DeepMind researchers have developed GraphCast, a machine learning-based weather prediction program that can forecast weather variables for the next 10 days in less than one minute. According to a report, GraphCast has demonstrated a 90% verification rate, outperforming traditional weather prediction technologies.
The AI-powered program operates by analyzing “the two most recent states of Earth’s weather,” incorporating variables from the current time and six hours prior. With this data, GraphCast can predict the weather state six hours into the future.
In practical applications, the AI has proven its effectiveness, accurately predicting the landfall of Hurricane Lee on Long Island a full 10 days in advance, surpassing the performance of traditional meteorological tools. Traditional weather simulations, which rely on complex physics and fluid dynamics, can take longer to generate forecasts.
GraphCast not only excels in terms of speed and scalability compared to traditional technologies but also demonstrates the capability to predict severe weather events, including tropical cyclones and extreme temperature fluctuations. The algorithm’s adaptability, as it can be re-trained with recent data, suggests continuous improvement in predicting weather oscillations aligned with broader climate change trends.
There are indications that GraphCast, or at least the underlying AI algorithm, may find integration into more mainstream services. Google is reportedly exploring the possibility of incorporating GraphCast into its products. This development aligns with the increasing demand for improved storm modeling, with supercomputers, like those developed by the National Oceanic and Atmospheric Administration (NOAA), working towards more accurate predictions of severe weather events and hurricane intensity forecasts.