ORIGINAL PAPER

Nuno Gil Æ Iris D. Tommelein Æ Sara Beckman

Postponing design processes in unpredictable environments

Received: 1 February 2003 / Revised: 25 March 2004 /Accepted: 28 June 2004 / Published online: 17 August 2004� Springer-Verlag London Limited 2004

Abstract This work explores the effectiveness of designpostponement in the concept development of large-scaleengineering projects. Our empirical research shows lim-ited use of postponement in semiconductor fabricationfacility (‘fab’) projects despite evidence that the cus-tomer inevitably requests design criteria changes in theproject’s life. We simulate fab concept development as a2-stage process—conceptualization followed by design.We find that postponing the start of design in relation tothe completion of conceptualization reduces the averageresources spent on design and the variability in theconcept development duration but increases the averageconcept development duration. A sensitivity analysis onthe postponement lag duration indicates, however, thatsome degree of postponement may allow reducing de-sign rework without increasing the risk of overrunningthe project completion date, in comparison to the riskwith early commitment. Further, simulation indicatesthat the effectiveness of postponement decreases asdesigners’ capability to reuse work increases.

Keywords Design Æ Postponement Æ Large-scaleengineering design Æ Uncertainty Æ Projectmanagement Æ Design reuse Æ Change

1 Introduction

Project organizations involved in new product develop-ment—in computing, automotive, and other indus-tries—have to cope with increasingly dynamic,uncertain, high-velocity and turbulent market environ-ments. To accommodate these fast-moving markets,project organizations seek both to compress time-to-market and to accommodate the many design changesthat arise during the project development cycle (e.g.,Clark and Fujimoto 1991; Eisenhardt and Tabrizi 1995;Iansiti 1995; Ward et al. 1995; Thomke and Reinertsen1998). Empirical research shows that effective projectorganizations overlap the two major phases of theproduct development process—concept developmentand implementation—and postpone finalizing designdecisions as late as possible to accommodate change(Iansiti 1995). By postponing commitments on criticaldesign features until as late as possible during imple-mentation, project organizations can not only accountfor implementation issues in their design but also gainflexibility for reacting to unanticipated changes.

Our work resides within the domain of large-scaleengineering projects, and more specifically, investigatesthe concept development process for semiconductorfabrication facilities or ‘fabs.’ Fabs are complex high-tech buildings that house semiconductor manufacturingtools used either to research and develop new chiptechnologies (R&D fabs) or to mass-produce chips(high-volume manufacturing fabs). This research ex-plores phenomena in this domain similar to those ob-served in new product development projects. First, aneed exists to compress the fab project time (i.e., thetime between the start of concept development and thedate when the fab can start being used to produce chips)to facilitate early introduction of new products and topreempt competitive products. Second, during the longlead-time associated with designing, building, andramping up new fabs multiple exogenous events arise,creating a need to change design criteria. These events

N. Gil (&)Division of Operations,Technology and Innovation Management,Manchester Business School,The University of Manchester, UKE-mail: nuno.gil@umist.ac.ukTel.: +44-161-2004632Fax: +44-161-2004646

I. D. TommeleinEngineering and Project Management Programm,Civ. 1 and Environmental Engineering Department,U.C. Berkeley, USA

S. BeckmanHaas School of Buisness, U.C. Berkeley, USA

Research in Engineering Design (2004) 15: 139–154DOI 10.1007/s00163-004-0049-5

are difficult for fab designers to anticipate because theytend to relate to external changes in the chip manufac-turing technology and in the forecasts of market demandfor chips. Other design changes are generated inside thefab project environment when designers unearth incor-rect assumptions about product characteristics causedby lack of timely communication among designspecialists and by poor understanding of work inter-dependencies. In contrast with related work in engi-neering design that has focused on the ability ofproject organizations to accommodate internal changes(e.g., Smith and Eppinger 1997a, b; Yassine et al.2003), our research focuses on the ability of projectorganizations to accommodate significant, unpredictedexternal changes.

Intriguingly, our empirical investigation found lim-ited evidence in fab projects of decision postponement asreported in the product development literature. Hence,we employed our empirical findings about the decisionsmade in fab concept development and about the fre-quency of customer-requested changes in design criteriato build a model of the concept development process.We, jointly with practitioners, numerically calibrated thesimulation model for the case of R&D fabs and used itto investigate the pros and cons of using postponementin fab design projects.

Our research frames the development phases some-what differently than the product development researchwe cite. We focus on what the product developmentliterature calls concept development, and examine twophases within concept development: conceptualizationand design. Conceptualization corresponds to the de-signers’ initial effort to develop a fab technical concept,using rules of thumb and historical data. Design corre-sponds to the development of the technical concept intomore complete drawings and specifications, usingsophisticated analytical tools. Specifically, we use com-puter simulation to investigate the impact of postponingthe start of design in relation to the end of conceptual-ization (see Fig. 1a and b).

In this work, we seek to answer the following ques-tions:

1. When should fab designers start to design in relationto the completion of conceptualization?

2. How is the performance of concept development(conceptualization plus design) affected by postpon-ing the start of design in terms of lead-time, resourcesspent, and design process reliability?

3. Is the effectiveness of postponement (delaying thestart of design relative to the end of conceptualiza-tion) affected by the ability of designers to learn andreuse work developed prior to any significant change,and if so, how?

This paper is organized as follows. We review relatedliterature, and report our empirical findings on theconcept development process of fabs. Then, we describethe simulation model and discuss the main modelingassumptions. Finally, we use numerical experiments to

investigate the pros and cons of postponement formanaging concept development of fab projects; we alsodiscuss the managerial insights as well as the limitationsof the simulation exercise.

2 Related work

To delay decision-making for supporting large-scaleengineering projects in uncertain environments is not anew managerial concept. For the case of weapon systemsdevelopment, Klein and Meckling (1957), for example,propose that designers postpone final commitments oncritical design features until they gain more confidence intheir estimates to allow efficient use of limited resources.Recently, van Hoeck (2001) reviewed the literature onpostponement dating back to the 1960s to concludethat the recent growing interest in, and managerialapplication of postponement in domains as diverse asengineering design and operations/supply chain man-agement, should be interpreted as a rediscovery of theconcept. In the domain of engineering and constructionprojects, researchers have also long criticized the pref-erence of project organizations to pursue early com-mitment strategies and to focus on single-point designsolutions, which frequently result in missing promisedproject due dates and in performing extensive rework(e.g., Tommelein et al. 1991; Pietroforte 1997; Lottazet al. 1999).

Related to our approach is research that has usedanalytical constructs to study the effectiveness ofpostponement for managing product developmentprojects in unpredictable environments. Bhattacharyaet al. (1998) claim that having a sharp product defi-nition early on may not be desirable or even feasiblefor managing product development projects in high-velocity environments. Instead, they propose that firmsdelay commitments and gradually refine their productsolutions, according to the levels of uncertainty theyexpect, their own risk profiles, the difficulty in makingchanges to the product solutions, and the value of newcustomer information. Along the same lines, Terwieschand Loch (1999) use an analytical concurrent engi-neering model to demonstrate that uncertainty due toengineering changes and to interdependency betweentasks may make concurrency less attractive. Theysuggest that managers trade off savings in projectduration that result from overlapping activities againstrework delays caused by changes of preliminaryinformation.

Specific to the fab building environment, Wood(1997) uses an analytical model to analyze the effec-tiveness of modular tooling of fabs in accelerating thestart of manufacturing and in meeting the manufac-turer’s need for flexibility. Modular tooling gives man-agement the option to postpone decisions on tooling toas late as possible by tooling up the fab sparsely initiallyand then installing modules of tools as needed. Woodshows that postponement decreases risks associated with

140

obsolescence of capital equipment and inventories sinceit allows for more accurate matching of fab capacitywith demand and technology. This practice is onlyeffective, however, if lead-times to add tooling capacityare short.

Other related work has developed modeling tools tosupport engineering design management in uncertainenvironments, such as work using the design structurematrix (e.g., Eppinger et al. 1994; Yassine et al. 2003).This research has primarily focused however on internaluncertainty caused by work dependencies between de-sign specialties and between design and implementation.In contrast, our investigation uses discrete-event simu-lation to probe into the project performance impacts ofsignificant changes caused by events exogenous to thedesign team.

3 Empirical research

We started to investigate the dynamics of fab projects bygathering empirical qualitative and quantitative data,primarily through three methods: interviews with prac-titioners, attendance at design and construction coordi-nation meetings, and analysis of project records. Thisdata gathering work was made possible through a closecollaboration with Industrial Design Corporation(IDC), a leading design-construction firm specializing inhigh-tech facilities. Specifically, we used semi-structuredquestionnaires to interview senior people including leaddesigners of the mechanical, architecture, electrical,chemical, and structural specialties, construction-, de-sign-, and project-managers, and customer representa-tives. We questioned practitioners about the criticaldecisions they make throughout conceptualization anddesign, the sequences and durations of tasks, and ex-pected patterns, frequency, and implications of cus-

Fig. 1 Two models of concept development for large-scale engi-neering projects (adapted from Iansiti 1995)

141

tomer-requested changes in design criteria. We inter-viewed 22 design-related people and 10 owner repre-sentatives. The interviews lasted approximately 1 to 2 h.We carried out follow-up interviews with all the inter-viewees.1 In addition, we attended various constructionand design coordination meetings for fab projects on-going at that time as well as studied the records forseveral fab projects, including project proposals, meetingminutes, schedules, logs of design change orders, andproject drawings and specifications.

To promote the validity and reliability of the empiricalfindings, we triangulated the data. Triangulation is ‘avehicle for cross validation when two or more distinctmethods are found to be congruent and yield comparabledata’ (Jick 1979). van Hoeck (2001) suggests the adoptionof triangulation for enhancing the richness of the findingson postponement research. From our empirical research,we developed two conceptual constructs: (1) a model ofproject concept development, and (2) a probabilisticmodel of the sequences of significant changes in the courseof a fab project, as perceived by practitioners.

3.1 Construct 1: Concept development model

Figure 2 represents the concept development process forfab building systems (e.g., electrical, structural, mechan-ical, architectural, process, etc.) as a two-phase genericmodel: conceptualization followed by design. Duringconceptualization, designers use empirical rules, histori-cal data, and customer requirements to set forth the designcriteria and agree on the design definition for eachbuilding system.They estimate, for example, design loads,sizes of critical cross-sections, space requirements, andmajor equipment needs based on preliminary informationabout the expected area for the cleanroom or about the

expected number of wafer2 starts per month. The con-ceptualization phase deliverable is a report that, for eachbuilding system, summarizes the main features of the de-sign definition and may include order-of-magnitude esti-mates for the cost and schedule for construction.

During design, designers use sophisticated computer-based tools to refine the design definition developed inconceptualization. Design is an iterative loop of threetasks: load-, section-, and layout-development. Loaddevelopment is the process during which designers cal-culate the loads that the system should support. Sectiondevelopment is the process during which designers sizethe cross-sections of the main system given the designloads. Layout development is the process during which,based on the cross-section sizes, designers decide therouting of each utility system and the location of majorpieces of equipment. During design, designers also sizeand procure equipment with long delivery times (for thesake of simplicity, we excluded this activity from thescope of the model). Typical design phase deliverablesare sets of drawings and specifications for each buildingsystem, and more detailed estimates on cost and sche-dule for the construction phase. Conceptualization anddesign for each building system do not necessarily pro-gress at the same pace. Instead, some work (such asfoundation design) may go ahead of other work (such asarchitectural design) to allow construction work to startas early as possible.

3.2 Construct 2: External changes model

Diverse events exogenous to fab concept developmentprojects, such as new or updated forecasts of marketdemand for future chips and modifications in chipmanufacturing technology, affect the fab design criteriaand the design definition. Figure 3 illustrates these causeand effect relationships. These updates or modifications

Fig. 2 Concept developmentmodel

1Many professionals interviewed worked in several high-techdesign or contracting firms or even at customer organizations priorto their job position at the time of the interview; therefore, theknowledge we gathered largely reflects current practices in the high-tech engineering and construction industry.

2Wafers are the basic units of production in a fab. They are discs of(usually) silicon, on which the semiconductors are etched. Wafersare then sliced into what we know as semiconductor chips.

142

create the need to change design criteria such as thecapacities for the cleanroom and utilities. We observedthat designers group external changes in fab design cri-teria into three categories: (1) Full changes, which causedesigners to redo both conceptualization and design; (2)Partial changes, which affect work done during designbut have less impact on the work done during concep-tualization; and (3) Small changes, which have limitedrework implications. Full and partial changes affect thefab design definition in terms of cleanroom dimensions,design loads, cross-sections of routings, equipment sizes,and layouts of the utility systems.

Regrettably, we found that (at IDC, at least) projectmanagers seldom keep consistent logs of external designchanges and of their impacts on concept development. Asearch in IDC’s project database found consistent butlimited data on design changes, logged on a monthlybasis, for only one project, Fab X. Table 1 shows, forthree specialties, the initial estimates of work-hours forconceptualization and design and detailing, as well asthe number of additional hours spent in design changes.Note that significantly fewer hours are spent in con-ceptualization than in design and detailing, and work-loads differ across design specialties.

Figure 4 shows how the design change work hoursfor the chemical specialty in Table 1 were spreadmonthly in the course of the project. Note the pre-dominance in frequency of small changes, which causeless than 100 h of design rework per change. Full andpartial changes occurred only sporadically: the thirdchange in the first month of the project (caused by amodification of the design criteria for the industrialwater system), the second change in the eighth month(caused by release of a new tool layout), and the secondchange in the eleventh month (caused by unanticipatedneed for a copper lab) all significantly affected the workdeveloped by the chemical specialty. While the pre-dominance of small external changes in the project’s lifesuggests their project performance impacts merit furtherinvestigation, this work is limited to investigating theimpacts of full and partial changes.

Jointly with senior designers, we developed a modelof the set of possible and exclusive sequences of full andpartial changes in a fab project’s life (Fig. 5). The modelreflects the following premises:

1. Full and partial changes happen independently.2. Partial changes are more likely, and are likely to

occur earlier than full changes.3. The occurrence of the first change within a time

interval after the project start conditions (i.e., affectsthe probability of) the occurrence of the secondchange of the same type after a time lag. In turn, thesecond change conditions the occurrence of a thirdchange of that type, and so on. Senior designersinterpret an early significant change in the designcriteria as a signal that subsequent significant changesare somewhat likely to occur. In contrast, designersdeem the scenario of a first, significant change late inconcept development extremely unlikely.

4. The conditional probability of each subsequent sig-nificant change is lower than that of the immediatelypreceding change of the same type primarily becausethe more work that has been done, the more costly itwill be for project organizations to execute anothersignificant change, and thereby organizations will beincreasingly reluctant to do so.

5. The time when a significant change occurs is moredifficult to predict for significant changes that occurlater in the process.

Table 1 Estimates of work hours for the design process and hoursspent in external design changes (full, partial, and small) for themechanical/hvac, electrical, and chemical specialties (Fab X)

Mechanical/HVAC

Electrical Chemical

Conceptualization 290 340 660Design and detailing* 11,955 10,710 14,626Total conceptualization,design and detailing workhours

12,245 11,050 15,286

Design change work hours 1,723 3,215 4,993Design change as a percentof total work hours

14% 29% 33%

*An estimate of work-hours for design separate from detailing wasnot available; in detailing, designers develop the technical designconcept into drawings and specifications for construction andprecise bill of materials

Fig. 3 Cause and effect relationships between exogenous eventsand fab design

143

4 Simulating concept development

4.1 Modeling conceptualization and design

Simple models have been used to develop insights intothe pros and cons of various managerial propositions forproduct development projects in uncertain environments(Ha and Porteus 1995; Krishnan et al. 1997: Bhattach-arya et al. 1998; Loch and Terwiesch 1998; Terwieschand Loch 1999; Krishnan and Bhattacharya 2002).Simple models reduce the number of parameters thatneed to be estimated, and thereby provide a reasonablestarting point to research. Modeling simplicity alsofacilitates understanding of how different factors inter-play. Once the workings of a simple model are wellunderstood, it can be developed further by adding fac-tors initially left out but perhaps judged important bypractitioners. As a result, insights from simple modelscan be useful to complement research with more com-plex models.

Regrettably, our empirical research found that thehigh-tech engineering and construction sector seldomproduces reliable quantitative data on the conceptdevelopment process. The scarcity of hard data as wellas the need to ensure the simulation results were trac-table determined our choice to simulate the fab conceptdevelopment process with a simple event-graph model(Fig. 6). The interested reader can find detailed repre-sentations of this process in Gil (2001). The modelintegrates the concept development and uncertaintyconstructs described above, and was implemented inSIGMA, a discrete-event simulation environment based

on event scheduling (Schruben and Schruben 1999).Event-scheduling systems work by ‘‘identifying [a sys-tem’s] characteristic events and then writing a set ofevent routines that give a detailed description of thestate changes taking place at the time of each event’’(Law and Kelton 2000, p. 205).

In the description that follows, words in all-caps de-note geometric shapes in the figure, which representevents. Specifically, rectangles denote the beginning orend of design tasks, circles denote the start and end ofsimulation replications, and diamonds denote (full orpartial) changes of design criteria. The arrows representrelationships between the events they connect. Associatedwith each arrow is a set of conditions.A solid arrowmeansthat the event from which the arrow emanates schedulesthe event to which the arrow points after a time delay(Dt‡0), if edge conditions are met. A dashed arrow meansthat the event from which the arrow emanates cancels theevent towhich the arrowpoints after a timedelay (Dt‡0), ifthe latter is scheduled to occur and the edge conditions aremet.

Simulation starts with the START REPLICATEevent, which schedules the START CONCEPTUALI-ZATION event. The START REPLICATE event mayalso schedule the first PARTIAL CHANGE and the firstFULL CHANGE, with independent probabilities andstochastic delays. When a change event occurs, itschedules a subsequent change of the same type. STARTDESIGN LOAD may take place immediately afterEND CONCEPTUALIZATION or it may be post-poned. (Whether to postpone is a choice made by theuser, as discussed later). A FULL CHANGE uncondi-tionally cancels all scheduled events related to concep-tualization and design and schedules a new STARTCONCEPTUALIZATION event. Similarly, a PAR-TIAL CHANGE unconditionally cancels all scheduled

Fig. 4 Additional design work-hours per change for the chemicalspecialty over the duration of the project (Fab X)

144

events related to design and schedules a new STARTDESIGN LOAD event. Note that Fig. 6 only depictstwo sequences of change events, but other sequences canbe easily added to the model.

The model assumes that designers consider all designcriteria changes (full or partial) that may occur before aset project horizon, whether or not design is completedby the time the change occurs. Changes occurring afterthat horizon, however, are ignored to simulate a designfreeze at a specific deadline. This milestone is a decisionvariable that we purposely set far away from the STARTCONCEPTUALIZATION event in order to model arealistic situation that considers most significant chan-ges. Once concept development is completed and thesimulation time exceeds the set project horizon, theEND REPLICATE event is scheduled. This event col-lects the values of the performance variables for thesimulation run, cancels any changes that may still bescheduled to occur, resets all the simulation variables(except those that store data for purposes of statisticalanalysis), and schedules a START REPLICATE eventfor a new independent, identically distributed simulationrun.

4.2 Modeling rework

When full or partial changes occur, designers return totasks they have started and perhaps completed earlierand repeat them. The designers we interviewed suggestedthat some efficiencies typically exist as they repeat tasks;they believed that they could reduce the duration of atask by 50% or so the first time they repeated it, and thatfurther reductions in the duration of any task wouldgrow smaller and smaller as the task was repeated.However, no quantitative data was available to informon the actual gains achieved by designers from reusingwork. This led us to model two hypothetical algorithmsto simulate designers’ ability to reuse work (Fig. 7).

4.2.1 Algorithm 1: No design reuse

The first algorithm simulates an extreme situation inwhich designers do not reuse work. Whenever a full orpartial change occurs, the expected duration for a taskthat needs to be repeated is equal to its initial duration.This scenario can be written as

Fig. 5 Partial model for therange of possible sequences offull and partial changes

145

Diþ1;n ¼ Di;nþ1 ¼ Di;n ¼ D1;0; 8i; n ð1Þ

wherei, n Number of times (i=1, 2, 3,...) designers have

started to perform the task, given the number oftimes (n=0, 1, 2, 3,...) designers already com-pletely executed the task

Di,n Expected duration of the task in iteration i if nodesign change interrupts its execution, given thatdesigners already completely executed the task ntimes (days)

4.2.2 Algorithm 2: Limited design reuse

The second algorithm matches the designers’ intuitionson their ability to reuse work. The duration of a task in arework cycle is determined by prorating its duration inthe preceding cycle:

1. If the task was concluded when the change occurred:

Di;nþ1 ¼nþ 1ð Þ � Di;n

nþ 2; 8n; i ð2Þ

2. If the change interrupted the execution of the task:

Diþ1;n ¼ Di;n � Ti;n þi � Ti;n

iþ 1¼ Di;n �

Ti;n

iþ 1; 8n; i ð3Þ

whereTi,n Time designers spent working on iteration i before

a change interrupted its execution, given that theyhad already completely executed the task n times(days)

This algorithm assumes that if a change occurs aftercompletion of a task, the duration of this task decreasesin the next iteration due to efficiency gains and learning.If a change interrupts a task, in the next iteration thework that is repeated benefits similarly but the work thatremains will last what it was previously estimated to last.In addition, it assumes that these benefits decrease be-tween successive iterations of complete and incompletetasks. Figure 8 illustrates this algorithm for a sequenceof complete task iterations. This algorithm mimics alearning curve, in which the three basic assumptions are:(1) the amount of time required to complete a given taskwill be less each time the task is undertaken; (2) thisamount of time will decrease at a decreasing rate; and (3)the reduction in time will follow a predictable pattern(Chase et al. 1998, p. 446).

4.3 Measuring concept development performance

We applied three metrics to assess the performance ofthe development process in our simulations: conceptdevelopment duration, resources spent in design, andnumber of repetitions for each design task start event

Fig. 6 Event-graph model for concept development

146

(Table 2). Monitoring the concept development dura-tion is critical for understanding the repercussions ofpostponing the start of design on this critical projectdimension. Monitoring resources spent during design isequally critical because professionals skilled in designing

high-tech fabs are a scarce resource. The third met-ric—number of repetitions of the start of designtasks—provides a key measure of the rework that has tobe done.

5 Numerical applications

5.1 Calibrating external changes

Jointly with practitioners, we developed a stochasticmathematical model that matched their perceptions ofthe possible random sequences of partial and fullchanges for the case of R&D fabs for leading-edge mi-croprocessors and application specific integrated circuits(ASICS). We used re-scaled and shifted symmetric betarandom variables [a+(b)a)ÆBeta(a1=2,a2=2)] to ex-press the variability around the time when a changeoccurs. We employed the beta distribution—a parameterinput distribution—because the richness of shapes that itcan take on with simple changes of its a1 and a2parameters was needed to accurately align the mathe-matical modeling with practitioners’ perceptions. Thisflexibility is frequently exploited in simulation studieswhere a subjective approach to fit a distribution is

Fig. 8 Reduction in task duration due to efficiency gains forhypothetical sequence of complete task iterations

Fig. 7 Graphical representation of the design reuse algorithms

147

needed because reliable hard data is not available (Lawand Kelton 2000 p. 309).

A first set of interviews allowed us to quantify theparameters in the beta distributions (using Perry andGreig’s [1975] formulae for estimating the mean andvariance of subjective distributions), as well as to esti-mate the parameters A, B, and C in Table 3. Subse-quently, jointly with practitioners, we analyzed thesimulated histograms of changes (illustrated in Fig. 9) toascertain that the modeling assumptions were consistentwith their beliefs. The conditional probabilities and thetemporal relationships between changes of the sametype, within any stream of changes, can be stated as

P change1ð Þ ¼ A ð4Þ

P change2jchange1ð Þ ¼ A1þ B � 1:0 ð5Þ

or in general:

P changeijchangei�1ð Þ ¼ A1þ B � i� 1ð Þ ; i > 2 ð6Þ

P changeijchangei�1� �

¼ 0; i > 2 ð7Þ

T1 ¼ C 1þ Beta1 a1 ¼ 2; a2 ¼ 2ð Þ½ � daysð Þ ð8Þ

T2 ¼ T1 þ C 1þ Beta2 a1 ¼ 2; a2 ¼ 2ð Þ � 1þ Bð Þ½ � ðdaysÞð9Þ

or in general:

Ti ¼ C �"

iþXi

s¼1fBetas

�a1 ¼ 2; a2 ¼ 2

�

� ð1þ B � s� 1ð ÞÞ�#

daysð Þ; i>1 ð10Þ

where

P(i) Probability of change i occurring

P(i|i)1) Probability of change i conditional tooccurrence of change i)1

Ti Time when change i occurs (days)

Betai(a1=2,a2=2)

Symmetric beta random variable that issampled for every value of i

A Probability of a first change

B Declining rate in likelihood and in timepredictability for external changes (variedbetween 0 and 1)

C Mean and standard deviation for time delaybefore occurrence of first change (days)

Note that the conditional probabilities of the secondchange and of the subsequent changes of a given type aresmaller than the probability of the first change. Simi-larly, the model increases the time intervals betweensubsequent changes. Table 3 presents the designers’estimates for the parameters A, B, and C for the case ofR&D fab projects. Specifically, designers assert that thefirst full change in any R&D fab project will occur be-tween 20 to 40 days after the start of the project with a50% probability and that the first partial change willoccur between 15 to 30 days after project start with a90% probability. The probability of subsequent chan-ges—as well as the predictability of the moment whenthey will occur—decreases faster for full changes thanfor partial changes.

We employed Eqs. 4, 5, 6, 7, 8, 9, and 10 in thesimulation model to generate the expected distributionof both full and partial changes over the duration of anR&D fab project (see Fig. 9). The distributions are asexpected: the probability of occurrence of a chance de-creases over time, and the time lag between occurrencesincreases over time.

5.2 Computational assumptions

The numerical experiments reflect the following com-putational assumptions:

1. Each task has a deterministic duration. Given thesequential nature of the model, with simple finish-

Table 3 Estimates of A,B, and C for R&D fab projects

Constant Full change Partial change

A 0.5 0.9B 0.5 0.25C 20 15

Table 2 Description ofperformance variables Performance variable Description

Concept developmentduration (days)

Time elapsed between theoccurrence of the first START CONCEPTUALIZATIONevent and the occurrence of the END DESIGN LAYOUTevent for the last design iteration

Resources spent in design(workdays)

Workdays spent executing design tasks

Number of repetitionsof each start design task event

Number of times each START DESIGNtask event was repeated, regardless ofwhether the task was completed beforebeing interrupted by design changes or not

148

to-start relationships, stochastic task durationswould not change the means of the performancevariables resulting from a large number of experi-ments (a consequence of the strong law of largenumbers), although the variability of the perfor-mance variables would increase somewhat.

2. As the focus of this work is on the project perfor-mance impacts of significant external changes, thedesign loop (including load, section, and layout) isdone only once unless the design criteria change. Wediscuss a possible extension of the model to includeinternal design iteration in the model limitationssection.

3. Resources—implicitly allocated by assuming specifictask durations—are available to execute the taskswhether or not design is postponed. In practice,obtaining sufficient resources later on in a projectmay not be trivial, as we also discuss later in thispaper.

4. We tested both algorithms (no design reuse andlimited design reuse) to simulate design rework.Practitioners judge algorithm 2, limited design reuse,to be the most realistic representation as it is rare theydo not reuse some work even after a significantchange.

5.3 Sensitivity analysis on postponement lag

Because we empirically found limited evidence of the useof postponement—i.e., of delaying critical decisionsuntil as late as possible in the concept developmentprocess—we constructed a simulation model to test thepotential for their use. We evaluated several postpone-ment strategies. To do so, we locked in the earliest dayto START DESIGN LOAD at different points in theconcept development process. One extreme scenarioassumes that designers START DESIGN LOAD

immediately after END CONCEPTUALIZATION.This means that designers start design on day 25 (con-ceptualization lasts 25 days if no full changes interruptit), or on whatever day conceptualization ends (if fullchanges occur in the mean time). The other extremescenario assumes that designers postpone START DE-SIGN LOAD up to day 110 (corresponding to a lag of85 days if conceptualization finishes on day 25). Wehypothesized that the latter scenario would allowdesigners to maximize the probability of developing thedesign in a single pass. To develop a sense for how thevarious factors interacted, we tested several strategies inbetween, using increments of 5 days to vary STARTDESIGN LOAD from day 25 to day 110. For eachscenario, we ran 1,000 simulations.

Figure 10 charts the relationship between the conceptdevelopment duration and the number of resourcesspent during design as the postponement lag increasesfor the no-design-reuse scenario. The mean and standarddeviation of each data point in the chart are calculatedwith their unbiased estimators applied to the results of1,000 independent, identically distributed simulationruns. This assumes that the 1,000 observations for eachof the two variables are approximately distributed asnormal random variables. This is acceptable since thecentral limit theorem says, in effect, that if the number ofobservations is sufficiently large, the observations areapproximately distributed as normal variables, regard-less of the underlying distribution of the correspondingvariables (Law and Kelton 2000, p. 248).

Figure 10 illustrates that as the postponement laggrows, the mean concept development duration in-creases, and the mean number of resources spent duringdesign decreases. Further, the shape of the curve showsthat as the postponement lag initially increases from ano-postponement strategy, the marginal reduction of themean number of resources spent is very steep while themarginal increase of the mean concept development

Fig. 9 Histograms of partialchanges for 1,000 simulationruns

149

duration is relatively small. As the postponement lagcontinues to increase, the marginal reduction of themean number of resources spent is less significant whilethe marginal increase of the mean concept developmentduration tends to equal the corresponding marginal in-crease of the postponement lag. Figure 10 also showsdecreasing variability in concept development durationas the postponement lag increases. Further, it shows thatthe one-standard deviation upper limit of the conceptdevelopment duration (lt+rt) remains more-or-lesssteady for postponement lags up to approximately 25 to35 days. Clearly, up to this lag, the marginal decrease inthe variability of the concept development durationcounterbalances the marginal increase of the meanconcept development duration. As the postponement lagcontinues to increase beyond 25–35 days, the marginalincrease of the mean concept development duration getsmore significant and the marginal decrease in its vari-ability no longer suffices to prevent the upper limitlt+rt from also increasing.

Postponement also decreases the mean and the vari-ability of the number of resources spent because fewerchanges fall during design, and thereby design tasks arerepeated less. Figure 10 shows a square that encom-passes a set of efficient postponement strategies that bestsimultaneously satisfy two conditions: (1) minimize themean number of resources spent during design (lr) andtheir variability (rr); and (2) do not increase the upperone-standard deviation limit of the concept developmentduration (lt+rt) beyond the value that lt+rt assumeswith a no-postponement strategy.

Figure 11 charts postponement lag versus (1) themean number of repetitions for the start design taskevents, (2) the mean number of changes falling withinthe postponement lag, and (3) the mean number ofchanges falling after completion of concept develop-ment. Table 4 shows the corresponding standard devi-ations of these means for selected scenarios. Figure 11shows that as the postponement lag increases, the mean

numbers of repetitions of the start design task eventsand of the changes falling after concept developmentdecrease to nearly zero, whereas the mean number ofchanges falling within the postponement lag increases.The graph shows that the mean numbers of repetitionsfor any start design task event do not decrease steadilybut rather oscillate slightly up and down along adecreasing trend line, ultimately reaching zero. Theselocal phenomena indicate that an incremental increase inthe postponement lag does not equally reduce the rep-etition of all the start design task events because of localinteractions between the time-dependent means aroundwhich changes occur and the durations of the post-ponement lag and of the tasks. These phenomena wouldhave been hard to anticipate without conducting asimulation, even for a process as simple as the one wepresent here. For more complex concept developmentprocesses, the effect of postponement will likely be evenmore difficult to gauge, since each specific postponementlag leads to unequal benefits for the various tasks. Giventhat the duration of design tasks and the timing of sig-nificant changes may differ for each design specialty,changes may force one specialty into doing more reworkeven though this rework does not reflect their own skillsand abilities. One specialty may also benefit less frompostponement than another, and therefore may be lesseager to buy into this strategy. Design managers must bemade aware of such phenomena so that they will rewardteam performance and not exclusively individual work.

5.4 Sensitivity analysis on designers’ ability toreuse work

Figure 12 provides a comparison of three scenarios. Thebaseline scenario is an extreme, unrealistic scenario thatassumes fixed design criteria for all the 1,000 replica-tions, thus eliminating external uncertainty and theoccurrence of full or partial changes. It is shown as the

Fig. 10 Concept developmentduration versus resources spentduring design, for alternativepostponement strategies(1,000 runs for each data point)

150

flat line across the bottom of the figure. The no-design-reuse curve developed in Fig. 10 is also shown; addi-tionally, the mean limited-design-reuse scenario is shownalong with a set of four curves that depict the variabilityaround that mean curve. Figure 12 illustrates that theeffects of postponement are more significant when thereis no design reuse than when limited design reuse is as-sumed. This is intuitive: an increase in the length of therework loop adds more work when design tasks have tobe repeated, and thereby designers would be better offpostponing design. Figure 12 also suggests that, whenlimited design reuse is assumed, the efficiency zone cor-responds to somewhat shorter postponement lags. Thisis an expected result since, in this scenario, repetition ofthe tasks takes less time but the time-dependent meansaround which external changes occur remain the same.

These results confirm what would be expected: designtools for enhancing designers’ ability to reusework—commonly used in software and chip designdevelopment (e.g., Jacome et al. 1999), but much less inconstruction—reduce the effect of external changes onthe concept development process. These tools can helpdesigners to cope better with changes and to estimatemore accurately the duration of the concept develop-ment process in unpredictable environments.

6 Managerial insights and numerical sensitivity

The numerical simulation experiments analyze conceptdevelopment performance as we vary two factors: theduration of the postponement lag and the ability ofdesigners to reuse work. The results show that earlycommitment in uncertain environments, though efficientfor compressing the mean duration of concept devel-opment (which may not be desirable if concept devel-

Table 4 Postponement effects on performance variables (mean±one standard deviation)

Performance variable No postponement Design cannotstart before day 45

Design cannotstart before day 90

[lag�20 days] [lag�65 days]

No. of repetitions of START DESIGN LOAD event 1.71±1.21 0.93±1.00 0.21±0.43No. of repetitions of START DESIGN SECTION event 1.22±1.03 0.74±0.90 0.18±0.39No. of repetitions of START DESIGN LAYOUT event 0.77±0.98 0.48±0.74 0.09±0.29No. of changes falling after concept development 0.34±0.65 0.25±0.52 0.02±0.12No. of changes falling within the postponement lag 0 0.78±0.64 1.50±0.97

Fig. 11 Variation of the mean numbers of repetitions of startdesign task events and of change occurrences for alternativepostponement strategies (1,000 runs for each data point)

151

opment is to overlap effectively with implementation),comes at a cost. First, it maximizes the average numberof times that designers repeat the design tasks, andconsequently maximizes the average resources spentduring design. Second, early commitment makes itharder to predict the duration of the concept develop-ment process and the number of resources it will con-sume. In contrast, the results show that as design isgradually postponed—up to a certain efficiency zo-ne—design iteration and resources spent in design can bereduced, while the risk that concept development maylast longer than a preset milestone date (lt+rt) remainsnearly the same. As the duration of the postponementlag increases beyond the efficiency zone, the mean con-cept development duration gradually increases propor-tionally to the increment in the postponement lag andthe savings in the resources spent in design decrease.

We performed sensitivity analyses to examine therobustness of the simulation results with differentnumerical estimates for the task durations and for thechange occurrences. For the sake of brevity, these resultsare not shown here but the simulation dynamics were asexpected: initial increments in the postponement lagtend to correspond to marginal gains in process pre-dictability and in savings in spent resources, whilecausing limited increases on the average concept devel-opment duration. Beyond a certain efficiency zone,increments on the postponement lag tend to increaseequally the average concept development duration andto bring negligible savings in resources spent in design.Logically, according to the underlying numerical esti-mates, the curve in Fig. 10 shifts in the (x, y) space andthe prominence of its concavity varies. For example, a

reduction in the frequency of changes reduces theeffectiveness of postponement and vice versa. Also, theeffectiveness of postponement increases if the levels ofuncertainty remain the same, but the design loop lastslonger as would be the case if internal iteration had beenmodeled or if tasks had been assumed to last longer.Figure 12 illustrates these dynamics when we vary theassumption on designers’ ability to reuse work. Readersinterested in experimenting further with the model areencouraged to contact the first author to obtain theexecutable program.

7 Model limitations

Our simulation model was purposefully simple to ensurethat results were tractable and to limit the number ofestimated parameters given the limited amount of harddata that we found available. It is clear that the modeldoes not encompass a broad range of issues importantfor engineering design.

First, difficulties in allocating resources—a factor notmodeled in this research—can be a potential drawbackin implementing postponement. Managers expressedconcern that if they let team members get involved withanother project during a postponement lag, they wouldhave difficulty getting their teams back together later.We include no provision in our model to represent thispossibility. Instead, our model in effect assumes suffi-cient capacity in the system to move resources as re-quired. Note that under loading resources (i.e., adding acapacity buffer), an approach commonly used by Japa-nese manufacturing organizations, allows resources toaccommodate variability in work demand and conse-quently increase workflow reliability throughout theprocess (Hopp and Spearman 1996, p. 157).

Fig. 12 Concept development duration versus resources spentduring design for different rework algorithms and for alternativepostponement strategies (1,000 runs for each data point)

152

Second, the model only focuses on the design processitself, not on its output. It cannot differentiate, forexample, the quality of a design solution that results fromseveral iterations from the quality of a solution that isengineered with mature design criteria. The model alsocannot reproduce the practice in the industry of over-engineering selected features in the design definition toshield it from external changes (as an alternative todesigning amore flexible design process as explored here).

Third, this work only models the concept developmentprocess for one fab building system. It would be inter-esting to model the process for two or more systems,including critical information hand-offs between special-ists. In doing so, simulation could generate insights onpossible cascading effects in which a significant externalchange affecting one system increases the likelihood ofsignificant changes affecting other systems. Such anexpanded model could also investigate the implicationsfor the design process of one system of postponing deci-sion-making in the design of another system. Anotherinteresting study would explore a situation in whichpostponement is applied only to a few selected designfeatures in a building system, while allowing design ofother features in that same system to start earlier.

Our numerical simulation experiments have otherlimitations. These experiments cannot guarantee that aspecific postponement lag that performs best on averagewill perform best for a given real world project, even ifthe process model is precisely calibrated. Simulationresults average a large number of realizations. In con-trast, in the real world, decision-makers have to decidewhether to postpone decisions without knowing for sureif external events will happen and cause significant de-sign criteria changes, even if they anticipate these arelikely to occur. Project organizations may nonethelesswish to consider our findings when deciding whether topostpone—and for how long—critical design decisions.

8 Concluding remarks

The main insight of this research on large-scale engi-neering design projects is that some decision-makingpostponement (when design criteria are uncertain in theearly project stages due to external events) can helpincrease the predictability of concept developmentduration and reduce resources spent in design withoutincreasing the risk that the project will overrun a com-pletion date. This insight may help designers and cus-tomers in certain project environments to think abouthow long they can afford to postpone design decisions asa function of: (1) the last possible moment when conceptdevelopment should be completed; (2) the risk they arewilling to incur of overrunning that completion date; (3)the resources they can afford to spend; and (4) the extentto which they are able to reuse design work after anysignificant change in design criteria.

In the case of fab projects, these results matterbecause the completion date of the fab design phase is a

critical project milestone. Once a customer and adesigner contractually agree on a completion date, thecustomer project manager very seldom allows the de-signer to slip that milestone, regardless of any latechanges requested by the customer. As a result, design-ers frequently have to resort to working overtime andunder intense pressure to meet contractual milestones ifmajor external changes indeed occur. This damages thejob reputation and may contribute to scarcity of skilleddesigners wanting to work in high-tech projects. Inter-estingly, the designers we interviewed expressed convic-tion that postponement would jeopardize their ability tomeet project milestones even as they acknowledged thatthey repeated the same tasks several times because ofcustomer-requested changes. This work shows, however,that in certain project environments, some degree ofpostponement may actually help designers to meet thecustomer requirements. Fab designers and customersmay want to consider these findings before committingto when the fab design will be completed and, accord-ingly, to allocating the necessary resources.

Acknowledgements We gratefully acknowledge several grants thatfunded this research: grant SBR-9811052 from the National Sci-ence Foundation, and grants awarded to Mr. Nuno Gil from theFundacao para a Cie�ncia e Tecnologia (FCT) and from the Fun-dacao Luso-Americana para o Desenvolvimento (FLAD). Anyopinions, findings, conclusions, or recommendations expressed inthis report are those of the authors and do not necessarily reflectthe views of the foundations. The authors thank two anonymousreviewers for their useful comments and suggestions.

References

Bhattacharya S, Krishnan V, and Mahajan V (1998) Managingnew product definition in high velocity environments. ManageSci 44(11):50–64

Chase RB, Aquilano NJ, Jacobs FR (1998) Production and oper-ations management: Manufacturing and services. McGraw-Hill, New York

Clark KB, Fujimoto T (1991) Product development performance:Strategy, organization, and management in the world autoindustry. Harvard Business School Press, Boston, MA

Eisenhardt MK, Tabrizi BN (1995) Accelerating adaptive pro-cesses: Product innovation in the global computer industry.Adm Sci Q 40:84–110

Eppinger SD, Whitney DE, Smith R, Gebala D (1994) A model-based method for organizing tasks in product development. ResEng Des 6(1):1–13

Gil N (2001) Product-process development simulation to supportspecialty contractor involvement in early design. Dissertation,Civil and Environmental Engineering, University of Californiaat Berkeley

Ha AY, Porteus EL (1995) Optimal timing of reviews in concurrentdesign for manufacturability. Manage Sci 41(9):1431–1447

Hopp WJ, Spearman ML (1996) Factory physics: Foundations ofmanufacturing management. McGraw-Hill, New York

Iansiti M (1995) Shooting the rapids: Managing product develop-ment in turbulent environments. Calif Manage Rev 38(1):37–58

Jacome M, Peixoto H, Royo A, Lopez J (1999) The design spacelayer: Supporting early design space exploration for core-baseddesigns. In: Design, automation and test in Europe (DATE ‘99),Germany, pp 676–685

Jick TD (1979) Mixing qualitative and quantitative methods: Tri-angulation in action. Adm Sci Q 24 Dec(4):602–612

153

Klein B, Meckling W (1957) Application of operations research todevelopment decisions. Presented at the twelfth national meet-ing of the Operations Research Society of America, Pittsburgh,Pennsylvania, 15 Nov

Krishnan V, Eppinger SD, Whitney DE (1997) A model-basedframework to overlap product development activities. ManageSci 43(4):437–451

Krishnan V, Bhattacharya S (2002) Technology selection andcommitment in new product development: The role of uncer-tainty and design flexibility. Manage Sci 48(3):313–327

Law AM, Kelton WD (2000) Simulation modeling and analysis.McGraw-Hill, New York

Loch CH, Terwiesch C (1998) Communication and uncertainty inconcurrent engineering. Manage Sci 44(8):1032–1048

Lottaz C, Clement D, Faltings B, Smith I (1999) Constraint-basedsupport for collaboration in design and construction. J ComputCiv Eng ASCE 13(1):23–35

Perry C, Greig ID (1975) Estimating the mean and variance ofsubjective distributions in PERT and decision analyses. ManageSci 21(12):1477–1480

Pietroforte R (1997) Communication and governance in thebuilding process. Constr Manage Econ 15:71–82

Schruben DA, Schruben LW (1999) Event graph modeling usingSIGMA. Custom Simulations. http://www.customsimula-tions.com. Cited 10 Dec 2002

Smith RP, Eppinger SD (1997a) A predictive model of sequentialiteration in engineering design. Manage Sci 43(8):1104–1121

Smith RP, Eppinger S (1997b) Identifying controlling features ofengineering design iteration. Manage Sci 43(3):276–292

Terwiesch C, Loch CH (1999) Measuring the effectiveness ofoverlapping development activities. Manage Sci 45(4):455–465

Thomke S, Reinertsen D (1998) Agile product development:Managing development flexibility in uncertain environmentsCalif Manage Rev 41(1):8–30

Tommelein ID, Levitt RE, Hayes-Roth B, Confrey T (1991)SightPlan experiments: Alternate strategies for site layout de-sign. J Comput Civ Eng ASCE 5(1):42–63

van Hoek RI (2001) The rediscovery of postponement: A literaturereview and directions for research. J Oper Manage 19:161–184

Ward A, Liker JK, Cristiano JJ, Sobek DK II (1995) The secondToyota paradox: How delaying decisions can make better carsfaster. Sloan Manage Rev Spring:43–61

Wood SC (1997) The impact of tool delivery times on the optimalcapacity and value of semiconductor wafer fabs. In: Proceed-ings of the IEMT symposium, Austin, Texas, 13–15 Oct

Yassine A, Joglekar N, Braha D, Eppinger S, Whitney D (2003)Information hiding in product development: The design churneffect. Res Eng Des 14(3):31–144. DOI 10.1007/s0016300300362

154