Fast and accurate! DeepMind releases AI weather forecast model, beating the world's most advanced system in 90% of indicators
For the first time, artificial intelligence (AI) has surpassed traditional forecasting methods in predicting global weather conditions for the next 10 days. Google DeepMind said in a paper published in Science on Tuesday that its AI model GraphCast AI "marks an inflection point in the field of weather forecasting."
Currently, the most advanced weather prediction system in the world is the one operated by the European Center for Medium-Range Weather Forecasts (ECMWF) (3 to 10 days in advance). But an extensive evaluation showed that GraphCast AI was more accurate than that system. It outperformed ECMWF's system in 90% of 1,380 indicators, including temperature, pressure, wind speed, wind direction, and humidity at different atmospheric levels.
Matthew Chantry, ECMWF machine learning coordinator, said progress on AI systems in meteorology is "much faster and more impressive than we expected two years ago." The agency has been using artificial intelligence models from Huawei, Nvidia and DeepMind, as well as its own comprehensive forecasting system for real-time weather forecasts.
Chantry agreed with DeepMind that its system is the most accurate. "We found that GraphCast AI consistently performed better than other machine learning models, and in many ways, it was more accurate than our own prediction system," he said.
GraphCast AI uses a machine learning architecture called a graph neural network to learn how weather systems develop and change around the world from more than 40 years of ECMWF data.
The input data for ECMWF forecasts is the global atmospheric state at the current time and 6 hours ago, which is collected by ECMWF from global weather observations. GraphCast AI runs on a Google TPU v4 cloud computer and can generate a 10-day weather forecast in one minute.
Unlike this data-derived "black box" approach, the traditional approach used by ECMWF and meteorological offices around the world, known as numerical weather prediction, uses supercomputers to calculate equations based on scientific knowledge of atmospheric physics, a process that takes several hours. energy-intensive process.
Chantry said: "GraphCast AI, once trained, is very cheap to operate. In terms of energy consumption, probably 1,000 times cheaper. This is a miraculous improvement."
Chantry said the next step for ECMWF will be to build its own artificial intelligence model and consider integrating it with numerical weather prediction systems. "We can inject an understanding of physics into these machine learning systems, which may look like black boxes," he said.
The Met Office last month announced a partnership with the UK's artificial intelligence research center, the Alan Turing Institute, to develop its own graphical neural network for weather forecasting and incorporate it into existing supercomputer infrastructure.
Simon Vosper, director of science at the Met Office, pointed out that climate change needs to be taken into account when forecasting. "It is unknown whether AI-based systems will be able to capture new extremes if they have only been 'trained' on previous weather conditions," he said.
But Vosper added: "Our goal is to leverage the benefits of artificial intelligence while using our traditional computer models based on atmospheric physics. We believe that in an era of rapid change, this convergence of technologies will will provide the most powerful and detailed weather forecast available.”