Just How Much Can We Trust A.I. to Predict Extreme Weather?
The stories on artificial intelligence’s improving ability to predict extreme weather often begin with powerful hurricanes. As Scientific American detailed this year, when Hurricane Lee was moving through the Atlantic Ocean last September, scientists using traditional weather models got a clear idea of where the hurricane would go just six days ahead of its Canadian landfall. Nine days before landfall, an experimental A.I. modeling system called GraphCast predicted that outcome. As William J. Broad recently wrote in the New York Times, when Hurricane Beryl blew through the Caribbean early this July, a European weather agency predicted that the storm would most likely hit Mexico, though other landfalls were possible. On the same day as the European weather agency prediction, A.I. software predicted landfall in Texas, and four days later that is where the hurricane hit.
So A.I. has predicted the correct paths for certain storms earlier than other methods. But before we examine the significance of such accomplishments, it helps to have some history about how more traditional weather forecasting and A.I. work.
For decades we have gotten our forecasts thanks to numerical weather prediction. Satellites, weather stations and buoys collect data—like temperature, humidity and other variables—that is fed into giant supercomputers. The supercomputers generate a grid of cubes representing the Earth’s atmosphere and employ physics to see how the cubes interact with each other. “You can’t just simulate the physics of the atmosphere directly, because the atmosphere is too complex,” says computer scientist and meteorologist Amy McGovern, of the University of Oklahoma. “You’re trying to simulate what’s happening in one little area of it, and then you try to figure out how it’s interacting with the other areas.”
In short, the models are simulations of the atmosphere that use the current weather to predict atmospheric conditions in the future. But these models may need to run on giant supercomputers for hours—making quadrillions of calculations—before they can generate a prediction. Then the predictions go to a meteorologist who refines the forecast for a specific area.
While the American Meteorological Society began promoting and advancing artificial intelligence in the 1980s, its use has really moved forward in the past two decades or so in three main ways. First, for over 20 years, forecasters have done something called post-processing, where they take numerical weather models that are wrong and use A.I. methods to improve the predictions. Secondly, in the last roughly ten years or so, experts developed hybrid models—where bits of A.I. were plugged into numerical weather prediction models to speed them up.
Those methods using A.I. still incorporate physics, but the third big revolution in A.I. that has gained momentum in the last few years uses data-driven models that don’t incorporate physics at all. A.I. models trained on roughly 40 years of freely available weather data use the same collected weather data that is fed into supercomputers and create forecasts. But rather than having to make quadrillions of calculations to come up with a forecast, they simply look for patterns in data. For this reason, A.I. models can run on modest computers, even regular laptops, and spit out forecasts in seconds. And since the runs take hardly any time, A.I. can generate thousands of forecasts in the time it takes a numerical weather model to make one, allowing meteorologists to see a wider range of possible outcomes.
But just how much can meteorologists trust those fast new A.I. forecasts of extreme weather?
McGovern, who leads the NSF A.I. Institute for Research on Trustworthy A.I. in Weather, Climate and Coastal Oceanography (AI2ES) at the University of Oklahoma, is the perfect person to ask. She has been studying A.I. and weather prediction for twenty years.
Growing up, McGovern was inspired by Sally Ride, the first American woman to rocket into space, and wanted to be an astronaut. As she got older, she became interested in earth science—and also in ways to make computers smarter in a way that helped people. She earned her bachelor’s at Carnegie Mellon University and her PhD at the University of Massachusetts Amherst in 2002—both in computer science. When the University of Oklahoma offered her a job to focus on A.I. and weather beginning in 2005, she jumped at the opportunity. Because even though she wanted to leave Earth, she’s always been interested in earth science—and the job would allow her to fulfill her goal of using computers to help people. “Weather is an application where we can actually save lives and save property,” she says. “And I was like, ‘This is a perfect match.’”
To find out what she’s learned in the past two decades about the promise of A.I. to forecast extreme weather, we asked her a few questions.
How will A.I. change or improve forecasting extreme weather events like tornadoes or hailstorms or hurricanes?
The tornadoes and the hailstorms are something I’ve studied a lot of, and I think A.I. is getting used more in the post-processing sense still. Right now, the average warning time on a tornado is 15 minutes. If you could get that warning time up to 30 or 45 minutes [using A.I.] and you could improve the spatial resolution so that there’s a warning zone, but I’m going to make it as narrow as possible so I know that this is exactly where the tornado is coming, I think you can improve saving lives and property. We’ve definitely got some promising results with A.I. that can do that for both tornadoes and hail.
Another one related to that will be convective initiation, which is the start of the thunderstorms. Convective initiation matters for turbulence forecasting for airplanes. They have onboard radar, but it’s not giving them a great view. But if they knew where a storm was about to form, they could avoid flying over an area, and that could help with turbulence.
Hurricanes are a larger-scale event, so they’re a little bit easier to forecast. And these new systems are starting to predict where the hurricanes are going and when they’re going to develop. Those can’t really do tornadoes right now because they’re too small-scale. I think those systems show some promise of being able to give us more days in advance than the current systems.
Are there things A.I. can do right now better than numerical weather prediction can do?
I had two students both working on similar ideas, one on the convective initiation and one on hail. They were ingesting in real time the observations that had happened in the last 30 minutes and then giving you a forecast in about 30 seconds. That’s pretty cool. And it’s something that A.I. can do that numerical weather prediction just isn’t going to be able to do, because we don’t have computers capable of doing that.
Is there a risk that A.I. might miss outlier events related to climate change because it’s learning from events that have already happened?
Apparently, yeah, because it doesn’t have the laws of physics. Everything is changing so much, then it’s going to be hard for it to predict. I mean, how do you predict a flooding event that you’ve never seen? If you have some physics in there, you can at least give yourself some confidence in it. I mean, it’s hard for numerical weather prediction events to predict events they’ve never seen either, but they can.
One of the things I’ve seen you say before is that you want to create A.I. that is trustworthy. Can you explain what you mean by that?
That’s what my NSF A.I. Institute focuses on. You saw our long name, NSF A.I. Institute for Trustworthy A.I. in Weather, Climate and Coastal Oceanography. We focus on understanding what it means to trust an A.I. model and what it means for the forecasters to trust it or distrust it. Many weather situations are life-and-death situations where you’re trying to decide what you’re doing about evacuating or calling a warning or something like that. A trustworthy model is one that you trust to give you proper input for making that decision.
It isn’t one that, in our case, that’s going to replace the humans. It’s providing input to those human forecasters, but they trust what it’s giving them.
And are we there yet with A.I.?
There are some forecasts that are in operation inside NOAA [the National Oceanic and Atmospheric Administration], and so those are clearly trustworthy. Are we there with everything? No.
You mentioned NOAA. Who else is already using A.I. to forecast extreme weather events?
Private industry is using A.I., too. Not all of them are going to tell you about what they’re doing. Some of the smaller companies are just flat-out telling you they’re using A.I., but they don’t tell you much about it. Caveat to all of that, I am advising in a private industry company that’s a startup right now, so I know what they’re doing. I know it’s happening. Not as many people are talking about the inside methods yet.
And the Europeans, are they using A.I.?
They have a model called AIFS, which is the A.I. version of their IFS, Integrated Forecasting System. The AIFS is a model comparable to the Google model and the NVIDIA model and all these others, and I think it’s going to long-term be able to give you some nice hurricane warnings, except for they call them tropical cyclones in Europe. And heat waves.
When people get information from the A.I. models, are they using it in combination with information from the numerical weather models?
Yes. It’s not being used by itself right now.
You’re working with certain institutions and colleges on a certificate for A.I. Can you just tell me a little bit about what that is and what you hope to accomplish with it?
That’s part of AI2ES’s work, which is our NSF A.I. Institute. The certificate is being developed by Del Mar College, a Hispanic-serving, minority-serving college down in South Texas. They’re in Corpus Christi. They’ve prototyped it. They’ve had a bunch of students who’ve already graduated with it. We’re trying to reach a different audience and trying to get A.I. to more people, get more people involved in A.I. By working with a Hispanic-serving, minority-serving institution, we’re reaching a different audience than we get at our typical universities—and trying to help diversify the field of A.I. and A.I. for weather.
And I also read that you’re doing work with K-12 students in A.I.
That’s also being led by Del Mar. They’re amazing. They have an outreach that they do for K-12 students to teach them just about coding in general, and then to teach them some basic A.I. concepts. I went down and saw their camp one year. They’re doing a great job. They’re teaching kids how to fly drones and teaching them how to program the drones, how to program their little robots to run around and do different tasks, and just getting the kids excited about the things they can do with A.I.
Do you think A.I. forecasting will ever replace human meteorologists?
Well, “ever” is a long time to predict. How about the next five to ten years?
All right. That sounds good.
In the next five to ten years, we still need human meteorologists, because we still need that human expertise on top of things. Their physics-based knowledge is still really important. I get asked a lot: Are you developing things to replace the meteorologists? No. We’re developing things to give them more options so that they can focus on turning the forecasts into actionable information for people to make better decisions. Humans are a lot better at that than A.I. is right now.
This interview has been edited for length and clarity.
Full image credit: Illustration by Emily Lankiewicz / ABI imagery from NOAA’s GOES-16 Satellite – AWS S3 Explorer / NASA / Matthew Dominick
Source link Report Story