Professional Documents
Culture Documents
Melissa L. Bowles
lntel Corporation
Douglas c. Montgomery
Department of Industrial Engineering
Arizona State University
Tempe. Arizona 85287
239
Copyright ~ 1997 by Marcel Dekker, Inc.
I
240 BOWLES AND MONTGOMERY
suppose that there are [hree variables (catalyst concemra of the th ingredient inthe mixture. Then, in light of [he
tion, reaction time, and temperarure) that affect the viscos aboye discussion, we must require that
ity of a chemical product. The levels of temperarure may
Xj ~ O, i = 1, 2, ... q, (1 )
be chosen independently of the levels of reaction time and
catalyst concentra[ion, and, consequently, we may [hink of and
[he region of experimentation as either a cube or a sphere .
The experimental design will consist of an appropriate set
of points over this cuboidal or spherical region (such as a
t
i=1
xj = XI + Xl + .. . + x q = 1. (2)
factorial design or a central composite design).
In a mixture experiment, the factors are the ingredients Constraint (2) makes the . levels of the factors X I
or components of the mixture, and [he response is a func nonindependent, and as a result, mixture experiments are
lion of the proportions of each ingredient. The proportion different than the usual types of experiments in which fac
ate amounts of each ingredient are typicalIy measured by torial (or fractional factorial) designs and response surface
weight, by volume, by mole ratio, and so forth. To illus designs are employed.
trate, consider the margarita formulation problem described Constraints (1) and (2) are shown graphicaIly in Figure
earlier . The formulation consis[s of blending aIl four ingre 1 for q = 2 and q = 3 components. For two components,
dients in a container in an effort to find an optimum (or at the feasible factor space for the mixture , experiment in
least a pleasing or effective) blend. Because the volume of eludes aH values of the two components for which Xl + x 2
the container is fixed, the experimem might consist of test = 1, which is the line segment shown in Figure la . With
ing various combinations of the four ingredients, where alI three components, the feasible space for the mixture experi
blends have the same volume. ment is the triangle in Figure 1b that has vertices corre
Let Xl' X 2' X 3' and X 4 represent the pro].ortiollS bl vol sponding to pure blends (i.e., mixtures made up of 1(){)%
ume o[ the. four.~r~-margarita mix, tequii, Triple
Sec, and ltme JUlce. Then sorne blends that mightbe of
interest are as follows:
Blend 1: XI =: 0.50 x2=: 0,25 x 3 = 0.15 x 4 = 0.10
Blend 2: Xl = 0.50 ~= 0.30 X3 = 0.10 x 4 = 0 .10 %2
,.
i :
.,: ..!
,;
.+ TUTORIAL ON EXPERIMENTS WITH MIXTURES 241
~f a .single ingredient) and edges that are binary blends. In an ingredient be present in the blend or that an ingredient
general, the experimental region of a mixture problem with cannot exceed a specified maximum proportion. The
q components is a simplex, which is a regularly-sided fig margarita formulation problem is a constrained mixture
ure with q vertices in q - l dimensions. For the margarita problem, because practical experience with the system in
problem, the simplex region is a regular tetrahedron; that dicates mat each one of the four ingredients must lie in a
is, a geometric figure with q = 4 vertices and each of the relatively narrow range for the beverage to be satisfactory.
four faces is an equilateral triangle . We will discuss and illustrate constrained mixture experi
The coordina te system for mixture proponions is a sim ments, using the margarita fonnulation problem as a prac
plex coordinate system. For example, with q = 3 compo tica! example. In sorne mixture problems, there are process
nents, the experimental region is shown in Figure 2. Each variables 2, 22' ... , zp in addition to the mixture ingre
of the three vertices in the equilateral triangle correspond dients. For example, in the margarita problem, suppose
to apure blend, and each of the three sides of the triangle that the serving temperature of the beverage and whether
represents a mixture that has only two of the three com or not it is served in a glass with a salted rim are al so
ponents. The nine grid Iines in each direction mark off considered as experimental factors. These are process vari
10% increments in the respective components. Interior ables, where z is the temperature and Z2 could take on the
points in the triangle represem mixtures in which all three value of -1 for a salted rim and + 1 for an unsa/ted rim.
ingredients are present at nonzero proportionate amounts. In a mixture-amount experiment, the response is a function
The centroid of me triangle corresponds to the mixture with of not only the component proportions but also the amount
equal proportions XI = Jj, x 2 = Jj, and x) = Jj of all of the mixture that is used. Such a problem could arise in
ingredients. developing an adhesive, where the pull-off force could be
In this article, we discuss experimental design, data a function of the adhesive fonnulation and the amount of
analysis, and model-building techniques for the original adhesive applied to the test parts. For more information on
mixture experiment defined by Eqs. (1) and (2). In this these and other types of experiments with mixtures, see
problem, the entire simplex region is investigated and the Refs . l and 2.
only design variables are the mixture components x,
x 2 ' . , x(r There are many variations of the original
mixture problem. One of these is me addition of upper and Simplex-Type Designs and Canonical Mixture
lower bounds on sorne of the component proportions. Polynomials
These upper or lower bounds occur beca use the problem
may require that at least a certan minimum proportion of
The prirnary differences between a standard factorial or
response surface experiment and a mixrure experiment are
that (1) a special type of design must be used and (2) the
form of the mixture polynomial is slightly different from
the "standard" polynomials used in factorial designs and
cJassical response surface work. In this section, we intro
duce designs that allow an appropriate response surface
model to be fit ayer the entire mixture space. Because the
mixture space is a simplex, al! design points must be at the
vertices, on the edges or faces, or in the interior of a sim
plex.
The two primary design types are the simplex-Iattice and
the simplex-centroid designs. A simplex /attice is just a
uniformly spaced set of points on a simplex. A {q, m} sim
plex-lattice design for q components consists of points
defined by the following coordinate settings: The propor
tions taken on by each component are the m + l equally
spaced values from O lo 1,
Figure 2. Trilinear coordinate system. Xi = O, 11m, 21m, ... ,l. i = 1,2 , ... , q, (3)
...
and all possible combinations (mixtures) of che proportions triangle. These binary blends are made up of equal parts
from chis equation are used. As an example, Jet q = 3 and of two of the three ingredients. Figure 3 also shows the {3,
m = 2; consequentiy, 3}, the {4, 2lo and [he {4, 3} simplex-lattice designs.
The notaton {q, m} mplies a simplex-lattice design in q /
Xi = O. l/, 1, j = 1,2.3,
components that will support a mixture polynomial of
and the simplex-Ianice design consists of the following six degree m.
points: A q component simplex-centroid design consists of
2q - 1 distinct design points . These design points are the
(XI' x2' x3) = (1, O. O), (O. l. O). (0, 0, 1), (lh, lh. O). q pennutations of (l, O, O, ... , O) or single component
(1/2,0, lh), (0, lh, l/).
blends, the (~) pennutation,s of (1h. 1/2 , O, . .. , O) or all
This is a {3, 2} simplex-lattice design, and it is shown in binary mixtures, the (~) pennutations of (,X', ,X', ,X', ... ,J
Figure 3a. The three vertices (1, 0, O), (0,1, O), and (O, O), and so forth and the overall centroid (1/q, lIq,
O, 1) are pure blends, and the points (1h, Ih, O), (lh, 0, lIq, . .. , llq). Figure 3b shows the simplex-centroid de
11z), and (O, 112, 112) are binary blends or tv.;o-component signs for q = 3 and q = 4 components . Note that the
mixtures located at the midpoints of the three edges of the design points are located at the centro id of the (q - 1)
x, =1
.x, - 1 ,r
I
r, - 1
(b) .r, = I
(i) (i i)
Figure 3. (a) Sorne simplex-Iattice designs for p = 3 and p = 4 cornponents. (b) Simplex-centroid designs with three and four
components (i) p = 3 and (ii) p = 4.
..
~
TUTORIAL ON EXPERIMENTS WITH MIXTURES 243
dimensional simplex and at che centroids of all the lower FuIl Cubic:
simplexes contained wichin che original simplex .
Now consider the types of models that can be fit when
the experimenter has performed a mixrure experiment. In
general, a first-order model in the design variables XI'
q
x2 ' . , xq is
+ LL oijxx/x - Xj)
q I<} j-2
Po I P;x;-
E(y) = +
;::>:.1
(4)
+ I<j I
jd k.J
f3jk XX j Xk (8)
Xq = 1-
:
1
xI E(y) =L
;-1
f3x + I L
i<j j=2
f3ijXX j
I~I
in Eq. (4), removing the dependency among the factors and (9)
producing unique estima tes of the parameters 130, 131' . o ,
+ LL <j jd
tt.3
PjkXXjX, (lO)
feasible subregion that is still a sirnplex, although this
srnaller sirnplex will be inverted. See Refs . 1 and 2 for
additional informaron and sorne examples.
+ ... +PI2 .. .qXIX2 Xq, Generally, when chere are boch lower and upper bounds
as in Eq. (11), che feasible mixture region is no longer a
which is a qth-order mixture polynomial. Occasionally, simplex bU( will be sorne polytope (or hyperpolytope) sub
models with quartic terms or inverse terms may be re region inside the original unconstrained simplex . There
quired . For more inforrnation. see Ref. l. fore, the standard simplex.-type designs cannot be used . In
Finally, note that the simplex-type designs are essen these cases, sorne type of computer-generated design must
tially boundary-point designs; that is, most of the design be employed.
points are on the edges and faces of the simplex, and on1y There are severa! approaches to constructing designs for
a few (if any) points are in che interior of the region. If the . constrained mixture experiments. One of the earliest was
experimenter wishes to make predictions in the interior, . the extreme vertiees approach of MeLean and Anderson
then sorne additional runs should be alIocated to that part (3). They suggested using the vertiees of the constrained
of the design space. We recommend augmenting simplex region as the basis of the design, along with a subser of che
type designs with the overall centroid (if it is not included) subregion centroids and the overall centroid . With the
and axial check blends located halfway between the cen development and introduction of effective eomputer soft
troid of the region and each vertex. Cornell (1) and Myers ware to support experiments with mixtures, other eom
and Montgomery (2) give examples of the use of simplex puter-aided design selection methods have become popular.
type designs and discuss the use ofaxial blends further. An exeellent anicle on computer-aided design of experi
- .
ments is Ref. 4.
Constrained Mixture Experiments The D-optirnal eriterion can be used to select points for
a mixture design in a constrained region. This criterion
In many mixture experiments, there are constraints on essentially selects points from a list of candidate points so
:he component proportions. These are ofien upper- amI/or that the variances of the regression coeffieients in the mix
ower-bound constraints of the form L ~ Xi ~ Vi' i = 1, - ture model are minnized. The effectiveness of this method
~ .... q, where L is the lower bound for the ith compo depends on the list of candidate points, the adequacy of the
lent and V is the upper bound for me ith component. proposed model, and the number of design points selected.
~ssentially, L represents a minimum proportion of the ith Cornell (1) and Myers and Montgomery (2) discuss these
omponent that must be present in the mixture, and Vi issues extensively.
~presents a maximum proportion of the ith component that Distance-based designs are also useful for mixture ex
JUst be present in the mixture. The general form of the periments in constrained regions . This criterion attempts to
Jnstrained mixture problem is spread the design points out unifonnly over the feasible
....
design space. The algorithm for selecting points starts with would be adequate to represent both responses. This model
the point tha is as c10se as possible to a vertex of the has 10 parameters (the l3's) that must be estimated. Snee
unconstrained region, then adds to this the point for which (5), Montgomery and Voth (6), Cornell (1), and Myers and
the Euclidean distance to the first point is a maximum. Al! Montgomery (2) all recommended that sorne additional
subsequent points are added simiJarly; that is, the point for runs be included so that the adequacy of the model fit can
which the minimum Euclidean distance to the other points be checked and an estimate of experimental error can be
in the design is maximum is chosen. The distance criterion obtained. We decided that three runs be added as replicates
requires a grid of candidate points. We recommend the and that two runs be added to test lack of fit. Thus, the
same candidate points that would be used for the D-opti design will have 15 total runs. Twelve of these will be
mal criterion. distinct design points and three runs will be replicated.
The experimental design was generated using the De
Designing the Margarita Experiment sign-Expert (7) software package. This particular package
allows the user 10 specify the number of extra design points
aboye the minimum required to fit the model, and how
The margarita formulation problem is a constrained
those extra runs are .10 be alIocated (either replicate runs
mixture experiment, as a beverage that contains 100 per
or rum to check lack of fit). The design is shown in Table
cent of any one of the four ingredients would obviously be
1. The left-hand panel shows the actual proportions for the
unsatisfacrory, and minimum proportions of each ingredi
mixture components used for each run in the designo The
ent are also desirabIe. Based on our practical experience
center panel shows the mixrure proportions expressed in
with this beverage, we decided on the following con
terms of pseudocomponents. If the actual mixture compo
straints:
nent is Xi' then the eorresponding pseudocomponent s
Margarita mix: 0.49 ~ XI ~ 0.55 defined as
Tequila: 0.25 ~ x2 ~ 0 .31
L,
Triple Sec : 0.08 ~ x3 $ 0.16 X ,----
_X, -
1 - ILi (13)
Lime juice: 0.04 ~ x 4 $ 0.10
and Xl + X2 + X 3 + X 4 = l. We are not the first authors It is customary 10 fit mixture models in terms of the
to consider such a problem; Sahrman et al. (4) applied pseudocomponents when the design has been run in a con
mixture experiments to optimizing the mixture for a strained region, as this reduces the natural ill-conditioning
Harvey Wallbanger. that is present when the method of least squares is used for
The response variable in this experiment should be a estimating the parameters. For more details about ill-con
measure of the enjoyment of the beverage. We decided to ditioning, see Refs. 1, 6, and 8.
use two responses. First, partieipants would be asked to The right-hand panel of Table 1 presents the ranking
evaluate ea eh reeipe on a scale of 1 to lO, with 10 being and goodness response data obtained when the experirnent
the most enjoyable . This is called the "goodness" response, was acrually runo The taste-test panel consisted of 15 mem
and it is a subjective evaluation of the participants' enjoy bers. Each panel member tasted the beverages in random
ment of the tlavor and intensity of the beverage. Second, order and completed a score sheet. The results shown in
each partieipant would be asked to compare the relative Table 1 are the averages of the goodness and ranking data
enjoyment of each recipe by preparing a forced ranking, across all panel members. We remind the readers that
with he best recipe ranked "1," the next best ranked "2," conducting a designed experiment is a complicated activ
and so on . A taste-test panel of 15 participants would be ity, and one should never underestimate the logistical and
formed and the average value of the forced ranks and rat other practical aspects of this effort. Preexperimental plan
ings would be used as the response variables. ning is necessary 10 ensure a reasonable chance of success.
The D-optimal criterion was used . for selecting the ex For further reading on preexperimental planning, see Refs .
perimental designo We assumed that the quadratic mixrure 9 and 10.
model
Response Surface Modeling
4 4
E(y) = L
i::::l
(3i X ' + I L
i<j j=2
l3jxx (12)
Mixture response surface models were built for both
(ponses. using !he design infomo"ion in Table l. For!he
- -- -- - ---- - - - - - - - --
... '.
XI Xl Xl x. XI Xl Xl X. YI Yl
1 14 49 25 16 10 0.0000 0 .0000 0.5714 0.4286 9 .33 5. 14
4
2 S SI )1 8 10 0.1429 0.4286 0.0000 0.4286 9.1 7 4.43
) n. 55 '31 28 \0 13 4 0.4286 0 .2143 0.3571 O.OOO ~ 9.57 3.96
'>( 4 10 55 25 10 10 0.4286 0 .0000 0. 1429 0.4286 \3.67 6 .29
~ 5
6 13
" 55 rJ~ 29 8 'O~ 108 0.4286 0 .2857 0.0000 0.2857 7 .33 5.07
')\51 JI 8 0. 1429 0.4286 0.0000 0.4286 9.00 4.12
~ 7 ~ 55 25 10 10 0.4286 0 .0000 0. 1429 0.4286 5.33 6.21
8 ",,~52 ~I 28 4 0.2143 0.2143 0.5714 0 .0000 12.33 4.21
10 '.i
"
9 H ' c.,<,52
49
1.~ 25
1.) 28 tI.. 13
16
'. e 16 JU 7
la
0.2143
0 .0000
0.0000
0 .2143
0.5714
0.3571
0.2143
0.4286
8.67
6 .50
5.54
4.92
11 I 55 25 16 4 0.4286 0.0000 0.5714 0.0000 8.00 4.60
..r' 12
9 55 31 29 8 b 8 0 .4286 0.2857 0.0000 0.2857 7 . 17 5.79
13 3 55 31 10 0.4286 0.4286 0.1429 0.0000 7.50 4.60
14 Cf 49 31 0.--, 13 q .1 "7 0.0000 0.4286 0.3571 0.2143 7.67 4 .99
15 'J.. 49 31 16 0.0000 0.4286 0.5714 0.0000 9 17 2..79
"
goodness response, we fit linear and quadratic mixture
models using least squares and selected the final model
based on certain regresson model evaluaton surnmary sta
tistics including a lack-of-fit test, the square root of the
are useful in explaining the ~sponse are added to the
model, R~ will increase, whereas if nonsignificant tenns
are added, R~ wiJl not increase and can actually decrease.
A model with a large value of the adjusted R2 statistic is
If
residual mean square, the adjusted R 2, and the PRESS usually preferred. Finally, the PRESS statistic is defined as
(prediction error sum of squares) statisc. We now briefly
n
surnmarize how these statistics are used in model-building.
For more complete accounts and descriptions of building. PRESS = I (Yi - Y(il f (15)
;=1
regression models, see Refs. 11 and 12.
The lack-of-fit test is based on partitioning the residual where Y(I) is the predicted value for lhe ith observed value
sum of squares for a particular model (linear, say) into a of the response using a model mat has been fit with the ith
component due to pure error that is computed using the observation deleted. Thus, PRESS is a measure of how
response values at the replicate runs in the design and a well a particular model is likely to perform as a prediction
lack-of-fit component that represents the contribution to the equation for new data . When comparing two or more
residual sum of squares coming from higher-order terms models, generaJly the model with the smallest value of the \
(such as quadratic) . The analysis-of-variance F-test can be PRESS statistic is preferable. r
:>erformed to determine if the higher-order tenns are sig
1ificant. If they are, this is evidence that the order of this
nodel should be increased. The square root of the residual
nean square measures the standard deviation of the vari
In addition to using these surnmary statistics, we also
used the residual plots to check overall model adequacy and
protect against violation of assumptions . Specifical1y, we
examined nonnal probability plots of the residuals and plots
I
bility not explained by the model. A good model will of residuals versus predicted response . This allowed a
ypically ha ve a small residual mean square. The adjusted check on the nonnality and constant variance assumptions
.2 is defined as and gave a visual test for response oUlliers.
We now summarize the model-building procedure for
the goodness response YI' The initial model selected for YI
(14) was the quadratic mixture model in Eq . (12) . Table 2
shows the summary statistics for this response for the can
here SSE is the residual sum of squares and SSTocal is the didate mixture models given in Eqs. (6)-(9). Note that the
tal sum of squares . Generally, if higher-order tenns that desigri does not contain enough runs to fit the full cubic or
...
Table 2. Mixture Model-Building Summary Statistics from Design-Expert for the Goodness
Response
SUM OF MEAN F
MODEL SQUARES DF SQUARE VALUE PROB > F
special cubic models-this results in the warning messages Table 3 shows the analysis-of-variance summary for the
in the table o Therefore, we will concentrate on choosing linear mixture model ft to the goodness response. This
between the linear and quadratic models. The upper panel output was produced by Design-Expert. Note that the usual
of Table 2 summarizes statistical tests on the polynomial {-[est employed for tesling he significance of individual
terms in the model. Note that only the linear terms are terms in a regressor model cannot be directJy applied to the
st3tistically significanL The middle panel presenrs lack-of linear blending terms in a mixture model, because the {-test
tit tests for each model. The F-value for lack of tit is rela relates to the hypothesis Ho: Pi = O versus H,: 13, ~ O,
tively small for a1l models . These tests indica te that a lin which is not relevant for a mixture modelo In linear mix
ear mixture model could be used ro describe the goodness ture models, we need to test the equality of the linear
response . blending terms (i.e., the null hypothesis is Ho: 131 = ~2 =
...
TOTAL 100.000
3 = 134 = {3), and ths is done wilh the analysis-of-vari Figure 4 shows a normal probability plot of Ihe residuals
lee F-test in the upper pare of Table 3. It is appropriate from this model, and Figure 5 is a pIot of residuals ver
, conduct (-tests on adjusted linear blending terros, which sus predicted goodness. Both residual plots are satisfactory
e defined as and indicate no model inadequacy. Figure 6 is a plot of the
q
contours of constant goodness with Triple Sec held constant
b =p - (q - l t l 2: Pj , at 11.33 % by volume. Note that there are several appar
j-I (16) ently good formulations that would result in high values of
j"
the goodness response.
is done in the lower part of Table 3. Note that the in We also tit a response surface mode1 for the ranking
:dient Triple Sec is a marginal contributor to the good response, allhough a slightly different procedure was used.
:s response . We began with he full quadratic model, but the residual
TUTORIAL ON EXPERIMENTS WITH MIXTURES 249
2.282 +
99
+ 1.659 .
95
90
~ 80 + ~ 1.035
+
:o1 + ::J
+
+
.D 70 + 'O
'c;;
o
~ 50 ~ 0.411
+ 'E + +
'E*'
c;; 30 +
+
a>
'C
~ -0.212
+
::; 20 +
z 10 +
-0.83i ++
5
+
+ +
-1.459 +
-1.459 -0.836 -0.212 0.411 1.035 1.659 2.282 3.291 3.nO 4.248 4.726 5.204 5.683 6.161
Studentized Residual Predicted as goodness
Figure 5. Plor of residuals versus predicted goodness,
Figure 4. Normal probablity plot of residuals; goodness
response .
plots for this model indicated that a transformation on the using the linear and quadratic terms as candidate variables .
response was necessary to correct problems with non-nor The details of the model fitting and the final model, which
mality and inequality of variance. This is a fairly cornmon is a reduced quadratic, are shown in Table 4.
problem with rank responses. We selected the square roer Figure 7 shows me response surface contours for rank
of the ranking response for constructing the model. Then ing (untransformed) with Triple Sec held constant al
we applied stepwise regression to the transformed response, 11.33% by volume. In this case, the lowest ranking cor
DESIGNEXPEAT Plet
Model:
Unear
Actual components:
=
Xl Marg mix
X2 = Teqilla
X3 = Ume)uice
Actual constants:
Triple_Sec = 11.33
X2 (35.7)
Figure 6. Response surface contour plot of goodness with Triple Sec al 11.33 % by volume,
..
I
1
250
I
BOWLES AND MONTGOMERY
Table. 4. Model-Building for [he Ranking Response Using a Square Root Transformation
,
STEP-WISE REGRESSION
Alpha To Enter = 0.1000
Alpha To Exit 0.1000 (
COEFFICIENT t FOR HO
COMPONENT ESTIMATE COEFFICIENT=O PROB > Itl R-SQUARED MSE
:-oreed
A 2.349
B 3.437
C 3 . 256
D 2.157
0.4710 9.31E02
Entered Be
A 2.016
B 4.415
e 3.877
D 1.713
SUM OF MEAN F
responds to the best beverage. Once again, we see that would want the ranking response ro be smaller than S and
there are several fonnula[ions that will produce a beverage the goodness response ro exceed 5.
with a satisfactory average ranking. There are several ways ro simultaneously oprimize sev
eral responses. When the number of mixture components
Xi is smalJ, then overlaying [he response surfaces and
The Optimum Margarita choosing the optimum conditions for each Xi by inspection
of the composite conrour plots is a simple and usually
The objective of most mixture problems is to find the highly effective method. Because our problem has four
levels of the components that optimize the product fonnu components, graphical optimization would be accomplished
lation. Based on knowledge of our specific problem, we by plotting response surface overlays for three components
TUTORIAL ON EXPERIMENTS WITH MIXTURES 251
Tah/e 4. Continued
Sqrt ranking
+
+
2.016
4.415 A
. e
B
+ 3.877
+ 1.713 D
4.306 BC
A very useful approach to the optimization of m differ and each response is transformed into an individual desir
en! responses is to maximize a desirability function D, ability di (O :5 di :5 1) by defining a set of levels a, ~ ti ~
Reduced Quadralic
Actualcomponents:
X2 = ToquIlla
X3 '" lime_Juice
Actual constants:
Triplu_Sec = 11.33
X2 (35.7) X3 (14.7)
di = ( ~J'.
b
li -
( 18)
Conclusion
Actual components:
Design Expert Graphical Optimization
X1 = Tequllla X1 (33.0)
X3 = Limejice .
Actual constants:
Marg_mx ;: 55.00