You are on page 1of 18

25 Design of Experiments

Design of experiments, referred to as DOE, is from time to time, because an idea that was originally
a systematic approach to understanding how process a good idea could not be accomplished because the
and product parameters affect response variables technology was not available or the project was
such as processability, physical properties, or product aborted because of lack of time and resources. Since
performance. It is a tool similar to any other tool, the original work, new technology has been developed
device, or procedure that makes the job easier. Unlike or resources are available that now make a solution
quality, mechanical, or process tools, DOE is possible. The advantage of using a DOE approach is
a mathematical tool used to define the importance of that systematic data are generated, summarized, and
specific processing and/or product variables, and how evaluated to definitively determine whether a project
to control them to optimize the system performance should be carried forward or if it is fundamentally
while maximizing properties. DOE uses statistical impossible to resolve and needs to be dropped.
methodology to analyze data and predict product Regardless of whether the DOE results are positive
property performance under all possible conditions (experiment showed desired response) or negative
within the limits selected for the experimental (experiment showed undesired response), it is
design. In addition to understanding how a particular important to complete the project and document the
variable affects product performance, interactions results so that the project will not be recycled at a later
between different process and product variables are time. The DOE results provide an understanding of the
identified. Design of experiments is a technique or processing and/or product parameters and their
procedure to generate the required information with interactions over the experimental space studied.
the minimum amount of experimentation, using the Design of experiments is not the procedure to use to
following: solve production problems on the floor during the
second shift when the product being produced must be
 Experimental limits shipped first thing the following morning to keep your
 Specific experimental conditions customer’s production running. To solve problems
 Mathematical analysis to predict the response at under these circumstances, change one variable at
any point within the experimental limits. a time, slow the production line down, do whatever is
necessary to produce the product to meet the
DOE is used to determine which factors or vari- customer’s requirements. After production is complete,
ables and interactions are significant in contributing to go back and evaluate the situation to determine if
the effect being measured, and those variables and a designed experiment can be used to define the process
interactions that are insignificant and do not limits and capabilities so that the next time the same
contribute to either a particular product property or product is run it can be produced more cost effectively
processing condition. Using DOE saves both time and at higher yields with fewer problems.
money by providing a usable understanding of the
properties and process. The best time to use a DOE is
25.1 DOE Process
during new product or process development, existing
product or process optimization, and while solving The process involved in conducting a successful
technical problems when more than one variable is design of experiments can be broken down into five
present. Where is DOE used? It is used for solving any steps:
technical problem when you want to fully understand
the response to different process or product variables  Define the problem.
that can be changed or controlled during the experi-  Plan the experiment.
mentation. Problems in industry typically get recycled  Run the experiment.

Extrusion. http://dx.doi.org/10.1016/B978-1-4377-3481-2.00025-9
Copyright Ó 2014 Elsevier Inc. All rights reserved. 291
292 E XTRUSION

 Analyze the data by using statistical methods. typical cylindrical shaped pellet. Customer Service
 Report the results. states that it cannot sell an oval shaped pellet
because all customers normally receive cylindrical
shaped pellets. Since the oval shaped pellets are
25.1.1 Defining the Problem free flowing and process the same in both injection
As with the discussion in Chapter 24, this may molding and extrusion equipment, working on the
seem obvious, but in practice, it is sometimes problem to make the pellets cylindrical is a waste
difficult to do. Input from different people may of valuable company resources that can be better
conflict, and deciding on the specific variables to spent on other developmental programs. Fortu-
control may be controversial. Some input may stress nately, the decision is made not to spend time or
the need for a DOE, whereas other suggestions may effort trying to solve the oval pellet problem, and
imply that a DOE is a waste of time. Verify that the the product is accepted in the field without
problem is properly defined and clearly understood comments.
before starting any work. As an example, assume an
extrusion profile at a customer is reported to have
unacceptable dimensional stability and you have 25.1.2 Plan the Experiment
been requested to solve the problem. Although After the problem is properly defined, the second
a DOE to understand the effect of processing step in the DOE process is to select the independent
conditions on part dimensions may be advisable, the variables, with their limits for evaluation, and the
first step is to define what is meant by “unacceptable dependent variables (responses to measure for each
dimensional stability”. Questions to ask are as experiment). Independent variables or factors are
follows: parameters of either processing or product that are set
at specific values (levels) and controlled in the
 Do dimensional changes occur in the x, y, or z experimental design. For extrusion, some potential
direction over time? independent factors are barrel and die temperatures,
 Is one dimension outside the specification limit, screw speed, raw material formulations, drying time,
and is this temperature related? drying temperature, puller speed, and downstream
 Do the parts warp inconsistently from piece to conditions such as roll temperatures, roll gaps, draw
piece? ratio, water temperature, oven temperature, vacuum
 Do the dimensions change in a controlled level, and so forth. Dependent variables or factors are
manner over time? the responses being measured for each experiment to
 Have the dimensions changed because of determine whether the independent variables have an
a change in raw materials? effect on properties or processing conditions being
 Are the dimensional issues due to shrinkage evaluated.
caused by the raw materials? Dependent factors might include physical prop-
 Are the dimensions different from run to run erties (impact, tensile, and flexural properties in the
because of process changes? machine and transverse directions, heat resistance,
and so forth), surface or appearance properties
Before deciding on the proper course of action and (surface roughness, gloss, distinctness of image,
whether a DOE is necessary, make sure the problem color, and transparency), formulation cost in
is properly defined. a product composition experiment, and dimensions
It is critical to fully understand the problem (shrinkage, warpage, thickness, length, and so
before you try to solve it. If the problem is only forth). Many processing variables in extrusion are
viewed as a problem in some people’s eyes, and dependent variables related to the selected inde-
after evaluating the issues you realize that solving pendent process variable; these include melt
the problem will not generate more business, temperature, melt pressure (die and barrel pressure),
improve customer satisfaction with the product, or and extruder torque. Melt temperature may be either
make the product better, you are probably wasting an independent or dependent response, depending
the company’s resources by addressing or spending on how it is used and/or controlled in the experi-
time on the problem. An example is a compounded mental design. In some processes, the melt
product produced in an oval shape rather than the temperature can be controlled and is treated as an
25: D ESIGN OF E XPERIMENTS 293

independent variable; in other processes, the barrel Dependent variables or responses are easy to
temperatures are controlled, and the melt tempera- measure and evaluate when using quantitative tests
ture is measured as a dependent response to the such as tensile, flexural, and impact, which provide
selected barrel temperatures and screw speed. Melt very specific values for each experiment. In some
pressure is normally a dependent variable or situations, responses are more qualitative than
response unless there is a valve to control and set the quantitative. Normal qualitative responses such as
melt pressure, at which time it becomes an inde- pass/fail, good/bad, or yes/no do not provide a good
pendent variable. In deciding whether a factor or response to model. A procedure has to be developed
variable is independent or dependent, ask the ques- to quantitatively differentiate between the different
tion, “Is there a knob to turn to set that variable to samples for that particular response. Instead of
a predetermined value?” If the answer is that no measuring the sample as pass or fail, develop
controller exists to set a particular variable in the a grading system that allows differentiation on
process to a fixed value, it is a dependent variable a basis of 1e10, even if 1e3 are considered failures,
and is measured as a response to factors that are 4e7 are mediocre, and 8e10 are good. The best
preset and controlled. results are obtained under processing conditions that
Independent variables are normally quantitative in produce a 10. The number of dependent variables or
nature, meaning that they are set at a specific factors that can be evaluated is almost limitless for
numerical value. However, in some experimental all experiments as the independent variables are
designs, the independent variables are qualitative, changed. If a particular property response is almost
meaning that a particular switch is on or off, a high or the same at all sets of experimental conditions, no
low setting is used, or a control variable is yes or no. significant effect exists for this property over the
It is better to use quantitative factors if possible. With range of processing or product conditions tested. If
quantitative variables, the high and low levels for each response value used in the analysis for
each factor are defined. As an example, the high/low a particular set of processing conditions is the
screw speed levels for a twin screw extruder experi- average of 10 measurements, the standard deviation
ment may be set at 250 and 150 rpm. This sets the obtained from the measurements at each set of
experimental space for that factor. After the experi- experimental conditions can be evaluated to deter-
mentation stage is complete, mathematical models mine which set of conditions generates the most
predict the response within the 150- to 250-rpm reproducible results.
range. As this is a linear model, it does not predict The number of data points evaluated for each
well outside the experimental limits, such as 100 or experimental response depends on the number of
300 rpm, if the response is nonlinear. Therefore, the tests required for statistical significance, based on the
experimental limits need to be carefully selected in precision and accuracy of the test. Precision is how
the planning stage and set as wide as is experimen- close the measurements are to each other. If the
tally practical. After the experimentation is complete, measurement precision is high, the number of
you do not want to come back and wish the limits samples needed to obtain an accurate mean is smaller
were wider. The points selected to test over the range than if the precision is low. Accuracy is how close the
of the experimental limits are called levels. Going average value for a particular test is to the true value.
back to the screw’s revolutions per minute, if If the accuracy is high, the value for the response in
experimental points are run at 150, 175, 200, 225, the test is very close to the true value; if accuracy is
and 250 rpm, there are five levels. In determining the low, the response is significantly different from the
experimental limits, do a few preliminary experi- true value. It is possible to have high precision and
ments to verify that the experiment can physically be low accuracy.
run when all the independent factors are at either How large does the difference in response vari-
their high or low values. If a particular experiment ables have to be to indicate a significant effect is
cannot be run, for example, at low screw speed and present in the data? This is determined by comparing
low barrel temperature because the torque exceeds the magnitude of the differences between the
the high limit on the extruder, this set of conditions is dependent responses at the various experimental
impractical and another set of low independent points in the matrix. Duplicate or replicate experi-
variables needs to be selected where the process can ments are run either at the center point or some other
be run practically. selected point within the experimental matrix. The
294 E XTRUSION

difference in dependent responses between the 25.1.3 Data Collection


replicated points determines the precision or varia-
tion within the data due to unassigned experimental The third step in the process is data collection.
error. The smaller the differences between the Once the experiments are defined, it is time to go into
response values at the replicated points, the more the laboratory or the production facility to execute
significant larger differences are between responses the experiments and collect the data. Keep in mind
at different experimental points within the design that it is easier to run the experiments in a nonrandom
matrix. This concept will be discussed in more detail order. However, this may introduce inherent errors
later. that can lead to the wrong conclusion. Randomize the
The experiments comprising the DOE are run in experimental order as much as possible and do not
random order to minimize systematic error that can run all the replicates or duplicate experiments one
lead to the wrong conclusions. As an example, after the other. Collect all available processing data
assume that a processing experiment is being run during the experimental phase on both the fixed and
with two different raw materials. On the first day, all manipulated independent variables. Measure the
the experiments are processed using resin A. On the response variables for each experimental test condi-
second day, the same experiments are run using resin tion or treatment combination.
B. If something happened to the dryer, heater bands
are burnt out on the extruder, the extruder or feed 25.1.4 Data Analysis
throat cooling is not operating properly, or the takeoff The fourth step in the DOE process is to analyze
equipment temperatures are higher the second day the data. There are good computer programs avail-
and the malfunction or discrepancy is not noticed, the able to do this. Computer programs predict which
results might indicate that there is a difference independent factors and interactions are significant
between resin A and resin B. If the experiments had (have an important effect) for a particular response,
been run in random order with some of resins A and generate models to predict the dependent response at
B processed on both days, the conclusions from the any experimental point or composition within the
DOE might be quite different. experimental matrix, plot the model equations to
This previous example raises a control question. provide a visual comparison of the data, predict the
Before doing any experimentation, verify that all experimental process condition or composition
components in the extrusion system are operating where the response is maximum, and predict or
properly. The temperature controllers, heaters, ther- define an experimental operating range or composi-
mocouples, and cooling in the various extruder zones tion where the properties meet the specifications or
must be operating properly. Temperature controllers, desired values. If the experimental design is for
heaters, and cooling associated with the downstream optimizing five different properties, one may define
equipment must be operating correctly. After all an experimental operating range where four of the
experimentation is complete, you do not want to have criteria meet the goal but not all five. Possibly only
to apologize for the data because something two out of the five response variables can be opti-
mechanical malfunctioned. DOEs require significant mized at a given set of processing conditions. The
time and effort to do properly, but the results are very five properties can be weighted to define a processing
rewarding, assuming that everything is operating region that maximizes the properties in the order of
properly during the experimental phase. importance. In other words, an operating range is
Independent variables or factors are either fixed or defined where properties 1 and 4 are maximum, 2 and
manipulated during the experimentation. Fixed 5 are slightly below their maximum value, and
variables are not changed in the experimental design; property 3 is average. The operating conditions to
these can be the extruder, downstream equipment, obtain this were selected on the basis that properties 1
raw materials, environmental temperature, feeders and 4 are the most important, followed by 2 and 5,
used for additives, screw design, screen pack, die and then 3.
design, cooling temperature, draw ratio, and so forth. Conclusions are based on statistical analysis and
The other approach is to manipulate the independent confidence levels. Statistics cannot prove that a factor
variables by changing them over a specified range has an effect, but they do verify reliability and test
that defines the experiment and the experimental validity. The confidence limit determines a particular
space. factor has an effect based on a specific level of
25: D ESIGN OF E XPERIMENTS 295

confidence. Most computer programs use a 95% Factorial design is used to screen process and/or
confidence limit as a means of saying a particular product variables to determine which factors are
factor has a significant effect. The confidence level significant in controlling the process. Normally,
can be manually varied if desired. However, 95% is factorial designs are two-level designs, meaning a high
a good confidence limit when determining which and low value for each factor is used to determine
factors or interactions have an effect. whether an effect is present. Fractional factorial
Some computer programs to set up DOE designs designs reduce the experimentation required and also
and analyze the data are Design Expert by Stat-Ease, decrease the information that can be obtained about
Inc. [1], Statgraphics Plus [2], S-Matrix [3], JMP [4], potential interactions between independent variables
and MiniTab [5]. in the experimental space. As the independent vari-
ables increase, the experiments required to understand
which factors and interactions are significant increase.
25.1.5 Report Conclusions Response surface designs are used after several
The fifth and final step in a good DOE is to report independent variables have been identified and one
the conclusions. Results generated but never reported wants a better description of their curvature and
or summarized did not happen. Without a summary interactions in the experimental space. Response
report, the work will probably be repeated in the surface experiments, unlike factorial design, show
future because people will not remember that the curvature. A response surface is a geometric repre-
work was done or the results were generated. With sentation obtained when a response variable is plotted
turnover in personnel, not only is the work forgotten as a function of one or more quantitative independent
but the people who did the work also may have factors. The results are graphed in two or three
moved on to different jobs. If the work is important dimensions, showing the response as a function of the
enough to do, it is important enough to write a report independent factors of interest. Response surfaces
and take credit for a job that is well done. typically replace three-level factorial designs (three-
level factorials contain a low, high, and center value for
each independent variable), because fewer experi-
25.2 Experimental Design ments are required to obtain the same information.
Mixture experiments are a special response surface
There are three experimental designs: mixture experiment.
experiments, factorial design, and response surface. Some terms used in design of experiments are
Each design is used in specific situations to gather defined below:
information from a particular set of independent
variables. Mixture designs are used to optimize  The statistical model is a model based on the
formulations where the independent variables are dependent responses or variables. The model is
ingredients of a mixture or recipe. Formulation based on observations and measurement error.
optimization is based on measuring property perfor-  Observation is the data obtained from the
mance (dependent variables or responses) over dependent responses for each set of independent
different compositions. Typical plastic or polymer variables evaluated.
formulation optimization includes the following:  Data are characterized by mathematical terms;
five terms are as follows:
 Selecting the correct additive level to provide e mean or average
the optimum performance profile for heat, UV e range
stability, or oxidative stability e standard deviation
 Evaluating the best filler concentration in one or e median
more polymers e coefficient of variance
 Identifying the optimum polymer ratio in blends
and alloys The mean is the average value and is defined by
 Determining optimum reinforcement and/or Eqn (25.1):
filler levels Pn
 Optimizing an existing formulation for physical Xi
i1
properties and cost effectiveness X ¼ (25.1)
n
296 E XTRUSION

where  Experimental space (factor space) is the region


defined by the high and low levels for the inde-
xi ¼ Response value pendent factors in an experiment.
n ¼ The number of x values  Interactions occur when the response is different
depending on the settings of two factors. Graphs
The range is the difference between the maximum of response y versus independent variable A at
and minimum response values. the high and low levels of B will appear with
The sample standard deviation (s), given by two nonparallel lines, indicating that the effect
Eqn (25.2), measures the variability within a given of one factor depends on the level of the other.
data set:  Duplicate tests or replications are when an
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
uP entire experimental trial or treatment combina-
un  2
u Xi  X tion with the same independent factor levels is
t
s ¼ i1 (25.2) repeated one or more times to determine the
n1 inherent error in the various tests.
 Degrees of freedom is the number of indepen-
The median is the middle value in a group of dent comparisons available to estimate a param-
samples. The coefficient of variance is a percentage, eter. It is usually the number of model
given by Eqn (25.3): parameters minus 1 or the number of free
s choices to estimate a parameter.
COV or CV ¼  100 (25.3)  Error comes from either assignable causes or
X random causes. Assignable causes are variations
due to changes in the independent factors that
 Experimental design (layout) lists the experi-
are known and can be identified during the
mental trials or treatment combinations. The
experimentation phase. Random errors are
design shows the randomization, replication,
uncontrolled variations that are inherent in the
and independent factor levels within the design.
measurement techniques and the test methods.
 Effect is the difference in the average response
These are unexplained errors that prevent
between two different levels of a given factor
responses from a similar set of experiments
or between two experimental runs. A positive
from producing identical results each time the
effect, shown in Figure 25.1, increases as the
experiment is run.
 Analysis of variance (ANOVA) is a mathematical
procedure to compare the response data with the
ffect error data to determine whether an independent
ive e
Posit variable or interactions are significant. When
Property

No effect analyzing a DOE, ANOVA uses the sum of


Nega
squares to compare effects with the error vari-
tive e ance to determine statistical significance.
ffect
 The F test is used to test for significance of
factors and interactions at a given probability
–A Factor +A
level. The F test is based on a sampling distribu-
Figure 25.1 Effect definition. tion that predicts whether the results are signifi-
cant at a given probability level, based on
a sample variance. At a predefined confidence
independent factor changes from its low to level (75%, 90%, 95%, 99%, and so forth), the
high value. A negative effect decreases as the F test predicts the probability that the F statistic
independent factor increases from its minimum is greater than the F value in an F value table. If
to maximum value. the F value is greater than the value in the table
 Factor effect is the average response for an inde- at a given probability level, the independent
pendent factor at its high value compared to the factor and/or interaction from which the F test
average response for the same factor at its low is calculated is deemed to have a significant
value. effect on the response.
25: D ESIGN OF E XPERIMENTS 297

 Model is an empirical mathematical expression Table 25.1 Degree and Description for Mixture
that describes the response variable in terms of Models
the independent variables and their interactions
Degree Description
within the experimental matrix. Typically,
models are linear, quadratic, or cubic. The Linear ABC.
number of experimental trials in the design Quadratic AB, AC, BC .
must exceed the number of terms in the model Special cubic ABC .
for the model to predict the response at any other
Cubic AB(AB) þ AC(AC)
point within the experimental matrix. A linear
.
response surface model is a combination of
independent factors given by Eqn (25.4): In mixture experiments, the degree of the model
indicates the presence of certain terms. Table 25.1
Yi ¼ C þ Ca A þ Cb B þ Cc C þ . (25.4) shows the degree and model description for linear,
quadratic, special cubic, and cubic models.
where
Mixture experiments are unique, as the sum of all
components equals 100%. This imparts special
Yi ¼ Dependent response variable
restrictions on the experimental points and formula-
C ¼ Constant tions. Taking the high and low values for the exper-
Ca ¼ Factor for independent variable A
imental design points does not work, because the
Cb ¼ Factor for independent variable B combination of these values must add to 100%. In
Cc ¼ Factor for independent variable C
most mixture experiments, a degree 2-quadratic or
degree 3-special cubic model fits most of the data. In
If more than three independent factors are evalu-
data analysis, always use the highest order model
ated, the final model can have as many terms plus the
(model with the most terms) that is statistically
constant as there are independent variables being
significant. Never use a model with more terms than
studied.
the number of experiments in the design matrix or
A quadratic model used to define the response evaluated during the experimental phase of the DOE.
surface is given by Eqn (25.5):
Typical mixture designs are Simplex-Lattice,
Simplex-Centroid, and Simplex Augmentation.
Yi ¼ C þ Linear Terms þ Cab AB þ Cac AC
Terms appearing in the computer analysis are as
þ Cac BC þ . þ Ca2 A2 þ Cb2 B2 þ Cc2 C 2 follows:
þ.  Residuals refer to experimental error.
(25.5)  ANOVA is the analysis of variance.
 Mean square is the sum of the squares for the
where the linear terms are those in the linear model; model divided by the degrees of freedom.
AC, AB, BC, . represent the potential two-factor  Root MSE or root mean square is the square root
interactions; and A2, B2, C2, . represent the potential of the mean square, which is the standard
square terms. deviation.
A cubic model for a response surface series of  R2 is the correlation coefficient. The closer the
experiments is given by Eqn (25.6): R2 is to 1, the better the model predicts the
response behavior.
Yi ¼ C þ Linear and Quadratic Terms þ Cabc ABC  Pure error is the error expected in a response
value if the experiment is completely rerun as
þ . þ Ca2b A2 B þ Ca2c A2 C þ Cb2c B2 C þ .
a replicate.
þ Ca3 A3 þ Cb3 B3 þ Cc3 C 3 þ .
(25.6) To determine if a difference in measured responses
is a true measure of a difference in the process, the
data are analyzed using an ANOVA. Analysis of
In response surface experiments, the quadratic variance is a mathematical method that uses statis-
model will fit approximately 95% of all problems. tical tools to take into account the data and the errors.
298 E XTRUSION

It gives you an understanding of what is a relevant The following output was generated with Mini-
difference versus a difference in data caused by errors TabÔ v. 13.3. From this output, one sees that the F
through the use of an F test. The F test allows you to value is 1.86, which gives a p value of 0.191. This
determine the significance of an independent variable means that there is an 80.9% chance that there is
on some dependent response at a specific confidence a significant difference in the data. The F value for
interval, such as 75%, 90%, 95%, or 99% confidence 95% confidence can be looked up in an F distribution
limits. If the F value calculated in the ANOVA is table for resin with three degrees of freedom, error at
greater than the value in the F tables, the independent 12 degrees of freedom, and 95% confidence limit.
factor or interaction is deemed to be significant at that The F value is 3.49 (Fm1, m2a ¼ F3, 12, 0.05 ¼ 3.49). If
probability or confidence limit [6]. the F value from the MiniTabÔ output is >3.49, then
The math for calculating an ANOVA is beyond the there would be a 95% chance that there is a difference
scope of this book. An effective way to calculate an in the data. Along with the calculations, MiniTabÔ
ANOVA is to use one of the computerized statistical provides a graphical output showing the four resins,
analysis programs. their averages, standard deviations, and a bar graph
As an example, assume that four materials are with the averages and 95% confidence interval. From
being compared for flow through a specific die under the numbers and observing the bar chart, one can see
a fixed set of processing conditions. The four resins that the confidence intervals overlap, which means
are designated as A, B, C, and D. Flow is critical to there is no difference in flow. In this case, one would
obtain the proper molecular orientation in the profile buy the cheapest resin.
and surface aesthetics in the final part. The However, let us assume that you see the large
management is going to purchase one of these four variability in the flow data for each resin and decide
materials, based on the die flow evaluation. The costs to look into the matter. On careful examination, you
of the four resins vary slightly, with A being $1.63/lb, find some transcribing errors and some calculation
B being $1.78/lb, C being $1.68/lb, and D costing errors. The revised data are presented in Table 25.3
$1.72/lb. Four flow rates are measured for each resin, with the corrected data highlighted in bold.
using the same processing conditions, and are given The following revised output (see the next page)
in Table 25.2 in units of lb/h. Based on the results, was generated with MiniTabÔ v. 13.3. From this
what material do you recommend to management for output, one sees that the F value is now 22.53 with a p
purchase and why? value of 0.000. This means that there is almost a 100%
chance that there is a significant difference in the data.
Table 25.2 Experimental Data on Resin Flow Looking at the averages and the bar chart, now we see
Resin Run 1 Run 2 Run 3 Run 4 that resin B has a statistically significant higher flow
than the other three resins. Even though resin B is
A 23.3 19.7 25.9 24.4
9.2% more expensive than the lowest cost resin, it has
B 25.9 25.8 34.3 30.5 a 41.6% higher flow. Thus, one has a case for rec-
C 18.3 28.4 26.4 25.8 ommending resin B because of its higher productivity.
D 23.3 32.8 26.7 26.9 One-way ANOVA: Flow versus Resin

Analysis of Variance for Flow


Source DF SS MS F p
Resin 3 81.9 27.3 1.86 0.191
Error 12 176.5 14.7
Total 15 258.5
Individual 95% CIs For Mean
Based on Pooled StDev
Level N Mean StDev –––––+–––––––––+–––––––––+–––––––––+–
A 4 23.325 2.641 (–––––––––*–––––––––)
B 4 29.125 4.088 (–––––––––*–––––––––)
C 4 24.725 4.425 (–––––––––*–––––––––)
D 4 27.425 3.946 (–––––––––*–––––––––)
–––––+–––––––––+–––––––––+–––––––––+–
Pooled StDev = 3.835 20.0 24.0 28.0 32.0
25: D ESIGN OF E XPERIMENTS 299

One-way ANOVA: Flow 2 versus Resin

Analysis of Variance for Flow 2


Source DF SS MS F p
Resin 3 198.74 66.25 22.53 0.000
Error 12 35.29 2.94
Total 15 234.03
Individual 95% CIs For Mean
Based on Pooled StDev
Level N Mean StDev –––––+–––––––––+–––––––––+–––––––––+–
A 4 24.025 1.473 (–––––*–––––)
B 4 33.350 2.261 (–––––*–––––)
C 4 27.225 1.323 (–––––*–––––)
D 4 25.675 1.654 (–––––*–––––)
–––––+–––––––––+–––––––––+–––––––––+–
Pooled StDev = 1.715 24.0 28.0 32.0 36.0

Table 25.3 Revised Resin Flow Data


thick, extruded on a single screw extruder using
Resin Run 1 Run 2 Run 3 Run 4 a three-roll stack and a flexible lip die. An under-
A 23.3 22.5 25.9 24.4 standing of how processing conditions affect product
B 35.9 32.8 34.3 30.5 properties is desired. The processing conditions
(independent variables) being evaluated are through-
C 28.3 28.4 26.4 25.8
put rate, screw speed, barrel temperature, die lip
D 23.3 25.8 26.7 26.9 opening, and roll temperatures. Physical properties
Corrected data in bold.
and process parameters (response variables) being
measured are color, gloss, impact, tensile properties in
both the machine and transverse directions, melt
The above example shows an ANOVA analysis
temperature, torque, and die pressure. What is the
and how to interpret the results. When doing DOE
best procedure to optimize the process? A factorial
experiments, the ANOVA is performed by the
design provides a systematic approach to determine
computer program. The same computer programs set
which experimental factors and interactions are
up the experimental design and do the math neces-
significant in controlling the different responses being
sary to generate the equations and models. To
evaluated.
summarize the procedure, to do DOE experiments in
The first step is to define the experimental range
extrusion, define the problem, establish the experi-
for each dependent variable. Let us choose the
mental limits, run the experiments, input the depen-
following processing limits:
dent responses into the program to analyze the data,
and interpret the results.
 Throughput rate: 200e300 lb/h
 Screw speed: 250e400 rpm
25.3 Experiments  Barrel temperatures: High and low temperature
profile
The last section in this chapter sets up some  Die lip opening: 0.125e0.250 in. (3.2e6.4 mm)
experimental design examples for different DOE  Roll temperatures: 125e175  F (52e79  C)
experiments.
Using a full factorial design to identify all the
25.3.1 Factorial Design significant effects and interactions, there are 2k
Factorial experiments are used to screen many experiments, where k is the number of independent
independent variables to determine what independent variables. For five independent variables, there are
factors are significant in controlling the performance 25 or 32 experiments. Fractional factorial designs
of various response variables being evaluated. As an require less experimentation but do not define all the
example, consider a plastic sheet 0.125 in. (3.2 mm) interactions that can be present. A 1/2 factorial with
300 E XTRUSION

five independent variables requires 16 experiments, independent variables. To set up the experiments
while a 1/4 factorial requires eight experiments. In using a full factorial design, generate a table with
laying out full factorial experiments, the high and a column for each factor. In the first column, insert
low values for each independent factor are combined the high independent factor value in the first n/2
with the high and low values of all other (where n ¼ 2k) rows and the value for the low

Table 25.4 Full Factorial Experiments for Five Factor Examples

Throughput, Screw Speed, Barrel Die Lip, Roll Temperature,


Experiment
lb/h rpm Temperature,  F in. 
F
1 300 400 Hi 0.250 175
2 300 400 Hi 0.250 125
3 300 400 Hi 0.125 175
4 300 400 Hi 0.125 125
5 300 400 Lo 0.250 175
6 300 400 Lo 0.250 125
7 300 400 Lo 0.125 175
8 300 400 Lo 0.125 125
9 300 250 Hi 0.250 175
10 300 250 Hi 0.250 125
11 300 250 Hi 0.125 175
12 300 250 Hi 0.125 125
13 300 250 Lo 0.250 175
14 300 250 Lo 0.250 125
15 300 250 Lo 0.125 175
16 300 250 Lo 0.125 125
17 200 400 Hi 0.250 175
18 200 400 Hi 0.250 125
19 200 400 Hi 0.125 175
20 200 400 Hi 0.125 125
21 200 400 Lo 0.250 175
22 200 400 Lo 0.250 125
23 200 400 Lo 0.125 175
24 200 400 Lo 0.125 125
25 200 250 Hi 0.250 175
26 200 250 Hi 0.250 125
27 200 250 Hi 0.125 175
28 200 250 Hi 0.125 125
29 200 250 Lo 0.250 175
30 200 250 Lo 0.250 125
31 200 250 Lo 0.125 175
32 200 250 Lo 0.125 125
25: D ESIGN OF E XPERIMENTS 301

independent factor in the second set of n/2 rows. For variable along with the high and low points, requiring
the second independent factor, set the first n/4 rows three experiments for each independent variable.
at the high value, the second n/4 rows at the low This is called a three-level factorial design because of
value, the third n/4 rows at the high value, and the the third factor level. Inclusion of the third factor
final n/4 rows at the low value. The same procedure greatly increases the number of experiments. In the
is followed with the third column, except that the previous factorial design with five variables, there are
first n/8 rows contain the high value and the next n/8 2k or 25 ¼ 32 experiments. The same independent
rows contain the low value, and so forth, until the factors using a three-level factorial design has 3k or
column is full. The fourth factor is treated the same 35 ¼ 243 experiments. Qualitative variables with
way, except that the number of rows alternated is only two levels (high/low, on/off, yes/no) cannot be
n/16. The fifth factor rows alternate every other row, used in a three-level factorial design.
with one being the high value and the other low, or A three-level, two-factor design with 32 or 9
n/32 rows with the high value followed by n/32 rows experiments is shown in Table 25.5, with the high and
with the low value. low values for factors A and B being represented as 1
This procedure produces all possible combinations and 1, and the middle value as 0.
of maximum and minimum values. As three-level factorial designs generate large
Table 25.4 contains the 32 experiments for the treatment combinations or experimental points,
full factorial design to screen the five independent alternate response surface designs were developed to
processing factors. Response variables are collected reduce the number of experiments required while
for all 32 experiments or treatment combinations. providing similar information. Response surface
The response model contains the main effects plus analysis generates models that are used to plot
all the interactions. Factorials or screening experi- contours in one or two dimensions, depending on the
ments are used to screen many variables to deter- factors used to describe the response variable
mine which factors and interactions are significant. throughout the experimental space. The experiments
Fractional factorials are used to determine which in response surface designs are special cases of
main effects are important and the effect of some factorial design that include center points in the
interactions, while minimizing the number of experimental space plus edge center points or face
experiments. Fractional factorials are normally one center points. The multilevel experimental points
of the first steps in an evaluation procedure. The allow fitting the responses to quadratic or cubic
important variables can be identified and then equations, which provides a better model of the
evaluated in more detail. After screening experi- response variable in the experimental space. Designs
ments are completed, the next step is to take the that are more practical than the three-level factorial
significant factors (hopefully a smaller set of designs are Central Composite and Box Behnken
independent variables) and perform a response designs.
surface DOE to define the optimum processing
conditions, or complete a factorial design with
a center point and replicate experiments. Qualitative Table 25.5 Three-Level, Two-Factor Factorial
Design
variablesdsuch as the barrel temperature in the
example, are high or low, on or off, and yes or Experiment Factor A Factor B
nodcan be used in factorial design but not in
1 1 1
response surface designs.
2 1 0
3 1 1
25.3.2 Response Surface
4 0 1
A response surface is a geometrical representation
5 0 0
of a response variable plotted as a function of the
independent variables. This experiment provides 6 0 1
more information about a dependent or response 7 1 1
variable than a two-level factorial or frac- 8 1 0
tionalefactorial design. A three-level factorial design
has a center point included for each independent 9 1 1
302 E XTRUSION

The Central Composite design is based on a two- Centroid


level factorial design with the addition of 2k (k is
the number of independent variables) points (star
points) between the axes plus repeat points at the
centroid. As with all good experimental designs,
the experiments are randomized. Consider an
example similar to that used for the factorial
design, with the exception that the variables are as
follows: Figure 25.3 Box Behnken design.

 Screw speed (250e400 rpm)


experimental space is replicated. A full three-level
 Throughput (200e300 lb/h)
factorial design with three independent variables
 Die lip opening (0.125e0.250 in. [3.2e6.4 mm])
requires 3k experiments or 33 ¼ 27 experiments
(Figure 25.4). In addition to the high and low
What are the experiments and what is the
appearance of the experimental space? Figure 25.2

Star points
In plane Out of plane

Figure 25.4 Three-level factorial.

Experimental points in the experimental space

Figure 25.2 Central Composite design with star experimental values, the three-level factorial design
points. includes center points on each edge, the center point
of each face, and the center point plus replications to
determine experimental error. A three-level factorial
shows the experimental space for a three-factor design to obtain similar information as a Central
Central Composite design. Depending on the Composite or Box Behnken design contains 32
design, the star points or the center points in each experiments.
face may extend out past the face, as shown in Using the example to determine the effect of
Figure 25.2. In most plastic or extrusion designs, screw speed, throughput, and die lip opening on
the star points have to be located in the plane with physical properties, the randomized experimental
the other coordinate points (shown in the left-hand layout for the three designs is given in Tables
drawing of Figure 25.2), because star points 25.6e25.9, with Tables 25.6 and 25.7 showing
outside the plane may not be practical processing Central Composite design with the star points in
conditions. In the current example, high values and out of the plane, corresponding to the left- and
associated with the star points on the throughput right-hand schematics in Figure 25.2. Table 25.8
may exceed the torque or screw speed limitations provides a Box Behnken design, and Table 25.9
of the equipment; the very high screw speed shows the three-level factorial design. In Table
would be about 450 rpm. In any design, all 25.7, the star points that are not within the plane
experimental points must be physically possible to represent experiments 3, 4, 6, 9, 13, and 16;
process. experiment 4, making a 0.125-in.-thick sheet
The other common response surface design is product with a 0.08-in. die opening, is probably
a Box Behnken design, where the experimental impossible to achieve. Experiment 9, requiring
points are located at the center points of the edges 250 lb/h throughput rate with a 199-rpm screw
(Figure 25.3). The centroid or center point of the speed, may be impossible to run without torqueing
25: D ESIGN OF E XPERIMENTS 303

Table 25.6 Central Composite Design with Star Points in the Plane

Experiment Screw Speed, rpm Throughput, lb/h Die Opening, in.


1 325 250 0.25
2 250 300 0.12
3 400 300 0.25
4 250 250 0.19
5 400 200 0.25
6 325 250 0.19
7 325 250 0.19
8 325 200 0.19
9 250 300 0.25
10 325 250 0.19
11 250 200 0.25
12 325 300 0.19
13 325 250 0.19
14 400 250 0.19
15 325 250 0.19
16 325 250 0.19
17 400 200 0.12
18 400 300 0.12
19 250 200 0.12
20 325 250 0.12

out the extruder. Experiment 6, with a 451-rpm response surfaces is that fewer experiments are
screw speed, may exceed the extruder screw speed required.
capabilities. The die opening in experiment 13 may Using either the Central Composite or Box
be too large to draw down the sheet to 0.125-in. Behnken design to fit a cubic model requires more
thickness without excessive molecular orientation experimentation so that there is at least one
or necking in the sheet. Central Composite exper- experiment for each term in the model. Since
iments laid out in Table 25.6 are more practical to quadratic models fit 95% of all experimental
use for the experimental design. The center points designs, starting with a quadratic model makes
repeated in Table 25.6 are experiments 6, 7, 10, 13, sense. Additional experiments can be added later if
and 16. Table 25.8 shows the experimental points a higher order model is required to predict the
using a Box Behnken design which has 17 responses.
experiments versus 20 required for the Central
Composite designs. As with the other designs, the 25.3.3 Mixture Experiments
experimental order is randomized, and the center
point is repeated, experiments 2, 5, 10, 13, and 17. Mixture experiments are used when experiment-
Table 25.9 shows the experimental points with ing with composition ingredients. Typical uses
a three-level factorial design with the center point include the following:
replicated six times in experiments 4, 8, 13, 19, 21,
and 31. The advantage of using either the Central  Formulation optimization and optimizing prop-
Composite or Box Behnken designs to generate erty profiles with different additive levels
304 E XTRUSION

Table 25.7 Central Composite Design with Star Points Outside the Plane

Experiment Screw Speed, rpm Throughput, lb/h Die Opening, in.


1 325 250 0.19
2 325 250 0.19
3 325 334 0.19
4 325 250 0.08
5 325 250 0.19
6 451 250 0.19
7 400 200 0.25
8 325 250 0.19
9 199 250 0.19
10 250 300 0.25
11 325 250 0.19
12 250 200 0.12
13 325 250 0.29
14 400 200 0.12
15 250 300 0.12
16 325 166 0.19
17 325 250 0.19
18 400 300 0.12
19 250 200 0.25
20 400 300 0.25

 Combinations of different resins in alloys or you want to understand the proper proportions of
blends each ingredient to optimize tensile strength, flexural
 Determination of impact modifier concentrations modulus, and impact strength. The experimental
 Flame retardant formulation optimization limits are defined as follows:
 Effect of filler and/or reinforcement concentrations
 Effect of regrind levels  D is present from 38% to 72%.
 E ¼ 5%e85%.
Experimental responses include the following:  Talc ¼ 10%e55%.

 Physical properties What experiments are required to establish the


 Cost property profiles at any composition within the
 Flame retardance design matrix, and what is the appearance of the
 Chemical resistance experimental space?
 UV stability Figure 25.5 shows the experimental space for
 Thermal stability a mixture experiment with three variables. Trian-
 Dimensional stability gular coordinates are used to satisfy the mixture
constraint, which requires all components to add to
As an example of a mixture experiment, assume 100% for each experimental condition. Any point
that resin D, resin E, and talc are being mixed, and within the experimental matrix in Figure 25.5
25: D ESIGN OF E XPERIMENTS 305

Table 25.8 Box Behnken Design

Experiment Screw Speed, rpm Throughput, lb/h Die Opening, in.


1 400 300 0.19
2 325 250 0.19
3 400 250 0.12
4 400 200 0.19
5 325 250 0.19
6 250 250 0.25
7 400 250 0.25
8 325 200 0.12
9 250 200 0.19
10 325 250 0.19
11 325 200 0.25
12 325 300 0.12
13 325 250 0.19
14 325 250 0.19
15 250 250 0.12
16 325 300 0.25
17 325 250 0.19

totals to 100%. The horizontal axis starting at the Once the dependent responses are determined and
bottom of the triangle is 0% talc, with each models are generated, a processing or compositional
horizontal line representing a 10% change until the window can be defined where the dependent
apex is reached at 100% talc. The other axes are responses are better. In most situations, it is
interpreted in the same manner. Consequently, any a compromise, because all response variables or, in
point in the experimental space is a combina- this case, physical properties are not normally
tion of the three components. Using the three- maximized at the same compositional or processing
component coordinate space in the mixture example, point. Optimization of one dependent response may
the experimental space is defined by the limits shown be at a completely different location within the
in Figure 25.6. Combinations of independent experimental space compared to the maximum for
factors to define the response variables within the next dependent response. A decision is required
the experimental space are the vertices, the mid- to determine where the best property balance is
points of long edges, the centroid, and a check- obtained within the experimental matrix. If
point. With the duplicate experiments included a composition is selected where all the responses are
to define the experimental error, the coordinates not at their maximum value, does the compromise
of the experimental compositions are given in meet the customer requirements?
Table 25.10. Thirteen experiments provide sufficient In any factorial, response surface, or mixture
data to generate models and response curves experiment, once the optimum combination of pro-
to predict the dependent response at any compo- cessing conditions or compositional ingredients is
sition within the experimental limits. Experiments chosen based on a predicted response, additional
1 and 2, 4 and 8, 5 and 13, and 8 and 12 are experimentation is needed to verify the predictions
repeated experiments used to determine experimental and verify the response profile predicted is routinely
error. attainable in practice.
306 E XTRUSION

Table 25.9 Three-Level Factorial Design

Experiment Screw Speed, rpm Throughput, lb/h Die Opening, in.


1 250 200 0.12
2 250 250 0.25
3 400 250 0.19
4 325 250 0.19
5 400 200 0.12
6 400 300 0.12
7 400 250 0.25
8 325 250 0.19
9 325 250 0.25
10 400 200 0.19
11 325 200 0.12
12 250 300 0.25
13 325 250 0.19
14 250 300 0.12
15 325 200 0.25
16 250 250 0.19
17 250 200 0.19
18 325 300 0.25
19 325 250 0.19
20 325 300 0.19
21 325 250 0.19
22 250 300 0.19
23 400 300 0.19
24 250 250 0.12
25 400 200 0.25
26 325 200 0.19
27 250 200 0.25
28 325 250 0.12
29 325 300 0.12
30 400 250 0.12
31 325 250 0.19
32 400 300 0.25
25: D ESIGN OF E XPERIMENTS 307

Talc = 100% Table 25.10 Mixture Experiment Example


Experimental point is
40% Talc, 30% Resin E, Talc, Resin Resin
and 30% Resin D Experiment
% E, % D, %
1 32.5 29.5 38.0
2 32.5 29.5 38.0
%

Re
=0

s
3 10.0 35.0 55.0

in
E

D
in

=0
4 10.0 52.0 38.0
s
Re

%
5 10.0 18.0 72.0
6 30.6 17.4 52.0
Re
s

0%
in

7 39.0 5.0 56.0


D

10
=1

8 55.0 7.0 38.0

E=
Talc = 0%
00

sin
%

Re 9 10.0 52.0 38.0


10 20.3 34.7 45.0
Figure 25.5 Experimental coordinates for a three-
11 23.0 5.0 72.0
component mixture experiment.
12 55.0 7.0 38.0
13 10.0 18.0 72.0
5
E=

Talc
sin
Re

Re
sin
D=
38

Talc=55
References
X [1] Design Expert. Stat-Ease, Inc., 2021 E. Hennepin
Re
sin

Avenue, Minneapolis, MN.


85
Ex

X
D=

E=
pe

[2] Statgraphics Plus. Manugistics, Inc., 2115 East


72

X
sin
rim

X Jefferson Street, Rockville, MD.


sp

Re

X
en

X
ac

Talc=10 [3] CARD. S-Matrix Corporation, 317 Third Street,


tal
e

X X X Suite 5, Eureka, CA.


Re

[4] JMP. SAS Institute, Inc., SAS Campus Drive,


E
sin

sin
D

Cary, NC.
Re

[5] MiniTab. MiniTab, Inc., State College, PA.


Figure 25.6 Mixture experiment exampled [6] H.F. Giles, SPE Plastics Technician’s Toolbox,
composition points and definition experimental Society of Plastics Engineers, Brookfield, CT,
space. 2002 (Chapter 1) 50.
308 E XTRUSION

Review Questions
1. What is the definition of experimental space?
2. What are three experimental designs and where is each used?
3. Why use a DOE?
4. Where should DOE experiments be used and where are they inappropriate?
5. What are the five steps in doing a DOE? What is the objective of each step?
6. What is the difference between independent and dependent variables? Give some examples in
extrusion applications.
7. What is the difference between an independent variable and a factor?
8. Why is it critical to run experiments in random order?
9. Why is it necessary to run duplicate experiments or repetitions?
10. What is a statistical model and how is it used?
11. What are four mathematical terms describing variability, and how do they differ from one another?
12. What is the definition of an effect? What is the difference between positive, negative, and no effects?
13. Why is an F test important in DOE?
14. How is analysis of variance used in DOE?
15. In response surface experiments, what terms are included in quadratic, linear, and cubic models?
16. In mixture experiments, what is the difference between quadratic, special cubic, and cubic models?
17. What type of experimental design can use qualitative variables?
18. How many experiments are required in a full two-level factorial with six variables, in a 1/2 factorial,
and in a 1/4 factorial?
19. What is lost by doing fractional factorials?
20. What is a three-level factorial design and how many experiments are required with four independent
variables?
21. What is the difference between a Central Composite and a Box Behnken design?
22. What constraint is present in mixture experiments, and how does this affect the experimental space?
23. How are data analyzed in DOEs?

You might also like