Skip to main content
Partly Cloudy icon
60º

AI weather forecasting the future

Google GenCast claims victory

GenCast, a groundbreaking ML-based probabilistic model, it delivers faster and more accurate 15-day global forecasts than the world's top ensemble system, excelling in extreme weather prediction and tropical cyclone tracking.

JACKSONVILLE, FLA – A new AI-assisted weather forecasting model called GenCast is changing the game for predicting weather, making forecasts more accurate and faster than ever before.

Google is behind GenCast which is outpacing traditional weather models and reshaping how we predict everything from sunny days to severe storms.

Weather forecasting is like solving a giant puzzle made up of billions of tiny pieces—each representing data like wind speed, air pressure, and temperature. Computers do this by running trillions of calculations to put everything together. The more pieces you have, or the further into the future you want to predict, the more complicated it gets. Adding more details or predicting far ahead means more pieces and more calculations, which requires super-powerful computers.

One of the biggest challenges in forecasting is making sure you have enough accurate data to start with. Think of it like trying to complete a jigsaw puzzle without all the pieces—without good data, the forecast is incomplete and less reliable. The weather models also have limited time to process this data and give a forecast before a storm or heatwave hits. Even the best computers take hours to produce a 10-day forecast, which isn’t ideal when you need information in real time.

GenCast, however, has a distinct advantage. It combines current weather data with 40 years of historical records, including temperature, wind speed, and air pressure readings from around the world. This vast amount of information allows GenCast to “learn” global weather patterns, giving it a head start compared to older models that only work with real-time data.

In a comparison of accuracy, GenCast outperformed the leading weather model in the world—the ECMWF-ENS (from the European Centre for Medium-Range Weather Forecasts)—97.2% of the time, according to a peer-reviewed study. And for forecasts beyond 36 hours, GenCast was accurate 99.8% of the time. This includes predicting extreme weather events like hurricanes and cyclones with impressive precision.

What’s truly remarkable is the speed at which GenCast works. It can produce a 15-day forecast in just 8 minutes. By comparison, traditional weather models using supercomputers can take hours to complete a forecast.

This study was based on data from 2019, before the ECMWF upgraded its model in 2023, which now runs at a higher spatial resolution of 11 km. Results from the ECMWF’s higher resolutions put it ahead of GenCast for surface weather predictions during this past summer and in line with GenCast for upper air predictions.

Surface weather verification scores through time. Lower Y axis values have lower errors.
Upper air error scores show less variability with GraphCast having sligltly lower errors through day 5.

Google isn’t the only one working on advanced AI weather models. The ECMWF itself has its own AI model called AIFS. Other companies, like Huawei (with its Pangu-Weather model) and Nvidia (with FourCastNet), are also testing AI-based weather forecasting. Even Microsoft has entered the field with its model, Aurora, which works with GFS forecasts and runs at ECMWF’s higher resolution.


About the Author
Mark Collins headshot

After covering the weather from every corner of Florida and doing marine research in the Gulf, Mark Collins settled in Jacksonville to forecast weather for The First Coast.

Loading...