the crash in the machine

If you live in a developed country,chances are good that you ride inan automobile every day without

thinking much about the risks. Indeed,several factors have made car travelsafer since the mid-1980s: the inclusionof air bags and other design improve-ments in vehicles, the use of seat beltsand even the increasing maturity of thedriving population have combined tolower the fatality rate on U.S. highwaysby 29 percent since 1987.

Nevertheless, driving remains a rela-tively risky means of transportation. Inthe U.S. alone in 1997 there were

6,764,000 accidents reported to the po-lice, according to the National HighwayTraffic Safety Administration (NHTSA).These accidents killed 41,967 peopleand injured nearly 3,400,000. It is asad fact that motor vehicle accidentsare the leading cause of death foryoung people between the ages of sixand 27. In addition to this incalculableloss of life, there are enormous eco-nomic costs. In 1994 the NHTSA esti-mated that the annual cost of motor

vehicle crashes exceeded a staggering$150 billion.

Cars that are better designed to pro-tect their human occupants in a crashare a major reason the rate of fatalitiesis lower today than it was in the mid-1980s. Unfortunately, though, compet-itive pressures in the automobile indus-try are forcing most companies tospend less money and time developingnew automobiles. In short, at a timewhen customers and governments are

The Crash in the Machine92 Scientific American March 1999

THE CRASH

by Stefan Thomke, Michael Holzner and Touraj Gholami

SIMULATED FRONTAL COLLISION of a BMW 5-series hitting a barrier at 64 kilo-meters per hour (40 miles per hour) shows an end result (above) similar to that of anactual prototype frontal crash of the same vehicle at the same speed (right). The testswere carried out at BMW’s Research and Engineering Center in Munich, Germany.

Copyright 1999 Scientific American, Inc.

demanding safer cars, the budgets todesign such cars are shrinking.

In computer technology, many au-tomakers are finding a way out of thisdilemma. Increasingly, these companiesare replacing their traditional crashtests—in which they verify new engi-neering concepts by running heavily in-strumented prototype cars into con-crete barriers—with “virtual” crashes,in which high-performance computerssimulate a collision. Over the past 10years, tremendous increases in comput-er speed and improved software haveadvanced crash simulation to the point

where the results are trusted with ahigh degree of confidence. The resultingsurge in the use of computers is revolu-tionizing the way vehicles are designed.

The savings in time and money havebeen impressive. For a traditional crashtest, the first step is building the proto-type vehicle, which generally takes fourto six months and costs hundreds ofthousands of dollars. Then it must beoutfitted with several crash-test dum-mies, which have embedded electronicsensors to record acceleration and cancost $65,000 apiece. A variety of in-struments, including high-speed cam-

eras, record the crash. Unfortunately,glass and other debris often partiallyobstruct the view, and crash dummiessometimes accelerate through interiorregions that are not covered by thecameras. Thus, the postcrash films usu-ally give engineers precious little thatthey can use to improve a design.

A simulated test, on the other hand,can be conceived, programmed on acomputer and carried out in days orweeks, and the main expense is payingthe salaries of the simulation engineers.True, the computers are typically eithertop-of-the-line workstations costing tens

The Crash in the Machine Scientific American March 1999 93

IN THE MACHINEIncreasingly, automakers are relying on computer simulations of accidents to develop safer cars more quickly and efficiently

BMW SIMULATION DEPARTMENT


of thousands of dollars or supercom-puters that cost millions. But unlike thecrash-test prototype vehicles, the com-puters are used over and over again andsometimes can have other applicationswithin the company apart from verify-ing crashworthiness.

Perhaps most important, computersimulations let design engineers work inways that would be impossible other-wise. For example, in a relatively shortperiod, they can carry out a barrage oftests aimed at improving a structuralpiece—such as one of the “pillars” thatconnects the roof of a car to the chassisbelow the windows—that strongly af-fects the crashworthiness of the entirevehicle. They can “replay” a simulationas slowly as they like and zoom in onany structural element or even on asmall piece of a structural element to seehow it reacts.

Such capabilities not only generate awealth of useful detail, they also enableengineers to make the most of the ex-pensive prototype collision tests. With agood set of simulated crashes, the de-velopment team can reduce the chancesthat an actual prototype crash test willgo poorly and require another round ofcostly redesign and retesting.

When Cars Collide

Although their economy and other ad-vantages are earning them a larger

share of the design and developmentprocess, computer simulations will com-plement, rather than replace, traditionalcrash tests for the foreseeable future.Steady increases in computer processingpower have let programmers achieve aremarkable level of fidelity and detail,but simulations do have inherent limita-tions. A fundamental one is that each in-dividual simulation can answer only aspecific question, such as: What effectwould a pillar 7 percent thinner have ona side impact at 50 kilometers per hour(30 miles per hour)?

Indeed, the kinds of questions thatcan be answered by simulation are lim-ited by the range of phenomena thatcan be modeled. For example, it is verydifficult today to simulate and predictthe outcome of rollover accidents be-cause of their duration and complexity.A rollover can take a full three seconds,as opposed to 100 to 150 millisecondsfor a more typical smashup. To simu-late that much time requires prodigiouscomputer power. The behavior of a carin a rollover can also be difficult to pre-dict, because it depends on road fric-tion and other factors. It is also essen-tially impossible to use computers to dis-cover whether any parts of the car willpresent a fire hazard in an accident—forexample, whether a fuel tank is prone toexplode.

Another reason prototype crash testsare not likely to become obsolete any-time soon is that government trafficsafety agencies in most developed coun-tries still require data from them. In theU.S. the NHTSA works with other orga-nizations to develop safety regulationsthat carmakers must meet in order tosell vehicles. In Europe the specific regu-lations are somewhat different, but thelegislative process is similar; the UnitedNations’s Economic Commission for Eu-rope issues regulations to its members,which the European Union may thenadopt, and vice versa.

These laws require automakers torecord data from prototype crashes inthe three main accident categories:frontal, rear and side impact. The regula-tions are detailed—specifying, for exam-ple, a frontal collision with a concretebarrier at up to 48 kilometers per hour.Typically these tests are first applied toearly prototypes during automotive de-velopment and later used by governmentagencies to sample the safety of produc-tion vehicles. Automakers often augmentthe standard prototype collisions withtheir own (more stringent) tests or withthose of the Insurance Institute for High-way Safety in the U.S., which permitthem to attain levels of crashworthinesswell beyond the government mandates.

The government requirements were

developed based on accident statistics. Inthe U.S. in 1997, according to the trans-portation safety administration, 45.2percent of all crashes involving passen-ger cars were frontal, 33.9 percent wereside impact, and 19.6 percent were rearimpact. Yet frontal crashes were dispro-portionately deadly: 61.9 percent of fatalaccidents involved frontal crashes, 25.2percent were side impact, and only 5.5percent were rear impact. These figureswere fairly typical; thus, in their effortsto improve crashworthiness, govern-ment safety agencies have traditionallyfocused on head-to-head collisions. Theresult has been seat-belt laws and the in-creasing use of air bags. In recent years,though, automakers have been payingmore attention to improving the abilityof vehicles to protect passengers in side-impact crashes, generally through theuse of head air bags [see bottom illustra-tion on next page] and through redesignof the pillars and other key structuralpieces in the side of a car.

To understand how designers usecrash data to make cars safer requiressome knowledge of the physics of colli-sions. An automobile accident is basi-cally just a transformation of energy:from the kinetic energy of the movingvehicles to the energy used in deform-ing their bodies during the crash. Thesingle most influential characteristic isthe vehicle’s speed at impact, becausethe absorbed crash energy goes up withthe square of velocity (a crash at 90kilometers per hour is four times moreenergetic than a crash at 45).

Weight is another key factor. Al-though weight is a disadvantage in asingle-vehicle crash, it can be a plus in amultivehicle accident. When heavy andlight vehicles collide, passengers in theheavy ones generally fare better—unlessthe light vehicles are built with stiffermaterials, relative to the heavy cars, inkey impact areas. In fact, this issuepoints to an unfortunate situation: gov-ernment-mandated prototype crashtests require crashworthiness to be eval-uated in isolation, despite the fact that63 percent of all fatal crashes involvetwo or more vehicles. If automakers


CHRONOLOGY OF A CRASH shows asimulated collision with a deformable barrierat 64 kilometers per hour. The interval be-tween successive images is 10 milliseconds.


had to minimize the damage to all thevehicles in a crash—for example, bybalancing weight and stiffness—sport-utility and other relatively heavy pas-senger vehicles would be built with ma-terials that were somewhat more pliant.Indeed, a few automakers are just nowstarting to incorporate this principleinto their designs.

Basically, all injuries that occur in anaccident can be traced to one of twocauses: the body’s collision with objects—

the steering wheel, for one—resulting inexternal injuries such as bruises or punc-tures; or the body’s sudden accelerationduring the crash, which causes injuriesinside the body such as bone fracturesand organ ruptures accompanied by in-ternal bleeding. In a prototype crash test,sensors in the dummies record peak ac-celeration; a lower acceleration indicatesbetter crashworthiness and thus a lowerprobability of death or severe injury.

This acceleration comes from mo-mentum, which the vehicle and whatev-er it hits transfer to each other in a col-lision. For safety purposes, one of themost significant factors is the rate atwhich the momentum is transferred tothe vehicle. This factor in turn dependson many variables; in the vehicle, mate-rial strength and stiffness, structuralsupports, the position of the engine andthe rigidity of the steering wheel col-

umn—to name a few of many designparameters—can all influence the de-gree to which a collision causes injury.

Piece by Piece

The computer programs that modelall these parameters are based on an

algorithmic technique known as finite el-ement analysis. With this method, pro-grammers represent each piece of thestructure as a group of finite elements,each of which is a polygon that has asso-ciated with it a mathematical descriptionof its physical and material properties,such as stiffness and tensile strength. Fora crash test, the complete model general-ly consists of several components: thebody of the vehicle, its seats, the engineand the passengers. Each of these piecesis further broken down. The vehicle, forexample, consists of door panels, win-dows, pillars, struts and other parts; pro-grammers represent each of these as agroup of finite elements.

The more finite elements in the model,the more closely it simulates reality. Cur-rently engineers use high-end worksta-tions or supercomputers, which arepowerful enough to simulate a vehiclemodel with 200,000 to 300,000 finite el-ements. The seats, engine and passengerscan add another 100,000 to 200,000polygons. Limitations in computer pow-

er have forced programmers to modelthe passengers as rigid, jointed figures,much like crash dummies; this is still thestandard practice in the industry. Buthigher computing speeds are finally en-abling some university researchers tosimulate occupants with more realisticfeatures, such as soft tissue and bones.The work is important because as com-puters continue to become more power-ful, it will only be a matter of time beforesimulation engineers will be able to com-pute the acceleration of specific organsin the body during a crash. This capabil-ity would be another significant advan-tage for computer-based simulations, be-cause although crash-test dummies haveembedded accelerometers, these sensorsmerely measure the increase in speed ofparts of the dummy. They cannot predicthow a specific organ suspended in thebody, a largely fluid medium, will move.

To generate the many thousands offinite elements in a model, engineers usedata from the computer-aided designprograms that are created early in thedevelopment process. Then they associ-ate with each element the physical prop-erties (mass, density, stiffness and so on)and contact conditions relative to the el-ements that surround it.

As they connect the elements to cre-ate a model, including the passengermodules, engineers fine-tune it, making

sure that the mass distribu-tion and the resulting centerof gravity represent reality asclosely as possible. Thefinished model is a complexpiece of software that com-putes how kinetic energy istransformed into deforma-tions, acceleration forces andother parameters during acollision.

Before a simulation, engi-neers create the crash condi-tions by setting the velocities,just before impact, for the ve-hicle and whatever it hits. Onimpact, the kinetic energy isconverted into deformationenergy according to the lawsof Newtonian physics. Cal-


PROTOTYPE CRASH TEST was recorded by high-speed still cameras. These images, of a side-impactcrash, represent the best possible visual results because they are unobscured by debris. During thecollision, the head air bag deployed, keeping the dummy’s head from hitting the side of the car.

BM

WSI

MU

LATI

ON

DEP

AR

TMEN

T

BM

WPR

OTO

TYPE

TES

TIN

G


culating the conversion from kinetic todeformation involves representing themovement within and between themany finite elements, using simple rela-tions. In effect, the programs sum theforces over all the elements, which re-sults in a system of equations that issolved using various mathematical andnumerical methods. The stress withinthe elements is determined using stan-dard principles of material behavior.The simulation is time-dependent,meaning that the system of equations issolved over and over again, each timeupdating the position and stress levels ofevery element. Each new iteration takesas its initial conditions the final resultsof the previous iteration. The conver-sion goes on, iteration after iteration,until there is no more kinetic energy leftto convert—or, in other words, until allmoving pieces have come to rest.

During the simulation, programmerscan determine the velocities and defor-mations at the vertices of the finite ele-ment polygons. They can then use thesevalues to determine the stress to whicheach finite element is subjected. For thepassenger components of the model(the “software dummies”), they mea-sure accelerations, movements andforces rather than levels of stress.

The three major simulation programsused by auto firms today are PAM-CRASH, LS-DYNA3D and RADIOSS.All three are based on programs that weredeveloped in the late 1960s for militarypurposes in the U.S. They all work on thefinite element principles outlined aboveand differ from one another subtly in theassortment of materials they can easilysimulate, the way they handle the simu-lated surfaces that come into “contact”with one another in the collision andthe software support they provide dur-ing model-building (preprocessing) andcrash-analysis (postprocessing) phases.

The programs grew out of alliancesbetween automotive firms and softwarevendors. PAMCRASH, for example,

was the result of a European effort in-volving Volkswagen, Ford, Opel andthe French software company Engi-neering Systems International (ESI).PAMCRASH is also widely used byJapanese car companies. The main pro-gram among U.S. automakers is LS-DYNA3D, which was based on codewritten at Lawrence Livermore Nation-al Laboratory for modeling nuclearblasts. RADIOSS was developed atMecallog, a French firm founded bysome former employees of ESI.

Exploring Simulation’s Potential

The power of simulation was well il-lustrated by a recent project at

BMW, in which a team of designers, asimulation engineer and a test engineerattempted to develop technical con-cepts that could improve the side-im-pact safety for all BMW ve-hicles. The team set out in1995 to explore the poten-tial of simulation (they usedthe PAMCRASH program),deciding to limit prototypetesting to only two crashesat the end to verify theirfinal design concepts.

An existing productionmodel, a 1995 5-series vehi-cle, served as the project’sstarting point. After each sim-ulation, the team met, ana-lyzed the results and designedanother experiment. As ex-pected, the team enjoyedquick feedback, enabling themembers to try out an ideaand accept or reject it withindays. The surprise was that asthe trials began to accrue, thewhole was more than thesum of the iterations; thegroup was increasing its fun-damental understanding ofthe underlying mechanics.

One notably fruitful exam-ple of this improvement in-volved the so-called B pillar,one of the six structuralmembers that connects the

roof of a car to the chassis below thewindows. (There are three such pillarson each side of any car; from front toback, they are labeled A, B and C.) Byanalyzing the records of prototype side-impact crashes from earlier developmentprojects, engineers on the team hadfound that in crash after crash, a smallsection of the B pillar folded. The sectionwas next to the bottom of the pillar [seeillustration on opposite page]. The fold-ing bothered them because when a pillarbuckles, its value as a barrier is compro-mised and the probability of passengerinjuries goes up.

The engineers assumed that addingmetal would strengthen the bottom ofthe pillar, making the car more resistantto penetration from the side. None ofthem felt that it was necessary to testthis assumption. One developmentteam member, however, insisted on a


IMPACT ENERGY of a 64-kilometer-per-hour head-oncrash during a collision with a rigid barrier is roughlyequivalent to the automobile dropping vertically from 16meters—about the height of a six-story building.

CRASH CHRONOLOGY continues as thevehicle’s front end is smashed in by about 70centimeters. More crumpling could thrust theengine into the passenger compartment.

DA

NIE

LS &

DA

NIE

LS


verification, pointing out that it wouldbe neither difficult nor expensive on thecomputer. When the program was run,the group was shocked to discover thatstrengthening the folded area actuallydecreased crashworthiness significantly.

Initially, none of the team memberscould explain the phenomenon. Aftermore iterations and careful analysis,though, they found the cause. Rein-forcement of the lower part of the B pil-lar, they discovered, would cause thepillar to be prone to folding higher up,above the reinforced area. Thus, thepassenger compartment would be morepenetrable higher up—closer to themidsection, chest and head area of pas-sengers. So the solution to the foldingB-pillar problem turned out to be com-pletely counterintuitive: weaken thelower B pillar rather than reinforce it.

Equipped with that knowledge, thegroup undertook a reevaluation of allthe reinforced areas in the bodies of allBMW vehicles then in production orunder development. The project im-proved to varying degrees the crash-

worthiness of all those automobiles.The team finished its work in 1996,

after it had carried out 91 virtual acci-dents and two prototype crashes inabout a year. For the developmentalvehicles that were redesigned, side-impact crashworthiness advanced anaverage of 30 percent over the initialdesign. This improvement was mea-sured in several ways, such as by cal-

culating and comparing the acceleration,in both virtual and actual crashes, of sim-ulated or dummy body parts, such as thepelvis and chest. It is worth noting thatthe two prototype crashes at the end ofthe project strongly confirmed the simu-lation results and also the economics oftesting: at a total of about $300,000, thetwo prototypes cost more to build, pre-pare and test than did the entire series ofthe 91 virtual crashes.

Similar projects at BMW focused onfrontal crashes and were successfulenough to win a commendation fromthe Insurance Institute of Highway Safe-ty. In 1997 the institute bestowed itshighest crashworthiness rating on theBMW 5-series, one of the cars whose de-velopment benefited significantly fromsimulations.

Virtual Crashes in the Next Millennium

During the next five to 10 years,software and design engineers will

be producing crash simulation modelswith several million finite elements. The

week it now takes to execute another it-eration in a series of tests will be downto half a day. These and other advanceswill bring about some importantbenefits. For one, software dummies willbecome considerably more detailed,mimicking human physiology and pro-viding data that no crash-test dummyever could. Automotive corporationswill probably also be able at last to sim-ulate rollover accidents. And such com-puting power will let engineers modelmore realistic accident scenarios, such asmultiple-vehicle crashes, including onesthat occur at various angles of incidence.

Moreover, automotive engineers willbe able to use computers to model theperformance of so-called smart safetysystems, such as air bags that detect apassenger’s position, weight and heightand use the information to adjust theforce and speed at which they deploy.Only with fast and inexpensive simula-tion will automakers be able to carryout the massive experimentation neces-sary to optimize these complex safetydevices.

Automotive safety experts have reallyjust begun exploiting the power of com-puter software and hardware. Over thenext decade, some major advances willcontinue to expand the role of computermodeling in the development process.And as simulation technology transformscrashworthiness, the success of this revo-lution will be measured in the number oflives saved.


The Authors

STEFAN THOMKE, MICHAEL HOLZNER and TOURAJGHOLAMI met three years ago at BMW’s Research and Engineer-ing Center in Munich, Germany. Thomke, an assistant professor oftechnology and operations management at Harvard BusinessSchool, was at the center doing field research for an academic proj-ect. At Harvard, his work has been mainly on the management ofresearch and development in the automotive, electronics and phar-maceutical industries. Holzner is the head of crash simulation at theBMW center. He holds a doctorate in mechanical engineering fromthe Technical University of Munich. Gholami is group leader in thecrash simulation department. His master’s degree is in engineeringfrom the University of Berlin.

Further Reading

Simulation, Learning and R&D Performance: Evidencefrom Automotive Development. Stefan Thomke in ResearchPolicy, Vol. 27, No. 1, pages 55–74; May 1998.

Modes of Experimentation: An Innovation Process—andCompetitive—Variable. Stefan Thomke, Eric Von Hippel andRoland Franke in Research Policy, Vol. 27, No. 3, pages 315–332;July 1998.

The World Wide Web sites for the National Highway Traffic Safe-ty Association (NHTSA) at www.nhtsa.dot.gov and the InsuranceInstitute for Highway Safety at www.hwysafety.org includesearchable databases with information on traffic safety and thecrashworthiness of vehicles.

SA

LOWER B-PILLAR AREA (green) of a se-dan needed to be weakened, not strength-ened, to protect passengers better in a crash.

BM

WSI

MU

LATI

ON

DEP

AR

TMEN

T

BM

WSI

MU

LATI

ON

DEP

AR

TMEN

T


the crash in the machine

Documents