AI Weather Forecasting System Beats Conventional Weather Forecasting Organizations

By: Carey Zhang

An AI weather forecasting experimental program running on desktop computers achieved great success in predicting the route of Hurricane Beryl, which struck North America on July 1st. The model was shown to be even more accurate than the top weather agencies with advanced mega computers, with data collected from planes, weather balloons and spacecraft.

Source: The New York Times
The graph above shows the prediction routes of Hurricane Beryl with the black line marking the actual route, blue lines showing the prediction routes by the ECMW and the NOAA, and the yellow line showing the result of the GraphCast AI weather forecasting system.

The result shown on much smaller computers operated by AI predicted using existing data in history is more accurate and much more closely resembles the actual path of the hurricane, compared to the top agencies with high-end NWP systems and mega computers.

GraphCast AI, an AI weather forecasting program designed in London by Google Deepmind, collected over 40 years of historical weather data for the AI to process, and then used the pattern to draw conclusions and make predictions.

Conventional weather forecasting systems, which use conditions and weather patterns on Earth to calculate the weather with equations that mix factors like wind, rain, clouds, temperature, air pressure and different zones. But the data can easily change, and one factor altered can impact the result greatly, which causes the forecasts to fail over days. To try to get the most accurate forecast, the mega computers have to process calculations and different factors involving fluid mechanics, making the process take hours even with advanced devices.

GraphCast AI on the other hand, will only identify the current stage of the atmosphere and “plug in” the data into existing models made from the patterns and data samples collected over the past decades. And as a result, the calculations are much more simple, making the system available on much smaller computers and work much faster. Microsoft Aurora Weather Forecasting system, for example, is said to be 5,000 times faster than average NWP weather forecasting systems running on mega computers.

The GraphCast AI can also be downloaded and is designed for daily use. It can predict the weather of the next 10 days within a minute, and is available for download through Github.