Sunny with a Chance of Accuracy

Google's DeepMind yesterday unveiled an AI model capable of predicting the weather more accurately than existing forecasting systems. The breakthrough marks an advancement in the field, promising quicker and cheaper results. 

DeepMind researchers trained their model, GenCast, on data from 1979 to 2018 to predict the weather for 2019. In a study published in Nature, they found GenCast outperformed the world's leading atmospheric predictor—the ensemble system from the European Centre for Medium-Range Weather Forecasts—over 97% of the time on factors like wind speeds, temperatures, and precipitation. GenCast was also fast, producing 15-day predictions in eight minutes rather than the hours conventional systems can take. GenCast also outperformed in predicting fast-changing hurricane paths but underperformed in predictions of hurricane intensity.

At least 35 countries rely on Europe's ensemble system, which forecasts weather using room-sized supercomputers that calculate global observations from satellites and weather stations. In contrast, GenCast runs on smaller machines and uses historical data to spot patterns and present scenarios.