % of test variables for future weather.

Better warnings of harsh weather events

Our analyzes revealed that Graphast can also determine the harsh weather events early in traditional prediction models, although they are not trained to search for them. This is a major example of how to help GraphCast prepare to save lives and reduce the effect of storms and harsh weather on societies.

By applying a simple hurricane directly to Graphast’s expectations, we can predict the hurricane movement more accurately than the HRES model. In September, there was a direct version of the Graphast model available to the audience, which was published on the EcMWF, which was expected accurately nine days before Hurricane Lee would make land in Nova Scotia. On the contrary, traditional predictions had a greater contrast in the place where the landing would occur, and he was only imprisoned in Nova Scotia about six days ago.

GraphCast can also describe weather rivers – narrow areas of the atmosphere that transport most of the water vapor outside the tropical areas. The intensity of the atmosphere can indicate whether it will lead to useful rain or flood flood. GraphCast predictions can help describe weather rivers, which can help plan emergency responses with artificial intelligence models to predict the floods.

Finally, the prediction of extreme temperatures is of increasing importance in our world of warming. GraphCast can distinguish when the heat is set at the high historical temperatures of any specific location on the ground. This is especially useful in expecting heat waves, sabotage and dangerous events that have become increasingly common.

The future of the artificial intelligence of the weather

GraphCast is now the 10 -day global weather prediction system in the world, and it can predict harsh weather in the future than it was previously possible. With the development of weather patterns in a changing climate, GraphCast will develop and improve with high quality data.

To make AI energy prediction more easy, we have opened our model model code. ECMWF is already experimenting with GRAPHCAST predictions for 10 days, and we are excited to see the possibilities it opens to researchers-from allocating the model to the specific weather phenomena to improving it to different parts of the world.

GraphCast joins the Google DeepMind and Google Research systems, including a regional model now that results from up to 90 minutes, and Metnet-3, a regional prediction model throughout the United States and Europe produced 24 hours more accurate than any other system.

The pioneer in the use of artificial intelligence to predict weather will benefit billions of people in their daily lives. But our broader research is not only about expecting weather – it is related to understanding the broader patterns of our climate. By developing new tools and accelerating research, we hope that Amnesty International will enable the global community to face our greatest environmental challenges.