Google DeepMind isn’t the only major technology company applying AI to weather forecasting. Nvidia released FourCastNet in 2022. And in 2023, Huawei developed the Pangu-Weather model, trained on 39 years of data. This produces a definitive forecast that provides a single number rather than a range, such as predicting tomorrow’s temperature to be 30 °F or 0.7 inches of precipitation.
GenCast differs from Pangu-Weather in that it produces probabilistic predictions, or probabilities of different weather outcomes, rather than precise predictions. For example, a prediction might be, “There is a 40% chance that the temperature will reach a low of 30 degrees Fahrenheit,” or “There is a 60% chance that tomorrow’s precipitation will be 0.7 inches.” This type of analysis helps authorities understand the likelihood of different weather events and plan accordingly.
These results do not mean the end of traditional meteorology as a field. Because the model is trained on past weather conditions, applying it far into the future can lead to inaccurate predictions of a changing and increasingly unstable climate.
Aaron Hill, an assistant professor in the University of Oklahoma’s Department of Meteorology, said GenCast still relies on datasets like ERA5, hourly estimates of various atmospheric variables dating back to 1940, but this study is not involved. . “The backbone of ERA5 is a physics-based model,” he says.
Additionally, there are many variables in the atmosphere that we don’t directly observe, so meteorologists use physical equations to come up with estimates. These estimates, combined with accessible observational data, feed into models such as GenCast, which constantly require new data. “A model trained by 2018 will perform worse in 2024 than a model trained by 2023 will do in 2024,” says DeepMind researcher and one of the creators of GenCast. says one Ilan Price.
In the future, DeepMind plans to test the model directly using data such as wind and humidity measurements to see how feasible it is to make predictions based solely on observational data.