Oklahoma, where the wind comes sweepin’ down the plain. When singing and dancing cowboys deliver this song, you might think wind in Oklahoma is a rejuvenating experience no one should miss. Oklahomans know, however, that springtime on the plains is not a musical comedy, not when cumulus clouds roll up in dark, stormy fists, and conditions are right — or wrong — for the wind to curl into tight spirals of energy with high-speed updraft.

Severe storms spawn about 800 tornadoes a year in the United States, most of them on the Great Plains. The yearly toll in property loss and economic cost runs to billions of dollars, and that doesn’t count an average of 1,500 injuries and 80 deaths.

Better forecasts would reduce loss and save lives, but tornados, by their nature, are notoriously hard to predict. Forecasters can identify storms with tornadic potential, but warning time with current technology is seldom more than half an hour, and the warnings are imprecise. A major problem, furthermore, is forecasts that cry wolf.

Photo of
                  Kelvin Droegemeier.

Kelvin Droegemeier, University of Oklahoma.

Photo of Ming Xue.

Ming Xue, University of Oklahoma.

“We’ve been getting much better at detecting tornados and their precursor cyclones,” says Kelvin Droegemeier, director of the Center for Analysis and Prediction of Storms (CAPS) at the University of Oklahoma, “but three out of four tornado warnings is a false alarm.” Droegemeier and his CAPS colleagues have developed a storm-scale prediction capability, the Advanced Regional Prediction System (ARPS), that has set milestones in accurate forecasting of severe storms. But tornados remain a persistent challenge, and a big part of the problem is forecast data inadequate to the task.

Help is on the way. To prepare for a new, more comprehensive approach to the data-gathering stage of storm forecasting, Droegemeier’s colleague, Ming Xue, used LeMieux, Pittsburgh Supercomputing Center’s terascale system, to carry out the largest tornado simulation ever done. With 2,048 LeMieux processors and the ARPS model, he successfully reproduced a 1977 storm and the high-intensity tornado it spawned.

The results — which capture the tornado’s vortex structure, with wind speed of 260 miles per hour — represent a watershed in tornado research. “This is the highest resolution simulation ever done,” says Droegemeier, “of an entire thunderstorm and its tornado.”

Ground-Level Data

Warnings for no-show tornados may be inevitable with storm forecasting, and in large part people accept that it’s much better to err on the side of caution. Nevertheless, false alarms force businesses to close and people to stay home from work or flee the area. One false-alarm study in Wichita, Kansas found an economic cost of a million dollars for one evening.

A major factor in the high false-alarm rate has been the inherent limitations of NEXRAD, the national system of Doppler radar that feeds atmospheric data to weather-forecasting computer models. One hundred forty-two NEXRAD units cover the United States, but their conical beams focus at long-range, and because of Earth’s curvature, they look over the heads of tornados.

Illustration of NEXRAD and CASA radar systems. Filling the Gaps

Because of Earth’s curvature, the NEXRAD Doppler radar system doesn’t observe 72 percent of the atmosphere below one kilometer. More finely spaced CASA radars, mounted on cell-phone towers, will observe storms at ground level.

“NEXRAD sees maybe two kilometers above the ground,” says Droegemeier. “They’re missing a good chunk of the storm, especially near the ground, where tornados occur. That’s why we have this high false-alarm rate.”

As a solution to this problem, CAPS participates in a project called CASA (Center for Collaborative Adaptive Sensing of the Atmosphere). An NSF engineering research center, led by the University of Massachusetts, Amherst, CASA will deploy small, inexpensive Doppler radar units on cell phone towers. These units will aim at the ground and fill-in what NEXRAD doesn’t see.

Equally important, notes Droegemeier, CASA radar units will have sophisticated built-in intelligence. Each will collaborate with neighbor units in a network to track multiple atmospheric phenomena and adapt to rapidly changing weather. In contrast to NEXRAD, moreover, they will meet the needs of multiple end users, who at any given time will determine what data is collected.

First, however, the smart CASA detectors need to learn what kind of lower atmosphere information is associated with the origin of tornados. To begin to provide this information, which CAPS will use to develop scanning algorithms for CASA radar, Xue turned to LeMieux to simulate the 1977 tornado.

“The only way we can develop algorithms to hunt for these kinds of features,” says Droegemeier, “is to simulate them and then say ‘OK, let’s suppose these are real data’ and then apply the algorithms to the simulated data.”

Simulated Twister, Frame 1. Simulated Twister, Frame 2. Simulated Twister, Frame 3.

Simulated Twister
This sequence of frames (l to r) from an animation, created by PSC visualization specialist Greg Foss, show the simulated tornado as it moves within the grid volume and evolves. (Darkness of gray shading corresponds to water concentration.)

Tornado on the Grid

Along with the difficult problem of developing algorithms for CASA, Xue’s simulation also aims at better understanding of tornado dynamics. Why does one supercell thunderstorm — an intense storm with rotating updraft — produce a tornado while another one doesn’t? “We don’t really know,” says Droegemeier, “if there is a uniform dynamical mechanism or multiple mechanisms. Storms are much different from one another, and there’s no reason to believe that all tornados form in the same way.”

Computing limitations have hampered previous attempts to simulate tornados, with grid resolution held to a relatively coarse 100 meters, too porous to reliably capture a forming tornado — which have a median path width of only 46 meters. Other simulations have increased resolution locally — superimposing a finer mesh over a small area of the coarse grid. While somewhat successful, localized grids tend to skew the results, forcing the simulated tornado to form where it doesn’t in reality.

With LeMieux, Xue wasn’t bound by these limits. “I am basically able to run a uniform high-resolution grid that covers the entire storm system,” says Xue, “as well as being able to resolve the very small tornado in a system. And there is no uncertainty about where the tornado forms.”

To capture the entire storm, Xue laid out an area 50 kilometers on each side to an altitude of 16 kilometers. He subdivided this volume horizontally in 25-meter squares, with 20-meter vertical resolution at ground level, increasing with altitude, for a total of a third of a billion cells. At the center of each cell, ARPS calculated temperature, pressure and air speed every second, requiring 24 hours of computing time, with 2,048 LeMieux processors, for an hour of simulated storm, yielding 20 terabytes of data.


“The use of a uniform resolution grid large enough to contain the entire parent storm is a first,” notes Xue, “and eliminates the uncertainties of artificial human control associated with nested grids. In fact, the most intense tornado that developed in these simulations did so at an unexpected location within the model domain.”

Horizontal wind-speed and direction and radar reflectivity in a tornado. The Spiral Hook

This image shows horizontal wind-speed and direction (vectors) and radar reflectivity, which corresponds to rain-water concentration (increasing from light blue through green and yellow to red), in a horizontal plane (7 x 7 kilometers) 10 meters above ground. The spiral hook of reflectivity at the center is characteristic of tornado structure.

Xue and Droegemeier watched as the simulation produced a high- intensity (F5) tornado with wind speeds over 260 mph and a pressure drop of 80 millibars — the most intense tornado ever seen in simulation. Significantly, they could see the internal structure of the tornado, its twisting vortex and unstable wave patterns, and the coalescence of multiple vortices. “Basically,” says Xue, “the entire evolution of the tornado, including how it collapsed, appears very realistic.”

From where does a tornado get its whirling rotation? It’s one of the big questions of tornado dynamics, and Xue’s simulation provides a more detailed picture than has been available until now. Using the massive 3D data from ARPS, he calculated and tracked the path of air parcels in the storm. Air entering the storm, he found, undergoes an initial downdraft and flows along the ground before it feeds into the tornado vortex.

“The air turns abruptly in vertical directions in a very concentrated area,” says Xue. “It carries vorticity from the direction of the ground-level flow, which becomes vertical vorticity when it turns upward abruptly. This has not been well understood.”

Droegemeier expects to incorporate findings from Xue’s simulations into algorithms for CASA Doppler radar by spring 2006, when he plans to begin testing a set of four of these units on cell phone towers in Oklahoma. Each will cover a diameter of about 60 kilometers. “With these new radars,” says Droegemeier, “we are going to have fine-scale information to couple lower-atmosphere physics with wind and temperature structures at the ground. This will be another leap forward in numerical weather prediction. We believe CASA radars hold the promise of significantly reducing tornado false alarms — from the current 75 percent to 25 percent.”

© Pittsburgh Supercomputing Center, Carnegie Mellon University, University of Pittsburgh
300 S. Craig Street, Pittsburgh, PA 15213 Phone: 412.268.4960 Fax: 412.268.5832

This page last updated: May 18, 2012