Google DeepMind has built an AI weather forecast that is both faster and more accurate than the best system used today. The model, called GenCast, predicts global weather more accurately than the leading ENS system run by the European Centre for Medium-Range Weather Forecasts, long regarded as the world leader. It also produces a full forecast in minutes rather than hours. The work appears in the peer-reviewed journal Nature.
A better AI weather forecast is not just a technical win. Clearer predictions of cold spells, heatwaves, high winds, and storms help people plan, help governments warn the public, and help energy companies decide how much power their windfarms will produce.
How GenCast beat the top weather system
Researchers compared the AI model and the traditional system on the same forecasting task. In the authors’ evaluation, GenCast had greater skill than ENS on 97.2% of 1,320 targets they measured. Reporting the same study, The Guardian noted the model performed up to 20% better than the ENS forecast overall.
The gains showed up where they matter most. GenCast was better at predicting extreme weather and tropical cyclone tracks, including where powerful storms would make landfall. That kind of lead time can change how communities prepare.
Fifteen-day forecasts in about eight minutes
Speed is the second advance. GenCast produces 15-day global forecasts at 0.25 degree resolution, which works out to squares of about 28 kilometers across. Each run covers more than 80 surface and atmospheric variables and finishes in roughly eight minutes on a single chip built for machine learning.
The contrast with older methods is large. A traditional forecast can take hours on a supercomputer with tens of thousands of processors. Doing similar work in minutes on one machine makes it practical to run the forecast far more often.
How the model learned to forecast
GenCast did not solve the physics of the atmosphere step by step. Instead it was trained on decades of past weather data, from 1979 to 2018, learning how conditions such as wind, temperature, pressure, and humidity tend to change over time.
It also handles uncertainty differently from its predecessor. GenCast builds on GraphCast, an earlier Google model that produced a single best guess, by generating a range of possible forecasts and estimating how likely each one is. This mirrors a wider shift in which AI models now tackle problems once left to hand-built equations, while human experts still check the results.
What better forecasts could change
More accurate and frequent forecasts have practical value. GenCast is designed to help predict wind power production, which matters for anyone planning renewable power such as wind and wave energy on a grid that must balance supply and demand.
For now the AI is expected to support traditional forecasts, not replace them. “Outperforming ENS marks something of an inflection point in the advance of AI for weather prediction,” said Ilan Price, a research scientist at Google DeepMind, who added that these models will sit alongside existing approaches in the short term. The ECMWF called the work a significant advance and said some parts of GenCast already feed into one of its own AI forecasts.
Limitations and open questions
The results are promising, but experts point to real gaps. Sarah Dance, a professor of data assimilation at the University of Reading, said it is not yet clear whether the model captures the “butterfly effect,” the way tiny errors can grow quickly and reshape a forecast. She also noted that GenCast still leans on physics-based “hindcasts” to fill gaps in historic data, so it does not yet go straight from raw observations to a finished forecast.
There is also the plain fact that every forecast can be wrong. The team behind GenCast acknowledges that its model, like any other, can make mistakes. Readers can treat AI forecasts as one more useful input, check the sources below, and follow how national weather services adopt these tools over time.
Sources and related information
Nature – Probabilistic weather forecasting with machine learning – 2024
The peer-reviewed paper reports that GenCast is faster and more skilful than ENS, the top operational medium-range ensemble forecast, on most of the targets the authors tested. It also documents the 15-day range, resolution, and eight-minute runtime.
Google DeepMind – GenCast predicts weather and the risks of extreme conditions – 2024
Google DeepMind’s announcement describes how GenCast forecasts extreme conditions and wind power, the training data behind the model, and its relationship to the earlier GraphCast system.
The Guardian – Google DeepMind predicts weather more accurately than leading system – 2024
Ian Sample’s report explains that GenCast beat the ENS forecast by up to 20% and gathers comment from the Met Office, the ECMWF, and the University of Reading on what the advance does and does not settle.
