North Carolina
THE Ford Taurus is cruising along peacefully, when an identical vehicle
suddenly appears directly in its path. Neither driver has time to change course,
and the inevitable happens: the two vehicles collide head-on. The impact sets
off airbags that inflate and cushion the upper bodies of both drivers, but
nothing protects their legs, and their feet slam into the floor like pistons on
the go. Not surprisingly, the crash causes serious injury.
But this is no ordinary accident. There is no shattering of glass or
crunching of metal, but silence broken only by the soft hum of a powerful
computer. This is a virtual crash, identical in almost every respect to the real
thing. Despite the damage to the vehicles and their occupants, they will live to
crash another day, and another.
Advertisement
This virtual crash took place earlier this year at the National Crash
Analysis Center (NCAC) at the George Washington University in Ashburn, Virginia.
Here, a small group of computer engineers are pioneering methods for creating
virtual cars that crash in a realistic way. They have had considerable success,
but their most formidable task lies ahead: to create a virtual human鈥攁
kind of computer crash dummy鈥攖hat can accurately reveal the effects of
crashes on humans. 鈥淲e are addressing the safety problem from all angles and
that means creating simulated crashes,鈥� says Azim Eskandarian, the centre鈥檚
deputy director.
Safer cars and roads are desperately needed. In 1990, road traffic accidents
ranked ninth in the international league tables of causes of death and
disability. And in 1995, more than 40 000 people were killed in road traffic
accidents in the US alone, while numbers are rising in many other parts of the
world. While virtual crashes can鈥檛 improve driving skills, they can help to
reduce the chance of serious injury when crashes do occur.
The NCAC produces its simulations by making a model of a vehicle in its
digitisation laboratory. This room鈥攁bout the size of a two-car
garage鈥攈as a split identity, housing spanners on one side and computers on
the other. Car bonnets and panels hang from the walls, and shelves are jammed
with cardboard boxes filled with car parts.
In the centre of the laboratory sits a small white hatchback鈥攁 1997 Geo
Metro鈥攚ith its passenger side covered in a web of masking tape. For the
most part, the web runs in an orderly pattern across the car鈥檚 surface, almost
like the rules on graph paper. This grid is the first step in making a virtual
model of the vehicle.
The researchers use a technique known as finite element analysis, in which
every part of the car is divided into many small pieces, or elements, something
like a jigsaw puzzle but with more regular shapes. One 鈥渟quare鈥� of the grid of
masking tape represents each element.
To create a 3D model of the car on computer, the researchers must measure the
position of each intersection on the grid. The researchers make the measurements
with an articulated arm: it looks like a heavier version of the ones used to
support dentists鈥� tools but its function is to record the position of its tip.
Dhafer Marzougui, a member of this team, uses the arm by leaning over the
Metro鈥檚 bonnet and touching its pointer on every intersection. This records
their position in space to within a few tenths of a millimetre.
Every nut and bolt
The data are collated on a desktop computer that runs a conventional
computer-aided design program to model the vehicle鈥檚 surfaces. When the outside
of the car is finished, researchers start on the inside. They dismantle the car
and digitise every piece, down to the smallest nut or bolt. In the end, a model
may consist of up to 200 000 elements.
Throughout this process, the researchers must make compromises. While smaller
elements obviously allow surfaces to be modelled more accurately, every detail
increases the computer time required to simulate a crash. So the elements used
on areas of relatively smooth topography, such as the main surfaces of doors,
can be several centimetres across. More complex parts, such as the wheels, are
modelled with smaller elements that can be measured in millimetres.
With a completed model, the NCAC crew can begin simulating crashes using a
computer to calculate exactly what happens to each element. For the sake of
simplicity, each element is treated as a solid, undeformable shape. However, the
researchers must specify how strongly each element is linked to its neighbours
so that the bulk properties of the material such as its stiffness can be
accurately simulated. As long as the shape and size of the elements are chosen
carefully, the virtual crash can bear a remarkable resemblance to the real
thing.
First the team decide on a scenario鈥攕ay, the car hitting a wall at 50
kilometres per hour. On impact, the computer calculates the initial force. Then
the computer applies that force to the elements in the vehicle鈥檚 bumper which
touch the wall first. Those elements apply a force to the other elements in
contact with it, which apply forces to another series of elements, and so on.
The computer also works out how each element would move under these forces. In
this way, it can mimic the wave of destruction that tears through the vehicle as
it crashes.
The team then check their results against the real thing by comparing the
amount and pattern of buckling, and the final position and angle of vehicles, in
virtual and real crashes. Finally, they work out the energy balance of the
crash, based on the assumption that the energy required to twist and buckle all
the parts in the virtual car can be no more than the vehicle鈥檚 kinetic energy
just before the crash. The assumption is that if the energy going in equals the
energy coming out, the model is probably right.
Getting a simulation to work well takes considerable fine-tuning. When the
NCAC team 鈥渃rashed鈥� its first virtual model, a 1994 C1500 Chevrolet pickup
truck, it crumpled in a noticeably unrealistic way and they found that the
engine was anchored too strongly. 鈥淭hat usually means something鈥檚 too stiff,鈥�
says Steve Kan, director of modelling and simulation at the NCAC. So he and his
colleagues took a close look at the real vehicle鈥檚 front rail, on which the
engine sits. Sure enough, they found two holes in this rail, which had been
neglected in the model. Adding the holes to the model made the engine move much
more realistically.
No amount of refining can remove every error from a model. But perfection may
not be required. As Kan says, 鈥淪ome errors create positive effects and some
create negative effects, so they can offset each other.鈥� Real head-on crashes
between two Ford Taurus vehicles are surprisingly similar to a simulation鈥檚
pattern of damage.
The entire process鈥攆rom collecting 3D data to refining details鈥攊s
extremely time-consuming. The virtual model of the C1500 pickup truck took a
year to finish. But practice pays off, and progress on the two current
models鈥攁 1996 F800 Ford truck and the Metro鈥攊s speeding up. 鈥淲e are
aiming to finish a vehicle model in three months,鈥� says Marzougui.
Reaping actual benefits from the tests will also come with practice. The
simulations have already pointed out changes that would improve safety for
passengers. Modern safety devices such as airbags, for instance, do not protect
a passenger鈥檚 legs, but simulations show that leg injuries might be reduced,
perhaps significantly, by padding a vehicle鈥檚 floor.
To discover yet more ways of protecting passengers, the NCAC engineers also
simulate people. They began by building a 15 000-element computer model of a
Hybrid III dummy, which is used in crash tests. This is a huge improvement. The
Hybrid III only shows the effects of acceleration at a few places but the
computer model shows what happens at every one of its 15 000 elements. This
information reveals so many points of vulnerability that it gives a much more
comprehensive picture of how a passenger could be protected during a crash.
Flesh and bone
The team is not stopping here, however. Graduate students Paul Bedewi and
George Bahouth have made the first steps toward an even more realistic
model鈥攂y simulating a human leg.
A leg is a complex object, and a model of it must include every bone, and be
able to mimic the forces around joints created by muscles. 鈥淢uscle is very
complex. It is controlled by the brain, and neural activation dictates the
stiffness of the material,鈥� says Bahouth. In other words, a muscle鈥檚 mechanical
properties depend on signals sent by the brain. So modelling the muscle forces
around the joints in a leg requires knowledge about the orientation of the
muscles relative to the bones, the mechanical properties of bones and muscles,
and the brain鈥檚 output to the muscles. The researchers even need to know which
muscles are likely to be flexed or relaxed in a particular situation.
Bedewi and Bahouth replaced the computer dummy鈥檚 legs with a pair more
closely resembling human ones. With this improved model they can predict which
types of accident will produce enough force to break leg bones or tear soft
tissues like ligaments. Their model already offers one piece of advice: stay
relaxed during a crash, because tensing up increases the forces around joints,
which can lead to breaks and tears. Eventually, the NCAC team hopes to model the
entire human body, but that will require more research and a possible 500 000
elements.
There is more to an accident than just vehicles and passengers, however. The
NCAC team also concentrates on the scene of the accident, modelling roadside
hardware ranging from guard rails to lamp posts. For example, Leonard
Meczkowski, a road safety expert at the Federal Highway Administration in
McLean, Virginia, and his colleagues at the NCAC have studied what is called a
鈥渟lip-base鈥� for lamp posts. This is a device in which a lamp post is connected
to a base by bolts that pass through triangular notches. If a vehicle hits the
lamp post, the notches should let the base slip away from the bolts, leaving the
vehicle relatively undamaged.
The problem is that if the lamp post is hit from the side, the bolts can lock
in the notches, making the post almost as hard to knock down as a wall. By
simulating what happens to this base during an impact, the NCAC team arrived at
two simple improvements: make the notches wider and loosen the bolts. Those
modifications allow the base to slip when hit by a vehicle coming from any
direction.
Whether vehicles, people or roadside hardware, all these models demand
massive computing power. Imagine a simple model of a vehicle, made of only
30 000 elements, running into a rigid barrier鈥攖he easiest obstacle to model
because it doesn鈥檛 deform. To simulate 150 milliseconds of that crash on a
Pentium-powered computer running at 200 megahertz would take more than 250 hours
of computing time. But the NCAC team, equipped with two Silicon Graphics Power
Challenge XL machines, cranks out the same simulation in less than eight hours.
Faster computers will reduce this time even further in future.
Despite all the improvements in the time it takes to make models or run
simulations, much work remains for the NCAC researchers. They hope to expand
their range of virtual models until they have at least one vehicle in most of
the general classes, such as pickup trucks, small family cars and hatchbacks.
And they will continue to fine-tune their existing simulations, making even
better representations of crashing vehicles.
So what will the NCAC provide for future generations of drivers? 鈥淗opefully,
a safer environment, where you don鈥檛 pay for a mistake with your life,鈥� says
Meczkowski. 鈥淲e like to work on the concept of forgiving highways, which allow
you to make a mistake out there.鈥�
One day, an equation from the NCAC may let you make a driving mistake and
live to make another.