System runs crashes for safety's sake
The National Crash Analysis Center is giving the crash test dummies some time off. The center, which is funded by the Federal Highway Administration and the National Highway Traffic Administration, is using supercomputers to wreck virtual vehicles and improve the safety of real ones. The wrecks are brutal and very detailed. Each day virtual cars and trucks slam into walls, guard rails, road signs and each other.
The National Crash Analysis Center is giving the crash test dummies some time off.
The center, which is funded by the Federal Highway Administration and the National
Highway Traffic Administration, is using supercomputers to wreck virtual vehicles and
improve the safety of real ones.
The wrecks are brutal and very detailed. Each day virtual cars and trucks slam into
walls, guard rails, road signs and each other.
Often the crashes would prove fatal to passengers and destructive enough to mangle a
real vehicle. But by simulating them, scientists can get accurate results from crashes
without destroying a single car.
Since the Ashburn, Va., facility was founded in 1992, scientists have worked on two
missions. The first was to become the federal depository for 14,000 films of car crashes
around the country.
The second was to improve vehicle safety. Steve Kan, a scientist who runs crash
simulations, said the films help researchers improve the computer models.
"We can run a simulation based on one of the films and then go back and see if the
model showed the same results," Kan said.
But the detailed graphics Kan watches today are highly advanced compared to past
models.
"In the past, a full vehicle model had about 28,000 elements in it," he said.
"Today, there are half a million elements."
An element is a small area of a vehicle for which the computer can predict behavior
based on surrounding elements. Like pixels on a screen, the more elements a model has, the
more detailed a prediction will be.
With only 28,000 elements, the simulations appeared blocky and were inaccurate, Kan
said. But today, sections of a vehicle's tires can be plotted during a simulated crash.
The center uses two Silicon Graphics Inc. Power Challenge XL supercomputers to run the
simulations. One system has 10 processors, and the other has six. Kan said he used to run
simulations on an IBM Corp. RS/6000 workstation, but they took too long.
"To run a Ford Taurus into a wall took one month," he said. "On the
Silicon Graphics machines, it takes 16 hours."
Mike Sheh, marketing director for Cray Research Inc., said the use of supercomputers
for crash models has grown since the first simulations were run in 1986.
"It's always evolving as computers get faster and add more memory," Sheh
said. "Unlike a calculator, which is pretty much the same today as it was 10 years
ago, the crash models have increased in size, resolution and reliability."
Sheh said adding more elements to crash models lets scientists examine action too tiny
to see with the naked eye and too precise for computer models of just a few years ago.
For example, when a person tries to bend a piece of metal, it does not appear to move.
But actually, Sheh said, any substance gives a little when force is applied. A computer
could model small changes in the metal that could not otherwise be observed.
"Today's models are more closely related to physics," Sheh said. "In the
past, we had to make simple assumptions about the behavior of materials. They were often
too simple to be useful."
The center has had several potentially life-saving successes. The Oregon Transportation
Department asked for the center's help with its state road signs. Although the signs were
meant to break if a car hit them, they were sometimes holding firm and leading to serious
injuries and deaths.
"If you can imagine going 60 mph and coming to a stop in half a second," Kan
said, "you can see the type of damage that would do."
Kan built models of the Oregon road signs that included the torque force of bolts, the
type of bolts, metal plates and the friction of all the elements against one another.
The road sign model had 10,000 elements. Virtual cars were then relentlessly smashed
into the signs to find out why the signs held up to violent impacts.
It turns out that rain and other elements could damage the signs, and extremely cold
weather made them harden. The center developed a sign model that was more forgiving.
Between the state departments of transportation and private industry's requests, Kan
said, the center stays busy. Not a moment goes by anymore when the computers are not
smashing up virtual vehicles.
Both Kan and Sheh said despite recent supercomputing advances, simulation still has its
limits. Even with accurate models, real cars will always be needed to test crash
simulations, just less often.
"If you are doing a very simple model, you might not need simulation," Sheh
said. "But if you need to probe a complex system, it's the right tool for the right
problem."