For those who keep an eye on the elements, the outlook is bright: Researchers have developed an artificial intelligence-based weather forecast that delivers faster and more accurate forecasts than the best system available today.

GenCast, an AI weather program from Google DeepMind, performed up to 20% better than the ENS forecast from the European Center for Medium-Range Weather Forecasts (ECMWF), widely considered the world leader.

In the short term, GenCast is expected to support traditional forecasts rather than replace them, but even in a supporting role it could provide clarity about future cold squalls, heat waves and strong winds, and help energy companies predict how much electricity they will generate from wind farms.

In direct comparison, the program provided more accurate forecasts than ENS of daily weather and extreme events up to 15 days in advance and was better able to predict the path of destructive hurricanes and other tropical cyclones. including where they would land.

“Exceeding ENS marks something of a turning point in the progress of AI for weather forecasting,” said Ilan Price, a researcher at Google DeepMind. “At least in the short term, these models will accompany and complement existing, traditional approaches.”

Traditional physics-based weather forecasters solve a variety of equations to create their predictions, but GenCast learned how global weather evolves by training on 40 years of historical data generated between 1979 and 2018. These included wind speed, temperature, pressure, humidity and dozens more variables at different altitudes.

Based on the latest weather data, GenCast predicts how conditions around the planet will change in squares of up to 28 km x 28 km for the next 15 days in 12-hour increments.

While a traditional forecast takes hours to run on a supercomputer with tens of thousands of processors, GenCast takes just eight minutes on a single Google Cloud TPU, a chip designed for machine learning. Details are published in Nature.

Google has released a number of AI-powered weather predictions in recent years, the result of researchers looking at different approaches. In July, the company announced NeuralGCM, which combines AI and traditional physics for long-term predictions and climate modeling.

In 2023, Google DeepMind introduced GraphCast, which produces a single best-guess prediction at a time. GenCast builds on GraphCast by generating an ensemble of 50 or more forecasts and assigning probabilities for various upcoming weather events.

Meteorologists welcomed the progress. Steven Ramsdale, the Met Office’s chief forecaster in charge of AI, said the work was “exciting”, while an ECMWF spokesman called it a “significant advance” and added that components of GenCast would be used in one of its AI forecasts.

“Weather forecasting is facing a fundamental change in methodology,” said Sarah Dance, professor of data assimilation at the University of Reading.

“This opens up the opportunity for national weather services to produce much larger forecast ensembles, providing more robust estimates of forecast reliability, particularly for extreme events.”

But questions remain. “The authors have not answered whether their system has the physical realism to capture the ‘butterfly effect,’ the cascade of rapidly growing uncertainties that is critical to effective ensemble forecasting,” said Prof. Dance.

“There is still a long way to go before machine learning approaches can fully replace physics-based forecasting,” she added.

The data trained by GenCast combines previous observations with physics-based “hindcasts” that require sophisticated mathematics to fill in gaps in historical data, she said.

“It remains to be seen whether generative machine learning can replace this step and move directly from recent raw observations to a 15-day forecast,” Dance said.

The performance is promising, but is there a “Michael Fish moment” lurking on the horizon? “Will AI predictions be immune?” Price said. “All predictive models would have the opportunity to make a mistake, and GenCast is no different.”