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

net/publication/229091254

Axial-Flow Ventilation Fan Design Through Multi-Objective Optimization to

Enhance Aerodynamic Performance

Article in Journal of Fluids Engineering · October 2011

DOI: 10.1115/1.4004906

CITATIONS READS

59 8,341

3 authors, including:

Jin-Hyuk KIM Kwang-Yong Kim

Korea Institute of Industrial Technology Inha University
292 PUBLICATIONS 2,549 CITATIONS 712 PUBLICATIONS 10,952 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Kwang-Yong Kim on 21 November 2014.

The user has requested enhancement of the downloaded file.

 Axial-Flow Ventilation Fan
Design Through Multi-Objective
Optimization to Enhance
Aerodynamic Performance
Jin-Hyuk Kim
This paper presents an optimization procedure for axial-flow ventilation fan design
Jae-Woo Kim through a hybrid multiobjective evolutionary algorithm (MOEA) coupled with a response
surface approximation (RSA) surrogate model. Numerical analysis of a preliminary fan
Kwang-Yong Kim1 design is conducted by solving three-dimensional (3-D) Reynolds-averaged Navier-
e-mail: kykim@inha.ac.kr Stokes (RANS) equations with the shear stress transport (SST) turbulence model. The
multiobjective optimization processes are performed twice to understand the coupled
Department of Mechanical Engineering, effects of diverse variables. The first multiobjective optimization process is conducted
Inha University, with three design variables defining stagger angles at the hub, mid-span, and tip, and the
253 Yonghyun-Dong, Nam-Gu, second is conducted with five design variables defining hub-to-tip ratio, hub cap installa-
Incheon 402-751, Rep. of Korea tion distance, hub cap ratio, and the stagger angles at the mid-span and tip. Two aerody-
namic performance parameters, the total efficiency and total pressure rise, are selected
as the objective functions for optimization. These objective functions are numerically
assessed through 3-D RANS analysis at design points sampled by Latin hypercube sam-
pling in the design space. The optimization yields a maximum increase in efficiency of
1.8% and a 31.4% improvement in the pressure rise. The off-design performance is also
improved in most of the optimum designs except in the region of low flow rate.
[DOI: 10.1115/1.4004906]

Keywords: axial-flow ventilation fan, numerical optimization, response surface

approximation, hybrid multiobjective evolutionary algorithm, Pareto-optimal solutions

1 Introduction gelmans [2] performed a parametric study on the blade stacking

line in a linear compressor cascade. Beiler and Carolus [3] con-
With the recent focus on a low-carbon economy and green
ducted a systematic study to understand the flow phenomena
growth, high-efficiency technology for turbomachinery, which is
within an axial-flow fan with skewed blades.
widely used for energy transfer, is highly desirable. The flow field
Systematic optimization is the most efficient way to find the
of turbomachinery is extremely complicated owing to a three-
optimal shape of turbomachinery for improving performance.
dimensional (3-D) unsteady vertical flow structure in the blade
Many researchers have optimized turbomachinery. Larocca [4]
passage. However, with recent developments in computational
presented an optimization procedure for the design of entire multi-
fluid dynamics, the accuracy of the analysis of turbomachinery
stage axial turbines and compressors based on a method for solv-
flows using 3-D Reynolds-averaged Navier-Stokes equations
ing inverse problems. Oyama et al. [5] developed a reliable and
(RANS) becomes adequate for use in practical design processes.
efficient aerodynamic design optimization tool for transonic com-
Optimization techniques associated with RANS analysis have
pressor blades; entropy production was minimized as the objective
recently become a promising tool for high-efficiency turbomachi-
function to reduce flow losses at the rotor. Chen et al. [6] proposed
nery design. Numerical optimization strategies, coupled with vari-
a strategy of 3-D blade design optimization using various methods
ous surrogate models, are widely applied to the design of
and increased the adiabatic efficiency of a single-rotor compressor
turbomachinery to satisfy the design requirements. In particular,
by 2.96%. Shu et al. [7] developed an efficient multipoint optimi-
optimization techniques based on response surface approximation
zation approach for optimizing the geometry of a blower with a
(RSA) [1] as the surrogate model have been widely applied to tur-
very low specific speed and successfully applied it to the redesign
bomachinery design by many researchers. The RSA model can
of an industrial centrifugal blower blade. Wang et al. [8] estab-
use information collected from various sources and by different
lished the Kriging model for design optimization of a centrifugal
tools; thus, it is effective for both single and multioptimization
compressor impeller. They demonstrated the accuracy of the Krig-
problems.
ing model and proposed a method to further improve its accuracy.
Blade shape is one of the most important factors affecting the
As the objective function for optimizations, efficiency has been
aerodynamic performance of turbomachinery such as fans and
widely used, since it is directly related to the aerodynamic per-
compressors. With the aid of 3-D analysis, it becomes possible to
formance of turbomachinery. Sørensen et al. [9] maximized the
analyze the effects of the blade staking line on performance,
efficiency through a numerical design optimization combined
which was not possible with the approximate analysis methods
with an arbitrary vortex flow model for rotor-only axial fans by
available in the past. Therefore, many recent studies have exam-
using a standard method. Kipouros et al. [10] optimized the design
ined the 3-D shapes of turbomachinery blades. Sasaki and Breu-
of axial compressors and tested enhancements to the optimizer
that improve the efficiency. Other objective functions are also
1
Corresponding author. used for optimization. For example, Kim et al. [11] enhanced the
Contributed by the Fluids Engineering Division of ASME for publication in the total-to-total pressure ratio of an impeller with four design varia-
JOURNAL OF FLUIDS ENGINEERING. Manuscript received November 22, 2010; final
manuscript received August 11, 2011; published online October 4, 2011. Assoc. bles by constructing the curves of the shroud and the hub by Bez-
Editor: Chunill Hah. ier curves. Samad et al. [12] demonstrated the advantage of using

Journal of Fluids Engineering Copyright V

C 2011 by ASME OCTOBER 2011, Vol. 133 / 101101-1

Downloaded 04 Oct 2011 to 165.246.100.201. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
 multiple surrogates in the shape optimization of a transonic axial
compressor blade and selected adiabatic efficiency, total pressure,
and total temperature as objective functions.
Furthermore, several multidisciplinary studies have been con-
ducted to improve the aerodynamic performance of turbomachi-
nery by applying multiobjective optimization procedures. Dennis
et al. [13] illustrated an automatic multiobjective design optimiza-
tion of a two-dimensional airfoil cascade row to simultaneously
minimize the total pressure loss, maximize the total aerodynamic
loading, and minimize the number of airfoils in a finite cascade
row. Oksuz and Akmandor [14] performed a genetic-algorithm
(GA) based aerodynamic shape optimization of an axial turbine
blade for two different maximization objectives, adiabatic effi-
ciency and torque. Arabnia and Ghaly [15] applied an optimiza-
tion method that combines a GA with an artificial neural network
Fig. 1 Sections of blade profiles from hub to tip
(ANN) to an axial turbine stage for two different optimization
objectives. Lotfi et al. [16] demonstrated that a method that links a
commercial computational fluid dynamics (CFD) code and a GA-
tional mesh generation are performed through Blade-Gen and
based multiobjective optimization offers a promising approach for
Turbo-Grid, respectively. CFX-Pre, CFX-Solver, and CFX-Post
industrial axial fan designers. Benini [17] developed a method for
are applied for defining boundary conditions, solving, and post-
multiobjective design optimization of a transonic compressor and
processing, respectively. The 3-D steady-state incompressible
obtained the optimal rotor configurations corresponding to the
RANS equations are discretized using the finite volume method
maximum pressure ratio and maximum efficiency at the design
(FVM). A high-resolution scheme that is second-order accurate in
point. Lian and Liou [18] brought together design of experiment,
space is used to solve the convection-diffusion equations.
an RSA model, a GA, and CFD analysis tools to provide an inte-
The k-x-based SST model [25] is used to account for the trans-
grated optimization system. Kim et al. [19] performed design opti-
port of the turbulent shear stress and to produce highly accurate
mization to enhance the isentropic efficiency and total pressure
results for the start and the amount of flow separation under an
ratio of a centrifugal compressor by using a hybrid multiobjective
adverse pressure gradient. The prediction in terms of flow separa-
evolutionary algorithm (MOEA) [20,21].
tion is much improved by using the SST model [26]. This model
This work presents a numerical optimization procedure for an
uses a k-x model in the near-wall region and a k-e model in the
axial-flow ventilation fan that uses a hybrid MOEA coupled with
bulk domain, and a blending function ensures smooth transitions
an RSA model [1] for surrogate modeling. Two multiobjective
between the two. The accuracy of the numerical scheme for turbu-
optimization processes, the first multiobjective optimization (first
lent flow depends strongly on the treatment of the wall shear
MOPT) and the second multiobjective optimization (second
stress. In the present study, to benefit from the SST model, a reso-
MOPT), are conducted to understand the coupled effects of
lution of the boundary layer with more than 10 mesh points is
diverse variables on the enhancement of aerodynamic perform-
required, and the near-wall grid resolution is adjusted to keep
ance. The objective functions are numerically assessed through 3-
yþ  2 to accurately capture wall shear stress and also to imple-
D RANS analysis at each design point sampled by Latin hyper-
ment low-Reynolds-number SST model [24].
cube sampling (LHS). Each clustered optimal solution (COS)
The computational domain for the numerical analysis is a single
from the global Pareto-optimal solution (POS) is analyzed to com-
passage of the axial-flow fan, as shown in Fig. 2. The passage
pare its aerodynamic performance with that of a reference model.
includes a single blade consisting of one rotating domain and two
stationary domains (inlet and outlet domains); it is analyzed by
2 Preliminary Design of an Axial-Flow Fan assuming that the flow between two adjacent main blades is peri-
odic in the direction of rotation. The total pressure and the
A computerized preliminary design system [22] for an axial-
designed mass flow rate are set at the inlet and outlet of the com-
flow fan is used to construct the basic 3-D blade geometry and
putational domain, respectively. Air at 25  C is considered as the
predict its aerodynamic performance on the basis of simple ana-
working fluid. The solid surfaces in the computational domain are
lytical methods. The system [22] used in this work was previously
considered to be hydraulically smooth with adiabatic and no-slip
applied to actual aircraft and air-conditioning fans, and the results
conditions. A frozen rotor method [24] is used at the interface
of its predictions are compared with corresponding experimental
between the stationary and rotating domains.
results to verify the prediction accuracy of the design system
A structured grid system is constructed in the computational do-
[22,23].
main, which has O-type grids near the blade surfaces and H-type
The preliminary aerodynamic blade design of the axial-flow
grids in other regions. The inlet and outlet domains are each con-
fan, shown in Fig. 1, is conducted by using a through-flow model-
structed using approximately 52,000 grid points, whereas the
ing technique to successively determine the blade angle distribu-
rotating domain is constructed using approximately 116,000 grid
tion, camber line, and airfoil thickness distribution, and finally
points. A grid-dependency test was performed at various grids;
stack the blade profiles in the spanwise direction. Before creating
the results are shown in Fig. 3 for the static pressure distributions
the preliminary design, however, the four design variables (e.g.,
at the trailing edge (TE) of the blade. The static pressure at the TE
tip diameter, hub-to-tip ratio, chord length, number of blades) are
decided by setting the design requirements.
The required pressure rise is 100 Pa at a volume flow rate of 60
m3=min. The axial-flow fan is operated at a speed of 1170 rpm, Table 1 Design specifications for axial-flow fan
the number of fan blades is 10, the tip diameter is 0.510 m, and
the hub-to-tip ratio is 0.294. More detailed specifications are given Design volume flow rate, m3=min 60
in Table 1. Rotational speed, rpm 1170
Pressure rise, Pa 100
Tip diameter, mm 510
3 Computational Methods Hub-to-tip ratio 0.294
Tip clearance, mm 2
The flow of the axial-flow fan is analyzed using the commercial Number of blades 10
code ANSYS CFX 11.0 [24]. Blade profile creation and computa-

101101-2 / Vol. 133, OCTOBER 2011 Transactions of the ASME

Downloaded 04 Oct 2011 to 165.246.100.201. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
 Fig. 4 Validation of flow analysis

occurs near the blade suction surface (SS). If the diffusion factor
Fig. 2 Three-dimensional geometry and computational do- is designed to be more than 0.5, the viscous loss due to locally
main of axial-flow fan occurring separation generally increases rapidly. As shown in Fig.
5(a), the diffusion factor is more than 0.5 in the region from the
does not change as the grid size varies from approximately hub to near 0.25 times the length of the span, and a local separa-
220,000 to 430,000. Thus, about 220,j000 grid points are used to tion region is detected near the TE of the hub, as shown in Fig.
define the passage. Figure 2 shows a typical example of the sur- 5(b). Therefore, the numerical analysis is considered to have reli-
face grid system used for the axial-flow fan in the present study. able accuracy considering the results of these validations.
Root-mean-square (RMS) values of the equation residuals for
convergence criteria are specified to be at least 104 for all equa- 5 Parametric Study
tions. The physical time scale is set to 1=x, where x is the angular Six variables related to the geometry of the axial-flow fan are
velocity of the blades. The converged solutions are obtained after tested to investigate their effects on the fan’s efficiency. These vari-
500 iterations. The computations were performed on a PC with an ables are the stagger angles (fh; , fm , and ft ) at the hub, mid-span,
Intel Pentium IV CPU and a clock speed of 3.0 GHz. The compu- and tip locations, respectively, the hub-to-tip ratio (Rh–t), the hub
tational time is approximately 4 hrs, depending upon the geometry cap installation distance (Dhc), and the hub cap ratio (Rhc). These
considered and the rate of convergence. variables are described in detail below.

4 Validation of Numerical Solutions

To evaluate the accuracy of the numerical analysis, the results
of the flow analysis should be validated before design optimiza-
tion. The reference fan model used for this validation is the basic
model designed by a computerized preliminary design system
[22]. The validity of the numerical analysis is verified by compari-
son with the results of the prediction of this system. Figure 4
shows the results of validation by comparing all the system curves
for the pressure rise and the efficiency for the preliminary design
predictions and the RANS analyses. The efficiency and the pres-
sure rise show acceptable levels of agreement. The results for the
pressure rise in particular tend to be highly similar.
The analysis is also validated by comparing the internal flow
fields determined by the numerical analysis and the quantitative
results provided by the preliminary design, as shown in Fig. 5.
Figure 5(a) shows the diffusion factor distribution along the span-
wise direction. When a blade is designed for turbomachinery, the
diffusion factor can serve as the criterion for whether separation

Fig. 5 Comparison of the results of RANS analysis and

the preliminary design: (a) Diffusion factor along the spanwise
direction (preliminary design), (b) Separation region near trailing
Fig. 3 Results of grid-dependency test edge of hub (RANS analysis)

Journal of Fluids Engineering OCTOBER 2011, Vol. 133 / 101101-3

Downloaded 04 Oct 2011 to 165.246.100.201. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
 Fig. 6 Definition of design variables: (a) Definition of the stagger angle, (b) a angle distribution according to a Bezier
curve, (c) Blade generation by interpolation with B-spline curve, (d) Schematic diagram defining the hub-to-tip ratio,
(e) Meridional view defining the hub cap

5.1 Stagger Angles at the Hub, Mid-Span, and Tip varied to change Rh–t. Here, the variation in dh is limited by two
Locations. The stagger angle is illustrated in Fig. 6(a). Stagger constraints on Rh–t; one is the root length (lr) for the coupling
angles (fh; , fm , and ft at the hub, mid-span, and tip, respectively) between the blades and the hub, and the other is the shaft diameter
are defined by the variation in chord length, which changes (ds) for the coupling between the motor shaft and the hub. As dh
according to the a angle distribution, which consists of the fourth- decreases, the blade length is extended by the amount by which dh
order Bezier curve shown in Fig. 6(b). Here, a is defined as the decreases, as shown in Figs. 6(c) and 6(d). Consequently, the
angle between the rotational axis and a tangent of the camber line. blade length is extended from the original hub span to the modi-
When each stagger angle is changed at hub, mid-span, and tip, the fied hub span by using interpolation with a B-spline curve, as
blade profiles of other locations except for the designated three shown in Fig. 6(c). Figure 6(c) illustrates that the values at undes-
span locations are interpolated by the B-spline curve from hub to ignated blade spans (other than those at the hub, mid-span, and
tip, as shown in Fig. 6(c). The advantage of using the Bezier curve tip) are interpolated by the B-spline curve constructed using the
for shape parameterization is that only control points located control points at the hub, mid-span, and tip.
along the curves can control the curves. When one control point
of the Bezier curve is moved vertically, the others remain fixed.
5.3 Hub Cap Installation Distance and Hub Cap
Thus, each control point is controlled independently, and all these
Ratio. The axial-flow fan model used in this work originally did
points can be considered design variables. In the present study, to
not have a hub cap. However, to reduce distortion of the inlet
limit the number of design variables, all control points are fixed
flow, a conventional hub cap having a rounded front was installed
except for control point P3, which showed the most sensitive
[27]. Here, two additional variables are selected. The hub cap in-
results among the control points during a sensitivity test. Figure
stallation distance (Dhc) is defined as shown in Fig. 6(e). The di-
6(a) shows an example of variation in the blade profile with the
ameter of the hub cap depends on Rh–t, whereas Dhc is
stagger angle according to the vertical movement of a control
independent of the other design variables. The other variable, the
point of the Bezier curve.
hub cap ratio (Rhc), is defined as the ratio of the hub cap’s axial
length (Daxial) to its radius (Dradial), as shown in Fig. 6(e).
5.2 Hub-to-Tip Ratio. The hub-to-tip ratio (Rh–t) is defined
as the ratio of the hub diameter (dh) to the tip diameter (dt); dt is 5.4 Sensitivity Tests. Sensitivity tests were conducted by
fixed to limit the size of the axial-flow fan casing, and only dh is changing the value of each variable with the other variables fixed at

101101-4 / Vol. 133, OCTOBER 2011 Transactions of the ASME

Downloaded 04 Oct 2011 to 165.246.100.201. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
 Fig. 7 Results of sensitivity tests: (a) Stagger angle at hub, (b) Stagger angle at mid-span, (c) Stagger
angle at tip, (d) Hub-to-tip ratio, (e) Hub cap installation distance, (f) Hub cap ratio

the reference values. Figure 7 shows the results of the sensitivity test. On the basis of the results of the sensitivity test, design
tests for each variable. The efficiency is sensitive to the stagger variables for the optimizations are selected from among the geo-
angles fm and ft at mid-span and tip, respectively. However, the metric variables tested. The first MOPT process is conducted
stagger angle (fh ) at the hub does not greatly affect the efficiency, as with three design variables, stagger angles fh; , fm , and ft and
shown in Fig. 7(a). The minimum hub-to-tip ratio (Rh–t), as shown the second MOPT process is conducted with five design
in Fig. 7(d), is restricted owing to the mechanical constraint dis- variables, Rh–t, Dhc, Rhc, and two of the previous design variables,
cussed above. The peak efficiency is shown at minimum Rh–t, 0.235. fm and ft.
The effects of the hub cap installation distance (Dhc) and hub cap ra- The purpose of the present multiobjective optimization process
tio on the efficiency are shown in Figs. 7(e) and 7(f), respectively. is to maximize both the total efficiency (g) and the pressure rise
(Pr), which are defined, respectively, as
6 Optimization Techniques
Before the optimization process, the influence of each geomet- ðPout  Pin Þ  Q
g¼ (1)
ric variable on the efficiency was evaluated by a sensitivity sx

Journal of Fluids Engineering OCTOBER 2011, Vol. 133 / 101101-5

Downloaded 04 Oct 2011 to 165.246.100.201. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
 Table 2 Ranges of design variables Table 3 Results of ANOVA and regression analysis

(a) First MOPT (a) First MOPT

Variables Lower bound Reference Upper bound Objective functions R2 R2adj RMSE Cross-validation errors

fh (deg) 32.5 36.3 40.3 g 0.999 0.998 3.75  102 2.08  103
fm (deg) 62.7 67.9 69.0 Pr 0.961 0.938 1.48  103 5.32  103
ft (deg) 67.9 74.7 75.9
(b) Second MOPT
(b) Second MOPT g 0.935 0.927 1.94  103 2.77  103
fm (deg) 62.7 67.9 69.0 Pr 0.922 0.998 3.61  102 5.20  102
ft (deg) 67.9 74.7 75.9
Rh–t 0.235 0.294 0.304
Dhc (mm) 5  75
Rhc 0.333  3

Pr ¼ Pout  Pin (2)

where P, Q, s, and x are the total pressure, volumetric flow rate,

torque, and angular velocity, respectively. The subscripts in and
out represent the inlet and outlet, respectively. Thus, the total effi-
ciency and pressure rise are expected to be maximized simultane-
ously through the multiobjective optimization.
For design optimization, it is important to find the feasible
design space formed by ranges of design variables, which are
determined by the sensitivity tests. In each multiobjective optimi-
zation process, three or five design variables are selected, and the
design space is determined, as shown in Table 2. 20 and 45 design
points for the first and second MOPTs, respectively, are generated
within the design space with the help of LHS [28] as design of Fig. 9 Results of the first and second MOPTs
experiment. The objective functions at these design points are
evaluated by RANS analysis.
Multiobjective optimization based on evolutionary algorithms The procedure that uses the multiobjective optimization meth-
combined with a surrogate model requires many evaluations of odology to generate the global POS is shown in Fig. 8. First,
the objective functions in the search for optimal solutions. In the design variables that influence the performance are identified
present study, RSA is employed as the surrogate model and through analyses of the internal flow field of the reference model,
applied to predict the objective function values in the design and the objective functions are defined by considering the system
space. RSA is a methodology of fitting a polynomial function to performance parameters to be optimized. The objective functions
discrete responses obtained from numerical calculations. It repre- are evaluated for the design points obtained from LHS. Using
sents the association between design variables and response func- these numerical solutions, an RSA model is constructed to ap-
tions. The constructed response of a second-order polynomial proximate the objective function values for an evolutionary algo-
RSA can be expressed as rithm, and a hybrid MOEA is used to obtain global POSes [21].
Approximate POSes are obtained by using the real-coded fast and
elitist nondominated sorting genetic algorithm (NSGA-II) devel-
X
N X
N XX
N
f ðxÞ ¼ b0 þ b j xj þ bjj x2j þ bij xi xj (3) oped by Deb [20] for two objective functions, the efficiency and
j¼1 j¼1 i6¼j
the pressure rise. Here, “real-coded” indicates that crossovers and
mutations are performed in real space to obtain a response from
NSGA-II. The POSes are then refined by searching for a local
where b, N, and x represent the regression analysis coefficients, optimal solution for each objective function over all the NSGA-II-
the number of design variables, and a set of design variables, derived optimal solutions; the search uses sequential quadratic
respectively, and the number of regression analysis coefficients programming (SQP) [29] with NSGA-II solutions as initial
(b0, bi, etc.) is (N þ 1)  (N þ 2)=2. guesses. SQP represents the state of the art in nonlinear

Fig. 8 Multiobjective optimization methodology

101101-6 / Vol. 133, OCTOBER 2011 Transactions of the ASME

Downloaded 04 Oct 2011 to 165.246.100.201. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
 Table 4 Design variables corresponding to the clustered
optimal solutions from the global POSes

(a) First MOPT

Design variables

Designs fh (deg) fm (deg) ft (deg)

Reference 36.3 67.9 74.7

COS A 40.3 63.0 69.8
COS B 40.3 63.3 70.7
COS C 40.3 63.8 72.0
COS D 40.2 64.6 74.0
(b) Second MOPT

Design variables

Designs fm (deg) ft (deg) Rh–t Dhc (mm) Rhc

Reference 67.9 74.7 0.294 – –

COS E 63.3 71.6 0.235 5 2.544
COS F 63.7 72.0 0.235 5 2.596
COS G 64.1 72.4 0.235 5 2.652
COS H 64.7 72.9 0.235 5 2.719
COS I 65.6 73.6 0.235 5 2.785

programming methods. SQP is a generalization of Newton’s

method, which is a gradient-based optimization technique. At
each major iteration, an approximation is made of the Hessian of
the Lagrangian function using a quasi-Newton updating method.
This is then used to generate a quadratic programming sub-
problem whose solution is used to form a search direction for a Fig. 10 Performance curves of clustered optimal solutions in
line search procedure [30]. Two approaches are usually applied to the first and second MOPTs: (a) Efficiency curves, (b) Pressure
perform a local search [21]. In one approach, all the objectives are rise curves
combined into a single composite objective, and the optimum is
sought. In the other, one objective is optimized by treating the
others as equality constraints, and the process is repeated for all
objectives. 7 Results and Discussion
In the present study, the first objective is optimized, whereas
the second objective is fixed. Then the local search is repeated To perform multiobjective optimizations of an axial-flow ven-
for the second objective by keeping the first objective at a fixed tilation fan, the RSA model is trained for both objective func-
value. This process produces two new sets of optimal solutions, tions, the efficiency and the pressure rise, by using RANS
which are then merged with the NSGA-II solutions. From analysis results at the design points selected by LHS. In the RSA
these solutions, dominated solutions are discarded, and then dupli- model, an analysis of variance (ANOVA) and a regression analy-
cate solutions are removed to produce the global POSes. Subse- sis including t-statistics [1] are conducted to measure the uncer-
quently, a local search is conducted to improve the quality of the tainty in a set of coefficients in a polynomial. The values of R2
POSes. and R2adj for second-order curve-fitting and the root mean square

Table 5 Optimization results

(a) First MOPT

MOEA Prediction RANS Prediction Error (%) Increment

Designs g Pr (Pa) g Pr (Pa) g Pr g Pr (Pa)

Reference – – 0.7421 85.433 – – – –

COS A 0.7206 110.919 0.7157 110.264 0.68 0.59 0.0264 24.831
COS B 0.7328 110.111 0.7305 109.001 0.31 1.01 0.0116 23.568
COS C 0.7447 108.007 0.7458 107.991 0.15 0.01 0.0037 22.558
COS D 0.7539 102.522 0.7536 102.538 0.04 0.02 0.0115 17.105

(b) Second MOPT

Reference – – 0.7421 85.433 – – – –
COS E 0.7349 111.614 0.7395 112.239 0.63 0.56 0.0026 26.806
COS F 0.7423 111.306 0.7459 111.752 0.49 0.40 0.0038 26.319
COS G 0.7495 110.526 0.7537 110.701 0.56 0.16 0.0116 25.268
COS H 0.7567 108.764 0.7577 108.781 0.13 0.02 0.0156 23.348
COS I 0.7622 105.077 0.7603 104.803 0.25 0.26 0.0182 19.370

Journal of Fluids Engineering OCTOBER 2011, Vol. 133 / 101101-7

Downloaded 04 Oct 2011 to 165.246.100.201. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
 Fig. 12 Axial velocity contours at trailing edge (unit: m=s): (a)
Reference, (b) COS A, (c) COS I
Fig. 11 Axial velocity contours at leading edge (unit: m=s): (a)
Reference, (b) COS A, (c) COS I

error (RMSE) for the RSA model are listed in Table 3. Here, R2 g1 ¼  0:725 þ 0:0019x1  0:0831x2  0:054x3
and R2adj indicate the correlation coefficient in least squares sur- þ 0:0081x1 x2 þ 0:0001x1 x3 þ 0:0683x2 x3
face fitting and the adjusted correlation coefficient, respectively.
The values of R2adj for each objective function are 0.998 and  0:0021x21 þ 0:0589x22 þ 0:0424x23 (4)
0.938 in the first MOPT and 0.927 and 0.998 in the second Pr1 ¼  73:0067  0:3909x1  51:3667x2  31:3927x3
MOPT, respectively. These values are reliable according to the
0.9 < R2adj < 1.0 range suggested by Giunta [31] for accurate pre- þ 2:5478x1 x2 þ 1:5488x1 x3 þ 12:0728x2 x3
diction of RSA models. Leave-one-out cross-validation (CV)  2:1864x21 þ 21:0517x22 þ 12:0397x23 (5)
[32] is also conducted to assess the accuracy of the RSA models.
Although it is uncertain how well the CV is correlated with the g2 ¼  0:7392 þ 0:0157x1  0:0634x2  0:0321x3
accuracy of the RSA model, the estimation of the generalization  0:0093x4 þ 0:0094x5  0:02x1 x2  0:0109x1 x3
errors is nearly unbiased, as it takes into account the CV of the
RSA model at every design point. The estimated CV errors are þ 0:0057x1 x4 þ 0:0027x1 x5 þ 0:0354x2 x3
shown in Table 3. The functional forms of both objective values  0:0029x2 x4  0:0052x2 x5 þ 0:0002x3 x4
for the first and second MOPTs that are obtained by the RSA  0:0009x3 x5 þ 0:0015x4 x5 þ 0:0023x21 þ 0:0626x22
model can be expressed in terms of normalized design variables,
respectively, as þ 0:0215x23 þ 0:0056x24  0:0055x25 (6)

101101-8 / Vol. 133, OCTOBER 2011 Transactions of the ASME

Downloaded 04 Oct 2011 to 165.246.100.201. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
 Fig. 14 Static and total pressure distributions at trailing edge:
(a) Static pressure, (b) Total pressure

Fig. 9, the POSes of the second MOPT have better results for both
objectives than those of the first MOPT. The reason is thought to
Fig. 13 Static pressure distributions on blade surfaces: (a) At
15% span, (b) At 50% span, (c) At 85% span be that the second MOPT has better combinations with more
diverse design variables than the first MOPT, which used only the
stagger angles.
Pr2 ¼  80:0527 þ 5:1242x1  47:9398x2  18:8794x3 A trade-off analysis shows an obvious correlation between the
efficiency and the pressure rise: Higher efficiency is obtained at a
 1:9081x4 þ 2:1579x5  4:5287x1 x2  2:1445x1 x3
lower pressure rise and vice versa. Therefore, when a preliminary
þ 1:4495x1 x4 þ 0:6066x1 x5 þ 7:1442x2 x3 design is required, designers can meet their own design require-
 0:4171x2 x4  0:9314x2 x5  0:1732x3 x4 ments with regard to both the efficiency and the pressure rise by
the POSes obtained from a hybrid MOEA. Meanwhile, the POSes
 0:0447x3 x5 þ 0:1612x4 x5 þ 1:5244x21 þ 22:7559x22 are grouped through K-means clustering [33] to determine the rep-
þ 6:5992x23 þ 1:4945x24  1:0429x25 (7) resentative solution for a group of solutions. In the present study,
four and five representative clusters are formed for the first
where the subscripts, 1 and 2 indicate the first and second MOPTs, and second MOPTs, respectively, and distributed equally among
respectively. the POSes. They are also numerically evaluated through RANS
analysis. The values of the design variables corresponding to the
7.1 Results of the Multiobjective Optimization. A hybrid clustered optimal solutions (COSes) in each optimization are
MOEA based on the response constructed by the RSA is used to shown in Table 4. Table 4(a) shows that two design variables, fm
obtain the global POSes through the real-coded NSGA-II. In the and ft, tend to increase, but, fh has almost the same optimized
two multiobjective optimizations, namely, the first and second value at every COS. In Table 4(b), three design variables, fm, ft,
MOPTs, the real-coded NSGA-IIs are invoked to obtain well- and Rhc, tend to increase, while Rh–t and Dhc have the same opti-
spread approximate POSes with 250 generations and 100 popula- mized value at every COS. Rh–t and Dhc are believed to be con-
tions, and crossover and mutation probabilities of 0.95 and 0.25, strained by the mechanical manufacturing conditions discussed
respectively, are chosen. The crossover and mutation parameters above.
are set to 10 and 50, respectively. Here, these parameters are The values of the objective functions at the COSes and the
adjusted one by one to suit the nature of the problem. objective function values calculated by RANS analysis are shown
Figure 9 represents the results of the first and second MOPTs, in Table 5, along with the reference values. The errors of the
consisting of the global POSes generated by a hybrid MOEA, the hybrid MOEA predictions compared to the RANS calculations are
evaluations at the design points, and the reference values. As both also shown. The maximum errors of the MOEA predictions for
objectives are to be maximized, the POSes show a convex shape. the objective functions at the first and second MOPTs are 1.01
Every POS is thought to have its own optimized conditions for the and 0.63%, respectively. These results represent quite accurate
objective functions. Extreme ends of the POSes represent a pair of predictions. RANS calculations of the objective functions for both
the highest value of one objective function and the lowest value of the first and second MOPTs are compared with the reference val-
the other objective function. Any improvement of one objective ues. In particular, the COSes except for A, B, and E are superior
function leads to the deterioration of the other. As shown in to the reference values for both the efficiency and the pressure

Journal of Fluids Engineering OCTOBER 2011, Vol. 133 / 101101-9

Downloaded 04 Oct 2011 to 165.246.100.201. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
 Fig. 15 Static pressure contours on the passage surface (unit: Pa): (a) At 15% span, (b) At 50% span, (c) At 85% span

rise. The results of COSes D and I represent the highest efficiency in the overall flow-rate range, as shown in Fig. 10(b). COSes A
values for the MOPTs. The increments are 0.0115 and 0.0182 for and E show higher pressure rises than COSes D and I over most
the first and second MOPT, respectively, relative to the efficiency of the flow-rate range. Therefore, the off-design performance is
of the reference model without a hub cap, 0.7421. The results of mostly improved by the present optimizations except in the region
COSes A and E show the highest pressure rise values, 110.264 of low flow rate.
and 112.239 Pa, respectively, among the results of each MOPT,
but the pressure rise of COS A is less than those of COSes E, F,
and G in the second MOPT. The second MOPT is considered to 7.3 Analyses of the Internal Flow Fields. To find the main
yield better results because it uses a more diverse range of design factors responsible for improving the fan’s performance, the inter-
variables than the first MOPT. nal flow fields of representative clusters A and I are compared
with the reference in Figs. 11–15. As shown in Fig. 9 and Table 5,
COS A has the highest pressure rise among the COSes in the first
7.2 Off-Design Performance. To evaluate the performance MOPT, whereas it has the lowest efficiency among all of the
of the fans at off-design points, flow analyses were also performed COSes in the first and second MOPTs. On the other hand, COS I
at the several off-design points for the reference and the COSes at has the highest efficiency among all of the COSes in the first and
the extreme ends of the POSes, A, D, E, and I. Figure 10(a) shows second MOPTs. Thus, COSes A and I are selected as the represen-
the efficiency curves of the COSes in the first and second MOPTs. tative high-pressure and high-efficiency designs of the axial-flow
COSes D and I both show a wide range of high uniform effi- fan, respectively.
ciency, which is desirable for stable economic operation; the effi- Figures 11 and 12 show the axial velocity contours at the
ciencies for these COSes are higher than 0.7 throughout the blade’s leading and trailing edges, respectively. The inflows into
overall range of the flow rate in Fig. 10(a). However, they have the blade passage in COSes A and I show reduced axial velocity
lower efficiencies than the reference in the low-flow-rate region, contours near the center of the blade passage, as shown in Fig. 11.
in which the flow rate is less than 55 m3=min. COSes A and E The stability of inflows are expected to be enhanced by the opti-
yield efficiencies that are even higher than those of the other two mized stagger angles. Among the three COSes considered, COS A
COSes near the free-delivery region but show rapid decreases as has the most uniform inflow contours of the axial velocity compo-
the flow rate decreases and much lower efficiencies in the low- nent from hub to tip owing to the decrease in the axial velocity
flow-rate region than the other COSes, including the reference. component in the tip region. The hub diameter is reduced with the
All the COSes have remarkably higher values of the pressure rise decrease in Rh–t in COS I, reducing the axial velocity near the

101101-10 / Vol. 133, OCTOBER 2011 Transactions of the ASME

Downloaded 04 Oct 2011 to 165.246.100.201. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
 hub. The outflows from the blade passage of COSes A and I have ning (ITEP), that is funded by the Ministry of Knowledge
similar stable axial velocity contours with a radial velocity gradi- Economy.
ent. For COS I, the axial velocity near the hub region is slightly
reduced owing to the reduced Rh–t, as shown in Fig. 12. Nomenclature
Figure 13 shows the static pressure distributions in the stream-
wise direction at blade spans of 15, 50, and 85% on the suction ANN ¼ artificial neural network
and pressure surfaces for the reference and COSes A and I. COS ANOVA ¼ analysis of variance
A, which features the highest overall pressure rise, shows the larg- CFD ¼ computational fluid dynamics
est pressure difference between the suction and pressure surfaces COS ¼ clustered optimal solution
in most of the spanwise locations along the blade spans except for CV ¼ cross-validation
15%, where it shows the largest pressure difference only at nor- Daxial ¼ axial length of the hub cap
malized streamwise locations of less than about 0.5. At a blade Dhc ¼ hub cap installation distance
span of 15%, COS I shows the highest pressure distribution on the dh, s and t ¼ hub, shaft, and tip diameters, respectively
SS among the three COSes, as shown in Fig. 13(a). The reason is Dradial ¼ radius of the hub cap
the installation of the optimized hub cap, even though the radius FVM ¼ finite volume method
of the hub of COS I is relatively small owing to the decrease in GA ¼ genetic algorithm
Rh–t relative to the other COSes. LE ¼ leading edge
Figure 14 shows the distributions of the static and total pres- LHS ¼ latin hypercube sampling
sures in the spanwise direction at the TE. In Fig. 14(a), the static lr ¼ root length
pressures for COSes A and I are similar, especially near the max ¼ maximum value
mid-span, and these values are much larger than those for the ref- MOEA ¼ multiobjective evolutionary algorithm
erence at most of the spanwise locations. However, the total pres- MOPT ¼ multiobjective optimization
sure for COS I becomes lower than that for the reference in the N¼ number of design variables
range of normalized spanwise locations of less than 0.1, as shown NSGA-II ¼ fast and elitist nondominated sorting genetic
in Fig. 14(b). Both the COSes show values much larger than the algorithm
reference except at spanwise locations of less than about 0.1. As P¼ total pressure
discussed earlier, this is the effect of the hub cap installation. POS ¼ pareto-optimal solution
Figure 15 shows the static pressure contours for the reference Pr ¼ pressure rise
and for COSes A and I at blade spans of 15, 50, and 85%. The PS ¼ pressure surface
three COSes show similarly shaped contours between the blades, P1–5 ¼ control-point-generated Bezier curve
but COSes A and I show static pressures in the blade passage that Q¼ volumetric flow rate
are much larger than the reference at blade spans of 50 and 85%, R2 ¼ correlation coefficient in least squares surface
as shown in Figs. 15(b) and 15(c). The variations in the stagger fitting
angles at the mid-span and tip are thought to contribute to increas- R2adj ¼ adjusted correlation coefficient
ing the static pressure. RANS ¼ Reynolds-averaged Navier-Stokes
ref ¼ reference model
Rhc ¼ hub cap ratio
8 Concluding Remarks Rh–t ¼ hub-to-tip ratio
Optimizations of an axial-flow ventilation fan with a conven- RMS ¼ root mean square
tional hub cap were performed by using a hybrid MOEA and RSA RMSE ¼ root mean square error
surrogate modeling through 3-D RANS analysis. The multiobjec- RSA ¼ response surface approximation
tive optimization processes were performed twice to understand SQP ¼ sequential quadratic programming
the coupled effects of diverse design variables associated with SS ¼ suction surface
enhanced aerodynamic performances. The first optimization pro- SST ¼ shear stress transport
cess was conducted with three design variables defining the stag- TE ¼ trailing edge
ger angles at the hub, mid-span, and tip, and the second x¼ set of design variables
optimization process was conducted with five design variables a¼ angle between the rotational axis and a tangent
defining the hub-to-tip ratio, hub cap installation distance, hub cap of the camber line
ratio, and the stagger angles at the mid-span and tip. Before the b¼ regression analysis coefficient
design variables were selected, sensitivity tests for the diverse var- fh; ; fm , and ft ¼ stagger angles at hub, mid-span, and tip,
iables were conducted. The relative errors of the objective func- respectively
tions predictions by the RSA model were generally less than 1.0% g¼ total efficiency
relative to the RANS calculations. The highest increments in the s¼ torque
efficiency through the optimizations were 0.0115 and 0.0182 for x¼ angular velocity
the first and second optimizations, respectively, compared with 3-D ¼ three-dimensional
the efficiency of the reference model without a hub cap, 0.7421.
The highest values of the pressure rise for the first and second Subscripts
optimizations were 110.264 and 112.239 Pa, respectively, which
are much larger than the reference value, 85.433 Pa. However, the in ¼ inlet
improvements in performance are attributed in part to the installa- out ¼ outlet
tion of the hub cap. By analyzing the extreme clustered optimal ref ¼ reference shape
solutions that represent the highest-pressure and highest-
efficiency designs among the POSes, we found that the off-design References
performance is mostly improved by the present optimizations [1] Myers, R. H., and Montgomery, D. C., 1995, Response Surface Methodology:
Process and Product Optimization Using Designed Experiments, John Wiley &
except in the region of low flow rate. Sons Inc, New York, USA.
[2] Sasaki, T., and Breugelmans, F., 1998, “Comparison of Sweep and Dihedral
Effects on Compressor Cascade Performance,” ASME J. Turbomach., 120(3),
Acknowledgment 454–464.
[3] Beiler, M. G., and Carolus, T. H., 1999, “Computation and Measurement of the
This research was supported by a grant (No. 10031771) from Flow in Axial Flow Fans With Skewed Blades,” ASME J. Turbomach., 121(1),
the Korea Institute of Industrial Technology Evaluation and Plan- 59–66.

Journal of Fluids Engineering OCTOBER 2011, Vol. 133 / 101101-11

Downloaded 04 Oct 2011 to 165.246.100.201. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
 [4] Larocca, F., 2008, “Multiple Objective Optimization and Inverse Design of Axial [19] Kim, J. H., Choi, J. H., Husain, A., and Kim, K. Y., 2010, “Multi-Objective
Turbomachinery Blade,” AIAA J. of Propulsion and Power, 24(5), 1093–1099. Optimization of a Centrifugal Compressor Impeller Through Evolutionary
[5] Oyama, A., Liou, M. S., and Obayashi, S., 2002, “Transonic Axial-Flow Blade Algorithms,” Inst. Mech. Eng., Part A: J. Power and Energy, 224(5),
Shape Optimization Using Evolutionary Algorithm and Three-Dimensional Nav- 711–721.
ier-Stokes Solver,” 9th AIAA=ISSMO Symposium and Exhibit on Multidiscipli- [20] Deb, K., 2001, Multi-Objective Optimization Using Evolutionary Algorithms,
nary Analysis and Optimization, Atlanta, USA, AIAA paper No. 2002-5642. John Wiley & Sons Inc, Chichester, UK.
[6] Chen, N. X., Zhang, H. W., Xu, Q., and Huang, W. G., 2009, “Application of [21] Goel, T., Vaidyanathan, R., Haftka, R. T., Shyy, W. Queipo, N. V., and Tucker,
Simple Gradient-Based Method and Multi-Section Blade Parameterization K., 2007, “Response Surface Approximation of Pareto Optimal Front in Multi-
Technique to Aerodynamic Design Optimization of a 3D Transonic Single Objective Optimization,” Comp. Meth. in App. Mech. and Eng., 196(4–6), pp.
Rotor Compressor,” ASME GT2009-59783. 879–893.
[7] Shu, X., Gu, C., Wang, T., and Yang, B., 2009, “Optimum Design and Experi- [22] Lee, C., and Kil, H. G., 2010, “A Computerized Design System of the Axial
mental Study of a Very Low-Specific-Speed Centrifugal Blower Blade,” ASME Fan Considering Performance and Noise Characteristics,” J. Fluid Mach. (in
GT2009-59823. Korea), 13(2), 48–53.
[8] Wang, X. F., Xi, G., and Wang, Z. H., 2006, “Aerodynamic Optimization [23] Lee, C., and Kil, H. G., 2009, “A Preliminary Axial Fan Design Method With
Design of Centrifugal Compressor’s Impeller With Kriging Model,” Inst. the Consideration of Performance and Noise Characteristics,” The 10th Asian
Mech. Eng., Part A: J. Power and Energy, 220(6), 589–597. International Conference on Fluid Machinery, Kuala Lumpur, Malaysia,
[9] Sørensen, D. N., Thompson, M. C., and Sørensen, J. N., 2000, “Toward AICFM0095.
Improved Rotor-Only Axial Fans-Part II: Design Optimization for Maximum [24] ANSYS CFX-11.0, 2006, ANSYS CFX-Solver Theory Guide, ANSYS Inc.
Efficiency,” ASME J. Fluids Eng., 122(2), 324–329. [25] Menter, F. R., 1994, “Two-Equation Eddy-Viscosity Turbulence Models for
[10] Kipouros, T., Jaeggi, D. M., Dawes, W. N., Parks, G. T., Savill, A. M., and Engineering Application,” AIAA J., 32(8), 1598–1605.
Clakson, P. J., 2008, “Biobjective Design Optimization for Axial Compressors [26] Bardina, J. E., Huang, P. G., and Coakley, T. J., 1997, “Turbulence Modeling
Using Tabu Search,” AIAA J., 46(3), 701–711. Validation,” 28th AIAA Fluid Dynamics Conference, Snowmass Village, USA,
[11] Kim, J. H., Choi, J. H., and Kim, K. Y., 2010, “Surrogate Modeling for Optimization AIAA-1997-2121.
of a Centrifugal Compressor Impeller,” Int. J. Fluid Mach. and Syst., 3(1), 29–38. [27] Jang, C. M., Choi, S. M., and Kim, K. Y., 2008, “Effects of Inflow Distortion
[12] Samad, A., Kim, K. Y., Goel, T., Haftka, R. T., and Shyy, W., 2008, “Multiple due to Hub Cap’s Shape on the Performance of Axial Flow Fan,” J. Fluid Sci.
Surrogate Modeling for Axial Compressor Blade Shape Optimization,” J. Prop. Tech., 3(5), 598–609.
and Power, 24(2), 302–310. [28] McKay, M. D., Beckman, R. J., and Conover, W. J., 2000, “A Comparison of
[13] Dennis, B. H., Egorov, I. N., Han, Z. X., Dulikravich, G. S., and Poloni, C., Three Methods for Selecting Values of Input Variables in the Analysis of Out-
2000, “Multi-Objective Optimization of Turbomachinery Cascades for Mini- put From a Computer Code,” Technometrics, 42(1), 55–61.
mum Loss, Maximum Loading, and Maximum Gap-To-Chord Ratio,” 8th [29] MATLABV R , 2004, The Language of Technical Computing, Release 14, The

AIAA=NASA=USAF=ISSMO Symposium on Multidisciplinary Analysis and Math Work Inc.

Optimization, Long Beach, USA, AIAA 2000-4876. [30] Hock, W., and Schittkowski, K., 1983, “A Comparative Performance
[14] Oksuz, O., and Akmandor, I. S., 2008, “Axial Turbine Blade Aerodynamic Opti- Evaluation of 27 Nolinear Programming Code,” Computing, 30(4),
mization Using a Novel Multi-Level Genetic Algorithm,” ASME GT2008-50519. 335–358.
[15] Arabnia, M., and Ghaly, W., 2010, “On the Use of Blades Stagger and Stacking [31] Giunta, A. A., 1997, “Aircraft Multidisciplinary Design Optimization Using
in Turbine Stage Optimization,” ASME GT2010-23399. Design of Experimental Theory and Response Surface Modeling Methods,”
[16] Lotfi, O., Teixeira, J. A., Ivey, P. C., Sheard, G., and Kinghorn, I. R., 2005, Ph.D. thesis, Department of Aerospace Engineering, Virginia Polytechnic Insti-
“Aerodynamic Optimization of Industrial Fan Blade Cascades,” ASME tute and State University, Blacksburg, VA.
GT2005-68385. [32] Queipo, N. V., Haftka, R. T., Shyy, W., Goel, T., Vaidyanathan, R., and Tucker,
[17] Benini, E., 2004, “Three-Dimensional Multi-Objective Design Optimization of P. K., 2005, “Surrogate-Based Analysis and Optimization,” Prog. Aerospace
a Transonic Compressor Rotor,” J. Prop. and Power, 20(3), 559–565. Sci., 41(1), 1–28.
[18] Lian, Y., and Liou, M. S., 2005, “Multiobjective Optimization Using Coupled [33] JMP, 2005, The Statistical Discovery Software, Version 6.0.0, SAS Institute
Response Surface Model and Evolutionary Algorithm,” AIAA J., 43(6), 1316–1325. Inc., Cary, NC, USA.

101101-12 / Vol. 133, OCTOBER 2011 Transactions of the ASME

Downloaded
View publication stats 04 Oct 2011 to 165.246.100.201. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm