As climate change makes weather more unpredictable and extreme, we need more reliable forecasts to help us prepare and prevent disasters. Today, meteorologists use massive computer simulations to make their forecasts. They take hours to complete, because scientists have to analyze weather variables such as temperature, precipitation, pressure, wind, humidity, and cloudiness one by one. 

However, new artificial-intelligence systems could significantly speed up that process and make forecasts—and extreme-weather warnings—more accurate, two papers published in Nature today suggest. 

The first, developed by Huawei, details how its new AI model, Pangu-Weather, can predict weekly weather patterns around the world much more quickly than traditional forecasting methods, but with comparable accuracy. 

The second demonstrates how a deep-learning algorithm was able to predict extreme rainfall more accurately and with more notice than other leading methods, ranking first around 70% of the time in tests against similar existing systems.  

If adopted, these models could be used alongside conventional weather predicting methods to improve authorities’ ability to prepare for bad weather, says Lingxi Xie, a senior researcher at Huawei.  

To build Pangu-Weather, researchers at Huawei built a deep neural network trained on 39 years of reanalysis data, which combines historical weather observations with modern models. Unlike conventional methods that analyze weather variables one at a time, which could take hours, Pangu-Weather is able to analyze all of them at the same time in mere seconds.

The researchers tested Pangu-Weather against one of the leading conventional weather prediction systems in the world, the operational integrated forecasting system of the European Centre for Medium-Range Weather Forecasts (ECMWF), and found that it produced similar accuracy.

Pangu-Weather was also able to accurately track the path of a tropical cyclone, despite not having been trained with data on tropical cyclones. This finding shows that machine-learning models are able to pick up on the physical processes of weather and generalize them to situations they haven’t seen before, says Oliver Fuhrer, the head of the numerical prediction department at MeteoSwiss, the Swiss Federal Office of Meteorology and Climatology. He was not involved in the research. 

Pangu-Weather is exciting because it can forecast weather much faster than scientists were able to before and forecast things that weren’t in its original training data, says Fuhrer.

In the past year, multiple tech companies have unveiled AI models that aim to improve weather forecasting. Pangu-Weather and similar models, such as Nvidia’s FourcastNet and Google-DeepMind’s GraphCast, are making meteorologists “reconsider how we use machine learning and weather forecasts,” says Peter Dueben, head of Earth system modeling at ECMWF. He was not involved in the research but has tested Pangu-Weather.

Before, machine learning was seen as more of a “toy” project, Dueben says. But now it looks likely that meteorologists will be able to use it alongside conventional methods to make their forecasts. 

Time will tell how well these systems perform in practice. Conventional weather prediction systems are trained on observational data, whereas Pangu-Weather relies on reanalysis data. Xie says that they hope to train their model on observational data in the future. 

And while AI can help predict where tropical cyclones are heading, it cannot forecast how intense they will be. “AI will tend to underestimate extreme weather,” says Xie. 

However, other AI models might assist with that. A physics-based generative AI model called NowcastNet can predict extreme rain with a longer lead time than existing conventional methods. 

Existing deep-learning rain prediction tools, such as DeepMind’s DGMR, can predict the likelihood of all rain in the next 90 minutes. NowcastNet is able to predict extreme rain, a tougher task, up to three hours in advance. Sixty-two Chinese meteorologists evaluated the system against other similar systems and concluded it was the best rain prediction method in around 70% of cases. 

The team built a deep generative model that is trained on data collected from different weather radars and other technologies, such as sensors and satellites, Jordan says. The model is also trained on the principles of atmospheric physics—gravity, for example—and fed data from radars, which offer snapshots of weather patterns. The model can then generate the next likely scenario for the weather pattern. 

Because other models, such as DGMR, are trained only on radar data, they have only a partial snapshot of the atmosphere. That leads to less accurate results for rare events like extreme rainfall. Because NowcastNet is anchored in physics, the researchers say, their model is able to get a more comprehensive view of rain and how it might behave, leading to more accurate predictions. 

AI could help people buy more time when it comes to short-term predictions about weather events such as rainfall. Extreme rain causes massive death and destruction, and being able to predict it in a time frame that gives people a chance to prepare is important, says Michael I. Jordan, a computer scientist at the University of California, Berkeley, who worked on the study.

It’s still early days for AI-based weather forecasting, and it remains to be seen how useful these systems really will be in practice. Climate change might also complicate the picture, says Dueben. 

“The climate system is changing quite drastically. So suddenly all the ice in the Arctic disappears—no one knows what a model like Pangu-Weather will do,” he says.