Determining Optimal Moulding Process Parameters by Two

Level Factorial Design with Center Points

S.Rajalingam1* Awang Bono2 and Jumat bin Sulaiman3

1
School of Engineering and Science, Curtin University, CDT 250, 98009,
Miri, Sarawak, Malaysia, Email:[email protected]
2
Universiti Malaysia Sabah, JalanUMS, 88400, Kota Kinabalu, Sabah,
Malaysia, Email:[email protected]; [email protected]

ABSTRACT

Determining optimal process parameter setting critically influences

productivity, quality and cost of production in the injection moulding
industry. Previously production engineers used trial and error method to
determine optimal process parameter setting. Inappropriate machine
parameter settings can cause production and quality problems. In this paper
the authors used a case study to investigate the moulding machine
parameters which will affect the dimensions (length and width) in a plastic
component. The machine process setting in use currently caused variations in
the dimensions exceeding the specification limit. Therefore the experiment is
needed to identify the optimal machine parameter setting which could be set
to maintain the dimensions closest to the target value with smallest possible
variation. A design of experiments (two level factorial design with center
points) was conducted to study the effect of three injection moulding process
parameters (mould temperature, injection speed and injection pressure)
versus dimensions (length and width). Finally, the optimal process
parameters to maintain the dimensions closest to the target values were
identified. Statistical results and analysis are used to provide better
interpretation of the experiment. The models are form from ANOVA and the
models passed the tests for normality and independence assumptions.

Keywords: Process, Parameters, Response, Factors, ANOVA, Target Values,

Moulding.

63
1. Introduction

In recent years, the communication products like cell phone are widely
applied throughout the world. The designs of cell phone have a tendency to
be thin, light and small and more convenient style. Therefore, the shapes of
smart cell phone are changing, and more features have to be tightly packed
into smaller volumes within the housing. In order to procure more space for
the tightly packed components inside, the wall thickness of the housing parts
must originally be reduced to less than 3 mm in thickness [1]. Plastic
injection moulding is one of the most important procedures applied for
forming an injection-moulded thermoplastic part with a thin-shell feature.
Machining parameters, moulding material, product and mould designs are
major factors affecting the quality of thermoplastic parts produced by
injection moulding. Due to the complexity of injection moulding, numerous
mathematical models have been proposed and extensively developed by a
growing numbers of studies for the analysis of different stages of the
injection moulding process. The process of injection moulding includes
filling, packing, cooling, opening the mould cavity, injecting and closing the
mould cavity. However, the selection of appropriate machining parameters
for the injection moulding operation becomes more difficult as the wall
thickness of plastic parts gets thinner [2].

Warpage and shrinkage are among the most significant defects of thin shell
thermoplastic part in terms of quality in the process of injection moulding.
The level of warpage and shrinkage is highly related to the injection machine
parameters setting. Leo and Cuvelliez [3] investigated the influence of the
packing parameters and gate geometry on the final dimensions of a moulded
part by experiment. Huang and Tai [4] used the computer simulation and the
experimental design of the Taguchi method to analyze the effective factors
of warpage in an injection moulded part with a thin-shell feature. Ozcelik
and Erzumlu [5] explored the determination of efficient minimization of
warpage on thin-shell plastic parts by integrating the response surface
method and genetic algorithm. Recently, the simulation of the injection
moulding process provides visual and numerical feedback of the part
behaviour and eliminates the traditional trial and error approach for
optimization. Proper interpretation of the results from simulation can help
selecting a suitable material; reduce cycle time and costs on mould
modification. Patcharaphun,et al. [6] used a commercial software package
(Moldflow) to predict the fiber orientation distribution within the weldline
area of push–pull-processing parts. Song, et al. [7] had applied the
orthogonal experiment method and numerical simulation software

64
(Moldflow), discussed the influence of different process parameters on the
molding process for ultra-thin wall plastic parts.

The role and implications of DOE

Design of Experiment (DOE) was invented by R.A. Fisher in England in

early 1920. It was Fisher’s idea that it was much better to vary all the factors
simultaneously in what he called a “factorial design”. The purpose of
experimental design is to minimize the number of experiments (runs)
required to identify which experiment is significantly related to the desire
output variables (responses). This allows for a large number of variables to
be studied and analyzed easily and economically. For further information on
DOE refer to the work done by of Montgomery [8]. Additionally, there were
several papers addressing the importance of DOE. Cupello [9] stated that
DOE is one of the few reliable and effective tools for developing globally
competitive products and processes. Rowand [10] argued that employees
from equipment operators to the plant managers could benefit from the use
of DOE. At the same time, various industries have employed the DOE
method over the years to improve products and processes, increase
productivity and reduce costs. It is a powerful and effective method to solve
challenging quality problems [11, 12]. Many quality techniques and concepts
such as statistical process control (SPC) are directed toward improving
quality at the downstream processes and operations by emphasizing control
charts and inspection. DOE, on the other hand, emphasizes the upstream
processes and operations, focusing on developing products and processes
that are well-designed.

The role of DOE in the optimization of plastic injection moulding

The role of DOE in the optimization of plastic injection moulding processes

has received attention in the literature. The nature of DOE applications as
well as their goals has been quite different. The following represents the
literature related to the broad use of DOE in the plastic injection moulding
industry. Mapleston [13] in a brief paper indicated that the optimization has
produced dramatic improvements in injection moulding processes at a
company. Morgan [14] argued that the optimization process for injection
moulding industry does not need to stop at the quality control department in
order for the full range of benefits to be obtained. Speight [15] discussed the
moulding process control system that resulted in a faster turnaround, higher
quality, and lower costs for revising a design and modification of a mould.

65
Objective of the study

In this paper the authors used a case study to investigate the injection
moulding machine parameters which will affect the dimensions (length and
width) of a plastic cell phone shell. The machine process setting currently
caused variations in the dimensions exceeding the specification limit.
Therefore the experiment is needed to identify the optimal machine
parameter setting which could be maintain the dimensions (length and width)
closest to the target value with smallest possible variation. Statistical results
and analysis are used to provide better interpretation of the experiment.

2. Methedology

Factorial design is a collection of mathematical and statistical techniques

which used in the modelling and analysing of problems. The objective is to
optimise the response [8]. Factorial design also quantifies relationships
among one or more measured responses and the input factors [16].

In this study, the approximation of the mathematical model will be proposed

using the fitted second-order regression model. The necessary data for
building the response model are generally collected by the experimental
design. The experimental design adopts the full factorial design with center
point. The experimental factorial design is combinations of the factors at two
levels (high and low) with two center points. The factorial design is a
sequential procedure for determining the optimal injection machine
parameters[16,17].

In order to determine if there exist a relationship between the factors and the
responses, the data must be analysed in a statistically manner using
regression. A regression is performed in order to describe the data collected
whereby an observed, response variable is approximated based on a
functional relationship between the estimated response ( and factors
( , ,…, .In this case study exist a second order relationship between
response and three factors. A second order equation,
error, was used to describe
the functional relationship between the estimated response, and the
factors , , and .The least square technique is being used to fit a model
equation containing the said factors by minimising the residual error
measured by the sum of square deviations between the actual and the
estimated responses. This involves the calculation of estimates for the
regression coefficients, i.e. the coefficients of the model variables including
the intercept or constant term. The calculated coefficients or the model

66
equation need to however be tested for statistical significance. In this respect,
the following tests are performed by using Design Expert Software Version 7
(Stat-Ease Inc., Minneapolis, USA) [16].

Analysis of variance (ANOVA)

The method used to compare the magnitude of estimated effects of factors

with the magnitude of experimental error is called ANOVA. If the magnitude
of a factor effect is large when compared with experimental error, it is
decided that the changes in the selected response cannot occur by chance and
those changes in the response can be considered to be the effects of the
factors. The factors causing a variation in the response are called significant.
In this study, F-test was used in the ANOVA.

Test for significance of the regression model

This test is performed as an ANOVA procedure by calculating the F-ratio,

which is the ratio between the regression mean square and the mean square
error. The F-ratio, also called the variance ratio, is the ratio of variance due
to the effect of a factor (in this case the model) and variance due to the error
term. This ratio is used to measure the significance of the model under
investigation with respect to the variance of all the terms included in the
error term at the desired significance level, α. A significant model is desired.

Test for significance on individual model coefficients

This test forms the basis for model optimisation by adding or deleting
coefficients through backward elimination, forward addition or stepwise
elimination/addition/exchange. It involves the determination of the P-value
or probability value, usually relating the risk of falsely rejecting a given
hypothesis. For example, a “Prob. > F” value on an F-test tells the
proportion of time you would expect to get the stated F-value if no factor
effects are significant. The “Prob. > F” value determined can be compared
with the desired probability or α-level. In general, the lowest order
polynomial would be chosen to adequately describe the system.

Test for lack-of-fit

The test statistic for lack-of-fit is the ratio between the lack-of-fit mean
square and the pure error mean square. As previously, this F-test statistic can
be used to determine as to whether the lack-of-fit error is significant or
otherwise at the desired significance level, α. Insignificant lack-of-fit is

67
desired as significant lack-of-fit indicates that there might be contributions in
the factor–response relationship that are not accounted for by the model.
Additionally, checks need to be made in order to determine whether the
model actually describes the experimental data. The checks performed here
include determining the various coefficient of determination, . These
coefficients have values between 0 and 1. The lack-of-fit must be
insignificant. The various coefficient of determination, values should be
close to 1.

Graphical residuals analysis

The adequacy of the model is also investigated by the examination of

residuals [8]. The residuals, which are the difference between the respective,
observe responses and the predicted responses are examined using the
normal probability plots of the residuals and the plots of the residuals versus
the predicted response. If the model is adequate, the points on the normal
probability plots of the residuals should form a straight line. On the other
hand the plots of the residuals versus the predicted response should contain
no obvious patterns.

3. Expeimental Set-up

The material used in this experiment was commercially available

polycarbonate (PC) material. The material was pre-conditioned at 120 for
four hours using a dehumidifying drier before moulding.In this experiment
three factors are being studied and their levels are given in Table 1. The
average value of a factor is equal to center point. The three factors are mould
temperature, injection pressure and screw rotation speed and it is labeled as
A, B and C respectively.

The levels of the factors determined according to our experience about the
process and from the literature research. The injection moulding machine
process setting in use currently caused variations in the dimensions
exceeding the specification limit of length and width of cell phone shell. The
specification limit for the length and width are 93.49 ± 0.2 mm and 45.93 ±
0.2 mm respectively. The objective of this experiment is to identify the
optimal machine parameter setting which could be set to maintain the
dimensions closest to the target values with smallest possible variation. The
target values for length and width are 93.49 mm and 45.93 mm respectively.

68
 Table 1: Experimental control factors and levels

Factor Units Low Level (-) High Level (+)

A. Mould Temperature 85 95
B. Injection Pressure kg/cm2 2250 2400
C. Screw Rotation Speed mm/Sec 110 140

A full two level factorial experimental design with center point was carried
out to study on how the above three factors will influence the responses. The
number of run (combination of machine parameter setting) needed according
to full two-level factorial experimental design method [8,17] for three factors
was 2 =8 and the run was repeated twice. At the same time the authors like
to add another two centre points to provide sufficient information on possible
curvature in the system. Therefore a total 18 experimental runs were required
for these study. At each run, the team leader will record the dimensions of
the responses. The dimensions were measured by using digital smart scope
machine. Experimental design matrix constructed according to standard
order rule which was given in Table 2. Experiments were executed randomly
to provide protection against the extraneous factors, which could effect the
measured responses. In all experimental runs, the reaction time was
considered as 5 min. The resulting responses values are shown in Tables 2 as
well. The actual experiment was conducted in the factory with some help
from the staff, taking one working day to be completed.

4. Results and Discussion

The experimental results as per the experimental plan are shown in Table 2.
These results were input into the Design Expert software version 7 for
further analysis [16]. Without performing any transformation on the
responses, the half normal plot Figure 1-2 revealed for both responses
(length and width). The half normal plots shows the effects of factors and the
factors lie along the line are negligible. Figure 1 show that, five factors seem
to be significant for length those are A, B, C (main factors) AB and AC (two
factor interactions). Figure 2 show, four factors seem to be significant for
width those are A, B C (main factor) and AC (two factor interactions). The
main effect B is the most significant factor associated with length and width.

69
 Table 2: Experimental design matrix and results
Random Standard Factors Responses
order order A B C Length (mm) Width (mm)
1 16 95 2400 140 93.470 45.822
2 12 95 2250 140 93.440 45.813
3 8 95 2400 110 93.455 45.823
4 11 95 2250 140 93.443 45.813
5 10 85 2250 140 93.429 45.800
6 14 85 2400 140 93.443 45.810
7 3 95 2250 110 93.428 45.815
8 7 95 2400 110 93.456 45.822
9 17 90 2325 125 93.487 45.844
10 5 85 2400 110 93.442 45.821
11 4 95 2250 110 93.427 45.812
12 6 85 2400 110 93.441 45.820
13 15 95 2400 140 93.468 45.820
14 2 85 2250 110 93.426 45.810
15 18 90 2325 125 93.486 45.844
16 1 85 2250 110 93.428 45.810
17 13 85 2400 140 93.441 45.808
18 9 85 2250 140 93.428 45.800

Figure 1: Half normal plot for length

70
 Figure 2: Half normal plot for width

ANOVA analysis

Table 3 shows the ANOVA table for response length (after backward
elimination). The value of “Prob. > F” in table for model is less than 0.05
which indicates that the model is significant, which is desirable as it
indicates that the terms in the model have a significant effect on the
response. Table 3 show the resulting ANOVA table for the improved model
for response length. By selecting the backward elimination procedure in
Design Expert software it will automatically reduce the terms that are not
significant. These insignificant model term (BC) can be removed and result
in an improved model. Results from Table 3 indicate that the model is still
significant. However, the main effect of mould temperature (A), injection
pressure (B) and speed (C) and the two-level interaction of mould
temperature and injection pressure (AB) and mould temperature and speed
(AC) are the significant model terms. The lack-of-fit can still be said to be
insignificant.

71
 Table 3: The ANOVA table for the length (after backward elimination)

Source Sum of df Mean F Prob > F

Squares Square
Model 3.06E-03 5 6.11E-04 493.25 < 0.0001 Sig.
A 7.43E-04 1 7.43E-04 599.5 < 0.0001
B 1.74E-03 1 1.74E-03 1407.24 < 0.0001
C 2.18E-04 1 2.18E-04 175.65 < 0.0001
AB 1.89E-04 1 1.89E-04 152.64 < 0.0001
AC 1.63E-04 1 1.63E-04 131.24 < 0.0001
Curvature 3.59E-03 1 3.59E-03 2898.35 < 0.0001 Sig.
Residual 1.36E-05 11 1.24E-06
Lack of Fit 6.25E-07 2 3.13E-07 0.22 0.8095 not sig.
Pure Error 1.30E-05 9 1.44E-06
Cor Total 6.66E-03 17
Std. Dev. 1.11E-03 R-Sq 0.9956
Mean 93.45 Adj R-Sq 0.9935
C.V. % 1.19E-03 Pred R-Sq 0.9884
PRESS 3.56E-05 Adeq Prec. 85.37

The same procedure is applied on response width and the resulting ANOVA
table for the improved model is shown in Table 4. For width, the main
effects of mould temperature (A) injection pressure (B) and screw rotation
speed (C) and the two-level interaction of mould temperature and screw
rotation speed (AC) are the significant model terms. The lack-of-fit also said
to be insignificant. Both model from ANOVA Table 3 and 4 shows that the
value is high, close to 1, which is desirable. The predicted is in
reasonable agreement with the adjusted . The adjusted value is
particularly useful when comparing models with different number of terms.
Adequate precision compares the range of the predicted values at the design
points to the average prediction error. Ratios greater than 4 indicate adequate
model discrimination. In this particular case the value is well above 4.The
following equations are the final models equation in terms of actual factors
for both responses:

Length = 95.361 - 0.025(A) - 6.86E-4(B) - 3.579E-3(C) + 9.167E-6(A B) +

4.250E-5(AC)

Width = 45.994 - 3.300E-(A) + 6.083E-5(B) - 3.121E-3(C) + 3.250E-(AC)

72
 Table 4: The ANOVA table for the width (after backward elimination)

Source Sum of Sq. df Mean Square F Prob > F

Model 7.99E-04 4 2.00E-04 188.87 < 0.0001 Sig.
A 2.33E-04 1 2.33E-04 219.96 < 0.0001
B 3.33E-04 1 3.33E-04 315.01 < 0.0001
C 1.38E-04 1 1.38E-04 130.58 < 0.0001
AC 9.51E-05 1 9.51E-05 89.91 < 0.0001
Curvature 1.63E-03 1 1.63E-03 1544.99 < 0.0001 Sig.
Residual 1.27E-05 12 1.06E-06
Lack of Fit 3.19E-06 3 1.06E-06 1.01 0.4337 not sig.
Pure Error 9.50E-06 9 1.06E-06
Cor Total 2.45E-03 17

Std. Dev. 1.03E-03 R-Sq. 0.9844

Mean 45.82 Adj R-Sq 0.9792
C.V. % 2.24E-03 Pred R-Sq 0.9669

Graphical residuals analysis

The normal probability plots of the residuals and the plots of the residuals
versus the predicted response for length and width are shown in Figures 3–6.
A check on the plots in Figures 3 and 5 revealed that the residuals generally
fall on a straight line implying that the errors are distributed normally. Also
Figures 4 and 6 revealed that they have no obvious pattern and unusual
structure. This implies that the models proposed are adequate and there is no
reason to suspect any violation of the independence or constant variance
assumption.

Figure 3: Normal probability plot of residuals for length.

73
 Figure 4: Plot of residuals vs. predicted for length.

Figure 5: Normal probability plot of residuals for width.

Figure 6: Plot of residuals vs. predicted for width

74
Examine main effect and interactions

Figure 7: Interaction graph of A versus B (length)

Figure 8: Interaction graph of A versus C (length).

Figure 9: Interaction graph of A versus C (width).

75
 Figure 10: Main effect B for width

In the interaction graph those points that have non overlapping intervals are
significantly different and the points overlapping intervals not significantly
different. Figure 7 show the interaction graph of AB for length. The spread
of the points on the right side of the graph (where mould temperature is high)
is higher than the spread between the points at the left side of the graph
(where mould temperature is low.) In other words, the effect of length is
more significant at the high level of mould temperature (A+) and high
injection pressure (B+) to increase the length current dimension to target
value (93.49 mm). Figure 8 show the interaction graph of AC for length. The
points on the right side of the graph (where mould temperature is high) is
significantly different and the points at the left side of the graph (where
mould temperature is low) are not significantly different. In other words, the
effect of length is more significant at the high level of mould temperature
(A+) and high level of speed (C+) to increase the current length dimension to
target value (93.49 mm). Figure 9 show the interaction graph of AC for
width. The points on the left side of the graph (where mould temperature is
low) is significantly different and the points at the right side of the graph
(where mould temperature is high) are not significantly different. In other
words, the effect of width is more significant at the low level of mould
temperature (A-) and low level of speed (C-) to increase the current width
dimension to target value (45.93 mm). Figure 10 show that the effect of main
factor effect injection pressure (B) with width. It is very clear that factor B
do not have any significant effect on the response width. Therefore
economically it is very clear that the width can be maintained close to target
value (45.93 mm) by reducing the injection pressure (B-).

76
Figures 7 - 10 show that the effect (main and interaction factor effect)
does not have any significant effect on the responses length and width
because both line (red and black) almost super impose each other and the
gradient of the lines do not have any significant different. In other word,
Figure 7 – 9 show that the gradient of the red and black lines are almost
same and the lines are almost horizontal. Therefore economically it is
very clear that the response dimensions can be maintained close to target
values (93.49 mm and 45.93 mm) by setting the three injection machine
parameters mould temperature, injection pressure and screw rotation
speed at lower level (A-, B- and C-).If look carefully the Figure 7 -10 it is
clear that the most nearest points (Design Points) to the target values of
the length and width are obtainable when the factors A, B and C are at
middle of the experimented range, that is when A is 90 , B is 2325
km/cm2 and C is 125 mm/sec. The same observation can also be made
from the 3D surface graphs for length and width which are shown in
Figure 11 – 12. Therefore practically the authors decide that the center
parameter setting (90 , 2325 km/cm2 and 125 mm/sec) is the optimised
parameter to achieve the target dimensions for responses length and
width.

Figure 11: 3D view AB interaction for length

77
 Figure 12: 3D view AC interaction for width.

Confirmation Run

Before switch to entire manufacturing operation to the above setting and

produce high volume of parts, need to some verification runs. For this
experiment the authors decided to produce 25 samples. A different lot
Polycarbonate (PC) material was used in this verification run. The
purpose of this verification run was to validate that the center parameter
setting (A is 90 , B is 2325 km/cm2 and C is 125 mm/sec) is the
optimised parameter to achieve the target dimension. The result show
(Figure 13 -14) that the 25 samples the length and width are near to target
values.

Figure 13: Verification Run for response – Length

78
 Figure 14: Verification Run for response - Width

5. Conclusion

In this paper a DOE approach (2-level factorial design with center point)
for optimising the injection moulding process was investigated. The
injection moulding machine parameters which will affect the dimensions
(length and width) in a plastic cell phone shell studied. The significant
factors affecting the length and width of the cell phone shell were
identified from ANOVA. The optimal process parameters to maintain the
dimensions closest to the target value were identified (mould temperature
is 90 , injection pressure 2325 km/cm2 and screw rotation speed is 125
mm/sec) from interaction graphs and 3D views. Statistical results and
analysis are used to provide better interpretation of the experiment. The
models are form from ANOVA and the models passed the tests for
normality and independence assumptions. Confirmation run with the
above center parameter setting determined that target dimensions for
responses length and width were achievable. It is noted here that the
results obtained in this study were quite satisfactory for the concerned
industry since they were able to achieve the target values for the length
and width of cell phone shell.

Acknowledgement

The main author gratefully acknowledges the financial support of Curtin

University of Technology, Sarawak Campus, Research and Development
Department. He also would like to thank Prof. Awang Bono and A/Prof.
Jumat Sulaiman from University Malaysia Sabah, Mr. Prem Kumar and

79
Mr. Loh Kim Long for their contribution and technical support of the
work.

References

1. Thin T (1996) Asian plastics news, July/August, Wiley, New

York, pp 12–14
2. Crawford RJ (1989) Plastic engineering, 2nd edn. Pergamon,
Oxford
3. Leo V, Cuvelliez Ch (1996) The effect of the packing parameters,
gate geometry, and mold elasticity on the final dimensions of a
molded part. Polym Eng Sci 36:1968–1977
4. Huang MC, Tai CC (2001) The effective factors in the warpage
problem of an injection-molded part with a thin shell feature. J
Mater Process Technol 110:1–9
5. Ozcelik B, Erzumlu T (2005) Determination of effecting
dimensional parameters on warpage of thin shell plastic parts
using integrated response surface method and genetic algorithm.
Int Commun Heat Mass Transf 32:1085–1094
6. Patcharaphun, S., Zhang, B., & Mennig, G. (2007). Simulation of
three-dimensional fiber orientation in weldline areas during push–
pull-processing. Journal of Reinforced Plastics and Composites,
26(10), 977–985.
7. Song, M. C., Liu, Z., Wang, M. J., Yu, T. M., & Zhao, D. Y.
(2007). Research on effects of injection process parameters on the
molding process for ultra-thin wall plastic parts. Journal of
Materials Processing Technology, 178(1–3), 148–153.
8. D.C. Montgomery, Design and Analysis of Experiments, 6th ed.,
Wiley, New York, 2005.
9. Cupello, J.M. (1999), “Training technologies in experimental
design”, Research Technology Management, Vol. 42 No. 5, pp.
47-50.
10. Rowand, R. (1988), “A matter of survival”, Automotive News,
September 5, p. E38.
11. Sofuoglu, H. (2006). A technical note on the role of process
parameters in predicting flow behavior of plasticine using design
of experiment. Journal of Materials Processing Technology,
178(1–3), 148–153.
12. Yang, Y. K. (2006b). Optimization of injection-molding process
of short glass fiber and polytetrafluoroethylene reinforced

80
 polycarbonate composites via design of experiments method: A
case study. Materials and Manufacturing Processes, 21(8), 915–
921.
13. Mapleston, P. (1999), “Real-time process control is said to
provide perfect shots”, Modern Plastics, Vol. 29 No. 8, pp. 29-30.
14. Morgan, C. (1999), “Auxiliary equipment: a management issue”,
Molding Systems, Vol. 57 No. 2, pp. 34-8.
15. Speight, R.G. (1998), “Molding process optimization comes of
age”, Molding Systems, Vol. 56 No. 8, pp. 20-3.
16. Design-Expert Software, Version 7, User’s Guide, Technical
Manual, Stat-Ease Inc., Minneapolis, MN, 2000.
17. Myers RH, Montgomery DC (1995) Response surface
methodology: process and product optimization using designed
experiments. Wiley, New York
18. D.D. Steppan, J. Werner, R.P. Yeater, Essential regression and
experimental design for chemists and engineers, 1998.
http://geocities.com/ iliconValley/Network/1032/CGPage1.html.

81
82