Stay informed with free updates
Simply log in to Artificial intelligence myFT summary — delivered straight to your inbox.
Google DeepMind presented an artificial intelligence weather forecasting model that outperforms traditional methods on forecasts up to 15 days and is better at predicting extreme events.
The tool, known as GenCast, measures the likelihood of multiple scenarios to accurately assess trends from wind power generation to tropical cyclone movement.
GenCast’s probabilistic technique is a new milestone in the rapid advancement in the use of artificial intelligence for better and faster everyday weather projections, an approach that is increasingly being embraced by large traditional forecasters.
“(This) marks something of a turning point in the advancement of AI for weather forecasting, with the most advanced raw forecasts now coming from machine learning models,” said Ilan Price, a researcher at Google DeepMind.
“GenCast could be incorporated as part of operational weather forecasting systems, offering valuable insights to help decision makers better understand and prepare for upcoming weather events.”
GenCast’s innovation over previous machine learning models is its use of so-called “ensemble” predictions that represent different outcomes, a technique used in state-of-the-art traditional forecasting. GenCast was trained on four decades of data from the European Center for Medium-Range Weather Forecasts (ECMWF).
The model outperformed ECMWF’s 15-day forecast on 97.2 percent of 1,320 variables, such as temperature, wind speed and humidity, according to the paper published in Nature on Wednesday.
The results are a further improvement in the accuracy and scope of Google DeepMind’s discoveries GraphCast model presented last year. GraphCast outperformed ECMWF forecasts on about 90 percent of the metrics for forecasts three to 10 days ahead.
AI forecasting models tend to be faster and potentially more efficient than standard forecasting methods, which rely on massive computing power to process equations derived from atmospheric physics. GenCast can generate its forecast in just eight minutes, compared to hours for a traditional forecast – and with a fraction of the electronic processing required.
The GenCast model could be further improved in areas such as its ability to predict the intensity of major storms, the researchers said. The resolution of its data could be increased to match upgrades made by the ECMWF this year.
ECMWF said the development of GenCast was “a significant milestone in the evolution of weather forecasting”. It said it had integrated “key components” of the GenCast approach into a version of its own AI forecasting system, with live ensemble forecasts available from June.
The innovative machine learning science behind GenCast has yet to be tested in extreme weather, ECMWF added.
The development of GenCast will further fuel the debate about how much AI should be used in forecasting, with many scientists favoring a hybrid technique for some purposes.
Google presented in July NeuralGCM modelwhich combines machine learning and traditional physics to achieve better results than artificial intelligence alone for long-term forecasts and climate trends.
“There are open questions and debate about the optimal balance between physics and predictive machine learning systems. The broad scientific community, including (us), is actively investigating this,” ECMWF said.
Britain’s Met Office, the national weather service, is exploring how to use the “exciting” developments in its own AI-driven forecasting models, said Steven Ramsdale, chief forecaster responsible for AI.
“We believe the greatest value comes from a hybrid approach, combining human judgment, traditional physics-based models and AI-based weather forecasting,” he added.
Climate capital
Where climate change meets business, markets and politics. Explore the FT’s coverage here.
Curious about FT’s commitment to environmental sustainability? Learn more about our science-based goals here
