It’s able to accurately track the pathway of tropical cyclones, the researchers found. And it’s even slightly more accurate than the European Centre for Medium-Range Weather Forecasts, one of the world’s leading weather centers.
Still, Pangu-Weather has some limitations. The researchers didn’t investigate its performance on precipitation — a major weather variable and one of the trickiest to accurately capture in models.
The second paper, on the other hand, deals primarily with rainfall. It describes an AI system known as NowcastNet, a program that specializes in short-term forecasts maxing out just a few hours into the future. The researchers found that NowcastNet was capable of outperforming many of its leading competitors.
Pangu-Weather and NowcastNet are some of the latest in a recent wave of new AI weather models, many of them developed by private corporations rather than the government entities that traditionally dominate the weather. These programs differ from conventional forecasting systems in some fundamental ways.
Conventional forecasts rely on a system known as numerical weather prediction. It’s a kind of mathematical model that uses complex equations to predict the way weather systems change over time and space. These equations describe the actual physics behind the movement of air and water in the atmosphere and the oceans.
Because there’s so much math and physics involved, numerical weather models require extremely high levels of computational power. That makes them expensive and time-consuming to run. It also limits the fine-scale processes that these models can accurately capture. Things like the physics of individual clouds, for instance, are difficult to simulate in models that are making large-scale global predictions.
Scientists have come up with various ways to get around these difficulties in traditional models. One strategy is a method known as parameterization — that’s when scientists replace the actual physical equations in a model with a simplified program that generally captures the process without forcing the model to represent the actual physics.
But artificial intelligence could replace these workarounds, enthusiasts argue, with potentially faster and more accurate results.
AI models don’t have to represent actual physics in the form of mathematical equations. Instead, they ingest large amounts of historical weather data and learn to recognize patterns. They then use these patterns to make predictions when presented with new data on present-day weather conditions.
For several decades, scientists have worked to integrate AI components into traditional weather models in an attempt to make them faster and cheaper to run. And some firms are now developing all-AI models — such as Pangu-Weather and NowcastNet — that can entirely replace the numerical model system.
It’s a swiftly evolving field. Just two years ago, in a paper published in a Royal Society journal, scientists suggested that there “might be potential” for AI weather models to produce equal or better results than numerical models.
“We think that it is not inconceivable that numerical weather models may one day become obsolete, but a number of fundamental breakthroughs are needed before this goal comes into reach,” the researchers stated.
Emerging approaches like Pangu-Weather and NowcastNet suggest that such breakthroughs are in progress. And there’s potential for the field, said Colorado State University researchers Imme Ebert-Uphoff and Kyle Hilburn in a comment on the new research, also published Wednesday in Nature.
In principle, the much faster computational speed exhibited by models such as Pangu-Weather “could yield immense benefits,” they write.
On the other hand, there are still some potential obstacles for AI systems — especially as the planet grows warmer.
AI models may run into problems simulating extreme weather events as they grow more intense because of climate change, experts warn.
Heat waves, droughts, hurricanes, wildfires and a myriad of other climate-related events are all growing more extreme as temperatures rise, and some of them are veering into unprecedented territory. In the last week alone, heat records toppled all over the globe while scientists warned that the planet was likely experiencing its hottest days in human history.
Accurately forecasting extreme weather events is one of the most crucial functions for weather models, enabling decisionmakers to issue public safety announcements or facilitate evacuations with enough time to protect vulnerable populations. But AI models learn how to produce forecasts using historical weather data — and as the weather grows more extreme, there may be fewer examples of such intense events in the historical record.
That means AI systems might not have enough data to accurately simulate unprecedented extremes in the future. In fact, if they’re presented with weather conditions that are entirely foreign to them, it may be hard to predict how they’ll react.
The behavior of AI systems “is often unpredictable when the program operates under conditions that it has never encountered before,” Ebert-Uphoff and Hilburn warned in their comment. “An extreme weather event might therefore trigger highly erratic predictions.”
Other experts have raised similar concerns.
The authors of the 2021 Royal Society paper note that the “scarcity of extreme events” in the historical record poses a challenge for AI weather models. They also point out that while a few studies have attempted to evaluate the performance of AI systems when it comes to capturing extremes with limited data, they’ve produced mixed results — some have performed well while others have faltered.
“The question of how AI models will perform in a warming climate is a very interesting one, and to my knowledge hasn’t been explored very thoroughly at this point,” said Russ Schumacher, Colorado’s state climatologist and a scientist at Colorado State University, in an email to E&E News. Schumacher’s own research group has applied artificial intelligence to models predicting storms and other hazardous weather conditions.
Hybrid models that include both AI components and numerical model components may run into fewer difficulties with record-breaking events, he suggested. But for models entirely driven by AI, he said, “it’s not totally clear how it will respond to situations that fall entirely outside of the historical record.”
These are important evaluations to consider as researchers continue developing AI weather models, he added. They must investigate not only the way the models perform on routine, daily weather forecasts but also on dangerous, high-impact events.
In general, he suggests AI weather models have potential. But he noted that they may not entirely replace conventional approaches either. Numerical models and AI models may end up with different strengths, and human experience will remain valuable for synthesizing and communicating information about the weather.
“In my mind, we ideally get to a point where the field of meteorology can take advantage of the strengths of all of the approaches,” he said.
Reprinted from E&E News with permission from POLITICO, LLC. Copyright 2023. E&E News provides essential news for energy and environment professionals.
ABOUT THE AUTHOR(S)
Chelsea Harvey covers climate science for Climatewire. She tracks the big questions being asked by researchers and explains what’s known, and what needs to be, about global temperatures. Chelsea began writing about climate science in 2014. Her work has appeared in The Washington Post, Popular Science, Men’s Journal and others.
