ARTICLE IN PRESS
Ocean Engineering 36 (2009) 237–247
Contents lists available at ScienceDirect
Ocean Engineering
journal homepage: www.elsevier.com/locate/oceaneng
Combined use of dimensional analysis and modern experimental design
methodologies in hydrodynamics experiments
Mohammed F. Islam , L.M. Lye
Faculty of Engineering and Applied Science, Memorial University of Newfoundland, St. John’s, Newfoundland, Canada A1B 3X5
a r t i c l e in fo
abstract
Article history:
Received 23 June 2008
Accepted 18 November 2008
Available online 6 December 2008
In this paper, a combined use of dimensional analysis (DA) and modern statistical design of experiment
(DOE) methodologies is proposed for a hydrodynamics experiment where there are a large number of
variables. While DA is well-known, DOE is still unfamiliar to most ocean engineers although it has been
shown to be useful in many engineering and non-engineering applications. To introduce and illustrate
the method, a study concerning the thrust of a propeller is considered. Fourteen variables are involved
in the problem and after dimensional analysis this reduces to 11 dimensionless parameters. Then, a
two-level fractional factorial design was used to screen out parameters that do not significantly
contribute to explaining the dependent dimensionless parameter. With the remaining five statistically
significant dimensionless parameters, various response surface methodologies (RSM) were used to
obtain a functional relationship between the dependent dimensionless thrust coefficient, and the five
dimensionless parameters. The final model was found to be of reasonable accuracy when tested against
results not used to develop the model. The methodologies presented in the paper can be similarly
applied to systems with a large number of control variables to systematically derive approximate
mathematical models to predict the responses of the system economically and accurately.
& 2008 Elsevier Ltd. All rights reserved.
Keywords:
Dimensional analysis
Design of experiments
Fractional factorial design
Response surface methodology
Space-filling designs
Propellers
Propulsive performance
1. Introduction
Variables used in engineering are usually expressed in terms of
a limited number of basic dimensions namely mass, length, time,
and sometimes temperature. For certain phenomenon, a large
number of variables may be needed to describe or explain the
phenomenon. However, by the methods of dimensional or partial
analysis, the separate variables involved in the problem can be
reduced to a smaller set of independent dimensionless groups or
dimensionless parameters. Dimensional analysis (DA) has been
used in engineering particularly, fluid mechanics and hydraulics,
for about a hundred years. Many methods have been developed
for dimensional analysis and these are normally covered as part of
an undergraduate course in fluid mechanics or hydraulics.
Dimensional analysis is also called partial analysis because the
problem is only partially solved when the variables have been
suitably combined into independent dimensionless parameters.
To obtain a functional relationship between the parameters, a
series of experiments must be conducted to obtain data to relate
the dependent dimensionless parameter to the other dimensionless parameters. When there are several independent dimensional
parameters, a series of experiments coupled with statistical
analysis is usually required. How to conduct the experiment such
Corresponding author. Tel.: +1709 743 5627.
E-mail addresses: islam@engr.mun.ca (M.F. Islam), llye@mun.ca (L.M. Lye).
0029-8018/$ - see front matter & 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.oceaneng.2008.11.004
that experimental runs are minimized and yet obtain meaningful
results is often a challenge.
In this paper, a combined use of dimensional analysis and
modern statistical design of experiments (DOE) methodologies is
proposed for a hydrodynamics experiment where there are a large
number of variables. While DA is well-known, DOE is still unfamiliar
to most ocean engineers although it has been shown to be useful in
many engineering and non-engineering areas. To illustrate the
combined method, a study concerning the thrust of a propeller is
used as a case study. Fourteen potential variables are involved in this
problem. The methodology presented in the paper can potentially be
applied to most system with a large number of variables to derive
approximate mathematical models to predict the responses of the
system economically and accurately. In the following section,
dimensional analysis will be briefly introduced and performed for
the problem at hand followed by a description of statistical design of
experiment methodologies. The application of a fractional factorial
design followed by applications of various response surface designs
to the problem at hand will then be presented. The response surface
models developed are then validated against data not used to
develop the models. Finally, conclusions will follow.
2. Dimensional analysis
There are many methods of dimensional analysis and practically every book in fluid mechanics has a chapter or two devoted
ARTICLE IN PRESS
238
M.F. Islam, L.M. Lye / Ocean Engineering 36 (2009) 237–247
to the topic. For example, Munson et al. (1994) provided a list of
about 15 books on the subject. More recently, a 970-page book on
the subject was written by Szirtes (1997). Popular methods of
dimensional analysis include Rayleigh’s method, Buckingham P
theorem, matrix method, and method of synthesis. All these
methods are described in Sharp (1981), among others.
Dimensional analysis is essentially a means of utilizing partial
knowledge of a problem when the details are too obscure to
permit an exact analysis. For example, the thrust producing
phenomenon of a propeller is very complicated and is dependent
upon many parameters that it is quite difficult to derive a
complete analytical relationship expressing it in terms of the
variables of control. For the case study under consideration, the
objective is to derive a functional relationship between propulsive
performance parameter of a propeller, expressed as the thrust
coefficient, KT, and the variables upon which this parameter
depends. A detailed derivation and formulation of the problem
under consideration is as follows:
2.1. Physical knowledge of the phenomenon
The thrust generated by the propeller is dependent on the
propeller’s operating environment, the geometrical characteristics
of the propeller, and the relative motion between water and
propeller. Based on experience, the dependent variable of the
phenomenon may be described in terms of relevant fluid properties, blade and pod geometry, and operating conditions as given in
(1) (assuming that the blade sectional shapes are fixed).
T ¼ f ðD; N; n; V A ; g; m; rw ; p; P 0:7R ; c; ys ; yr ; Dh ; aÞ.
(1)
The variables that are involved in this phenomenon are
summarized in Table 1. There are some terms having length or
angle dimensions, which can be excluded from the equation and
can be added to the final non-dimensional equation directly. Thus
simplifying the above equation results in
f ðT; n; V A ; g; m; p; rw ; DÞ ¼ 0.
(2)
2.2. Application of Barr’s method of dimensional analysis
To develop the non-dimensional equations for each of the
variables of interest, a dimensional analysis method, ‘‘the method
of synthesis’’ developed by Barr (1969) was used. The details of
Barr’s method are also given in Sharp (1981). According to Barr’s
Table 1
List of the variables affecting propulsive performance of a propeller.
Symbol
Dimensions
Size of the propeller (represented by diameter)
Number of blades
Rotational speed (RPM)
Speed of advance
Gravitational acceleration
Dynamic viscosity
Mass density of water
Pressure of the fluid
Pitch at 0.7R (R is the radius of the propeller)
Mean chord length
Mean skew
Mean rake
Hub diameter
Hub taper angle
D
N
n
VA
g
a
L
–
1/T
L/T
L/T2
M/(LT)
M/L3
M/(LT2)
L
L
Deg.
Deg.
L
Deg.
Response variables
Propeller thrust
Symbol
T
Dimensions
ML2/T2
p
P0.7R
c
ys
yr
Dh
T 1=4
TD T nD2 n2 D2 nrw D3
;
;
;
;
;
V A rw nD g rw D mn pD V A
g
m
r
!
n2 rw D3
m V 2A V 2A rw D m2=3 g rw D2 mnD
;
;
D
¼ 0,
;
;
;
;
;
2=3
p
p
rw V A g
p
p
g 1=3 rw
1=4 1=2
w n
;
T
;
T
2
;
(3)
where, from the alternative choice of 16 parameters, any five can
be selected, provided T, n, VA, g, m and r are each included at least
once.
2.3. Choose the right coefficients
One possible conventional solution leads to Eq. (4).
!
T 1=4
nD2
m V 2A V 2A rw D
;
;
;
;
;
D
¼ 0.
1=2 V A rw V A g
p
r1=4
w n
(4)
A further choice is now available in forming the non-dimensional equation. The one that leads to a meaningful functional
relationship is
!
T
nD
m
V 2A V 2A rw
;
;
;
;
¼ 0.
(5)
rw n2 D4 V A Drw V A gD p
Another method such as Buckingham’s P theorem could have
been used to obtain the dimensionless terms in Eq. (5).
2.4. Rearrangement of the dimensionless parameters
Rearranging the above dimensionless parameters and adding
the linear terms that were excluded in non-dimensional form
results in the dimensionless Eq. (6) for the thrust of a propeller
operating in open water.
!
T
V A rw DV A V 2A
p
P0:7R c
Dh
;
;
; ; ys ; yr ;
; a; N .
¼f
;
;
(6)
nD
m
gD rw V 2A D D
D
rw n2 D4
In the above equation there are several standard numbers
and commonly used
pffiffiffiffiffiffipropeller related terms. These are Froude
number, F n ¼ V A = gD; Reynolds number, Rn ¼ rwDVA/m or DVA/n,
where n is the kinematic viscosity of the fluid; propeller advance
coefficient, J ¼ VA/nD; pitch ratio ¼ P0.7R/D; skew ¼ ys; rake ¼ yr;
and thrust coefficient, KT ¼ T/rwn2D4.
2.5. Further study of the dimensionless parameters
Variable of control
m
rw
method, linear proportionalities are formed by combining the first
six variables in pairs. Introducing rw as required to eliminate mass
dimensions and using the reference tables provided in Barr
(1969), Eq. (3) was obtained.
Eq. (6) states in effect that if all the parameters on the righthand side have the same values for two geometrically similar but
different sized propellers, the flow patterns will be similar and the
value of T/rwn2D4 will be the same for each. When one is testing
model propellers to predict the performance of a prototype
propeller it is necessary to fulfill a number of conditions (laws of
similitude) to ensure similarity between the full-scale and the
model scale results. Here, it should be noted that the response
now depends on 11 dimensionless parameters instead of 14 that
existed before applying partial analysis to the system. In general, if
there are n quantities and m fundamental dimensions (e.g., M, L,
and T), there will be (n–m) dimensionless terms. However, to
develop a functional relationship with 11 independent dimensionless parameters would require an extensive systematic
experimentation program to generate the data for statistical
analysis and modelling. As such, it may be prudent to further
reduce the dimensionality of the problem. That is, use only those
ARTICLE IN PRESS
239
M.F. Islam, L.M. Lye / Ocean Engineering 36 (2009) 237–247
parameters that will significantly contribute to explaining the
variation in the dependent parameter. An efficient technique for
screening a large number of parameters based on modern
experimental design methodologies will be discussed next.
3. Statistical design of experiments
Before describing the experimental design and the subsequent
analysis to screen out the statistically insignificant parameters for
the problem at hand, a brief note about the statistical design of
experiments or DOE is warranted.
Engineers in general carry out a fair amount of physical
experimentation in the laboratory and on the computer using a
variety of numerical models. Experiments are carried out to (1)
evaluate and compare basic design configurations, (2) evaluate
material alternatives, (3) select design parameters so that the
design will work well under a wide variety of field conditions
(robust design), and (4) determine the key design parameters that
impact performance. As with most engineering problems, time
and budget are often limited. Hence it is necessary to gain as
much information as possible and do so as efficiently as possible
from an experimental program.
In engineering, one often-used approach is the best-guess
(with engineering judgment) approach. Another strategy of
experimentation that is prevalent in practice is the one-factorat-a-time or OFAT approach. The OFAT method was considered the
standard, systematic, and accepted method of scientific experimentation. Both of these methods have been shown to be
inefficient and in fact can be disastrous (Montgomery, 2005).
These methods of experimentation became outdated in the early
1920s when Ronald A. Fisher discovered much more efficient
methods of experimentation based on factorial designs. This class
of experimental designs includes the general factorial, two-level
factorial, fractional factorial, and response surface designs among
others. These statistically based experimental design methods are
now simply called design of experiment methods or DOE
methods. A recent application of DOE methods in ocean
engineering can be found in Hawkins and Lye (2006), among
others.
Basically, DOE is a methodology for systematically applying
statistics to experimentation. DOE lets experimenters develop a
mathematical model that predicts how input variables interact to
create output variables or responses in a process or system. DOE
can be used for a wide range of experiments for various purposes
including nearly all fields of engineering and science and even in
marketing studies. The use of statistics is important in DOE but
not absolutely necessary. In general, by using DOE, one can
learn about the process being investigated;
screen important factors;
determine whether factors interact;
build a mathematical model for prediction; and
optimize the response(s), if required.
DOE methods are also useful as a strategy for building
mechanistic models, and they have the additional advantage that
no complicated calculations are needed to analyze the data
produced from the designed experiment. It has now been
recognized that the factorial-based DOE is the correct and the
most efficient method of conducting multi-factored experiments;
they allow a large number of factors to be investigated in few
experimental runs. The efficiency stems from using settings of the
independent factors that are completely uncorrelated with each
other. That is, the experimental designs are orthogonal. The
consequence of the orthogonal design is that the main effect of
each experiment factor, and also the interactions between factors,
can be estimated independent of the other effects. As stated
earlier, many industries have recognized this fact and design of
experiment methodologies is a key component of the Six-Sigma
quality program used by many major corporations. Yet it is
surprising that after about 90 years since the invention of modern
experimental design it is still not widely taught in schools of
engineering or science in our universities (Box et al., 2006). The
wide variety of experimental designs and their statistical details
can be found in many excellent texts including Montgomery
(2005), Myers and Montgomery (2002), Box et al. (2005),
Ryan (2007), Antony (2006), Box et al. (2006) and Berger and
Maurer (2002), among others. In the next section, a two-level
quarter-fractional factorial or in short, a 2k2 factorial design will
be used to study the effect of the 11 factors (dimensionless
parameters) on the thrust coefficient, KT of a propeller.
4. Application of fractional factorial design
The following steps were followed to study the thrust
coefficient of the propeller using the factorial design method.
4.1. Statement of the problem
Evaluate how the geometric and motion parameters and their
interactions affect the thrust coefficient of a propeller in open
water conditions.
4.2. Choice of factors, levels, ranges and response variables
From Eq. (6), the terms rwDVA/m, VA2/gD, and p/rwVA2 can be
removed since they are extraneous in the sense that for constant g,
p and rw, these terms mainly depends on speed of advance VA.
Also it is difficult to take into account the viscosity m because it is
hard to control. The modified equations take the form of Eq. (7):
T
V A P0:7R c ys yr Dh
;
;
;
;
;
;
¼
f
a
;
N
.
(7)
nD D D D D D
rw n2 D4
In this reduced and simplified form there are still eight
independent factors to deal with (Table 1). Before proceeding
with the factorial design it is important that the response and the
factors’ ranges are properly defined. The factors (variables of
control) are given in Table 2 with the corresponding low and high
values to be considered in the two-level fractional factorial design.
The low and high values of the factors are chosen from a practical
viewpoint. It is to be noted that the resulting model for the
response is only valid within the ranges of the factors. Factors
involving angles are modelled as actual angles rather than ratios.
In traditional propeller experimental program design, variables
are usually divided into two separate groups: geometry and
motion parameters. Geometry parameters determine the number
of propeller models that need to be built (materials and
Table 2
The variables of control with their ranges (low and high values).
A
B
C
D
E
F
G
H
Factors
Low (1)
High (+1)
Number of blades, N
Chord-diameter ratio, c/D
Pitch-diameter ratio, p0.7R/D
Rake angle, yr
Skew angle, ys
Hub taper angle (HTA), a
Hub-diameter ratio, Dh/D
Propeller advance coeff., J
3
0.33
0.80
151
01
151
0.20
0.2
5
0.52
1.20
151
901
151
0.30
1.0
ARTICLE IN PRESS
240
M.F. Islam, L.M. Lye / Ocean Engineering 36 (2009) 237–247
fabrication cost) and motion parameters determine the required
number of runs (facility time and staff labor cost). It is possible to
deal with the parameters in the two groups separately using the
proposed method and then integrate these two to give an
integrated solution. However, in the current study, the two groups
of parameters were studied together. Furthermore, the advance
coefficient, J is the ‘‘primary’’ independent variable for the dependent variable, thrust coefficient, because propulsion performance
charts are plotted as curves of thrust coefficient versus advance
coefficient. One can argue that the advance coefficient should be
definitely included in the final analysis without needing any DOE
analysis. However, in the present study the advance coefficient
was included in the analysis to illustrate the method for any
hydrodynamic problem for completeness. The significance of the
advance coefficient will automatically be confirmed from the
subsequent analysis if it is indeed the most important parameter.
More importantly, how it will interact with other parameters will
also be quantified which cannot be done if it is analyzed
on its own.
4.3. Choice of experimental design
A fractional factorial design (FFD) was used to design the
experiments to minimize the runs. With eight factors, the quarterfractional two-level factorial design (282) requires a combination
of experimental 64 runs or calculation points. The 64 run
combinations for the 282 design and the responses are shown
in Table 3. The design is a Resolution V design, which means that
all main effects and two-factor interactions can be estimated
without ambiguity (Montgomery, 2005).
4.4. Perform the experiments
The experimental design requires that the responses at high
and low values of the control factors (independent variables)
be obtained at different combinations. This would require a
large number of model propellers and test runs to obtain the
responses for the numerous combinations for all the control
factors. In the current study, a computer simulation software
package, PROPELLA, developed by Liu (2003–2008) was used to
obtain the responses for the run combinations. PROPELLA is a time
domain panel method computer code, which can calculate the
propeller thrust, KT for all the combinations of experimental runs.
The geometry part of the code can model the propeller with any
practical geometry and the code has been validated against actual
measurements with good agreement. All the calculations were
done with a propeller rpm of 600, zero shaft angles and at a shaft
depth of 2 diameters. Details of the geometry of the propellers are
presented by Liu (2006). The predictions given by PROPELLA might
or might not be as accurate as real physical measurements, but are
sufficient to establish a standard procedure of implementing
design of experiments techniques in a quantitative analysis of
propulsive performance. Fig. 1 shows a few model propellers
simulated using PROPELLA.
Table 3
FFD data sheet.
Std. order
N
c/D
P0.7R/D
yr
ys
a
Dh/D
J
KT
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
3
5
3
5
3
5
3
5
3
5
3
5
3
5
3
5
3
5
3
5
3
5
3
5
3
5
3
5
3
5
3
5
3
5
3
5
3
5
3
5
3
5
3
5
3
5
3
5
3
5
3
5
3
5
3
5
3
5
3
5
3
5
3
5
0.3337
0.3337
0.5157
0.5157
0.3337
0.3337
0.5157
0.5157
0.3337
0.3337
0.5157
0.5157
0.3337
0.3337
0.5157
0.5157
0.3337
0.3337
0.5157
0.5157
0.3337
0.3337
0.5157
0.5157
0.3337
0.3337
0.5157
0.5157
0.3337
0.3337
0.5157
0.5157
0.3337
0.3337
0.5157
0.5157
0.3337
0.3337
0.5157
0.5157
0.3337
0.3337
0.5157
0.5157
0.3337
0.3337
0.5157
0.5157
0.3337
0.3337
0.5157
0.5157
0.3337
0.3337
0.5157
0.5157
0.3337
0.3337
0.5157
0.5157
0.3337
0.3337
0.5157
0.5157
0.8
0.8
0.8
0.8
1.2
1.2
1.2
1.2
0.8
0.8
0.8
0.8
1.2
1.2
1.2
1.2
0.8
0.8
0.8
0.8
1.2
1.2
1.2
1.2
0.8
0.8
0.8
0.8
1.2
1.2
1.2
1.2
0.8
0.8
0.8
0.8
1.2
1.2
1.2
1.2
0.8
0.8
0.8
0.8
1.2
1.2
1.2
1.2
0.8
0.8
0.8
0.8
1.2
1.2
1.2
1.2
0.8
0.8
0.8
0.8
1.2
1.2
1.2
1.2
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
90
90
90
90
90
90
90
90
90
90
90
90
90
90
90
90
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
90
90
90
90
90
90
90
90
90
90
90
90
90
90
90
90
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
15
0.3
0.2
0.2
0.3
0.2
0.3
0.3
0.2
0.2
0.3
0.3
0.2
0.3
0.2
0.2
0.3
0.3
0.2
0.2
0.3
0.2
0.3
0.3
0.2
0.2
0.3
0.3
0.2
0.3
0.2
0.2
0.3
0.3
0.2
0.2
0.3
0.2
0.3
0.3
0.2
0.2
0.3
0.3
0.2
0.3
0.2
0.2
0.3
0.3
0.2
0.2
0.3
0.2
0.3
0.3
0.2
0.2
0.3
0.3
0.2
0.3
0.2
0.2
0.3
0.8
0.2
0.2
0.8
0.8
0.2
0.2
0.8
0.8
0.2
0.2
0.8
0.8
0.2
0.2
0.8
0.2
0.8
0.8
0.2
0.2
0.8
0.8
0.2
0.2
0.8
0.8
0.2
0.2
0.8
0.8
0.2
0.2
0.8
0.8
0.2
0.2
0.8
0.8
0.2
0.2
0.8
0.8
0.2
0.2
0.8
0.8
0.2
0.8
0.2
0.2
0.8
0.8
0.2
0.2
0.8
0.8
0.2
0.2
0.8
0.8
0.2
0.2
0.8
0.0115
0.3118
0.2597
0.0029
0.1685
0.4616
0.3718
0.2233
0.0195
0.2832
0.2382
0.0131
0.1469
0.4529
0.3737
0.1974
0.1193
0.0595
0.0637
0.3571
0.1704
0.1665
0.1381
0.3264
0.1308
0.059
0.0497
0.3144
0.1558
0.1733
0.157
0.3634
0.2251
0.0298
0.0466
0.2622
0.3362
0.2052
0.1821
0.4064
0.226
0.041
0.0605
0.2715
0.3296
0.2064
0.1774
0.4266
0.0334
0.227
0.1815
0.0663
0.122
0.3099
0.2732
0.2452
0.0371
0.2738
0.2119
0.1039
0.1331
0.305
0.3074
0.2618
4.5. Statistical analysis of the two-level fractional factorial design
The main focus of this part is to determine the most significant
factors among the eight factors in Table 2 for the propulsive
performance of a propeller. A reduced number of factors will then
be used to obtain a response surface design for fitting a secondorder polynomial model to the thrust coefficient. Design Experts
7.03 from Statease, a stand-alone software for design of experiments was used to design the experiment and analyze the results.
The Pareto chart for the response thrust coefficient is shown in
Fig. 2. This is a plot of the ordered absolute value of the effects
estimates. The important main effects in descending order that
emerge from this analysis are advance coefficient, J, pitch ratio,
P0.7R/D, number of blades (N), skew angle (ys), and chord-diameter
ratio (c/D). The two-factor interactions that stood out included,
interaction of skew angle and advance coefficient, interaction of
ARTICLE IN PRESS
M.F. Islam, L.M. Lye / Ocean Engineering 36 (2009) 237–247
241
together with the other parameters, how it will interact with the
other parameters could not be determined and quantified.
4.6. Further study of the significant parameters
Having now identified the most significant dimensionless
parameters that explained the variation of the dependent
dimensionless parameter of interest, follow up experiments can
now be carried out to refine the models so that quadratic terms
can be added to the model presented by Eq. (8). Adding quadratic
terms would allow nonlinear effects to be modelled. Experimental
designs for fitting second-order models are known as response
surface methodology (RSM) in DOE terminology. This will be
discussed in the next section.
5. Response surface modelling
Fig. 1. Propeller models simulated by PROPELLA. From left to right (row-wise),
Standard run # 1, 22, 35, 46, 55 and 64, as shown in Table 3.
Fig. 2. Pareto chart for thrust coefficient in the FFD analysis.
pitch ratio and skew angle, interaction of number of blades and
advance coefficient, interaction of chord-diameter ratio and skew
angle, interaction of pitch ratio and advance coefficient, and
interaction of number of blades and skew. The model for KT
obtained from the fractional factorial design in coded units is
given in Eq. (8).
c
P0:7R
þ 0:058
0:017ys 0:088J
D
D
c
P0:7R
þ0:011N ys 0:015NJ þ 0:014 ys 0:016
ys
D
D
P0:7R
J þ 0:021ys J.
þ0:011
(8)
D
K T ¼ 0:200 þ 0:030N þ 0:016
This simple linear model gave a R2 value of about 96% and
predicted R2 of about 94%.
It should be noted that the five main factors were identified to
be: number of blades (N), chord-diameter ratio (c/D), pitch ratio
(P0.7R/D), skew (ys), and the advance coefficient (J) with the
advance coefficient being the most significant factor. From Eq. (8)
it is clear that the advance coefficient, skew, and pitch ratio and
their interactions play a dominant role in determining the thrust
coefficient, KT, with the other terms playing a minor role. It should
be pointed out that if the advanced coefficient was not analyzed
Response surface methodology is a strategy to achieve a goal
that involves experimentation, modelling, data analysis, and
optimization. Usually a sequential experimentation strategy is
considered. This facilitates an efficient search of the input factors
to optimize the response by using a first order experiment, as was
done above using a fractional factorial design, followed by a
second-order experiment. The second-order design allows one to
approximate the response surface relationship with a fitted
second-order regression model to include nonlinearities in the
response surface. There are several second-order designs available. Among the most popular designs are the central-composite
designs or CCD and the Box-Behnken design or BBD. There are also
special space-filling designs that are meant for computer experiments. Some of these newer designs are discussed in Ryan (2007).
Details of the RSM design phases and theoretical background on
the CCD and BBD are presented in Myers and Montgomery (2002)
among others. In this paper, three second-order response surface
designs were considered with five factors. They are the conventional Face-Centered Central-Composite design and the BoxBehnken design; and a space-filling design called Uniform Design
(Fang et al., 2006).
The response functions to be considered are given in Eq. (9)
after screening out the insignificant factors, which are, rake angle
(yr), hub taper angle (a), and the hub-diameter ratio (Dh/D).
T
V A c P 0:7R
;
;
;
¼
f
y
;
N
,
(9)
s
nD D D
rw n2 D4
where all terms are as defined earlier and in Table 2. The final five
factors used to develop the second-order response surface and
their ranges considered are given in Table 4. Other dimensions in
defining propeller geometry and the operating condition are held
constant. The thrust coefficient, KT of the propeller was calculated
using the numerical simulation program PROPELLA described
earlier for all the test points.
5.1. Central-composite design (CCD)
The central-composite design is a quadratic design that
contains an embedded factorial or fractional factorial design. In
this design the treatment combinations are at the factorial or
fractional factorial points plus axial points and at the center. These
designs are rotatable or near rotatable and can give up to 5 levels
of each factor. The axial points can be located at the face of the
design space or located at a distance designed to give a rotatable
design depending on the number of factors. Future details can be
obtained from Myers and Montgomery (2002). The CCD is one of
the most popular experimental designs for fitting a second-order
ARTICLE IN PRESS
242
M.F. Islam, L.M. Lye / Ocean Engineering 36 (2009) 237–247
Table 4
Control factors and their ranges for the response surface designs.
A
B
C
D
E
Parameter name
Symbol
Low value (1)
Mid value (0)
High value (+1)
Number of blades
Chord-diameter ratio
Pitch-diameter ratio
Skew
Propeller advance coeff.
N
c/D
P0.7R/D
3
0.33
0.80
01
0.20
4
0.43
1.00
451
0.50
5
0.52
1.20
901
0.80
ys
J ¼ VA/nD
Table 5
CCD data sheet.
Std. order
N
c/D
P0.7R/D
ys
J
KT
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
3
5
3
5
3
5
3
5
3
5
3
5
3
5
3
5
3
5
4
4
4
4
4
4
4
4
4
0.33
0.33
0.52
0.52
0.33
0.33
0.52
0.52
0.33
0.33
0.52
0.52
0.33
0.33
0.52
0.52
0.425
0.425
0.33
0.52
0.425
0.425
0.425
0.425
0.425
0.425
0.425
0.8
0.8
0.8
0.8
1.2
1.2
1.2
1.2
0.8
0.8
0.8
0.8
1.2
1.2
1.2
1.2
1
1
1
1
0.8
1.2
1
1
1
1
1
0
0
0
0
0
0
0
0
90
90
90
90
90
90
90
90
45
45
45
45
45
45
0
90
45
45
45
0.8
0.2
0.2
0.8
0.2
0.8
0.8
0.2
0.2
0.8
0.8
0.2
0.8
0.2
0.2
0.8
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.2
0.8
0.5
0.024
0.2926
0.4719
0.0459
0.3421
0.2117
0.1744
0.46
0.1368
0.0577
0.081
0.2235
0.1127
0.2625
0.1952
0.2329
0.1976
0.2798
0.2211
0.264
0.1746
0.2988
0.2349
0.1721
0.3205
0.1434
0.2434
Fig. 3. Actual versus predicted values for the response, thrust coefficient in the
CCD model.
model. For this experiment, a half fractional factorial with five
factors, 10 axial points at the face, and a center point were used
giving a total of 27 design points. When the axial points are placed
at the low and high values, the design is called a face-centered
central-composite design. The design and corresponding responses are given in Table 5.
The analysis part of the RSM consists of different distinct steps
to ensure the predicted model acts well and the statistical
assumptions are valid. Using a step-wise regression approach
and using a significance level of 5%, the final quadratic regression
equation in terms of the actual factors for the thrust coefficient, KT,
is shown in Eq. (10).
pffiffiffiffiffiffi
c
P0:7R
K T ¼ 0:3623 0:0292N þ 1:8144 0:1602
0:0012ys
D
D
c
P0:7R
0:9981J 0:1917N: þ 0:1192N:
þ 0:0003N:ys
D
D
c P
P
P
0:7416 : 0:7R 0:0010 0:7R :ys þ 0:6826 0:7R :J þ 0:0037ys J
D D
D
D
2
0:00002ys 0:1983J 2 .
(10)
The R2 is about 0.995 and the prediction R2 is about 0.953. The
actual versus predicted values for the model is shown in Fig. 3.
The addition of the quadratic terms is clearly reflected in the
curved interaction plots between the skew and advance coefficient, and between the skew and the advance coefficient as shown
in Figs. 4 and 5, respectively.
Fig. 4. Interaction plot between number of blades and pitch-diameter ratio for the
response, thrust coefficient in the CCD analysis.
5.2. Box-Behnken design (BBD)
The Box-Behnken design is an independent quadratic design in
that it does not contain an embedded factorial or fractional
factorial design. In this design the treatment combinations are at
the midpoints of edges of the process space and at the center.
These designs are rotatable (or near rotatable) and require 3 levels
of each factor. However, the use of the BBD should be confined to
situations in which one is not interested in predicting response at
ARTICLE IN PRESS
243
M.F. Islam, L.M. Lye / Ocean Engineering 36 (2009) 237–247
Table 6
BBD data sheet.
Fig. 5. Interaction plot between skew and advance coefficient for the response,
thrust coefficient in the CCD analysis.
the extremes, that is, at the corners of the cube. Futher details can
be obtained from Myers and Montgomery (2002). For this design,
a total of 46 experimental runs were required and the run
combinations and responses are shown in Table 6. For the BBD,
the final quadratic regression equation in terms of the actual
factors for the thrust coefficient, KT, is shown in Eq. (11).
pffiffiffiffiffiffi
c
P 0:7R
þ 0:0006ys
K T ¼ 0:0157 þ 0:0406N þ 0:0998 þ 0:6844
D
D
c
P 0:7R
P 0:7R
0:7259J þ 0:0021 :ys 0:0018
:ys þ 0:5044
:J
D
D
D
2
P
2
þ 0:0026ys J 0:2553 0:7R 0:00002ys 0:2158J2 (11)
D
The R2 is about 0.992 and the prediction R2 is about 0.981. The
actual versus predicted values for the model is shown in Fig. 6.
5.3. Uniform design (UD)
The uniform design is an efficient, near orthogonal, and robust
fractional factorial design for experiments where there are a large
number of factors (four or more) and levels (three or more). It was
first proposed by Kai-Tai Fang and Wang Yuan in 1980 and is
described in great detail in Fang and Lin (2003) and in Fang et al.
(2006). The uniform design is an important class of space-filling
designs, which is useful for computer and industrial experiments
when there is little or no information about the effect of factors on
the response but the true model is suspected to be highly
nonlinear. Design points are essentially regularly spaced over the
design region.
The construction of uniform designs for more than two factors
is not as straightforward as the classical response surface designs.
The construction depends on the measure of uniformity used.
Fortunately, uniform designs developed by Fang et al. have been
tabulated and are available at http://www.math.hkbu.edu.hk/
UniformDesign. Also, unlike classical designs, there is a wide
choice of sample sizes for a given measure of uniformity, number
of levels, and number of factors. For example, for a Un(q)s uniform
design based on the centered L2-discrepancy measure, sample size
n can range from 9 to 51 when the number of levels q is 3 and
number of factors s is 5. The choice of the sample size hence
depends on balancing the need for minimizing the number of
experimental runs with the complexity of the model to be fitted
and the type of modeling approach used. To implement the
Std. order
N
c/D
P0.7R/D
ys
J
KT
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
3
5
3
5
4
4
4
4
4
4
4
4
3
5
3
5
4
4
4
4
4
4
4
4
3
5
3
5
4
4
4
4
3
5
3
5
4
4
4
4
4
4
4
4
4
4
0.33
0.33
0.52
0.52
0.425
0.425
0.425
0.425
0.33
0.52
0.33
0.52
0.425
0.425
0.425
0.425
0.425
0.425
0.425
0.425
0.33
0.52
0.33
0.52
0.425
0.425
0.425
0.425
0.425
0.425
0.425
0.425
0.425
0.425
0.425
0.425
0.33
0.52
0.33
0.52
0.425
0.425
0.425
0.425
0.425
0.425
1
1
1
1
0.8
1.2
0.8
1.2
1
1
1
1
0.8
0.8
1.2
1.2
1
1
1
1
0.8
0.8
1.2
1.2
1
1
1
1
0.8
1.2
0.8
1.2
1
1
1
1
1
1
1
1
1
1
1
1
1
1
45
45
45
45
0
0
90
90
45
45
45
45
45
45
45
45
0
90
0
90
45
45
45
45
0
0
90
90
45
45
45
45
45
45
45
45
0
0
90
90
45
45
45
45
45
45
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.2
0.2
0.8
0.8
0.5
0.5
0.5
0.5
0.2
0.2
0.8
0.8
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.2
0.2
0.8
0.8
0.2
0.2
0.8
0.8
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.18
0.2562
0.2109
0.307
0.1546
0.3095
0.129
0.209
0.3017
0.3441
0.1224
0.1594
0.1402
0.1987
0.2435
0.3514
0.3509
0.2049
0.1165
0.1169
0.1538
0.1903
0.2773
0.3239
0.2067
0.2577
0.138
0.2075
0.2677
0.3731
0.0621
0.215
0.2646
0.3707
0.1151
0.1682
0.2367
0.238
0.1596
0.1906
0.2434
0.2434
0.2434
0.2434
0.2434
0.2434
Fig. 6. Actual versus predicted values for the response, thrust coefficient in the
BBD model.
ARTICLE IN PRESS
244
M.F. Islam, L.M. Lye / Ocean Engineering 36 (2009) 237–247
pffiffiffiffiffiffi
hci
P 0:7R
K T ¼ 0:5060 þ 0:0476½N þ 0:0869
0:00052½ys
þ 0:3840
Table 7
UD data sheet.
Std. order
N
c/D
P0.7R/D
ys
J
KT
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
4
4
3
5
4
3
3
4
4
5
3
3
5
5
3
3
5
5
4
4
4
4
5
3
3
5
5
0.43
0.43
0.52
0.43
0.52
0.52
0.43
0.33
0.43
0.43
0.33
0.33
0.52
0.52
0.33
0.33
0.43
0.52
0.33
0.52
0.52
0.33
0.33
0.52
0.43
0.33
0.33
1.2
1
1
1.2
0.8
1.2
1
1
0.8
0.8
0.8
1.2
0.8
1.2
1
1
1
1
1.2
1.2
1
0.8
0.8
0.8
0.8
1.2
1
45
45
45
0
0
0
45
0
90
90
45
0
45
90
90
90
0
0
90
45
90
0
45
90
0
45
90
0.2
0.5
0.2
0.5
0.2
0.5
0.5
0.8
0.2
0.8
0.8
0.2
0.5
0.2
0.2
0.2
0.2
0.8
0.5
0.8
0.8
0.5
0.2
0.5
0.8
0.8
0.5
0.3731
0.2434
0.2809
0.3407
0.1694
0.2728
0.1976
0.111
0.1733
0.0898
0.0371
0.3421
0.2268
0.302
0.1591
0.1591
0.3799
0.1305
0.1923
0.2364
0.1372
0.1544
0.2817
0.1165
0.0271
0.2489
0.2075
uniform design in industrial experiments the following steps are
necessary (Fang and Lin, 2003):
1. Choose factors and experimental domain as well as determine
suitable number of levels for each factor.
2. Choose a suitable UD to accommodate the number of factors
and levels.
3. From the UD table, randomize the run order if necessary and
conduct the experiments.
4. Find a suitable model to fit the data. E.g., regression analysis,
neural networks, wavelets, multivariate splines, etc., and
5. Knowledge discovery from the built model. Optimize response
if required.
As can be seen, the use of uniform designs is similar to the use
of the classical fractional factorial designs or response surface
designs except that there is a wider choice of levels, number of
runs, and model fitting methods. This means that results obtained
from a uniform design may be quite different depending on the
various choices made at each step.
For the problem at hand, a uniform design based on the
centered L2-discrepancy uniformity measure that uses the same
number of factors and levels as the CCD and BBD was used. The
sample size chosen was 27. This was the same number as the half
fractional face-centered central-composite design used earlier.
The experimental run combinations and responses obtained are
given in Table 7. A step-wise second-order centered quadratic
regression at the 5% significance level was used to obtain the
relationship between the factors and the response. The final
quadratic regression equation in terms of the actual factors for the
thrust coefficient, KT, is shown in Eq. (12). In this equation, the
terms with square brackets are centered values. That is,
½N ¼ N 4:000;
hci
D
¼
c
P 0:7R
P 0:7R
0:9926,
¼
0:4230;
D
D
D
½ys ¼ ys 45:00; ½J ¼ J 0:4778
D
D
hci
P 0:7R
P 0:7R
:½ys 0:0022
:½ys þ 0:1996
:½J
D
D
D
hci
hci
þ 0:0014½ys ½J 0:1140½N:
þ 0:4540
½J
D
D
P 0:7R 2
0:3710
0:000025½ys 2 0:3340½J2 .
(12)
D
0:2870½J þ 0:0051
The R2 is about 0.992 and the prediction R2 is about 0.942. The
actual versus predicted values for the model is shown in Fig. 7.
Square-root transformation is again necessary for the response to
give a better fit to the assumptions of regression analysis.
Centering the explanatory variables reduces the multicollinearity
problem considerably.
5.4. Comparison of results
From the above sections, one can see that all second-order
models gave reasonably high R2 and predicted R2 values, with the
BBD giving the highest values and the UD giving the lowest values.
However, it is important to note that these R2 values are not the
true measure on how well the models actually perform when
tested against data that have not been used in developing the
models. To compare the results among the three second-order
experimental designs, 20 sets of the five factors (N, c/D, P0.7R/D, ys
and J) within their respective ranges were randomly generated as
shown in Table 8. As shown in Table 8, the propeller pitch ratios
might be unusual numbers and might not normally be chosen for
a propeller model. The random values of the parameters were
merely used to test the accuracy of the models, not to provide
practical propeller design parameter values. The responses of
the propellers using the random parameter values were then
calculated using PROPELLA. The responses were then compared to
the predictions obtained from each of the regression models. The
criterion used for the comparison is the mean absolute percentage
error (MAPE). The most accurate model would be the one with the
smallest MAPE. The results of the comparison are shown in Table
8 and the plots of actual versus predicted values for the thrust
coefficient based on each model are shown in Figs. 8–10,
respectively. The results show that the most accurate predictions
for KT are those based on the BBD, followed by UD. The models
based on the CCD although gave higher R2 and prediction R2
values than the models for the UD, did not perform as well for the
20 unseen data. The MAPE for the KT predictions were 11.96%,
5.35%, and 3.11% for the CCD, UD, and BBD, respectively. In terms
of maximum percentage error, they were 33.61%, 16.01%, and
9.90% for the CCD, UD, and BBD, respectively. If we choose the best
model, the MAPE is about 3% with a maximum percentage error of
less than 10%, which would be quite acceptable for hydrodynamic
experiments.
It should also be noted that while the regression equations
for all three experimental designs looked similar at first glance,
they however do not possess the same terms. It is likely that if
another uniform design was used, say one with a larger sample
size, a different regression equation would result. Also, if a
different fitting method was used, say multivariate splines or
neural network, again a better goodness of fit could be obtained.
Another point to make is that with five independent parameters, the CCD or UD with 27 experimental runs, one can study
the linear and quadratic effects of each of the five parameters and
their interactions. If the traditional one-factor-at-a time or OFAT
approach has been used, 48 runs would be required just to study
the effects of each of the five parameters with the same reliability
as the fractional factorial design. Furthermore, the OFAT design
would not inform us about the presence of interaction effects; this
ARTICLE IN PRESS
245
M.F. Islam, L.M. Lye / Ocean Engineering 36 (2009) 237–247
Fig. 7. Actual versus predicted values for the response, thrust coefficient in the UD model.
Table 8
Comparison of results.
Run
N
c/D
P0.7R/D
ys
J
KT
KT_CCD
KT_BBD
KT_UD
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
5
3
3
3
4
5
5
5
3
3
4
3
4
3
3
3
3
4
5
5
0.3687
0.3402
0.5194
0.4485
0.4554
0.3409
0.4598
0.4869
0.3919
0.4416
0.3794
0.3495
0.4777
0.4775
0.4783
0.3346
0.4820
0.4341
0.4609
0.4422
1.1346
1.0068
0.9772
0.8830
1.0259
1.1796
0.9575
0.9044
1.0783
0.8445
1.1565
1.1926
0.9060
1.0619
1.1149
1.0160
0.8834
1.1879
0.8388
1.0007
80.7882
87.3684
18.7508
73.8626
19.9483
16.7385
60.5916
64.7757
59.1554
37.0964
61.9974
1.4986
80.8365
18.2799
42.4605
69.8868
73.1526
10.5408
83.3857
48.0798
0.7084
0.5169
0.3502
0.6057
0.7281
0.6936
0.5412
0.7282
0.5516
0.6955
0.7350
0.3373
0.5762
0.7010
0.6728
0.7514
0.5758
0.3714
0.7711
0.7358
0.1984
0.1293
0.2608
0.1196
0.1795
0.2574
0.2488
0.1749
0.1813
0.0991
0.1931
0.3054
0.1624
0.1719
0.1933
0.1027
0.1333
0.3573
0.1238
0.1995
0.2336
0.1139
0.3484
0.1411
0.1655
0.2703
0.2235
0.1430
0.1857
0.1240
0.2034
0.3019
0.1644
0.1840
0.2090
0.0954
0.1660
0.3582
0.0949
0.1773
0.2060
0.1273
0.2651
0.1189
0.1719
0.2715
0.2450
0.1655
0.1912
0.0927
0.1983
0.3124
0.1610
0.1643
0.1978
0.1056
0.1330
0.3582
0.1115
0.1988
0.2109
0.1228
0.2356
0.1250
0.1763
0.2986
0.2612
0.1784
0.1969
0.0967
0.1962
0.3109
0.1722
0.1651
0.2046
0.0931
0.1426
0.3554
0.1206
0.2135
Run
%KT_CCD
%KT_BBD
%KT_UD
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
17.72
11.95
33.61
17.99
7.80
5.03
10.18
18.22
2.41
25.15
5.33
1.16
1.25
7.05
8.10
7.10
24.51
0.24
23.34
11.12
3.81
1.56
1.63
0.57
4.22
5.47
1.54
5.38
5.46
6.41
2.69
2.30
0.85
4.40
2.30
2.78
0.21
0.25
9.90
0.37
6.32
5.01
9.68
4.49
1.78
16.01
4.97
1.98
8.63
2.41
1.61
1.80
6.02
3.95
5.86
9.34
6.96
0.54
2.56
7.03
ARTICLE IN PRESS
246
M.F. Islam, L.M. Lye / Ocean Engineering 36 (2009) 237–247
Fig. 8. Actual versus predicted values of thrust coefficient in the CCD model.
Fig. 9. Actual versus predicted values of thrust coefficient in the BBD model.
is a serious defect. Even if we were to say that there might not be
interaction, there is no decline in the merit of the factorial design.
Indeed, in the OFAT design, there would be no way to determine
that there is no interaction (Berger and Maurer, 2002). Hence, not
only there would be cost savings in using factorial designs, it is the
more correct and complete method of experimental design.
6. Conclusions
This paper has presented a combined dimensional analysis and
statistical design of experiments approach to reduce the high
number of variables to a reduced set of significant variables. The
system under consideration was a propeller operating in open
water whose performance depends on many variables. A total
number of fourteen variables were considered to control the
thrust coefficient. The number of variables were first reduced by
three using dimensional analysis, which also changes the
variables to a more useful format in terms of dimensionless
Fig. 10. Actual versus predicted values of thrust coefficient in the UDD model.
parameters. A two-level fractional factorial design was then used
to screen out any statistically insignificant parameters. A secondorder experimental design was then used to obtain information to
define a functional relationship between the thrust coefficient and
the five parameters. In developing the functional relationships,
three response surface experimental designs were considered.
Two of them, namely the CCD and BBD are well-known classical
designs, whereas the UD is a rather new type of space-filling
design that have been found to be useful in both industrial and
computer experiments. The UD provides for a greater choice of
designs and modeling approaches, which also means that the user
would need a much greater range of statistical and mathematical
expertise to use it. It was also seen in the comparison of the
results that a different experimental design might give rise to
different significant terms in the regression. For the problem at
hand, the BBD gave the best prediction model. The final models as
presented may not be as accurate as classical regression methods
for interpolating and extrapolating the thrust coefficient. The
purpose of this paper however is not to find the best models for
the thrust coefficient but to illustrate how a problem with a large
number of variables can be systematically handled to obtain
functional relationships using a variety of modern statistical
design of experiment tools combined with the more well-known
dimensional analysis.
Acknowledgments
The authors would like to thank Pengfei Liu of Institute for
Ocean Technology of National Research Council for giving the
permission to use his panel method code PROPELLA, which was
used to carry out all the numerical predictions on the performance characteristics of the propellers at various geometric
configurations required by the DOE analysis process. Thanks
are also extended to Brian Veitch of Memorial University of
Newfoundland for all his support and suggestions in carrying out
the work.
References
Antony, J., 2006. Design of Experiments for Engineers and Scientists. Elsevier,
Butterworth-Heinemann.
Barr, D.I.H., 1969. Method of synthesis: basic procedures for the new approach to
similitude. Water Power, 148–153 pp. 183–8.
ARTICLE IN PRESS
M.F. Islam, L.M. Lye / Ocean Engineering 36 (2009) 237–247
Berger, P.D., Maurer, R.E., 2002. Experimental Design with Applications in
Management, Engineering, and the Sciences. Duxbury Press.
Box, G., et al., 2006. Improving Almost Anything. Wiley, New York.
Box, G., Hunter, Hunter, 2005. Statistics for Experimenters, second ed. Wiley, New York.
Fang, K.T., Lin, D.K.J., 2003. Uniform experimental designs and their applications in
industry. In: Khattree, R., Rao, C.R. (Eds.), Handbook of Statistics, Vol. 22.
Elsevier, Amsterdam.
Fang, K.T., Li, R., Sudjianto, A., 2006. Design and Modeling for Computer
Experiments. Chapman and Hall/CRC Press, London.
Hawkins, D., Lye, L.M., 2006. Use of DOE methodology for investigating conditions
that influence the tension in marine risers for FPSO ships. In: 1st International
Structural Specialty Conference, Calgary, Alberta.
Liu, P., 2003–2008. PROPELLA: User Manual, Institute of Ocean Technology. St.
John’s, NL, p. 24.
247
Liu, P., 2006. The design of a podded propeller base model geometry and prediction
of its hydrodynamics. Technical Report no. TR-2006-16, Institute for Ocean
Technology, National Research Council, Canada, p. 16.
Montgomery, D.C., 2005. Design and Analysis of Experiments, fifth ed. Wiley,
New York.
Munson, B.R., Young, D.F., Okiishi, T.H., 1994. Fundamentals of Fluid Mechanics,
second ed. Wiley, New York.
Myers, R.H., Montgomery, D.C., 2002. Response Surface Methodology: Process
and Product Optimization Using Designed Experiments, second ed. Wiley,
New York.
Ryan, T.P., 2007. Modern Experimental Design. Wiley, New York.
Sharp, J.J., 1981. Hydraulic Modelling. Butterworth’s, London.
Szirtes, T., 1997. Applied Dimensional Analysis and Modeling. McGraw Hill,
New York.