Application Note

The Optimization of Injection Molding Processes

Using Design of Experiments
Problem
Manufacturers have three primary goals: 1) produce goods
that meet customer specifications; 2) improve process
efficiency in cycle times, labor cost and energy consumption;
3) increase process robustness by reducing sensitivity to small
changes in process or material parameters. To meet these
goals, manufacturers often employ one-variable-at-a-time
investigations to construct models relating independent
process parameters to product quality. The models are used
to devise iterative process modifications that lead to a process
having a reasonable trade-off between process efficiency
and product quality. Unfortunately, this method of process
development does not explicitly test the robustness of the
process.
Figure 1 - Process window definition
Background
A manufacturing process must exhibit an adequate breadth The identification of robust process settings is best
of operating conditions, sometimes referred to as the process accomplished with Design of Experiments (DOE) approaches.
window [1]. Figure 1 shows the process window for an DOE is a structured, efficient method that simultaneously
injection molding process in terms of the injection velocity investigates multiple process factors using a minimal number
and pack pressure. The size and shape of the process window of experiments [4-6].
is determined by certain constraining boundaries. Boundaries
1, 2 and 5 are defined by the machine’s injection velocity Consider: To treat n = 10 independent variables at m = 2
and pack pressure limits; 3 defines the combined upper levels requires 210 or 1024 experiments for a full high/low
limits of injection velocity and pack pressure beyond which evaluation. The prospect of performing such a high number of
unacceptable levels of flash are produced; and 4 determines experiments is daunting, especially in a manufacturing setting.
the lower limits of pack pressure and injection velocity The use of DOE drastically reduces the size of experimental
beyond which undesirable short shots and sink marks occur. matrices while retaining a balance in the points tested in
Viable production processes lie within the process window parameter space. This balance permits valid statistical analyses
determined by these boundaries. of the results and determination of the main effects of process
parameters.
Process set-points that might be chosen by a plastics manu-
facturer are shown as circles within the process window. A very useful DOE approach, D-optimal designs, can be
While the robustness of the process can be tested by changing generated using a numerical optimization technique that
one process variable, depending on the initial settings, such maximizes the “volume” of the investigated process parameter
investigations can lead to different process limits. Such studies space (as measured by the determinant of the design matrix
also ignore the possibility of interactions between the process multiplied with its transpose) [6]:
variables which are common, and ignoring them can result D = det(X' X) (1)
in selecting process conditions too close to a process window This application note will use injection molding production
boundary (denoted by *, Figure 1). data to show how DOE can be used to determine process
robustness and how it can be effective in identifying process
While such investigations may lead to operable processes, they
window boundaries and their relation to process settings.
do not necessarily lead to robust processes nor do they provide
reliable, quantitative models of the process that can be used for
yield prediction, process optimization and quality control [2,
3].
Page 2
Application Note
SOLUTION Run 1 2 3 4 5 6 7 8 9 10

Evaluating DOEs for Minimum Runs 1 -1 -1 -1 -1 -1 -1 -1 -1 1 1

2 -1 -1 -1 1 -1 1 1 1 -1 1
In the DOE approach, a set of conditions for a stable injection 3 -1 -1 1 -1 1 1 1 -1 -1 1
molding process that produces acceptable molded parts with
4 -1 -1 1 1 1 -1 -1 1 1 1
repeatable part weights and thicknesses is first determined.
5 -1 1 -1 -1 1 1 -1 1 -1 -1
Then ranges are defined around these process settings to
6 -1 1 -1 1 1 -1 1 -1 1 -1
encompass the expected long term variation for, in this
7 -1 1 1 -1 -1 -1 1 1 1 -1
example, 10 important process factors (Table 1). These factors
8 -1 1 1 1 -1 1 -1 -1 -1 -1
are the independent variables in the process model. The
9 1 -1 -1 -1 1 -1 1 1 -1 -1
combination of minima and maxima of these different factors
10 1 -1 -1 1 1 1 -1 -1 1 -1
(coded -1 and +1) are used to set up the experimental designs.
11 1 -1 1 -1 -1 1 -1 1 1 -1
# Factor Min Max 12 1 -1 1 1 -1 1 -1 -1 -1 -1
1 Pack Time (s) 2 4 13 1 1 -1 -1 -1 1 1 -1 1 1
2 Material Blend (%) 0 40 14 1 1 -1 1 -1 -1 -1 1 -1 1
3 Barrel Temperature (c) 195 205 15 1 1 1 -1 1 -1 -1 -1 -1 1
4 Coolant Temperature (c) 25 35 16 1 1 1 1 1 1 1 1 1 1
5 Injection Velocity (mm/s) 60 80
Table 2.1 - 16 run, 2 10-6
matrix
6 Pack Pressure (MPa) 60 80
7 Shot Size (mm) 20.75 20.75 Run 1 2 3 4 5 6 7 8 9 10

8 Back Pressure (MPa) 20 26 1 1 -1 1 1 1 1 1 -1 1 1

2 -1 1 1 1 -1 1 -1 -1 -1 -1
9 Cooling Time (s) 6 10
3 -1 -1 -1 -1 1 1 1 -1 -1 1
10 Screw Speed (RPM) 100 150
4 1 1 -1 -1 -1 1 -1 -1 1 -1
Table 1
5 1 -1 -1 1 1 -1 -1 1 -1 -1
6 -1 1 -1 1 1 1 1 1 1 -1
Tables 2.1 – 2.4 show the four design approaches evaluated
in this study. Note that the 6-run design in Table 2.4 is not 7 -1 -1 1 -1 -1 -1 -1 1 1 -1

normally used as all factors are confounded and such designs 8 -1 -1 -1 1 -1 1 -1 1 1 1

do not normally produce reasonable models. Occasionally, 9 1 1 -1 1 -1 -1 1 -1 -1 1

however, such designs can perform well, as is discussed below. 10 1 1 1 -1 1 1 -1 1 -1 1

MKS’ SenseLink QM was used for data acquisition, analysis, 11 1 -1 1 -1 -1 1 1 1 -1 -1

and quality control. Table 2.2 - 11 run, D-optimal matrix

The results of regression analysis are shown in Figure 2. The Run 1 2 3 4 5 6 7 8 9 10

dark line in each graph is the best estimate of the main effects 1 -1 -1 -1 -1 -1 -1 -1 -1 1 1
derived using an aggregation of all DOEs. Figure 2 shows that 2 -1 -1 -1 1 -1 1 1 1 -1 1
regression models for the fractional factorial design and the 3 -1 -1 1 -1 1 1 1 -1 -1 1
D-optimal design closely match the best estimate while those 4 -1 -1 1 1 1 -1 -1 1 1 1
of the 6- and 8-run DOEs fail. This is because the 6- and 5 -1 1 -1 -1 1 1 1 1 1 1
8-run DOEs suffer from confounded parameters, (pack time 6 1 1 1 1 1 1 -1 1 -1 -1
with pack pressure; material blend with shot size; back pressure 7 -1 1 1 -1 -1 -1 1 1 1 -1
with cooling time and screw RPM in the 6-run DOE). The
8 -1 1 1 1 -1 1 -1 -1 -1 -1

Table 2.3 - 8 run, 2 7-4

DOE
Page 3
Application Note
Run 1 2 3 4 5 6 7 8 9 10 Multivariate analyses (MVA), on the other hand, can
1 1 -1 -1 1 -1 1 -1 -1 1 -1 determine main effects from limited data. Principal
2 1 1 -1 -1 -1 1 1 -1 -1 -1 Components Analysis (PCA) and Projection to Latent
3 -1 1 -1 1 -1 -1 1 -1 1 -1
Structures (PLS) methods evaluate process behavior using
4 -1 -1 1 -1 1 -1 -1 1 -1 1
Distance to the Model (DModX) and Hotelling t-squared (T2)
5 1 1 1 1 -1 1 1 1 1 -1
scores [8]. Figures 3.1 to 3.4 show the loadings scatter plots
derived using the DOEs in Table 2. Each loadings plot shows
6 1 1 -1 1 1 1 1 -1 1 1
the amount of the observed process behavior modeled by
Table 2.3 - 6 run, oversaturated DOE the first two principal components due to each process state.
Process states further from the origin have greater influence
8-run DOE provides a better estimation of the main effects, on the process behavior. The plot also indicates the correlation
however, the effect of pack time, cool time, and screw pack of each process state with respect to other process states. For
time is not investigated and cool time and screw RPM are example, the upper right quadrant of Figure 3.1 includes the
confounded with other factors. weight, thickness, and other process states calculated from
These results demonstrate the need for extensive experimental the machine process traces (i.e. M10, M11, M15, M17, and
designs for the estimation of main effects when using multiple M40 correspond to screw velocity during packing, average
regression techniques. pack pressure, injection energy, and integral of pressure during
packing, respectively).

Figure 2 - DOE regression models

Page 4
Application Note

Figure 3:1 - 16-run DOE Figure 3:2 - D-optimized DOE

Figure 3:3 - 8-run DOE Figure 3:4 - 6-run DOE

Evaluating DOEs for Robustness

Inspection of the loadings plots shows that the 11-run DOE An 8 run, 3 factor full factorial design (Table 3) was used to
(Figure 3.2) to be similar to the best estimate 16-run DOE investigate process robustness. Melt temperature, T, injection
(Figure 3.1) even though less data was used (thickness, weight, velocity, v, and pack pressure, P were chosen to test short shot
M2, M18, M15, and other states are similarly placed). The behavior. Run number 8 in Table 3 is the standard process
two models should therefore provide similar predictions. The with all settings at their high value. Low values were chosen so
loadings plot for the 8-run DOE (Figure 3.3) and the 6 run short shots would occur when any two of the process factors
DOE (Figure 3.4) vary significantly from the best estimate. were jointly operated at their lower settings. Several common
Specifically, the loadings plot for the 8 run DOE places M2, process faults were imposed on the process in the validation
a process state for pack time, very close to the origin, and also DOE of Table 3. Figure 4 shows the distance from the process
clusters many other process states together. The 8-run DOE data to the models generated from the four experimental
fails because the effect of pack time was not investigated. designs and a model generated from the aggregate data. Higher
values of DModX indicate process deviation from the reference
model. The results show that the 5 most critical faults were
detected by the models for the 41-run aggregate dataset, the
16-run DOE, 11-run DOE, and 6-run DOE. The model for
the 8-run DOE failed since: 1) all of the DModX values are
Page 5
Application Note
Barrel Temp Injection Velocity Pack Pressure
Run
(C) (mm/s) (MPa)
1 172.5 15 15
2 172.5 15 70
3 172.5 70 15
4 172.5 70 70
5 200 15 15
6 200 15 70
7 200 70 15
8 200 70 70

Table 3

very high, and 2) the DModX only captured 3 of the 8 defects.

The 6-run DOE performed surprisingly well, given that each
of the ten process factors were confounded with one other
process factor; the performance far surpassed that of the 8-run
DOE despite fewer runs. The multivariate analysis was thus
able to model the principal components effectively.
The factorial DOE developed to evaluate the robustness of
multivariate models generated 80 observations for the part
Figure 4 - Process fault diagnostics
weight. These data were imported into a multivariate analysis
program (SIMCA® P+ v. 11, Umetrics) and the main effects of Acceptable process windows were defined using multivariate
the process settings on the part weight were calculated (Figure analysis to build the model with interaction terms (Umetrics
5). The results showed that pack pressure, P, injection velocity, MODDE® v.7). The model predicted the product weight
v, and melt temperature, T, were all significant. as two of the factors were varied over a specified range while
holding the third factor constant (Figures 7.1-7.3). The
However, the correlation coefficient for the model, R2, was calculated product weights greater than 0.21 grams were found
only 0.81, implying that 81% of the process behavior was not to produce a short shot. These were deemed to be within
accounted for by the model. This could be significantly the acceptable process window. No upper limit was specified.
improved.
The behavior of the predicted product weights as the process
The model was therefore expanded to include interaction strays from the set point (the lines emanating from the grey
terms. The analysis yielded the main and interaction effects circles in Figure 7) is instructive. The predicted magnitude of
shown in Figure 6, and the R2 value for the model was displacement for a single process variable that shifts the process
0.99. Thus this model is a much better estimate than the out of the acceptable window is significantly greater than
purely linear model. It should be noted that the addition the shift predicted for the simultaneous displacement of two
of interaction terms significantly alters the model behavior. variables. This implies that it is more likely that defects will
Specifically, the effect of injection velocity, which was strong occur through small changes in multiple process factors than
in the purely linear model, becomes insignificant. Instead, the through large changes in a single process factor. This likelihood
part weight is greatly reduced only when the injection velocity can be quantified. If we define process limits as the number
and melt temperature are changed in combination. Other of standard deviations (s) from set point before reaching the
interaction terms were also found to be significant. The impact boundary of the defined process window:
of this change can be fully appreciated when one considers
its implications for the definition of the acceptable process (2)
window.
Page 6
Application Note
Tables 3.1 to 3.3 show values for n in the injection molding barrel temperature and pack pressure; ~5.8% for combinations
process in which the standard deviation values for the barrel of lower injection velocity and pack pressure. If the constraints
temperature, injection velocity, and pack pressure were were tighter or the process variation was higher, then the
assumed to be 5°C, 20 mm/s, and 10 MPa, respectively. The likelihood of short shots would be significantly greater. The use
probability of a short shot occurring can be estimated for the of multivariate analyses shows that the process is far less robust
univariate analysis as: than is indicated by the univariate analyses. The multivariate
Prob =1− normsdist (n) (3) approach is therefore well suited as a means of measuring the
robustness of a process within a defined process window.
where normsdist is the standard normal cumulative distribution
function.

For the multivariate analysis, the number of standard

deviations for each of the two process states, n1 and n2, were
first calculated from the process settings along the bold arrows
in Figures 7.1-7.3. Assuming mutual independence of the
process factors, the joint probability of the short shot occurring
in the multivariate analysis is then given by:
·
Prob =1− normsdist (n1) normsdist (n2) (4)

Barrel Temp (C) Inj Velocity (mm/s)

Probability
Limit # Std Dev Limit # Std Dev
Univariate 200 0 10 3 0.135%
Univariate 170 6 70 0 0.000%
Multivariate 187 2.6 28 2.1 2.244%

Table 4.1 Temperature and Velocity Limits at P=70 MPa

Figure 5 - The (scaled) main effects of the process settings on the part
Barrel Temp (C) Pack Pressure (MPa)
Probability weight in the estimation of robustness
Limit # Std Dev Limit # Std Dev
Univariate 200 0 18 5.2 0.000%
Univariate 170 6 70 0 0.000%
Multivariate 186 2.8 23 4.7 0.256%

Table 4.2 Temperature and Pressure Limits at V=70 mm/s

Inj Velocity (mm/s) Pack Pressure (MPa)

Probability
Limit # Std Dev Limit # Std Dev
Univariate 10 3 70 0 0.135%
Univariate 70 0 22 4.8 0.000%
Multivariate 38 1.6 43 2.7 5.808%

Table 4.3 Velocity and Pressure Limits at T=200°C

The results in Tables 4.1-4.3 indicate a significant discrepancy

between the predictions of univariate analysis compared
to those of multivariate analyses. The univariate analyses
indicate only one very small probability of producing short
shots (0.135% for velocity violations of the process window
Figure 6 - The main and interactive effects of the process settings on
boundary — Figs. 7.1 and 7.3); no short shots were predicted the part weight in the estimation of robustness
for any values of melt temperature or pack pressure. By
comparison, the multivariate analyses predict a much higher
probability of short shots: ~2.2% for combinations of lower
Page 7
Application Note

Figure 7:1 - Temperature-velocity window; Figure 7:2 - Temperature-pressure window; Figure 7:3 - Pressure-velocity window;
P=70 MPa V=70 mm/sec T=200°C

Conclusion
The number of characterization experiments needed to develop
reliable model-based process and quality control is a serious
impediment to implementation in plastics manufacturing.
This note shows how the effective use of Design of
Experiments can facilitate the development of models through
the use of robust approaches such as multivariate analyses.
Principal Components Analysis and Projection to Latent
Structure methods can employ data from oversaturated DOEs
to effectively predict process faults and to characterize the
variational effect of many process factors simultaneously.
While such DOEs are inadequate for effective multivariate
regression analysis, PCA models provide reasonable fidelity
to setup the process and evaluate its robustness. The case
study for plastics injection molding in this report shows how
neglecting the interaction affects results in poor identification
of the process boundaries and gross over estimations of the
process robustness.
Page 8
Application Note
REFERENCES
1. D. Kazmer, Plastics Manufacturing Systems Engineering,
Hanser Verlag, p. 352-288 (2009).
2. D. Kazmer and C. Roser, "Evaluation of Product and
Process Design Robustness," Research in Engineering
Design, 11, 20, (1999).
3. D. O. Kazmer, S. Westerdale, and D. Hazen, "A
Comparison of Statistical Process Control (SPC) and
On-Line Multivariate Analyses (MVA) for Injection
Molding," International Polymer Processing, 23, 447,
(2008).
4. G. E. P. Box and N. R. Draper, Empirical Model Building
and Response Surfaces, (1986).
5. R. H. Myers and D. C. Montgomery, "Response Surface
Methodology: Process and Product Optimization Using
Designed Experiments," in Wiley Series in Probability and
Statistics: Wiley Interscience, p. 248 (1995).
6. R. H. Myers, D. C. Montgomery, G. Geoffrey Vining,
C. M. Borror, and S. M. Kowalski, "Response Surface
Methodology: A Retrospective and Literature Survey,"
Journal of Quality Technology, 36, 53, (2004).
7. J. S. Jung and B. J. Yum, "Construction of Exact D-optimal
Designs by Tabu Search," Computational Statistics & Data
Analysis, 21, 181, (1996).
8. S. Wold and M. Josefson, "Multivariate Calibration of
Analytical Data," Encyclopedia of Analytical Chemistry
(Meyers RA, ed). Chichester, UK: John Wiley & Sons,
pp. 9710–9736, 2000.

For further information, call your local MKS Sales Engineer or contact the MKS Applications Engineering Group at 800-227-8766.
SIMCA® and MODDE® are registered trademarks of MKS Instruments, Inc., Andover, MA.

App. Note 06/12 - 7/12 MKS Global Headquarters

© 2012 MKS Instruments, Inc. 2 Tech Drive, Suite 201
All rights reserved. Andover, MA 01810
978.645.5500
800.227.8766 (within USA)
www.mksinst.com