See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/221913304

General Introduction to Design of Experiments (DOE)

Chapter · July 2011

DOI: 10.5772/23878 · Source: InTech

CITATIONS READS
6 12,561

1 author:

Ahmed Ibrahim Badr Eldin

Sigma Pharmaceutical Corp. - Egypt
29 PUBLICATIONS   129 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

PhD thesis on green analytical chemistry View project

Development and characterization of a novel liposomal formulation for oral delivery of glucosamine and chondroitin sulphates View
project

All content following this page was uploaded by Ahmed Ibrahim Badr Eldin on 02 June 2014.

The user has requested enhancement of the downloaded file.

 2

General Introduction to
Design of Experiments (DOE)
Ahmed Badr Eldin
Sigma Pharmaceutical Corp.,
Egypt

1. Introduction
Experimental design and optimization are tools that are used to systematically examine
different types of problems that arise within, e.g., research, development and production. It
is obvious that if experiments are performed randomly the result obtained will also be
random. Therefore, it is a necessity to plan the experiments in such a way that the
interesting information will be obtained.

2. Terminology
Experimental domain: the experimental ‘area’ that is investigated (defined by the variation of
the experimental variables).
Factors: experimental variables that can be changed independently of each other
Independent Variables: same as factors
Continuous Variables: independent variables that can be changed continuously
Discrete Variables: independent variables that are changed step-wise, e.g., type of solvent.
Responses: the measured value of the result(s). from experiments
Residual: the difference between the calculated and the experimental result

3. Empirical models
It is reasonable to assume that the outcome of an experiment is dependent on the
experimental conditions. This means that the result can be described as a function based on
the experimental variables[2],
Y= (f) x. The function (f) x. is approximated by a polynomial function and represents a good
description of the relationship between the experimental variables and the responses within
a limited experimental domain. Three types of polynomial models will be discussed and
exemplified with two variables, x1 and x2.
The simplest polynomial model contains only linear terms and describes only the linear
relationship between the experimental variables and the responses. In a linear model, the two
variables x1 and x2 are expressed as:

y = b0 + b1x1 + b2 x2 + residual.
22 Wide Spectra of Quality Control

The next level of polynomial models contains additional terms that describe the interaction
between different experimental variables. Thus, a second order interaction model contains the
following terms:

y = b0 + b1x1 + b2 x2 + b12 x1x2 + residual.

The two models above are mainly used to investigate the experimental system, i.e., with
screening studies, robustness tests or similar.
To be able to determine an optimum (maximum or minimum). quadratic terms have to be
introduced in the model. By introducing these terms in the model, it is possible to determine
non-linear relationships between the experimental variables and responses. The polynomial
function below describes a quadratic model with two variables:

y = b0 + b1x1 + b2 x2 + b11x12 + b22 x22 + b12 x1x2 + residual.

The polynomial functions described above contain a number of unknown parameters

(b0 , b1 , b2 , etc.) that are to be determined. For the different models different types of
experimental designs are needed.

4. Screening experiments
In any experimental procedure, several experimental variables or factors may influence the
result. A screening experiment is performed in order to determine the experimental
variables and interactions that have significant influence on the result, measured in one or
several responses.[3]

5. Factorial design[4]
In a factorial design the influences of all experimental variables, factors, and interaction
effects on the response or responses are investigated. If the combinations of k factors are
investigated at two levels, a factorial design will consist of 2k experiments. In Table 1, the
factorial designs for 2, 3 and 4 experimental variables are
shown. To continue the example with higher numbers, six variables would give 26 = 64
experiments, seven variables would render 27 = 128 experiments, etc. The levels of the
factors are given by – (minus) for low level and + (plus) for high level. A zero-level is also
included, a centre, in which all variables are set at their mid
value. Three or four centre experiments should always be included in factorial designs, for
the following reasons:
• The risk of missing non-linear relationships in the middle of the intervals is minimised,
and
• Repetition allows for determination of confidence intervals.
What - and + should correspond to for each variable is defined from what is assumed to
be a reasonable variation to investigate. In this way the size of the experimental domain
has been settled. For two and three variables the experimental domain and design can be
illustrated in a simple way. For two variables the experiments will describe the corners in
a quadrate (Fig. 1), while in a design with three variables they are the corners in a cube
(Fig. 2).
General Introduction to Design of Experiments (DOE) 23

Table 1. Factorial designs

−+ ++
X2

−− +−

X1
Fig. 1. The experiment in a design with two variables

6. Signs of interaction effects[5]

The sign for the interaction effect between variable 1 and variable 2 is defined as the sign for
the product of variable 1 and variable 2 (Table 2). The signs are obtained according to
normal multiplication rules. By using these rules it is possible to construct sign columns for
all the interactions in factorial designs.
Example 1: A ‘work-through’ example with three variables
This example illustrates how the sign tables are used to calculate the main effects and the
interaction effects from a factorial design. The example is from an investigation of the
influence from three experimental variables.
24 Wide Spectra of Quality Control

−++ +++
X3
−−+ +−+

−+− ++−
X2
−−− +−−

X1
Fig. 2. The experiment in a design with three variables

7. Fractional factorial design

To investigate the effects of k variables in a full factorial design, 2k experiments are needed.
Then, the main effects as well as all interaction effects can be estimated. To investigate seven
experimental variables, 128 experiment will be needed; for 10 variables, 1024 experiments
have to be performed; with 15 variables, 32,768
experiments will be necessary. It is obvious that the limit for the number of experiments it is
possible to perform will easily be exceeded, when the number of variables increases. In most
investigations it is reasonable to assume that the influence of the interactions of third order
or higher are very small or negligible and can then be excluded from the polynomial model.
This means that 128 experiments
are too many to estimate the mean value, seven main effects and 21 second order interaction
effects, all together 29 parameters. To achieve this, exactly 29 experiments are enough. On
the following pages it is shown how the fractions (1/2, 1/4, 1/8, 1/16 . . . 1/2 p) of a
factorial design with 2 k-p experiments are defined, where
k is the number of variables and p the size of the fraction. The size of the fraction will
influence the possible number of effects to estimate and, of course, the number of
experiments needed. If only the main effects are to be determined it is sufficient to perform
only 4 experiments to investigate 3 variables, 8 experiments for 7 variables, 16 experiments
for 15 variables, etc. This corresponds to the following
response function:

v = β n + ∑ β i xi + ε

It is always possible to add experiments in order to separate and estimate interaction effects,
if it is reasonable to assume that they influence the result. This corresponds to the following
second order response function:

y = β 0 + ∑ β i xi + ∑∑ β ij xi x j + ε

In most cases, it is not necessary to investigate the interactions between all of the variables
included from the beginning. In the first screening it is recommended to evaluate the result
General Introduction to Design of Experiments (DOE) 25

and estimate the main effects according to a linear model (if it is possible to calculate
additional effects they should of course be estimated as well.).
After this evaluation the variables that have the largest influence on the result are selected
for new studies. Thus, a large number of experimental variables can be investigated without
having to increase the number of experiments to the extreme.

8. Optimization
In this part, two different strategies for optimization will be introduced; simplex
optimization and response surface methodology. An exact optimum can only be determined
by response surface methodology, while the simplex method will encircle the optimum.
simplex is a geometric figure with (k+1) corners where k is equal to the number of variables
in a k-dimensional experimental domain. When the number of variables is equal to two the
simplex is a triangle (Fig. 16.).

Var. 2

3
1

Var. 1
Fig. 3. A simplex in two variables
Simplex optimization is a stepwise strategy. This means that the experiments are performed
one by one. The exception is the starting simplex in which all experiments can be run in
parallel. The principles for a simplex optimization are illustrated in Fig. 17. To maximize the
yield in a chemical synthesis, for example, the first step is to run k+1 experiments to obtain
the starting simplex. The yield in each corner of the simplex is analyzed and the corner
showing the least desirable result is mirrored through the geometrical midpoint of the other
corners. In this way, a new simplex is obtained. The co-ordinates (i.e., the experimental
settings) for the new corner are calculated and the experiment is performed. When the yield
is determined,
the worst of the three corners is mirrored in the same way as earlier and another new
simplex is obtained, etc. In this way, the optimization continues until the simplex has
rotated and the optimum is encircled. A fully rotated simplex can be used to calculate a
response surface. The type of design described by a rotated simplex is called a Doehlert
design.
 26 Wide Spectra of Quality Control

Var. 2

12

13 11
9
7
5 10
3 8
6
1
4
2

Var. 1

Fig. 4. Illustration of a simplex optimization with two variables

9. Rules for a simplex optimization

With k variables k+1 experiments are performed with the variable settings determined by
the co-ordinates in the simplex. For two variables the simplex forms a triangle. For three
variables it is recommended to use a 2 3-1 fractional factorial design as a start simplex.

10. References
[1] Experimental design and optimization, Chemometrics and Intelligent Laboratory
Systems 42 _1998. 3–40
[2] R. Sundberg, Interpretation of unreplicated two-level factorial experiments, Chemometrics
and intelligent laboratory system, 24 _1994. 1–17.
[3] Atkinson, A. C. and Donev, A. N. Optimum Experimental Designs Clarendon Press,
Oxford p.148.
[4] Kowalski, S.M., Cornell, J.A., and Vining, G.G. (2002) “Split Plot Designs and Estimation
Methods for Mixture Experiments with Process Variables,” Technometrics 44: 72-
79.
[5] Goos, P. (2002) The Optimal Design of Blocked and Split-Plot Experiments, New York:
Springer

View publication stats