Accurately predicting the weather is hard — really hard, but a new AI-powered forecast model just hit a milestone that has experts saying your forecast could soon get more accurate, and further out, too.
It takes a Herculean effort to keep pace with weather in an atmosphere constantly in flux. The task is so difficult and complex that a reliable forecast more than a couple of days in advance was unheard of just a few decades ago.
A five-day forecast in the early 1980s was only accurate about 65% of the time. But better weather observations, more robust computing power and innovations in the way weather around the globe is modeled by computers has improved forecasts by leaps and bounds. Today the same forecast hits the mark nine times out of 10.
Forecasts took another step forward this month, experts said, thanks to GenCast, a new artificial intelligence forecast model by Google’s DeepMind. Its forecasts through 15 days were significantly more accurate than one of the most well-respected traditional non-AI forecast models, according to a study published by DeepMind in the journal Nature.
“It’s an impressive result,” said Peter Dueben, a machine learning expert and head of Earth system modeling at the European Centre for Medium-Range Weather Forecasts, home to the model bested by GenCast. “It’s a big step.”
GenCast isn’t ready for the public yet. It and other AI models still have a few key kinks to work out, particularly in forecasting the more frequent and severe weather of a warming world, before they change forecasting and save lives in the process.
Man vs. machine
The skill and usefulness of weather forecast models has always been closely tied to technology.
The majority of weather forecast models used today are based on a complex series of mathematical equations that model the physics of the atmosphere and use hundreds of millions of datapoints from real-time weather observations to paint a picture of how the weather will play out a day, a week or even a season from now.
This process of numerical weather prediction was first conceived in the early 1900s and needed to be done by hand, a method so slow that the weather had already happened long before calculations were finished.
Early computers improved forecasting in the 1950s and 1960s, but it wasn’t until 1974 that the first model able to pull in data from around the globe and generate a rudimentary forecast became operational.
Skip to the current day and supercomputers are performing a nearly unfathomable number of calculations daily to produce highly detailed weather forecasts many days into the future around the globe.
But current forecast models still have limitations. The most robust ones can only be run every few hours because of how long it takes to crunch the complex calculations. They also demand a lot of computing power and energy that make them costly.
And they have limitations when it comes to forecasting, too. The farther out in time they get from observations of the atmosphere, the more difficult it is to get a clear idea of what’s to come because the atmosphere never stops changing.
Most AI weather forecasting models like Google’s GenCast take a different approach. Rather than relying on observations plugged into physics-based equations, they predict how Earth’s atmosphere might behave in the future by analyzing verified past weather data to understand how the atmosphere behaved in similar situations. This helps improve accuracy over traditional models by eliminating errors from real-time weather data.
AI forecast models also run simulations a lot faster and use less computing power and energy than traditional models once they’re trained and ready to go. This means they can be run more frequently and model a wider range of possibilities, improving forecasts as they do.
Google’s game changer
AI weather modeling before Google’s GenCast has been limited to models that spit out a singular forecast without any indication of how likely it is to happen. It’s essentially a best guess that’s most useful for predicting common weather variables like temperature, precipitation and wind a handful of days in advance.
But GenCast runs dozens of simulations simultaneously.
“Once you have multiple possible futures it gives you a sense of both the range of what might happen and it also lets you calculate how likely some (futures) are rather than others,” according to Ilan Price, the lead author of the new study and a senior research scientist with DeepMind.
This type of modeling approach is highly regarded because it lends more confidence to weather forecasts for around five to 15 days in the future.
The European Centre for Medium-Range Weather Forecasts’s model is widely considered to be the gold standard. It was what Google wanted to beat with its first-of-its-kind AI version — and it did.
Researchers trained GenCast on 40 years of weather data up to 2018. They then used the trained model to predict more than 1,300 combinations of conditions like temperatures, precipitation and wind speeds, in 2019’s weather.
The AI model produced more accurate forecasts than the ECMWF’s traditional model for more than 97% of these variables within a 15-day timeframe, but showed particular skill within the first week of forecasts.
It showed anywhere from a 10 to 30% accuracy improvement on forecasts in the three-to-five-day range, depending on the exact combination of variables tested, according to Price. GenCast also had more accurate forecasts than the ECMWF’s model up to 15 days in the future, the study said.
The AI model could better capture some forms of extreme weather, including exceptionally high and low temperatures and extreme wind speeds. GenCast also needed less than 10 minutes to run on a supercomputer, compared to the hours necessary for traditional models.
The results mark an “inflection point” in AI weather modeling technology, Price said.
“AI-based weather forecasting is ready for prime time,” Price added. “It’s ready to start being incorporated alongside… traditional models in operation.”
GenCast is not in operation yet, but the DeepMind team plans to take another step toward it by releasing its present-day forecasts and an archive of its past forecasts, according to Price.
A big problem to solve
GenCast is a critical advancement in modeling, but like any other weather forecast model, it isn’t perfect.
AI models introduce a new potential issue since they predict the future based on what they’ve seen in past data.
“The machine learning model… doesn’t know anything about physics,” Dueben explained.
This can make it difficult for AI to conceive of future extremes that haven’t occurred in the recent past. Can an AI model trained on only 40 years of data accurately predict the types of extremes happening at a record pace in a changing climate, like a once-in-100 year or once-in-1,000 year torrential rainfall event?
“It turns out that actually these models are more robust to those extreme events than you would think,” Dueben said. The ECMWF has tested AI models against real-time weather for more than a year now and has seen improvements in their overall accuracy, even with extreme events, he explained.
But AI models can start inventing impossible-on-Earth physics the farther out in time they look, according to Dueben.
Other prediction issues remain, particularly with one of the most destructive weather phenomena: tropical cyclones.
Accurately predicting how strong a tropical cyclone like a hurricane or typhoon could become is an issue that plagues all models. It’s a crucial problem to solve as tropical systems get stronger and rapidly intensify more frequently in a world warming due to fossil fuel pollution.
GenCast showed better skill than traditional models when predicting the tracks of tropical systems but struggled to accurately capture intensity, according to Price.
In part, that’s because some of the recent notable record-breaking systems weren’t included in the 40 years of data GenCast was trained on, Price noted.
It’s an issue Price is “quite confident” can be overcome in the future as the model trains on more data.
In pictures: Hurricane Helene unleashes its fury
There are also models in development combining machine learning with real-world physics — known as hybrid models — that could be the solution to some of these problems.
Each step forward with this nascent technology adds another tool human weather forecasters can use to craft accurate forecasts people rely on for almost every aspect of their lives.
“You can be as skeptical as you want against machine learning forecasts in principle,” Dueben said. “These models will make a positive impact on our weather predictions; there’s no question there.”
