In experimental investigation we study the relationship between the input factors and the response

(output) of any process or system. The purpose may be to optimize the response or to understand
the system. If the input factors are quantitative and are a few, Response Surface Methodology
(RSM) can be used to study the relationship. Suppose we want to determine the levels of temperature
(X1 ) and pressure (X2) that maximize the yield (Y) of a chemical process. That is, the process
yield is a function of the levels of temperature and pressure which can be represented as
Y = f(X1, X2) + e
where e is the observed error in the response Y.
If the expected response E(Y) = f(X1, X2) = , then the surface is represented by
= f(X1, X2)
An example of a response surface is shown in Figure 8.1.

55
Response

50

45
140
40
120
Pressure (B)
60
65 100
70
Temperature ( A)
 In RSM, a sequential experimentation strategy is followed. This involves the search of input
factor space by using first-order experiment followed by a second-order experiment. The first
order equation is given by
Y= 0 + 1X1 + 2 X2 + ... + k Xk + e (8.1)
If there is curvature in the system, higher order polynomial, usually a second-order model
given below is used.
K K
2
Y = 0 +  i Xi +  ii X i +  ij X i X j +e (8.2)
i 1 i 1 i< j

The method of least squares is used to estimate the parameters (discussed in Chapter 2). The
response surface analysis is then performed using the fitted surface. The objective of RSM is to
determine the optimum operating conditions. This analysis is generally performed using computer
software.
The model parameters can be estimated effectively if proper experimental designs are
used to collect data. Such designs are called response surface designs. Some of the efficient
designs used in second-order experiment are the Central Composite Design (CCD) and the
Box–Behnken Design.

Selecting an appropriate design for fitting and analysing the response surface is important. These
designs should allow minimum number of experimental runs but provide complete information
required for fitting the model and checking model adequacy.

Suppose we want to fit a first-order model in k variables

k
Y = 0 +  i Xi +e (8.3)
i 1

These designs include 2k factorial and fractions of 2k series in which main effects are not
aliased with each other. Here the low and high levels of the factors are denoted by 1 notation.
We will not obtain error estimate in these designs unless some runs are replicated. A common
method of including replications is to take some observations at the center point. This design has
already been discussed in Chapter 5.

The central composite design for k factors include:

(i) nF factorial points (corner/cube points) each at two levels indicated by (–1, +1)
(ii) nC centre points indicated by 0.
(iii) 2k axial points (star points)
In selecting a CCD, the following three issues are to be addressed:
1. Choosing the factorial portion of the design
2. Number of center points
3. Determining the value for the axial points

Choice of factorial portion of the design: A CCD should have the total number of distinct
(k + 1)(k + 2)
design points N = nF + 2k + 1, must be at least
2
where,
k = number of factors,
nF = number of factorial points,
2k = number of star points
(k + 1)(k + 2)
= smallest number of points required for the second order model to be estimated.
2
Table 8.1 can be used to select a CCD with economical run size for 2 k 4.

k (k + 1) (k + 2)/2 N nF Factorial portion (cube points)

2 6 7 2 22–1(I = AB)
2 6 9 4 22
3–1
3 10 11 4 2III (I = ABC)
3
3 10 15 8 2
4–1
4 15 17 8 2III (I = ABD)
4 15 20 11 11 4 sub matrix of 12 run PB design*
4 15 25 16 24
*Obtained by taking the first four columns and deleting row 3 of the 12-run Plackett–Burman design.

Further discussion on small composite designs on the Placket–Burman designs can be found
in Draper and Lin (1990).

In general value is selected between 1 and k . A design that produces contours of constant
standard deviation of predicted response is called rotatable design. A CCD is made rotatable by
selecting
= (nF)1/4 (8.4)
This design provides good predictions throughout the region of interest. If the region of interest
is spherical, the best choice of = k . This design is often called spherical CCD.
In general, the choice of depends on the geometric nature of and practical constraints on
the design region. Unless some practical considerations dictate the choice of , Eq. (8.4) serves
as a useful guideline.
The rule of thumb is to have three to five center points when is close to k . If the purpose
is to obtain the error it is better to have more than 4 or 5 runs. Further discussion on choice of
and the number of runs at the center point can be found in Box and Draper (1987). The
graphical representation of CCD for 2 factors is shown in Figure 8.2. The CCD for k = 2 and
k = 3 with five centre points is given in Tables 8.2 and 8.3 respectively.

2 2

Run Factors
A B
1 –1 –1
2 1 –1
3 –1 1
4 1 1
5 –1.41 0
6 1.41 0
7 0 –1.14
8 0 1.14
9 0 0
10 0 0
11 0 0
12 0 0
13 0 0

( = k and nc = 5)
 =

Run Factors
A B C
1 –1 –1 –1
2 1 –1 –1
3 –1 1 –1
4 1 1 –1
5 –1 –1 1
6 1 –1 1
7 –1 1 1
8 1 1 1
9 –1.73 0 0
10 1.73 0 0
11 0 –1.73 0
12 0 1.73 0
13 0 0 –1.73
14 0 0 1.73
15 0 0 0
16 0 0 0
17 0 0 0
18 0 0 0
19 0 0 0

( = k and nc = 5)

Box and Behnken (1960) developed a set of three-level second-order response surface designs.
These were developed by combining two-level factorial designs with balanced or partially balanced
incomplete block designs. The construction of the smallest design for three factors is explained
below. A balanced incomplete block design with three treatments and three blocks is given by

Block Treatments
1 2 3
1 X X
2 X X
3 X X

The three treatments are considered as three factors A, B, C in the experiment. The two
crosses (Xs) in each block are replaced by the two columns of the 22 factorial design and a
column of zeros are inserted in the places where the cross do not appear. This procedure is
repeated for the next two blocks and some centre points are added resulting in the Box-Behnken
design for k = 3 which is given in Table 8.4. Similarly, designs for k = 4 and k = 5 can be
obtained. The advantage of this design is that each factor requires only three levels.
 =

Run A B C
1 –1 –1 0
2 –1 1 0
3 1 –1 0
4 1 1 0
5 –1 0 –1
6 –1 0 1
7 1 0 –1
8 1 0 1
9 0 –1 –1
10 0 –1 1
11 0 1 –1
12 0 1 1
13 0 0 0
14 0 0 0
15 0 0 0

In this section we will discuss how polynomial models can be developed using 2k factorial
designs.
Fitting of a polynomial model can be treated as a particular case of multiple linear regression.
The experimental designs used to develop the regression models must facilitate easy formation
of least squares normal equations so that solution can be obtained with least difficulty.

Let us consider a 23 factorial design to illustrate a general approach to fit a linear equation.
Suppose we have the following three factors studied each at two levels in an experiment.

Factors Level
Low (–1) High (+)
Time (min) (A) 6 8
Temperature (B) 240°C 300°C
Concentration (%) (C) 10 14

Let the first-order model to be fitted is

Yˆ = 0 + 1 X1 + 2 X2 + 3 X3 (8.5)
where X1, X2, X3 are the coded variables with 1 as high and low levels of the factors A, B and
C respectively.
 The relation between the coded variable (X) and the natural variables is given by

V  (h + l )/2
(8.6)
(h  l )/2

V = natural/actual variable such as time, temperature etc.

h = value of high level of the factor
l = value of low level of the factor
Accordingly, we have

Time  7 Temperature  270 Concentration  12

X1 = , X2 = , X3 =
1.0 30 2.0
Suppose we have the following experimental design and data (Table 8.5).

Run Factors Y
X1 X2 X3
(1) –1 –1 –1 61
a 1 –1 –1 83
b –1 1 –1 51
ab 1 1 –1 70
c –1 –1 1 66
ac 1 –1 1 92
bc –1 1 1 56
abc 1 1 1 83

The data is analysed as follows. In order to facilitate easy computation of model parameters
using least-square method, a column X0 with +1s are added to the design in Table 8.5 as given
in Table 8.6.

X0 X1 X2 X3 y

1 –1 –1 –1 61
1 1 –1 –1 83
1 –1 1 –1 51
1 1 1 –1 70
1 –1 –1 1 66
1 1 –1 1 92
1 –1 1 1 56
1 1 1 1 83
 Obtain sum of product of each column with response y.
0y = X0 y = 562 (Grand total)
1y = X1 y = 94
2y = X 2 y = – 42
3y = X3 y = 32
This design (Table 8.5) is simply a 23 design. Hence these sums are nothing but contrast
totals.
That is, CA = 94, CB = – 42, CC = 32
Therefore, the regression coefficients can be estimated as follows:

0 =y
562
= = 70.25
8

1 1  CA 
1 = (Effect of A) =  
2 2  n 2 k 1 
where,
y = grand mean
CA = contrast A = 94
n = number of replications = 1
1 = 94/8 = 11.75

Similarly, 2 = – 42/8 = –5.25 and 3 = 32/8 = 4.00

The fitted regression model is
Ŷ = 70.25 + 11.75 X1 – 5.25 X2 + 4.0 X3 (8.7)

Analysis of variance: Usually, the total sum of squares is partitioned into SS due to regression
(linear model) and residual sum of squares. And
SS residual (SS R) = Experimental error SS or pure SS + SS lack of fit
In this illustration, the experiment is not replicated and hence pure SS is zero. Therefore,
SS Total = SS Linear model + SS Lack of fit (8.8)
These sum of squares are computed as follow:

N
SS Total =  yi2 – CF
i =1

(562) 2
= {(61)2 + (83)2 + ... + (83)2 } 
8
= 40956.0 – 39480.5 = 1475.50
 SS Model = All factor sum of squares. This can be computed from the contrast totals, which
is given below:
(C A ) 2 (94)2
SSA = = = 1104.5 (Since n = 1 and k = 3)
n2 k 8

(  42)2 (32) 2
Similarly, SSB = = 220.5 and SS C = = 128.0
8 8

SS Linear model = 1104.5 + 220.5 + 128.0 = 1453

The ANOVA for the first-order model is given in Table 8.7.

Source Sum of Degrees Mean F0

squares of freedom square

Linear model 1453.00 3 484.33 86.02

Lack of fit 22.50 4 5.63
(residual)
Total 1475.50 7

When tested with the lack of fit, the linear model is a good fit to the data.
Note that SS lack of fit = SSAB + SS AC + SS BC + SS ABC.
As these types of designs do not provide any estimate of experimental error variance, the
lack of fit cannot be tested. A significant lack of fit indicates the model inadequacy. In practice
we need experimental error when first-order or second-order designs are used. Otherwise the
experimenter does not know whether the equation adequately represents the true surface. To
obtain the experimental error either the whole experiment is replicated or some center points be
added to the 2k factorial design. Addition of center points do not alter the estimate of the regression
coefficients except that 0 becomes the mean of the whole experiment. Suppose nc number of
center points are added. These observations are used to measure experimental error and will have
nc – 1 degrees of freedom. If YC is the mean of these observations and YF is the mean of factorial
observations, ( YF – YC ) gives an additional degree of freedom for measuring the lack of fit. The
sum of squares for lack of fit is given by

nc nF
(YF  YC )2 (8.9)
nc + nF

where nF is the number of factorial points.

The effects are estimated using contrasts. The effects and coefficients for the first-order
model are summarized in Table 8.8.
 Term in the model Effect estimate Coefficient
Constant 70.25
Time (A) 23.50 11.75
Temperature (B) –10.50 –5.25
Concentration (C) 8.00 4.00
AB – 0.50 – 0.25
AC 3.00 1.50
BC 1.00 0.50
ABC 1.00 0.50

Figures 8.3 and 8.4 show the normal and half-normal plot of the effects respectively for the
first-order design given in Table 8.5. From the regression model [Eq. (8.7)], it is evident that high
value of time and concentration and low value of temperature produces high response. This is
also evident from the contour plot of response (Figure 8.5). Because the fitted model is first order
(includes only main effects), the surface plot of the response is a plane (Figure 8.6). The response
surface can be analysed using computer software to determine the optimum condition.

A
 Temperature * Time Concentration * Time
298 14

286 13

274
12
262
11
250

10
6.0 6.5 7.0 7.5 8.0 6.0 6.5 7.0 7.5 8.0
Concentration * Temperature
 80 90
70 Y 80
Y
60 300 70 14
275 12
50 60
250 Temperature 10 Concentration
6 6
7 8 7 8
Time
Time

65
Y 50
55 14
50 12
10 Concentration
250
275 300
Temperature

The general form of a second-order polynomial with two variables is:

2 2
Y= 0 + 1 X1 + 2 X2 + 11 X1 + 22 X 2 + 12 X1 X 2 +e (8.10)

In order to estimate the regression coefficients, we normally use CCD with two factors
(Table 8.2). This design has k = 2, nF = 4, nC = 5 and = 2 . The CCD for two factors with
data is given in Table 8.9.
For conducting the experiment, the coded variables (Xi) are converted into natural variables
(temperature, time, pressure etc.) using the following relation:

 (AH + AL )/2
Xi = (8.11)
(AH  AL )/2

where,
= natural variable
AH = value of high level of factor A
AL = value of low level of factor A
Suppose, the factor A is the temperature with low level = 240° and high level = 300°.

Temperature  270
Then, Xi = (8.12)
30
Now when Xi = 0 (center point), the level of the factor is set at 270° [from Eq. (8.12)]
Similarly, when Xi = 1.682 (the axial point), the factor level will be about 320°.
 Following this procedure, the experimental design can be converted into an experimental
layout.

Step 1: To the CCD (Table 8.2) add column X0 to the left and X12, X22, X1X2 and y (response)
to the right side as given in Table 8.9.
Step 2: Obtain sum of product of each column with y. Denote these sum of products as
(0y) for the X0 column, 1y for X1 column and so on. That is,
0y = X0 y, 1y = X1 y, 2y = X2 y, 11y = X12 y, 22y = X22 y and 12y = X 1X2 y
Step 3: Obtain regression coefficients using the following formulae:

0 = 0.2 (0y) – 0.10 iiy

where,
iiy = 11y + 22y
i = 0.125 (iy)
ii = 0.125 (iiy) + 0.01875 iiy – 0.10 (0y)
ij = 0.25 (ijy)

The quantity iiy is sum of cross products of all the squared terms with y.
In this model iiy is X12 y + X22 y
For the data shown, the parameters of the model are estimated as follows:
0y = X0 y = 110.20
1y = X1 y = 7.96
2y = X2 y = 4.12
11y = X12 y = 59
22y = X22 y = 62
12y = X1X 2 y = 1.00

X0 X1 X2 X12 X22 X1X2 y

1 –1 –1 1 1 1 6.5
1 1 –1 1 1 –1 8.0
1 –1 1 1 1 –1 7.0
1 1 1 1 1 1 9.5
1 –1.414 0 2 0 0 5.6
1 1.414 0 2 0 0 8.4
1 0 –1.414 0 2 0 7.0
1 0 1.414 0 2 0 8.5
1 0 0 0 0 0 9.9
1 0 0 0 0 0 10.3
1 0 0 0 0 0 10.0
1 0 0 0 0 0 9.7
1 0 0 0 0 0 9.8
 The regression coefficients are estimated as follows:

0 = 0.2 (0y) – 0.10 iiy = 0.2 (110.20) – 0.10 (59 + 62) = 9.94
1 = 0.125 (1y) = 0.125 (7.96) = 0.995
2 = 0.125 (2y) = 0.125 (4.12) = 0.515
11 = 0.125 (iiy) + 0.01875 iiy – 0.10 (0y)
= 0.125 (59) + 0.01875 (121) – 0.10 (110.2) = –1.38
22 = 0.125 (62) + 0.01875 (121) – 0.10 (110.2) = –1.00
12 = 0.25 (ijy)
= 0.25 (1.00) = 0.25
Therefore, the fitted model is
Ŷ = 9.94 + 0.995 X1 + 0.515 X2 – 1.38 X12 – 1.00 X22 + 0.25 X1X2 (8.13)
Equation (8.13) can be used to study the response surface. Usually, we use computer software
for studying these designs. The software gives the model analysis including all the statistics
and the relevant plots. Figure 8.7 and 8.8 show the contour plot and surface plot of response
respectively for the Central Composite Design (CCD) with two factors A and B discussed
above.

1.0

Y
0.5 <4
4–5
5–6
B 0.0 6–7
7–8
8–9
–0.5
9–10
> 10

–1.0

–1.0 –0.5 0.0 0.5 1.0

A
 10

8
Y
6
1
4
0
B
–1 –1
0
A 1

Central Composite Design with three variables (factors) is given in Table 8.10. Usually, we
use computer software for solution of these models which will give the fitted model including
ANOVA and other related statistics and generate contour and response surface plots. Note that
y in Table 8.10 is the response column to be obtained from the experiment.

X0 X1 X2 X3 X12 X22 X32 X 1X 2 X1X3 X2X3 y

1 –1 –1 –1 1 1 1 1 1 1
1 1 –1 –1 1 1 1 –1 –1 1
1 –1 1 –1 1 1 1 –1 1 –1
1 1 1 –1 1 1 1 1 –1 –1
1 –1 –1 1 1 1 1 1 –1 –1
1 1 –1 1 1 1 1 –1 1 –1
1 –1 1 1 1 1 1 –1 –1 1
1 1 1 1 1 1 1 1 1 1
1 –1.682 0 0 2.828 0 0 0 0 0
1 1.682 0 0 2.828 0 0 0 0 0
1 0 –1.682 0 0 2.828 0 0 0 0
1 0 1.682 0 0 2.828 0 0 0 0
1 0 0 –1.682 0 0 2.828 0 0 0
1 0 0 1.682 0 0 2.828 0 0 0
1 0 0 0 0 0 0 0 0 0
1 0 0 0 0 0 0 0 0 0
1 0 0 0 0 0 0 0 0 0
1 0 0 0 0 0 0 0 0 0
1 0 0 0 0 0 0 0 0 0
1 0 0 0 0 0 0 0 0 0
8.1 Use the data from Table 8.11 23 design and fit a first-order model.

Factors
X1 X2 X3 Y
–1 –1 –1 31
1 –1 –1 43
–1 1 –1 34
1 1 –1 47
–1 –1 1 45
1 –1 1 37
–1 1 1 50
1 1 1 41

8.2 The data given in Table 8.12 have been collected from a Central Composite Design with
two factors. Fit a second-order model.

X1 X2 y
–1 –1 34
1 –1 26
–1 1 13
1 1 26
–1.414 0 30
1.414 0 31
0 –1.414 18
0 1.414 30
0 0 20
0 0 18
0 0 23
0 0 22
0 0 20