Artificial intelligence “predicts” the weather more accurately than meteorologists – and issues a 10-day forecast in a minute
DeepMind’s GraphCast program outperformed a traditional European weather forecasting system by 99% in 12,000 measurements.
The AI model runs from a desktop computer and makes accurate forecasts in just a few minutes, while the most powerful traditional meteorological tools take hours to run, according to the Nature article.
“GraphCast is currently leading the race among artificial intelligence models,” says computer scientist Aditya Grover from the University of California, Los Angeles.
Today, the world’s weather forecasters use what is known as numerical weather prediction (NWP), an approach that uses mathematical models and data from buoys, satellites, and weather stations. Calculations show quite accurately how heat, air, and water vapor move in the atmosphere, but such instruments are quite expensive and energy-intensive.
Several well-known technology companies have already developed alternative AI tools, including DeepMind, computer chip maker NVIDIA, Chinese tech company Huawei, and a number of startups such as Atmo from Berkeley, California. Artificial intelligence is 1000-10,000 times faster than conventional NWP models, leaving more time for interpretation and communication of forecasts.
Huawei’s Pangu weather model is currently the most powerful competitor to the standard NWP system of the European Center for Medium-Range Weather Forecasts (ECMWF) in Reading, UK, which provides the world’s best weather forecasts up to 15 days in advance. At the same time, both tools seem to have already been outperformed by DeepMind’s GraphCast, which was trained on weather data from 1979 to 2017 to memorize the relationships between weather variables such as atmospheric pressure, wind, temperature, and humidity.
DeepMind found that GraphCast could also use global weather forecasts from 2018 to make 10-day forecasts in less than a minute, and they were more accurate than the High Resolution Forecasting System (HRES), a version of NWP that takes hours to forecast.
“In the troposphere, GraphCast outperforms HRES in more than 99% of the 12,000 measurements we’ve made,” says computer scientist Remy Lam of DeepMind in London.
At all levels of the atmosphere, GraphCast outperformed HRES in 90% of predictions. The model also effectively identified extreme weather events, such as the movement of tropical cyclones, severe cold or heat – in one particular example, the tool predicted Hurricane Lee’s approach to Long Island 10 days before it happened, while traditional weather forecasting technologies used by meteorologists at the time lagged behind.
Compared to Huawei’s model, GraphCast was better in 99% of predictions.
At the same time, AI models can improve certain types of weather forecasting that standard tools cannot handle, such as predicting the amount of precipitation that will fall on the ground within a few hours.
“Standard physical models are still needed to produce global weather estimates that are initially used to train AI models,” the researchers say. “We anticipate that it will take another two to five years before people can use machine learning predictions to make real-world decisions.”
Soon, GraphCast, or at least the basis of the artificial intelligence algorithm that provides the tool’s predictions, may appear in more mass services. According to Wired, Google may now be exploring how to integrate the model into its products.
