Google DeepMind isn’t the sole massive tech entity leveraging AI for climate forecasting endeavors. Nvidia launched in 2022. In 2023, Huawei unveiled a sophisticated AI model that leveraged its vast repository of data accumulated over the course of nearly four decades? The system yields precise predictions, providing a solitary value rather than a range, such as forecasting a temperature of precisely 30°F or 0.7 inches of precipitation for the following day.
While GenCast diverges from Pangu-Climate in its approach to generating climate forecasts, the primary distinction lies in the former’s ability to produce probabilistic outcomes – a range of possible climate scenarios rather than singular, exact predictions. The forecast is likely to be “There’s a 40% probability of temperatures dropping as low as 30 °F” or “There’s a 60% likelihood of 0.7 inches of rainfall tomorrow.” Such an evaluation enables officers to assess the probability of various weather events and plan accordingly, ultimately enhancing their preparedness and response strategies.
These outcomes do not necessarily imply the end of traditional meteorology as a field. While the mannequin excels at modeling past climate patterns, its reliance on historical data may lead to flawed projections about an increasingly unpredictable and dynamic global climate, potentially yielding inaccurate forecasts for the distant future.
According to Aaron Hill, an assistant professor at the University of Oklahoma’s Department of Meteorology, GenCast still relies heavily on the ERA5 dataset, a comprehensive hourly estimate of atmospheric variables dating back to 1940. “The ERA5 dataset’s foundation is a sophisticated physics-based model,” he remarks.
While there are numerous unobservable environmental variables, meteorologists rely on physical laws to generate predictions, leveraging complex equations to estimate conditions. These estimates are combined with readily available observational data to inform models such as GenCast, and ongoing updates are consistently necessary. According to Ilan Worth, researcher at DeepMind and co-creator of GenCast, “A model trained to the level of 2018 would perform worse in 2024 than a model trained to the level of 2023 will do in that same year.”
Soon, DeepMind intends to leverage real-time data, including atmospheric conditions like wind and humidity, to predict the likelihood of making accurate statements solely based on weather patterns.
