The AI program GenCast proved to be better at predicting daily weather and hurricane and cyclone tracks than the European Center for Medium-Range Weather Forecasts (ECMWF), which is considered the world leader.
The AI program GenCast proved to be better at predicting daily weather and hurricane and cyclone tracks than the European Center for Medium-Range Weather Forecasts (ECMWF), which is considered the world leader.
As The Guardian reports, in the near future GenCast will help forecasters make forecasts, not replace them completely. But already it can improve forecasts of future cold, heat and strong winds, and help power companies predict how much power they can generate from wind farms.
In a direct comparison, the program produced more accurate forecasts than the traditional ENS system for daily weather and extreme events 15 days ahead, and better predicted the paths of devastating hurricanes and other tropical cyclones, including their landfall locations.
«Overcoming ENS marks a certain turning point in the development of AI for weather forecasting. At least in the short term, these models will accompany existing traditional approaches,» said Ilan Price, Google DeepMind Research Fellow.
Traditional physics-based weather forecasting solves a huge number of equations, but GenCast learned how global weather is evolving by learning from 40 years of historical data generated between 1979 and 2018. This includes wind speed, temperature, pressure, humidity, and dozens of other variables at different altitudes.
Using the latest weather data, GenCast predicts how conditions will change across the planet in squares up to 28 km by 28 km over the next 15 days in 12-hour increments.
While traditional prediction takes hours on a supercomputer with tens of thousands of processors, GenCast runs just eight minutes on a single Google Cloud TPU, a chip designed for machine learning.
In 2023, Google DeepMind introduced GraphCast, which produces a single best prediction. GenCast builds on GraphCast by generating an ensemble of 50 or more forecasts, assigning probabilities to various future weather events.
Forecasters welcomed this development of technology, notes The Guardian. Stephen Ramsdale, the Met Office’s chief forecaster in charge of AI, said the work was «exciting» and an ECMWF spokesman called it a «significant achievement», adding that GenCast components are used in one of its AI-based forecasts.
«This opens up the opportunity for national weather services to produce much larger forecast ensembles, providing more robust estimates of forecast probabilities, particularly for extreme events,» said Sarah Dance, professor of data assimilation at the University of Reading.
However, she noted that GenCast’s developers did not address whether their system has the physical realism to account for the «butterfly effect,» a cascade of rapidly growing uncertainties critical to effective ensemble forecasting. There is still a long way to go before machine learning approaches can fully replace physics-based prediction,» said Professor Dance.
The data that GenCast was trained on combines past observations with physical «hindcasts» that require complex math to fill in the gaps in historical data, she said.
Previously, Demis Hassabis, co-founder and CEO of Google DeepMind, became one of three researchers to win the Nobel Prize in Chemistry for work on protein structure prediction.
