See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/220164909

Grand Challenges in Modeling and Simulation of Complex Manufacturing

Systems

Article in SIMULATION: Transactions of The Society for Modeling and Simulation International · September 2004
DOI: 10.1177/0037549704044324 · Source: DBLP

CITATIONS READS
281 4,675

2 authors:

J.W. Fowler Oliver Rose

Arizona State University University of the Bundeswehr Munich
300 PUBLICATIONS 9,182 CITATIONS 201 PUBLICATIONS 3,436 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Oliver Rose on 15 May 2014.

The user has requested enhancement of the downloaded file.

 SIMULATION
http://sim.sagepub.com

Grand Challenges in Modeling and Simulation of Complex Manufacturing Systems

John W. Fowler and Oliver Rose
SIMULATION 2004; 80; 469
DOI: 10.1177/0037549704044324

The online version of this article can be found at:

http://sim.sagepub.com/cgi/content/abstract/80/9/469

Published by:

http://www.sagepublications.com

On behalf of:

Society for Modeling and Simulation International (SCS)

Additional services and information for SIMULATION can be found at:

Email Alerts: http://sim.sagepub.com/cgi/alerts

Subscriptions: http://sim.sagepub.com/subscriptions

Reprints: http://www.sagepub.com/journalsReprints.nav

Permissions: http://www.sagepub.com/journalsPermissions.nav

Downloaded from http://sim.sagepub.com at PENNSYLVANIA STATE UNIV on February 5, 2008

© 2004 Simulation Councils Inc.. All rights reserved. Not for commercial use or unauthorized distribution.
Grand Challenges in Modeling
and Simulation of Complex
Manufacturing Systems
John W. Fowler
Department of Industrial Engineering
Arizona State University
Tempe, AZ 85287-5906
[email protected]

Oliver Rose
Lehrstuhl für Informatik III
Am Hubland
D-97074 Würzburg
[email protected]

Even though we have moved beyond the Industrial Age and into the Information Age, manufacturing
remains an important part of the global economy.There is a need for the pervasive use of modeling and
simulation for decision support in current and future manufacturing systems, and several challenges
need to be addressed by the simulation community to realize this vision. First, an order of magnitude
reduction in problem-solving cycles is needed. The second grand challenge is the development of
real-time, simulation-based problem-solving capability. The third grand challenge is the need for true
plug-and-play interoperability of simulations and supporting software. Finally, there is the biggest
challenge facing modeling and simulation analysts today: that of convincing management to sponsor
modeling and simulation projects instead of, or in addition to, more commonly used manufacturing
system design and improvement methods such as lean manufacturing and six sigma.
Keywords: Modeling, simulation, manufacturing systems, real-time simulation, interoperability

1. Manufacturing Systems Models are intended to support management decisions

about the system, and a single model often will not be
Modern high-technology manufacturing systems, such as capable of supporting all decisions. Rather, different de-
those in the electronics, semiconductor, aerospace, and au- cisions require different models because various aspects
tomotive industries, can be extremely complex. The com- of the design and operation of the system will be impor-
plexity of these systems is due to factors such as multiple tant for the questions being asked of the model. While
part types made in the same facility/line, numerous man- spreadsheet and queuing models are useful for answer-
ufacturing steps (300-500 steps is not uncommon), batch ing basic questions about manufacturing systems, discrete
processing, very complex equipment that leads to high lev- event simulation models are often needed to answer de-
els of preventive maintenance and downtime, and multiple tailed questions about how a complex manufacturing sys-
levels of subassemblies, just to name a few. This complex- tem will perform [1]. Simulation models lend themselves
ity, combined with the high cost of setting up and maintain- to incorporating additional details about the manufacturing
ing such a system, necessitates the use of formal models of system and therefore often give more accurate estimates of
the system, rather than just relying on experience or sim- manufacturing system behavior than the simpler models
ple rules of thumb for performance evaluation and decision mentioned above, but usually at the cost of more computa-
making. tion. In general, simulation is a practical methodology for
understanding the high-level dynamics of a complex man-
| ufacturing system. According to Yücesan and Fowler [2],
| simulation has several strengths, including the following:
SIMULATION, Vol. 80, Issue 9, September 2004 469-476 |
© 2004 The Society for Modeling and Simulation International |
• Time compression—the potential to simulate years of real
DOI: 10.1177/0037549704044324 system operation in a much shorter time.

Downloaded from http://sim.sagepub.com at PENNSYLVANIA STATE UNIV on February 5, 2008

© 2004 Simulation Councils Inc.. All rights reserved. Not for commercial use or unauthorized distribution.
 Fowler and Rose

• Component integration—the ability to integrate complex A. Model Design

system components to study their interactions.
• Risk avoidance—hypothetical or potentially dangerous 1. Identify the issues to be addressed.
systems can be studied without the financial or physical
risks that may be involved in building and studying a real 2. Plan the project.
system.
• Physical scaling—the ability to study much larger or 3. Develop the conceptual model.
smaller versions of a system.
• Repeatability—the ability to study different systems in B. Model Development
identical environments or the same system in different
environments. 1. Choose a modeling approach.
• Control—everything in a simulated environment can be
2. Build and test the model.
precisely monitored and exactly controlled.
3. Verify and validate the model.

C. Model Deployment
2. Grand Challenges
1. Experiment with the model.
A grand challenge is a problem that (1) is difficult, with
the solution requiring one or more orders-of-magnitude 2. Analyze the results.
improvement in capability along one or more dimensions;
(2) should not be provably insolvable; and (3) has a solu- 3. Implement the results for decision making.
tion that results in a significant economical and/or social
impact. We see three grand challenges in the modeling Opportunities for efficiency improvements for each
and simulation of complex manufacturing systems. Each phase will be discussed below.
of them will be discussed below, with the grandest chal-
lenge discussed first. 2.1.2 Model Design
The model design phase is a very important but often over-
2.1 Grandest Challenge 1: An Order-of-Magnitude looked part of the simulation process. In this phase, the
Reduction in Problem-Solving Cycles project participants are identified, the project goals clearly
delineated, and the basic project plan developed. Chance,
It currently takes too long to design, collect informa- Robinson, and Fowler [1] describe project management
tion/data, build, execute, and analyze simulation models techniques that can significantly improve this portion of
to support manufacturing decision making. This leads to a the process and the likelihood of having an overall suc-
smaller number of analysis cycles than is desirable. While cessful project.
there are opportunities for efficiency improvements in all If these activities are not done well, the model developed
phases of the simulation process, we particularly see oppor- will likely be a very detailed model. While incorporation of
tunities to reduce the time needed to collect and synthesize detail may increase the credibility of the model, excessive
the required information and data, as well as opportuni- levels of detail may render a model hard to build, debug,
ties to reduce the time to carry out the experimentation. understand, deploy, and maintain. The determination of
The reduction in information/data collection and synthesis how much detail to add to the model is a primary goal
time can partially be achieved by proactive data analysis of the design stage. Experienced simulationists seem to
and by instilling better factory discipline in maintaining instinctively know the proper amount of detail. Perhaps
current information systems. The reduction in experimen- an expert system or at least a system that keeps track of
tation time can be approached from a number of different previous analyses and the level of detail used could be
angles, including exploring models of reduced complexity developed to assist in the model design phase.
that still give high-fidelity results, employing variance re- The issue of determining the appropriate level of detail
duction techniques, and using distributed and parallel sim- becomes even more important when one models a manu-
ulation. We discuss this further using the overall process facturing supply chain. While a lot of work has been done
for simulating manufacturing systems as a guide. to model operations at the machine and factory levels, con-
siderably less has been done at the supply chain level. For
2.1.1 The Simulation Process for Manufacturing the most part, models are used at just one of these levels,
Systems Analysis and little information is shared between them. In the future,
it will become increasingly necessary for these models to
The process of simulating manufacturing systems involves be used in conjunction with one another. To do this, several
the following phases and steps (modified from Chance, key questions remain: what is the right level of abstraction
Robinson, and Fowler [1] and Yücesan and Fowler [2]): for each model? Can parallel and distributed simulation

470 SIMULATION Volume 80, Number 9

Downloaded from http://sim.sagepub.com at PENNSYLVANIA STATE UNIV on February 5, 2008

© 2004 Simulation Councils Inc.. All rights reserved. Not for commercial use or unauthorized distribution.
 MODELING AND SIMULATION OF COMPLEX MANUFACTURING SYSTEMS

capabilities be employed? What is the right way to share of approach can make a large difference in the subsequent
information between the levels? model building and model execution times.
While manufacturing was one of the earliest simulation Most conventional simulation software packages used
application areas [3], simulation of manufacturing supply for modeling manufacturing systems take a “job-driven”
chains is just starting to become widely used. Supply chains worldview (also called a process interaction worldview).
can be very complex, spanning multiple manufacturing In this approach, manufacturing jobs are the active system
sites in various locations around the globe. For simplicity, entities while system resources, such as machines, are pas-
most researchers have limited the scope of their analyses sive. The simulation model is created by describing how
to only a few broad categorical links in the chain and have jobs move through their processing steps, seizing available
taken an aggregate view of the supply chain, making vari- resources whenever they are needed. A separate record for
ous assumptions about the internal workings of each chain every job in the system is created and maintained for track-
member. Supply chain analyses are often restricted to a ing wafers or lots through the factory. A lot of execution
single representative product. time can be consumed when sorting lists of these jobs in
Barker [4] indicates that it is the lack of models a given queue or in searching the queue for a given job.
and frameworks for analyzing the value-adding capabil- Therefore, the speed and space complexity of these simu-
ity throughout the supply chain that represents the greatest lations must be at least on the order of some polynomial
weakness in our supply chain knowledge base and current of the number of jobs in the factory.
literature. The dearth of supply chain models is not indus- An alternative simulation methodology focuses on re-
try specific but rather seems endemic to all industries due source cycles (see Schruben and Roeder [10] for more de-
to a lack of overall comprehension of the intricacies of the tails). In a “resource-driven” simulation, individual jobs
supply chain. are passive and are “moved” or “processed” by active sys-
Traditionally, simulation models of these systems have tem resources such as machines and operators. Rather than
been either (1) a discrete event simulation model that tracks maintaining a record of every job in the system, only in-
lots through each factory in the supply chain by consider- teger counts of the numbers of jobs of particular types at
ing the queuing at various work centers or (2) a high-level different steps are necessary. The state of the system is de-
continuous simulation model that does not track individual scribed by the status of resources and these job counts, and
lots through the factories but simply considers the gross the execution speed and memory footprint does not change
output of each factory. The first approach can be quite significantly as the system becomes more congested.
accurate, but it generally takes a long time to build the The events in a resource-driven simulation involve sim-
model, and the execution of the model is extremely slow, ple elementary integer operations, typically incrementing
making the exploration of many different scenarios pro- or decrementing job counts and numbers of available re-
hibitive. Umeda and Jones [5] present an example of this sources. Very large and highly congested queuing networks
type of model. Lendermann et al. [6] discuss a distributed can be modeled this way with a relatively small, finite
simulation approach to detailed (high-fidelity) modeling set of integers. In addition to their simplicity, resource-
of semiconductor manufacturing supply chains. This will driven simulations have a number of advantages over con-
be discussed further in the third grand challenge. ventional job-driven process flow models that describe the
Models of the second type can generally be built fairly paths of individual jobs. Resource-driven simulations of
quickly and their execution is much faster, but a large highly congested systems have been created that execute
amount of accuracy is generally lost. Heita [7] presents orders of magnitude faster than corresponding job-driven
an example of this type of model. Jain et al. [8] examine process flow simulations [10].
the trade-off between these two approaches. Duarte et al. The hope is that resource-driven factory simulations can
[9] discuss an effort that combines the two approaches by be developed that are able to provide much of the same in-
continuing to track lots through the factory but treating formation as job-driven simulators while executing many
the daily output of the factory in a somewhat gross fash- times faster. The expectation is that the two types of fac-
ion. Preliminary results comparing the accuracy and speed tory models can be used together. For example, high-speed,
of the approach to the traditional approaches are quite resource-driven factory simulators would be used for large-
promising. scale experiments that identify key opportunities for im-
provement. These opportunities can then be studied with
detailed job-driven simulators.
2.1.3 Model Development
After determining the modeling approach, the next step
After the preliminary work is done and the conceptual is to build the model and to determine the appropriate distri-
model designed, the next phase is to develop the model. butions to use for various system elements. Model building
This involves choosing the modeling approach, building can be very time-consuming and is an area where there po-
the model, and doing verification and validation of the tentially is room for significant efficiency improvements.
model. While there typically is not a lot of efficiency to There is a rich history of efforts to improve the efficiency
be gained in choosing the modeling approach, the choice of the model-building process, with much of this work led

Volume 80, Number 9 SIMULATION 471

Downloaded from http://sim.sagepub.com at PENNSYLVANIA STATE UNIV on February 5, 2008

© 2004 Simulation Councils Inc.. All rights reserved. Not for commercial use or unauthorized distribution.
 Fowler and Rose

by the software community. In the early days of comput- such as interarrival times, processing times, and so forth.
ing, most simulations of manufacturing systems used an Today, when data from the actual system are available, this
event orientation and were written in assembly language is typically done manually using software packages such as
or a high-level programming language such as FORTRAN. BestFit [13] or ExpertFit [14]. However, it may be possible
This led to models that were relatively efficient to execute to also automate this activity. Both the automatic genera-
because only the events desired were modeled. However, tion of the simulation model and the automated determi-
each new model development effort led to new model code nation of distributions should provide significant speed-up
being written (i.e., there was little reusability of code). possibilities in the overall time required to build a simula-
In the next wave, simulation languages were developed tion model.
to speed up the modeling process by allowing reuse of The automatic generation of the simulation model
simulation constructs. These languages included GPSS, should also reduce the time needed to verify a model by
SIMSCRIPT, GASP, SLAM, SIMAN, and so on. In these reducing the time required to debug the code. The time
languages, very basic logic/functionality was included in required to perform verification for some manufacturing
each construct. In addition, simulation overhead activ- systems has been reduced by the use of animation. Valida-
ities, such as managing the event calendar, generating tion, on the other hand, generally requires expert opinion,
(pseudo)-random numbers and random variates, and col- so there is probably not a lot that can be done to reduce this
lecting statistics, were included as a part of the package. time other than to make sure that the output of the model is
The academic community played a major role in determin- in a format that facilitates understanding the performance
ing efficient ways to accomplish the simulation overhead of the model.
activities.
The widespread availability of microcomputers in the 2.1.4 Model Deployment
early to mid-1980s led to an increased use of sim-
ulation to model manufacturing systems. It also led In the model deployment phase, the main area for efficiency
the simulation software community to develop two re- improvement is in executing the model. While simulations
lated but somewhat different approaches to making of some manufacturing systems may not take much time to
model building more efficient. The first of these was run, models of complex manufacturing systems may take
the development of user interfaces that allow multiple several hours for a single replication, particularly when de-
base constructs to be combined into higher level con- tails of automated material handling systems (AMHS) are
structs such as QUEUE-SEIZE_RESOURCE-DELAY- included. Shikalgar, Fronckowiak, and MacNair [15] indi-
FREE_RESOURCE to model a workstation. ARENA was cate that the average time for a 160-day simulation run of
perhaps the most successful of these systems. The second an IBM wafer fabrication facility with AMHS is approxi-
approach was the development of simulators. In a simula- mately 24 hours, but removing the AMHS system may re-
tor, the model is already built, and the user simply supplies duce this to 3 to 4 hours. Mercier, Bonnin, and Vialletelle
the appropriate data. Obviously, simulators can save con- [16] report that a single replication of their wafer fabri-
siderable model-building time if the underlying model of cation model takes 1 to 2 hours of CPU time. Anecdotal
the simulator can adequately model the system being stud- evidence suggests that these runtimes are not uncommon
ied.AutoSchedAP and Factory Explorer are two simulators for models of wafer fabrication facilities. Clearly, simula-
that are used to model semiconductor manufacturing sys- tion runtimes of the magnitudes mentioned above limit the
tems. A nice overview of the history of simulation language number of problem-solving cycles that can be achieved.
development can be found in Nance [11]. While all of the In fact, the IMTI Integrated Manufacturing Technology
developments described above have significantly reduced Roadmapping Project [17] report on modeling and simu-
the time to build some models, there is still considerable lation identifies that a key goal is the need for “fast, accurate
room for improvement. exploration of many more product and process design op-
One potential opportunity for improvement is in using tions” and the “requirement for simulation techniques that
existing sources of data available about a particular manu- enable complex simulations to run orders of magnitude
facturing system. Indeed, some analysts have done this to faster and more cost effectively than today.”
some degree already. In many industries, it is very com- As mentioned above, the modeling approach (job driven
mon for the process-routing information to be kept elec- vs. resource driven) may make a difference here. In ad-
tronically in the manufacturing execution system (MES). dition, the use of variance reduction techniques (VRTs)
This information can sometimes be extracted to populate offers some potential savings (see Law and Kelton [18,
a simulator, or a simulation model could be built automat- chap. 11] for more information onVRTs). While these tech-
ically from the data using model generation techniques. niques have successfully been used in some manufacturing
Ozdemirel, Mackulak, and Cochran [12] developed an ex- simulations, we believe their potential has not been fully
pert system called MASCOT that generated a model based exploited. Specifically, the automated use of common ran-
on these data and the system performances measures of in- dom numbers and antithetic variates could be better sup-
terest. Existing data sources can also be used to determine ported by simulation packages. Finally, parallel simulation
the appropriate distribution to use for random processes offers some hope for reduced simulation execution times

472 SIMULATION Volume 80, Number 9

Downloaded from http://sim.sagepub.com at PENNSYLVANIA STATE UNIV on February 5, 2008

© 2004 Simulation Councils Inc.. All rights reserved. Not for commercial use or unauthorized distribution.
 MODELING AND SIMULATION OF COMPLEX MANUFACTURING SYSTEMS

[19], particularly when modeling the manufacturing sup- based solution time requirements for the operational prob-
ply chain. lem are considerably more aggressive (shorter) than for
The time to analyze the results of the simulation can classical approaches. However, providing accurate results
be shortened by automatic generation of graphs, charts, is still important for the results to be accepted. Because of
confidence intervals, paired t-tests, and so forth. However, this time challenge, we see two ways to realize real-time
ultimately, human judgment is critical, and this part of the simulation capabilities:
analysis process probably does not provide an opportunity
for significant time savings. • use of a simulation model that is permanently running and
synchronized to the factory,
• automated building of a model from the factory databases.
2.2 Emerging Grand Challenge 2: Development
of Real-Time Simulation-Based Problem-Solving
Capability 2.2.1 Permanent, Always-On, Synchronized
Currently, most simulation models are used in single Factory Models
projects for tactical or strategic decision support (i.e., with These models mimic the behavior of the real factory. The
a long time horizon). Often, simulation models are used to models are continuously updated and synchronized with
plan equipment purchases or to evaluate planned changes factory data. If an additional model is required for decision
of the material flow control. To build these models from making, a clone is generated from the prototype model. The
scratch requires considerable effort. It is not uncommon main problem of this approach is the availability of on-time
that the models are not used again after the decision is and correct data from the factory. Although announced for
made. Even when this is the case, the time and money years, all-in-one factory databases do not exist yet. Due to
spent on these traditional simulation projects leads, in most database infrastructures that grew (evolved) over several
cases, to a large return on investment [1]. years and the fact that the single databases are implemented
Relatively little is known, however, about approaches to in different (sometimes incompatible) software products,
use simulation for operational (real-time) decisions in man- it requires considerable effort to obtain up-to-date, con-
ufacturing. Simulation models for this purpose are possible sistent, and useful data. The different database standards
today due to the increased amount of data and information and access techniques lead to a wide variety of interfaces
that is collected and maintained by current shop-floor in- between simulation software and factory databases. Often,
formation systems. The literature on real-time applications these databases are not physically in one location, and data
of simulation in manufacturing is sparse, but there are pa- transfer times of several minutes may arise for updating the
pers about simulation-based scheduling [20-22], order re- factory model. Even if the interface issues between simu-
lease [23], forecasting [10], and exception management lator and factory databases are solved, the updates of the
[24]. In Peters et al. [21], the focus is on the requirements databases from the factory workstations and MES have to
for an online simulation tool, whereas Rogers and Gor- be considered. The data collection, processing, and transfer
don [22] provide an overview of the principal problems of capabilities of the tools range from very basic to highly so-
simulation-based scheduling. phisticated. Some only report machine states, while others
In this context, real-time, simulation-based problem- provide all types of information a simulation model needs.
solving capability means that if the status of the factory As a result, the following problems have to be solved:
changes abruptly, we can run a simulation nearly instan-
taneously to decide about appropriate actions to be taken. • A clear definition of the term factory state is missing: be-
For example, when a key piece of manufacturing equip- fore we can start to collect data, we have to know exactly
ment fails, plans for minimizing the impact of the failure what kind of data we require to obtain a clear picture of
need to be developed. Simulations of various recovery al- the factory.
ternatives could help determine the best course of action. • There is a lack of data: data required by the simulation
Another example deals with batch operations in wafer fab- are not available and cannot be generated in an automated
rication. In this case, when there is a partial batch available way from other factory information.
to begin processing, a decision needs to be made whether • The data quality is too low: for example, the simulation
model requires a histogram, and the tool can only provide
to start the partial batch now or to wait for future arrivals. average values of a parameter of interest.
Fowler, Hogg, and Phillips [25] describe non-simulation-
• The update frequency is too low: the tool can only generate
based techniques that can be helpful in making this deci- data reports after a given and long time interval (say, once
sion, but the capability to generate a set of simulations to a day).
see what might happen for either decision (start now or
wait) could be extremely beneficial since these decisions It depends on the objective of the simulation-based
have a major influence on the performance of the system. problem solver whether these problems have an effect on
Due to the need for a quick response to the types of the results. For instance, if only a coarse overview on the
problems described above, model-building and data collec- factory behavior for the next few days is required, the prob-
tion times must be very short. In particular, the simulation- lems have only a minor impact on the results. If, however,

Volume 80, Number 9 SIMULATION 473

Downloaded from http://sim.sagepub.com at PENNSYLVANIA STATE UNIV on February 5, 2008

© 2004 Simulation Councils Inc.. All rights reserved. Not for commercial use or unauthorized distribution.
 Fowler and Rose

the intention is to provide a simulation-based scheduling approach is more appropriate because cloning from a mas-
system, the factory data have to be very accurate, and all of ter copy can be done in seconds. To simplify and speed up
the issues above must be solved, or there must be at least an the model generation, the required data should not be stored
estimation of the error in the schedule induced by wrong, in the databases in raw format but preprocessed and ready
missing, or late data. to use for the simulation package. For instance, it causes
To mimic the real factory’s behavior, the simulation unnecessary delays if machine availability histograms or
model must have appropriate MES functionality. This can distributions have to be computed when the model is being
be achieved either by using a copy of the real MES with built. This can be done proactively right after each database
interfaces to the simulator or by rebuilding the MES inside update from the factory.
the simulator. The second approach is time-consuming and
error prone because most MES software companies will
not be willing to provide detailed information about the 2.3 Emerging Grand Challenge 3: True Plug-and-
implementation of their products. Therefore, the rebuilt Play Interoperability of Simulations and Supporting
MES will only behave approximately like the real one. Software within a Specific Application Domain
If all data problems are solved, we have to find a simu-
lation package that supports real-time capabilities, such If we begin to have persistent models of the manufacturing
as Arena RT from Rockwell Software. At the moment, system, it is likely that there will actually be models of
however, there is a lack of off-the-shelf software that can many different subsets of the factory. These models will
be used for large-scale manufacturing simulation models. need a seamless way to interact with each other. In addi-
During recent years, more effort has been spent trying to tion, more and more of the information/data will be pro-
speed up simulation engines than to try to run them in real vided from other manufacturing support software such as
time. In addition, the models must have the ability to be the MES, available-to-promise systems, analysis software
synchronized with the factory state. The factory state not packages, and so on. It will be increasingly more important
only summarizes tool states and lot positions but also must for all of these systems to be able to quickly communicate
comprehend adding or removing equipment. The persis- with each other and the outside world in an unambiguous
tent, constantly synchronized factory simulation model is way. The High-Level Architecture (HLA) is one partial
the master copy for all simulation models required for de- solution to this challenge.
cision making. As soon as an operational problem arises, The HLA is a general-purpose architecture for simula-
a clone is created from the master copy, and the analy- tion reuse and interoperability. In the HLA, each simula-
sis is only based on this clone. At the moment, there does tion or other software system is run as a separate federate
not appear to be a simulation package that provides this (process), and the collection of all federates is called a fed-
feature. eration. Each federate can be developed independently and
implemented using different software languages and dif-
2.2.2 On-Demand, Automatically Built Factory ferent hardware platforms. Thus, faster cycles of analysis
Models may occur due to less time spent developing new inte-
grated software and less execution time due to the distri-
Taking into account the problems discussed in the previ- bution of the software on multiple processors. The HLA
ous section, particularly the lack of real-time simulation was developed under the leadership of the U.S. Defense
software packages, it is worthwhile to consider a different Modeling and Simulation Office (DMSO) to support reuse
approach. Instead of generating clones from a factory sim- and interoperability across a wide range of different types
ulation model that has to be perpetually synchronized, the of simulations. The HLA Baseline Definition was com-
experimenter generates the model on demand directly from pleted in 1996 and was adopted as the Facility for Dis-
the factory databases. In this case, the requirements for the tributed Simulation Systems 1.0 by the Object Manage-
factory databases remain the same, but the simulation soft- ment Group (OMG) in 1998. The HLA was approved as an
ware can be far less sophisticated. Most of the simulation open standard through the Institute of Electrical and Elec-
packages on the market today facilitate building models tronic Engineers (IEEE)—IEEE Standard 1516—in 2000.
from formatted text files. Thus, we only need an application Unfortunately, the software community has been slow to
that retrieves the current factory data from the databases incorporate the HLA in their product offerings. This is due
and then transforms them into simulation model files. Be- primarily to lack of pull from their users, which is at least
cause considerable amounts of factory data, including tool in part due to a lack of understanding what the HLA can
sets and their properties, current and future orders, prod- provide.
uct specifications and routing, and so forth, have to be re- However, some users have applied the HLA to study
trieved and transferred, it may take considerable time to their manufacturing systems. Schumann et al. [26] present
finish the model-building process. As a consequence, this an example of using the HLA for factory simulation. Len-
approach is less suited for time-critical decisions in which dermann et al. [6] discuss a distributed simulation proto-
the simulation results have to be available within minutes type of a semiconductor manufacturing supply chain based
or seconds after the problem occurred. In that case, the first on the HLA.

474 SIMULATION Volume 80, Number 9

Downloaded from http://sim.sagepub.com at PENNSYLVANIA STATE UNIV on February 5, 2008

© 2004 Simulation Councils Inc.. All rights reserved. Not for commercial use or unauthorized distribution.
 MODELING AND SIMULATION OF COMPLEX MANUFACTURING SYSTEMS

Straßburger et al. [27] discuss the fact that while the 4. References
HLA is helpful in the application of the distributed simu-
[1] Chance, F., J. Robinson, and J. Fowler. 1996. Supporting manufac-
lation of the “digital factory,” it is not sufficient. Specifi- turing with simulation: Model design, development, and deploy-
cally, they point out that it is unfortunate that (1) there are ment. In Proceedings of the 1996 Winter Simulation Conference,
no time-managed versions, (2) there is not a way to trans- San Diego, CA, pp. 1-8.
fer an object/attribute to a specific receipt, and (3) there is [2] Yücesan, E., and J. Fowler. 2000. Simulation analysis of manufactur-
no possibility of transferring the final state of an object re- ing and logistics systems. In Encyclopedia of production and man-
ufacturing management, edited by P. Swamidass, 687-97. Boston:
garding attribute values. They refer to papers by Fujimoto Kluwer Academic.
and Tacic [28], Sauerborn et al. [29], and Myjak et al. [30] [3] Naylor, T. H., J. L. Balintfy, D. S. Burdick, and K. Chu. 1996. Com-
that study these deficiencies but indicate that no solutions puter simulation techniques. New York: John Wiley.
have been implemented in the HLA interface specification. [4] Barker, R. C. 1996. Value chain development: An account of some
implementation problems. International Journal of Operations &
While it is not clear if the HLA will ultimately provide (or Production Management 16 (10): 23-36.
be a major part of providing) true plug-and-play interoper- [5] Umeda, S., and A. Jones. 1998. An integration test-bed system for
ability, it is clear that this functionality is needed as analysts supply chain management. In Proceedings of the 1998 Winter
strive to adequately model supply chain operations using Simulation Conference, pp. 1377-85.
the logic embedded in the software used in the real system. [6] Lendermann, P., N. Julka, B. P. Gan, D. Chen, L. F. McGinnis, and J. P.
McGinnis. 2003. Distributed supply chain simulation as a decision
support tool for the semiconductor industry. Transactions of the
Society for Modeling and Simulation International 79 (3): 126-38.
3. Big Challenge 4: Greater Acceptance of [7] Heita, S. 1998. Supply chain simulation with LOGSIM-simulator.
In Proceedings of the 1998 Winter Simulation Conference, pp.
Modeling and Simulation within Industry 323-26.
[8] Jain, S., C. Lim, B. Gan, andY. Low. 1999. Criticality of detailed mod-
We decided to label this challenge as a “big” challenge in- eling in semiconductor supply chain simulation. In Proceedings
of the 1999 Winter Simulation Conference, pp. 888-96.
stead of as a “grand” challenge because it is not really a [9] Duarte, B. M., J. W. Fowler, K. Knutson, E. Gel, and D. Shunk. 2002.
technical challenge but more of a social challenge. In fact, Parameterization of fast and accurate simulations for complex
this may make this the most difficult challenge of all the supply networks. In Proceedings of the 2002 Winter Simulation
challenges discussed in this paper. While the use of mod- Conference, pp. 1327- 36.
eling and simulation in manufacturing is steadily gaining [10] Schruben, L., and T. Roeder. 2003. Fast simulations of large-scale
highly congested systems. Transactions of the Society for Model-
acceptance for certain applications (such as capacity plan- ing and Simulation International 79 (3): 115-25.
ning), there is still a long way to go before it is commonly [11] Nance, R. E. 1993. A history of discrete event simulation program-
applied for a multitude of other applications. Currently, ming languages. Proceedings of the Second ACM SIGPLAN His-
modelers often spend much of their time convincing man- tory of Programming Languages Conference 23 (3): 149-75.
[12] Ozdemirel, N. E., G. T. Mackulak, and J. K. Cochran. 1993. A group
agement of the need for these services. Simulation is gen- technology classification and coding scheme for discrete manu-
erally only one of several manufacturing system design facturing simulation models. International Journal of Production
and improvement approaches that are presented to man- Research 33 (3): 579-601.
agement for possible implementation. Other approaches [13] Jankauskas, L., and S. McLafferty. 1996. BESTFIT, distribution fit-
include lean manufacturing, six sigma, just-in-time man- ting software by Palisade Corporation. In Proceedings of the 1996
Winter Simulation Conference, pp. 551-55.
ufacturing, total quality management, and so on, and the [14] Law, A. M., and M. G. McComas. 1989. Pitfalls to avoid in the sim-
simulationist should not try to convince management that ulation of manufacturing systems. Industrial Engineering 31:28-
simulation is better than these techniques. Indeed, simu- 31.
lation, by itself, does not improve the performance of a [15] Shikalgar, S. T., D. Fronckowiak, and E. A. MacNair. 2002. 300mm
wafer fabrication line simulation model. In Proceedings of the
manufacturing system. It is only by the use of the model to 2002 Winter Simulation Conference, pp. 1365-68.
answer specific questions about ways to change the system [16] Mercier, D., O. Bonnin, and P. Vialletelle. 2003. Achieving optimal
that realizable improvements are identified. Thus, simula- cycle time improvement in a 300mm semiconductor fab using
tionists should try to convince management that simulation dynamic simulation and design of experiements. Transactions of
is complementary to the other approaches mentioned above the Society for Modeling and Simulation International 79 (3):
171-79.
and that it can be used to assess the potential improvements [17] Integrated Manufacturing Technology Initiative. 2000. Integrated
that can be made to the system when the other approaches Manufacturing Technology Roadmapping Project: Modeling
are employed. Finally, simulationists must be careful to re- & Simulation. Integrated Manufacturing Technology Initiative
sist the temptation to oversell the use of a model’s results; (IMTI): Oak Ridge, USA.
[18] Law, A. M., and D. W. Kelton. 2000. Simulation modeling and anal-
this may be a good short-term strategy, but it can have very ysis. 3d ed. New York: McGraw-Hill.
negative long-term consequences if the expectations of the [19] Low, M.Y. H. 2002. Manufacturing simulation using BSP time warp
users of the model results are not met. Instead, simulation- with variable number of processors. In Proceedings of the 14th
ists should indicate that while a particular model could be European Simulation Symposium.
extended to help to answer numerous questions, including [20] Chong, S., A. Sivakumar, and R. Gay. 2002. Design, development
and application of an object oriented simulation toolkit for real-
those that it was not originally designed for, these questions time semiconductor manufacturing scheduling. In Proceedings of
should be tackled in a systematic and planned manner. the 2002 Winter Simulation Conference, pp. 1849-56.

Volume 80, Number 9 SIMULATION 475

Downloaded from http://sim.sagepub.com at PENNSYLVANIA STATE UNIV on February 5, 2008

© 2004 Simulation Councils Inc.. All rights reserved. Not for commercial use or unauthorized distribution.
 Fowler and Rose

[21] Peters, B., J. Smith, J. Curry, C. LaJimodiere, and G. Drake. 1996. [30] Myjak, M., S. Sharp, T. Lake, and K. Briggs. 1999. Object trans-
Advanced tutorial: Simulation-based scheduling and control. In fer in HLA. Paper no. 99S-SIW-140. In 1999 Spring Simulation
Proceedings of the 1996 Winter Simulation Conference, pp. 194- Interoperability Workshop, March.
98.
[22] Rogers, P., and R. Gordon. 1993. Simulation for real-time decision
making in manufacturing systems. In Proceedings of the 1993
Winter Simulation Conference, pp. 866-74. John W. Fowler is Professor of Industrial Engineering at Ari-
[23] Muller, D., J. Jackman, and C. Fitzwater. 1990. A simulation-based zona State University and is the Center Director for the Fac-
work order release mechanism for a flexible manufacturing sys-
tem. In Proceedings of the 1990 Winter Simulation Conference,
tory Operations Research Center that is jointly funded by the
pp. 599-602. National Science Foundation, International SEMATECH and
[24] Katz, D., and S. Manivannan. 1993. Exception management on a the Semiconductor Research Corporation. He received his PhD
shop floor using online simulation. In Proceedings of the 1993 from Texas A&M University and was previously a Senior Mem-
Winter Simulation Conference, pp. 888-96. ber of Technical Staff in the Modeling, CAD, and Statistical
[25] Fowler, J. W., G. L. Hogg, and D. T. Phillips. 2000. Control of
multiproduct bulk server diffusion/oxidation processes part two: Methods Division of SEMATECH. His research interests in-
Multiple servers. IIE Transactions on Scheduling and Logistics clude modeling, analysis, and control of manufacturing (es-
32 (2): 167-76. pecially semiconductor) systems. He is a member of ASEE,
[26] Schumann, M., E. Blümel, T. Schulze, S. Straßburger, and K.- C. Rit- IEEE, IIE, INFORMS, POMS, and SCSI. His web address is
ter. 1998. Using HLA for factory simulation. In Proceedings of
the 1998 Fall Simulation Interoperability Workshop, September,
http://www.eas.asu.edu/∼masmlab/fowler/.
Orlando, FL.
[27] Straßburger, S., A. Hamm, G. Schmidgall, and S. Haasis. 2002. Us- Oliver Rose is Professor for Modeling and Simulation in the De-
ing HLA ownership management in distributed material flow sim- partment of Computer Science at the Dresden University of Tech-
ulations. In Proceedings of the 2002 European Simulation Inter- nology, Germany. He received an MS degree in applied mathe-
operability Workshop, June, London. matics (’92) and a PhD degree in computer science (’97) from
[28] Fujimoto, R., and I. Tacic. 1999. Time management of unsynchro-
nized HLA services. Paper no. 99F-SIW-165. In 1999 Fall Simu-
the University of Würzburg, Germany. His research focuses on the
lation Interoperability Workshop, September. operational modeling, simulation and analysis of complex manu-
[29] Sauerborn, B., G. Tan, G. Moss, P. Oxenberg, F. Moradi, and R. facturing systems with applications in the semiconductor indus-
Ayani. 2000. HLA ownership management services: We almost try. He is a member of IEEE, INFORMS College on Simulation,
got it right. Paper no. 00F-SIW-076. In 2000 Fall Simulation In- and ASIM. His web address is www.iai.inf.tu-dresden.de/ms/.
teroperability Workshop, September.

476 SIMULATION Volume 80, Number 9

Downloaded from http://sim.sagepub.com at PENNSYLVANIA STATE UNIV on February 5, 2008

© 2004 Simulation Councils Inc.. All rights reserved. Not for commercial use or unauthorized distribution.

View publication stats