Medium-term weather forecasting is key to decision making in many social and economic fields and artificial intelligence is playing an increasingly important role. Google now presents a model capable of making a “faster and more accurate” forecast up to ten days in advance.
The model, based on machine learning, is called GraphCast and, according to its managers, from the company DeepMind, “significantly outperforms” traditional systems and also serves to provide earlier warnings of extreme weather phenomena.
It predicts dozens of ten-day weather variables for the entire planet in less than a minute, they say.
Details are published in the journal Science. GraphCast – open source – “takes an important step forward in artificial intelligence (AI) for weather forecasting, offering more accurate and efficient forecasts, and opening up avenues to support critical decision making,” say the authors, who include Spaniards Álvaro Sánchez González and Ferran Alet.
At present, forecasts are generally based on what is called numerical weather prediction (NWP), which starts with carefully defined physical equations that are then translated into algorithms run on supercomputers.
While this traditional method has been a “triumph of science and engineering,” designing the equations and algorithms is time-consuming, requires great expertise and expensive computers to achieve accurate forecasts, explains DeepMind on its blog, which argues that deep learning offers a different approach.
It involves, he adds, using data instead of physical equations to create a weather forecasting system.
GraphCast is trained on decades of historical weather data to learn a model of the cause-effect relationships that govern the evolution of weather on Earth, from the present to the future.
This AI covers the entire Earth’s surface and predicts five variables over it, including temperature, wind speed, wind direction and mean sea level pressure, and six atmospheric quantities at each of 37 altitude levels, including humidity.
Making 10-day forecasts with GraphCast takes less than a minute, compared to the computational hours of traditional approaches, according to DeepMind, which tested the model against HRES, the model developed by the European Centre for Weather Forecasting.
According to the company, GraphCast provided more accurate predictions on more than 90% of the 1,380 verified targets.
The model can also identify severe weather events earlier than other approaches and more accurately, say the authors led by Remi Lam. In addition to cyclones or temperature extremes, it is able to characterize atmospheric rivers, narrow regions of the atmosphere that transfer most of the water vapor out of the tropics.
The intensity of an atmospheric river can indicate whether it will bring beneficial rains or a deluge that will cause flooding.
“GraphCast is the most accurate 10-day global weather forecasting system and can predict extreme weather events farther into the future than was previously possible,” according to its developers, who say it will continue to evolve and improve.