You are on page 1of 45

EXPERIMENTAL

DESIGN

Instructor: Lê Ngọc Liễu

Email: lnlieu@hcmiu.edu.vn
Office: A1.706
Content

 Chapter 1: Introduction and Basics

 Chapter 2: Comparative Experiments

 Chapter 3: Factorial Design

 Chapter 4: Optimization

 Chapter 5: Modeling

 Chapter 6: Design of experimental flowchart


Chapter 3
Screening Experiments
Four general fields applying Design of Experiment

 Comparative

 Screening

 Optimization

 Modeling
Screening Experiments

Used when you want to see the effect of a whole range of factors so as to know
which one(s) are most important.

Screening experiments:

 Simple design: one factor at a time

 Full factorial design: 2k

 Half factorial design: 2k-1

 Plackett-Burnam design
Simple design

 Start with a configuration and vary one factor at a time

 Given k factors and the i-th factor having ni levels

 The required number of experiments

Example:
Simple design: one factor at a time
Example:

Investigate the effects  Factor 1: enzyme concentration: 0.01; 0.02; 0.03%


of enzymatic treatment
on fruit juice yield  Factor 2: treatment duration: 30, 60, 90, 120 min
 Factor 3: treatment temperature: 30, 60 oC

1 2 3 4 5 6 7
Enzyme concentration 0.02 0.02 0.01 0.03 0.03 0.03 0.03
Treatment duration 90 90 90 90 30 60 120
Treatment temperature 30 60 60 60 60 60 60

How to determine the number of factors and number of levels?


Simple design: one factor at a time
Common procedure

 Identify a key factor

 Make changes to the key factor while trying to hold all other factors constant.

 Evaluate the effect

 Alter the process (maybe)

 Repeat
Simple design: one factor at a time
Principle to be considered in determination of levels:

 Each numeric factor should generally have at least 3 levels

 There should exist a peak or plateau in the trend of change

 Statistical analysis should be applied to determine the peak/ plateau

 Control sample should be included


peak
Simple design: one factor at a time
90 90 a 90 a
80 80 b 80 b
70 70 c 70 c c
60 60 60

Yield (%)
Yield (%)

Yield (%)
50 50 50
40 40 40
30 30 30
20
20 20
10
10 10
0
0 0 30 60 90 120
30 60 30 60 90
Treatment duration (min)
Treatment duration (min) Treatment duration (min)

plateau
90 90 a
a 80 b
80 a
70 b 70 c c
60 60

Yield (%)
d
Yield (%)

50 50
40 40
30 30
20 20
10 10
0 0
30 60 90 0 30 60 90 120
Treatment duration (min) Treatment duration (min)
Simple design: one factor at a time
Disadvantages

 Require more runs for the same precision in effect estimation

 Cannot estimate interactions

 Can miss optimal settings of factors

1 2 3 4 5 6 7
Enzyme concentration 0.02 0.02 0.01 0.03 0.03 0.03 0.03
Treatment duration 90 90 90 90 30 60 120
Treatment temperature 30 60 60 60 60 60 60
Full factorial design

 Use all possible combinations at all levels of all factors

 Given k factors and the i-th factor having ni levels

 The required number of experiments

Example:
Fractional factorial design

 When full factorial design results in a huge number of experiments, it


may be not possible to run all
 Use subsets of levels of factors and the possible combinations of these

 Given k factors and the i-th factor having ni levels; and selected subsets of
levels mi ≤ ni.

 The required number of experiments

Example:
Comparation of experiment runs
Main effect and interaction

 The main effect of a factor is defined to be the change in response


produced by a change in the level of a factor.

 In some experiments we may find that the difference in response


between the levels of one factor is not the same at all levels of the other
factor. When this occurs, there is an interaction between the factors.
Main effect and interaction

 Graphs are frequently useful in interpreting significant interactions.

 When an interaction is large, the main effects have little practical


meaning.

 A significant interaction will often mask the significance of main effects.


Advantages of Factorials

 They are more efficient than one-factor-at-a-time experiments.

 A factorial design is necessary when interactions may be present to


avoid misleading conclusions.

 Factorial designs allow the effects of a factor to be estimated at several


levels of the other factors, yielding conclusions that are valid over a
range of experimental conditions.
Full factorial design

 Easy design to create. Each factor is tested at each condition of the factor

 Popcorn microwave making example:


2 brands x 6 power settings x 3 time settings = 36 trials

 Results analyzed with a linear model

 Cost: resources, time, materials. What if you can’t afford 36 trials? What if
you only have 16 bags of popcorn?
2k Factorial Designs

 Test each of k factors at 2 levels

 Begin by experimenting at the minimum and maximum level of a factor


 two levels
 A total of 2k experiments are required
 Advantages • Less experiment runs

• Help to identify important factors (screening)

 Disadvantages • Cannot identify nonlinear effects

• Follow-up with additional sampling on important factors


(augment design)
Standard 23 Sign Table

Randomization Run A B C Response


1 - - -
2 - - +
3 - + -
4 - + +
5 + - -
6 + - +
7 + + -
8 + + +
Standard 23 Sign Table (with result)

Run Temp (oC) Time (min) Ingredient Yield (%)


1 200 1 1 65.3
2 200 1 2 81.3
3 200 2 1 53.3
4 200 2 2 69.9
5 250 1 1 61.8
6 250 1 2 77.4
7 250 2 1 73.9
8 250 2 2 89.9
2k Factorial Designs
Principle to be considered

 Be aggressive in setting factor levels, but not exceed the extreme

 Always randomize run order


2k Factorial Designs
Example: popcorn microwave making

Be aggressive in setting factor levels, but not exceed the extreme (burn
popcorn)
2k Factorial Designs
Example: popcorn microwave making

Always randomize run order


2k-p Fractional Factorial Designs

 To do a full 2k factorial experiment, you must do at least 2k runs

• Testing 4 factors costs twice as much as testing 3 factors

 When the number of factors is large, a full factorial design requires a


large number of experiments

 In that case fractional factorial design can be used

• Require fewer experiments, e.g., 2k-1 requires half of the experiments as


a full factorial design
2k-p Fractional Factorial Designs

 Condition combinations are chosen to provide sufficient information to


determine the factor effect

 Key point, what subset of factor combinations?

• Pick the wrong set and you’re likely to miss interactions

• DoE software can help


2k-p Fractional Factorial Designs

 Confounding
• with fractional factorial design some of the effects can not be determined

• only combined effects of several factors can be computed

 A fractional factorial design is not unique

 Design resolution

• The resolution of a design is measured by the order of effects that are


aliased (or confounded)

• A design of higher resolution is considered a better design, but require


more runs
Plackett-Burman design

 This design screens out unimportant factors (noise), which means that
you avoid collecting large amounts of data on relatively unimportant
factors.

 Number of factor

• Only be used for experiments that are multiples of 4 with 8 as the


starting point (N = 8, 12, 16, 20, 24, 28, 32, 36)

• A minimum of 4n experiments is needed for estimating main effects


for 4n-1 factors

For example, 4, 5, 6, or 7 factors would require at least 8 experimental


runs, 8, 9, 10, or 11 would require at least 12 runs, and so on.
Plackett-Burman design

Advantage

 Sort out which factors to concentrate on, greatly reducing the amount of
data you have to collect.

• For example, with 15 factors in your design, 20 data points needed in


a Plackett-Burman. A full factorial design would require over a
thousand times that amount (32,768 data points).
Plackett-Burman design

It is particularly helpful to use Plackett-Burman design:

 In screening
 When neglecting higher order interactions is possible
 In two-level multi-factor experiments
 When there are more than four factors (if there are between two to four
variables, a full factorial can be performed)
 To economically detect large main effects
 For N = 12, 20, 24, 28 and 36 (where N = the number of experiments)
Plackett-Burman design

Disadvantage

 They do not verify if the effect of one factor depends on another factor.
 If you run the smallest design you can, it does not follow that enough data
has been collected to know what those effects are precisely.

Plackett-Burman should be used as a starting point for further


experiments. Once the important factors have been identified, a
full factorial or fractional design should be run to study those
factors more.
2k-1 Fractional Factorial
Design Example

Case study 3.1


Chang, Siu Hua, Tjoon Tow Teng, and Norli Ismail.
"Screening of factors influencing Cu (II) extraction
by soybean oil-based organic solvents using
fractional factorial design." Journal of environmental
management 92.10 (2011): 2580-2585
Analysis procedure for a factorial design
 Estimate factor effects
 Statistical testing (ANOVA) + Formulate model
 Refine the model
 Analyze residuals (graphical)
 Interpret results
Estimate factor effects
Estimate factor effects
Estimate factor effects
Statistical testing (ANOVA) + Formulate model
Regression analysis

High-order interactions are neglected

Regression model for %E (y)


Refine model

Reduced model for %E (in coded units)

Reduced model for %E in uncoded units

%E = -239.4 + 2.823 B + 74.03 E - 0.6160 BE


Analyze residuals (graphical)
Total number of experiments: 25 = 32

Full factorial Table 1: Five Factor Analyses with Two Levels (Full Factorial Design)
A B C D E Y
Total number of experiments: 12
vs. -1
1
-1
-1
-1
-1
-1
-1
-1
-1
75.87
76.14
-1 1 -1 -1 -1 109.95
Plackett- 1 1 -1 -1 -1 109.55
Table 2: Five Factor Analysis with Two Levels (Plackett-Burman Design)
A B C D E Y
-1 -1 1 -1 -1 80.17
Burman 1
-1
-1
1
1
1
-1
-1
-1
-1
80.3
114.07
1 -1 1
-1
-1 -1
-1
75.67
1 1 1 102.4
1 1 1 -1 -1 114.05
-1 -1 -1 1 -1 68.77 -1 1 1 -1 1 113.71
1 -1 -1 1 -1 68.63 1 -1 1 1 -1 72.95
5 factors, 2 levels -1 1 -1 1 -1 102.41
1 1 -1 1 1 102.32
1 1 -1 1 -1 102.27
-1 -1 1 1 -1 72.95 1 1 1 -1 1 113.85
1 -1 1 1 -1 72.68 -1 1 1 1 -1 106.51
-1 1 1 1 -1 106.98
-1 -1 1 1 1 72.66
1 1 1 1 -1 106.65
-1 -1 -1 -1 1 75.79 -1 -1 -1 1 1 68.62
1 -1 -1 -1 1 75.71 1 -1 -1 -1 1 75.54
-1 1 -1 -1 1 110.08
1 1 -1 -1 1 110.04 -1 1 -1 -1 -1 109.28
-1 -1 1 -1 1 80.49 -1 -1 -1 -1 -1 75.69
1 -1 1 -1 1 80.86
-1 1 1 -1 1 113.49
1 1 1 -1 1 114.06
-1 -1 -1 1 1 68.22
1 -1 -1 1 1 68.6
-1 1 -1 1 1 102.5
1 1 -1 1 1 102.3
-1 -1 1 1 1 73.03
1 -1 1 1 1 73.3
-1 1 1 1 1 106.08
1 1 1 1 1 106.7
Full factorial vs. Plackett-Burman

High-order interactions: only consider second-order


Full factorial vs. Plackett-Burman
Full factorial vs. Plackett-Burman

Comparison of Results
Full Factorial Design Plackett-Burman Design
Number of experiments 32 12
Significant factors B, C, D B, C, D
Significant levels B: +1, C: +1, D: -1 B: +1, C: +1, D: -1
Significant interactions (B, C) and (A, E)* Non-detectable
Optimized value of Y 114.13 113.66
*If considered at p < 0.1
Half fractional factorial design – extra example
Half factorial design was used to screen the effect of 5 factors on the response to obtain the following table. Use
Minitab to compute the regression equation of the response. Provide the effect order (Pareto chart) of all factors
and their interaction.

StdOrder RunOrder CenterPt Blocks A B C D E Response


7 1 1 1 2 80 80 30 100 43.02
5 2 1 1 2 20 80 30 10 5.62
13 3 1 1 2 20 80 300 100 31.88
2 4 1 1 6 20 30 30 10 16.7
17 5 0 1 4 50 55 165 55 32.38
3 6 1 1 2 80 30 30 10 18.15
16 7 1 1 6 80 80 300 100 48.88
10 8 1 1 6 20 30 300 100 29.02
15 9 1 1 2 80 80 300 10 9.4
19 10 0 1 4 50 55 165 55 31.52
12 11 1 1 6 80 30 300 10 17.29
14 12 1 1 6 20 80 300 10 13.52
8 13 1 1 6 80 80 30 10 14.4
18 14 0 1 4 50 55 165 55 33.33
11 15 1 1 2 80 30 300 100 41.73
6 16 1 1 6 20 80 30 100 38.73
4 17 1 1 6 80 30 30 100 24.59
9 18 1 1 2 20 30 300 10 17.28
1 19 1 1 2 20 30 30 100 16.3
Half fractional factorial design – extra example

You might also like