Tests
Our cutting-edge model provides 10-day weather forecasts with unprecedented accuracy in less than one minute
The weather affects us all, in gigantic and diminutive ways. It can determine what we wear in the morning, provide us with green energy, and in the worst cases, cause storms that can destroy communities. In a world of increasingly extreme weather, quick and exact forecasts have never been more critical.
In paper published in Science., we present GraphCast, a state-of-the-art artificial intelligence model that is capable of producing medium-range weather forecasts with unprecedented accuracy. GraphCast predicts weather conditions up to 10 days in advance more accurately and significantly faster than the industry standard weather simulation system, High Resolution Forecast (HRES), developed by the European Center for Medium-Range Weather Forecasts (ECMWF).
GraphCast can also offer advance warnings about extreme weather events. It can predict cyclone tracks into the future with high accuracy, identifies atmospheric rivers associated with flood risk, and predicts the occurrence of extreme temperatures. This ability can save lives through greater preparedness.
GraphCast takes a significant step forward in artificial intelligence for weather prediction, offering more exact and competent forecasts and opening pathways to support decision-making critical to the needs of our industries and societies. And by open source model code for GraphCast, we enable scientists and forecasters around the world to benefit billions of people in their everyday lives. GraphCast is already in apply by weather agencies, including ECMWF, which is running a live experiment our model’s predictions on your website.
A selection of GraphCast forecasts covering 10 days, showing specific humidity of 700 hectopascals (approximately 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 demanding scientific endeavors. Medium-term forecasts are critical to support key decisions in a variety of sectors, from renewable energy to event logistics, but are challenging to execute accurately and efficiently.
Forecasts typically rely on a numerical weather forecast (NWP), which starts with carefully defined physical equations that are then translated into computer algorithms run on supercomputers. While this classic approach has been a triumph of science and engineering, designing equations and algorithms is time-consuming and requires deep expertise as well as costly computational resources to produce exact predictions.
Deep learning offers a different approach: using data instead of physical equations to create a weather forecast system. GraphCast uses decades of historical weather data to model the cause-and-effect relationships that govern the evolution of Earth’s weather, from the present to the future.
Most importantly, GraphCast and classic approaches go hand in hand: we trained GraphCast on four decades of weather reanalysis data from ECMWF’s ERA5 dataset. This collection relies on historical weather observations such as satellite imagery, radar, and weather stations that apply classic NWP to “fill in the gaps” where observations are incomplete, in order to reconstruct the world’s affluent record of historical weather.
GraphCast: an artificial intelligence model for weather prediction
GraphCast is a weather forecasting system based on machine learning and Graph Neural Networks (GNN), which provide a particularly useful architecture for processing spatially structured data.
GraphCast produces forecasts with a high resolution of 0.25 degrees longitude/latitude (28 km x 28 km at the equator). That’s over a million grid points covering the entire Earth’s surface. At each grid point, the model predicts five Earth surface variables – including temperature, wind speed and direction, and mean sea level pressure – and six atmospheric variables at each of 37 elevation levels, including specific humidity, wind speed and direction, and temperature.
Although GraphCast training was computationally intensive, the resulting prediction model is very competent. Creating 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 require hours of computation on a supercomputer with hundreds of machines.
As part of a comprehensive performance evaluation against the gold standard HRES deterministic system, GraphCast provided more exact predictions for over 90% of 1,380 test variables and predicted execution times (see our Science article for details). When we restricted the evaluation to the troposphere, the region of the atmosphere closest to the Earth’s surface at altitudes of 6–20 km where exact forecasting is most critical, our model outperformed HRES on 99.7% of the future weather test variables.
For input, GraphCast only needs two sets of data: the weather state from 6 hours ago and the current weather state. The model then predicts the weather for 6 hours in the future. This process can then continue at 6-hour intervals to provide cutting-edge forecasts up to 10 days in advance.
Better warnings about extreme weather events
Our analyzes showed that GraphCast can also identify severe weather events earlier than classic forecast models, even though it has not been trained to look for them. This is a great example of how GraphCast can assist prepare to save lives and reduce the impact of storms and extreme weather on communities.
By applying a uncomplicated cyclone tracker directly to GraphCast forecasts, we were able to predict cyclone movement more accurately than the HRES model. In September, a live version of our publicly available GraphCast model, deployed on the ECMWF website, accurately predicted nine days in advance that Hurricane Lee would make landfall in Nova Scotia. In contrast, classic forecasts had greater variability in the location and time of landfall and only covered Nova Scotia 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 produce beneficial rain or flooding. GraphCast forecasts can assist characterize atmospheric rivers, which can aid in emergency response planning AI models for flood forecasting.
Finally, predicting extreme temperatures is increasingly critical in a warming world. GraphCast can characterize when temperatures rise above historic highs anywhere on Earth. This is particularly useful in predicting heat waves and the nuisance and unsafe events that are becoming more common.
Predicting Major Events – Comparing GraphCast and HRES.
Left: Cyclone tracking performance. As the time to predict cyclone movements increases, GraphCast maintains greater accuracy than HRES.
Right: Atmospheric river forecasts. GraphCast prediction errors are significantly lower than HRES errors for all 10-day forecasts
The future of artificial intelligence for weather
GraphCast is currently the world’s most exact 10-day global weather forecast system, capable of predicting extreme weather events further into the future 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-based weather forecasting more accessible, we have introduced open source code of our model. ECMWF is here experimenting with GraphCast’s 10-day forecasts We’re excited about the possibilities it opens up for researchers – from adapting the model to specific weather phenomena to optimizing it for different parts of the world.
GraphCast combines other cutting-edge weather prediction systems from Google DeepMind and Google Research, including the regional Nowcasting model that generates forecasts up to 90 minutes in advance, and MetNet-3a regional weather forecasting model that already operates in the US and Europe and produces more exact 24-hour forecasts than any other system.
The pioneering apply of artificial intelligence in weather forecasting will benefit billions of people in their everyday lives. But our broader research isn’t just about predicting the weather – it’s about understanding the broader patterns of our climate. We hope that by developing fresh tools and accelerating research, artificial intelligence can empower the global community to address its greatest environmental challenges.