Proceedings of DETC03:

ASME 2003 Design Engineering Technical Conferences and

Computer and Information in Engineering Conference
Chicago, Illinois USA, September 2-6, 2003

DETC2003/CIE-48198
A GENERAL DECISION-MAKING METHOD FOR THE RAPID MANUFACTURING OF
CUSTOMIZED PARTS
Christopher B. Williams, Jitesh H. Panchal, and David W. Rosen1
Systems Realization Laboratory
G. W. Woodruff School of Mechanical Engineering
Georgia Institute of Technology
Atlanta, GA 30332-0405

ABSTRACT
In this paper we propose a general method for making
process parameter decisions in the development of a rapid
manufacturing system. Given a level of demand, the designer
can use our method to select the appropriate type of rapid
prototyping machine, the number of machines, the batch size,
and other process parameters (layer thickness, road width, etc.)
in order to achieve an ideal cost, throughput, and quality. We
illustrate our method through the application of Stratasys
Fused Deposition Modeling technology to rapidly manufacture
customized hearing aid shells. We close with a look ahead to a
larger problem: the use of our method to select the proper rapid
prototyping technology for use in rapid manufacturing.
KEYWORDS
Rapid Prototyping, Rapid Manufacturing, Mass Customization
1

INTRODUCTION

1.1 Geometric Customization of Products via Rapid

Prototyping
Mass customization is the production of customized
products at efficiencies and speeds close to those of mass
production [1]. Mass customized production involves building
a product in response to a particular customers desire, as
opposed to producing a generic product to be placed in an
inventory in hopes that it will later be purchased. It is
impossible to cost effectively cope with customers demands
for product variety through an increase in inventory;
manufacturing enterprises are recognizing that product and
process design present the best control over offering product
variety [1]. Whether customized through shape (i.e., fits

perfectly in the hand of the user), or customized in function, a

company cannot cost-effectively offer customization to the
point of personalization through traditional manufacturing
methods. CNC laser cutting is one of the most often used
processes for achieving customized geometry in parts for mass
customized production; however it can only offer customized
geometry in two dimensions.
The advent of rapid prototyping opens a new and exciting
method of offering variety for designers: the opportunity to
offer affordable customized geometry in three dimensions.
Rapid prototyping (RP), a layer-based additive manufacturing
system, gives a designer the power to build almost any
conceivable geometry. The designer simply generates a three
dimensional CAD model of the prototype. The model is then
sliced into layers via software, and then sent to the machine,
which constructs each cross-section until the part is complete.
While all RP technologies share the above in common,
their solution principles are vastly different. 3D Systems
Stereolithography (SLA) technology uses a laser to cure
photopolymer resin layer by layer. Stratasys Fused Deposition
Modeling (FDM) technology draws part cross-sections with a
nozzle that extrudes a heated polymer filament that cools and
solidifies with the previously deposited layer. The medium of
3D Systems Selective Laser Sintering (SLS) technology is a
powder that sinters together upon contact with the drawing
laser.
When compared to traditional hard tooling alternatives,
rapid prototyping offers a number of advantages in the quest to
create customizable geometry affordably:
Rapid prototyping gives the manufacturer the ability to
create multiple, different geometries in a single build. A
traditional casting process would require the creation of a
separate mold for each desired geometry. The RP engineer

Associate Professor of Mechanical Engineering and Corresponding Author

Phone: (404) 894-9668; Fax: (404) 894-9342; Email: david.rosen@me.gatech.edu

Copyright 2003 by ASME

is capable of producing many different parts in one build by

simply adding each of their 3D CAD models to the build
file.
Rapid prototyping does not require a change in tooling for
the creation of different geometries. Each technology uses a
central method in which to draw and fill the cross section
of each part. Unless the parts are of different materials,
many parts can be manufactured in one common build.
The desire to create customized geometries with rapid
prototyping does not require extra skilled operator input. To
manufacture different geometries with RP technology, the
operator is simply required to orient the part in the build
file.
Currently there are two industrial examples of the
application of RP technology to create geometrically masscustomized parts. Align Technologies [2] offers customized
clear braces. After obtaining a laser scan of the customers
dental clay impression, the company rapidly manufactures a
mold for each set of braces with SLA. An FDA approved
polymer is then thermoformed over the SLA molds to create
the braces. By rapidly manufacturing the molds, Align
Technologies is able to produce a new set of customized braces
every two weeks for thousands of customers around the globe.
Siemens and Phonak have also taken advantage of this
technology to create geometrically customized products [3].
Through the use of SLS technology, Siemens and Phonak
rapidly manufacture hearing aid shells that are customized to
each individual customers ear canals. After scanning a clay
impression of the ear canal, the technicians use CAD software
to create a hollow shell model (to be later filled with hearing
aid components), which is then sent to a SLS machine. The
result is a customized hearing aid that fits perfectly in the
customers ear.
1.2 The Intriguing Problem of Rapid Manufacturing
This concept of using rapid manufacturing to create masscustomized products presents an interesting problem. Not only
are there numerous product and process variables that must be
selected in order to deliver a personalized product within a
reasonable time, but these variables must be chosen to satisfy
the processs conflicting goals: minimum cost, maximum
quality, and maximum throughput. While engineers will want
to choose the process variables (batch size, number of
machines, etc.) that provide the minimum cost per part, an
industrial engineer will wish to design a process with a
maximum throughput for a given demand, while a RP engineer
will want to maximize the quality of the part. Cost, quality, and
throughput present major tradeoffs in this problem one cannot

Demand
Various Geometries
Various Colors

Various Materials
Capturing information about
the problem from scheduling
and manufacturing perspective

simultaneously maximize one without minimizing another (and

vice versa). A graphical overview of the overall problem is
provided below in Figure 1.
As a result of these numerous, coupled design decisions,
we present the question: How can a rapid manufacturing
process be methodically designed to provide a mass-customized
product to achieve a desired balance among cost, throughput,
and quality?
In this paper, we propose a general method for making
decisions in the process design of a customized product via
rapid manufacturing. In Figure 1 it is shown that the
progression from demand to scheduling to process planning is
iterative. In this paper, we investigate the scheduling and
process planning aspects of the rapid manufacturing product
development cycle and come up with a means for reducing the
iteration. In Section 2 we present the theoretical foundations of
our work. Our solution strategy for this type of problem is
presented in Section 3. The FDM process model and its
associated metrics are discussed in Section 4. We then attempt
to validate our solution method through an example of the rapid
manufacturing of hearing aid shells in Section 5. Results of
this example are presented in Section 6. We offer closure
through suggestions for an extension of this work and its
application to a broader problem in Section 7.
2

THEORETICAL FOUNDATION: THE UTILITYBASED COMPROMISE DECISION SUPPORT

PROBLEM

The coupled decisions that are encountered in this process

parameter selection method presented a major problem. Here
we present the theoretical foundation of our solution method:
the utility based compromise Decision Support Problem.
Compromise and selection Decision Support Problems
(DSP) are general formulations for decision problems that are
encountered in engineering design. The compromise DSP
(cDSP) is a general framework for solving multi-objective,
non-linear, optimization problems [4]. Mathematically, the
cDSP is a multi-objective decision model which is a hybrid
formulation based on Mathematical Programming and Goal
Programming [5]. It is used to determine the values of the
design variables, which satisfy a set of constraints and bounds
and achieve as closely as possible a set of conflicting goals.
The structure of the cDSP is as follows:
Given:

A feasible alternative, assumptions, parameter

values and goals.

Scheduling
(Industrial Engineer)

Manufacturing
(RP Engineer)

- Based on the demand, the

process and material are fixed.
-Find an appropriate batch size
for the production based on the
demand. This is based on the
assumed time and cost required
for build.
- Decisions:
- Machine type
- No. of machines
- Batch size

- Find an appropriate combination of

processing parameters for the machine
- For the given batch size and processing
parameters, find the time required to
build
- Send the information about time back to
the scheduling department so that they
can adjust the batch size

- Decisions:
- Layer thickness
- Road width
- Offsets

Process Planning

Figure 1: Graphical Representation of Problem

2

Copyright 2003 by ASME

Find:
Satisfy:

Values of design and deviation variables.

System constraints, system goals, and bounds on
variables
Minimize: Deviation variable that measures distance between
goal targets and design points
Seepersad et al. recently augmented the cDSP with utility
theory to create a multi-objective decision support model with
an objective function derived from utility theory [6]. The
utility-based compromise DSP is a decision support construct
that is based on utility theory and permits mathematically
rigorous modeling of designer preferences such that decisions
can be guided by expected utility in the context of risk or
uncertainty associated with the outcome of a decision. We
discuss the application of this theory in the next section.

SOLUTION STRATEGY

3.1 Problem Characteristics

To solve the problem of scheduling and manufacturing in a
rapid manufacturing system, we have broken up the process
into two events and their corresponding decisions. The two
events in the process are: scheduling and RP process parameter
selection. The decision associated with scheduling is the
selection of the number and type of machines and the
determination of desired number of parts per build. The
decision associated with RP process parameter selection event
is the selection of best combination of process parameters.
These decisions are shown in the graphical representation of the
problem (see Figure 1).
In both these decisions, some of the system parameters are
selected, which has an effect on the system goals. The system
parameters and goals in the overall process are listed in Figure
2.
Overall System Parameters:
1. Type of machine to be used
2. Number of machines to be used
3. Number of builds/Number of parts to be built
4. Layer thickness
5. Road widths
6. Scan Speeds
Overall System Goals:
1. Cost
2. Time
3. Quality
Figure 2: Overall System Parameters and Goals
The characteristics of the problem are:
1. Tradeoffs between goals
2. Coupling between decisions
These characteristics, and the manner in which we handle them
in our solution strategy, are discussed next.
3.2 Tradeoffs Between Goals
The system goals are to minimize cost, minimize time per
build, and to maximize quality. There is a tradeoff between
these goals. Increasing the number of parts per build leads to

an increase in the time required for manufacturing. Decreasing

the number of builds increases the cost per build. Similarly, for
the second decision, larger road width and layer thickness
imply a poor quality but a much lower build time.
In this paper, utility theory (see Section 2) is used for
quantifying the preferences for various goals and then
combining the preferences of different designers towards
different goals. A utility value, that reflects the preference of a
designer, is assigned for various values of goals. The higher the
utility value, the higher the designers preference towards the
goal and vice versa. Generally, a utility of 1 is assigned to the
most preferable value of the goal and a utility of 0 is assigned
to the least desirable value. Utility theory also helps in
incorporating attitude towards risk. A risk-prone attitude means
that a small change in the value of a goal from the most
preferable value results in a large reduction in the utility. A
risk-averse attitude means that a small change in the value of
goal results in a relatively small change in the utility.
Fernandez et al. have argued that utility theory can be used
for making decisions consistently [7]. Consistency in decisionmaking is important in the problem of RP process planning
because decisions are made at different stages in the process
and different people are involved in making the decisions.
The tradeoffs between these goals are considered in this
project using utility functions for each system goal. Separate
utility functions are assigned to each goal and the overall utility
is evaluated as a weighted average of the individual utilities.
The weighted average utility function is chosen based on the
assumption that the risk aversion for a goal is independent of
the values of the other goals (utility independence) and there is
no interaction between preferences for different goals (additive
independence). The details of the conditions under which this
simplified form can be employed are provided by [8]. It is
assumed that the engineer is risk averse to these goals. The
objective of the solution is now focused on maximizing this
overall utility so that all of our conflicting goals to their fullest
extent in a manner that is consistent with our preferences.
The decisions make use of these utility functions whenever
these types of tradeoffs between goals are present.
3.3 Coupling Between Decisions
The design decisions of this problem are coupled with one
another. For example, a decision about build size affects the
value of machine parameters, while a decision on machine
parameters affects the decision of appropriate build size. These
decisions are shown in Figure 1 along with the information
flowing from one decision to another.
To overcome this problem of coupling, we use a
combination of Response Surface Modeling (RSM) technique
and Compromise DSP. Our approach to decoupling decisions
is to construct response surfaces as surrogate models of design
spaces and then use these surrogate models in subsequent
decisions.
In Figure 3, we present a graphical view of the use of
response surfaces for decision-making. The details of the
solution strategy are presented in Section 3.4. For each
decision, a response surface model is fit between the decision
variables and the goals. For example, in the RP process
parameter selection decision (Decision 2), a response surface
model is fit between machine parameters (layer thickness, road
width and scan speed) and the goals (cost, build time). In the

Copyright 2003 by ASME

Forward

layer thickness, road

width, scan speed

Step 3

Step 2

Step 4

Parts per build, #

machines, machine type

Demand

Backward

goals (cost, time)

goals (cost, time)

Parts per build, #

machines, machine type

scheduling decision (Decision 1), a similar response surface

model is fitted between decision variables (machine type,
number of machines and parts per build) and goals. These
response surfaces from different decisions are used together to
come up with values of decision parameters that achieve a
balance among the cost, build time and quality. While the
response surfaces do not completely decouple the decisions,
they do help in capturing and combining the knowledge from
various domains.

layer thickness, road

width, scan speed

Parts per build, #

machines, machine type

Decision 1: Selection of machine,

number of machines and batch size

Step 1

Decision 2: Selection of layer

thickness, road width, scan speed

Figure 3: Graphical Representation of the Solution Strategy

The details of two decisions in the process are provided
next. Their word formulations are presented in Figures 4 and 5.
Decision 1: Find the number and type of machines to be used
Given:

Find:

Demand for different varieties of parts

Maximum number of machines that can
process different types of parts
K-Values for waiting time, build time, and cost goals
Utility function for waiting time, build time, and cost
Type of machine to be used
Number of machines to be used
Number of parts in a batch
Number of batches

Satisfy:
System Constraints:
(1) Time constraints (parts need to be produced within a
week of order arrival)
(2) Machine constraints (batch size is limited by the
machine capability)
(3) Cost constraints (cost per piece is not too expensive)
(4) The number of parts to be manufactured (all of the
demand needs to be met)
System Goals:
Minimize overall cost
Minimize production time
Maximize:
Expected Utility

Figure 4: Decision 1 cDSP

Decision 2: Find the machine parameters (layer thickness, road

width, etc.)
Given:

Number of parts to be manufactured

Machine type
Material type
Cost model
Quality model
Build time model
K-values for cost, quality, and time goals
Utility function for cost, quality, and time

Find:
Machine parameters
o Layer thickness
o Road widths
o Scan velocity
Satisfy:
System Constraints:
o Cost constraints (cost per piece is not too expensive)
o Machine constraints (batch size is limited by the machine
capability)
System Goals:
Minimize part cost
Minimize production time
Minimize quality metric
Maximize:
Expected Utility

Figure 5: Decision 2 cDSP

Decision 2 (D2)
For this decision, the inputs (i.e., parameters coming from
the previous decision) are the number of parts to be built, the
selected machine, and the material selected. In this decision,
we calculate the relation between the machine parameters and
the goals (cost, time, quality). For each combination of inputs,
we find the best values of machine parameters that will
maximize an expected utility function (a weighted sum of the
utility functions of cost, time and quality). Response surfaces
are then created establish a relation between these best values
and the system goals. These surfaces are then sent to the
previous stage (Decision 1).
Decision 1 (D1)
The second step in the decision process involves selecting
the best machine, number of machines, and parts per build for a
given demand scenario. Response surfaces are created to
discover the relation between the number of parts per build, the
production time, and the expected utility. Using these response
surfaces and the preferences of engineers, the machine type,
number of machines and parts per build are obtained that give
the maximum utility for the system goals.
3.4 Solution Methodology
With the general approach for solving our problem in
place, we are able to devise a concrete decision-making
method. As stated above, there are two levels of decisions in
our problem. We have adopted a method similar to dynamic
programming [9] we first develop response surfaces for the
two decisions (second decision followed by the first decision)
in a backward manner; we then use these surfaces to progress
through the problem in order. In other words, the first decision
uses the response surfaces from the second decision.

Copyright 2003 by ASME

The solution methodology involves four steps (see Figure

3):
Step 1. (D2) For various combinations of number of parts per
build, and different combinations of machine parameters,
the cost, quality, and time metrics (system goals) are
evaluated. A response surface is fit between the machine
parameters and the system goals.
Step 2. (D1) For a given demand and varying combinations of
machine types, number of machines, and parts per build, the
system goals are evaluated and a response surface is plotted.
This response surface, along with the response surface built
in Step 1 is used for the final two steps.
Step 3. (D1) Based on the demand and the response surfaces
developed in Steps 1 and 2, the best machine type, the
number of machines and the number of parts per build are
obtained.
Step 4. (D2) Based on the number of parts and the machine
selected, the best set of machine parameters is obtained.
Steps 1 and 2 are the backward progression steps in which
ranges of values are passed forward and the response surfaces
are built. Steps 3 and 4 are the forward progression steps in
which the response surfaces of goals are used for making the
decision at each step.
With the general solution strategy in place, we begin its
illustration through an application: the rapid manufacturing of
customized hearing aids.

MODELS FOR THE RAPID MANUFACTURING OF

HEARING AID SHELLS

We wish to explore an extension of the Siemens/Phonak

hearing aid shell scenario discussed earlier in Section 1. We
imagine a company that wishes to offer customized hearing
aids with a rapid manufacturing process based on Stratasys
FDM technology. We want to offer customers a hearing aid
with a personalized geometry, a choice of color (three different
tones), and a choice of material: ABS or elastomer.
This extended level of variety provides a more complex
problem to investigate. With three different colors and two
different materials available, our problem becomes one of
identifying proper process parameters for six different types of
products.
In this section we present our modeling efforts. We present
the specifics of the hearing aid itself in Section 4.1. We
provide an overview of the FDM process and our method of
modeling it in the following section. Before moving on to the
solution of the specific problem in Section 5, we discuss how
demand, utility, and the problems constraints are modeled.

Figure 6: Hearing Aid Shell (Masters, 2002)

Table 1: Hearing Aid Dimensions

Hearing Aid Model

Upper Ellipse
Minor Diameter
10.72 mm
Lower Ellipse
Minor Diameter
6.36 mm
Part Height
21.78 mm
Volume of Material

Upper Ellipse
Major Diameter
18.7 mm
Lower Ellipse
Major Diameter
10.5 mm
Wall Thickness
1.11 mm
229 mm3

This model provides insight into the selection of some

process parameters. For example, the extremely fine wall
thickness provides a constraint on the feasible road width
values. With this information presented, we move on with a
discussion of the FDM process and our method of modeling it.
4.2 The FDM Process Model
While our motivation is to present a general decisionmaking method for the rapid manufacturing of customized
parts, our validation is based in the technology of Fused
Deposition Modeling (FDM). The FDM process from Stratasys
[10] produces prototype parts by extruding a heated filament
through a nozzle that lays down the parts cross-section layer
by layer. The material filament is fed through a heating
element where it becomes semi-molten. The filament is then
fed through a nozzle and is deposited onto the partially
constructed part. The newly deposited material fuses with
adjacent material that has already been deposited. The head
deposits a layer by moving around in the X-Y plane. Multiple
layers are fused together to complete the prototype by lowering
the build platform along the Z-axis.
FDM Head
Direction

4.1 Hearing Aid Shell Model

We are interested in offering customized geometry of the
hollow shell of the hearing aid. A render of a typical hearing
aid shell is presented in Figure 6.
Although each hearing aid will be geometrically different,
for modeling purposes we approximate its geometry as a
truncated elliptical cone. Measurements, taken directly from an
actual shell with micro-calipers, are displayed in Table 1.

Model
X

Support

Figure 7: FDM Process

As can be seen in Figure 7, FDM machines have a second
nozzle that extrudes support material to enable the construction
of overhanging structure.

Copyright 2003 by ASME

4.2.1 FDM Machines

For this paper, three FDM machines were investigated: the
Prodigy Plus, the FDM Titan, and the FDM Maxum. Each
machine was assigned specific process characteristics based on
information received from the manufacturer. The maximum
number of parts was estimated with the knowledge of average
hearing aid dimensions and the dimension of the build volume
of each machine. This data is presented below in Table 2.
Table 2: Machine Process Characteristics
(Stratasys, 2003)
Machine

Max. Parts

Scan Speed

Prodigy Plus

234

64 mm/s

Titan

818

127 mm/s

Maxum

1703

254 mm/s

Layer
Thickness
0.178
0.33 mm

Road
width
0.19
0.21 mm

Cost

0.24
0.26 mm
0.127
0.25 mm

0.19
0.21 mm
0.193
0.965 mm

$210,000

$70,000

$260,000

These characteristics will provide the foundation for the

process constraints that will be presented later.

Clabor = cost of labor ($25/hour)

Coperation = hourly cost of machine operation ($80/hour)
t = time to build a single part (hours)
Cmaint = annual machine maintenance cost ($50,000/year)
Nmachine = the number of machines
Nppy = the number of parts produced per year
The values listed for Clabor, Coperation, Cmaint are rough
estimates. Their values do not change the fundamental results
presented in Section 6.
4.2.4 Quality Model
For many products, quality is a metric with a subjective
nature. When manufacturing with rapid prototyping, quality
can be quantified by the stair stepping affect the inability to
produce smoothed surfaces in the z-direction due to the layerbased additive manufacturing method. For our purposes, the
quality metric is dependent on the selection of the layer
thickness (LT) and road width (RW) process parameters. In
Figure 8, we show how different road width and layer thickness
affect the stair stepping affect.

4.2.2 Throughput Model

The total time to build a set of hearing aids is
approximated as:

t = t s + tl Nl + tbase

(hours)

(1)

where:
ts = Set up time that includes heating the filament, setting
foam etc. ( 0.5 hr.)
tl = Time to build each layer
Nl = Number of layers
tbase = Time to build the base of the part (generally, 4 layers
are built)
The number of layers needed to complete a part is a
function of the layer thickness process parameter. The time to
build each layer is approximated as:

tl = tlower + tboundary + t fill + t supp

(hours)

(2)

where:
tlower = Time to move down z-axis (hours)
tboundary = Time to draw outer and inner boundaries (hours)
tfill = Time to fill the inner portion by raster scan (hours)
tsupp = Time to draw support for this layer (hours)
tfill and tsupp are functions of the number of parts per layer, the
scan speed, and the road width. All of these variables are
unique to each machine type.
While it is nave to assume that the time to build each layer
will be the same for each layer of the inverted cone model, we
believe this will not significantly influence the overall results.
4.2.3 Cost Model
The cost of a single part is estimated as:
C N
C Mfg = Vmaterial Cmaterial + ( Clabor + Coperation ) t + maint machine
N ppy

($)

where:
Vmaterial = volume of material used to build one part
(~= 229 mm3)
CMaterial = cost of material
(includes build and support material) ($)

(3)

LT

LT

RW

RW

Desired Part Geometry

Actual Part Geometry

Figure 8: Quality Metric Diagram

As can be seen, the best quality is achieved with a minimal
layer thickness and minimal road width. Our quality metric is
simply the sum of these two process variables. The engineer
would like to reduce this value of quality metric for obtaining a
good quality part.
The metric is determined by the equation below:

Q = tlayer + wroad

(mm)

(4)

where:
Q = quality (mm)
tlayer = layer thickness (mm)
wroad = road width (mm)
This method of quantifying quality is the most relevant for
this application. Ideally one would compare the error between
the actual part and the model via an exact representation of the
curvature of the hearing aid.
4.2.5 Modeling Demand
Because we are working with constant demand in this
phase of our research, it is beneficial to capture demand
information in units of demand per day. This simulates an
arrival rate of demand to the RP machines. Each different
product variant can have a different demand per day. The
demand at a specific time can be evaluated though the relation:
Dd
D ( tw ) = i tw
(7)
24
where:
D(tw) = demand as a function of wait time (number of parts)
Ddi = demand per day of product variant i
tw = wait time (hours)

Copyright 2003 by ASME

The wait time, tw, is the manner in which we model the

machines creation of batches. tw is the amount of time the
machine waits to begin building a batch. As time passes, more
and more parts arrive at the workstation (as a function of
demand per day).
The number of parts per batch to meet this demand is
calculated through:
D ( tw )
N PPB =
(8)
N batches
where:
NPPB = number of parts per batch
D(tw) = demand as a function of wait time
Nbatches = number of batches
By substituting this into the cost and time equations previously
developed for each machine and material combination, we are
able to relate demand and the number of batches to time and
cost.
4.2.6 Applying Constraints
In order to meet the requirements of the problem statement,
we must impose some constraints (see Figure 4) to be certain
that:
(1) the parts are produced within a week of their order arrival
(2) the batch size is not greater than the maximum allowed in
the machine
(3) the cost per part is not too expensive
(4) each demand per day of all product variants can be met
To assure that the parts are produced within a week of their
order arrival, we introduce the notion of allotted time:
hrs

t allotted =

(168 week )( Dd )
i

Dd

(9)

zero. This applies bounds on the time and cost per part we
cannot accept any combination that is too expensive or cannot
produce the demand by the determined lead-time of one week
(3) and (4).
4.2.7 Evaluating Utility
In order to evaluate the utility functions at each decision,
we must comprehensively assess and quantitatively capture a
decision-makers preferences. In our modeling efforts we
assume that most designers under most circumstances are risk
averse - they prefer alternatives that offer on-target outcomes to
those that have considerable chances of yielding undesirable
results.
We begin modeling utility quantitatively by specifying
points along the utility curve so that a utility function can be fit
to the data to represent the decision-makers preferences. The
designer begins by identifying an upper bound (ideal - a utility
of 1) and a lower bound (unacceptable - a utility of 0) to each
variable. The designer is then asked a series of 50-50 lottery
questions in order to assess the utility values in between these
bounds. A utility function is then fit to the points assessed in
during the questioning.
For our efforts, the utility functions are formulated using
the mathematical form: U = a + b.x + c.exp(d.x). The variable
x is the goal for which utility is given by this function. The
coefficients a, b, c, and d are constants and determined by
fitting the utility curve to the points identified by the designer.
(For more information, please consult a standard reference on
utility theory, e.g., [8]).
With the process model, metrics, and constraints fully
defined, we move on to the series of decisions in which we
select the best values of process parameters.

i =1

where:
tallotted = time allotted for each product variants order to be
met in a week (hours)
Ddi = demand per day of product variant i
i = index for product variants (six different combinations of
material and color)
(1) Given a daily demand for each product variant and
the fact that the order must be completed within a week we
can use this idea of allotted time to assign each variant a
number of hours that it must be completed by. As will be
shown in Section 5.2, this metric is used in the determination of
the utility of build time. Only those build times that are
feasible within this allotted time are given a positive utility.
Without this constraint, the solution algorithm would look for
large batch sizes in order to cope with demand, while
completely ignoring the fact that larger batch sizes take longer
times to complete.
(2) As presented earlier (Table 2), each machine has an
estimate of the maximum number of parts per batch that it can
physically build. For each combination of parameters explored,
any combination that provides an estimate of parts per batch
greater than the machines capacity is assigned an expected
utility of zero.
The use of utility metrics is another manner in which we
apply constraints to our solution space. Any combination of
build parameters that provide a negative utility value is
immediately ignored by setting the expected utility value to

SELECTING PROCESS PARAMETERS

In the following section, we solve the decision problems

for the rapid manufacturing of hearing aid shells presented in
the Section 3.
5.1 Response Surface Generation for Decision 2
Using the process models discussed above, response
surfaces are needed to approximate the general trends of the
process parameters and their effects on the overall utility. With
these response surfaces developed, we will be able to quickly
choose a solution for Decision 2 selection of process
parameters (road width and layer thickness). This is Step 1 in
our solution methodology presented in Section 3.4.
5.1.1 General Response Equations
By running a DOE on the many RP process parameters of
FDM, we are able to calculate the time, cost, quality, and
overall utility for each process parameter combination for
different number of batch sizes for each machine. With this
data we create response surfaces for the different combinations
of road width and layer thickness for a single machine and a
single batch size. This is then repeated for each machine,
material, and build size.
The resultant response surfaces provide a graphical
representation of the expected metric trade-offs. For example,
both cost and build-time increase with small layer thicknesses

Copyright 2003 by ASME

and small road widths (thus lowering their utilities); but the
utility of quality is a maximum at for these values.
With these response surfaces for each metric established,
we are able to combine them into one surface to display the
overall utility. Overall utility is expressed by:
Overall Utility (D2) = (Kcost.Ucost + Ktime.Utime + KQuality.UQuality)

(5)

where:
kcost = k-value for cost = 0.25
Ucost = utility of cost (the price per part of the batch)

0.685C

Cmax

= U cos t = 1.95089 exp

ktime = k-value for build time = 0.25

Utime = utility of build time

combine them in order to develop approximate values that

provided the largest overall utility.
5.1.2 Largest Utility Response Equations
For each surface generated in the previously discussed
section, there is a maximum point on that surface a
combination of road width and layer thickness for each build
size of each machine and material type that provides the largest
possible overall utility. When combining these values for each
machine and material combination, we are able to obtain an
equation that approximates the best cost and time for each build
size. These response surface equations are presented in Table
3.
Table 3: Decision 2 Response Equations for Largest Overall
Utility
Machine
Type

0.6861t

tmax

Response Equation

t = 0.0357n + 0.5917
C = -9E-07n3 + 0.0004n2 - 0.0563n + 6.0287
t = 0.0357n + 0.5917
Elastomer
C = -9E-07n3 + 0.0004n2 - 0.0563n + 6.0588
t = 0.0235n + 0.6141
ABS
C = 2E-06n2 - 0.0019n + 2.6578
FDM Titan
t = 0.0235n + 0.6141
Elastomer
C = 2E-06n2 - 0.0019n + 2.6879
t = 0.0168n + 0.6717
ABS
FDM
C = -3E-10n3 + 9E-07n2 - 0.0011n + 1.9304
Maxum
t = 0.0168n + 0.6717
Elastomer
C = -3E-10n3 + 9E-07n2 - 0.0011n + 1.9605

= U t = 1.951 exp

Prodigy
Plus

kuality = k-value for quality = 0.5

Uquality = utility of quality
= U Q = 1.96268 exp ( 2.3516Q )
The manner in which the coefficients of the individual utility
functions were calculated is presented in Section 4.2.7.
The k-values represent the preference of a decision maker
towards various goals. The k-values can also be viewed as the
weights to utilities of goals. In this problem, we believe the
metrics of cost and time are dependent. Hence, we reduce their
k-values to 0.25 and assign a k-value of 0.5 to quality so that
equal weights are assigned to quality and cost/time related
goals. Another decision maker may choose different k-values
based on his/her preferences with corresponding effect on the
results.
A response surface is then created for an individual
machine, material, and build size combination. In Figure 9 we
present a response surface of the overall utility for the FDM
Titan elastomer material for a build size of 851 parts.

Material
ABS

where:
t = the total time for a build (hours)
C = cost per part ($)
n = number of parts per batch
As expected, the cost per part decreases as the build size
increases. With very large build sizes the cost per part becomes
constant as the number of parts exceeds the machine cost.
With these response surfaces, it is much easier to capture
the trends of the metrics for various machines, materials, and
batch sizes. With this information we will be able to quickly
and efficiently formulate the response surfaces required of
Decision 1.

0.9
0.85
0.8
0.75
0.7

Utility

0.65
0.6

0.193

0.55

0.45031
0.70762
0.2377

0.2254

0.2131

0.2008

0.1885

0.1516

0.127

0.1393

0.1762

0.5

0.965

0.1639

RW (mm)

5.2 Formulating Response Surface for Decision 1

The second step in the general solution methodology is to
develop response surfaces for Decision 1.
Given the
approximations for cost and time of each batch size developed
in Decision 2, the task in this decision step is to find the best
batch size, the best machine for the job, and the appropriate
number of machines to minimize an expected utility function.
The expected utility expression for this stage in the
problem is similar to that developed earlier (Equation 5):

LT (mm)

Figure 9: Response Surface for Overall Utility

Titan, Elastomer, 851 parts
With these numerous response surfaces generated for each
machine, material combination, and size of build, we are able to

Overall Utility (D1) = (Kwait.Uwait + Kbuild.Ubuild + Kcost.Ucost) (6)

where:
kwait = k-value for wait time = 0.25
Uwait = utility of wait time (the amount of time a batch waits
to accumulate demand)
kbuild = k-value for build time = 0.25
Ubuild = utility of build time (the amount it takes to build the
batch)
kcost = k-value for cost = 0.5
Ucost = utility of cost (the price per part of the batch)

Copyright 2003 by ASME

Similar to Decision 2, we reduced the k-values for metrics

that are dependent on another. The wait time and build time
metrics are set to 0.25, with cost set to 0.5; this was done in
order to give each metric equal importance.
The numerical values for wait time, build time and cost
used for evaluation of utility functions are presented in Table 4.
Table 4: Values of goals for utility evaluation
Goal
Wait time
Build time

Attitude towards risk

Risk Averse
Risk Averse

K-value
0.25
0.25

Cost

Risk Averse

0.50

Value at Utility=0
168 hrs
Time allotted for
the batch (in hrs)
$5.00

Value at Utility=1
0.0 hrs
0.0 hrs
$0.00

The wait time utility function is formulated as:

U wait

0.6861t wait
= 1.951 exp

hrs
168 week

Figure 10: Expected Utility Surface - Titan

(10)

This function evaluates the utility involved with waiting

for more demand to arrive to process larger batches. This value
is greater when the process does not wait very long i.e., the
quicker the process the better. The utility value for wait time of
0 hrs is 1 and a wait time of 168 hrs (one week) is assigned a
utility value of 0.
The build time utility function is formulated as:

0.6861tbuild N batch

t allotted

U build = 1.951 exp

(11)

This utility function uses the maximum utility response

curves from Decision 2 to estimate the build time, tbuild. When
multiplied by the number of batches, the total build time is
calculated. As mentioned earlier, the utility function is
normalized by the allotted time for a certain product variant
thus only build times that are feasible within this allotted time
are given a positive utility (all negative values are immediately
thrown out by setting the overall utility function equal to zero).
Obviously higher utility is assigned for those average build
times that are shorter.
Finally, the cost utility function is formulated as:
0.685C part
(12)
U cos t = 1.95089 exp

$5
A cost of $5.00 is assigned the utility value of 0 assuming
that a cost of the hearing aid shell has to be below 5 dollars.
Larger utilities are given for those prices per part that are lower.
With these individual metric utilities determined, they are
simply multiplied by their respective k-values and then
summed to produce a value for overall utility (Equation 6).
The utility is calculated with every combination of number of
batches and wait time for each machine/material combination.
After applying the constraints presented in Section 4.2.6, a
three-dimensional surface is plotted for each machine/material
combination. An example (for the FDM Titan) is presented in
Figure 10.
These surfaces can be generated for any values of daily
demands of the six product variants. On each surface a point of
maximum utility is found; this point represents the best number
of batches and wait time (of which, number of parts per batch
can be extrapolated) for the given demand of the product
variant on a certain machine.

The expected utility surfaces are a graphical combination

of the utility surfaces of the individual metrics Uwait, Ubuild, and
Ucost. The flat portions at the bottom of the surface represent
those configurations of process parameters that produced an
expected value of zero usually due to the inability to meet the
imposed constraints. The broad curved dome shape shown in
Figure 10 is typical of those scenarios in which the machines
capacity can easily handle the given demand scenario. Demand
scenarios that border the upper limit of the machines capacity
produce expected utility surfaces of steeper slope and smaller
areas of positive utility.
With this established, we are now ready to move on to
Steps 3 and 4 in the solution methodology outlined in Section
3.4.
5.3 Iterating through Response Surfaces
With all of our response surfaces generated for each
decision step, we are now able to present the specific steps of
solving the overall problem.
Decision 1: This corresponds to the Step 3 in our solution
methodology (see Section 3.4)
Given a set of daily demands for each product variant, a set
of response surfaces can be generated for each type of
machine.
From these surfaces, a point of maximum utility can be
identified.
This point of maximum utility represents the best
combination of wait time and number of batches to
accomplish the given demand of each product variant.
Decision 2: This corresponds to the Step 4 in our solution
methodology (see Section 3.4)
Given the numbers of parts per batch for all six product
variants, and the type of machine used, search response
surfaces to determine best road width and layer thickness
that maximizes quality, minimizes cost, and minimizes
build time.
If no options satisfy constraints (i.e., no point on the
response surface has a utility greater than zero), one must
investigate purchasing more than one machine. To make the
investigation less time-intensive we assume that only the same
type of machine can be purchased thus doubling the capacity.

Copyright 2003 by ASME

For these scenarios, the analysis is simply run again, except at

half of the original demand and double the price per machine.
6

RESULTS

6.1 Rapid Manufacturing With Fused Deposition

Modeling
With our response surfaces generated and coded for easy
repetition, we are able to find the best machine type, machine
number, batch size, and process parameters for any given
demand. For this paper we present one demand scenario,
summarized in Table 5.
Table 5: Demand Scenario
Material;
Color

ABS;
Color A

ABS;
Color B

ABS;
Color C

Elastomer;
Color A

Elastomer;
Color B

Elastomer;
Color C

Demand

150 ppd

150 ppd

150 ppd

120 ppd

120 ppd

120 ppd

This demand scenario represents an annual demand of

295,650 parts. Although we are able to handle a problem with
a different daily demand value for each product variant, we
chose these similar values to maintain brevity. Since the color
of the product variant does not affect the process parameter
selection, and each material type has equivalent values of
demand, we find that the results are the same among products
with similar materials. The best process parameters for this
demand scenario are presented below in Table 6.

6.2 Extending the Work: Technology Selection

As mentioned earlier, we would like to extend our
developed method to select a proper technology for a rapid
manufacturing scenario. Through the development of a proper
rapid prototyping model, one can easily compare results for
different technologies for a variety of demand scenarios. From
a simple analysis of 3D Systems Selective Laser Sintering
(SLS) sample build times, we are able to offer the following
results for the same demand scenario presented in Section 6.1.
We analyzed the SLS process with two materials:
1) DuraForm glass-filled (DFGF) A thermoplastic nylon
material
2) SOMOS 201 a thermoplastic elastomer material that
has a rubber-like characteristics
We present the overall utility response surface of the SLS
Vanguard HS machine with the DFGF material in Figure 11.
The response surfaces for the SLS Vanguard HS machine and
with the SOMOS 201 material are very similar. The best set of
process parameter values are shown in Table 7.

Table 6: Best Process Parameters

Machine

Prodigy
Plus
Titan
Maxum

Product
Variant

ABS
Elastomer
ABS
Elastomer
ABS
Elastomer

Optimal
Waiting Time
(hrs)

Batch Size

Cost per piece

($)

Layer
Thickness
(mm)

Road Width
(mm)

22
21
44
48
40
42

69
52
275
240
250
210

$4.70
$5.02
$3.00
$3.06
$2.56
$2.65

0.33
0.33
0.24
0.24
0.176
0.176

0.208
0.208
0.208
0.208
0.193
0.193

Figure 11: Overall Utility Response Surface Vanguard HS

Table 7: SLS Best Process Parameters
Product Variant

Comparing the maximum expected utility across all

product variants for each machine, the Industrial Engineer is
able to choose the best alternative, which in this scenario, is the
FDM Maxum. The Maxums fast scan speed and large build
volume also presents opportunity for the lowest cost per piece,
another metric for selection.
The cost per piece metric presented in the table above
includes the cost of each type of machine. The Prodigy Plus
large cost per piece can be attributed to the fact that two of
these machines were required to satisfy this large demand
scenario.
The crux of the rapid manufacturing problem is accurately
represented in these results. The industrial engineer would like
to wait for a large build size in order to make the build become
more cost effective. The longer the engineer waits, however,
the longer the build time becomes this combined time could
exceed the week-long lead-time constraint. A machines build
size capacity must also be big enough to build a cost effective
(large) batch size.
It is also important to note that the relative flat expected
utility surfaces of thee machines (Figure 10, as an example)
show that this selection is very robust to changes in demand or
wait time.

DFGF, Vanguard
SOMOS 201, Vanguard
DFGF, Vanguard HS
SOMOS 201, Vanguard HS

Optimal Waiting
Time (hrs)
35
29
30
25

Batch
Size
218
145
188
125

Cost per
piece ($)
$1.67
$1.67
$1.94
$1.94

Layer Thickness
(mm)
0.1524
0.1778
0.1524
0.1778

By comparing the results of this SLS analysis with that of

FDM, an educated decision on which technology one should
choose for a rapid manufacturing application can be made.
Here we notice that both SLS options present much cheaper
costs than any of those for FDM.
7 CLOSURE
Our goal in this paper was to answer the question earlier
presented in Section 1.2, namely:
How can a rapid manufacturing process be methodically
designed to provide a mass-customized product to achieve a
desired balance among cost, throughput, and quality?
In this paper, we proposed a general method for making
decisions in the process design of a customized product via
rapid manufacturing. We demonstrated our method on the
problem of rapidly manufacturing customized hearing aid shells

10

Copyright 2003 by ASME

with FDM technology. We then briefly presented results of the

same problem using SLS technology. This extension of our
work served as a preliminary investigation into the use of our
method for technology selection in a rapid manufacturing
scenario. Our solution has successfully linked the idea of
collaborative engineering, coupled decision-making, using
utility theory for tradeoff decisions, and rapid prototyping
process models.
Many interesting discoveries were made through the
analysis of our results. During the development of the response
surfaces of Decision 2, we noted that the road width and layer
thickness values were the same for each batch size for every
machine. In other words, the machine parameters were
independent of the batch size. Because of the extremely
different capabilities of each machine, the utility surfaces were
not specific enough to display preference over one value or
another. As a result, we conclude that one cannot confidently
use an expected utility function to select between alternatives
with large differences.
The goal spaces are relatively flat for broad ranges of
system variables for Decisions 1 and 2. That is, the graphs of
expected utility for Decision 2 (see Figure 10) are relatively flat
across ranges of wait time and number of builds. This indicates
two items of note. First, the trade-offs among competing goals
cancel out one another. A different result would occur for
different relative importances on the goals. Second, process
solutions are very robust to changes in wait time or number of
builds. For example, a 10 percent change in build size only
provides a 3% change in overall utility. This is a very desirable
result.
Currently, our work is limited to dealing with constant
demand. In the future we wish to rework our decision process
to deal with a probabilistic demand scenario through the use of
a Monte Carlo analysis to find the best values for output of
each decision. We have begun to implement our work over the
web to support distributed engineers. Through the use of XML
files, we are creating a template for decision-making in a rapid
manufacturing scenario. Finally, we are investigating this work
as a potential manner in which to model the development of a
process family a baseline of process parameters from which a
family of processes can be generated to efficiently deal with
varying demand scenarios.
ACKNOWLEDGMENTS
For providing SLS process estimates we thank Tim Gornet,
CAE Consultant at the University of Louisville. We thank

Martin Masters of Siemens Hearing Instruments for supporting

our investigation. Christopher Williams is sponsored by NSF
Grant DMI #9900259, NSF Grant DMI #0085136, and by a
Georgia Tech Presidential Fellowship. Jitesh Panchal is
sponsored by NSF Grant DMII #0100123. The cost of
computer time was underwritten by the Systems Realization
Laboratory at the Georgia Institute of Technology.
REFERENCES
[1] Anderson, D. M., 1997, Agile Product Development for
Mass Customization, IRWIN, Chicago, Illinois.
[2] Align Technologies, February 2003,
http://www.invisalign.com/US/html/explore/Tech_Behind_
Invisalign.html
[3] Masters, M., 2002, Direct Manufacturing of Custom-Made
Hearing Instruments, SME Rapid Prototyping Conference
and Exhibition, Cincinnati, OH.
[4] Mistree, F., W. F. Smith, B. A. Bras, 1993, A Decision
Based Approach to Concurrent Engineering, in Handbook
of Concurrent Engineering. Chapman and Hall: New York.
[5] Mistree, F., Hughes, O., and Bras, B., 1993, The
Compromise Decision Support Problem and the Adaptive
Linear Programming Algorithm, Structural Optimization:
Status and Promise, M.P. Kamat, Editor. AIAA:
Washington, D.C., pp.247-286.
[6] Seepersad, C.C., F. Mistree, and J.K. Allen, 2002, A
Quantitative Approach for Designing Multiple Product
Platforms for an Evolving Portfolio of Products,
Proceedings of the 2002 ASME Design Automation
Conference, Montreal, Canada, ASME, DETC2002/DAC34096.
[7] Fernandez, M.G., C.C. Seepersad, D.W. Rosen, J.K. Allen,
and F. Mistree, 2001, Utility-Based Decision Support for
Selection in Engineering Design, Proceedings of the 2001
ASME Design Automation Conference, Pittsburgh, PA,
ASME, DETC2001/DAC-21106.
[8] Keeney, R. L. and H. Raiffa, 1976, Decisions with Multiple
Objectives: Preferences and Value Tradeoffs, John Wiley
and Sons, New York.
[9] Jacobs, O. L. R., 1967, An Introduction to Dynamic
Programming: The Theory of Multistage Decision
Processes, Chapman and Hall, London.
[10] Stratasys, February 20003, http://www.stratasys.com

11

Copyright 2003 by ASME