Google Reveals New A.I. Model That Predicts Weather Better Than the Best Traditional Forecasts

There’s nothing more frustrating than making plans based on the weather forecast, then experiencing the prediction’s error firsthand when, for instance, it starts hailing during a picnic.

Now, Google’s DeepMind artificial intelligence research lab has presented GenCast: an A.I.-based weather forecast model that provides faster and more accurate 15-day weather predictions than the best traditional forecasting systems. The model was detailed in a study published last week in the journal Nature.

“GenCast is a machine learning-based weather model, which learns directly from historical weather data. This is in contrast to traditional models, which make forecasts by solving physics equations,” Ilan Price, DeepMind researcher and a co-author of the study, tells the Register’s Thomas Claburn. “One limitation of these traditional models is that the equations they solve are only approximations of the atmospheric dynamics. GenCast is not limited to learning dynamics/patterns that are known exactly and can be written down in an equation.”

The research team trained GenCast on historical weather data from the European Center for Medium-Range Weather Forecasts (ECMWF), spanning from 1979 to 2018. These included temperature, wind speed and air pressure measurements at different altitudes. They then tested the model’s 15-day forecast prediction on 1,320 weather events from 2019, and compared it to predictions by the ECMWF, “the world leader in atmospheric prediction,” per the New York Times’ William J. Broad.

GenCast predicted both day-to-day weather and extreme events more accurately than the ECMWF model, called ENS, outdoing it a whopping 97.2 percent of the time.

“Outperforming ENS marks something of an inflection point in the advance of A.I. for weather prediction,” Price tells the Guardian’s Ian Sample. “At least in the short term, these models are going to accompany and be alongside existing, traditional approaches.”

GenCast makes weather predictions based on a snapshot of the latest atmospheric conditions and provides them at 0.25-degree resolution, meaning in squares of approximately 17.4 miles by 17.4 miles, per the Guardian. It builds on DeepMind’s previous A.I. weather models such as GraphCast, a deterministic model, which means it only gave a single best weather prediction. GenCast, on the other hand, generates 50 or more forecasts, which provide a range of weather scenarios.

This method allows the program to express uncertainty in a forecast. “If most predictions show a cyclone hitting the same area, uncertainty is low. But if they predict different locations, uncertainty is higher,” study co-author Remi Lam explains in a Google DeepMind blog post. “GenCast strikes the right balance, avoiding both overstating or understating its confidence in its forecasts.”

GenCast can generate its 15-day ensemble forecast in eight minutes on a single Google Cloud TPU v5, a computer chip specialized in processing A.I. and machine learning. In contrast, traditional ensemble forecasts take hours to make predictions with supercomputers, according to the post.

“It’s a big deal,” says Kerry Emanuel, a climate scientist at MIT who was not involved in DeepMind’s development of GenCast, to the New York Times. “It’s an important step forward.”

Quick and accurate atmospheric predictions, especially in the case of extreme weather, are essential to saving lives, infrastructure and money. More accurate atmospheric predictions could also hold implications for green energy infrastructure, such as by helping predict how much power could be generated from wind farms.

“Weather forecasting is on the brink of a fundamental shift in methodology,” Sarah Dance, an applied mathematician at the University of Reading in England, tells the Guardian. She also points out, though, that GenCast’s training set included physics-based data to fill holes in past records, and this still has “a long way to go before machine learning approaches can completely replace physics-based forecasting.”

The research team, however, plans on making their data public in support of the global weather forecasting community, in line with their aim of enhancing traditional forecasting methods rather than replacing them.

“This is a really great contribution to open science,” Matthew Chantry, a machine-learning coordinator at ECMWF, tells Nature News’ Alix Soliman.

Margherita Bassi | READ MORE

Margherita Bassi is a freelance journalist and trilingual storyteller. Her work has appeared in publications including BBC Travel, Discover magazine, Live Science, Atlas Obscura and Hidden Compass.