Google DeepMind is not the only big tech firm that’s applying AI to weather forecasting. Nvidia released FourCastNet in 2022. And in 2023 Huawei developed its Pangu-Weather model, which trained on 39 years of information. It produces deterministic forecasts—those providing a single number fairly than a variety, like a prediction that tomorrow can have a temperature of 30 °F or 0.7 inches of rainfall.
GenCast differs from Pangu-Weather in that it produces probabilistic forecasts—likelihoods for various weather outcomes fairly than precise predictions. For instance, the forecast is perhaps “There may be a 40% probability of the temperature hitting a low of 30 °F” or “There may be a 60% probability of 0.7 inches of rainfall tomorrow.” Any such evaluation helps officials understand the likelihood of various weather events and plan accordingly.
These results don’t mean the top of conventional meteorology as a field. The model is trained on past weather conditions, and applying them to the far future may result in inaccurate predictions for a changing and increasingly erratic climate.
GenCast remains to be reliant on an information set like ERA5, which is an hourly estimate of assorted atmospheric variables going back to 1940, says Aaron Hill, an assistant professor on the School of Meteorology on the University of Oklahoma, who was not involved on this research. “The backbone of ERA5 is a physics-based model,” he says.
As well as, there are lots of variables in our atmosphere that we don’t directly observe, so meteorologists use physics equations to work out estimates. These estimates are combined with accessible observational data to feed right into a model like GenCast, and recent data will all the time be required. “A model that was trained as much as 2018 will do worse in 2024 than a model trained as much as 2023 will do in 2024,” says Ilan Price, researcher at DeepMind and considered one of the creators of GenCast.
In the longer term, DeepMind plans to check models directly using data akin to wind or humidity readings to see how feasible it’s to make predictions on commentary data alone.
