Research Methodology-

Design of Experiments

Universiti Kuala Lumpur

1
 Presentation Topics
1. Introduction to Design of Experiment
2. General Factorial Design: ANOVA (One-
way and Two-way)
3. Regression Modeling & Analysis
4. 2k Factorial Design
5. Screening design: Fractional factorial
6. Design optimization: RSM
7. (Extra)Robust design: Taguchi concept
2
 Learning Outcomes at the end of
this presentation

The participants will be able to:

◼ Explain some basic principles and
terminology
◼ Apply the strategy of experimentation
◼ Utilize Guidelines for planning,
conducting and analyzing experiments

3
Part 1. Introduction to DOE

4
 Introduction to DOE
◼ An experiment is a test or a series of tests
◼ Experiments are used widely in the
engineering world
 Process characterization & optimization
 Evaluation of material properties
 Product design & development
 Component & system tolerance determination

◼ “All experiments are designed experiments,

some are poorly designed, some are well-
designed” 5
 Causes of project failures
◼ No DOE (50%)
◼ Poor Planning (30%)
◼ Poor Execution
◼ Poor Design
◼ Poor Analysis
◼ Others

6
 Engineering Experiments
Benefits:
◼ Reduce time to
design/develop new
products & processes
◼ Improve performance of
existing processes
◼ Improve reliability and
performance of products
◼ Achieve product & process
robustness
◼ Evaluation of materials,
General model of a process or system
design alternatives, setting
component & system
tolerances, etc.
7
 The Basic Principles of DOE

◼ Randomization
 Running the trials in an experiment in random order
 Notion of balancing out effects of “lurking” variables

◼ Replication :
❖ Completely re-run experiment with same input levels.
❖ Used to determine impact of measurement error.

◼ Blocking
 Dealing with nuisance factors

8
 Terminology
◼ Response variable : Measured output value.
◼ Factors : Input variable that can be changed.
◼ Levels : Specific value of factors (inputs).
Continues or discrete (type of system).
◼ Interaction : Effect of one input factor depends
on level of another input factor.

9
 Types of Experiments
◼ Treatment Comparisons : To compare
several treatments of a factor.
◼ Variable screening : Have a large number
of factors, but only a few are important.
◼ Response Surface Exploration : Important
factors impact on the system is explored,
regression model building.

10
 Types of Experiments (cont’d)

◼ System Optimization : Interested in

determining the optimum conditions.
◼ System Robustness : Wish to optimize a
system and also reduce the impact of
uncontrollable factors (noise)

11
Strategy of Experimentation

◼ Build-test-fix
 thetinkerer’s approach
 “pound it to fit, paint it to match”
 impossible to know if true optimum
achieved
◼ you quit when it works!
 consistently slow
◼ requires intuition, luck, rework
◼ reoptimization and continual fire-fighting

12
Strategy of Experimentation (Cont’d)

13
Strategy of Experimentation (Cont’d)
◼ One-factor-at-a-time
 Sometimes associated with the “scientific” or
“engineering” method
 Devastated by interaction, also very inefficient
 procedure (2 level example)
◼ run all factors at one condition
◼ repeat, changing condition of one factor
◼ continuing to hold that factor at that condition, rerun with
another factor at its second condition
◼ repeat until all factors at their optimum conditions
 slow, expensive: many tests

14
 Strategy of Experimentation (Cont’d)

Process: Yield = f(temperature, pressure)

50% yield
30%
yield

Max yield: 50% at 78C, 130 psi?

15
Strategy of Experimentation (Cont’d)

A better view of the maximum yield!

Optimized yield is over 85%

Process: Yield = f(temperature, pressure)

16
 Strategy of Experimentation (Cont’d)

◼ Statistically designed experiments

 Based on Fisher’s factorial concept
 A statistics-based approach to designed
experiments
 A methodology to achieve a predictive
knowledge of a complex, multi-variable
process with the fewest trials possible
 An optimization of the experimental process
itself
17
 Objectives of DOE
◼ Design a new system (DESIGN).
◼ Finding out the most important variables that
affect a performance measure (SCREENING).
◼ Fine-tuning a process (OPTIMIZATION).
◼ Reducing the variability in a process
(VARIANCE REDUCTION).
◼ Finding a set of parameters that minimize the
noise in a system (ROBUSTNESS)

18
 Factorial Design (2 Factors)
◼ In a factorial experiment,
all possible
combinations of factor
levels are tested
◼ The golf experiment:
 Type of driver
 Type of ball
 Walking vs. riding
 Type of beverage
 Time of round
 Weather
 Type of golf spike
 Etc, etc, etc…
19
Factorial Design (2 factors-cont’d)

20
Factorial Design (3 Factors)

DOX 5E Montgomery 21
21
Factorial Designs (Fractional 4
Factorial)

22
 Planning, Conducting &
Analyzing an Experiment
1. Recognition of & statement of problem
2. Choice of factors, levels, and ranges
3. Selection of the response variable(s)
4. Choice of design
5. Conducting the experiment
6. Statistical analysis
7. Drawing conclusions, recommendations
23
Determine Outputs & Inputs

24
 Simple Test: 2 Sample t-Test
(One factor, 2 levels)
Portland Cement Formulation

Observation Modified Mortar Unmodified Mortar

(sample), j (Formulation 1) y1 j (Formulation 2) y2 j

1 16.85 17.50
2 16.40 17.63
3 17.21 18.25
4 16.35 18.00
5 16.52 17.86
6 17.04 17.75
7 16.96 18.22
8 17.15 17.90
9 16.59 17.96
10 16.57 18.15
25
 Dotplots of Form 1 and Form 2 Boxplots of Form 1 and Form 2
(means are indicated by lines) (means are indicated by solid circles)

18.5
18.3

17.5
17.3

16.5
16.3

Form 1 Form 2 Form 1 Form 2

26
The Hypothesis Testing Framework

◼ Sampling from a normal distribution

◼ Statistical hypotheses: H 0 : 1 =  2
H1 : 1   2
27
 Estimation of Parameters
1 n
y =  yi estimates the population mean 
n i =1
n
1
S =
2

n − 1 i =1
( yi − y ) estimates the variance 
2 2

28
 Summary Statistics
Formulation 1 Formulation 2

“New recipe” “Original recipe”

y1 = 16.76 y2 = 17.92
S = 0.100
1
2 S 22 = 0.061
S1 = 0.316 S 2 = 0.247
n1 = 10 n2 = 10

29
 How the Two-Sample t-Test Works:
Use S and S to estimate  and 
1
2 2
2
2
1
2
2

y1 − y2
The previous ratio becomes
2 2
S S
1
+ 2
n1 n2
However, we have the case where  =  =  2
1
2
2
2

Pool the individual sample variances:

( n − 1) S 2
+ ( n − 1) S 2
Sp = 1
2 1 2 2
n1 + n2 − 2
30
 How the Two-Sample t-Test Works:
The test statistic is
y1 − y2
t0 =
1 1
Sp +
n1 n2

◼ Values of t0 that are near zero are consistent with the

null hypothesis
◼ Values of t0 that are very different from zero are
consistent with the alternative hypothesis
◼ t0 is a “distance” measure-how far apart the averages
are expressed in standard deviation units

31
 The Two-Sample (Pooled) t-Test
(n1 − 1) S12 + (n2 − 1) S 22 9(0.100) + 9(0.061)
S =
2
= = 0.081
n1 + n2 − 2 10 + 10 − 2
p

S p = 0.284

y1 − y2 16.76 − 17.92
t0 = = = −9.13
1 1 1 1
Sp + 0.284 +
n1 n2 10 10

The two sample means are about 9 standard deviations apart

Is this a "large" difference?

32
 The Two-Sample (Pooled) t-Test

◼ So far, we haven’t really

done any “statistics”
◼ We need an objective
basis for deciding how
large the test statistic t0
really is
◼ In 1908, W. S. Gosset
derived the reference
distribution for t0 …
called the t distribution
◼ Tables of the t distribution

33
 The Two-Sample (Pooled) t-Test
◼ A value of t0 between
–2.101 and 2.101 is
consistent with equality
of means
◼ It is possible for the
means to be equal and
t0 to exceed either 2.101
or –2.101, but it would
be a “rare event” …
leads to the conclusion
that the means are
different
◼ Could also use the P-
value approach

34
The Two-Sample (Pooled) t-Test

◼ The P-value is the risk of wrongly rejecting the null hypothesis of

equal means (it measures rareness of the event)
◼ The P-value in our problem is P = 3.68E-8

35
Minitab Two-Sample t-Test Results
Two-Sample T-Test and CI: Form 1, Form 2
Two-sample T for Form 1 vs Form 2

N Mean StDev SE Mean

Form 1 10 16.764 0.316 0.10
Form 2 10 17.922 0.248 0.078

Difference = mu Form 1 - mu Form 2

Estimate for difference: -1.158
95% CI for difference: (-1.425, -0.891)
T-Test of difference = 0 (vs not =): T-Value = -9.11
P-Value = 0.000 DF = 18
Both use Pooled StDev = 0.284

36
 Checking Assumptions –
The Normal Probability Plot
Tension Bond Strength Data
ML Estimates

Form 1
99
Form 2

95 Goodness of Fit
AD*
90
1.209
80 1.387

70
Percent

60
50
40
30
20

10
5

16.5 17.5 18.5

Data

37
 Importance of the t-Test
◼ Provides an objective framework for
simple comparative experiments
◼ Could be used to test all relevant
hypotheses in a two-level factorial design,
because all of these hypotheses involve
the mean response at one “side” of the
cube versus the mean response at the
opposite “side” of the cube
38
Part 2. General Factorial Design:
ANOVA (One-way and Two-way)

39
 What If There Are More Than
Two Factor Levels?
◼ The t-test does not directly apply
◼ There are lots of practical situations where there are
either more than two levels of interest, or there are
several factors of simultaneous interest
◼ The analysis of variance (ANOVA) is the appropriate
analysis “engine” for these types of experiments
◼ The ANOVA was developed by Fisher in the early
1920s, and initially applied to agricultural experiments
◼ Used extensively today for industrial experiments

40
 An Example
◼ An engineer is interested in investigating the relationship
between the RF power setting and the etch rate for this tool. The
objective of an experiment like this is to model the relationship
between etch rate and RF power, and to specify the power
setting that will give a desired target etch rate.
◼ The response variable is etch rate.
◼ She is interested in a particular gas (C2F6) and gap (0.80 cm),
and wants to test four levels of RF power: 160W, 180W, 200W,
and 220W. She decided to test five wafers at each level of RF
power.
◼ The experimenter chooses 4 levels of RF power 160W, 180W,
200W, and 220W
◼ The experiment is replicated 5 times – runs made in random
order

41
42
An Example

43
• Does changing the power change the
mean etch rate?
• Is there an optimum level for power?
• We would like to have an objective
way to answer these questions
• The t-test really doesn’t apply here –
more than two factor levels

44
 The Analysis of Variance

◼ In general, there will be a levels of the factor, or a treatments,

and n replicates of the experiment, run in random order…a
completely randomized design (CRD)
◼ N = an total runs
◼ We consider the fixed effects case…the random effects case
will be discussed later
◼ Objective is to test hypotheses about the equality of the a
treatment means
45
 The Analysis of Variance
◼ The name “analysis of variance” stems from a
partitioning of the total variability in the
response variable into components that are
consistent with a model for the experiment
◼ The basic single-factor ANOVA model is
 i = 1, 2,..., a
yij =  +  i +  ij , 
 j = 1, 2,..., n

 = an overall mean,  i = ith treatment effect,

 ij = experimental error, NID(0,  2 )
46
 Models for the Data

There are several ways to write a model

for the data:
yij =  +  i +  ij is called the effects model
Let i =  +  i , then
yij = i +  ij is called the means model
Regression models can also be employed

47
 The Analysis of Variance
◼ Total variability is measured by the total
sum of squares:
a n
SST =  ( yij − y.. ) 2
i =1 j =1

◼ The basic ANOVA partitioning is:

a n a n

 ij ..  i. .. ij i.
( y − y
i =1 j =1
) 2
= [( y − y ) +
i =1 j =1
( y − y )]2

a a n
= n ( yi. − y.. ) +  ( yij − yi. ) 2
2

i =1 i =1 j =1

SST = SSTreatments + SS E
48
 The Analysis of Variance

SST = SSTreatments + SS E
◼ A large value of SSTreatments reflects large differences in
treatment means
◼ A small value of SSTreatments likely indicates no
differences in treatment means
◼ Formal statistical hypotheses are:
H 0 : 1 =  2 = = a
H1 : At least one mean is different

49
 The Analysis of Variance
◼ While sums of squares cannot be directly compared to
test the hypothesis of equal means, mean squares can
be compared.
◼ A mean square is a sum of squares divided by its degrees
of freedom:
dfTotal = dfTreatments + df Error
an − 1 = a − 1 + a (n − 1)
SSTreatments SS E
MSTreatments = , MS E =
a −1 a (n − 1)

◼ If the treatment means are equal, the treatment and error

mean squares will be (theoretically) equal.
◼ If treatment means differ, the treatment mean square will
be larger than the error mean square.
50
 The Analysis of Variance is
Summarized in a Table

◼ The reference distribution for F0 is the Fa-1, a(n-1) distribution

◼ Reject the null hypothesis (equal treatment means) if

F0  F ,a −1,a ( n −1)
51
An Example

52
ANOVA Table

53
The Reference Distribution:

P-value

54
 ANOVA calculations are usually done via
computer

◼ Popular software packages, Design-Expert,

JMP, SAS, SPSS, StatGraph and Minitab
◼ May consider HyperStudy (Altair Hyperworks),
ModeFrontier, InSight

55
 Model Adequacy Checking in
the ANOVA

◼ Checking assumptions is important

◼ Normality
◼ Constant variance
◼ Independence
◼ Have we fit the right model?
◼ Later we will talk about what to do if
some of these assumptions are
violated 56
 Model Adequacy Checking in the ANOVA

◼ Examination of
residuals
eij = yij − yˆij
= yij − yi.

◼ Computer software
generates the residuals
◼ Residual plots are very
useful
◼ Normal probability plot
of residuals

57
Other Important Residual Plots

58
 Post-ANOVA Comparison of
Means
◼ The analysis of variance tests the hypothesis of equal
treatment means
◼ Assume that residual analysis is satisfactory
◼ If that hypothesis is rejected, we don’t know which
specific means are different
◼ Determining which specific means differ following an
ANOVA is called the multiple comparisons problem
◼ There are lots of ways to do
◼ We will use pair wise t-tests on means…sometimes
called Fisher’s Least Significant Difference (or Fisher’s
LSD) Method

59
Design-Expert Output

60
Graphical Comparison of Means

61
 Two-way ANOVA

Definition of a factor effect: The change in the mean response when the
factor is changed from low to high
40 + 52 20 + 30
A = y A+ − y A− = − = 21
2 2
30 + 52 20 + 40
B = yB+ − yB− = − = 11
2 2
52 + 20 30 + 40
AB = − = −1
2 2 62
The Case of Interaction:

50 + 12 20 + 40
A = y A+ − y A− = − =1
2 2
40 + 12 20 + 50
B = yB+ − yB− = − = −9
2 2
12 + 20 40 + 50
AB = − = −29
2 2
63
 Regression Model & The
Associated Response
Surface

y =  0 + 1 x1 +  2 x2
+ 12 x1 x2 + 
The least squares fit is
yˆ = 35.5 + 10.5 x1 + 5.5 x2
+0.5 x1 x2
 35.5 + 10.5 x1 + 5.5 x2

64
 The Effect of
Interaction on the
Response Surface

Suppose that we add an

interaction term to the model:

yˆ = 35.5 + 10.5 x1 + 5.5 x2

+8 x1 x2

Interaction is actually
a form of curvature

65
 Statistical (effects) model:
 i = 1, 2,..., a

yijk =  +  i +  j + ( )ij +  ijk  j = 1, 2,..., b
k = 1, 2,..., n

Other models (means model, regression models) can be useful 66
 Extension of the ANOVA to Factorials

a b n a b

 ijk ...

( y − y )
i =1 j =1 k =1
= 2
bn  i.. ...
( y − y
i =1
) + an  . j. ...
( 2
y − y ) 2

j =1
a b a b n
+ n ( yij . − yi.. − y. j . + y... ) +  ( yijk − yij . ) 2
2

i =1 j =1 i =1 j =1 k =1

SST = SS A + SS B + SS AB + SS E
df breakdown:
abn − 1 = a − 1 + b − 1 + (a − 1)(b − 1) + ab(n − 1)
67
ANOVA Table

68
 Examples (One-way ANOVA)

A rental car company wants to investigate whether the type of car

rented affects the length of the rental period. An experiment is run
for one week at a particular location, and 10 rental contracts are
selected at random for each car type. The results are shown in the
following table.

69
 Examples (One-way ANOVA) (cont’d)

1. Is there evidence to support a claim that the type of car rented affects the
length of the rental contract? Use  = 0.05. If so, which types of cars are
responsible for the difference?

2. Analyze the residuals from this experiment and comment on

the model adequacy.

70
 Examples (One-way ANOVA) (cont’d)

Solution no.1

71
 Examples (One-way ANOVA) (cont’d)

Solution no.2

There is nothing unusual about the residuals

72
 Examples (Two-way ANOVA)
An engineer suspects that the surface finish of a metal part is influenced by
the feed rate and the depth of cut. She selects three feed rates and four
depths of cut. She then conducts a factorial experiment and obtains the
following data:

73
 Examples (Two-way ANOVA) (cont’d)

Questions:

1. Analyze the data and draw conclusions. Use  = 0.05

2. Prepare appropriate residual plots and comment on the

model’s adequacy.

74
 Examples (Two-way ANOVA) (cont’d)

Solution no.1

75
 Examples (Two-way ANOVA) (cont’d)
Solution no.2

There is nothing unusual about the residuals

76
Part 3. REGRESSION MODELING
AND ANALYSIS

77
Introduction
• Regression analysis is the most often applied technique of
statistical analysis and modeling.
• In general, it is used to model a response variable (Y) as a
function of one or more driver variables (X1, X2, ..., Xp).
• The functional form used is:
• Yi = 0 + 1X1i + 2X2i + ... + pXpi + 

Regression models are empirical

Regression modeling is often performed on undesigned or unplanned
data
Regression modeling is also used extensively to build models to data
from designed experiments
78
 Introduction cont’d)

❖ If there is only one driver variable, X, then we usually

speak of “simple” linear regression analysis.
❖When the model involves
❖(a) multiple driver variables,
❖(b) a driver variable in multiple forms, or
❖(c) a mixture of these, then we speak of “multiple linear
regression analysis”.
❖ The “linear” portion of the terminology refers to the
response variable being expressed as a “linear
combination” of the driver variables.

79
 Introduction (cont’d)

❖Model assumptions are stated in terms of

the random errors, , as follows:
❖the errors are normally distributed,
❖with mean = zero, and
❖constant variance 2, that does not depend on
the settings of the driver variables, and
❖the errors are independent of one another.
❖This is often summarized symbolically as: 
is NID(0, 2)
80
 Model Estimation
• Ordinarily the regression coefficients (the ’s) are of
unknown value and must be estimated from sample
information. The estimate of a given coefficient, i, is often
symbolized by i.
• Although there are well-established statistical/ mathematical
methods for determining estimates by these methods, they are
generally tedious and are well-suited to performance by a
computer. The resulting estimated model is:

^
Yi = ^0 + ^1X1i + ^2X2i + ... + ^kXki
^i - Yi
• The random error term, i, is then estimated by ei = Y
81
 Interval Estimation
• Estimates will vary from sample to sample and it is useful to
have estimates of the standard deviations of these estimates, Si.
These estimated standard deviations tend to be included in the
regression output of most statistical packages and can be used in
the formation of confidence intervals for the true value of i,
that is:
i +/- t/2,n-(k+1)Si
• Where t/2,n-(p+1) is the value of Student’s t distribution with n-
(p+1) degrees of freedom that places a proportion /2 in the
upper tail of the t-distribution.

82
 Interval Estimation: Key Concept

➢ When examining a confidence interval for a

particular regression coefficient, j, we will want to
know whether the interval includes the value zero.
➢ If zero is included in the interval then, conceivably,
j = 0, which would imply that the model could be
simplified by dropping the corresponding term, jXj
from the model.
➢ Otherwise, the corresponding variable, Xj is
considered to be a potentially important predictor or
determinant of Y.

83
Written in matrix notation as:

84
 Where :

Test for significance of regression;

85
The test uses an ANOVA approach. The regression or model sum
of squares is

86
 Coefficient of multiple determination R2, where:

Just as in designed experiments, R2 is a measure of the amount of reduction in the

variability of y obtained by using the regressor variables x1, x2, . . . ., xk in the model, i.e.,
how good is the model.

If R2 and R2adj differ significantly: most probably nonsignificant terms are added. R2adj
should not be less than 0.60 or 0.70, as a rule of thumb.

87
 Testing for Lack of Fit
Very important test in both general regression modeling and in
analysis of a designed experiment
Does the chosen model adequately fit the data, or should higher-
order terms be considered?
A statistical test can be performed provided that the error term can
be decomposed into the two components shown below:

where;
The sum of squares for lack of fit is

The sum of squares for pure error is

The test statistic for lack of fit is

88
 Example
The brake horsepower developed by an automobile engine on a dynamometer is
thought to be a function of the engine speed in revolutions per minute (rpm), the road
octane number of the fuel, and the engine compression. An experiment is run in the
laboratory and the data that follow are collected.

89
Example (cont’d)

0.002 90
Part 4. 2k FACTORIAL DESIGN

91
 The 2k Factorial Design
◼ Special case of the general factorial design; k
factors, all at two levels
◼ The two levels are usually called low and high
(they could be either quantitative or qualitative)
◼ Very widely used in industrial experimentation
◼ Form a basic “building block” for other very
useful experimental designs (DNA)
◼ Special (short-cut) methods for analysis

92
The Simplest Case: The 22
“-” and “+” denote
the low and high
levels of a factor,
respectively
Low and high are
arbitrary terms
Geometrically, the
four runs form the
corners of a square
Factors can be
quantitative or
qualitative, although
their treatment in the
final model will be
different
93
 Chemical Process Example

A = reactant concentration, B = catalyst amount,

y = recovery

94
 Analysis Procedure for a Factorial Design
◼ Estimate factor effects
◼ Formulate model
 With replication, use full model
 With an unreplicated design, use normal
probability plots
◼ Statistical testing (ANOVA)
◼ Refine the model
◼ Analyze residuals (graphical)
◼ Interpret results

95
 Estimation of Factor Effects
A = y A+ − y A−
The effect estimates are:
ab + a b + (1)
= − A = 8.33, B = -5.00, AB = 1.67
2n 2n
= 1n [ab + a − b − (1)]
B = yB+ − yB−
ab + b a + (1)
= −
2n 2n
= 1n [ab + b − a − (1)]
ab + (1) a + b
AB = −
2n 2n
= 1n [ab + (1) − a − b]

96
Estimation of Factor Effects
Form Tentative Model

Term Effect SumSqr % Contribution

Model Intercept
Model A 8.33333 208.333 64.4995
Model B -5 75 23.2198
Model AB 1.66667 8.33333 2.57998
Error Lack Of Fit 0 0
Error P Error 31.3333 9.70072

97
 Statistical Testing - ANOVA
Response: Conversion
ANOVA for Selected Factorial Model
Analysis of variance table [Partial sum of squares]
Sum of Mean F
Source Squares DF Square Value Prob > F
Model 291.67 3 97.22 24.82 0.0002
A 208.33 1 208.33 53.19 < 0.0001
B 75.00 1 75.00 19.15 0.0024
AB 8.33 1 8.33 2.13 0.1828
Pure Error 31.33 8 3.92
Cor Total 323.00 11

Std. Dev. 1.98 R-Squared 0.9030

Mean 27.50 Adj R-Squared 0.8666
C.V. 7.20 Pred R-Squared 0.7817
PRESS 70.50 Adeq Precision 11.669

The F-test for the “model” source is testing the significance of the
overall model; that is, is either A, B, or AB or some combination of
these effects important?
98
 Statistical Testing - ANOVA

Coefficient Standard 95% CI 95% CI

Factor Estimate DF Error Low High VIF
Intercept 27.50 1 0.57 26.18 28.82
A-Concent 4.17 1 0.57 2.85 5.48 1.00
B-Catalyst -2.50 1 0.57 -3.82 -1.18 1.00
AB 0.83 1 0.57 -0.48 2.15 1.00

General formulas for the standard errors of the model coefficients and
the confidence intervals are available.

99
 Refine Model
Response: Conversion
ANOVA for Selected Factorial Model
Analysis of variance table [Partial sum of squares]
Sum of Mean F
Source Squares DF Square Value Prob > F
Model 283.33 2 141.67 32.14 < 0.0001
A 208.33 1 208.33 47.27 < 0.0001
B 75.00 1 75.00 17.02 0.0026
Residual 39.67 9 4.41
Lack of Fit 8.33 1 8.33 2.13 0.1828
Pure Error 31.33 8 3.92
Cor Total 323.00 11

Std. Dev. 2.10 R-Squared 0.8772

Mean 27.50 Adj R-Squared 0.8499
C.V. 7.63 Pred R-Squared 0.7817
PRESS 70.52 Adeq Precision 12.702

There is now a residual sum of squares, partitioned into a “lack of fit”

component (the AB interaction) and a “pure error” component
100
 Regression Model for the Process
Coefficient Standard 95% CI 95% CI
Factor Estimate DF Error Low High VIF
Intercept 27.5 1 0.60604 26.12904 28.87096
A-Concentration
4.166667 1 0.60604 2.79571 5.537623 1
B-Catalyst -2.5 1 0.60604 -3.87096 -1.12904 1

Final Equation in Terms of Coded Factors:

Coded Factors: Regressor
Conversion = variables coded/formulated in
27.5
such a way high level will
4.166667 * A
-2.5 * B
become ‘+1’ and low level as ‘-1’
& therefore shows orthogonality
Final Equation in Terms of Actual Factors:

Conversion =
18.33333
0.833333 * Concentration
-5 * Catalyst

101
 Residuals and Diagnostic
Checking
DESIGN-EXPERT Plot Normal plot of residuals DESIGN-EXPERT Plot Residuals vs. Predicted
Conversion Conversion
2.16667

99
N o r m a l % p r o b a b i l i ty

95
0.916667
90

80

R e s id u a ls
70

50 -0.333333

30
20

10
-1.58333
5

2
1

-2.83333

20.83 24.17 27.50 30.83 34.17

-2.83333 -1.58333 -0.333333 0.916667 2.16667

Predicted
Res idual

102
 The Response Surface DESIGN-EXPERT Plot

PERT Plot Conversion

2.00
3 Conversion 3
X = A: Concentration
Y = B: Catalyst
entration
lyst
34.1667
23.0556
Points
1.75 30.8333

27.5

C o n v e rsi o n
B : C a ta l ys t

24.1667
25.2778
27.5
1.50 20.8333
29.7222

1.25
31.9444
2.00
25.00
1.75
22.50
3 3
1.00 1.50
20.00
15.00 17.50 20.00 22.50 25.00
B: Catalyst 1.25 17.50
A: Concentration
1.00 15.00 103
A: Concentration
The 23 Factorial Design

104
Effects in The 23 Factorial Design

A = y A+ − y A−
B = yB+ − yB−
C = yC + − yC −
etc, etc, ...

105
 An Example of a 23 Factorial Design

A = carbonation, B = pressure, C = speed, y = fill deviation

106
 Table of – and + Signs for the 23 Factorial Design

Factorial Effect

Treatment I A B AB C AC BC ABC
Combination

(1) = -4 + - - + - + + -

a=1 + + - - - - + +

b = -1 + - + - - + - +

ab = 5 + + + + - - - -

c = -1 + - - + + - - +

ac = 3 + + - - + + - -

bc = 2 + - + - + - + -

abc = 11 + + + + + + + +

Contrast 24 18 6 14 2 4 4

Effect 3.00 2.25 0.75 1.75 0.25 0.50 0.50

107
 Properties of the Table

◼ Except for column I, every column has an equal number of +

and – signs
◼ The sum of the product of signs in any two columns is zero
◼ Multiplying any column by I leaves that column unchanged
(identity element)
◼ The product of any two columns yields a column in the table:
A  B = AB
AB  BC = AB 2C = AC
◼ Orthogonal design
◼ Orthogonality is an important property shared by all factorial
designs
◼ All 2k designs are orthogonal designs
108
Estimation of Factor Effects

Term Effect SumSqr % Contribution

Model Intercept
Model A 3 36 46.1538
Model B 2.25 20.25 25.9615
Model C 1.75 12.25 15.7051
Model AB 0.75 2.25 2.88462
Model AC 0.25 0.25 0.320513
Model BC 0.5 1 1.28205
Model ABC 0.5 1 1.28205
Error LOF 0
Error P Error 5 6.41026

109
ANOVA Summary – Full Model

Response:Fill-deviation
ANOVA for Selected Factorial Model
Analysis of variance table [Partial sum of squares]
Sum of Mean F
Source Squares DF Square Value Prob > F
Model 73.00 7 10.43 16.69 0.0003
A 36.00 1 36.00 57.60 < 0.0001
B 20.25 1 20.25 32.40 0.0005
C 12.25 1 12.25 19.60 0.0022
AB 2.25 1 2.25 3.60 0.0943
AC 0.25 1 0.25 0.40 0.5447
BC 1.00 1 1.00 1.60 0.2415
ABC 1.00 1 1.00 1.60 0.2415
Pure Error 5.00 8 0.63
Cor Total 78.00 15

Std. Dev. 0.79 R-Squared 0.9359

Mean 1.00 Adj R-Squared 0.8798
C.V. 79.06 Pred R-Squared 0.7436
PRESS 20.00 Adeq Precision 13.416
110
Model Coefficients – Full Model

Coefficient
Standard 95% CI 95% CI
Factor Estimate DF Error Low High VIF
Intercept 1.00 1 0.20 0.54 1.46
A-Carbonation 1.50 1 0.20 1.04 1.96 1.00
B-Pressure 1.13 1 0.20 0.67 1.58 1.00
C-Speed 0.88 1 0.20 0.42 1.33 1.00
AB 0.38 1 0.20 -0.081 0.83 1.00
AC 0.13 1 0.20 -0.33 0.58 1.00
BC 0.25 1 0.20 -0.21 0.71 1.00
ABC 0.25 1 0.20 -0.21 0.71 1.00

111
Refine Model – Remove Nonsignificant Factors
Response: Fill-deviation
ANOVA for Selected Factorial Model
Analysis of variance table [Partial sum of squares]
Sum of Mean F
Source Squares DF Square Value Prob > F
Model 70.75 4 17.69 26.84 < 0.0001
A 36.00 1 36.00 54.62 < 0.0001
B 20.25 1 20.25 30.72 0.0002
C 12.25 1 12.25 18.59 0.0012
AB 2.25 1 2.25 3.41 0.0917
Residual 7.25 11 0.66
LOF 2.25 3 0.75 1.20 0.3700
Pure E 5.00 8 0.63
C Total 78.00 15

Std. Dev. 0.81 R-Squared 0.9071

Mean 1.00 Adj R-Squared 0.8733
C.V. 81.18 Pred R-Squared 0.8033
PRESS 15.34 Adeq Precision 15.424
112
Model Coefficients – Reduced Model

Coefficient Standard 95% CI 95% CI

Factor Estimate DF Error Low High
Intercept 1.00 1 0.20 0.55 1.45
A-Carbonation 1.50 1 0.20 1.05 1.95
B-Pressure 1.13 1 0.20 0.68 1.57
C-Speed 0.88 1 0.20 0.43 1.32
AB 0.38 1 0.20 -0.072 0.82

113
 Model Summary Statistics

◼ R2 and adjusted R2
SS Model 73.00
R =
2
= = 0.9359
SST 78.00
SS E / df E 5.00 / 8
R 2
Adj = 1− = 1− = 0.8798
SST / dfT 78.00 /15

◼ R2 for prediction (based on PRESS)

PRESS 20.00
RPred = 1 −
2
= 1− = 0.7436
SST 78.00 114
 Model Summary Statistics

◼ Standard error of model coefficients

2 MS E 0.625
se( ˆ ) = V ( ˆ ) = k
= k
= = 0.20
n2 n2 2(8)

◼ Confidence interval on model coefficients

ˆ − t / 2,df se( ˆ )    ˆ + t / 2,df se( ˆ )

E E

115
The Regression Model
Final Equation in Terms of Coded Factors:

Fill-deviation =
+1.00
+1.50 *A
+1.13 *B
+0.88 *C
+0.38 *A*B

Final Equation in Terms of Actual Factors:

Fill-deviation =
+9.62500
-2.62500 * Carbonation
-1.20000 * Pressure
+0.035000 * Speed
+0.15000 * Carbonation * Pressure

116
Residual Plots are Satisfactory
DESIGN-EXPERT Plot Normal plot of residuals
Fill-deviation

99

N o r m a l % p r o b a b i l i ty
95
90

80
70

50

30
20

10
5

-1.67 -0.84 0.00 0.84 1.67

Studentized Res iduals

117
 Model Interpretation
DESIGN-EXPERT Plot Interaction Graph
Fill-deviation B: Pres s ure
6

X = A: Carbonation Moderate
Y = B: Pressure

B- 25.000
interaction between
3.75
B+ 30.000
Actual Factor
carbonation level
F i l l - d e vi a ti o n

C: Speed = 225.00
and pressure
1.5

-0.75

-3

10.00 10.50 11.00 11.50 12.00

A: Carbonation
118
 Model Interpretation

DESIGN-EXPERT Plot Cube Graph

Fill-deviation Fill-deviation
X = A: Carbonation
Y = B: Pressure 1.13 4.88 Cube plots are
Z = C: Speed
often useful visual
displays of
B+ -0.63 3.13 experimental
results
B : P re s s u re

-0.37 1.88 C+

C: Speed

B- -2.13 0.12 C-
A- A+
A: Carbonation

119
 Contour & Response Surface Plots –
Speed at the High Level DESIGN-EXPERT Plot

DESIGN-EXPERT Plot 2 Fill-deviationFill-deviation2

30.00
X = A: Carbonation
Fill-deviation
Y = B: Pressure
X = A: Carbonation
Y = B: Pressure
Actual Factor
4.875
C: Speed = 250.00
Design Points
28.75
3.125
3.5625
Actual Factor
C: Speed = 250.00 2.25

F i l l -d e v i a t i o n
B : P re s s u re

0.9375
2.25
27.50
1.375 -0.375

0.5
26.25

30.00
12.00
28.75
11.50
2 2
25.00 27.50
11.00
10.00 10.50 11.00 11.50 12.00
B: Pressure 26.25 10.50
A: Carbonation
25.00 10.00
A: Carbonation
120
 The General 2k Factorial Design
There will be k main effects, and
k 
  two-factor interactions
 2
k 
  three-factor interactions
 3

1 k − factor interaction

121
 Unreplicated 2k Factorial Designs

◼ These are 2k factorial designs with one

observation at each corner of the “cube”
◼ An unreplicated 2k factorial design is also
sometimes called a “single replicate” of the
2k
◼ These designs are very widely used
◼ Risks…if there is only one observation at
each corner, is there a chance of unusual
response observations spoiling the results?

122
Spacing of
Factor Levels in
the Unreplicated
2k Factorial
Designs
If the factors are spaced
too closely, it increases
the chances that the
noise will overwhelm
the signal in the data
More aggressive
spacing is usually best

123
 Unreplicated 2k Factorial Designs

◼ Lack of replication causes potential problems

in statistical testing
 Replication admits an estimate of “pure error” (a
better phrase is an internal estimate of error)
 With no replication, fitting the full model results in
zero degrees of freedom for error

124
 Example of an Unreplicated 2k Design

◼ A 24 factorial was used to investigate the

effects of four factors on the filtration rate
of a resin
◼ The factors are A = temperature, B =
pressure, C = mole ratio, D= stirring rate
◼ Experiment was performed in a pilot plant

125
The Resin Plant Experiment

126
The Resin Plant Experiment

127
Estimates of the Effects
Term Effect SumSqr % Contribution
Model Intercept
Model A 21.625 1870.56 32.6397
Model B 3.125 39.0625 0.681608
Model C 9.875 390.062 6.80626
Model D 14.625 855.563 14.9288
Model AB 0.125 0.0625 0.00109057
Model AC -18.125 1314.06 22.9293
Model AD 16.625 1105.56 19.2911
Model BC 2.375 22.5625 0.393696
Model BD -0.375 0.5625 0.00981515
Model CD -1.125 5.0625 0.0883363
Model ABC 1.875 14.0625 0.245379
Model ABD 4.125 68.0625 1.18763
Model ACD -1.625 10.5625 0.184307
Model BCD -2.625 27.5625 0.480942
Model ABCD 1.375 7.5625 0.131959

128
The Normal Probability Plot of Effects
DESIGN-EXPERT Plot Normal plot
Filtration Rate

A: Temperature
B: Pressure 99
C: Concentration
D: Stirring Rate 95 A

N o r m a l % p r o b a b i l i ty
90
AD
80 D
70 C

50

30
20

10
5
AC
1

-18.12 -8.19 1.75 11.69 21.62

129
ANOVA Summary for the Model
Response: Filtration Rate
ANOVA for Selected Factorial Model
Analysis of variance table [Partial sum of squares]
Sum of Mean F
Source Squares DF Square Value Prob >F
Model 5535.81 5 1107.16 56.74 < 0.0001
A 1870.56 1 1870.56 95.86 < 0.0001
C 390.06 1 390.06 19.99 0.0012
D 855.56 1 855.56 43.85 < 0.0001
AC 1314.06 1 1314.06 67.34 < 0.0001
AD 1105.56 1 1105.56 56.66 < 0.0001
Residual 195.12 10 19.51
Cor Total 5730.94 15

Std. Dev. 4.42 R-Squared 0.9660

Mean 70.06 Adj R-Squared 0.9489
C.V. 6.30 Pred R-Squared 0.9128
PRESS 499.52 Adeq Precision 20.841

130
The Regression Model

Final Equation in Terms of Coded Factors:

Filtration Rate =
+70.06250
+10.81250 * Temperature
+4.93750 * Concentration
+7.31250 * Stirring Rate
-9.06250 * Temperature * Concentration
+8.31250 * Temperature * Stirring Rate

131
Model Residuals are Satisfactory
DESIGN-EXPERT Plot Normal plot of residuals
Filtration Rate

99

N o r m a l % p r o b a b i l i ty
95
90

80
70

50

30
20

10
5

-1.83 -0.96 -0.09 0.78 1.65

Studentized Res iduals

132
 Model Interpretation – Interactions

DESIGN-EXPERT Plot Interaction Graph DESIGN-EXPERT Plot Interaction Graph

Filtration Rate C: Concentration D: Stirring Rate
104 Filtration Rate 104

X = A: Temperature X = A: Temperature
Y = C: Concentration Y = D: Stirring Rate

C- -1.000 88.4426 D- -1.000 88.75

C+ 1.000 D+ 1.000
F i l tr a ti o n R a te

F i l tr a ti o n R a te
Actual Factors Actual Factors
B: Pressure = 0.00 B: Pressure = 0.00
D: Stirring Rate = 0.00 C: Concentration = 0.00
72.8851 73.5

57.3277 58.25

41.7702 43

-1.00 -0.50 0.00 0.50 1.00 -1.00 -0.50 0.00 0.50 1.00

A: Tem perature A: Tem perature

133
 Model Interpretation – Cube Plot
DESIGN-EXPERT Plot Cube Graph If one factor is
Filtration Rate
X = A: Temperature
Filtration Rate
dropped, the
72.25 92.38
Y = C: Concentration
Z = D: Stirring Rate unreplicated 24
Actual Factor
B: Pressure = 0.00
design will project
C+ 74.25 61.13
into two replicates
of a 23
C : C o n c e n tr a ti o n

Design projection is
44.25 100.63 D+
an extremely useful
D: Stirring Rate property, carrying
C- 46.25 69.38 D-
over into fractional
A- A+
A: Tem perature factorials

134
 Model Interpretation – Response Surface Plots

DESIGN-EXPERT Plot
DESIGN-EXPERT Plot Filtration Rate Filtration Rate
1.00
Filtration Rate
X = A: Temperature
X = A: Temperature Y = D: Stirring Rate
Y = D: Stirring Rate
90.125 Actual Factors
100.625
Actual Factors B: Pressure = 0.00
0.50 C: Concentration = 86.5313
-1.00
B: Pressure = 0.00 83.75
C: Concentration = -1.00
D : S ti r r i n g R a te

F i l t ra t i o n R a t e
72.4375
77.375
58.3438
0.00
51.9395 71
56.935
64.625 44.25

-0.50

1.00
1.00
0.50
0.50
-1.00
0.00
-1.00 -0.50 0.00 0.50 1.00 0.00
D: Stirring Rate-0.50 -0.50
A: Temperature
A: Tem perature -1.00 -1.00

With concentration at either the low or high level, high temperature and
high stirring rate results in high filtration rates
135
 The Drilling Experiment

A = drill load, B = flow, C = speed, D = type of mud,

y = advance rate of the drill
136
Effect Estimates - The Drilling Experiment

Term Effect SumSqr % Contribution

Model Intercept
Model A 0.9175 3.36722 1.28072
Model B 6.4375 165.766 63.0489
Model C 3.2925 43.3622 16.4928
Model D 2.29 20.9764 7.97837
Model AB 0.59 1.3924 0.529599
Model AC 0.155 0.0961 0.0365516
Model AD 0.8375 2.80563 1.06712
Model BC 1.51 9.1204 3.46894
Model BD 1.5925 10.1442 3.85835
Model CD 0.4475 0.801025 0.30467
Model ABC 0.1625 0.105625 0.0401744
Model ABD 0.76 2.3104 0.87876
Model ACD 0.585 1.3689 0.520661
Model BCD 0.175 0.1225 0.0465928
Model ABCD 0.5425 1.17722 0.447757

137
 Residual Plots

GN-EXPERT Plot DESIGN-EXPERT Plot Residuals vs. Predicted

_rate
Normal plot of residuals
adv._rate
2.58625

99
N o r m a l % p r o b a b i l i ty

95
1.44875
90

80

R e s id u a ls
70

50 0.31125
DESIGN-EXPERT Plot Residuals vs. Predicted
adv._rate
2.58625

1.44875

R e s id u a ls
0.31125

30
-0.82625

-1.96375

1.69 4.70 7.70 10.71 13.71

Predicted

20

10
-0.82625
5

-1.96375

-1.96375 -0.82625 0.31125 1.44875 2.58625

1.69 4.70 7.70 10.71 13.71

Res idual Predicted 138

 Addition of Center Points to a 2k Designs

◼ Based on the idea of replicating some

of the runs in a factorial design
◼ Runs at the center provide an estimate
of error and allow the experimenter to
distinguish between two possible
models: k k k
First-order model (interaction) y =  0 +   i xi +   ij xi x j + 
i =1 i =1 j i
k k k k
Second-order model y =  0 +   i xi +   ij xi x j +   ii xi2 + 
i =1 i =1 j i i =1

139
yF = yC  no "curvature"
The hypotheses are:
k
H 0 :   ii = 0
i =1
k
H1 :   ii  0
i =1

nF nC ( yF − yC ) 2
SS Pure Quad =
nF + nC

This sum of squares has a

single degree of freedom

140
Example

nC = 5
Usually between 3
and 6 center points
will work well
Design-Expert
provides the analysis,
including the F-test
for pure quadratic
curvature
141
ANOVA for Example
Response: yield
ANOVA for Selected Factorial Model
Analysis of variance table [Partial sum of squares]
Sum of Mean F
Source Squares DF Square Value Prob > F
Model 2.83 3 0.94 21.92 0.0060
A 2.40 1 2.40 55.87 0.0017
B 0.42 1 0.42 9.83 0.0350
AB 2.500E-003 1 2.500E-003 0.058 0.8213
Curvature 2.722E-003 1 2.722E-003 0.063 0.8137
Pure Error 0.17 4 0.043
Cor Total 3.00 8

Std. Dev. 0.21 R-Squared 0.9427

Mean 40.44 Adj R-Squared 0.8996
C.V. 0.51 Pred R-Squared N/A
PRESS N/A Adeq Precision 14.234

142
If curvature is significant, augment the design with
axial runs to create a central composite design.
The CCD is a very effective design for fitting a
second-order response surface model

143
Part 5. 2k-p FRACTIONAL FACTORIAL
DESIGN

144
 The 2k-p Fractional Factorial Design
◼ Motivation for fractional factorials is obvious; as the
number of factors becomes large enough to be
“interesting”, the size of the designs grows very
quickly
◼ Emphasis is on factor screening; efficiently identify
the factors with large effects
◼ There may be many variables (often because we
don’t know much about the system)
◼ Almost always run as unreplicated factorials, but
often with center points

145
 The One-Half Fraction of the 2k

◼ Notation: because the design has 2k/2 runs, it’s referred to as a

2k-1
◼ Consider a really simple case, the 23-1
◼ Note that I =ABC (where I = defining relation, ABC = generator)

146
The One-Half Fraction of the 23
For the principal fraction,
notice that the contrast for
estimating the main effect A
is exactly the same as the
contrast used for estimating
the BC interaction.
This phenomena is called
aliasing and it occurs in all
fractional designs
Aliases can be found
directly from the columns in
the table of + and - signs

The two one-half fractions of the 23

147
Construction of a One-half Fraction

The basic design; the design generator

148
The One-Quarter Fraction of the 2k

149
 A One-Quarter Fraction of the 26-2:
Example

◼ Injection molding process with six factors

◼ Design matrix
◼ Calculation of effects, normal probability plot of
effects
◼ Two factors (A, B) and the AB interaction are
important
◼ Residual analysis indicates there are some
dispersion effects

150
 The General 2k-p Fractional
Factorial Design
◼ 2k-1 = one-half fraction, 2k-2 = one-quarter fraction, 2k-3
= one-eighth fraction, …, 2k-p = 1/ 2p fraction
◼ Add p columns to the basic design; select p
independent generators
◼ Important to select generators so as to maximize
resolution
◼ Projection – a design of resolution R contains full
factorials in any R – 1 of the factors
◼ Blocking

151
 The General 2k-p Design: Resolution may
not be Sufficient

◼ Minimum aberration designs

152
 Resolution III Designs

◼ Designs with main effects aliased with two-factor

interactions
◼ Used for screening (5 – 7 variables in 8 runs, 9
- 15 variables in 16 runs, for example)
◼ A saturated design has k = N – 1 variables 27III− 4

153
Resolution III Designs

154
 Plackett-Burman Designs

◼ These are a different class of resolution III

design
◼ The number of runs, N, need only be a multiple
of four
◼ N = 4, 8, 12, 16, 20, 24, 28, 32, 36, 40, …
◼ The designs where N = 12, 20, 24, etc. are
called non-geometric PB designs

155
Plackett-Burman Designs

156
Plackett-Burman
Designs

Projection of the
12-run design into
3 and 4 factors
All PB designs
have projectivity 3
(contrast with other
resolution III
fractions)

157
Part 6. DESIGN OPTIMIZATION APPROACH
(RSM)

158
 Overview of Response Surface Methods

◼ Primary focus of previous slides is factor

screening
 Two-level factorials, fractional factorials are widely
used
◼ Objective of RSM is optimization
◼ RSM dates from the 1950s; early applications in
chemical industry

159
 RSM is a Sequential Procedure

◼ Factor screening
◼ Finding the region
of the optimum
◼ Modeling &
Optimization of the
response

160
 Response Surface Models

◼ Screening
y =  0 + 1 x1 +  2 x2 + 12 x1 x2 + 
◼ Steepest ascent
y =  0 + 1 x1 +  2 x2 + 
◼ Optimization
y =  0 + 1 x1 +  2 x2 + 12 x1 x2 +  x +  x + 
2
11 1
2
22 2

161
 The Method of Steepest Ascent

A procedure for moving

sequentially from an initial
“guess” towards to region of
the optimum
Based on the fitted first-order
model
ŷ = ˆ0 + ˆ1 x1 + ˆ2 x2

Steepest ascent is a gradient

procedure

162
 An Example of Steepest Ascent

Factor 1 Factor 2 Response 1

Std Run Block A:Time B:Temp yield
minutes degC percent
1 7 {1} -1 -1 39.3
2 6 {1} 1 -1 40.9
3 5 {1} -1 1 40
4 2 {1} 1 1 41.5
5 9 {1} 0 0 40.3
6 4 {1} 0 0 40.5
7 1 {1} 0 0 40.7
8 3 {1} 0 0 40.2
9 8 {1} 0 0 40.6

yˆ = 40.44 + 0.775 x1 + 0.325 x2

163
 An Example of Steepest Ascent
PERT Plot yield
2.000
An approximate step
tion T ime
p size and path can be
oints
1.250
determined graphically
Formal methods can
B : Tem p

40.5 41 41.5
also be used/
0.500

40 5
Types of experiments
along the path:
-0.250

•Single runs
39.5

-1.000 •Replicated runs

-1.000 -0.250 0.500 1.250 2.000

A: Reaction Tim e
164
 Results from the Example

The step size is 5 minutes of reaction time and 2 degrees F

What happens at the conclusion of steepest ascent?
165
 Analysis of the Second-Order
Response Surface Model
Factor 1 Factor 2 Response 1 Response 2 Response 3
Std Run Block A:time B:temp yield viscosity Mn
min degF
1 8 Block 1 80 170 76.5 62 2940
2 6 Block 1 90 170 78 66 3680
3 9 Block 1 80 180 77 60 3470
4 11 Block 1 90 180 79.5 59 3890
5 12 Block 1 77.93 175 75.6 71 3020
6 10 Block 1 92.07 175 78.4 68 3360
7 7 Block 1 85 167.93 77 57 3150
8 1 Block 1 85 182.07 78.5 58 3630
9 5 Block 1 85 175 79.9 72 3480
10 3 Block 1 85 175 80.3 69 3200
11 13 Block 1 85 175 80 68 3410
12 2 Block 1 85 175 79.7 70 3290
13 4 Block 1 85 175 79.8 71 3500

This is a central composite design

166
 The Second-Order Response
Surface Model
y =  0 + 1 x1 +  2 x2 + 12 x1 x2 +  x +  x + 
2
11 1
2
22 2

• These models are used widely in practice

• The Taylor series analogy
• Fitting the model is easy, some nice designs are
available
• Optimization is easy
• There is a lot of empirical evidence that they
work very well
167
Example

Sequential Model Sum of Squares

Sum of Mean F
Source Squares DF Square Value Prob > F
Mean 80062.16 1 80062.16
Linear 10.04 2 5.02 2.69 0.1166
2FI 0.25 1 0.25 0.12 0.7350
Quadratic 17.95 2 8.98 126.88 < 0.0001 Suggested
Cubic 2.042E-003 2 1.021E-003 0.010 0.9897 Aliased
Residual 0.49 5 0.099
Total 80090.90 13 6160.84

Model Summary Statistics

Std. Adjusted Predicted
Source Dev. R-Squared R-Squared R-Squared PRESS
Linear 1.37 0.3494 0.2193 -0.0435 29.99
2FI 1.43 0.3581 0.1441 -0.2730 36.59
Quadratic 0.27 0.9828 0.9705 0.9184 2.35 Suggested
Cubic 0.31 0.9828 0.9588 0.3622 18.33 Aliased

168
Example
ANOVA for Response Surface Quadratic Model
Analysis of variance table [Partial sum of squares]
Sum of Mean F
Source Squares DF Square Value Prob > F
Model 28.25 5 5.65 79.85 < 0.0001
A 7.92 1 7.92 111.93 < 0.0001
B 2.12 1 2.12 30.01 0.0009
A2 13.18 1 13.18 186.22 < 0.0001
B2 6.97 1 6.97 98.56 < 0.0001
AB 0.25 1 0.25 3.53 0.1022
Residual 0.50 7 0.071
Lack of Fit 0.28 3 0.094 1.78 0.2897
Pure Error 0.21 4 0.053
Cor Total 28.74 12

yˆ = 79.94 + 0.99 x1 + 0.52 x2 + 0.25 x1 x2

−1.38 x12 + −1.00 x22

169
 Contour Plots for Example
XPERT Pl ot yield
180.00
The contour plot is
e
p given in the natural
Poi nts
177.50 variables
The optimum is at
B : te m p

5
175.00
78.2573
about 87 minutes and
79.5606 176.5 degrees
78.9089
172.50
77.6056 Formal optimization
76.954
methods can also be
78.2573
170.00 used (particularly
80.00 82.50 85.00 87.50 90.00
when k > 2)
A: tim e

170
 Multiple Responses
◼ The example illustrated three response variables (yield,
viscosity and molecular weight)
◼ Multiple responses are common in practice
◼ Typically, we want to simultaneously optimize all
responses, or find a set of conditions where certain
product properties are achieved
◼ A simple approach is to model all responses and overlay
the contour plots

171
 Designs for Fitting Response
Surface Models
◼ For the first-order model, two-level factorials
(and fractional factorials) augmented with center
points are appropriate choices
◼ The central composite design is the most
widely used design for fitting the second-order
model
◼ Selection of a second-order design is an
interesting problem
◼ There are numerous excellent second-order
designs available
172
Part 7. TAGUCHI CONCEPT

173
 What is Robust Design
Robust design: a design whose performance is insensitive to variations.
Example: We want to pick x to maximize F

Simply doing a trade study to optimize the value of F

F would lead the designer to pick this point

What if I pick this

point instead?
This means that
values of F as
low as this can
be expected!

174
 What is Robust Design
◼ The robust design process is frequently formalized
through “six-sigma” approaches (or lean/kaizen
approaches)
◼ Six Sigma is a business improvement methodology
developed at Motorola in 1986 aimed at defect reduction
in manufacturing.
◼ Numerous aerospace organizations that have
implemented these systems, including:
 Department of Defense
 NASA
 Boeing
 Northrop Grumman 175
 Taguchi Method for Robust Design

◼ Systemized statistical approach to product and process

improvement developed by Dr. G. Taguchi
◼ Approach emphasizes moving quality upstream to the
design phase
◼ Based on the notion that minimizing variation is the
primary means of improving quality
◼ Special attention is given to designing systems such that
their performance is insensitive to environmental
changes

176
The Basic Idea Behind Robust Design
ROBUSTNESS ≡ QUALITY

Reduce
Variability

Increase Reduce
Quality Cost

177
 Any Deviation is Bad: Loss Functions
The traditional view states that there is no In Robust Design, any deviation from the
loss in quality (and therefore value) as target performance is considered a loss in
long as the product performance is within quality → the goal is to minimize this loss.
some tolerance of the target value.

Loss = k(x-xT)2

No
Loss Loss
Loss

xLSL xT xUSL x xLSL xT xUSL x

xT = Target Value xLSL = Lower Specification Limit xUSL = Upper Specification Limit178
 Overview of Taguchi Parameter Design Method

1. Brainstorming

Design Parameters: Variables under your

2. Identify Design Parameters control
and Noise Factors
Noise Factors: Variables you cannot control or
variables that are too expensive
3. Construct Design of to control
Experiments (DOEs)
Ideally, you would like to investigate all
possible combinations of design parameters
4. Perform Experiments and noise factors and then pick the best
design parameters. Unfortunately, cost and
schedule constraints frequently prevent us
from performing this many test cases – this
is where DOEs come in!
5. Analyze Results
179
 Design of Experiments (DOE)
Design of Experiments: An information gathering exercise. DOE is a
structured method for determining the relationship between process inputs
and process outputs.
L9(34) Orthogonal Array
Here, our objective is to intelligently choose the information
we gather so that we can determine the relationship Exp. Variables
between the inputs and outputs with the least amount of
effort
Num X1 X2 X3 X4
1 1 1 1 1
L4(23) Orthogonal Array
2 1 2 2 2
Number of Exp. Variables
Number of 3 1 3 3 3
Variable Levels Variables Num X1 X2 X3
4 2 1 2 3
L4(23) 1 1 1 1
5 2 2 3 1
Number of
2 1 2 2
6 2 3 1 2
Experiments 3 2 1 2
7 3 1 3 2
4 2 2 1
8 3 2 1 3
Num of Experiments must be ≥ system degrees-of-freedom:
DOF = 1 + (# variables)*(# of levels – 1) 9 3 3 2
180180 1
 N3 1 2 2 1
Inner & Outer Arrays N2 1 2 1 2
N1 1 1 2 2
1 2 3 4
Design Parameters Noise X1 X2 X3 X4
Experiment Num
1 1 1 1 1
2 1 2 2 2 y11 = f {X1(1), X2(1),
X3(1), X4(1),
Experiment Number

Performance 3 1 3 3 3 N1(1), N2(1), N3(1)}

Characteristic
evaluated at the 4 2 1 2 3
specified design
parameter and 5 2 2 3 1
noise factor values
6 2 3 1 2 y52 = f {X1(2), X2(2),
X3(3), X4(1),
7 3 1 3 2 N1(1), N2(2), N3(2)}
8 3 2 1 3
9 3 3 2 1
Inner Array – design parameter matrix
Outer Array – noise factor matrix
181
 Processing the Results (1 of 2)

Noise
Design Parameters Experiment Num

Compute signal-to-noise (S/N) for

each row
Experiment Number

1 n 2
Performance

S / N i = −10 log  yij 

Characteristic Minimizing performance
evaluated at the
specified design characteristic  n j =1 
parameter and
noise factor values

Maximizing performance 1 n 1 
S / N i = −10 log  2 
characteristic  n j =1 y 
 ij 

Inner Array – design parameter matrix

Outer Array – noise factor matrix
182
 Processing the Results (2 of 2)
Isolate the instances of each design parameter at each level and average the corresponding
S/N values.
Design Parameters X X X X
1 2 3 4
1 1 1 1 1 S/N
1
Signal-to-Noise (S/N)
Experiment Number

2 1 2 2 2 S/N X2 is at level 1 in
2
experiments 1, 4, & 7
3 1 3 3 3 S/N
3
4 2 1 2 3 S/N
4
5 2 2 3 1 S/N
S / N1 + S / N 4 + S / N 7
Avg5 S / N T1 (1) =
6 2 3 1 2 S/N 3
6 183
Visualizing the Results
Plot average S/N for each design parameter

ALWAYS aim to maximize S/N

In this example, these are the best cases.

184
 Pareto Plots and the 80/20 Rule
20% of the variables in any given system control 80% of the variability in
the dependent variable (in this case, the performance characteristic).
Individual design parameter effects
Cumulative effect
0% 20% 40% 60% 80% 100%

X1
X2
X3
X4
X5
X6
20% of the variables
X7
X8
X9
X10

80% of the variability in

the dependent variable
185
 Conclusions
◼ Decisions made early in the design process cost very
little in terms of the overall product cost but have a major
effect on the cost of the product
◼ Quality cannot be built into a product unless it is
designed into it in the beginning
◼ Robust design methodologies provide a way for the
designer to develop a system that is (relatively)
insensitive variations

186
Extra Slides

187
Practical Use of Center Points

◼ Use current operating conditions as the

center point
◼ Check for “abnormal” conditions during the
time the experiment was conducted
◼ Check for time trends
◼ Use center points as the first few runs when
there is little or no information available about
the magnitude of error
◼ Center points and qualitative factors?

188
 Center Points and Qualitative Factors

A 23 factorial design with one qualitative factor and center points

189