A huge dark cloud with a swirling funnel in southwest Texas

A tornado touches down in Southwest Texas.

The United States just experienced the busiest stretch of tornado activity in more than a decade. Scientists are now trying out new forecasting methods powered by artificial intelligence that could yield precious lead time ahead of these capricious and deadly storms.

Between April 25 and May 27, there were only two days when tornadoes didn’t touch down. According to a preliminary tally from the National Centers for Environmental Information, 1,117 tornadoes were detected between January and May of this year, the highest count over this time frame since 2011.

These menacing funnels of spinning air are deadly. Twisters over Memorial Day weekend killed at least 21 people across states including Kentucky, Arkansas, Oklahoma, and Texas. They’ve racked up billions of dollars in damages. They’ve also dropped down from the sky in places that rarely see them, like central California and outside of Washington, DC, forcing people who may have never experienced these storms before to seek shelter that may not exist.

Tornadoes remain one of the most dangerous weather events. And they buck an otherwise promising trend: While many types of natural disasters are killing fewer people over time thanks to better forecasting and stronger infrastructure, tornadoes can catch people off guard.

The lead time for tornado warnings is often less than 10 minutes, and progress has been frustratingly slow, especially when compared to other types of severe weather. (Forecasters can, for instance, predict the path of a hurricane far more accurately than they used to — three days in advance compared to just one day ahead in the 1990s.)

And troublingly, tornado patterns are shifting. Over the past 40 years, the number of tornadoes occurring in states like Arkansas, Mississippi, and Tennessee — places more densely populated than tornado hotspots over Texas and Oklahoma — are on the rise. Tornadoes also appear to be clustering together more often, with single thunderstorms spawning multiple twisters.

In the last few years, scientists have made progress in anticipating when the next twisters will touch down. In particular, forecasters are now testing a new set of tools built on machine learning, an artificial intelligence technique that trains computers to detect patterns without explicitly programming them.

Such forecasts won’t be able to tell a specific resident that their home is in the path of danger, but they are capable of a lot: These AI-driven programs can advise airlines to reroute traffic ahead of disruptions, allow farmers decide whether to hold off on watering their crops, and help disaster responders figure out where they should have additional emergency crews on standby.

These algorithms depend on good data to teach them, and that poses a major challenge for getting ahead of this particularly confounding phenomenon: As global average temperatures rise and as land use changes, past tornado activity might not reflect how these storms will whip through cities in the future.

Why tornadoes are so tricky to predict

One of the biggest obstacles to forecasting tornadoes is their size. “In the grand scheme of the atmosphere, they’re very small-scale,” said Russ Schumacher, a professor of atmospheric science at Colorado State University. “The biggest ones might be a mile wide. Most of them are smaller than that.” Tornadoes can rip entire homes off their foundations while houses a few blocks away are left unscathed.

Tornadoes are also short-lived, often just a few minutes. Detecting tornadoes with instruments like Doppler radars requires looking for subtle cues and still needs verification from storm spotters on the ground. Weather monitoring stations are often spaced too far apart to pick up smaller tornadoes before they form.

The complex physics powering these whirling towers of wind requires the processing power of supercomputers to simulate. Once they form, tornadoes can touch down, lift up, and change direction with little notice. That makes it hard to issue tornado warnings more than a few minutes in advance.

Hurricanes, in contrast, gather strength over days, can span hundreds of miles, and are visible to satellites, yielding ample time and information to generate useful forecasts, issue alerts, and get people out of the way. “I don’t think we’re ever going to have the level of specificity of forecasts for tornadoes that we do for hurricanes,” Schumacher said.

Most tornadoes erupt from a particular type of thunderstorm known as a supercell, which contains a rotating column of air that moves upward. According to Schumacher, they need four basic ingredients to form: a lifting mechanism that pushes air upward, instability in the atmosphere that allows that air to climb higher, a large quantity of moisture to fuel the thunderstorm, and wind shear that changes direction with altitude, thus causing the storm to rotate.

But not every supercell leads to tornadoes, and not every tornado hatches from a supercell. The specific strengths and quantities of the ingredients have to be just right. A little more wind here, or a bit more moisture there, can make the difference between an ordinary thunderstorm and a rampaging swarm of twisters.

“Forecasters now are really good at identifying the days when the ingredients are in place, when the potential is there for a lot of tornadoes to happen,” Schumacher said. “But it’s still really difficult to identify which of those storms is going to make a tornado.”

Could AI eventually hack the twister problem?

While it’s been difficult, there have been improvements in tornado forecasting over the past decade, and artificial intelligence has sped up progress more recently. Scientists have already developed AI weather forecasting systems that can outperform conventional techniques in some respects, but tornadoes remain a challenging test case. “That has the potential to make big advances but it’s still kind of in its very early stages in terms of evaluation,” Schumacher said. “This part of the field has evolved just in the last two years, so it’ll be really interesting to see in two or five years from now where it is.”

One of the conventional ways to predict weather is using numerical models, where scientists plug their observations into complicated physics equations that generate a prediction of how weather will play out. They require good measurements, a robust understanding of the mechanisms at work, and a lot of time-consuming computational horsepower.

Researchers refined these models and enhanced their resolution in the past decade, creating a sharper picture of how severe weather forms, particularly the kinds of storms that allow the convection needed to create supercells.

Scientists have also developed a better understanding of how tornadoes are influenced by broader global factors. The recent burst of tornado activity was influenced by the shift away from the Pacific Ocean’s warm phase of its temperature cycle, known as El Niño. Right now, the world is coming out of one of the strongest El Niños on record, and the Pacific Ocean is shifting into La Niña, its cool phase. As this shift happens, water temperature in the equatorial Pacific tends to introduce disruptions in the atmosphere above the continental US, creating a fertile breeding ground for tornadoes.

“When El Niño decays, the atmospheric waves change and can become wavier, so they have a greater amplitude,” wrote meteorology researcher Jana Lesak Houser in The Conversation. “The US often sees more frequent tornadoes when the climate is transitioning out of El Niño.”

Since the Pacific Ocean begins to telegraph when it’s likely to shift gears months in advance, this swing between El Niño and La Niña can be a warning sign that more tornadoes are brewing. Similarly, changes in the Indian Ocean’s temperature cycles can create ripples that lead to more spinning storms over North America. Known as the Madden-Julian Oscillation (MJO), these cycles create atmospheric disturbances over shorter time scales that move eastward across the world and over the continental US.

“El Niño sets the stage and then the MJO is the conductor of the orchestra,” explained Victor Gensini, a meteorology professor at Northern Illinois University who studies tornadoes. “We had several MJO cycles this year.” The intense heat wave over Central America and Mexico last month then evaporated plenty of water into the atmosphere that served as fuel for convective storms.

Now scientists are taking these historical records, present weather measurments, and computer simulations and feeding them into machine learning models to better predict tornadoes. One such forecasting model that’s currently undergoing testing at the National Weather Service’s Storm Prediction Center could anticipate heightened tornado activity over a region several days in advance of a strike.

The idea is to use past predictions from numerical models and line them up with historical observations of tornadoes. The machine learning algorithm then connects the dots between the meteorological starting conditions and where severe weather later emerges.

Schumacher said the machine learning system has proven especially useful roughly three to seven days ahead of a storm — a period when forecasters don’t have a lot of other tools that can make useful predictions in that time frame.

In this aerial view, a home is crushed by a fallen tree knocked down by a tornado in the Olde Towne neighborhood in Gaithersburg, Maryland.

Forecasters don’t want to overpromise and underdeliver when it comes to chalking out where the threats may emerge, but the machine learning model doesn’t have any compunction about drawing specific contour lines on a map of where it thinks tornadoes will crop up days from now. “I think the human forecasters tend to be a bit conservative,” Schumacher said. “[The machine learning tool] tends to be a little bit more bullish even at those longer lead times, but it’s turned out that a lot of the time it’s right.”

But scientists don’t want to take their hands off the radars and leave everything up to the AI just yet either. Gensini dubbed the current strategy “human-in-the-loop AI,” where a meteorologist evaluates predictions from the machine learning model to ensure they line up with the laws of physics. At the same time, researchers also want to keep an open mind and an eye out for any new, previously unrecognized relationships in weather that can cause tornadoes that might show up in the AI forecast.

“As an expert, you look at some of these and you’re like, ‘That doesn’t make any sense. Why is the model weighting that?” Gensini said. “Maybe it’s picking up on something.”

The big challenge for machine-learning forecasts, however, is that they’re learning from history.

Robust tornado records don’t go back that far and there are lots of gaps in sensor networks. And as humans alter the flows of rivers, cut down forests, and change the climate, future tornadoes will arise in a regime that looks less like the past. “If you’re seeing something or trying to forecast something that’s never happened before, then the model gets into some trouble,” Gensini said.

That’s why a key part of developing better tornado forecasts is gaining better observations.

That requires more Doppler radars, more monitoring stations, more weather balloons, more computer networks to collect, synthesize, and share this information. To catch the tornado of the future, we need more eyes on the weather of the present.

This story originally appeared in Today, Explained, Vox’s flagship daily newsletter. Sign up here for future editions.

You may also like...

Leave a Reply

Your email address will not be published. Required fields are marked *