You are on page 1of 13

Mixture-Process Variable Experiments

with Noise Variables


HEIDI B. GOLDFARB
The Dial Corporation, Scottsdale, AZ 85254

CONNIE M. BORROR and DOUGLAS C. MONGOMERY


Arizona State University, Tempe, AZ 85287

In a mixture experiment, the design factors are mixture components whose proportions are varied, and
the response variables are assumed to depend only on these component proportions. In addition to the
mixture components, the experimenter may be interested in other variables that can be varied independently
of one another and of the mixture components. We consider the case where one or more of these variables
is a noise variable, or a variable that cannot be controlled in practice. We develop models for these robust
mixture formulation problems. We then derive mean and variance functions and illustrate their use in
formulation optimization. Cases of uncorrelated and correlated noise variables are addressed.

Introduction There may be instances where, in addition to the


dependent mixture components, we have other fac-
F ORMULATION mixture experiments are commonly
encountered in the chemical, food, pharmaceu-
tical, and consumer products industries, as well as
tors and/or process variables that can be controlled
independently of one another and of the mixture
components. For example, in a chemical production
in other industrial settings. In these experiments
system the composition of the chemical formulation
the factors are ingredients or components of a mix-
involves mixture variables, while the settings of the
ture, and the responses are typically assumed to be
manufacturing equipment are process variables.
functions of the component proportions. Thus, if
0 xi 1 is the proportion of the ith component, Cornell (2002) gave a comprehensive presentation
then of all aspects of mixture experiments, including those
 q
xi = 1 where both mixture components and process vari-
i=1
ables are included in the same experiment or where
the mixture components and the total amount of the
is the usual constraint on these proportions. Because
mixture are varied. Cornell (1995) provided guidance
of this constraint, a design characteristic of mixture
on tting models to data from experiments with both
experiments is that the design factors (the compo-
mixture and process variables. Piepel and Cornell
nents) cannot be adjusted independently of one an-
(1985, 1987) proposed design and modeling strategies
other.
specically for those experiments where the response
depends on the total amount of the mixture. Cornell
and Gorman (1984) addressed the topic of fractional
Dr. Goldfarb is a Principal Statistician in the Research
and Development Department. She is a Member of ASQ. Her
designs for experiments with both mixture and pro-
e-mail address is goldfarb@dialcorp.com. cess variables. An excellent annotated bibliography
of papers on mixture designs with process variables,
Dr. Borror is a Senior Lecturer in the Department of In- as well as various other types of mixture experiments,
dustrial Engineering. She is a Senior Member of ASQ. Her
was created by Piepel and Cornell (2001).
e-mail address is conni@asu.edu.

Dr. Montgomery is a Professor in the Department of Indus- Another important area in industrial experimen-
trial Engineering. He is a Fellow of ASQ. His e-mail address tation is that of robust design or process robustness
is doug.montgomery@asu.edu. studies. Genichi Taguchi rst introduced the con-

Vol. 35, No. 4, October 2003 393 www.asq.org


394 HEIDI B. GOLDFARB, CONNIE M. BORROR, AND DOUGLAS C. MONTGOMERY

cept of robust design in the United States in the dratic loss function that took into account both the
1980s. His design and analysis ideas can be found in mean and the variance of the response. Then, so-
Taguchi (1986, 1987), and Taguchi and Wu (1985). lutions to the resulting constrained nonlinear opti-
In robust design experiments there are two types of mization problem were found using standard tech-
variablesthose that we can control and those that niques. We use the Derringer and Suich (1980) desir-
we cannot. These uncontrollable variables are of- ability function approach to solve the simultaneous
ten referred to as noise variables. We assume that optimization problem rather than the loss function
they can be controlled for the purposes of an exper- approach. In this approach, a desirability function
iment (i.e., xed eects), but that when the process is built for each response based on whether we wish
is in normal operation they will vary in a random to minimize, maximize, or achieve a target level for
fashion (i.e., random eects). The overall goal in ro- the response. Then, the geometric mean of these in-
bust design is to nd settings of the controllable vari- dividual desirability functions is taken to form a sin-
ables that are least sensitive to changes in the noise gle composite desirability function to be maximized.
variables, while still yielding acceptable mean levels This procedure is very sensitive to the weights used
of the responses. While Taguchis introduction and within the individual desirability functions as well as
description of the robust design problem have been the weight used for the geometric mean. More de-
of great value, his strategies for design and analysis tails and examples of this procedure can be found
have been critiqued by many. Nair (1992), Myers in Myers and Montgomery (2002). For the examples
and Montgomery (2002), and Montgomery (1999) in this paper, the Design-Expert software package
provided summaries of some of the more prevalent (Stat-Ease (2002)) was used to simultaneously opti-
views on Taguchis ideas, including philosophy, de- mize the mean response and minimize the standard
sign, and analysis. deviation response using the desirability approach.
For our examples, linear desirability functions were
Taguchis basic approach was to build separate ex-
used, and equal weights were given to the mean and
perimental designs for the sets of controllable and
standard deviation responses. Desirability function
noise variables, and then form the Cartesian product
optimization has been included in several other soft-
of the two designs in order to produce a crossed-
ware packages, including Statistica, Minitab, and
array design. His approach to analysis was to ex-
JMP.
amine signal-to-noise ratios. An alternative to the
crossed-array design is to include both sets of vari- The model used in Steiner and Hamada (1997)
ables in a single design called a combined-array de- contains interaction terms involving the noise vari-
sign. The suggested analysis for this type of design ables. To be more consistent with the typical as-
is based on response surface methods. More details sumptions of the robust design eld, we do not in-
on this design and analysis approach can be found in clude these terms in our model. Steiner and Hamada
Myers, Khuri, and Vining (1992), Lucas (1994), and also assumed independence among the noise vari-
Myers and Montgomery (2002). ables, while we consider the situation where the noise
In this paper we address experiments involving variables are correlated, which is a common situation
both of the situations mentioned previously. Specif- in many industrial settings. For example, we con-
ically, we consider the analysis of experiments with sider a soap-making process where two noise vari-
mixture components, controllable process variables, ables, temperature and humidity, are known to af-
and uncontrollable (or dicult to control) process fect many attributes of the resulting product, such
or environmental variables (which we consider to be as machinability, yield, and hardness. In a soap pro-
noise variables). The overall objective of such exper- duction plant, these two noise variables are often cor-
iments is to nd levels of the mixture components related, sometimes to a great degree.
and controllable variables that are most robust to
The paper is organized as follows. First, we intro-
variation in the noise variables while simultaneously
duce the notation to be used and develop the general
providing an acceptable mean response value.
model. We then apply the model to a classic and a
Steiner and Hamada (1997) addressed a very new example. A model is also developed for the situ-
similar problem. After tting a combined mixture- ation in which there is a correlation among the noise
process-noise variable model, they constructed and variables. An example is used to demonstrate the
solved an optimization problem to minimize a qua- eects of the correlation.

Journal of Quality Technology Vol. 35, No. 4, October 2003


MIXTURE-PROCESS VARIABLE EXPERIMENTS WITH NOISE VARIABLES 395

Notation and Model Development be centered using historical data or, in some cases,
theoretical information about the process. Similarly,
We suppose there are q mixture components, c 2
the variances of the noise variables, N oise , are typi-
controllable process variables, and n noise variables.
cally estimated from historical data about the range
We also suppose that in model form, there are m
of variation of the noise variables or estimated from
mixture component terms, where m will depend on
external knowledge about the process. Then, by set-
the number of components and the degree of the
ting the low and high levels of the noise variables
model. We then have the mixture components xi ,
in the design at 1 N oise , in coded form we obtain
i = 1, 2, . . . , q, controllable process variables wp ,
z2 = 1. Myers and Montgomery (2002, Section 11.4)

p = 1, 2, . . . , c, and noise variables zt , t = 1, 2, . . . , n.
included more details about estimating variances of
The general models proposed contain quadratic or
the noise variables.
special cubic mixture terms (although a generaliza-
tion to other models follows naturally), interactions Equation (1) can be expressed in matrix form as
between the mixture components and the control-
lable process variables, interactions between the mix-
Y = f (x, w, z) = x + x Aw+
ture components and the noise variables, and inter-
actions among all three types of variables. This type x z + x Vz + , (2)
of model allows the determinination of the settings of
the mixture variables and controllable process vari- where
ables that will be most robust to variation in the
x1 1
noise variables. x2 2
..
..
The general form of the model, assuming a . .

quadratic mixture model, is xq q

x1 x2 12
Y = f (x, w, z)
  xm1 = x1 x3 , m1 = 13 ,
= i xi + ij xi xj .. ..

. .
i

i<j

xq1 xq q1,q
+ ip xi wp + ijp xi xj wp
x1 x2 x3 123
i p i<j p .. ..
  .
.
+ it xi zt + ijt xi xj zt (1) xq2 xq1 xq q2 q1 q
i t i<j t


+ ipt xi wp zt w1 0 0
i p t .. ..
 w2 . .
+ ijpt xi xj wp zt + , . .. ..
.
. . .
i<j p t ..
wc 0 .
where the s are the mixture model coecients, the

..
s are the coecients of the interactions of the mix- 0 w1
.

ture and controllable process variables, the s are the ..
..
. w2 .
coecients of the interactions of the mixture and the
w1 .. .. ..
noise variables, and the s are the coecients of the . . .
w2 .
interactions of the mixture, controllable process, and wc1 =
... , Vcnn =
.. wc
,

..
.
noise variables. In this model, the process and noise . ..
.
variables appear with only linear terms. Higher- wc . 0 .
. .. .. ..
order terms are not included, although they could .. . . .

be incorporated if appropriate. This model high- .. ..
. . 0
.
lights the importance of interactions with the mix- . ..
. . w1
ture components to develop the methodology, and it . ..
.
is a reasonable model in many industrial settings. . . w2
. .. ..
It is initially assumed that the noise variables are .. . .
centered so that they have means of zero. They can 0 0 wc

Vol. 35, No. 4, October 2003 www.asq.org


396 HEIDI B. GOLDFARB, CONNIE M. BORROR, AND DOUGLAS C. MONTGOMERY


11 12 1c
21 21 2c

.. ..
. .

q1 qc

121 12c

Amc =

131 13c ,

.. ..
. .

q1,q,1 q1,q,c

1231 123c
.. ..
. .
q2,q1,1 q2,q1,2 q2,q1,c FIGURE 1. The Matrix mcn .

z1
z2
zn1 =
... ,
where z = diag(z21 , z22 , z2n ) is an nn diagonal
matrix with the variances of the noise variables on
zn the diagonal and 0s elsewhere.

11 12 1n
Since z only appears in a rst-order term in our
21 22 2n
model, these results are exact rather than approxi-
.. ..
. . mate. Using the previously dened model for Y in

q1 qn matrix form, we nd that

121 12n

mn = 131 13n , f (x, w, z = 0) = x + x Aw
.. ..
f  (x, w, z = 0) =  x + V  x,
. .

q1,q,1 q1,q,n
so
1231 123n

.
.. .
.. E(Y ) = x + x w, (3)
q2,q1,q,1 q2,q1,q,n and
and mcn is given in Figure 1. 
Var(Y ) = [ x + V  x] x [ x + V  x] , (4)
The expected value and variance of Y are found
using the delta method, which employs a rst-order where it is assumed that the zs are normally distri-
Taylor series approximation of the model around the buted and centered so that they have a mean of 0.
mean of z. The delta method will also work for mod-
els with higher-order terms than the present model. Note that the noise variables are treated as ran-
See Rice (1995) for more details on the delta method. dom eects for the purpose of deriving the mean and
We assume that the noise variables are uncorrelated, variance expressions in Equations (3) and (4). How-
that the model errors () are also uncorrelated, and ever, for design construction and experimental exe-
that the noise variables and the model errors are un- cution, they are treated as xed eects. This is a
correlated with each other. Expanding f (x, w, z) us- standard approach to robust modeling. The noise
ing a rst-order Taylor series around z = 0, which factors are uncontrollable during normal process op-
represents the vector of centered means of the noise eration, but it is assumed that they can be controlled
variables, yields for experimental purposes.

Y = f (x, w, z = 0) + (z 0)f  (x, w, z = 0) + , Example 1: Fish Patties


where f  (x, w, z = 0) represents the vector of partial Consider the sh patty texture experiment from
derivatives with respect to z evaluated at z = 0. Cornell (2002). In this classic example there are
Expectation and variance operators are applied to three mixture components (mullet = x1 , sheepshead
obtain = x2 , and croaker = x3 ) and three process variables.
E(Y ) f (x, w, z = 0) The response variable is the texture of the sh pat-

Var(Y ) [f  (x, w, z = 0)] z [f  (x, w, z = 0)] + 2 , ties. The experimental design is a 23 factorial in the

Journal of Quality Technology Vol. 35, No. 4, October 2003


MIXTURE-PROCESS VARIABLE EXPERIMENTS WITH NOISE VARIABLES 397

process variables and a simplex-centroid in the mix- assigning a desirability value to the mean texture
ture components (7 runs) at each combination of the which increases linearly as the texture increases, and
process variables, giving a total of 56 trials. We as- a desirability value to the standard deviation of the
sume that one of the process variables is controllable texture which increases linearly as the standard de-
(deep fat frying time = w1 ) and that the other two viation decreases. The ranges used here for the de-
are noise variables (oven temperature = z1 and oven sirability function are the full ranges of values from
baking time = z2 ). Using the notation introduced in the design itself, so the highest texture value is as-
the previous section, q = 3, c = 1, and n = 2. A signed a desirability of 1 and the lowest texture is
quadratic model in the mixture components is t, so assigned a desirability of 0. The geometric mean of
m = 6. these two desirability values is then maximized to
nd the optimum.
The data given in Cornell (2002, p. 360, Table
7.1) are the averages of two patties made under each We assume that z22 = 1 for demonstrative
z21 =
condition, each measured three times. Using the sig- purposes, since no such variances are given in the
nicant model coecients shown in Cornell (2002, p. original example. For the later examples, we give
363, Table 7.2) we obtain the following model: strategies for estimating the variances. We also set
y(x, w, z) = 2.86x1 + 1.11x2 + 2.03x3 0.99x1 x2 2 = 0.0184, the value reported in Cornell (2002).

For maximum texture and minimum standard devi-
0.85x1 x3 + 0.49x1 z1 + 0.17x2 z1
ation, the following levels are optimal:
+ 0.24x3 z1 0.80x1 x2 z1 + 0.70x1 z2
+ 0.26x2 z2 + 0.39x3 z2 + 0.09x3 w1 z1 . Mullet (A or x1 ) = 0.0
Using Equations (3) and (4) developed in the previ-
ous section, we can estimate the mean and variance Sheepshead (B or x2 ) = 0.0
response functions for the mixture components at dif-
ferent levels of the controllable process variable, w1
Croaker (C or x3 ) = 1.0
to obtain

) = 2.86x1 + 1.11x2 + 2.03x3 Deep Fry Time (w1 ) = 25 seconds,


E(Y
0.99x1 x2 0.85x1 x3 , which yields a predicted texture of 2.03 with a stan-
and dard deviation of 0.440.

Contour plots for the average and standard devi-



)=
Var(Y z21 (0.49x1 + 0.17x2 + 0.24x3 ation of texture at w1 (deep fry time) = 25 seconds
0.80x1 x2 + 0.09x3 w1 )
2 are shown in Figures 2 and 3. The two plots could
be overlaid to simultaneously investigate obtaining
z22 (0.70x1 + 0.26x2 + 0.39x3 )2 +
+ 2 . a desired texture with a suitably small standard de-
Note that the expected value and variance equations viation. Since a high texture value is desired, there
are similar to those of Steiner and Hamada (1997). is a tradeo between the mean value and the stan-
There are some dierences in the variance equation dard deviation. As the mean texture values increase
because they included interactions among the noise (moving toward a higher percentage of mullet), the
variables in their model while we do not. Also note standard deviation also increases, indicating a de-
that the coecients for the process and noise vari- crease in the product robustness.
ables are in coded units.
To identify an area of high desirability, we sup-
Numerical optimization methods can be used to pose that the goal is a mean texture greater than
nd the optimum levels for the mixture variables to 2 and a standard deviation less than 0.7. A con-
yield the desired texture with the smallest standard tour overlay graph based on these criteria is shown
deviation. Design Expert V6.0 incorporates pattern in Figure 4. The area meeting both criteria is the
search techniques to do this. For a brief descrip- unshaded portion. In the shaded portion of the g-
tion of pattern search, see Carlyle, Montgomery, and ure, the mean is too small, the standard deviation
Runger (2000). The numerical solution is found by is too large, or both. This graph shows that there

Vol. 35, No. 4, October 2003 www.asq.org


398 HEIDI B. GOLDFARB, CONNIE M. BORROR, AND DOUGLAS C. MONTGOMERY

FIGURE 2. Design-Expert Contour Plot of Mean Tex- FIGURE 4. Design-Expert Response Overlay Plot for Ex-
ture Values for Example 1 at w1 (deep frying time) = 25 ample 1 at w1 (deep frying time) = 25 Seconds.
Seconds.

are many formulations of the sh patties that would


satisfy the requirements in addition to the numerical
solution found above. For example, a blend with
Mullet (A or x1 ) = 0.64, Sheepshead (B or x2 ) =
0.36, and Croaker (C or x3 ) = 0.0 yields a texture of
2.01 with a standard deviation of 0.598. This very
dierent blend has a desirability value close to that
of the optimum. This is valuable knowledge, since
there may be external factors that would make this
blend a more attractive option.

Example 2: Soap Manufacturing


This example involves a soap processing plant ap-
plication. We are interested in studying throughput,
or the amount of soap in pounds per hour that the
process yields. In making the soap blend, there are
three mixture components (soap = x1 , co-surfactant
= x2 , and ller = x3 ). We have the following restric-
tions on these variables:

0.20 x1 0.80
0.15 x2 0.50
0.05 x3 0.30
FIGURE 3. Design-Expert Contour Plot of Standard De- x1 + x2 + x3 = 1.
viation of Texture for Example 1 at w1 (deep frying time)
= 25 Seconds. In manufacturing mixture blends, there are two

Journal of Quality Technology Vol. 35, No. 4, October 2003


MIXTURE-PROCESS VARIABLE EXPERIMENTS WITH NOISE VARIABLES 399

processing variables of interest, plodder temperature the four factorial combinations of the two process
and mixing time. Mixing time (w1 ) is considered variables, for a total of 20 runs. The design and
a controllable variable that can be run at two lev- response data are shown in Table 1.
els, 0.5 hours and 1 hour. The plodder temperature
The chosen design supports a linear model in the
(z1 ) is not easy to control and is considered a noise
mixture terms crossed with a two-factor interaction
variable. Two levels of plodder temperature will be
model for the process variables. All higher-order
considered in the experiment, 15 C and 25 C. These
models would result in aliasing. This full crossed
levels cover about 70 percent of the variation in plod-
model was evaluated and found to have an excellent
der temperature around the average value of 20 C
t. After eliminating one non-signicant term at a
routinely observed in practice, and so represent low
0.05 level, the lack-of-t p-value is 0.90. The R2 and
and high levels that are about one standard deviation
adjusted R2 values are both greater than 0.99. The
of the noise variable from the mean. Thus, z21 = 1.
nal model is
If a reliable estimate of the standard deviation is not
easily obtainable, one approach that could be used y(x, w, z) = 438.36x1 254.01x2 + 765.78x3
would be to evaluate the problem over a reasonable
+ 34.55x1 w1 + 263.49x2 w1 203.75x3 w1
range of z21 values.
0.04x1 z1 + 5.27x2 z1 18.56x3 z1
Using the notation previously introduced, we have
3.90x2 w1 z1 + 4.69x3 w1 z1 .
q = 3, c = 1, and n = 1. Several previous ex-
periments performed on the mixture variables led to Using Equations (3) and (4) we can estimate the
models with only linear mixture components. There- mean and variance response surfaces for the mixture
fore, a linear model was used for the mixture com- components at dierent levels of the controllable pro-
ponents in this experiment, resulting in m = 3. cess variable, w1 to obtain
Design-Expert was used to generate a ve-run D- ) = 438.36x1 254.01x2 + 765.78x3
E(Y
optimal design for the linear mixture model. The
resulting ve-point design was repeated at each of + 34.55x1 w1 + 263.49x2 w1 203.74x3 w1

TABLE 1. Soap Processing Design and Data for Example 2 (Note: The Noise Variable has been Centered Around 0)

Standard Run Soap Co-surfactant Filler Plodder Mixing Time Throughput


Order Order (%) (%) (%) Temp ( C) (Hours) (lbs/hr)
1 3 0.2 0.5 0.3 5 0.5 245.2
2 2 0.8 0.15 0.05 5 0.5 381.5
3 15 0.8 0.15 0.05 5 0.5 381.0
4 19 0.55 0.15 0.3 5 0.5 453.3
5 12 0.45 0.5 0.05 5 0.5 172.9
6 18 0.2 0.5 0.3 5 1 285.4
7 6 0.8 0.15 0.05 5 1 409.1
8 20 0.8 0.15 0.05 5 1 411.7
9 8 0.55 0.15 0.3 5 1 450.0
10 13 0.45 0.5 0.05 5 1 245.8
11 10 0.2 0.5 0.3 5 0.5 213.8
12 1 0.8 0.15 0.05 5 0.5 378.4
13 16 0.8 0.15 0.05 5 0.5 377.3
14 9 0.55 0.15 0.3 5 0.5 408.8
15 5 0.45 0.5 0.05 5 0.5 180.6
16 7 0.2 0.5 0.3 5 1 250.6
17 14 0.8 0.15 0.05 5 1 404.2
18 17 0.8 0.15 0.05 5 1 406.6
19 4 0.55 0.15 0.3 5 1 410.1
20 11 0.45 0.5 0.05 5 1 245.4

Vol. 35, No. 4, October 2003 www.asq.org


400 HEIDI B. GOLDFARB, CONNIE M. BORROR, AND DOUGLAS C. MONTGOMERY


) = 2 (0.04x1 + 5.27x2 18.56x3
Var(Y z1
2
3.90x2 w1 + 4.69x3 w1 ) + 2 ,
where 2 = 1.07, the mean square error for the tted
model, is used as an estimate of 2 .
We would like to maximize the average through-
put while simultaneously minimizing the variability.
The desirability function optimization procedure of
Design-Expert 6.0 was used with a linear desirability
function over the ranges of the response values within
the design. The following optimum levels were ob-
tained:
Soap (A or x1 ) = 0.80

Co-surfactant (B or x2 ) = 0.15

Filler (C or x3 ) = 0.05

Mixing Time (w1 ) = 1 hour.


FIGURE 5. Design-Expert Contour Plot of Mean
The predicted throughput at this point is 408.5 with Throughput Values for Example 2 at w1 (mix time) = 1
a standard deviation of 0.52. Hour.
Contour plots of the average and standard devia-
tion of throughput are shown in Figures 5 and 6 for a and
mix time of 1 hour. The plots reveal clearly why the
optimum formulation has a high percentage of soap = [1 | 2 | | n ] ,
and a low percentage of ller.
where k =
Correlation Among the
11k 12k 1ck
Noise Variables 21k 22k 2ck

We now consider the case where there is correla- .. .. .. ..
. . . .
tion among some of the noise variables. It is conve-
q1k q2k qck
nient to dene the and matrices as partitioned
121k 122k 12ck
matrices, so that we have
131k 132k 13ck
.. .. ..
= [1 | 2 | | n ] ,
. . .

q1,q,1,1 q1,q,2,k q1,q,c,k
where
1231k 1232k 123ck
1k .. .. ..

2k . . .

.. q2,q1,q,1,k q2,q1,q,2,k q2,q1,q,c,k
.

qk for k = 1, 2, ...n.

12k

k =

13k ,
k = 1, 2, ...n, The response model for the correlated noise vari-
.. able problem is identical to the model for the in-
.
dependent noise variable problem, as is the ex-
q1,q,1
pected value of the response obtained from the delta
123k
.. method. However, the variance derived from the re-
. sponse model using the delta method is dierent. Be-
q2,q1,q,k cause we no longer have independence among the zs,

Journal of Quality Technology Vol. 35, No. 4, October 2003


MIXTURE-PROCESS VARIABLE EXPERIMENTS WITH NOISE VARIABLES 401

ature of the process. We know that temperature and


humidity are correlated and incorporate this fact into
the analysis. Let z1 be temperature and z2 be humid-
ity. From previous experience, we know that the av-
erage temperature is 20 C with a standard deviation
of 5 C, and the average humidity is 60% with a stan-
dard deviation of 20%. We set the low and high levels
for these variables at 1 N oise , and we center them
so that they have a mean of 0 and z21 =
z22 = 1. We
found that the correlation between temperature and
humidity, z1 z2 , can be as large as 0.9. Therefore,
z1 z2 = z1 z2
z1
z2 can be as large as 0.9(1)(1) =
0.9. The matrices of model parameters are now

x1  
z
x = x2 , w = [w1 ] , z = 1 ,
z2
x3

  1 11
w1 0
V= , = 2 , = 21 ,
0 w1
3 31

FIGURE 6. Design Expert Contour Plot of Standard De- 11 12 111 112
viation of Throughput for Example 2 at w1 (mix time) = 1 = 21 22 , and = 211 212 ,
Hour. 31 32 311 312
and the response model is
the variance term for Y now includes terms involving Y = f (x, w, z) = x + x w + x z + x Vz + .
the covariances for those zs that are correlated.
The expected value and variance of Y are given by
Specically, for all zi and zj that are correlated, Equations (6) and (5), respectively.
the variance equation has the additional term given
by A linear mixture model is once again assumed,
based on previous experience. The design here has
g() g()
2zi zj . a full 23 factorial structure for the noise and process
zi zj
variables crossed with a ve-point mixture design,
So, in general, for a total of 40 runs. The design and data are given

Var(Y ) = [ x + V  x] x [ x + V  x] in Table 2. The model above was t to the data
 and yielded an insignicant lack-of-t with a p-value
+2 zi zj [i x + w i x]
greater than 0.99. Additionally, the R2 value was
i<j
 also greater than 0.99. The nal model is given as
j x + w j x + 2 , (5) follows:
where z = diag(z21 , z22 , , z2n ) is an n n diag- y(x, w, z) = 1899.0x1 + 1626.4x2 + 1537.9x3
onal matrix with the variances of the noise variables
on the diagonal and 0s elsewhere. Note that for all zi + 39.6x1 w1 + 285.9x2 w1 + 24.7x3 w1
and zj that are not correlated, no additional term(s) + 10.3x1 z1 + 3.3x2 z1 + 18.8x3 z1
for those pairs would be needed. As before, We have + 1.10 1014 x1 z2 20.0x2 z2
E(Y ) = x + x Aw. (6) + 1.77 1014 x3 z2 + 15.7x1 w1 z1
+ 31.7x2 w1 z1 6.32 1015 x1 w1 z2
Example 3: Soap Processing 25.0x2 w1 z2 2.22 1014 x3 w1 z2 .
We next consider a second soap processing exam- While some of the coecients are quite small nu-
ple at a plant located in an area where we are con- merically, they do have a considerable impact on the
cerned about the humidity in addition to the temper- model t. The lack-of-t portion of the model mean

Vol. 35, No. 4, October 2003 www.asq.org


402 HEIDI B. GOLDFARB, CONNIE M. BORROR, AND DOUGLAS C. MONTGOMERY

TABLE 2. Soap Processing Design and Data for Example 3 (Note: The Noise Variable has been Centered Around 0)

Standard Run Soap Co-surfactant Filler Plodder Mixing Time Humidity Throughput
Order Order (%) (%) (%) Temp ( C) (Hours) (%RH) (lbs/hr)

1 33 0.2 0.5 0.3 5 0.5 20 1962.0


2 15 0.8 0.15 0.05 5 0.5 20 1887.7
3 4 0.8 0.15 0.05 5 0.5 20 1882.9
4 37 0.55 0.15 0.3 5 0.5 20 1784.4
5 27 0.45 0.5 0.05 5 0.5 20 2060.5
6 14 0.2 0.5 0.3 5 1 20 2124.2
7 35 0.8 0.15 0.05 5 1 20 1901.7
8 39 0.8 0.15 0.05 5 1 20 1927.8
9 7 0.55 0.15 0.3 5 1 20 1836.3
10 25 0.45 0.5 0.05 5 1 20 2201.6
11 24 0.2 0.5 0.3 5 0.5 20 2161.8
12 8 0.8 0.15 0.05 5 0.5 20 2074.6
13 1 0.8 0.15 0.05 5 0.5 20 2063.3
14 17 0.55 0.15 0.3 5 0.5 20 1975.4
15 26 0.45 0.5 0.05 5 0.5 20 2234.0
16 19 0.2 0.5 0.3 5 1 20 2402.8
17 40 0.8 0.15 0.05 5 1 20 2175.6
18 38 0.8 0.15 0.05 5 1 20 2199.8
19 11 0.55 0.15 0.3 5 1 20 2075.7
20 9 0.45 0.5 0.05 5 1 20 2510.0
21 13 0.2 0.5 0.3 5 0.5 20 1312.0
22 5 0.8 0.15 0.05 5 0.5 20 1692.7
23 23 0.8 0.15 0.05 5 0.5 20 1687.9
24 21 0.55 0.15 0.3 5 0.5 20 1589.4
25 2 0.45 0.5 0.05 5 0.5 20 1410.5
26 30 0.2 0.5 0.3 5 1 20 1224.2
27 3 0.8 0.15 0.05 5 1 20 1631.7
28 32 0.8 0.15 0.05 5 1 20 1657.8
29 31 0.55 0.15 0.3 5 1 20 1566.3
30 29 0.45 0.5 0.05 5 1 20 1301.6
31 28 0.2 0.5 0.3 5 0.5 20 1511.8
32 10 0.8 0.15 0.05 5 0.5 20 1879.6
33 36 0.8 0.15 0.05 5 0.5 20 1868.3
34 6 0.55 0.15 0.3 5 0.5 20 1780.4
35 34 0.45 0.5 0.05 5 0.5 20 1584.0
36 20 0.2 0.5 0.3 5 1 20 1502.8
37 12 0.8 0.15 0.05 5 1 20 1905.6
38 16 0.8 0.15 0.05 5 1 20 1929.8
39 18 0.55 0.15 0.3 5 1 20 1805.7
40 22 0.45 0.5 0.05 5 1 20 1610.0

square error is orders of magnitude larger and statis- and


tically signicant if these terms are omitted. With-
) = (10.3x1 + 3.3x2 + 18.8x3
Var(Y
out these terms, the widths of condence and pre- +15.7x1 w1 + 31.7x2 w1 )
2

diction intervals increase greatly. Estimates of the


+ (20.0x2 25.0x2 w1 )2
expected value and variance of Y are
+ 2(0.9) (10.3x1 + 3.3x2 + 18.8x3
) = 1899.0x1 + 1626.4x2 + 1537.9x3 + 39.6x1 w1
E(Y +15.7x1 w1 + 31.7x2 w1 )
+ 285.9x2 w1 + 24.7x3 w1 , (20.0x2 25.0x2 w1 ) +
2 ,

Journal of Quality Technology Vol. 35, No. 4, October 2003


MIXTURE-PROCESS VARIABLE EXPERIMENTS WITH NOISE VARIABLES 403

FIGURE 7. Design-Expert Contour Plot of Mean FIGURE 8. Design-Expert Contour Plot of Standard De-
Throughput Values for Example 3 at w1 (mix time) = 0.82 viation of Throughput for Example 3 with a Correlation of
Hours. 0.9 at w1 (mix time) = 0.82 Hours.

where 2 = 76.5, the mean square error for the t- will have the highest expected throughput, but also
ted model, is used as an estimate of 2 and we have a high standard deviation. Mixtures with a lower
z21 = 1 and
substituted z22 = 1. percentage of soap and a higher percentage of co-
surfactant have lower standard deviations but also
As in Example 2, we would like to maximize the
have lower expected throughputs. In practice, an
average throughput while minimizing the variabil-
engineer or manager may decide that increasing the
ity. The desirability function optimization procedure
output is worth the extra variation; however, this ex-
of Design-Expert 6.0 was used with linear desirabil-
ample is useful in demonstrating how to objectively
ity functions over the ranges of the response values
apply the methodology we have developed.
within the design. This provides the following opti-
mum levels: If we had performed the optimization without tak-
ing the correlation between temperature and humid-
Soap (A or x1 ) = 0.45
ity into account, then the standard deviation func-
tion would have changed and this would have led to
Co-surfactant (B or x2 ) = 0.50 dierent optimum levels for the mixture components
and the mixing time. The solution would have been
the following:
Filler (C or x3 ) = 0.05
Soap (A or x1 ) = 0.80
Mixing Time (w1 ) = 0.82 hours.
Co-surfactant (B or x2 ) = 0.15
The expected throughput is 1877.4 with a standard
deviation of 14.8. Filler (C or x3 ) = 0.05

Contour plots of the mean and standard deviation Mixing Time (w1 ) = 0.5 hours.
of throughput are shown in Figures 7 and 8 for a
mix time of 0.82 hours. From these plots we can This blend is quite dierent than the one obtained
see that a mixture with a high percentage of soap when the correlation was considered, and it yields

Vol. 35, No. 4, October 2003 www.asq.org


404 HEIDI B. GOLDFARB, CONNIE M. BORROR, AND DOUGLAS C. MONTGOMERY

FIGURE 9. Design-Expert Contour Plot of Mean FIGURE 10. Design-Expert Contour Plot of Standard De-
Throughput Values for Example 3 with a Correlation of 0 viation of Throughput for Example 3 with a Correlation of
at w1 (mix time) = 0.5 Hours. 0 at w1 (mix time) = 0.5 Hours.

an expected throughput of 1877.8 with a standard ing the delta method, we derived expressions for the
deviation of 20.9. Figures 9 and 10 show the mean expected value and variance of the response variable.
and standard deviation contour plots for the uncor- Using these expressions, an experimenter can nd
related scenario for a mix time of 0.5. By looking at variable levels that will lead to results that have an
the plots, it becomes clear that for the no-correlation optimal combination of mean and variance values.
case higher percentages of soap lead to blends that We have shown how both simple graphical methods
have high throughput as well as low standard devia- and more sophisticated non-linear optimization tech-
tions. A comparison of Figures 8 and 10 shows that niques can help nd these optimum conditions. The
increasing the level of co-surfactant causes the stan- desirability approach to optimization gives the ex-
dard deviation to increase when correlation is con- perimenter exibility in assigning weights to the val-
sidered but decrease when the correlation is ignored. ues within each response and weights among the re-
Thus, the correlation between the noise variables has sponses. The eects of correlation among the noise
had a dramatic impact on the results in this exam- variable were also demonstrated. We showed how ig-
ple. The impact will not be so large in all cases, and noring this correlation could lead to solutions that
we have seen other examples where the correlation are not optimal when correlation is present.
has a minimal impact. However, when correlation is
present it should be modeled, as this cannot do any Acknowledgments
harm to model correlation but can lead to erroneous
results if ignored. The authors would like to thank the Editor and
three anonymous referees for their helpful comments
Conclusions on an earlier draft of this paper.

We developed models for experiments contain- References


ing mixture components, controllable process (or Carlyle, W. M.; Montgomery, D. C.; and Runger, G.
other non-mixture) variables, and uncontrollable C. (2000). Optimization Problems and Methods in Quality
noise variables. Models were also developed that al- Control and Improvement. Journal of Quality Technology
low for correlations between the noise variables. Us- 32, pp. 131.

Journal of Quality Technology Vol. 35, No. 4, October 2003


MIXTURE-PROCESS VARIABLE EXPERIMENTS WITH NOISE VARIABLES 405

Cornell, J. A. (1995). Fitting Models to Data from Mixture Piepel, G. F. and Cornell, J. A. (1985). Models for Mix-
Experiments Containing Other Factors. Journal of Quality ture Experiments When the Response Depends on the Total
Technology 27, pp. 1333. Amount. Technometrics 27, pp. 219227.
Cornell, J. A. (2002). Experiments with Mixtures: Designs, Piepel, G. F. and Cornell, J. A. (1987). Designs for
Models, and the Analysis of Mixture Data, 3rd ed. John Wi- Mixture-Amount Experiments. Journal of Quality Technol-
ley & Sons, New York, NY. ogy 26, pp. 177196.
Cornell, J. A. and Gorman, J. W. (1984). Fractional De- Piepel, G. F. and Cornell, J. A. (2001). A Catalog of Mix-
sign Plans for Process Variables in Mixture Experiments. ture Experiment Examples. BN-SA-3298, Battelle, Pacic
Journal of Quality Technology 16, pp. 2038. Northwest Laboratories. Richland, WA.
Derringer, G. and Suich, R. (1980). Simultaneous Opti- Rice, J. (1995). Mathematical Statistics and Data Analysis,
mization of Several Response Variables. Journal of Quality 2nd ed. Wadsworth & Brooks/Cole, Pacic Grove, CA.
Technology 12, pp. 214219. Stat-Ease, Inc. (2002). Design-Expert Version 6.06. Min-
Lucas, J. M. (1994). How to Achieve a Robust Process Using neapolis, MN.
Response Surface Methodology. Journal of Quality Tech- Steiner, S. H. and Hamada, M. (1997). Making Mixtures
nology 26, pp. 248260. Robust to Noise and Mixing Measurement Errors. Journal
Montgomery, D. C. (1999). Experimental Designs for Prod- of Quality Technology 29, pp. 441450.
uct and Process Design and Development (with Commen- Taguchi, G. (1986). Introduction to Quality Engineering.
tary). Journal of the Royal Statistical Society D 48, pp. 159 Quality Resources, White Plains, NJ.
177. Taguchi, G. (1987). System of Experimental Design: Engi-
neering Methods to Optimize Quality and Minimize Cost.
Myers R. H.; Khuri, A. I.; and Vining, G. G. (1992). Re-
Quality Resources, White Plains, NJ.
sponse Surface Alternatives to the Taguchi Robust Param-
Taguchi, G. and Wu, Y. (1985). Introduction to O-Line
eter Design Approach. American Statistician 46, pp. 131
Quality Control. Central Japan Quality Control Association
139.
(available from American Supplier Institute, Dearborn, MI).
Myers R. H. and Montgomery, D. C. (2002). Response Sur-
face Methodology, 2nd ed. John Wiley & Sons, New York,
NY. Key Words: Mixture Experiments, Mixture-Process
Nair, V. N., (ed.) (1992). Taguchis Parameter Design: A Experiments, Noise Variables, Response Surface
Panel Discussion. Technometics 34, pp. 127161. Methodology, Robust Parameter Design.

Vol. 35, No. 4, October 2003 www.asq.org

You might also like