Castro2005 PDF

SIMULTANEOUS OPTIMIZATION OF MOLD DESIGN
AND PROCESSING CONDITIONS IN INJECTION

MOLDING
Carlos E. Castro 1 , Mauricio Cabrera Rios 2 , Blaine Lilly 1 , and Jose M.
Castro 1
1
Department
of Industrial,
Welding & Systems
The Ohio State

Columbus,
2
Graduate
Program
Universidad
Ohio, USA
in Systems
Autnoma
Engineering
University
43202
Engineering
de Nuevo
Leon
San Nicolas de los Garza, Nuevo Leon, Mexico,
66450
ABSTRACT:
Injection molding (IM) is considered the foremost process for massproducing plastic products. One of the biggest challenges facing injection
molders today is to determine the proper settings for the IM process
variables. Selecting the proper settings for an IM process is crucial because
the behavior of the polymeric material during shaping is highly influenced by
the process variables. Consequently, the process variables govern the quality
of the part produced. The difficulty of optimizing an IM process is that the
performance measures (PMs), such as surface quality or cycle time, that
characterize the adequacy of part, process, or machine to intended purposes,
usually show conflicting behavior. Therefore, a compromise must be found
between all of the PMs of interest. In the past, we have shown a method
comprised of Computer Aided Engineering, Artificial Neural Networks, and
Data Envelopment Analysis (DEA) that can be used to find the best
compromises between several performance measures. The analyses presented
in this paper are geared to make informed decisions on the compromises of
several
performance
measures.
These
analyses
also
allow
for
the
identification of robust variable settings that might help to define a starting

point for negotiation between multiple decision makers.
Future work will include adding information about the variability of PMs
on the DEA analysis and the determination of process windows with
efficiency considerations. This paper discusses the application of this method
459
Brought to you by | University of Michigan
Authenticated
Download Date | 6/15/15 4:59 PM
Simultaneous Optimization of Mold Design and

Processing Conditions in Injection Molding
Vol. 25, No. 6, 2005
to IM and how to exploit the results to determine robust process and design
settings.
INTRODUCTION
Injection Molding (IM) is considered the major process for massproducing plastic parts. According to the Society of the Plastics Industry,
over 75% of all plastics processing machines are IM machines, and close to
60% of all plastics processing facilities are injection molders /I/. Selecting
the proper IM process settings is crucial because the behavior of the
polymeric material during shaping is highly influenced by the process
variables. Consequently, the process variables govern the quality of the part
produced. A substantial amount of research has been directed towards
determining the process settings for the IM process as well as the optimal
location of the injection gate.
The challenge of optimizing an IM process is that the performance
measures often show conflicting behavior when they are functions of process
or design variables in common. For example, the cycle time and the part
warpage will both be affected by the ejection temperature. Increasing the
ejection temperature would
be favorable for minimizing cycle
time.
However, letting the part cool to a lower ejection temperature before

demolding would decrease the part warpage. Therefore, a compromise must
be found between these two performance measures to set the ejection
temperature. For this reason, when optimizing an IM process it is nearly
impossible to find one best solution. However, it is feasible to determine a set
of best compromises between multiple PMs.
The problem of considering several PMs simultaneously, i.e. finding the
best compromises, is referred to as a multiple criteria
optimization.
Conventional methods of multiple criteria optimization involve combining

individual weighted PMs into one objective function and optimizing that
function. These methods will converge to a solution, however it might prove
a challenge to determine if this solution lies in the efficient frontier,
especially in the case where the PMs show nonlinear behavior. In addition,
this solution is dependent on the bias of the user defining the weights. In
engineering practice it is often times impossible to define one optimal
solution to all criteria. Instead, it is both feasible and attractive to determine
the best compromises between PMs: that is the combinations of PMs that
460
Authenticated
Journal of Polymer
C.E. Castro et at.
Engineering
cannot be improved in one single dimension without harming another. Data

Envelopment Analysis (DEA) provides an unbiased way to find these
efficient compromises.
It is the purpose of this paper to demonstrate the determination of
efficient solutions (best compromises) in an IM context through a series of
case studies comprising several potential industrial applications. These
solutions prescribe the settings for IM process and design variables.
Additionally, the identification of robust solutions is discussed.
THE OPTIMIZATION STRATEGY

Proposed by Cabrera-Rios et al. 12,3/ the general strategy to find the best
compromises between several PMs consists of five steps:
Step 1) Define the physical system. Determine the phenomena of interest, the
performance measures, the controllable and non-controllable variables,
the experimental region, and the responses that will be included in the
study.
Step 2) Build physics-based models to represent the phenomena of interest in
the system. Define models that relate the controllable variables to the
responses of interest. If this is not feasible, skip this step.
Step 3) Run experimental designs. Create data sets by either systematically
running the models from the previous step, or by performing an actual
experiment in the physical system when a mathematical model is not
possible.
Step 4) Fit metamodels to the results of the experiments. Create empirical
expressions (metamodels) to mimic the functionality in the data sets.
Step 5) Optimize the physical system. Use the metamodels to obtain
predictions of the phenomena
of interest, and to find the
best
compromises among the PMs for the original system. The
best
compromises are identified here through DEA.

In the method outlined here, the metamodels are empirical
approximations of the functionality between the controllable (independent)
variables, and the responses (dependent variables). These metamodels are
used either for convenience or for necessity. Because DEA requires that
many response predictions be made, it is more convenient to- obtain these
predictions from metamodels rather than more complicated physics-based
461
Authenticated

Vol. 25, No. 6, 2005
models. In addition, when physics-based models are not available to

represent the phenomena of interest, the use of metamodels becomes
essential.
DATA ENVELOPMENT ANALYSIS (DEA)

Cabrera-Rios et al. /2,3/ have demonstrated the use of DEA to solve
multiple criteria optimization problems in polymer processing. DEA, a
technique created by Charnes, Cooper, and Rhodes [4], provides a way to
measure the efficiency of a given combination of PMs relative to a finite set
of combinations of similar nature. The efficiency of each combination is
computed through the use of two linearized versions of the following
mathematical programming problem in ratio form:
Find
,,0
to
Maximize
(1)
(+
vTYmin
s.t.
HTvmax
T - <. 11
(2)
vYj
*
8-l
Tvmin
v
Y
0
vT
v
(3)
t
8
.j = 1 l , . . . , n
min- '
(4)
free
(5)
where Yj1** and Ym are vectors containing the values of those PMs
currently under analysis to be maximized and minimized respectively, is a
vector of multipliers for the PMs to be maximized, is a vector of multipliers
for the PMs to be minimized, 0 is a scalar variable, is the number of total
combinations in the set, and is a very small constant usually set to a value
462

Authenticated
C.E. Castro et al.
Journal of Polymer Engineering
of lxlO"6. The solutions deemed efficient by the two linearized versions of

the model shown above represent the best compromises in the (finite) set of
combinations of PMs. A complete description of the linearization procedure
as well as the application of this model can be found in any of the references
1 through 5.
DEA is a non-parametric (i.e. data-driven) approach, as implied by its
name. This characteristic eliminates the need to decide upon specific
'weights' for the different PMs, thereby maintaining the objectivity of the
analysis. The linearized DEA versions described previously can be
effectively solved using the simplex method, which is included in practically
every major commercially-available optimization software package. The only
potential burden in the application of DEA might come from repeating the
task times (once per each of the combinations). However, by properly
computer-coding this routine, the analysis can be achieved in a reasonable
amount of time 161.
DETERMINATION OF SETTINGS OF PROCESS VARIABLES AND

INJECTION POINT
Consider the part shown in Figure 1. This part, which we introduced in
previous works /5,7,8/, represents a case where the location of the weld lines
is critical, and the part flatness plays a major role. The part is to be injection
molded using a Sumitomo IM machine using PET with a fixed flow rate of
9cc/s. Nine PMs were included in this study: (1) maximum injection
pressure, /, (2) time to freeze, tf, (3) maximum shear stress at the wall, Sw,
(4) deflection range in the z-direction, Rz, (5) time at which the flow front
touches hole A (a, (6) time at which the flow front touches hole B, tB, (7)
time at which the flow front touches the outer edge of the part, t, (8) the
vertical distance from edge 1 to the weld line, d,, and (9) the horizontal
distance from edge 2 to the weld line, d2. Figure 2 shows how the weld line
locations were measured. For production purposes it is desirable to minimize
Pt, tf, Sw, and Rz: Pi to keep the machine capacity unchallenged, tf to reduce
the total cycle time, Sw to minimize plastic degradation, and Rz to control the
part dimensions. It is desirable to maximize tA, tB, t^, d,, and d2. tA, tR, and t,K,
should be maximized in order to provide a balanced flow and a uniform
pressure distribution along the edges of the part. Uniform pressure along the
edges will minimize the potential for flash, d/ and d2 should be maximized to
Brought to you by | University of Michigan463

Authenticated

Vol. 25, No. 6, 2005
keep the weld lines away from corners which were assumed to be areas of
stress concentration. It was assumed that the top right and bottom left corners
of the part would be subjected to the highest stress during the intended
application.
= 5 cm
30
Edge
1 30 cm
Fig. 1: Part of constant thickness with cutouts
Fig. 2: Measuring the locations of the weld lines

Five controllable variables were varied at the levels shown in Table 1 in a
full factorial design. These controllable variables include: (a) the melt
464
Authenticated
Journal of Polymer
C.E. Castro et al.
Engineering
temperature, Tm, (b) the mold temperature, Tv, (c) the ejection temperature,
Te, (d) the horizontal coordinate of the injection point, x, and (e) the vertical
coordinate of the injection point, y. Te was only varied at two levels because a
preliminary study showed that a third level did not add any meaningful
variation. The injection point location is constrained to be in the region
shown in Figure 1, due to limitation of the IM machine. This point will be
characterized by the variables and y in a Cartesian coordinate system with
its origin at the lower left corner of the part.
Table 1
Levels of each of the controllable variables for the initial dataset
Tm
'
'
cm
cm
Label
-1
260
120
149
15
10
275
130
159
20
17.5
290
140
25
25
A finite element mesh of the part was created in Moldflow I M in order to

obtain estimates for the performance measures. An initial dataset of 162
experimental points was obtained from the full factorial design. Following
with the general optimization strategy, this initial dataset was used to create
metamodels to mimic the behavior of each of the performance measures. In
general, it is favorable to fit a simple model to the data. In this study, second
order
linear regressions were initially considered
as models for the
performance measures. When simple models do not suffice, then more

complicated models, in this case ANNs, become necessary. In order to
measure the prediction capability of the metamodels, a validation set was
used. This validation set was not incorporated into the training of the
metamodels. The levels used in the validation dataset are shown in Table 2.
The full factorial design for the validation set resulted in 16 runs.
Table 2
Levels of controllable variables for validation set
'm
Label
-0.5
267.5
125
0.5
282.5
135
1
e
154
cm
cm
17.5
13.75
22.5
21.25

Authenticated
465

Vol. 25, No. 6, 2005
In general the ANNs outperformed the second order linear regression for
every performance measure in terms of approximation quality and prediction
capability, and were therefore used to obtain predictions for each PM at
previously untried combinations of controllable variables. The results for the
performance of the regression models and the ANNs obtained can be found
Tables 3 and 4 respectively. The performance of the regression models was
measured by two means: the coefficient of determination (R2) and the mean
absolute percent prediction error (MAPE). The MAPE was found using the
equation:
MAPE =
100
v Z l Z
W
(6)
V;
where is the number of experimental points, y, and y, are the experimental

response and the prediction at the t h experimental point respectively. Also
reported in Tables 3 and 4 is the MAPE for the validation data set (the set of
points obtained experimentally not used to fit the metamodel). This second
MAPE is useful to understand the prediction capabilities of the metamodel
m.
Table 3
Summary of performance and results from residual analysis results for the
regression metamodels
Pi
R2 (%)
Sr
tf
Rz
tA
tot
d,
d2
96.6 98.98 99.45 96.17 99.34 99.57 78.29 85.98 97.4
MAPE (%)
"
MAPE
Validation
(%)
8.632 7.681 2.031 15.29 10.87 4.325 30.31 16.31 6.766
Normality
Yes
Yes
Yes
Yes
Yes
Yes
Yes
No
Yes
Independence Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
onstant
Variance
466
3.63
yeg
1.3
yeg
1.71 27.96 5.65
yeg
Ygs
Yeg
2.84 33.27 23.64 4.06
Yeg
Yes
No

Authenticated
Yeg
C.E. Castro et al.
Table 4
Summary of performance for the ANNs
R (%)
MAPE (%)
Sw
Pi
V
99.98 97.89 99.99
0.2 2.02 0.21
Rz
99.1
8.6
tA
100
0.25
to.
tB
99.98 100
0.43 0.36
MAPE
Validation
3.66
6.68
3.66
2.42
0>
(X
5.76
0.24
6.17
di
d2
97.04 99.43
7.26 1.29
4.87
1.38
(%)
The complete multiple criteria optimization problem originally posed for

this case contained all nine performance measures. To solve the optimization
problem, it was necessary to generate a large number of feasible level
combinations of the controllable variables. This was achieved by varying Tm
and Tw at five levels, and the rest of the variables at three levels within the
experimental region of interest (see Table 1) in a full factorial enumeration.
This experimental design resulted in a total of 675 combinations.
The results after applying DEA were that over 400 of the 675
combinations were found to be efficient. Such a large number of efficient
combinations can be explained by examining Table 5, which summarizes the
results of the analysis of variance of each PM in regression form. Notice that
the last five PMs are only dependent on the injection point position
determined by variables and y. Any specific combination of values (xV*)
will give the same result on all of these five PMs regardless of the values that
the rest of the other controllable variables Tm, Tw, and Te take. Having used a
full factorial enumeration with and y at three levels, it follows that we can
obtain only nine different values for these five PMs, but each of the nine
specific combinations (x,y) have in fact 75 combinations of the rest of the
controllable variables. In the high dimensionality of the problem, this
elevated amount of repetition results in a large number of efficient solutions.
In order to increase the discrimination power i.e. obtain fewer efficient
solutions, one can solve the DEA model shown in Eqs. 1 through 5 by setting
equal to zero. The resulting model is similar to Charnes-Cooper-Rhodes
(CCR) DEA model 191.
467

Authenticated

Vol. 25, No. 6, 2005
Table 5
The significant sources of variation (linear, quadratic and second order
interaction terms in the linear regression metamodel) to each performance
measure
Performance Measures
Pi
W
to
S
e
Zi
tf
>4
Sw
Rz
G>
t -
TO
to,
d,
d2
T *,
JC
::
* *
* . \ 1. t 1, ,
Second Order Interactions
tB
llIBl
tA
'
u p
IRl?
Using the simple modification described above, the number of efficient

combinations comes down to 149. It can be shown that these combinations
are a subset of those 4 0 0 plus found previously. These efficient combinations
are shown in terms of the PMs in Figure 3.
It is important to notice that we can exploit the information our methods
gave us about the functionality of the PMs in order to tailor the optimization
problem. To illustrate, five sub cases were defined for practical applications
of the conceptual part shown in Figure 1: (i) an excess capacity injection
machine application, (ii) a dimensional quality and economics
468

Authenticated
critical
C.E. Castro et al.
application,
(iii)
a structural
part
application,
(iv) a quality
critical
application, (v) and a case including PMs that are only dependent on the
injection location [8].
Position
Weld Line 2
m.i\
1 3 1 4 3 Iran
run 62 3 2 m m
Position
Weld Line 1
max= 124 67 mm
min= 31.37 mm
Time to touch
Outer Edge
Tune to touch
Hole
Time to touch
Hole A
Max Shear Stress at
the Wall
Deflection
Range. Rz
max= 0 163 MPa

m i n = 0 0 9 0 MPa
inax= 0 . 0 1 2 8 ivun
mlil=0.0003inin
Tune to
freeze
l n u = 30 92 MPa
Max Injection
Pressure
inln= 14 6 3 MPa
90 100 110 120 130 140
Efficient Solution
Fig. 3:
Levels of the PMs that corresponded to the efficient solutions when

all nine were included
Excess Capacity Injection Molding Machine:

For a case in which the injection-molding machine has excess capacity, it
would be possible not to consider the maximum injection pressure in the
optimization problem. For simplicity, in this case Sjp, t^, hi and t(>e were also
dropped from the optimization, leaving four performance measures. The
DEA model was again solved here by setting the constant 0 equal to zero in
order to improve the discrimination power of DEA. The functionality shown
in Table 5 called for inclusion of all variables, and the factorial enumeration
with 675 combinations was used. In this case, fourteen combinations were
found to be efficient. Figure 4 shows the levels of the PMs for the efficient
solutions. The compromise between the locations of the weld lines is evident.

Authenticated
Vol. 25, No. 6, 2005

A noticeable compromise also arises between tf and R,. This is an

understandable compromise, because the two depend oppositely on the
ejection temperature.
140
-a-Position Weld Line 2 Position Weld Line 1

Time to freeze
D e f l e c t i o n Range
10 11 12 13 14 15
Efficient Solution
Figure 4: Efficient solutions for the excess machine capacity application in
terms of the levels of the PMs considered.
Figure 5 shows the locations of the injection gate for the efficient solution.
The positions in this case help to define 'attractive' areas to locate the
injection port, sinte they tend to cluster in specific sections. In this case the
efficient injection locations clustered along right and bottom edges. The three
PMs that are affected by the location of the injection gate are the weld line
positions and the deflection in the z-direction. The additional PM here is the
time to freeze, which is not affected by the injection location according to the
analysis of variance.
470

Authenticated
C.E. Castro et al.
1.0
0.8
0.6
0.4
0.2
>-
0.0
-0.2
-0.4
-0.6
-0.8
-1.0
-1.0 -0.8 -0.6 -0.4 -0.2 0.0
0.2
0.4
0.6
0.8
1.0
X
Fig. 5:
Injection locations of the efficient solutions to the excess machine

capacity application transformed to fall between -1 and 1.
Table 6 shows the values for all of the controllable variables at the
efficient solutions. Notice that Tw and Tm were at 120 and 260 degrees Celsius
respectively for all of the efficient solutions. In industrial practice, if the PMs
involved in this case were the only ones of interest, this would be a good
indication that Tm and Tw should be set at these temperatures. Also notice that
the ejection temperature values of the efficient solutions vary over the entire
range. According to the analysis of variance, d, and d2 do not depend on the
ejection temperature, so this fact must be due to the compromise between R:
and //previously mentioned.
Dimensional Quality and Economics Critical Application:

In this case it was assumed that the economic concerns included
minimizing the cycle time and keeping the machine capacity untested in
order to have long machine life and smaller power consumption. These two
concerns are defined by // and P/ respectively. Rz defines the dimensional
quality. The analysis of variance shows that all of the controllable variables
affect at least one of these PMs, so the enumeration with 675 combinations
471
Authenticated

Vol. 25, No. 6, 2005
again was applied. Twenty-five efficient solutions were found. Since the
problem is three-dimensional the efficient frontier can be visualized. The
efficient points are shown in Figure 6 with respect to the rest of the data set.
Table 6
Efficient solutions for the excess machine capacity application
Controllable Variables
*e
c
*f
s
Rz
d,
d2
mm
mm
mm
120
260
149
20.89
0.0005
37.9
130.9
120
260
154
18.99
0.002
37.9
130.9
25
120
260
159
17.27
0.007
37.9
130.9
25
120
260
149
22.71
0.000
37.9
131.4
25
17.5
120
20.90
19.00
107.8
120
149
154
94.9
17.5
260
260
0.001
25
0.005
94.9
107.8
25
25
17.5
120
120
260
260
159
149
17.28
22.72
0.010
0.001
94.9
17.5
94.9
107.8
108.9
15
10
120
260
149
20.92
0.001
124.6
67.4
15
10
120
260
154
19.02
0.005
124.6
67.4
15
25
10
10
120
120
260
260
159
149
17.30
20.92
0.011
124.6
124.7
67.4
0.001
25
10
120
260
154
19.02
0.006
124.7
82.1
25
10
120
260
159
17.30
0.012
124.7
82.1
y
cm
25
25
25
25
25
25
*IV
c
m
c
:
cm
82.1
Figure 7 shows the efficient solutions in terms of the levels of the PMs.
The direct compromise between the time to freeze and deflection is
confirmed here. Notice that they follow opposite trends while it is favorable
to minimize both.
Figure 8 shows the locations of the injection gate for the efficient
solutions. This case contradicts the first case. In the large machine capacity
case, the 'attractive' areas for the injection gate were found at the bottom and
right edges of the feasible area, but in this case, the top edge and bottom left
472
Authenticated
Journal of Polymer
C.E. Castro et al.
Engineering
0.002
0 004
0.006
0.008
0 012
Time to freeze [si
Injection Pressure
f
[MPa]
Fig. 6: A visualization of the efficient frontier of the economics critical and

dimensional application
Number of efficient solution (sorted In i n c r e u t n g value of ,)

| - * - Injection Pressure
Time to freeze
Deflection Range jZ) ]
Fig. 7: Efficient solutions for the dimensional
quality and
economic
application in terms of the levels of the PMs considered.
473
Authenticated
Vol. 25. No. 6, 2005

corner proved to be the efficient locations. This is due to the fact that the
positions of the weld lines were not considered in this case. From these
results we can conclude that dt and d2 are the main drivers for keeping the
injection location on the right or bottom edge.
They are the only PMs
affected by and y that were included in the first case and not in this case.
1.0
0.8
0.6
0.4
0.2
>- 0.0
-0.2
-0.4
-0.6
-0.8
-1.0
-1.0 -0.8 -0.6 -0.4 -0.2 0.0 0.2 0.4 0.6 0.8
1.0
X
Fig. 8:
Injection locations of the efficient solutions to the dimensional

quality and economics critical application transformed to fall
between -1 and 1.
Table 7 shows the twenty-five combinations of the controllable variables

that proved to be efficient for the dimensional quality and economics critical
application. Eighteen out of the twenty-five efficient solutions had the
injection gate located at the upper left corner of the feasible region, which is
close to the center of the part. This is the most robust injection location for
this application. According to the analysis of variance, is affected by the
location of the injection gate. Locating the injection gate towards the center
474

Authenticated
C.E. Castro et al.
would favorably decrease Pj. Since dt and d2 were not included in this case
there were no negative effects of moving the injection gate towards the
center.
Table 7
Efficient Solutions for the dimensional quality and economics critical
application
cm
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
25
25
20
25
25
15
15
cm
25
25
25
25
25
25
25
25
25
25
25
25
25
25
25
25
25
25
25
25
25
25
25
10
10
' w
' m
Te
Pi
'/
290
290
282.5
282.5
275
275
267.5
267.5
260
260
260
275
260
260
275
260
260
260
260
260
260
260
260
282.5
275
159
149
159
149
149
159
159
149
154
159
149
149
149
159
149
154
149
159
149
149
159
159
154
149
149
MPa
9.35
9.35
9.55
9.55
9.75
9.75
9.96
9.%
10.17
10.17
10.17
12.25
14.69
14.69
16.00
16.61
16.61
16.61
17.46
19.43
26.92
27.16
27.16
28.50
29.21
s
26.7
37.9
23.8
32.0
27.4
21.9
20.4
24.5
20.9
19.1
22.7
26.5
21.4
17.8
25.5
19.0
20.9
17.3
22.7
22.4
17.3
17.3
19.0
29.8
25.5
140
140
140
140
140
140
140
140
140
140
140
130
125
125
120
120
120
120
140
135
120
120
120
120
120
mm
0.0098
0.0015
0.0099
0.0009
0.0006
0.0101
0.0102
0.0006
0.0045
0.0101
0.0006
0.0005
0.0005
0.0095
0.0004
0.0032
0.0005
0.0090
0.0004
0.0004
0.0071
0.0065
0.0020
0.0003
0.0003
A Structural Application
In this application, the PMs included were the. vertical distance from edge
1 to the weld line, d/, and the horizontal distance from edge 2 to the weld
line, d2. The location of weld lines is considered critical to design a
structurally sound part. From the analysis of variance, it was known that
these PMs depended only on the position of the injection gate, characterized
by variables and y. In order to avoid the repetition described in the full set,
a new dataset was created by varying and y at nine levels creating a finer

Authenticated
Vol. 25, No. 6, 2005

sampling grid for the injection location. The rest of the variables were set to a
value in the middle of their respective ranges. The levels of the controllable
variables for this dataset are shown in Table 8. The total number of
combinations of controllable variables in this dataset was 81.
Table 8
Levels of controllable variables used for the dataset for x,y dependent PMs
1 m
1
w
c
130
275
e
c
154
JC
cm
15
cm
10
11.875
16.25
17.5
13.75
18.75
15.625
17.5
20
21.25
19.375
22.5
21.25
23.75
25
23.125
25
The efficient frontier for this two-dimensional case is shown here in Figure 9.
140 130 120 -
I
N
c
;B
110 100 -
90 -
V)
S.
a
80 -
ffl 70 -
60 -
50
40 -
30
40
50
60
70
80
90
100
110
120
130
Weld Position 1 (mm)
Fig. 9: Visualization of the efficient frontier in the structural application
476

Authenticated
C.E. Castro et al.
Journal of Polymer
Engineering
The seven efficient solutions for a structural part are shown in Figure 10
in terms of the levels of the two PMs in increasing order of dh The
compromise between the positions of the weld lines is confirmed. We want to
maximize both of the weld line positions, but where one of them is at a
maximum, the other is at a minimum.
Efficient solution
Fig. 10:
Efficient solutions for the structural application in terms of the

weld line positions d and d2.
Figure 11 shows the positions of the injection gate corresponding to the

seven best compromises. The entire area shown is the feasible injection
region. In this case the 'attractive' clusters occur at the bottom right corner of
the feasible injection area and along the right edge of the feasible injection
region. These results tend to agree with the large machine capacity case.
Since the locations of the weld lines are independent of the other controllable
variables, any of these x,y pairs would obtain the same results for d, and d2
regardless of the temperature levels. In this case the efficient solutions are
defined by the injection location, so temperature levels are not shown. In
other words, had the temperatures been left at the maximum or minimum of
their respective feasible ranges, we would have arrived at the same results for
the locations of the weld lines.

477
Authenticated
Vol. 25. No. 6. 2005

Fig. 11:
Injection locations of the seven efficient solutions to the structural

application transformed to fall between -1 and 1.
A Quality Critical Application

Appearance in our test part we defined as related to the position of the
weld lines and the flatness of the part, i.e. d,, d2, and Rz. From the analysis of
variance in Table 5 it is known that dt and d2 depend only on the and y
position of the injection gate. However, the temperatures cannot be
disregarded in this case, because Rz depends on all three of them. Therefore,
we used the factorial enumeration already created for all the variables (x, y,
Tm Tw, and Te) with 675 combinations. The resulting sixteen efficient
solutions are shown _ in Figure 12 with respect to the rest of the dataset.
Notice that the data is organized into columns. The different columns
illustrate the repetitions that were referred to earlier. Each of the columns
corresponds to one x,y pair, and the variation in height of the data points in
these columns is determined by the controllable temperatures. Since only one
other PM was involved, Rz, this repetition did not cause a problem.
478

Authenticated
C.F Castro et al.
0.005
I
I
0.01
0.015
40
140
150
Weld Position 1
Weld Position 2
Fig. 12:
Visualization of the efficient frontier for the part quality

application
Figure 13 shows the efficient solutions with respect to the values of the
PMs in increasing order of dt. The compromise between d/ and d2 is again
evident.
Figure 14 analyzes the clusters of the design variables and y. Again, the
entire space shown is the feasible area for the injection gate. This case did not
use the same fine grid for the injection location that was used in the structural
application, so the 'attractive' clusters are not as well defined. However, it is
evident that the right and bottom edges would be the best areas to locate the
injection gate. This case agrees with the previous cases of the large machine
capacity, and the structural application.
Table 9 shows the levels of the controllable variables that correspond to
the efficient combinations of PMs. Notice that for all of the efficient
solutions, the value of Te was 149 degrees C. Allowing the part to cool to a
lower ejection temperature favorably affects the part deflection in the zdirection. In this case, the time to freeze was not considered. Allowing the
part to cool longer did not introduce any negative effects, so the efficient
ejection temperature was always at the minimum of the range.

479
Authenticated

Vol. 25, No. 6, 2005
Position W e l d Line 2, d2 e Deflection R a n g e . R z
- Position W e l d Line 1, d1
0.0050
0.0045
120
0.0040
100
0.0035
0.0030
0.0025
CO
c
S.
0.0020 t3
0.0015
0 0010
0.0005
0 -J
10
r,
11
12
13
14
15
0.0000
16
Efficient Solution
Efficient solutions for the part quality application in terms of the
Fig. 13:
position of the weld lines and deflection range.

1.0
0.8
0.6
0.4
0.2
>
0.0
-0.2
-0 .4 -0.6
-0.8
-1.0
-1.0 -0.8 -0.6 -0.4 -0.2
0.0
0.2
0.4
0.6
0.8
1.0
X
Fig. 14:
Injection Locations of the seven efficient solutions to the part

quality application transformed to fall between -1 and 1.
480

Authenticated
C.E. Castro et al.
Journal of Polymer
Engineering
Table 9
Efficient solutions for the quality critical application
X
Tm
Tw
Te
d2
cm
Rz
mm
dl
cm
mm
mm
15
15
20
20
25
25
25
25
25
25
25
25
25
25
25
25
10
10
10
10
10
10
17.5
25
10
17.5
25
25
10
17.5
10
17.5
130
135
120
130
120
125
125
125
130
130
130
130
135
135
140
140
260
260
260
260
260
260
260
260
260
260
267.5
275
260
260
260
260
149
149
149
149
149
149
149
149
149
149
149
149
149
149
149
149
0.0003
0.0003
0.0007
0.0008
0.0009
0.0009
0.0008
0.0004
0.0011
0.0009
0.0005
0.0006
0.0016
0.0013
0.0023
0.0020
109.7
109.7
1246
1246
1247
1247
94.9
37.9
1247
94.9
37.9
37.9
1247
94.9
1247
94.9
64.4
64.7
67.4
77.8
821
825
1089
13L4
829
11Q0
13 L6
13 L6
83.3
11Q9
83.6
11L8
Injection Location Dependent Performance Measures:

From the results of the analysis of variance we can see that there are some
PMs that are dependent only on the location of the injection gate. These PMs,
'a, te, toe, dh and d2, were considered in a separate case that is only concerned
with determining the location of the injection gate. For this case the factorial
enumeration of the levels of the controllable variables shown in Table 8 was
used. Again because of the high dimensionality of this case, a DEA model
with equal to zero was used. The resulting fourteen efficient solutions are
shown in Figure 15. Notice the compromises between the time to touch hole
A and the time to touch the outer edge. The peaks of these two PMs always
contrast with each other. On the other hand, the trend of the time to touch
hole follows a similar path as the time to touch hole A. As observed before,
the compromise between the weld line positions is evident.

481
Authenticated
Vol. 25, No. 6, 2005

Efficient Compromises
Fig. 15:
Efficient solutions for the case of determining the injection

location in terms of the levels of the PMs considered.
The efficient gate locations for this case are shown in Figure 16. This case
agreed with some of the earlier cases. The 'attractive' clusters for the
injection gate occurred along the bottom and right edges. Only a few new
injection locations resulted from introducing the flow times into this case on
top of the weld line locations, which were previously considered by
themselves in the structural applications. Additionally, these new injection
locations are still in the same general area. This implies that generally the
flow times as a group do not introduce definite compromises with respect to
the location of the injection gate with the locations of the weld lines. Here the
efficient solutions are only dependent on and y, so the levels of the
temperatures are not shown.
482

Authenticated
C.E. Castro et al.
1.0
0.8
0.6
0.4
0.2
>-
0.0
-0.2
-0.4
-0.6
-0.8
-1.0
-1.0 -0.8 -0.6 -0.4 -0.2 0.0 0.2 0.4 0.6 0.8 1.0
X
Fig. 16:
Injection locations of the efficient solutions for the application

considering only x,y dependent PMs
ANALYSIS OF ROBUST SOLUTIONS

The discussion of the different cases in the previous section leads to an
additional analysis: finding robust efficient solutions. Robust solutions can be
found within the individual cases, and some of those were discussed
previously. As can be inferred, a robust efficient solution is a combination of
controllable variable settings that remains efficient when analyzing different
subsets of performance measures. It is also beneficial to determine which
solutions were robust on a large scale, i.e., which combinations of process
variables were deemed efficient in several subsets of optimization. Indeed for
this case it was possible to identify that the combination of (x, y, Tm Tm Te) =
(20 cm, 10 cm, 120 C, 260 C, 149 C) is a robust efficient solution.
Determining a suitable location for the injection gate is a crucial decision.
The temperatures at which the process is run can be adjusted easily. On the
other hand, there is only one chance to decide where the injection gate will be
located. Selecting the proper location from the start can save time and money.
In this study, it was identified that the injection gate location at the top right
483
Authenticated
Vol. 25, No. 6. 2005

corner of the feasible injection area (x=20 cm, y=25 cm) is a robust solution.
This injection location was found in efficient solutions for all but one of the
subsets, and it was very close to the 'attractive' area in the subset in which it
did not appear.
This analysis might help to establish a 'common ground' among multiple
decision makers, to then move to the kind of compromises that can be taken
when presented with the rest of the efficient solutions /8/.
CONCLUSIONS AND FUTURE WORK

Finding the settings of process and design variables in Injection Molding
has been an active area of research. In this work, the coordinated use of CAE,
statistics, neural networks, and data envelopment analysis has been
demonstrated to find these settings in a multiple objective optimization
context. The optimization of a virtual part was presented for discussion, and
several sub cases were defined to further the details of practical applications
in the industry.
The analyses presented in this paper are geared to make informed
decisions on the compromises of several performance measures. These
analyses also allow for the identification of robust variable settings that might
help to define a starting point for negotiation between multiple decision
makers.
Future work will include adding information about the variability of PMs
on the DEA analysis and the determination of process windows with
efficiency considerations.
REFERENCES
1. The Society of the Plastics Industry (2003). spi: Plastics Data Source,
SPI, available at: www.plasticsindustry.org/ [accessed in July 2004]
2. Cabrera-Rios M, Zuyev K, Chen X, Castro JM, Straus EJ (2002) Polymer
Composites 23:5.
3. Cabrera-Rios M, Mount-Campbell CA, Castro JM (2002) Journal of
Polymer Engineering 22:5.
4. Charnes A Cooper WW, Rhodes (1978) European Journal of
Operational Research 2:6.
484
Authenticated
C.E. Castro et al.
5. Castro CE, Cabrera-Rios M, Lilly B, Castro JM, Mount-Campbell CA

(2003) Journal of Integrated Design & Process Science 7:1.
6. Cabrera-Rios M, Castro JM, Mount-Campbell CA (2004) Journal of
Polymer Engineering 24:4.
7. Castro CE, Bhagavatula N, Cabrera-Rios M, Lilly B, Castro JM (2003)
ANTEC Proceedings.
8. Castro CE, Cabrera-Rios M, Lilly B, Castro JM (2004)
ANTEC
Proceedings.
9. Charnes
A,
Cooper
WW,
Lewin
AY,
Seiford
LM
(1993)
Data
Envelopment Analysis: Theory, Methodology and Application, Boston:

Kluwer Academic Publishers.
485
Authenticated

Authenticated

Castro2005 PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Castro2005 PDF

Uploaded by

Copyright:

Available Formats

SIMULTANEOUS OPTIMIZATION OF MOLD DESIGN

AND PROCESSING CONDITIONS IN INJECTION

Welding & Systems

The Ohio State

San Nicolas de los Garza, Nuevo Leon, Mexico,

identification of robust variable settings that might help to define a starting

Simultaneous Optimization of Mold Design and

Vol. 25, No. 6, 2005

be favorable for minimizing cycle

However, letting the part cool to a lower ejection temperature before

Conventional methods of multiple criteria optimization involve combining

C.E. Castro et at.

cannot be improved in one single dimension without harming another. Data

THE OPTIMIZATION STRATEGY

of interest, and to find the

compromises among the PMs for the original system. The

compromises are identified here through DEA.

Simultaneous Optimization of Mold Design and

Vol. 25, No. 6, 2005

models. In addition, when physics-based models are not available to

DATA ENVELOPMENT ANALYSIS (DEA)

Brought to you by | University of Michigan

C.E. Castro et al.

Journal of Polymer Engineering

of lxlO"6. The solutions deemed efficient by the two linearized versions of

DETERMINATION OF SETTINGS OF PROCESS VARIABLES AND

Brought to you by | University of Michigan463

Simultaneous Optimization of Mold Design and

Vol. 25, No. 6, 2005

Fig. 1: Part of constant thickness with cutouts

Fig. 2: Measuring the locations of the weld lines

C.E. Castro et al.

A finite element mesh of the part was created in Moldflow I M in order to

linear regressions were initially considered

as models for the

performance measures. When simple models do not suffice, then more

Brought to you by | University of Michigan

Simultaneous Optimization of Mold Design and

Vol. 25, No. 6, 2005

where is the number of experimental points, y, and y, are the experimental

96.6 98.98 99.45 96.17 99.34 99.57 78.29 85.98 97.4

8.632 7.681 2.031 15.29 10.87 4.325 30.31 16.31 6.766

1.71 27.96 5.65

2.84 33.27 23.64 4.06

Brought to you by | University of Michigan

C.E. Castro et al.

Journal of Polymer Engineering

The complete multiple criteria optimization problem originally posed for

Brought to you by | University of Michigan

Simultaneous Optimization of Mold Design and

Vol. 25, No. 6, 2005

Second Order Interactions

Using the simple modification described above, the number of efficient

Brought to you by | University of Michigan

Journal of Polymer Engineering

C.E. Castro et al.

max= 0 163 MPa

90 100 110 120 130 140

Levels of the PMs that corresponded to the efficient solutions when

Excess Capacity Injection Molding Machine:

Brought to you by | University of Michigan469

Vol. 25, No. 6, 2005

Simultaneous Optimization of Mold Design and