Stay informed with free updates
Simply sign up to the Artificial intelligence myFT Digest — delivered directly to your inbox.
Google DeepMind has unveiled an artificial intelligence weather prediction model that outperforms traditional methods on forecasts up to 15 days and is better at foreseeing extreme events.
The tool, known as GenCast, gauges the likelihood of multiple scenarios to accurately estimate trends from wind power production to tropical cyclone movements.
GenCast’s probabilistic technique is a new milestone in the rapid progress in using AI to power better and quicker everyday weather projections, an approach big traditional forecasters are increasingly embracing.
“[This] marks something of an inflection point in the advance of AI for weather prediction, with state of the art raw forecasts now coming from machine-learning models,” said Ilan Price, a research scientist 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 novelty over previous machine-learning models is its use of the so-called “ensemble” predictions representing different outcomes, a technique deployed in state-of-the-art traditional forecasting. GenCast is trained on four decades of data from the European Centre for Medium-Range Weather Forecasting (ECMWF).
The model outperformed the ECMWF’s 15-day forecast on 97.2 per cent of 1,320 variables, such as temperature, wind speed and humidity, according to a paper published in Nature on Wednesday.
The results are a further improvement in accuracy and scope on Google DeepMind’s breakthrough GraphCast model unveiled last year. GraphCast outperformed the ECMWF’s forecasts on about 90 per cent of metrics for predictions three to 10 days ahead.
AI forecasting models are typically faster and potentially more efficient than standard forecasting methods, which rely on vast computing power to crunch equations derived from atmospheric physics. GenCast can generate its prediction in just eight minutes, compared with hours for the traditional forecast — and with a fraction of the electronic processing needs.
The GenCast model could be further improved in areas such as its ability to predict the intensity of big storms, the researchers said. The resolution of its data could be increased to match upgrades made this year by the ECMWF.
The 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 in a version of its own AI forecasting system, with live ensemble predictions available since June.
The innovative machine-learning science behind GenCast still needed to be tested on extreme weather events, the ECMWF added.
The development of GenCast will further fuel debate about how extensively AI should be deployed in forecasting, with many scientists preferring a hybrid technique for some purposes.
In July, Google unveiled the NeuralGCM model, which combines machine learning and traditional physics to achieve better results than AI alone for long-range forecasting and climate trends.
“There are open questions and discussion about the optimal balance between physics and machine learning forecasting systems. A wide scientific community including [us] is actively exploring this”, the ECMWF said.
The UK Met Office, the national weather service, is researching how to harness the “exciting” developments to its own AI-driven forecasting models, said Steven Ramsdale, chief forecaster with responsibility for AI.
“We maintain the greatest value comes from a hybrid approach, combining human assessment, traditional physics-based models and AI-based weather forecasting,” he added.
