For decades, it’s generally held that any weather forecast more than two weeks out will probably be unreliable due to atmospheric complications, an idea first coined as “the butterfly effect” in the 1960s. New algorithms, some of which use machine learning tools, are trying to break that cap on forecasting. Today, meaningful forecasts can be predicted at about ten days out, but the two-week limit is still generally held. Google produced an AI model called Graphcast trained on 40 years of reanalysis data, which managed to improve its own 10-day forecast by 86 percent when trained on initial conditions. Graphcast has since caused some excitement about a new approach by which that two-week mark might be beaten.
Comments