Google DeepMind announces AI 'GraphCast' that can predict the weather for 10 days in just 1 minute, more accurate than predictions that took several hours with a supercomputer

Google DeepMind announced weather prediction AI ' GraphCast '. To predict the weather for 10 days, conventional weather forecasting systems require a supercomputer to run for several hours, but GraphCast can predict the weather with accuracy exceeding that of traditional forecasting systems by running one machine for one minute. It is promoted as being predictable.

Learning skillful medium-range global weather forecasting | Science
https://www.science.org/doi/10.1126/science.adi2336

GraphCast: AI model for faster and more accurate global weather forecasting - Google DeepMind
https://deepmind.google/discover/blog/graphcast-ai-model-for-faster-and-more-accurate-global-weather-forecasting/

Conventional weather forecasting systems use a method called numerical weather forecasting (NWP), which calculates future weather based on observational data such as temperature and atmospheric pressure. However, NWP has problems such as ``expert knowledge is required to construct calculation algorithms'' and ``extremely costly calculation resources such as supercomputers are required to perform calculations''.

GraphCast, developed by Google DeepMind, learns the past 40 years of weather observation data included in the weather observation dataset ' ERA5 ' published by the European Center for Medium-Range Weather Forecasts (ECMWF) , and predicts future weather conditions. Predictions can be made quickly and with high accuracy. Predicting the weather using ECMWF's high-resolution 10-day atmospheric prediction model 'HRES' requires a supercomputer made up of hundreds of machines to operate for several hours, but GraphCast uses Google's AI specialized processor ' TPU v4 ” can be predicted just by running it for one minute.

GraphCast can predict the weather by dividing the earth into 0.25 degrees of latitude and 0.25 degrees of longitude. Below is a demo video that predicted the weather for 10 days using GraphCast.

A selection of GraphCast's predictions rolling across 10 days - YouTube


Google DeepMind compared the weather prediction accuracy of GraphCast and HRES using 1380 evaluation variables, and found that GraphCast was more accurate in 90.3% of the variables. Furthermore, when limiting predictions to altitudes of 6 to 20 km, which are important for weather forecasting, GraphCast was shown to be more accurate for 99.7% of variables.

The graph below shows the accuracy of predicting the hurricane's course using GraphCast (blue line) and HRES (black line). The horizontal axis shows the number of days counted from the prediction date, and the vertical axis shows the deviation from the actual course. Masu. Looking at the graph, you can see that the discrepancy between the predicted course and the actual course is smaller with GraphCast.

In addition, Google DeepMind has collaborated with ECMWF to publish a demo page where you can view a 10-day weather forecast using GraphCast.

ECMWF | Charts
https://charts.ecmwf.int/products/graphcast_medium-mslp-wind850

Additionally, the code used to learn GraphCast is published in the GitHub repository below.

GitHub - google-deepmind/graphcast
https://github.com/google-deepmind/graphcast

◆Forum now open
A forum related to this article has been set up on the GIGAZINE official Discord server . Anyone can write freely, so please feel free to comment! If you do not have a Discord account, please create one by referring to the article explaining how to create an account!

• Discord | 'Which service do you use for weather forecasts?' | GIGAZINE
https://discord.com/channels/1037961069903216680/1174276495523446814