Freezing rain is everyone’s least favorite winter weather forecast. No sledding, no snowmen—just slick sidewalks, icy highways, and downed power lines. Wouldn’t it be nice to know when you see that forecast for the dreaded ‘wintry mix’, on which side of freezing it’s going to fall?

That’s one of the big questions for the AI Institute for Research on Trustworthy AI in Weather, Climate, and Coastal Oceanography (AI2ES), a coalition of academic, government, and industry researchers who are employing machine learning to tackle this and a host of other tricky weather and climate problems.

“Leading experts from AI, atmospheric and ocean science, risk communication, and education, will work synergistically to develop and test trustworthy AI methods that will transform our understanding and prediction of the environment,” says Amy McGovern, AI2ES principal investigator and a professor of computer science at the University of Oklahoma.

Saving lives, saving property

Good weather can mean a fun day at the lake or more time to spend outdoors with friends while social-distancing. But bad weather doesn’t just ruin your day—it can be downright dangerous.

One of the superstars of violent weather is the tornado. Twisters kill around 80 people per year in the US and cause up to $20 billion in damage. McGovern says that humans already do pretty well forecasting large, violent tornadoes. It’s the smaller ones that they often miss—but those can still take lives.

“Violent tornadoes are usually forecasted well in advance. But if we can improve our prediction of the smaller ones, we can still save lives,” says McGovern. “There’s just so much data out there, that if we can use AI to find the patterns that humans are missing and improve the predictions, then we can improve resiliency, save lives, and save property.”

When it comes to the freezing rain mentioned above, the biggest forecasting difficulty lies in the one-degree Fahrenheit difference between regular rain and freezing rain. Additionally, if the temperature is hovering near 32 degrees, outcomes are very localized. It can be freezing and dangerous on one side of a city and safe on the other.

“If I give you a temperature forecast for today and I said it would be 72° and it turned out to be 73°, you wouldn't be mad,” says McGovern. “But if I said 33° and raining versus 32° and raining, there's a big impact difference. Current [weather] models don't really see that one-degree error as something significant.”

And it’s not just humans whose lives could be saved. Fluctuating temperatures have a big impact on wildlife too. McGovern gives the example of sea turtles that live off the coast of Texas. When a strong cold front comes through, the turtles in Corpus Christi Bay come to the surface, where they are run over by ships. If temperature predictions were more accurate, shipping traffic could be temporarily halted and up to 80% of the turtles saved.

How AI can help

In order for humans to accept machine predictions, the AI models must be trustworthy. McGovern defines trustworthiness as AI that forecasters, emergency managers, and the general public will want to use. “The different needs of those end users will drive what it means to be trustworthy,” says McGovern. 

One of the first things McGovern and her colleagues are hoping to develop is explainable AI. Many current machine-learning models are ‘black boxes’—meaning that their creators do not understand how they work. But McGovern says that using models that are explicable will help improve trust in their outcomes.

Another big challenge for the researchers of AI2ES is to create AI that reflects the physical reality of weather, climate, and oceanography.  

“Putting physics in [to a model] isn’t as simple as adding a little bit of math. If it was that simple, it would already be done,” says McGovern. “It tends to be that when you add physical constraints, you make things either not converge, or wildly diverge, or the model just doesn’t learn. So we have to figure out how to make it theoretically possible to put the constraints in while ensuring that we’re still getting the results we want.” 

The final piece is to develop robust AI that works under a variety of conditions. Meaning even if the data contains errors or omissions, the model retains a baseline of reliability and stability, and won’t predict something outrageous or impossible.

It won’t just be computer scientists and environmental scientists working towards these outcomes, but social scientists too. Risk communication researchers will interview end-users and the public in order to figure out what makes the AI trustworthy to them. “I think they are key to our success,” says McGovern.

Attracting the next generation

In addition to building trust, AI2ES also wants to build the next generation of climatologists and machine learning specialists. The Institute is working with Texas A&M—Corpus Christi and Del Mar College (both minority-serving institutions) to create an AI certificate program at the community college level, with the aim of improving diversity in both AI and geoscience careers.

“There’s a lot of research that shows minorities are interested in doing things that have an impact on the world,” says McGovern. “People really care about climate change right now. If we can show that using AI can improve something that relates to environmental science, we’re hoping we can hook more minority and under-represented students.”

Read more:

• Getting smarter about extreme weather
• Facial recognition improves hail forecasts
• How do tornadoes begin?