Abdus Samad

Turbomachinery Design and

Optimization
A systematic approach for turbomachinery blade
design, flow analysis and performance
optimization

~;L1 _ 13~~~A~ " __, I ,,__ , , t:0 _AB LISHME N

~/-\N I _ .~_0 : -~ i.:.. - j6 091
1 EC i-i:,,:C AL INFORMAT ~ ON CENTRE

i ,~-.cc. r')C _ _ ll5.1._ ___

! [H" - ~'
~-
~ - --~~ .._----.... - -- . --_.-
... ).' ~

ILAP ILAMflE.Rl AeadelililJi ~ P"b~iJsh~ng

Impressum/lmprint (nur fur Deutschland/only for Germany)
Bibliografische Information der Deutschen Nationalbibliothek: Die Deutsche
Nationalbibliothek verzeichnet diese Publikation in der Deutschen Nati ona lbi bllografie; ACK OWLEDGEMENTS
detaillierte bibliografische Daten sind im Internet uber http ://dnb.d-nb.de abrufbar.
Aile in diesem Buch genannten Marken und Produktnamen unterl iegen warenzeichen-,
marken- oder patentrechtlichem Schutz bzw. sind Warenzeichen oder eingetragene My most sincere acknowledgements go to advisor and chair of my advisor
Warenzeichen der jeweiligen Inhaber. Die Wiedergabe von Marken, Produktnamen,
Gebrauchsnamen , Handel snamen, Warenbezeichnungen u.s.w. in diesem Werk berecht igt committee, Prof. Kwang-Yong Kim, for his remarkable wisdom and val uable
auch ohne besondere Kennzeichnung nicht zu der Annahme, dass solche Namen irn Sinne
der Warenzeichen- und Markenschutzgesetzgebung als frei zu betrachten ware n und guidance. He has open door policy welcoming every idea and consistent support
daher von jederrnann benutzt werden durften .
in developing my career and motivation. I always believe that he has subtle but
Coverbild : www.ingimage.com
strong ways of encouraging all of his students and I have a lways felt lucky to be
Verlag: LAP LAMBERT Academic Publishing GmbH & Co. KG
Heinrich-B6cking-Str. 6-8, 66121 Saarbrucken, Deutsc hland
one of them.
Telefon +496813720-310, Telefax +49 6813720-3109 I am grateful to many individuals for their support in this research effort. The
Email : [email protected]
greater part of this work was made possible by the instruction of my teachers
Approved by: Incheon, Inha University, Thesis, 2008
and by the love and support of my family and friends. Sincere thanks to all of
Herstellung in Deutschland:
Schaltungsdienst Lange o .H.G., Berlin them.
Books on Demand GmbH, Nordersted t
Reha GmbH, Saarbrucken
I would like to thank my lab colleagues for their tremendous support in all
Amazon Distribution GrnbH, Leipzig aspect of my research work. I want to convey special thanks to my lab seniors;
ISBN : 978-3-8484-1612-7
Dr. C.M. lang (Korean Institute of Construction Technology) and Dr. l.H. Choi
Imprint (only for USA, GB)
Bibliographic information published by the Deutsche Nationalbibliothek: The Deutsche (Samsung Techwin Co., Ltd.).
Nationalbibliothek lists this publication in the Deutsche Nationalbibliografie; detailed
bibliographic data are available in the Internet at http ://d nb .d-nb.de . I would like to thanks to Prof. Wei Shyy (University of Michigan, USA) and
Any brand names and product names mentioned in th is book are subject to trademark,
Prof. R.T. Haftka and T. Goel (University of Florida, USA) for being a part of
brand or patent protection and are trademarks or registered trademarks of their respective
holders. The use of brand names, product names, cornmon names, trade names, product
my research work.
descriptions etc. even without a particular marking in this works is in no way to be
construed to mean that such names may be regarded as unrestricted in respect of
trademark and brand protection legislation and could thus be used by anyone.
AbdusSamad
Cover image: www.ingimage.com
August, 2008
Publisher: LAP LAMBERT Acad emic Publishing GmbH & Co . KG
Heinrich-B6cking-Str. 6-8, 66121 Saarbrucken, Germany
Phone +496813720-310, Fax +49 6813720-3109
Email : [email protected]

Printed in the U.S.A.

Printed in the U.K. by (see last page)
ISBN : 978-3-8484-1612-7

Copyright © 2012 by the author and LAP LAMBERT Academic Publi shing GmbH & Co . KG
and licensors
All rights reserved. Saarbrucken 2012
 NOMENCLATU RE Greek symbols
Roman sy mbols a Sweep variable
Bult) Bezier blending function fJ Lean variable
c Regress ion coefficient y Skew variable
Cf.1 Constant J Variable normal to chord line
E Error ( Variable along chord line
F Objective function ,;,K Constants of PBA surrogate model
F w( = I]ad+ W,P ,otal.er;/P ,olal. illlel . . . (p
Adiabatic efficiency = 'v,,,I.,·,,,
/P
"""1,;,,1,,
)' -1)" _ 1
I]ml
T,o/(d,( .I·it / T,(I/lIUlllel - I
k Ratio of specific heats (Equation 2.7)
Turbulent kinetic energy (Equation 2.5) I:: Turbulent e nergy dissipation rate

P Pressure f.11 Turbulent or eddy viscosity

Pi Coordinates of control points P Fluid density

p, q, r, S Coefficients of cubic polynomials (Equation 2. 13) T Shear stress, turbulence intensity

k'
Rr Turbulent Reynolds number, Rr = L
p c: Subscripts
S(x) Polynomia l segments EJJ Efficiency
T Temperature Expt Experimental
Normalized parameter ofa curve h, m and t Hub, mid span and tip, respectively
u, v, w velocity components (Equation 2. l to 2.5) inlet Inlet
w Weight (Equation 3.26) outlet Outlet
WI We ighting factor opt Optimized
x,y Spatial coordinates P Pressure ratio
+
y (PYIlI1 )/(TIIP) 1/2 Ref Reference
T Temperature ratio

total Total
wf Weighting factor
w Value at the wall

II \11
 Table captions
Abbreviations Table 1.1 Literatures on turbo machinery blade shape optimization
ANN Artificial neural network Table 2. 1 Design specifications of NASA rotor 37
CV Cross va lidation Table 4.1 Design ranges of blade sweep, lean, and skew
DOE Design of experiments Table 4.2(a) Weights for PBA model to construct weighted average model for
GA Genetic algorithm
F £fJ.
GMSE Genera li zed mean square error Table 4.2(b) Optimal designs suggested by various surrogates and
KRG Kriging corresponding predicted RANS results for objective F £ff-
LE Lead ing edge Table 4.3(a) We ights for PBA model to construct weighted average model for
LHS Latin hypercube sampling objective function, Fr.
MOEA Multi-Objective evolutionary algorithm Table 4.3(b) Optimal designs suggested by various surrogates and
MOGA Multi-Objecti ve ge netic algo rithm corresponding predicted RANS results for objective Fr.
NSGA Non-dominated sorting genetic algorithm Table 4.4(a) Weights for PBA model to construct weighted average model for
PBA PRESS based averaging objective function, F p .
PDE Partia l differential equation Table 4.4(b) Optimal designs suggested by various surrogates and
POF Pareto optimal front corresponding predicted RANS results for objective F p.
PRES S Predicted error sum o f squares Table 4.5 Root-mean-square averaged errors in predictions of the surrogates at
PS Pressure surface 12 optimal points.
RANS R eynolds average Navier Stokes Table 4.6 Design variables and ranges
RBNN Radial basis neural network Table 4.7 Results of optimizations: (a) Design variables and (b) Objective
RMS Root mean square functions
RS A Res ponse surface approx imation Table 4.8 Design space
RSM Response surface method Table 4.9 Res ul ts of optimizations: (a) Des ign variables and (b) Objective
SA Simp le averaging functions
SS Suction surface Table 4. 10 Weights assigned to surro gates to co nstruct PBA mode l: (a) For
SQP Sequential quadratic programming '1 ad, opl blade, (b) For F p.opl blade and (c) For F ulopi blade
TE Trailing edge

IV
V
 Fig. 4 .S Results of multiple-objectives optimization: (a) Design variables and
(b) Objective functi on values
Figure captions Fig. 4.6 Sensiti vity analyses for optimum shape by RSM
Fig. 2. 1 NASA rotor 37 Fig. 4. 7 Limiting streamlines on the blade sucti on surface: (a) Refere nce
Fig. 2.2 Meridional view of Rotor 37 blade and (b) Efficiency Optimized blade
Fig. 2.3 Computationa l grids: (a) Perspective view and (b) Grids at hub of Fig. 4.8 Mach number contours on the planes of 10, 7S and 90% span
LE and TE (interval of contour lines =0.1): (a) 90% span, (b) 7S% span and (c) 10%
Fig. 2.4 Defi nition of blade sweep: (a) Top view and span
(b) Side view Fig. 4.9 Temperature contours on surfaces of optimum blades: (a) Reference
Fig. 2.S Definition of blade lea n (top view) blade, (b) Efficiency optimized blade, (c) Pressure ratio optimized blade
Fig. 2.6 Definition of blade skew (front view) and (d) Temperature ratio optimized blade.
Fig. 2.7 Computational domain Fig. 4. 10 Pressure coefficient contours on surfaces of optimum blades: (a)
Fig. 2.8 Convergence plots: (a) Residua l history a nd Reference blade, (b) Efficiency optimized blade, (c) Pressure ratio
(b) Imba lances optimized blade and (d) Temperature ratio optimized blade.
Fig. 2.9 Definitions of thickness variables Fig. 4 . 11 Axial velocity norma lized by sound velocity (interval between
Fig. 3.1 Optimization procedure contours = O.OS): (a) 30 percent chord from blade leading edge and (b)
Fig. 3.2 Design of experiments: (a) Full factorial design and Blade trailing edge
(b) Latin hyper cube sampling Fig. 4.12 Distribution of vorticity on the quasi-orthogonal planes to the
Fig. 3.3 Radial bas is network (single ne uron) leakage vortex and leakage streamlines (interval between contours= 1.0):
Fig. 3.4 Rad ial basis function (a) Reference and (b) Optimum by RBNN
Fig. 3.S Multiobjective optimization procedure Fig. 4.13 Optimum blade shape
Fig. 4 . 1 Spanwise di stribution of effici ency Fig. 4.14 Pareto optimal design
Fig. 4.2 Adiabatic efficienc y acco rding to normalized mass flow rates Fig. 4. IS Spanwise effi c ie ncy distributions
(vertical solid line: design flow rate) Fig. 4.1 6 Spanwise total temperature ratio distributions
Fig. 4.3 RANS calculations at the optimum points predicted by different Fig. 4. 17 Spanwise total press ure ratio distributions
surrogates Fig. 4. 18 Spanwise Mach number distributions
Fig. 4.4 Errors in predictions of surrogates at different po ints in design space: Fig. 4.19 Mac h number contours at 80% span: (a) Reference blade, (b)
(a)FE{f, (b) FTand (c) F" Efficiency optimized blade and (c) Tota l pressure ratio optimized blade

VI V II
Fig. 4.20 Stream lines: (a) Reference blade, (b) Efficiency optimized blade TABLE OF CONTENTS
and (c) Pressure optimized blade
Fig. 4 .21 Result va lidations with experimental data ACKNOWLEDGEMENTS
Fig. 4 .22 Variables and objective function va lues with di fferent weighting NOMENCLATURES .... ......... ...... .. ..... ............. ............. ..... .. ... .. ........... ........ 11
fa ctors LIST OF TABLES
Fig. 4 .23 Reference and optimized blade shapes LIST OF FIGURES ... . . . Vl

Fig. 4 .24 Mach number contours at diffe rent span of blade: (a) 10% span, (b) CHAPTERS
50% span and (c) 90% span I. INTRODUCTION
Fig. 4.25 Mach number contours on blade suction surfaces : (a) Ref, I-I Literature survey
(b )I]ad, opl ,(c) F P.Opl and (d) F wfol'l
1-2 Objective and scope ... ....... .. .............. ......... ........... .. ... .... .... .......... .. 21
Fig . 4.26 Stream lines on blade surfaces: (a) Ref; (b) I] ad,opI> (c) F p.ol'l and (d) 2. NUMERICAL FORMULATION OF TURBOMACHINERY BLADE ... 23
F>~/:opl 2-1 Transonic axial compressor rotor
Fig. 4.27 Pressure contour at blade surface: (a) Ref, (b) 1]0'/.01'1 , (c) F p,upi and 2-2 Reynolds-Averaged Navier-Stokes Analysis ......................... 24
(d) F "f,opl
2-2.1 Zero-equation or algebraic eddy viscosity mode .... ........ 25
Fig. 4 .28 Blade loading: (a) 20% span, (b) 50% span and (c) 80% span. 2-2 .2 Two equation model ...... .. .... .... ... .. . .. . .............. ..... 26
2-3 Blade stacking line modification .......... .. ................ .................. ....... 28
2-3.1 Numerical analysis method .... ......... .. .. ......... ........ ............. 28
2-3.2 Objective functions and design variab les ............. .... .. ...... 30
2-4 Blade stacking line and thickness modification .... .... ......... .... .... .... ~
2-4. 1 Numerical analysis method
2-4.2 Objective functions and design variable ........................... 38
3. OPTIMIZATION PROCEDURE
3-1 Introduction
3-1 . 1 Design space ........ .. .... .......... ... .. .. .. .................... ..... .. ....... .41
3-1 .2 Design of experiments ...... ........ .. .. .................... ......... ...... 43
3-1 .3 Full factorial des ign .. .. ............ ... ..... ........ .. .... ... .. .... ........ .43
3-1.4 Latin hypercube sampling ............ .. ........ ... .... .... ... .... ..... .43
3-2 Optimization procedure ...... ..... .... ... .. .... ...... .... ....... ....... .... :......... .44
3-2. 1 Surrogate approach ......... ......... ......... ..... .... .. ......... .. .. ......45
VIII
IX
 3-2.1. 1 Response surface approximation mode ...... .46
3-2. 1.2 Rad ia l basis neura l network model... ...... ......... .47 CHAPTER I
3-2.1.3 Kri gi ng model... ... ..... ....... ...... .. .................. .. ..... 50 INTRODUCTION
3-2. 1.4 PBA model... ... ............ ...... ......... .... ... .. .. ........ .... 53
3-2. 1.5 Simp le average mode l. .. ...................... .... ....... ... 55 Des ign optimizatio n of fluid mac hinery based on compute r sim ulatio n has
3-2.2 Multi-objecti ve optimization ... ..... ........ ......... ................. 55 become a reality today because of deve lopment of high speed computers.
3-2.2. 1 Weighted sum approac h..................... ... .... ........ 56 Highly co mplex fl ow patterns are be ing predicted by so lving mass, mome ntum
3-2.2.2 Hyb rid MOEA approach .. ............. ..... .. .......... ... 56 and e nergy equati ons and near accura te soluti ons at an acce ptable leve l are
3-3 Optimiza tion algori thm ................ .... ......... .............. ......... ............60 achieved. On the other hand impleme ntation of optimizati on tec hniques to
4. R ESU LT AND DISCUSS ION des ign the turbo mac hinery syste ms has reduced the computationa l and
4-1 Blade stacking line modifica tion .. ..... ...... ............. .... .. ..... .. ... ........ 63 experimenta l expenses. The performance of turbo mac hinery is directly re lated
4-1. I Validatio n o f numerical simulatio n ...... ...... ................. .... 63 to reduction of consumptio n of fu e l, mass, vibration, etc. Hence, the thes is
4-1 .2 Optimization res ults .. ....... .................. ..... ... .. ................. .. 66 describes turbo machinery bl ade shape optimizatio n targeting to enha nce the
4- 1.3 Flow characteristics with refere nce and optimum blade aerodynamic performance and guidance to the fluid mac hinery des igners
Blades .... .... ............ .. ... ................... ......... ... .............. ..... ........ ....77 by a comparati ve analys is of surrogate based approx imation mode ls for
4-1.4 Multi-obj ecti ve optimizatio n through MO EA ...... .......... 88 optimi zatio n.
4-1 .5 Flow ana lys is .. .... ....... .... ............ ....................... .......... ... . 9 1
The turbo mac hinery des igns require enhancement o f performa nce in terms of
4-2 Blade stac king line and thickness modi fica tio n .. .. ............ .... ... ... 98
thermodyna mics para meters; e ffi cie ncy, pressure ratio and structura l
4-2. 1 Optimization res ults ............ ....... ... ....... .. ......... ............... 98
parameters; no ise, vibra tion, we ight, etc. These para meters are considered to be
4-2.2 Va lidation of numerical si mulatio n .......... ... ......... ........ .98
objectives of des igns a nd the geo metric parameters as va riables; such as blade
4-2.3 Flow characteristics with re ference and optimum blades
dimensio ns etc are changed. If the blade geo metry is deduced from the
III
objecti ves, the des ign is ca lled inve rse des ign. I f the geo metrica l change is used
5. CONC LUDI NG REMARK 11 7
to predi ct the objectives, the des ign ca ll ed di rect des ign. The prese nt wo rk
REFE RENCES 11 9
follows the direct des ign procedure.
AU THO RS PUB LICAT IONS 129

BIOG RAP HICAL SKETC H 136 Historica ll y, turbo mac hinery includi ng pumps, co mpressors, turb ines, etc . is
being deve loped fro m centuries . Engineer Heron of Alexa nd ria ( I 0-70AD) fi rst
deve loped steam turb ine and that engine was grad ually imp roved to catch the
modern day's requireme nt of high perfo rmance aerospace gas turbine or large

II III
 power plant. Initia l wind tunne l tests a nd later computat iona l power gave a new design optimi zatio n is the modeling fidelity. The surrogate models include
momentum for high efficiency designs. In last decades, owing to the rapid polynomia l response surface app rox imation [9], Kri ging [10], and radial basis
developments of high-speed computers and computational fl uid dynamics neural network [II] and, in addition, we ighted ave rage models based on globa l
(CFD) theories, the comp lex three-dimensiona l flows in turbomachinery could error meas ures are also implemented in shape optimization and des ign.
be ana lyzed more easi ly. Weighted ave rage mode ling is an effecti ve approac h to empl oy multipl e
surrogates, based on the same training data, to offer approximations from
Reynolds-averaged Navier-Stokes (RANS) equations have been effective ly used
alternati ve mode ling viewpoints [12].
in various fields like heat transfer or turbomac hinery app lications. These
equations are solved fo r various flow problems and different turbulence models The bas ic procedure for optimization followed in this thesis is des ign of
are used to predict turbulence structures. Differe nt applications of experiments (DOE) [9] , numerical s imulation, construction of surrogates and
co mputational codes are being reported considering different grid reso luti on, model validation. The DOE is the sampling plan a nd importa nt part is to se lect
numerica l algorithm, and turbulence mode ls etc. Despite the differences w hich sparsely di stributed design variables in design space. The key question in this
still ex ists betwee n numerica l simul atio ns and reality, it is possible to predict step is to find the design points where the numerical simulations can be
many of the flow characteristics a nd the losses due to the non-ise ntropic performed . The RANS equations are solved at the simulation stage to find
features of the flow for example shocks, visco us layers, tip clearance effects, obj ecti ve functions values. These values are then used to construct the
passage vortices etc. The accurate flow prediction inside a transonic ax ial surrogate based approximation models. Two questio ns are of interest for
compressor rotor us ing CFD is difficult due to its extreme ly complex features: surro gate constructi on step: mode l se lection and model identification . In prese nt
three-dimensional, unsteady and vo rtica l natu re in the blade passage. Howeve r, thes is, different surrogates are constructed and tried to find these questions. The
CFD has obvious advantages over the traditional experime nta l a na lys is. CFD optimum design is searched via a gradient based search algo rithm.
he lps us to ana lyze the effects of individ ua l feature more eas ily as compared to
1-1 Literature survey
the ex peri mental method [1-7].

Flow ana lysis for NASA rotor 37, a transonic axial compressor rotor has bee n
The Mathematical and statistica l tools fo r optimization are being used in
discussed in details by Reid and Moore [1 3] and in AGARD advisory report
optimization area in single as well as multidi sc iplinary design and optimization
[14] . As the test data are available in these refere nces, many resea rchers have
area. These tools combi ned w ith numerical ana lysis methods for flow fie ld have
made efforts to validate their computatio na l codes, and also to optimize the
reduced the experimenta l expenses to design turbo machinery blades in recent
rotor. The efforts for des ign optimization inc lude single obj ecti ve as we ll as
years. With the development of CFD analysis methods, accuracy of prediction
multi-obj ecti ve and multi-disciplinary optimizations. The blade shape defi ning
for the flow becomes acceptable for the purpose of blade design [8] .
stacki ng line, airfoil shapes, etc. are modified to get better efficiency, pressure
The surro gate based approac hes are exte nsive ly used in the design of structural
ratio, surge margin, etc.
and multi-disciplinary optimizations. A major issue in surrogate model-based
I 2 3

II I
Recently, the use of sweep, lean (d ihedra l), and skew (stac king li ne in rotat iona l fan w ith skewed blade . W ith the reduction of a secondary flo w and the
direction) in axial flow compressor rotor has become a matter of interest in the thickness of a rotor wake, they could reduce a broadband no ise. Fischer, et al.
design of turbo machinery blades [15-21] . The blade shape parameters formed a [21] observed the effect of bowed stators on the performance of a compressor,
three-dimensiona l stacking line is generally introduced to reduce shock losses, and showed that the separation was reduced in the bowed stator leading to
corner separation in the blade hub, and tip c leara nce losses in transonic increase in the stagnation pressure ratio and effic iency.
compressor rotor. For example, Gallimore et a l. [15] introduced three -
A set of papers [22 -63] contributed to single and multi-objective opti mizations
dimensional blade designs using a sweep and a lean in an ax ial flow compresso r
of turbo mac hines to enhance the ir performance as presented in Table 1. 1. It has
rotor for engine. They showed that the positive lean reduced a hub corner and
been reported that the effic iency is increased due to movement of separation
tip clearance losses excepting near the mid-span region. The improve ment in the
lines towards downstream direction reducing the separation vortex, end-wall
compressor performance as well as the large reduction in the cost and
losses, etc. These papers describe the blade shape optimization considering
timescales associated with a rig test was also obta ined together wi th CFD
stacking line modification in terms of sweeping, leaning or skew ing and airfoil
calculation.
shape modification in terms thickness, leading edge, tra iling edge modification
One of the significa nt design trends is the use of aerodyna mic sweepto improve etc.
the performance and stability of transonic compressor blades. The pio nee r study
The surrogate mode ls be ing used widely in multi-disc iplinary optimizations
on blade sweep in compressors has been done by Bliss [16]. The ma in objective
should be evaluated in two important aspects; computational eco nomy that
in this study was to reduce the noise leve l induced by shock waves. Hah, et al.
requires as few data points as possible for constructing a surrogate mode l, and
[17] studied both forward- and backwa rd-swept compressor blades, and showed
accuracy in 'represe nting the characteristics of the design space. Response
that a backward-swept blade could suppress the intensity of the shock loss and a
surface approximation (RSA) method [9] which is a global optimization method
forw ard-swept blade can suppress secondary flow and tip entropy ge neration.
is recentl y introduced as a tool of design optimization for turbomac hinery, This
Watanabe and Zangeneh [1 8] reported that the blade sweep in the des ign of a
is one of the simplest surro gate mode ls to apply to optimize the system because
transo nic turbomachinery blade was an effecti ve parameter to control the
this method does not require calculation of the local sensitivity of each des ign
strength and position of the shock wave at the tip of the transonic rotors . Dento n
variable, and is able to perform tasks in para lle l easi ly, The RS A can utilize
and Xu [1 9] investigated the effects of sweep and lean on the performance of a
informatio n collected from various sources and by different too l. Thus, this
transonic fa n, and showed that the sta ll margin was significantly improved with
method is effective for both of single- and multi-disciplinary optimization
the forward swept blade altho ugh a very little change in the peak effic iency was
problems [22-26]. Neura l network [II], which is anothe r surrogate
produced by the blade sweep or lean.
approximation model, and RSA based optimization were reported by Papila et
There are a number of studies on the advantages of a skewed rotor. Cai, et al. al. [31].
[20] studied on aerodynamic and aero-acoust ic characteristics of an axial flow

4 5
 Multi-disc iplinary a nd multiple operating point optimizations at fixed rotor prob lems usua ll y known as mu lti-obj ective problems requi re simultaneous
rotational speed were reported by Pierret et al. [50]. Oya ma et a l. [49] reported consideration of all objecti ve functions to optimize the system. There are
blade profile modification with the help of B-sp line curve of NASA rotor 67 to numbers of solution methods and algorithms ava ilab le for so lvi ng multi-
increase adiabatic efficiency by 2%. C hen et a l. [56] optimized camber line, objective opti mization problems [72 -78]. [n multi-objective optimizations of
thickness distribution and stacki ng line by polynomial curve to define turbomac hinery blade, effic iency, total pressure, static press ure, pressure loss,
compressor blade and gained 1.73% imp rovement of ad iabatic effic ie nc y. weight, stress, etc . are used as obj ectives, and variab les related to camber profil e
Maximum camber location effect was studied by Chen et a l. [5 6]. Benini [63 ] and/or stacking line of blade are employed as design variables [34, 38, 42, 50,
defined blade section profiles by Bezier curve using multi-objective 57, 58, 59, 63, 64]. A multi-obj ective optimization problem consists of man y
optimization considering total pressure ratio and adiabatic effic ienc y as optimal so lutions called Pa reto-optima l so lutions; therefore a designer's aim is
objectives for design of a compressor blade. He employed camber line and to find as many optimal so lutions within the design ran ge as possible. This help s
thickness profile as design parameters. Keskin and Bestle [64] reported Bezier designer to find a g lobal Pareto -optima l front. Each des ign set corresponding to
curve parameterization of blade shape to optimize w ith Pareto optima l design. optimal so lution represents a compromise of des ign objectives. Elitist Non-
Optimization of controlled diffus ion compressor blade has been reported by dominated Sorting Genetic Algorithm (NSGA-lI) given by Deb [76] generates
Sanger [65]. Pareto optimal so lution using evolutionary algorithm.

Queipo et al. [66] and Li and Padula [67] re viewed various surrogate based
mode ls used in aerospace applicatio ns . Shyy et a l. [68] presented global
optimization model applying in rocket propul s ion design. Zerpa et al. [69]
deve loped weighted average surrogate model for a lka line-surfactant-polymer
flo oding processes design using differe nt surrogates through point wise error
estimation. Goel et a l. [1 2] developed weighted average surro gate model using
globa l data based error and concluded that weighted ave rage surrogate mode l is
more reliable pred iction method tha n indi vidua l surrogates. Goe l et a l. [70] and
Samad et a l. [71] reported on the performances of several surrogate models in
optimi z ing a heat transfer augmentat ion device and cryogenics application and
presented that weighted ave rage surrogate model developed by Goel et al. [12]
is reliab le in prediction.

Enginee ring design generally involves multiple disciplines and simultaneous

optimization of multiple objectives related to each discipline . These design

6 7

II 11I11
 Table 1. 1 Literatures on turbomac hi nery blade shape optimization Table 1.1 continued ...
In vestigators Turbo- Desi gn Objective Comments
In vestigators Turbo- Design Objective Comments machines variables functions
machines variables functions Ka mmerer 1.5 stage 6 Bez ier Total Red uction of
Jang et al. Transonic 2 variables Efficiency 1.25% et a l. [27] turbine test parameter pressure tota I press ure
[22] ax ial (sweep) enhancement increase in case And stagger loss loss was
compressor efficiency. ang le reduction achieved by
Jang and Transonic 2 variables Efficiency 2.5% radial stagger
Kim [23] axial Stacking enhancement increase in angle va ri at ion.
comp ressor line of efficiency Cai and X u Axia l fan 4 kinds of Efficiency Forward skewed
stator blade et al. [28] (ex peri mental stacking line and Noise swept blade is
Ahn and Transonic 3 va ri ables Efficiency 0.7% study) (Line typ e, leve l better than the
Kim [24] ax ial Stack ing enha ncement increase in Constant backward
compressor line effic iency inclination skewed swept
(R3 7) ang le type, blade. Predicted
Lee and Single 3 Effic iency 1.1% arc type and effi ciency rise is
Kim [25] stage ax ia l dimensional enha ncement effic ie ncy Line-a rc 4.79% and noise
flow des ign inc reased combination reductio n is
compressor variab les of and total type) 3.6 IdB .
stacking pressure rise Gummeret Transonic Sweep and Red uction Pos iti ve sweep
line across both al. [29] ax ial dihedral of end- and Positive
rotor and co mpressor wall effect dihedral red uce
stator are stator end-wall losses
increased. and increase the
C hoi et al. Auto moti ve 2 variables I. Pressure Increased operating range
[26] cooling fa n Sweep coefficie nts press ure on compresso rs
(axia l fan) angle 2. Ration of coefficients stators.
product ion and G/ H So noda et al. High 42 Exit k-w model w ith
rate of rat io and [30] performance parameters deviation MOGA for
turbulent reduced Compressors are angle optimization is
kinetic noise. used in low optimized. constraint, used.
energy (G) Re . Blade Pressure Performance
to the contours are loss, four increased.
pressure defined by thickness
head (H) B-spline. constraints

8 9
 Table 1.1 continued
Investigators Turbo- Design Objective Comments In vestigators Turbo- Des ign Objective Comments
machines variables functions machines variab les functions
Papila et al. Supersonic 0(10) Performance Using global Li et al. [35] NASA rotor Variables Efficiency 0.4%
[3 I] turbine for design based on optimi zation 67 are from enhance ment. tncrease in
rocket variab les effic ienc y and methods, sweep and peak
propulsion weight shape is compound efficie ncy
optimized. lean.
Amano and Compressor Sweep Efficiency Impro ved Buche et al. Subsonic B-spline Operating Increased
Xu [32 ] single rotor enhanceme nt effic ie ncy. [36] compressor curve range operating
blade row Forward and blade opti mi zed increment, range and
backward with 57 reduction of reduced
sweep had design aerodynamic aerodynamic
different variables losses losses.
impact to the Evo lutionary
secondary strategy is
flow s and used.
shock wave Benini and NASA rotor Sweep and Effic ie nc y 0.5-0.6%
structures . Biollo [3 7] 37, Lean enhancement efficie ncy
Keskin Axial Cubic B- Minimi zation Blade profile Transonic variables enhanced
(2006) [33] compressor spline with of loss, is opti mi zed axial blade (Backward
8 control Max imization with multi- compressor sweep
points on of working objective performa nce
press ure range Pareto better than
side optima l forw ard
so lution . sweep).
Vob et al. Axia l B-sp line Reduction of Automated Efficie ncy
[34] compressor curve wit h total pressure multi- increment
23 design loss, objective 1.3% by
parameters Maximization optimization forward
are chosen of total is used. 2- lean.
to define pressure for 2 Dimensional
aerofoi l. off design blade profile
operating is opti mi zed.
point

--_ .- - -- -_ .. -

10 II
11 III1
 Tab le I. I continued ... Table 1.1 continued
Investigators Turbo- Design Objective Comments Investigators Turbo- Des ign Objective Comme nts
machines variab les functions machines variables functions
Jun et al. 1. 20 axial Bezier curve. \. Minimize the MOEA is used Oyama et al 4-stage 80 design Maximization Multi
[38] compressor I. Number of total pressure for [42] compressor parameters of isentropic objective
2.30 axial design loss, maximize optimization. from radial efficiency and optimization
compressor variables=8. static pressure distribution total pressure was applied.
2. Number of ri se. of total ratio
design 2. Maximize pressure and
variables=24. the pressure solidities at
rise and blade rotor trailing
load, minimize edges and
the rotational flow angles
IIII total pressure and
loss solidities at
Kammerer I.S-stage Stagger angle, Polytropic Performance stator are
et al. [39] 3D turbine stacking line efficiency enhanced. chosen as
blade and chord enhancement design
lengths are the variables.
parameters Lian and NASA 32 Maximize the Multi
Lampart and I. HP gas I. 6 geometric Efficiency Two HP Kim [43] rotor 67 deterministic stage pressure objective with
Yershov[40] and steam parameters of enhancement stages and one transonic design ratio and to Monte Carlo
turbine Highly loaded LP exit compressor variables are minimize the Simulation is
stages and gas turbine. turbine stage used to weight used for
LP exit 2. Nine were define the reliability
stage of geometric optimized shape of the analysis. New
stea m parameters of during the blades design with
turbine HP steam constrained lighter weight,
turbine . direct larger
3. LP exit efficiency- pressure ratio
turbine stage based and more
I parameters Optimization reliable
Burguburu Transonic Bezier curve Enhancement Efficiency performance
III', et al. [41] compressor of isentrop ic increased. than base line
and turbine efficiency rotor 67.
stator.

12 13

J
II
Tab le I. I continued
Investigators Turbo- Design Objective Comments Investigators Turbo- Design Obj ective Comme nts
machines vari ables functi ons machines va riables functions
Song et al. Low aspect 19 design Maximization ARDE is Oksuz [47] 3D gas Tota l 36 Maximization Genetic
[44] ratio variables of isentropic applied. turbine va riables of Algorithm
transonic for stator effici enc y 1.3% blade were used. aerodynamic based
turbine and 9 efficiency is efficienc y optimization
des ign increased. and torque to tools were
vari ables reduce the used . 2%
for rotor weight and blade
blade size and cost efficiency
Yi et al. [45] Transonic Sweeping Maximization 1.58% of the gas and 7.6%
axial and of adiabatic efficiency turbine torque were
compressor composite efficiency increased . engine. increased.
(NASA bowing Genetic Lampart High 9 geometric Minimization Stage loss
rotor 37) were Algorithm is [48] pressure parameters of enthalpy was
employed used. steam used : Stator loss decreased by
to redesign turbine and rotor 0.4%.
the leading blade
edge numbers and
stacking stagger
line. angle, rotor
Chung et al. Rotor Exit blade Lift-drag Regression blade twist
[46] blade of angle, ratio of the analysis and angle and
Curtis stagger blade section Genetic parameters of
turbine angle and is maximized. Algorithm is stator b lade
maximum used . compound
camber. lean at root
Rotor and tip.
profile is Constraints:
defined by mass flow
B-spline rate, exit
curve swirl angle,
and reaction.

14 15
II
Table I. I continued Table 1.1 continued
[ nvestigators Turbo- Design Objective Comments Investigators Turbo- Design Objective Comments
machines variables function s machines variab les functions
Oyama et a l. NASA Each blade To Reduced Papila [51] Supersonic First vane First va ne : RBNN
[49] rotor 67 shape is mInImIze entropy Turbine (7 design Objective: based
fan. represented the flow production variab les), Stage total to RSA
with 56 loss by more First Blade total model is
design manifested than 19% (11 design efficiency. deve loped.
parameters . vIa compared variab les), Constraints: 1.
entropy with the Second Min imum
generation. rotor 67. Vane (1 1 thic kness. First
Pierret et al. NASA Blade Effic iency Optimized design B lade:
[50] rotor 67 parameterized and at different variab les), Objective:
using Bezier pressure operating Second Stage total to
curve ratio point, Blade : total
enhanced. Ge netic Variables efficiency.
IIII algorithm are same Constraints: l.
used. as first Minimum.
vane. Second Vane:
Objective:
Stage tota l to
tota l
efficiency.
Constrai nts: 1.
Minimum
thickness.
Second Blade:
objectives and
constraints are
same as first
va ne.
Lofti [52] Fan Blade Camber Efficiency Genetic
line , Lean algorithm
and Sweep is used.
'-----

16 17
II
Table I. I continued Tab le I. I continued .. .
Investi gators Turbo- Design Obj ective Comments Investigators Turbo- Design Objecti ve Comme nts
machines vari ables functions machines variables functions
Xing and Wing 6 Wing: Kelner et al. Fa n blade Variables: Minimization: Genetic
Damodaran Design, Fan constra ints Minimize [58] Stagger Loss Algorithm,
[5 3] Blade drag/lift angle, coefficient Multi
Li and He Tra nsonic Intra row Flutter tangent to and torque obj ective
[54] Compressor gap fre e camber line Minimization : design
margin at trailing Static
Barankiewicz Low speed Stator Overa ll edge, pressure
and Multistage indexing efficiency, tange nt to
Hathaway Compressor peak camber at
[55] efficiency, leading
peak edge, ratio
pressure, of
pressure maximum
loss camber to
coefficient chord
length,
Chen et al. Transonic Maximum Adiabatic ratio to
[56] Axial camber efficiency maximum
Compressor location thickness
Dennis et Two 18 desi gn Minimum Multi to chord
al.[5 7] dimensional variables, 5 loss, objective length.
aerofoil nonlinear Maximum optimi zation Clarich et al. Axia l 16 Efficiency 0.85 to
cascade constraints loading, [59] compressor Variables : maximization 0.948%
maximum Aspect enhancement
gap-to- ratio, blade of efficiency
chord number,
Ratio. blade
thickness,
stagger,
chord
length,
blade
angle.

18 19
II
Table l. I continued .. . 1-2 Obj ective and scope
Investigators Turbo- Design variables Objective Comments
machines functions A low aspect ratio compressor blade which is designed by NASA is used for
Jang and Axial Variab les of Efficiency Enhanced present investigation. The compressor blade shape has been optimized by
Kim [60] compressor stacking line maximization efficiency
numerica l simulation coup ling with various surrogate models. The RANS
using
RSA equations are solved by finite volume method and sing le and multi-objective
model. optimizations have been performed. The blade shape is defined for its stacking
Yang et al. Low- Forward- Objectives: Neura l line and a irfoil shape. The stacki ng line is defined by lean, sweep and skew
[61 ] Pressure Skewed Blades Efficiency, network
while blade air foil shape is modified considering the maxi mum thickness. The
Axia l Fan pressure and
ratio, genetic blade secti ons are defined by parametric curves so that minimum numbers of
operati ng algorithm design variables can be considered to modify its shape. The objective functions
range, is used. to increase total pressure rise, adiabatic effic iency a nd tota l te mperature rise are
aerodynamic
considered for optimization. The si ngle objective optimization is carried out
noise
1111111111
Benini [62] Axial Blade stacking Effic ie nc y Genetic using the objectives via surrogate mode ls. The multi-obj ectives optimizations
Compressor line a nd ai rfoil and tota l algorithm are performed considering two obj ectives total pressure ratio and adiabatic
shape pressure ratio is used . efficiency. The optimizations show the improvement in blade performance in
enhance ment.
terms of the objectives considered. Hence, the object of the thesis is to modify
C he n et a l. Transonic Po lynomi a l Efficiency 1.73%
[63 ] compressor curves are used enhancement effic iency blade profile targeting the enhancement of performance with the via application
rotor to defi ne blade enhanced. and analysis of optimization methodology.
curves. RS A
mode l is Two set of problems are so lved considering in-house progra m code and
used. ANSYS -CFX 1l.0 solve r. The first probl em is formulated us ing stacking line
Keskin [64] Rolls- Annular Effic iency, Genetic only and the other proble m is formulated taking stackin g line as well as airfoil
I"'" Royce 9- geo metry pressure ratio algorithm
profile. The design points are ge nerated by three level full factoria l design and
I. stage parameterization and surge is used.
compressor by Bezier curve. ma rgin Latin hyper cube sampling design and surrogate approaches are app lied.
increment
The surrogates RS A, Kriging, neural network, and weighted ave rage models are
ana lyzed for app lication in turbomachinery app lications for their prediction
capabilities. The flow analysis is a lso presented for the blade performance
enhancement by optimizations.

21

20

II1I
Hence the thesis contains:
I. Blade shape optimization: Single and multi-objective optimization of blade CHAPTER 2
shape. The stacking and thickness variables are modified to enhance the NUMERICAL FORMULATION OF TURBOMACHINERY BLADE
effic iency, tota l temperature ratio and press ure ratio based objectives. RANS
equations are so lved to get the flow field and objective functions va lues.
2-1 Transonic axial compressor rotor
2. Surrogate analysis: The surrogate models RSA, Kri ging and Neural network
models along with weighted average and simpl e average surrogate models NASA Rotor 37 de ve loped by Nationa l Aeronautics and Space Admin istration

has been constructed and the blade shape has been modified to enhance the Lewis Research Center (USA) is an isolated axial fl ow compressor rotor [1 3].

performance. The surrogate performances are eva luated for app licability in This compressor is developed targeting the Aircraft turbine engine high pressure

turbo mac hinery blade. core compressor inlet stage. The airfoi l is defined by Multiple-Circular-Arc
(MCA). No inlet guide vane was used to develop this compressor. Rotor
polytropic effic iency and total pressure ratio are 88.9% a nd 2. 106,
respectively.The details experimental result was reported by Reid and Moore
[13] and in AGARD advisory report [14].

Fig. 2. 1 NASA roto r 37

22 23
 NASA rotor 37 which is low-aspect ratio axial-fl ow compressor rotor is used
a ap
-(p/.I i /.l j )= - -+-
~-
a [(alii
II - +aU
- - 2-au,
i
i/.l j J+Si
- 5ij J-P/.I-.. -;; (2.2)
for blade shape optimization in the present study and is shown in Fig. 2. 1. The aXj Ox) ox ) Ox) aX i 3 Ox,
detailed specifications of the compressor are summarized in Table 2.1 . The rotor
tip clearance is 0.356 mm (0.45 percent span). The measured choking mass flow where /.I i and /.I ;, are mean and fluctuating ve locities, respectively. Source term is

rate is 20.93 kg/s, which corresponds to 103.67% of the design flow rate. represented by s;' . These equations with turbulence models has been discretized

The meridiona l view of the axia l compressor is shown in Fig. 2.2. Total and solved. Turbulence closure requires modeling strategies for the ve locity

pressure, total temperature, and the adiabatic efficiency with relation to the mass correlation, /.I; u ~ .The c losure techniques can be classified as:

flow rates are measured at inlet (station I) and outlet (station 2) positions. The 1) Zero-equation or a lgebraic eddy viscosity models: The models employ
inlet and outlet positions are located at 41.9 mm upstream of the tip leading an a lgebraic form for the turbulent stress, u;u) .

edge of the rotor and at 101.9 mm downstream of the tip trailing edge of the
2) One equation models: These models emp loy an additional partia l
rotor, respectively. Two cases have been analyzed; one considering only
differential equation (POE) for a turbulent velocity scale.
stacking line and another considering the stacking line in combination with
3) Two equation models: One POE equation for turbulence length scale
airfoil profile. Two cases considering blade stacking line modification and
and another POE equation for turbulence velocity scale are employed.
stacking line and airfoi l shape modification have been presented here. The
4) Reynolds stress models : Several POEs are employed for all the terms for
problems are solved by Baldwin-Lomax and k-£ turbulence closure models with
turbulence stress tensor, - PU;/.I , .
two different programming codes . The descriptions of the numerical schemes
5) Oirect numerical simulation: 30 time dependent structure is resolved
are give n in the following sections.
through a numerical so lution of time-dependent Navier-Stokes equatio n.

2-2 Reynolds-Averaged Navier-Stokes Analysis Present thesis implements Zero-equation or a lgebra ic eddy viscosity model and
two equation model.
Turbulence modeling is a key issue in most CFO simulations. Virtually all
engineering applications are turbulent and hence require a turbulence model. 2-2.1 Zero-equation or algebraic eddy viscosity model

I For steady incompressible turbulent flows, the continuity and Reynolds-

The algebraic turbulence model of Baldwin and Lomax [79] has been employed
Averaged Navier Stokes equations are represented as follows:
to estimate the eddy viscosity. The Baldwin-Lomax model is a two-laye r
algebraic zero-equation model which gives the eddy viscosity as a function of
a (2 . 1) the local boundary layer velocity profile. This model is based on mixing length
aXi (pxi ) = 0
model. The model is suitable for high-speed flows with thin attached boundary-
layers, typically present in aerospace and turbomachinery applications. It is

III 24
25
!I

il l:
 commonly used in quick design iterations where robustness is more important Table 2.1 Design specifications of NASA rotor 37
than capturing all details of the flow physics. It is a low-Re model and as such it
requires a fairly well-resolved grid near the wall. The model is given in the form
Mass flow , kgls 20.19
of following equation:
Rotational speed, rpm 17190
O.16P.V 2[1 - exp( - y + / A+)J2140 S y S Y')

+5.5(~·:: JJ'.y;:: y'

Pressure ratio 2. 106
p, =
I 0.02688PF,,[t
(2.3)

Inlet hub-tip ratio 0.7

y+= ~p" r".y ,A+ = 26 (2.4)

Inlet tip relative Mach
1.4
Pw no.
Inlet hub relative Mach
1.13
no.
where, y is the distance normal to the wa ll. /l,,,a nd Tw are the shear stress and Tip so lidity 1. 29
molecular viscosity at the wall. OJ and Fware the functions of ve loc ities; u, v
Rotor aspect ratio 1.19
and w.
2-2.2 Two equation model Number of rotor blades 36

The K-eps ilon model which is a two-equation model is one of the most common
turbulence models. It includes two extra transport equations to represent the
I,
turbulent properties of the flow . This a llows a two equation model to account Station 1
I for history effects like convection and diffusion of turbulent energy.
I Station 2
--- 1----
I
---_
I
--- ,--- ---- -
I
--- ,--
I
- - -- --- - - ~ --
___ JI __ ___ __ _ ---- ,
I
I
I

I Hub

Fig. 2.2 Meridional view of Rotor 37

26 27
The two equation model employs equations governing dynamics or gross
properties of turbulence and it relates them to eddy viscosity through the
equation:

1', =I'Cf R"

!' !'
j'
!' =exp [1+-2.5]
RT 150 '
R
r
pk'
=----;;; (2.5)

where, e ll is constant and R r, k and f: are the turbulence Reynolds number,

turbulence kinetic energy and turbulence dissipation, respectively.

Inlet
2-3 Blade stacking line modification region
(H-GI'id)

2-3.1 Numerical analysis method

The three-dimensional thin-layer Navier-Stokes and energy equations are solved
on body-fitted grids using an explicit finite-difference scheme. An explicit
Runge-Kutta (RK) scheme proposed by Jameson et al. [SO] is used to solve flow
from initial to steady state with a spatially varying time step to accelerate
convergence. This RK scheme reduces the required storage but also reduces
time-accuracy of the schemes, usually to second order. Arti ficial dissipation
terms have been added to resolve shocks. Density based scheme is employed to
(a) Perspect ive view
solve the continuity equation by driving the density residual to zero. The
algebraic turbulence model of Baldwin and Lomax [SO] has been employed to
estimate the eddy viscosity.
(b) Grids at hub of LE and TE
Figure 2.3 shows computational grids. A composite grid syste m with structured
H-, C- , and O-type grids is adopted to represent the complicated configuration Fig. 2.3 Computational grids

of the ax ia l compressor. H-type grid consists of 60x36 x63 grids (in the
streamwise, pitchwise and spanwise directions, respectively) and is introduced for the blade passage. The grid embedded in the tip clearance consists of
for the inlet flow region. C-type grid consists of 350 x46 x63 grids, and is used IS2 X 13 x I3 grids. The who le grid system has about I, ISI,OOO grid points. The
average number of iterations and C PU time for the converged so lution are

28 29
i I;f '
, I
approx imate ly 3,000 and 3 .5 hours w ith supercomputer of NEe SX-6 (144
"
I'~i

GFLOPS), respectively. ,,
"'~
Mach numbers in each direction, total pressure, and total temperature are give n
,)TE
Swept blade
at the inlet. At ex it, the hub static pressure ratio has been specified, and the
radial equilibrium eq uation is so lved along the blade span. A periodic tip
clearance model is used to resolve the tip clearance flow explicitly. No-slip and
Reference
adiabatic wall conditions are used at all the wall boundaries. For reducing the blade Rotation

computational load, flow field in a single blade passage is simulated by ~

app lying periodic boundary cond ition in the tangential direction.

2-3.2 Objective functions and design variables

In the present study, the adiabatic efficiency, total temperature ratio, and total
pressure ratio are se lected as objective functions of the shape optimization of (a) Top view

the rotor blades. In order to impro ve the overall efficiency, the objective
function is defined as:

F Efr = 1- '1,,,, (2.6)

"'--- Swept blade

( pIO/a/,exil / Pfo({lf.jnle/ ) (k- I ) l k - 1

17(1(/ (2.7)
T,OIU{,/!Xil / ~()I(/I,il/lel - I

where, /)"d is adiabatic efficiency, inlet and exit are indicated in Fig. 2.2 ,

respectively. The total temperature ratio (F T ) and total pressure ratio (FI') are (b) Side view

also tested as objective functions for shape optimization, as follows:

Fig. 2.4 Definition of blade sweep

30 31
1; il!llill

FT = T 'Olal, exi, / T 'oral, inle, (2 .8)

-'
,-
,-
-' TE
,,
/,'
Reference ,,
Fp = P towl,eti/ P 'Olal,illle, (2,9) blade ,, ,,
,,
,, ,,
"
where, the location of inlet is not the same as in Equation (2 .6), and corresponds ""
Leaned blade ,','
to the location of inlet of C -grid in Fi g. 2.2. The intentio n of optimizatio n is to ~:: SS
,'
,,' ' '
,
increase the efficiency (therefore to reduce efficiency based objective function ,
FEU) ' total temperature ratio, and total pressure ratio. :/ I,'
Rotation

':"" ~

High temperature rise in a stage is desirable to minimize the number of stages

for a given overa ll pressure ratio in a compressor, provided that one can
maintain the same effic iency, Hence, as proposed in the present study, it is
Fig. 2.5 Definition of blade lean (top view)
important to address the total temperature and pressure rise along with the
efficiency consideration. To attain a higher total temperature rise, the designers + y (bend on blade
I pressure surface side)
must combine high blade speed, high axial velocity and high fluid deflection in
the rotor blade. High blade speed is limited due to blade stress as well as
Casing
aerodynamic reasons, and high axia l velocity and high fluid deflection are
limited due to adverse pressure gradient. Same set of experimental designs are
used to obtain the response data for the three objectives. Thus, inevitab ly, some
optimum points are located at the boundary of the design space .
, ,,
,,
Multi-objectives optimization to compromise between efficie ncy and total ....... __ ... -
pressure is also considered by emp loying the following objecti ve function,

F,,/= '7(f(t+ Wt P total,eti/ P ,otal, inle, (2 . 10)

Fig. 2.6 Definition of blade skew (front view)

where, w; is a we ighting facto r specified by des igner.

32 33
TI I',I

Three design variables are selected in these optimizations; one for sweep, The flow domain and meridional view of the compressor for simulation is
another for lean, and the other for skew. In Fig. 2.4 to 2.6, the definitions of the shown in Figs. 2.2 and 2.7, respectively. Flow parameters like total pressure,
variables are presented. Blade sweep, a in Fig. 2.4, is defined at the rotor tip, total temperature, adiabatic efficiency, etc.in relation to the mass flow rates are
and normalized by the axial tip chord (= 27.77 mm). The airfoil sections are measured at inlet (station I) and outlet (station 2). These stations are located at
moved towards downstream direction for positive sweep (a). The line of the 41.9 mm upstream of the tip lead ing edge of the rotor and at 101.9 mm
swept blade between the rotor tip and hub is linearly connected and tip downstream of the tip trailing edge of the rotor, respectively.
clearance gap is kept constant.
Structured grid system with O-type grids at near blade surface and H-type grids
Blade lean is shown in Fig. 2.5 by the top view. Lean is defined as the at the other zones are emp loyed. The O-grid having width factor 0.5 and 15
movement of aerofo il normal to the chord line at the tip. Lean also is elements is assigned to shroud tip clearance. Nodes in at blade to blade and
normalized by the axial tip chord. Here, lea n, ~ is taken as positive if the streamwise direction are 46 and 119, respectively. The inlet and outlet blocks
aerofoil sections are moved towards blade suction surface side. Lean (fJ) is taken contain 30,000 grid points each and the main passage contains 240,000 grid
as zero at hub and the blade lean is li nea rl y connected from hub to tip and tip points, approximately. Approximately, total 300,000 nodes are used for
Ii I
clearance is kept constant. computations fora blade passage consisting of single blade out of 36 blades of
I I I1 the compressor considered. The number of grid points is se lected after
Figure 2.6 presents the skewed blade stacking line, and a single skew angle (y)
generating different meshes with different resolution at near wall , tip and other
is defined o nl y at rotor tip. If the blade bends towards pressure surface side,
skew angle is taken as positive. The skew line is defined as a second order
polynomial. The constants and the coefficients are found by the constraints;
skew angle is zero at both hub and mid span, and y at tip of the blade.

2-4 Blade stacking line and thickness modification

2-4.1 Numerical analysis method

III In this case, the commercial code ANSYS-CFX 11.0 [81] is used for simulation.
Blade profile creation, computational mesh gene ration, initial and boundary
condition definitions and flow simulation are performed by Blade-Gen, Turbo-
Grid, CFX-Pre, CFX-Solver, respectively. (b) Blade passage

Fig. 2.7 Computationa l domain

34 35
II

Zones etc . mass imbala nce becomes lower. In ge nera l, residual target were set to 3. 0 E-
05(=RMS) and imbalances were kept below I.OE-02. Time scale was kept
The inlet total temperature and total pressure are set to 288.15K and !.Oatm,
smaller initially and gradually increased to make faster and better convergence.
respectively. The design flow rate which is 0.56083 kg/s per passage is set at
outlet. Steady state simu lation with a ir as ideal gas is considered. The three-
dimensional finite element based finite volume method is used to discretize the
governi ng equations. Thus, the solver retains the geometric flexibility of the
1.0 E·03 " , . , - - - - - - - - -- - - - -..,
finite element method as well as the important conservation properties of the
finite volume method. Each node is surrounded by a set of surfaces that define 7.5E·Q4

the control volume. All the so lution variables and fluid properties are stored at
5.0E·04 -I.. K I'I~ \\'·I'>hlm

the element nodes. However, various terms in the equations require solutions or
2.5E·Q4
solution gradients to be approximated at integration points. Finite element shape RM S [1-111 ,1'"

functions are consequently used to evaluate the solution and its variation within O.OE+OO I ~
o 125 250 375 sao
mesh elements. Adiabatic walls with no slip and hydraulically smooth walls are Accumulated time step

considered at solid boundaries. Periodic boundary is set at the blade passage

(a) Residual history
interface, and tip clearance is modeled. In boundary setting, viscous work is
included, and 5% turbulence intensity is assigned at inlet. The grid resolution is 1.0E+03 ~ ll -L Il~'rg~ illlh.tI,IIIIX (,'"j ill RI

made near wall region in such a way that scalable wall function can work with
k-E model. The k-E model is stable and numerically robust and has a well 1.0E+00

established regime of predictive capability [81].

1.0E·03

The most difficult task in transonic flow is to get properly converged solution .
1.0E·06 +I-----r-----r-----r--~
In AGARD advisory report [14] , the conditions reported for convergence are o 125 250 375 500
Accumulated time step
such that the calculated mass flow for rotor 37 blade at the inlet should match
the mass flow at the exit within 0.017%, and deviation in mass flow should be (b) Imbalances

within 0.004% for the last 200 iteratio ns. Hence, the similar conditions were
Fig. 2.8 Convergence plots
checked for the present calculations at the design flow rate . As current problem
contains number of different geometries to simul ate, some cases were producing
large mass imbalance although the residual criteria were fulfilled. Time scale,
RMS value and number of iteration for those cases were changed so that the
36 37

III
Number of iterations was about 500 for a single converged solu tion, and CPU
time for single simulation is approximately 8 hours in a personal computer with Here, t denotes the parameter of the curve normalized in [0 I] and Pi are the
an Intel Pentium IV CPU of 3.0 GHz speed. Benini [63] and Beheshti et al. [82] coordinates of the control points . The control points of Bezier curve are
also reported on the si mulations for NASA rotor 37 using sa me computer code. considered variables. In prese nt problem, to reduce the number of designs, some
They used approximate ly 240,000 grid points and set convergence criteria to control points are kept fixed and only fi ve control points are defined each for
RMS=5.0E I0-7 and I.3E-5 , respective ly. The mo mentum and mass residuals camber line and thickness distr ibutions.
are presented in Fig. 2.8(a) and mome ntum, mass and energy imbalances are
The blade tip control points are moved equally normal to the chord line so that
presented in Fig. 2.8(b).
blade can be leaned. Hence a single variable is considered for blade tip- leaning.
2-4.2 Objective functions and design variables Similar movement of control points of mid span camber line produces one more
variable. Therefore, two design variables are produced among 10 control
The objective functions are defined as presented in equations 2.6 to 2.10. The
parameters: the variab les at tip and mid span of blade to define lean (fJ) . The
blade profile is constructed by third order spline curve and Bez ier curve and
Fig. 2.4(a) shows the definition of blade lean. The blade sections from hub to tip
modified. The parameterization of the blade by Bezier a nd spline polynomial
are interpo lated by cubic spline curve .
curve gives two advantages: limited number of points can control the curves and
thus, these produce limited number of design variables and these produce The cubic spline interpolation is a piecew ise continuous curve pass ing through
smooth and discontinuity free blade profile . The variables from bla de thickness each of the coordinates. These splines are popular because they are easy to
and lean are collected from Bez ier control points. Blade stac king line is implement and produce a curve that appears to be seamless. Points [Xi, y;] , i= I,
modified by third order sp line curve which is used to connect the stacking from 2, . .. ,n for the function y=f(x) produces n+ I points with n intervals . There is
hub to tip . The blade shape can be changed by movi ng the contro l points of separate cubic polynomial for each interval, each with its own coefficients and
Bez ier curve. A Bez ier curve of order n is defined by the Bernstein the polynomials become:
polynomials :

Si(X) = Pi(X- Xi)3+q ;(X-Xi)2+ r i(X- x;)+Si, x E Ix; ,.\ +, 1 (2. 13)
C( I) = I B,,, (t)F: (2 .11 )
,-0

where, Bezier blending function is as follo ws:

where, coefficients P, q, r, and s are functions of x. Together, these polynomial
segments are denoted as Sex) . Since there are n interva ls and four coefficients
B,, (I) =(':} (I -I)"-' (2.12) for each, it requires total 4n parameters to define the spline Sex). To produce
piecewise continuous function two conditions ap plied at both ends: Si(Xi) = Yi
If Pi = (Xi, y;), then xU) = I x,n, "(t) and y (l) = I YA .,, (I) '
;=0

38 39
11';1

and Si(Xi+ l) = Yi+1. The other two conditions are prod uced by maki ng second and CHAPTER 3
third derivatives continuous : S;_I (x, ) = S;(x,l and s;(x, ) = S;~ , (x, ). OPTIMIZATION PROCEDURE
3-1 Introduction

Optimization is central of any problem involving decision making which entails

choosing among alternatives whether in engineering or economics . The measure
'"'" tJ
of goodness of the alternatives is explained by objective functions or
~
:£j performance indexes. Optimization methodology deals with the se lection of the
i;
~ best alternative in the sense of the given designs. In genera l opti mi zation

1
z
problem is defined as

a 20 40 60 80 100
M % (lE to TE)
Minimizej(x) (3.1 )
Subject to x E Q (n is subset of RI/ and is called the constraint set or feasible
Fig. 2.9 Definitions of thickness variab les set.)
The function f: f: R" ~ R is objective or cost function to be minimized. The

vector x is an n-vector of independent variables; i.e., x = [XI , X ;, " "XI/ ]T E R" . The
Airfoil thicknesses are a lso defined by Bezier curve and the definitions o f
variables XI, X ], .. . ,xl/are called decision variables.
variables from thickness are shown in Fig. 2.9. The multiple circular arc (MCA)
profile of blade is defined by Bezier polynomial of five control points; PI, P2, The optimization procedure is presented in Fig. 3.1. Initiall y design space is
P3, P4 and P5 . As similar curves are used for hub and tip, hence total control decided and Sampling designs in design space are selected by Design of
points are coming to be ten. The point P3 is moved normal to the chord line (15) Experiments (DOE) [9]. At next step, the designs are evaluated by RANS solver
and a long the chord line (0 and P3 at hub produces two design variab les. Other and single or multi-objective optimizations are performed.
control points (P I, P2 P4 and P5) are kept fixed. Similarly P3 at tip produce s
3-1.1 Design space
two more variables while other contro l points are kept fixed.
Design space is defined by the lower and upper bound of the factors or
Hence, the among 20 control parameters, 6 design variables are selected fo r
variables. The design space is defined by:
optimization. Variable (j is considered positive if the control point P3 is mo ved
from chord line and variab le ( is considered positive if the control point P3 is x: S X i $; x;1 (3 .2)
moved towards the trailing edge of the blade.

40 41
III

3-1.2 Design of experiments

(Proble m se tup)
Object ive fu ncti ons & Design varia bles
I A Design of Experiment (DOE) is a structured, orga nized method for
~ determinin g the re lationship betwee n factors affect ing a process and the
(Decision of de sig n space)
Lower and upper bou nd of va ri ab les are set ~
outputof that process. It is based on co nducting and ana lyz ing controlled tests to
evaluate the factors that control the va lue ofa parameter or gro up of para meters .
-+
The DOE includes custo m des ign, screening design, response surface des ign,
( Design of ex pe riments)
Selec ti on of desig n points by LHS/ FFDID-O pt ima l design fu ll facto rial des ign, Taguch i arrays, mi xture des ign, aug ment des ign and
Sample Size and Power [83]. In present thes is fu ll factoria l and space fi ll ing
+
( Numerical Ana lys is) designs are used and the ir exp la nations are presented be low.
Determinat ion o f th e va lues of objecti ve functi ons of th e designs by RANS solver

t
(Mu lt i-objective for mul atio n)
3-1.3 Full factor ial design
Weighted sum of m ultip le obj ectives to make sin gle object ive
problem
A fu ll factorial design [9] conta ins a ll possible combinations of a set of factors.

•
(C onstruction of surrogate)
This is the most conservative des ign approach, but it is a lso the most costl y in
experimental resources. The Fu ll Factoria l designer supports both continuo us
Surrogate model constructi on us ing the objec tives
factors and categorica l facto rs with arb itrary numbers of leve ls. If three level

•
(Search for optimal point)
Optima l poin t search from constru cted surrogate llsing opt im ization a lgorith m
fu ll factoria l design is presented in Fig. 3.2 (a) .

3-1.4 Latin Hyperc ube Sa mpling

.1
Is opt ima l point with in design space? No
Latin Hypercube Sampling (LHS) [84] , a n effecti ve sampling method in DACE,
is a matrix of m x n order, where m is the number of leve ls (sampling po ints) to

Yes be examined and n is the number of des ign variables. Each of the n co lumns of
the matrix containing the leve ls I , 2 . . ... m is randomly pa ired to form a Latin
Optima l Desig n
hypercube. It generates random sample points, ensuring that a ll portions o f the
deSign space are represe nted. For two va riables , LH S is prese nted in Fig. 3.2(b).

Fig. 3. 1 Optimization procedure Here, the differe nce is visible for the three -leve l full fac to ria l and Latin
Hypercube Sampling des igns as shown in Fig 3.2.
42 43
 IIIII

below . In multi-objective optimization surro gate approximatio n models are

coupled with genetic a lgorithm (MOEA) and optimum designs are predicted.
0.5
3-2.1 Surrogate approach
N
Q)
:c 0
'"
·c Most engineering design problems require experiments and/o r si mulations to
>'"
evaluate objectives and constraints as function of design va riab les. For many

. .
-0.5
real world problems, however, a single simulatio n can be expensive in terms of
time and cost. One way of alleviating this burden is by co nstructing
~ .
~ -0. 5 o 0.5 approximation models (known as surrogate mode ls) that mimic the behav ior of
Variable 1
the simu lation model as c losely as poss ible while being computationa ll y
(a) Fu ll factorial design
cheaper to evaluate. Surrogate mode ls are constructed using a data-d riven,
bottom-up approach .
•
0.5 A mode l is constructed based on modeling the response to a limited number of
N

:g·c
Q)

0
• intelligently chosen data points . An important distinction can be made between
two different applications of surrogate mode ls. The first involves building small
1IIIIIIIiiiil >'"
• and simple surrogates for use in optimization. Simple surrogates are used to
11111;11111 -0.5
guide the search towards a g loba l optimum. Once the optimum is found the
• surrogates are discarded . In the second case one is not interested in finding the
, 111111 -1
-1 -0.5 o 0.5 optimal parameter vector but rather in the global behavior o f the system. Here
II IIIII Vari able 1
the surrogate is tuned to mimic the underlying mode l as close ly as needed over

(b) Latin hyper cube sampling the comp lete des ign space. Such surrogates are a useful, che ap way to gain
ins ight into the globa l behavior of the system. Optimization can still occ ur as a
Fig. 3.2 Design of experiments
post process ing step.
3-2 Optimization procedure
The scientific cha llenge of surrogate modeling is the generation of a surro gate
Single as well as multi-objecti ve optimization procedures are done through that is as acc urate as poss ible, using as little simulati on eva luations as poss ible.
RANS simulated data. [n sing le objective optimization, surrogate approx imation The process comprises of three major steps which may be interleaved
mode ls are implemented. Three basic surrogate models as well as simp le iterative ly:
average (SA) a nd weighted ave rage (PBA) surrogate mode ls are exp lained
44 45
1

Sample se lection E(£)=Oa ndV(c) = all (3.6)

• Construction of the surrogate mode l a nd optimizing the mode l parameters wherey is a co lumn matrix of M responses and X is an Mxn matrix of design
Appraisal of the accuracy of the surrogate .
variable va lues. The least square estimate of f3 is

T he accuracy of the surrogate de pends on the number and location of samp les
(expe nsive experiments or si mulatio ns) in the des ign space. Various DOE (3 .7)

techniques cater to different sources of errors, in partic ular e rrors due to no ise in
the data or erro rs due to an improper surrogate mode l. Some basic surrogate The constructed second orde r po lynomi a l response can be expressed as:

mode ls are po lyno mi a l res ponse surface approxi matio n (RSA) [9], Kri ging
N N /Ii
(KR G) [10, 85 ], Supp ort vector mac hines (SVM) [86] a nd artifi cia l ne ural F(x) = Co + ~:>jXj +
jo:o. l
2:>jjX~ + I2:>ijx,x
j =1 i''"J
j
(3.8)
netwo rks (ANN) [11]. The basic surrogates used in prese nt pro blem are RSA,
KRG and ANN. The mode ls are described below.
For a second order po lynomial mode l, used in the c urrent study, the number o f
3-2.1.1 Response surface approximation model regress ion coe ffi c ie nts is(n + 1) x (n+2)!2.
Sum of squares of res idua ls is used to check the fi tness of data into curve and is
Res ponse Surface Approx imatio n is a methodology o f fittin g a polyno mi a l
formul ated as
function for discrete res po nses obta ined fro m numeri ca l ca lculations . It s igni fies
the associati on betwee n res ponse function and design va riables. For a set of n
design va riables xi> the linear respo nse functi o n is formulated as: (n-l
R 2 adj= l - - - ( l - R
n- p
') (3.9)

(3. 10)
N

F, = 2:>jxij + B, (3.3)
I:I

whe re errors, are independentl y distributed and with zero mean and va riance ~ . ,
(I.F )'
C;
w here,SS£ = L.. (F, -F, ) and SFF = F' F - - ,--, -'-
i =1 n
a 2, i.e.,
p is numbe r of mode l pa rameters. T his is spec ified by the number of regress ion

coeffici ents.
£(£,) = 0 (3.4)

In the matrix fo m1 equatio ns (3.3) and (3.4) can be expressed as:

3-2.1.2 Radial basis neural network model

F= X f3 + c (3 .5) Ne ural netwo rks are large nu mber of parallel computati ona l systems (s imple
processors) li nked by adj ustabl e interconnectio ns. The bas ic concept in
46 47
 neura lnetwork is to simu lat human functions of learn ing from ex perience, neuron acts as a detector that produces I whenever the input p is identical to its
predicting from previous data, etc. The processing nodes are called 'neurons' . weight vector p . The bias b allows the sensitivity of the radbas neuron to be
These neurons consist of multiple inputs each modified by a weight and a single adjusted. For example, if a neuron had a bias of 0.1 it would output 0.5 for any
outputs. The output of eac h node depends only on the information that is locall y input vector p at vector distance of 8.326 (0.8326/b) from its weight vector w. A
available at the node, either stored or arriving. The ability of the network to radial basis function is presented in Fig. 3.4.
predict is stored in the we ights that can be obtained from a set of training
patterns . The training of a network can be performed by repeated cycling
through the data.

The radial basis neural networks [II] are two-layered networks with a hidden
Output
layer of radial basis transfer function with linear output (Fig. 3.3). The hidden
layer consists of a set of radial basis functions that acts as activation function s,
',...2:::==:::::.c:==::::..., '-;=:::::::::::..c::::::::;J
o~teut laye~
the response of which varies with the distance between the input and the center.
Radial basi s hidden layer ILinear
The distance between two points is determined by the difference of their
coordinates and by a set of parameters. The main advantage of using the radial
Fig. 3.3 Radial basis network (single neuron)
basis approach is the ability to reduce the computational cost due to the linear
nature of the radial basis functions.

(3 . 11) 1.0

Here, the net input to the radbas transfer function is the vector distance between
n
its weight vector wand the input vector p, multiplied by the bias b. The transfer 0.0 -0.833 -0.833

function for a radial basis neuron is:

Fig. 3.4 Radial basis function

radbas(n) = e-n 1 (3.12)

The radial basis function has a maximum of I when its input is O. As the
distance between wand p decreases, the output increases. Thus, a radial basis There are several possibilities for the choice of basis functions. [fthe fixed basis
and other parameters are fixed through the training process, the model is linear.
48 49
However, if the basis function changes during the training/learning process the COV[Z(Xi),Z(Xi )] = 0-2R[R(Xi , xi )] , i, j= 1,2 ....... n, (3.14)
model is non-linear. The learning process is equi va lent to finding a surface in
multidimensional space that provides a best fit to the training data which
whereR is a corre latio n matrix consisting of a spatial correlation funct io n R(xi,
isfurther used to interpolate the test data. The design parameters for this
function are spread constant (SC) and a user de fin ed Error Goa l (EG). SC value
1) as its e lements. 0- 2 is the process va ri ance represent ing the scalar of the
spatial corre lation function (SCF) R(xi,l) quantifying the correlation between
is selected in such a way that should not be so large that each neuron wi ll not
respond sa me for the a ll input, and that should not be so small that the network
any two ns sampled data points Xi and 1 , and thereby contro ls the smoothness
of the Kriging model, the effect of nearby points, and differentiability of the
will be very high sensi ti ve for every input within design space. EG or mean
surface . The Gauss ian function used in this work is the most preferable SCF
square error goa l se lection is also important. A very sma ll e rror goal will
when used with a gradient based optimization a lgorithm as it provides a
produce over training of the network whi le a large error goal wi ll influence the
relativel y smooth and infinitely differentiable surface. The following Gaussian
accuracy of the model. Allowable error goal is decided from the a llowable error
correlation function is used to obtain the SCF.
from the mean input responses. In the present study, we used customized RBNN
function newrb ava ilab le in MATLAB.
.. "rI,· . . 2
R(x' ,Xl ) = exp[- LA Ix~ - x£1] (3.15)
k= 1
3-2.1.3Kriging model

The Kriging method in its basic formulation estimates the value of a function
where ndv is the number of design va riables and Ok is the unknown correlation
(response) at so me unsampled location as a combination of two components, the
parameter used to fit the model. The optimal correlation coefficients,
g loba l model and a systematic departure . Mathematically,
0=(0,...... ,0" ,), on the design space [0,1]"- are found by max imum li kehood
y(x) = J(x) + Z(x) (3.13 ) estimation (MLE).

wherey(r) is the unknown function to be estimated andf(x) is a known fu nction Let Y = [y(x' ), .........,y (x'" )l represent the output for a set of known inputs
(usuall y polynomial) representing the trend over the design space, also referred
X =[x', x' , ....... ,x", l. A matrix F is constructed by evaluati ng the regression
to as the ' globa l' model. The second part, Z(x), creates a localized deviation to
function f(x) at each of the design sites.
interpolate the samp led data points by quantifying the correlation of points with
a Gaussian corre lation having zero mean and nonzero covariance. The
(3.16)
covariance matrix of Z(x) is given by

50
IJ'
An another important term that is needed in order to represent the correlation
between ns sa mpled points and an unknown point x is represented as The MSE of the estimate y(x) is as follows.

rex) = [R (xj), R(x,x 2 ), ••• • •• , R(x,x"' )] (3.1 7)

MSE[y(x)] = 0- 2{[-[IT T J( ~ F: J(~~)J}
( x )r ( x ) (3 .23)

Consider a linear predictor of y{-t} at an unsamp led point x of the form :

Assuming a Gaussian process, the maximum likelihood estimator of the

)I(x) = e T (x)Y variance 0- ' is defined as

(3 . 18)

The mean square error (MSE) of the prediction is then given as c:T' = (y - F /3)R -' (y - F /3) 1n, (3.24)

'II,
MSE[y(x)] = E[e T (x)Y - Y(X)] 2 Any value of the correlation parameter (} = «()I." .... , ()".) in the Gaussian correlation
(3.19)
function creates an interpolative model. The best model is found by so lving the

The Kriging method finds the best linear unbiased predictor (BLUP), which nonlinear unconstrained optimization problem given by

IIIII!I minimizes the MSE subjected to the following unbiased constraint.

q;(B) = - ( det R)" n. c:T 2 (3.25)

E[eT (x)Y - y (x)] =0 (3.20)

3-2.1.4 PBA model

If e(x) is the required solution of equation (3.17) subjected to the unbiased A weighted average mode l proposed by Goel et al.[12] is adopted in the prese nt
T
constra int (equation (3 . 19», then c (x)Y represents the BLUP of y (x), given by investigation. It is based on the PRESS-based-averaging (PBA) model (termed
WTA3 by Goel et al.[12]). The predicted response is defined as follows for the

y (x) = I T(x)/3 + r T(x)R -' ( Y - F /3) PBA model:

(3 .21)

... N SM "

where iJ represents the least square estimate of f3. F wl.,,"I!(X) = L w,(x)F ,(x) (3.26)

jJ =(pT R- 1p)- 1pT R- l y (3.22)

52 53
Irl
where, N,Mis the number of basic surrogate mode ls used to construc t we ighted approximation terminology). Following equation represents a leave-one-out
average model. ( " surrogate model at design point x produces weight w; (x) , and calculation when the generalization error is described by the mean square error

~ (x) is the predicted response by /' surrogate model. (GMSE).

Weights are decided Llsing the guideline that the weights should reflect our
1
confidence in the surrogate model such that the surrogate which produces hi gh
GMSE = - I u;- ft;H)'
k

k ;~ I
(3.28)

error has low weight, and thus low contribution to the final weighted average
surrogate, and vice -versa. In this work, global weights are selection using where, ft; (- O represents the prediction at XU) usi ng the surrogate constructed with
generalized mean square cross-validation error (GMSE) or PRESS (in RS A
all sample points except (xU) ,F;).
terminology) that is a global data-based measure of good ness. Generalized mean Next RSA, Kri gi ng, RBNN, and PBA surrogate models are constructed using
square cross-validation error calculation procedure is given in the appendix. objective function values at design points . Eq uations 3.26 and 3.27 are used to
The weighting scheme used in PRESS-based averaging surrogate is given as
construct the weighted average surro gate model.
,1111 1
follows:

3-2.l.SSimple average model

( E.
K •

•
Wi = -'+~ J , W; = ~
W;
w; A simple averageing (SA) surrogate model when all the basis surrogates
E {/\'g L....
contribute equally is constructed to find the sensitivity of the weights calculated
N .\I/ £
Ea ,." = L - ;-; K < 0, q < 1 (3.27) for the PBA model. Hence the weigh for the surrogates are computed using the
;~l N SM

E; = .JGMSE;,i = 1,2, ... , N sM formula:

II II il Two constants ( and K are chosen as ( = 0.05 and K = -I (Goe l et al. [12]).
W= - - (3.29)
, N SM
Generalized mean square cross validation error is ca lculated to get the weights
for weighted average methods. In general, the data is divided into k subsets (k-
Hence, the equation 3.26 is used to formulate this surrogate.
fold cross-validation) of approximately equal size. A surrogate model is
constructed k times, each time leaving out one of the subsets from training, and
3-2.2Multi-objective optimization
using the omitted subset to compute the error measure of interest. The
generalization error estimate is computed using the k error measures obtained In multi-objective optimization procedure two approaches have been reported
(e.g., average). If k equals the sample size, thi s approach is called leave -one-out here : (a) weighted sum of objectives and (b) MOEA approach. Among these
cross-validation (also known as PRESS in the polynomial response surface
54 55
r I

weighted sum approach is simplest to use where as MOE A is based on genetic explored to get inside of the trade -off analysis between different confl icti ng
algorithm. These approaches are exp lained below. obj ecti ves.

3-2.2.1 Weighted sum approach Response surface approx imatio n method is app lied to get the polynomial
equations for eac h computed objective separately these equations are used for
Weighted-sum-of-objective-functions method [75] which is a lso known as
multi-objective Pareto optimal front generation. RSA is a curve fitting method
' na·ive approach' to multi-objective optimization is generall y used for mul ti-
by regression analysis using computationall y or experimentally generated data.
objective optimization problem to make a mono-objecti ve optimization
The second order polynomial function is constructed according to equatio n 3.8 .
problem. In present optimizations, objectives, FNII and Fr are li nearly combined
with a weighting factor, IVI to constitute a mono-objective, Fwf by na·ive approac h A multi-obj ective problem may be defined as:
and the final obj ecti ve (F'if) is defined by

Minimize 7(~) (M function to be optimi ze)

F"f = F" " + wr Ff (3.30) Subject to g (~) $ 0 (m inequality constraints)

h(~) = 0 (pequality constrai nts)

The weighting facto r, IVr is selected as per design requirement. In ge nera l, it is
the ' designer' s choice'.
where 7(~) = {;; (~) .;; (~).;; (~) •......... .;;/ (~)} is a vector of n rea l valued objective

3-2.2.2 Hybrid MOEA approach functions and x is a vector of n des ign variab les. ~ E R". g(~) E R"', h(~) E W . In

general , engineering problems are assoc iated with so me conflicting objectives

Multi-objective ap proach gives a set of optimal solutions instead of single
in which improveme nt of one objective leads to deterioration of other
opti ma l solution. No ne of the so lutions in this set of optimal so lutions can be
objectives. Eac h feasible so lution of the set ~ of multi-obj ective problem is
considered to be better than any other so lution with respect to a ll objectives
either dominated or non-dominated so luti on, in w hich all non-dominated
considered in the problem. These optimal solutions are ca lled Pareto-optimal
solutions are called Pareto-optimal solutio ns. Vector ~i dominates a vector ~j if
solutions and their funct iona l space representation is termed as Pareto optimal
front [67]. There are numbers of methods avai lable [66, 67] for solving multi - Xi is at least as good as ~j for a ll object ives and ~i is strictly better than ~J for

objective optimization problems but the classical way of tackling mul ti- at least one obj ecti ve.

objective problems is to convert mu lti-objective prob lem into single objecti ve

The methodology used to ge nerate globa l Pareto-optima l front is shown in Fig.
prob lem. The methodology of constructing a g lobal Pareto-optima l front is
3.5. Objective functions and constraints are defined mathematically and
evaluated on the data obtained from numerica l experiments. Polyno mial based

56 57
RSA model is constructed for the objective functions to red uce the complexity
of the multi-objective optimization problem. In the present study, hybrid MOE A I Problem fonnulation I
~
which is used to find Pareto-optimal solutions uses rea l coded Non-dominated I Design of expe riments
(Se lecti"n ofdcsign puints by DO E)
sorting of genetic algorithm (NSGA-II) with local search weighted sum method
[67]. ~
Numerical analysis I
Evaluation of designs by RA NS solver
In NSGA-II, different parameters like population size, generation crossover and
+
Construction of approxi mat ion fun cti ons
mutation are adjusted one by one to suit the nature of the problem.
RSA function

+
Evolutionary algorithm
Formulation of NSGA- II to lind Pareto optima l

r--
solutions
~
Initializt! Sclect parents for

I population mating

+
Fitness Crcate offspring. I
I evaluation crussover and
mutation

1
Insert Stop criteria
N(

I o ffspring meet?

Yes

!
Local search and clustering
..
I Non-domination check
DiscarJing JOl11inalcd solutions
~ Local search
Search
Search
~

r-
NSGA- II for
local optimal
Check for dURlicitx solutions

I Remo ve duplicate solutions

~
I Pareto o(2 li mal front Clusteril1g
I
+
.~

I Confidence check I
Numerical simulation at some clustered points
I I
58 Fig. 3.5 Multi-objectiv<:S'9}ptimization procedure
NSGA-JI giv s a s t of approximat Par to-optimal so lutions and th r for a
constrained optimization III the manner of an unconstrained optimization
weighted sum strategy of loca l search method is used to improve the quality of
problem. In its basic form it replaces the objective function with a quadratic
Pareto optimal solutions. In weighted sum strategy, a ll the objectives are
approximation and replaces the constraint function by linear approximation. Its
combined into one objective. As this strategy is easy to implement, and hence is
implementation consists of three steps, i.e. , updating the Hessian matrix of the
used in this problem. Weights associated with each objective are computed
Lagrangia n function using a quasi-Newton updating method, solution of the
using the formula :
quadratic programming sub-problem, and the formation of a search direction for
a line search procedure using a line search and merit function calculation.
(3.31)

And the objective becomes:

F= L/~ Wk
k=1
(3.3 2)

where HI is the weight for j'h objective, M is th e number of objectives. Fjlllill, Fj"ax
and Fj(X) are the sca led mini mum, maximum and initial va lues of j'h objectives,
respectively. This composite objective is locally optimi zed using Sequential
Quadratic Programming (SQP). These optimized so lutions are merged with
NSGA-\I obtained solutions a nd dominated so lutions are discarded. The global
Pareto-optimal solutions are ac hieved after removing duplicate sol utions from
the non-dominated so lutions. Since the present study deals with a two objecti ve
optimization problem, Pareto-optimal front can help the designer to trade-off
analysis.

3-3 Optimization algorithm

Sequential quadratic programming (SQP) (function jJnincon in MATLAB [79])

has been used as an optimization algorithm that follows Newton 's method fo r

60
61
r II

CHAPTER 4
RESUL T AND DISCUSSION

The design procedure for the two cases (cases I and 2) are same i.e. , initial set
up of problem, selection of design space , numerical simulation at sampling
designs and surrogate construction as shown in flow charts (Figs. 3. 1 and 3.5).
The results of the blade shape optimization and surrogate modeling are
described below.

4-1 Blade Stacking Line Modification

4-1.1 Validation of numerical simulation

The design space is presented in table 4.1. The design space is decided after
some preliminary calculation and three level full-factorial design is used to
generate the design points within design space. As three parametric variables
are selected, the full-factorial design produces 27 design points. Three cases are
not converged and hence 24 design points are used to construct the surrogates.

For the validation of the present numerical solutions, the spanwise distributions
of adiabatic efficiency for the reference blade are compared to the experimental
results at the design flow rate (=20.19 kgls) in Fig. 4.1. The figure shows thatthe
spanwise distributions of adiabatic efficiency, which is the object function in the
present study, match well with the experimental results for the reference shape.
Although the computational efficiency is locally overestimated compared to the
experimental one, the computed adiabatic efficiency has a maximum of 4
percent error with the experimental data at the design flow rate s.

62 63
Table 4.1 Design ranges o f blade sweep, lea n and skew

Figure 4.2 shows the comparison between computational and experi menta I
distributions of adiabatic efficiency with respect to mass flow rate. In the
Lower Upper
Variables Middle figure, the mass flow ratesare normalized with the chocking flow rate (20.93
Bound Bound
kg/ s) obtained from the experiment [13] . It is noted that the horizontal axishas
Sweep, a. 0 .0 0.13 0.25 narrow range around the design flow condition (= 0.965) because flow
characteristics in the present study is only considered in the design flow
Lean,(J -0.036 -0.018 0.000
condition. At the design flow rate (design mass flow/ choke flow rate = 0.965 in
Skew, y
0.0 0 .05 0.1 Fig. 4.2), the adiabatic efficiency obtained by the numerical analysis for the
(radian)
reference blade is 1.3 percent higher than that of the experimental efficiency.
The comparisons between the numerical and experimental results at the design

100

I
80 I
0.9
G

~
60
·13" 0.88
\.::
- ~I -------
I
~
<.0-
<!) I

~ 40~ f~
Reference Bla.de
I
.~ 0.86
Optimum by KRG on B RANS for ref
O ptimum by RBNN '"
'i3 e RANS forop!
--< 0.84
O ptimum by !<SA
e Expt. for ref
I
,
]
, , I
0 Experim ent 0.82
0.95 0.955 0.96 0.965 0.97 0.975
Mass flow/choke mass flow

0.6 0.7 0.8 0.9 1.0

Adiabatic Efficiency
Fig. 4 .2 Adiabatic efficiency according to normalized mass flow rates
(vertical solid line : design flow rate)
Fig. 4 . 1 Spanwise distribution of efficiency
flow rates show that the adiabatic efficiency of the compressor rotor issimulated
correctly by the present calculation. Consequently, flow characteristics in the
compressor rotor can be investigated by analyzing the numerical results in

detail.

64 65
 surrogates, and finding ways to form a more robust composite surrogate, as is
4-1.2 Optimization results done in this study.

In the Fig. 4. 1, spanwise distribution of adiabatic efficie ncy has almost same Table 4.2(a) Weights for PBA mode l to construct weig hted ave rage model for
va lue for three basic surro gate models . . Relative ly large increase of adiabatic F ElT·

effic iency is distributed between 50 and 90 percent span. The largest increase in
efficiency compared to the reference blade is observed near 75 percent span. MODEL Cross Validation Error, Ecv Weight

Tables 4.2 through 4A contai n the optimization resu lts for the three objective RBNN 3.09E-03 0.3 22

fun ctions (FED; Fp and F r) with different surrogate mode ls. The optimal RSA 2AOE-03 OA09
variab les given in these tables are normalized in the design ranges shown in
KRG 3.71 E-03 0.270
Table 4. 1, a nd have the values between 0 and 1. Tables 4.2(a), 4.3(a), and 4A (a) - -

show the cross validation errors and calculated weights for each surrogate. The
resul ts show that the RSA approac h gives the lowest cross validation error
among the three individual surrogates, and the KRG technique performs the
Table 4.2(b) Optimal des igns suggested by various surrogates and
wo rst for a ll three objectives. The low cross-va lidatio n error for RSA is like ly to
corresponding predicted RANS results for objective F EU:
indicate that the objectives are mode led well as quadratic polynomials. In PBA,
the weight associated with RSA is the largest.
Simple
Surrogate mode l PBA RBNN RSA KRG
average
Tables 4.2(b), 4.3(b), a nd 4A(b) compare the predictions of the optimal values
Optimal a opr (sweep) OA42 OAII OAOI OA65 OA42
predicted by different surro gate model s against the actua l RANS ca lculation at design
Popr (lean) 0.905 1.000 0.748 0.97 1 0.916
variables
the corresponding points. Consiste nt with the error estimate discussed above,
)'opr(skew) 0.651 0.6 15 0.632 0.635 0.652
KRG performs the worst. In case of F Elh RBNN gives the most accurate
F surrogute 0.0995 0.0994 0.0993 0.0990 0.0995
prediction of optimal objective function value, followed by PBA. But, in both
FRANS 0. 1010 0. 1008 0.1015 0. 1018 0. 1011
Fr and Fp , RSA gives the most acc urate predictions of the objective fu nctions at 1.50E- lAOE- 2.20E- 2.80E- 1. 64 E-
F RANS-F slIrrogtlle
the optimal points, and RBNN the next. Depending on the characteristics of the 03 03 03 03 03
Freference 0.1135
design space in each prob lem, certain surrogate models can perform more
F RANS-Freferelll"e -0.0125 -0 .0 127 -0.0 120 -0.011 7 -0.0 124
satisfactoril y than others. However, it is difficult to te ll a priori which surrogate
~eduction inFEb(%) 11.0 I 11.19 10.57 10.3 1 -10 .93
models are more appropriate. This is precisely the merits of adopting multiple

66 67
 Table 4.3(a) Weights for PBA model to construct weighted average model for
Table 4.4(a) Weights for PBA model to construct wei ghted average mode l for
obj ecti ve fun ction, FT.
obj ective function , Fp .

MODEL Cross Validation Error, Ecv Weight MODEL Cross Validation Error, Ecv Weight

RBNN 7.39E-03 0.359 RBNN 3.73E-02 0.373

RSA 6.72E-03 0.393 RSA 3.58E-02 0.388

KRG 1.09E-02 0.248 KRG 5.97E-02 0.239

Table 4 .3(b) Optimal designs suggested by various surrogates and Table 4.4(b) Optimal designs suggested by various surrogates and
corresponding predicted RANS results for objective FT. corresponding predicted RANS results for objective F p.

Simple
Simple Surrogate model PBA RBNN RSA KRG
Surrogate model PBA RBNN RSA KRG average
average
aopt(sweep) 0.000 0.000 0.000 0.000 0.000
aopt(sweep) 0.000 0.000 0.000 0.000 0.000 Optimal
Optima l design Popt (lean) 0.900 0.878 0.768 1.000 0.924
design popt(lean) 0.676 0.903 0.769 1.000 0.696 variab les
variables )'opt(skew) 0.227 0. 192 0. 195 0.346
)'opt(skew) 0.044 0.237 0. 165 0.392 0.049 0.237
F surrogatf! 2.1473 2.1585 2. 1505 2.1345 2. 1464
FSlIrrog ate 1.2730 1.2738 1.2744 1.2705 1.271 9
F IIA NS 2. 1525 2. 1536 2.149 1 2. 1494 2. 1512
F IIANS 1.2749 1.2748 1.2748 1.2738 1.2749
5.20E- -4 .90E-
- 1.49E- 4.81E-
1.90E- I.OOE- 4.00E- 3.30E- FRANS - F surmgu,e 1.40E-
F IIANS - FSllrrogate 2.00E-03 03 03 02 03
03 03 04 03 03
Fref erell ce 1.2721 Freferell('e 2.1334
F IIANS - Frefere/,ce 0.0028 0.0027 0.002 7 0.001 7 0.0028 F IlA NS - Frefere/,ce 0.0191 0.0202 0.015 7 0.016 0.01 78
Increase in FT(% ) 0.22 0.22 0.22 0.14 0.22 ~a se in Fp (%) 0.90 0.94 0.74 0.75 0.8343

68
69
 It should be pointed out that while the cross validation error estimate has
Table 4 .5 Root-mean-squareaveraged errors in predictionsofthe surrogates at 12 correctly predicted that KRG is the worst performing surrogate, RSA, with the
optimal points . lowest cross va lidation error, does not yield the best performance . The
discrepanc y between the cross-validation error and actual performance is due to
RMS errors produced by the surrogates two reasons. First, the cross validation error is an estimate of average square
Surrogates
error over the ent ire domain, while the optimal designs are found in a small sub-
FElT Fr Fe
region where one of the surrogates may have an advantage. Second, the cross-
PBA 1.24E-03 4.57E-03 2.12E-02
validation error is only an estimate of global performance. The multiple
RBNN 1.36E-03 3.93E-03 1.93E-02 surrogate model, PBA does not show the best performance in any case, but it

RSA 1.20E-03 3.50E-03 1.69E-02 gives satisfactory results in most cases and protects against the worst surrogate
(in this case KRG). Therefore, even though the multiple surrogate models do
KRG 1.43E-03 7.60E-03 3.63E-02
not necessarily guarantee the best performance, they seem to perform more
consistently, and hence can improve the robustness of optimization results at a
1.3 2.17
minimal computational cost.

~ The efficiency, total temperature, and total pressure ratios obtained by RANS
u 1.2 2.13
c analyses for the reference shape (Fre!erellce), i.e., NASA Rotor 37 are 88 .65 %
til C')
~

o
u..
(FEjFO.1135), 1.2721 , and 2.1334, respectively. Table 4.2(b) shows that RBNN
~ 1.1 2.09 found the optimum point that gives the largest reduction in FEjI, 11.19% (about
.--
u.. 1.43% improvement in efficiency) in comparison with the reference shape. And,

2.05 PBA predicts the optimum point with the second highest efficiency. While in
~lzl~10 CO~lzl~10
COzifJo::: ~lzl~10
zifJo::: CO zifJo:::
a..coo:::~a..coo:::~a..coo:::~
case of F r, PBA, RBNN, and RSA show similar performances, i.e., 0.22%
0::: 0::: 0:::
increase in F r, RBNN gives the best optimum point with 0.94% increase in F p ,
when F1 is when F2 is when F3 is
optim ized optim ized optim ized followed by PBA . Therefore, RBNN predicts the best optimum points in all of
the cases. It is noted that in case of F EI], the performances of the surrogate
models to pred ict the optimum point are consistent with the reliabilities of
Fig. 4.3 RANS calculations at the optimum points predicted by different thesurrogate models discussed above in comparison with the RANS predictions.
surrogates (FI =FEjl' F2=Fra nd F3=Fp) However, in case of F p, RS A with the best reliability predicts the worst
optimum point.

70 71
T

0.40 ~~~~~'----------Q...---R-B-N-N-" 3.00 ~~~~~--.---------~

o ___ KRG o --Q- RBNN
o o ---KRG
x

<{ Z <{ <.9 <{ Z <{ <.9 <{ Z <{ <.9

aJ Z (f)
0:: aJ Z (f)
0:: aJ Z (f)
0::
Q 0:: :,.::: Q 0:: :,.::: Q 0::
_ I 0:: _ I _I _ I 0:: _I _ I _I 0:: _I :,.:::
aJ aJ tIl
c:o zz (f) <.9
<{ <{ <{ z <{ <.9 <{ z <{ <.9
_I
0. _ I 0. 0. 0. _I 0. 0. 0. _I 0. 0. (L 0::: c:o Z (f) 0::: c:o Z (f)
0:::
0 0. 0 0 0 0. 0 0 0 0. 0 0 c:o 0::: :,.:::: (L c:o 0::: :,.:::: (L
_ I 0::: _I _I _I 0::: _I _ I _I c:o 0::: :,.::::
0 0 0 a. _I a. a. a. _I a. a. a. 0::: _I _I
0 a. 0 0 0
_I a. a.
When F1 is When F2 is When F3 is a. 0 0 0 a. 0 0
optimized optimized optimized 0 0 0
When F1 is When F2 is When F3 is
optimized optimized optimized

(a) F E! (c) Fp

Fig. 4.4 Errors in predictions of surrogates at different points in des ign space
(FI = PEfh F2= P T and F3 = P p )
1.20 ,....,~~~~--r---------Q---R-B-N--"N
o
o ---KRG
Except the weighted average surrogate model, PBA, the surrogate model s
predict similar optimum points for P T and PI'. However, the best efficiency
points, i.e., the optimum des igns for PEll; are ve ry different from the PTa nd PI'
designs as seen by comparing Tables 4.2(b), 4 .3(b) a nd 4.4(b), and a lso in Fig.
4.3, which shows the va lues of objective functions at each optimum point.
However, this is quite natural since enhancing the efficiency and reducing the
<{ Z <{ <.9 <{ z (f) <{ <.9 <{ Z <{ <.9
c:o Z (f)
0::: c:o Z 0::: c:o Z (f)
0:::
Q 0::: :,.:::: Q c:o 0::: :,.:::: Q c:o 0::: :,.:::: number of compressor stages by increasi ng the pressure ratio are different tasks.
_ I aJ _I _ I _ I 0::: _ I
_I _I _I _I
a. 0:::
_ I a. a. _ I a. a. a. a. 0:::
_ I a. a.
0 a. 0 0 0 a. 0 0 0 a. 0 0 There can be situatio ns with good te mperature profiles w hile terrible in tota l
0 0 0
pressure, an adiabatic shock flow being a n example. But, for this compressor
When F1 is When F2 is When F3 is
optimized optimized optimized application, tota l temperature and tota l pressure ratios are expected not much
different to each other.
(b) Fr

72 73
As for the reliability of the surrogates, further evaluations are made in Fig. 4.4
and Table 4.5. Relative errors of surrogate predictions at twel ve different Sweep
optimal points in design space are shown in Fig. 4.4. As the same set of data is Lean
0.8
Skew
used for the optimization, four surrogate models give total 12 optimal points in
] 06
design space for FElT, F T and Fp objectives . And, root-mean-square averaged .n '
.,;
.c;
errors based on RANS calculations and surrogate predictions at these points are >0.4
given in Table 4.5 . In all of the cases, RSA shows the best accuracy while KRG
0.2
shows the worst. This is consistent with the results that the cross validation
errors shown in Tab les 4.2(a), 4.3(a), and 4.4(a) ind icate. In comparison wit h o I I
o 0.03 0.06 0.09 0. 12 0. 15
the results at the optimum points (Tab les 4.2(b), 4.3(b) and 4.4(b» discussed wf

above, the surrogates with best accuracy are the same except for F EjJ, and the
magnitudes of the errors are substantiall y larger.
(a) Design variab les
To show the sensitivity of weights on PBA, a simple average surrogate model is
tested. In this model, all surrogates are weighted equally to construct the model.
Tables 4.2(b), 4.3(b) and 4.4(b) show that the results of the simple average
mode l are only slightly inferior to the resu lts of PBA. .g
e
0.9 ~e r..-. ~ ~ 9 e 9 9 ~ • • f 2. \ \ 1:!
The resu lts of optimization w ith multip le-objectives (Eq. 2.5) are shown in Fig. '"

]1 "
0.898 5.
4.5. The design variab les except lean are dependent strongly on weighting ;>,
u
.: :
,
:
.
1.8\
]
~ 0.896 B Weighted sum I.S \ ~
factor, and all of them decrease as the weighting factor increases in the range of '0
~:~~i~~~:n . _~ _~.:~
e
~
~

weighting factor shown in Fig. 4.5(a). This means that distortion of stacking It 0.894
1.2 \ '"
line is effective in enhancement of the efficiency rather than the total pressure. 0.892 ~ -~- :
'"
-0
2
0.9\ ~
Fig. 4 .5(b) shows that total pressure ratio increases as weighting factor increases
just as intended.
089
. 0I 0.03 0.06
wf
0.09 0.\2 0. \5 ';S"

(b) Objective function values

Fig. 4.5 Results of multiple-objectives optimization

74 75
Efficiency based objective is used to ana lyze the sensitiv ity of the variables to
4-1.3 Flow characteristics with reference and optimum blades
objective function at near optimal point and presented in Fig. 4.6. This ana lys is
for the three design variables is performed to evaluate the parameter effect on Flow characteristics ins ide the blade passage for the reference and optimum

the object function, adiabatic efficiency. The optimal efficiency is Fop/and blades using the results of RBNN are analyzed in the following text.

deviating variable values by 10% from optimal point and objective fun ctio n
Figure 4.7 shows the limiting stream lines on blade suction surface for the
values (F) are calculated by polynomial equation 3.8 generated for RSA model.
reference and efficiency optimized blades . In this figure, three solid lines are
As shown in the figure , the blade sweep is the most sensitive on the object
drawn to present the 10, 75 and 90% span of the blade. The separation and
function . This means that the shape optimization using blade sweep is the most
reattachment lines are generated near the mid chord. This separation line
effective to increase an adiabatic efficiency in axial compressor. The blade lean
appears due to interference between the passage shock and the suction surface
is not so much effective to increase the adiabatic efficiency among the three
boundary layer. Here, it can be seen that the separation line is moved
design variables.
downstream in the optimized blade . The outward radia l flow caused by the
centrifugal effect also moves downstream of separation line .

Tip

0..
2£-5

°l -'-'-' ~- ' -'-

- Tip
90","pOIl

"5"'1'""
~

-'//f(~
0
>.t.<
...... -2£-5 ~ Sep:u<1liOll Lute
~
...,
0..

&: I -4£-51 , " 1- - - - - sweepi

>.t.< ' - - -- Lean
'-' '
-6£-5 " Skew

-8£-5 I I I I
-10 -5 0 5 10 (a) Reference blade (b) Efficiency Optimized blade
Deviation from Optimal Point (%)

Fig. 4.6 Sensitivity a nal yses for optimum shape by RSM

Fig. 4.7 Limiting streamlines on the blade suction surface

76
77
 (c) 10% span

Reference blade Efficiency Optimized blade

Fig. 4.8 Mach number contours on the planes of 10, 75 and 90% span
(interval of contour lines =0.1)

(a) 90% span

number contours at 10, 75 and 90% span, depicts that the interference position
of the passage shock with the blade suction surface boundary layer is moved
towards downstream for the optimized blade as compared to that of the
reference blade at 75% and 90% span. At 10% of span, the figure shows that the
interference position of the passage shock did not change between the optimized
and the reference blades. This can also be seen in Fig. 4.6 that, at 10% span
there is not much difference between the reference and optimized blades for
separation and reattachment lines.

Figure 4.9 shows the temperature distributions on suction and pressure surfaces
of the reference and the optimum blades. Remarkable differences in temperature
(b)75% span distribution are not found between the cases.
Reference blade Effic iency Optimized blade
The distributions of static pressure on surfaces of optimum blades are shown in
Fig. 4.10.Pressure changes abruptly across the separation line on suction
surface. And, in case of efficiency optimized blade, it is noted that near tip of
leading edge on pressure surface the pressure shows the minimum while the

78 79
 Tip

H"b

Pressure surface Suction surface

Pressure surface Suction surface

(a) Reference blade

(c) Pressure ratio optimized blade

Pressure surface Suction surface

(b) Efficiency optimized blade

Pressure surface Suction surface

(d) Temperature ratio optimized blade.

Fig. 4.9 Temperature contours on surfaces of optimum blades

80 81
 Tip

1\

Pressure surface Suction surface

Pressure surface Suction surface

(c) Press ure ratio optimized blade

(a) Reference blade

,\

Pressure surface Suction surface Pressure surface Suction surface

(b) Efficie ncy optimized blade (d) Temperature ratio optimized blade

Fig. 4 . lO Pressure coefficient contours on surfaces of optimum blades

82 83
relatively high pres sure region is located in this place in the other cases. And, Casing Casing
also on suction surface, efficiency optimized blade shows that the pressure
decreases more rapidly as the tip is ap proached than the other optimum blades.
Therefore, it is found from this figure that enhancement of efficiency is 0.55

achieved by modifying the flow near the tip of the blade. This is consistent with
PSI rss s 0.6
the resu lts of Jang et a l. [60], who suggested that the optim ization of three - /
0.6
dimensional stacking line using the blade sweep, lean and skew reduces the ti p
losses. Gallimore et al. [15] also reported that a positive lean reduced a hub
........ Hub
corner and tip clearance losses except near the mid-span region. Since the blade
profiles are similar for total temperature ratio and total pressure ratio optimized
blades, the distributions shown in Figs. 4.10(c) and 4. 10(d) are also simil ar. (a) 30 percent chord from blade leading edge

Figure 4 . 11 , which is the perspective view from downstream, shows the

distributions of axial velocity normalized by sound velocity at the 30 percent
chord from blade leading edge and at the blade trailing edge. As shown in Fig. Casing Cas!:L-____....
...- , a~ _Sj

4.11 (a), relatively low axial velocity region observed near the blade suction side
0.55
of the casing for the reference blade is recovered by optimiz ing the blade "-0.55
stacking line . Uniform axial velocity is also observed at the blade trailing edge
for optimum blade.

Figure 4. 12 shows the distributions of the vorticityon six planes nearly

perpendicular to the tip leakage vortex and leakage streamlines surrounding the
Hub Hub
vortex core. In the figure, the vorticity is shown only from zero to seven. The
leakage vortex formed on the blade suction surface by the induced veloc ity
(b) Blade trailing edge
generated between a leakage jet flow and a main through flow is shown in the
blade passage. For both cases, the leakage vortex is tightl y rolled up at upstream
Reference Optimum by RBNN
of the bow shock with the high vorticity. The vorticity, shown upstream of the
bow shock, concentrates on the leakage vortex core. However, the di strib utio n
Fig. 4. 11 Axial velocity normalized by sound ve locity
of vorticity expands widely just downstream of the bow shock due to the
(interval between contours = 0.05)

84 85
 Interference between the leakage vortex and shock wave. As shown on the plane
IV, the regio n having a vo rtic ity above ze ro is spread out to abo ut 10 percent of
span from the rotor tip of the re ference bl ade . It is found that re lati ve ly small
region having a vorticity above ze ro is formed on the plane IV as co mpared to
the refe re nce blade. This is ma inl y caused by weake ned Mach number as shown
Leakage
Sire amline s in Fig. 4.8. It is noted that the red uced vertical flow in the optimum blade resu lts
in uniform ax ial ve loc ity as shown in Fig. 4 . 11 .

Fig. 4.1 3 shows the three-dimensional blade shapes of the refe rence a nd the
(a) Refere nce
optimum roto r blades. In the fi gure, solid line represents the reference blade. It

(b) Optimum by RBNN

Fig. 4. 12 Distributio n of vo rticity on the quas i-o rthogo na l planes to the leakage
Fig. 4. 13 Optimum a nd refe re nce blade shapes
vo rtex a nd leakage streamlines
(interva lbetwee n contours= 1.0)

is noted that the blade swept to backward and skewed to the blade pressure side
is mo re effecti ve to increase the adiabatic effi ciency in the ax ia l co mpressor.

86 87
From the above fi gures, the increase of adiabatic efficiency for the opti mum two objectives; obviously, the Pareto curve shows the improvement of these
blade is caused by reduction of the bow shock upstream of blade leading edge . obj ectives from the reference blade . To show the use of Pareto curve, one point
The reduced Mach number induces concentrated vorticity at downstrea m of the is ra ndomly selected on Pareto curve and computed using RANS anal ysis. The
bow shock, thus, makes a uniform axial velocity as compared to reference point is shown in the fig ure and the improvements of objectives are vis ible.
blade . Table 4.7 shows the optimal design variables and the objective function va lues
at these points. When only Fp is optimized, total pressure ratio is increase by
4-1.4 Multi-objective optimization through MOEA 1.62% while the efficiency is reduced by 0.04% in comparison with the
reference blade. Simil arly, if onl y effic iency is opti mized as si ngle objecti ve
As in the previous case the design variables are touching boundary of design
problem the increase in effic ienc y is IAI % and reduction in total pressure ratio
space (Table 4.1) fo r the total pressure based optimization (Tab le 4A), the
is 1.55%. Obviously, these two single objective optimi zations show if one
design space (Ta ble 4.6) is changed and optimized agai n consideri ng adiabatic
objective is optimized the other is reduced. In this table and on the Fig. 4. 14,
efficie ncy and tota l pressure obj ecti ves. Here si ngle objective optimization is
another point is se lected randomly from NSGA-ll results and computed by
also performed and a multi-objective optimizatio n by MOEA is applied. Here,
RANS analysis. The result shows the efficiency a nd tota l pressure rat io are
first, the single obj ect ives were used to construct RSA surrogates and used fo r
increase by 0.51 % and 1.25%, respectively. It is noted that, when the efficienc y
single objective optimi zation of these functions was used as input function of
is optimized, the blade sweep (a) is pos iti ve while when the total pressure ratio
NSGA-Il.
is optimized, a is negati ve.
The problem for two objecti ves was handled using single as well as multi-
objective optimi zation. The Pareto optimal front shown in Fig. 4.14 is obtained
using NSGA-Il w ith £-constra int strategy fo r two objectives; tota l pressure ratio Table4.6 Design va riables and ranges
(Fp) and adiabatic efficiency (I'/ad) . From the figure, it is seen that if one
Lower Upper
obj ective is maximized the other objecti ve va lue is reduced. The RANS anal ysis Variables
limit limit
resu lt for the reference blade is also shown in the figure . The single objective Sweep
-0. 126 0.252
(a)
optimization for I'/ad shows a ga in in efficiency and a loss in F p. Simi larly, single
Lean (fJ) -0.036 0.0
objective optimization for Fp shows a gain in Fp a nd loss in I'/ad. The single
Skew (y) 0.0 0.1
objective optim izations show that the values of th e objecti ves are located at
extreme ends of Pareto optima l front. The Pareto optimal front thus, presents
a choice to the designers to choose the obj ective function va lues acco rdi ng to
their design needs . The RANS computed points are shown in this fi gure for
these single objective optimizations. Since, NSGA-ll is the compromise of these
88 89
 Since the overall performance of turbomachinery blade is not a function of
Table 4.7 Resu lts of optimizations single objective, multi -objective optimi zation is the proper strategy for
turbomac hinery blade opti mizatio n. The Pareto optimal so lutions provide useful
(a) Design variables data for the designer who should select the final optimum design.

Optimal points 4-1.5 Flow analysis

Objectives
From Fig. 4.14, it is seen that the single objectives give optimal points at
0. fJ y (radians)
extreme ends of the Pareto optimal front. Hence, the followin g flow analyses
Fp -0.093 -0.007 0.017
have been performed for these two optimized blades as well as the the reference
IJad 0.102 -0.009 0.060 blade for the purpose of comparison.

A NSGA-II
-0.019 -0.004 0.Q35 Figs. 4. 15 shows the spanwise distribution s of adiabatic efficiency (IJad), total
Multi-objective result
temperature ratio (F r), total pressure ratio (Fp) , and Mach number, respectively,
for the reference and the two optimum blades in comparison with the
experimental res ults (Reid and Moore, 1978) for the reference blade . [t is found
that the computational results for the reference blade agree we ll with the

(b) Objective functions experimental results (Reid and Moore, 1978) throughout the span. The IJadin
Fig. 4. 15 is obtained by mass averaging the local values tangentially. The
maximum difference of the local IJad between the computational and
RANS calculated Objective function
I experimenta l results at the design flow rate is less than 4%. This fi gure shows
Objectives Values from
Refere nce shape % increased the IJad is largely increased near 75% span and reduced near 25% span when the
optimal shape
efficiency is optimized. The pressure optimized blade shows a lmost negligible
Fp Fp IJad Fp IJad Fp IJad
change in efficiency all a long the span. Fig. 4. 16. shows the significa nt
IJad 2. 168 0.886 1.62 -0.04
reduction in total temperature ratio, Fr near 75% span in the efficiency
A NSGA- ll 2. 1 0.899 2. 1334 0.8865
-1.55 IAI
optimized blade. Near 10% span, the Fr is increased for both the pressure and
Multi-objective the efficiency opt imized blades .
2. 16 0.891 1.25 0.51
result

90 91
 2.2 . - - - - - - - - - - - - - - - - - - - ,
N SGA-H ge nerated po int s
100 -r-.. . . .~-------....._~
2.18 75 +---------~
_ _ Pressure optimized
.Q c blade
- Q . - Efficiencyoptimized
'§2 .16

\. '"
'"fir 50 blade
l!! <f'. _ _ Refe rence blade
:::J
rJ)
~2 . 14 Press ure A NSGA-U
•
. result 25
a. o ptunum

~2 . 1 2 \ (RANS) (RANS)
f- o ~~~~~~~~~~-~
Efficiency • 1 .12 1.17 1.22 1.27 1.3 2
2.1 Reference
opt inlUm / Total temperature ratio
(RANS)
(RANS)
2.08 +----,----r--.---.----..,.--r--...-----,,.-J

0.885 0.889 0.893 0.897 0.901

Fig. 4. 16 Spanwise total temperature ratio distributions
Adiabatic efficiency

Fig. 4. 14 Pareto optima l design

75 -/----::----::--:--:-,bl,
__ Pressure optimized
blade
c - Q . - Effi ciency optimized
100 .,._ _ _ ~_- __(>----, 'fir" 50 blade
__ Reference blade
~

75 .t---------~:I..
~ Pressu re optimized 25
blade
c -"'- Effi ciency optimized
'"g- 50 blade
__ Reference blade
<f'.
1.45 1.75 2 .05 2.35
Total pressure ratio

o +----.----.---~~~w Fig. 4.17 Spanwise total pressure ratio distributions

0.69 0.76 0.83 0.9 0.97
Adiabatic effici ency

Fig. 4. 15 Spanwise effici ency distribution s

92 93
Large reduction in Fp is observed near 80% span when blade is optimi zed fo r Fig. 4. 18. It is visible that for the reference and the pressure optimized blades,
the efficiency as shown in Fig. 4.17. On the other hand, the Fr is increased in there is little change in Mach number profile. But, the efficiency optimized
the range from 25 to 80% span for the pressure optimized blade. blade (Fig. 4 . 19(b)) shows separation line moved downstream. This movement
of the separation line increases efficiency by reducing losses.
Fig. 4 . 18 shows the Mach number distributions along the span. The reference
and the pressure optimized blades show almost same results a ll a long the spa n Fig. 4.20 shows the limiting stream lines on pressure a nd suction surfaces of the
while the efficiency optimized blade shows the large reduction in Mach number three blades. Separation lines are formed near the mid c hord due to the
interference between the passage shock and the suction surface boundary layer.

100 r---.iii&A::---~ It is found that near the tip on suction surface the separation lines moves la rge ly
downstream for the efficiency optimized blade. This is consistent w ith the

75 +---------------~ results shown in Fig. 4. 15, where relative high increase in efficiency is observed
-+- Pressure optimized
blade near 75% spa n for the e fficienc y optimized blade. An attachment line is also
c: -0- Effi cie ncy optim ized
!tl
g- 50 blade observed behind the separatio n line for all the three blades.
-Ir- Reference blade

25

0.5 0.6 0.7 0.8 0.9

Mach number

Fig. 4.18 Spanwise Mach number distributions

near 80% span.

[n Fig. 4.19, Mach number distributions are shown for the reference (Fig.
4. 19(a)) and two optimum blades (Fig. 4. 19(b, c)) obtained from si ngle
obj ective optimi zations. Mach number contours are drawn at 80% of span,
where the Mach number is reduced largely by the efficiency optimized blade in

94 95
 -----
-lip -----"

~,
~E
LE_

Hub -------'

(a) Reference blade

\ \
"\

(a) Reference blade (b) Efficiency optimized blade

(b) Efficiency optimized blade

~
(c) Total pressure ratio optimized blade
~
(c) Pressure optimized blade
Fig. 4.19 Mac h number contours at 80% span Suction surface Pressure surface
Fig. 4.20 Stream lines

96 97
4-2 Blade stacking line thickness modificatio n

4-2.] Optimization res ults

Another optimization by ANSYS-CFX 11.0 has been performed considering Table 4.8 Design space
stacking line and airfo il shape. Validation of the present results of numerica l
Variables normal Variables along Blade lean
ana lysis wit h experimental data [13] for the reference blade has been performed Limits
to chord line (%) chord line (%) (radians)
of
as in Fig. 4.21. The numerical results for total pressure ratio (Fp) agree well variab les Ot Oil G G, Pt Pm
with experimenta l data for the reference shape. But, the resu lts for adiabatic
Lower -22 .52 -6.00 -5 -5 -0.105 -0.035
efficiency (FEfJ) are underestimated with in maximum 3% relative error. The
Upper 22.52 6.00 15 15 -0.035 0.035
previous works [14, 63, 82] a lso reported the underestimated results for the
effic iency.

Table 4.9 Resu lts of optim izations

4-2.2 Validation of numerical simulation
(a) design variables

Proper determination of desi gn space constituted by lower and upper limits of Variab les
Variables
norma l Blade lean
the design variab les is very important to reduce the iterat ion in present a long
Designs to chord line (radia ns)
chord line (%)
optimization procedure shown in Fig. 3. 1. To decide the range of each variable, (%)
effects of each variable on objective functions are tested first. This needs some Ot Oil G G, Pt Pm
pre liminary designs to s imulate to set the ra nges of variables. The design Ref 0.000 0.000 0.000 0.000 0.000 0.000
variables for blade lean are presented in radians, and those for blade thickness Ijad, opt -2 .74 1 -4.034 7.048 12.267 -0.10 1 0.002
are defined in the form of percentage change as shown in Tab le 4.8. F p.opt 6.379 4.804 0.932 6.675 -0.035 -0.003

The surrogates are problem dependent and proper distribution of design points F.,j.opt -1.432 -4 .016 8. 149 11.531 -0.094 -0.003

in design space is necessary for better constructions of surrogates. Better

constructed surrogate implies a better curve fitting in surrogate model and less
error in prediction of optimum design. As a lready stated, the PBA model
havingthe property of robustness in prediction is used in this problem.

98 99
 (b) Objective functions

Predicted RANS computed % improvement Table 4.10 Weights assigned to surrogates to construct PBA model.
Designs
F Sllrrogate Fp FEjl Fp F EU
Ref --- 2. 105 85.251 --- ---
(a) For /]ad,opl blade
I] ad, opt 86.74 2.055 86.684 -2.375 1.681
2.115 2.110 85.762 0.238 0.599 MODEL Cross validation error(Ecv ) Weight
Fl'. 01'1
RBNN 0.259 0.261
2.067 86.681 -1.805 1.677
Fwlop, 0.9966 RSA 0.165 0.400
FHr= 0.9908 FwF 1.23
KRG 0.197 0.339

The number of designs selected by LHS method on the basis of RSA model is
55 for 6 design variables.

(b) For FI'.op, blade

T he above designs are evaluated using RANS solver to get the FI', FE/] and FHI
objectives. In multi-objective optimized (PW!:Opl) design, Fw/ is calculated with MODE Cross validation
Weight
L error(Ecv )
wF 0.06.
RBNN 0.060 0.295
Tab le 4.9 represents the reference and optimum shapes and their objective RSA 0.060 0.295

function values. It is found that PBA gives excellent predictions for optimum KRG 0.042 0.410

efficiency and total pressure in comparison with the values calculated by RAN S
analysis. For efficiency optimized (/]ad,opl) blade, the blade lean shows that if
blade is moved towards the direction of rotation FE/T is increased. The blade
(c) For FHf,opl blade
thickness is little reduced and maximum thickness is moved toward trailing
MODE Cross validation
edge (TE). TheFEJT is increased by 1.681 % while F p is reduced by 2.375 % in Weight
L error(Ecv)
this case. RBNN 6.78E-03 0.220
RSA 4.13E-03 0.353
I(,~G 3.38E-03 0.426
- - - - - - - - - -------

100
101
 fn the same table (Ta ble 4.9), it is seen that the Fp and F EUare increased by
0.9 2.4 0.238% a nd 0.599%, respectively, in the case of total pressure ratio optimized
¢U__ - -_ ..... - - - - - --...... (F p.opt) blade. The blade moved towards the rotation direction. Maximum
» 0.86
u .S'
thickness is increased a little and maxim um thickness is moved towards the
c
Q) - 2.2 ~
~ 0.82 ~ trailing edge (TE). F "f,opt blade shows the objective function F wl (Eq. 2.5) value
Q)

u
c? ::::J
(f)
(f)

~ is increased by 1.225%while F Eff is increased by 1.677%, and F p is reduced by

20.78 - - - - Ref, Expt. ~
<:L

2 1.805%when wr O.06 is ass umed.

'"
'5 - - - Ref, RANS 0'"
<{ I-
0.74 --*"- 7lad,opt , RANS
_ _ Fpppl, RANS
----lIf-- Fwf.opt, RANS
0.7 1.8
0.94 0.95 0.96 0.97
Normalized mass flow

Fp.op,
Fig. 4.21 Result validations with experimental data

Fig. 4.22 Variables and objective function values with different weighting
factors

102 103
The objective function va lues of optimum designs computed by RANS analysis
presented in Fig. 4.21 . Obviously, the F Er( and Fp have increased in 'lad,opt and
F p.opt blades in the vast operating range by this optimization, respectively. The
F "fopt results show that the ad iabatic effic iency line is almost co inciding with
the 'lad,opt result line but tota l pressure ratio line is showing some difference wi th
the same result. This is a lso seen from the Table 4.9(b) that the F EfJ for the
I'/od,opt blade is same as that of the F "f.opt blade while that of the Fp is different fo r Ref f/ ad, opt

these cases.

F p.opt F "elopt

(a) 10% span

The RSA, KRG and RBNN surrogates are constructed and errors in data fitting
are calculated through their k-fold cross validation error. These errors are used
to find PRESS and weights. Finally PBA model is constructed by as per Eq .
3.26. In Tab le 4.9, the weights ass igned to each surro gate (RSA, KRG or
RBNN) to construct PBA model are presented.

104 105
 Ref 'lad, opt

Ref 11(/t/,opt

Fr. 01" Fllj:OPI

(b)50% span

F p,opt FI~jopl

(c) 90% span

The least cross validation error (Ecv) is produced by RSA model and hence the
weight ass igned to RSA to construct PBA model is highest for 1'/ad, opl blade
Fig. 4.24 Mac h number contours at different span of blade
(Table 4.10(a». In the case of Fp.opi blade (Table 4.10(b», the Ecv is equal for
RSA and RBNN and hence we ights are also equal for these sUlTogates .

106
107
KRG produced lowest Eev and hence the weight assigned to KRG is hi ghest. For In present problem, the third objective (F IV() which is the combination of FElr and
F uiopt case (Table 4.1 O( c)) also, KRG produced highest weight as it produced Fp is made through a weighting factor; w (. Fig. 4.22 shows variations of the six
least Eel" As better performing surrogate is having lower error (Ecv), and hence design variables and objective function (F'1) values with weighting factor, wJ
higher weight is assigned to the surrogate. This protects PBA model from poor
performance even when any surrogate performs bad for prediction. The
performance of indi v idual surrogate is strongl y dependent on the problem as
presented in case I.

(a) Ref

(a) Ref (b )'lad, opt

(b) 'lad, opt

(c) F p.opt (d) F u{:opt when F ul is optimized. The variables presented are normalized in 0 to I range.
With increasing the contribution of F" in F",I' the objective function value
increases. Variable of blade lean at tip (Yt ) gets increased while the variable of

Fig. 4.25 Mach number contours on blade suction surfaces blade lean at mid point (YIIl) is not affected so much with increase of wJ' Blade

l08 l09
 4-2.3 Flow characteristics with reference and optimum blades

< .~ Figure 4.24 shows the Mach number profiles at 10, 50 a nd 90% of span for the

~ reference and optimized blades (wr O.06). Near hub (10% span), the Mach

~
~
number profiles are almost same for the reference a nd optimized blades. Similar

~
trend is also shown for the 50% span for all the three reference and optimized
--
- ~~ blades. However, this trend is changed for the near tip (90% span) region. The
visible change is found in the case of 'lad,opi and F lV/ Opl blades . The shock lines
are moved towards TE, and the separati on zo ne is also moved towards TE. The
(c) F p.opl shock intensity is a lso reduced. For pressure optimized blade, the change is
visible but less as compared to 'lac/,opi or F lV/ Opl blade. Another figure of Mach

~
numbers (Fig. 4.25) on suction surfaces of the blades shows the shock beyond
50% span is moved towards TE in all cases . As the FI'.opl blade has the low

~
increase in Fp and that 's the reason why increase in pressure has not affect

~ much the flow structure.

==-~
~. Figure 4.26 shows that streamlines on suction surface are change noticeably by
the optimizations. The 'lac/, opi or F"oj;opi blade shows the separation line is moved
Largely towards TE beyond 50% span of the blade. However, F p.opi blade shows
(d) F lV/ Opl
only slight changes.
Pressure surface Suction surface
Pressure contours on blade suction surfaces are shown in Fig. 4.27. The
Fig. 4.26 Stream lines on blade surfaces contours near mid chord above 50% of span moved towards TE for the
optimized blades. This shows the same pattern as the Mach number contours on
suction surfaces shown in Fig.4.25. The shock loss decreases with en largin g the
front flat portion of the blade profile. The maximum thickness location is
thicknesses (a, and all) gets reduced while location of maximum thicknesses (P,
moved towards TE increasi ng the fl at portion and hence increasi ng efficiency.
and Ph) of blade doses not change remarkably with the increase of wI' Here , it
Stacking line and a irfoil profile modifications also produce the weaker shock
can be noticed that wrO.O implies the 'lad, opl blade . Reference and optimized
strength [60].
blade shapes are presented in Fig. 4.23.

11 0 III
 (c) F p.opt

(a) Ref

(b) 'lad. apt (d) F w/: opt

Pressure surface Suction surface

Fig. 4.27 Press ure contour at blade surface

As the blade thickness is changed and the blade is leaned due to optimization, a
structural ana lysis can be performed to check the endurance limit and flutter of
blade, this has not been studied in the present work. leading edge on pressure surface for 'lad. opt and F wf,opt blades which show a lmost
same loading profiles. Fp,opt blade shows only so me decrease in loadin g near
Blade loadings at di fferent span locations are presented in Fig, 4,28, At mid-
leading edge on pressure surface in compariso n with the reference blade ,
span, almost same loadings are shown for different blades. But, at near tip
location (at 80% span), a large variation in loading is shown especially near

112 11 3
 150000 - t 4 ; - - - - - - -- - - - - - ,
ro
0.. 100000
'2
ro
150000
~ 50000
ro 0~
0.. 100000 0
~
'2
ro CJ)
0
CL c:
(/) 50000 '6
0~ ~ -50000
0 -l
~ 0
CJ)
c: --Ref -100000 +1--~--.-----r--..,_-__1
'6
~ -50000 Eft_ opt o 0.2 0.4 0.6 0.8
-l 6 Pt opt Streamwi se location
. .. . ... MUlti_opt
-100000
0 0.2 0.4 0.6 0.8
Stream wise location (c) 80% span

Fig. 4.28 Blade loading

(a) 20% span

150000
ro
0.. 100000
'2
ro
CL
(/) 50000
~
0
0
~ 0
CJ)
c:
'6
~ -50000
-l

-100000
0 0.2 0.4 0.6 0.8
Streamwise location

(b) 50% span

114 115
 Chapter 5

CONCLUDING REMARK

The shape optimization of rotor blades for a transonic ax ial compressor has
been performed by the single- as well as multi-objective techniques and the
three-dimens ional Navier-Stokes analysis. The adiabatic efficiency, total
--- pressure ratio and tota l temperature ratio are considered to be the objectives of
the problem. Response surface approximation, Kriging, radial basis neural
network and weighted average models have been evaluated for their prediction
capabilities in turbo machinery applications.

For single objective optimization with stacking line modification of blade, about
1.4% of relative increase in blade efficiency is obtained . Other objective
functions , i.e ., total temperature and total pressure ratios are also increased. The
evaluation of surrogate models shows KRG gives the highest cross va lidation
error and performs the worst for all of the objective functions. RSA, with the
lowest cross validation error, shows the best reliability of the surrogate in some
cases. However, RBNN predicts the best optimum points in all cases. Even
though the weighted surrogate PBA is not the best in any case, it gives
satisfactory results in most cases and protects against choosing a poor surrogate
if only one is chosen. Therefore, the simultaneous application of multiple
surrogate model offers enhanced robustness in the optimization process. The
present approach can help address the multi-objective design on a rational basis
with quantifiable cost-benefit estimates.

In multi-objective optimizations by £-constraint NSGA-IJ a lgorithm coupling

with RANS analysis has been performed to modify stackin g line and optimized
for total press ure ratio and adiabatic efficiency. NSGA-JI w hich produces Pareto
optimal front of co nflicting objectives produces compromised sol utions which

116
11 7
gi e enhanced results of both objectives. Thes ingle objective optimizations REFERENCES
show that the va lues of the objectives are located at extreme ends of Pareto
[1] Japikse, D., 1976, "Review - Progress in Numerica l Turbomachinery
optimal front. Flow analysis results show that large reduction in total pressure
Anal ysis," ASME Transactions, Series I-Journal of Fluids Engineeri ng.
ratio is observed near blade tip for efficiency optimized blade whi lethe total
Vo l. 98, pp . 592-606.
pressure ration is increased in most of the spanwise region forthe pressure
[2] Horlock, J.H. an d Marsh, H. , 1982, "Fluid Mechanics of Turbomachines: A
optimized blade. In the efficiency optimization, the efficiency is improved by
Review,"Internat ional Jo urna l of Heat and Fluid Flow, Vol. 3, Issue I , pp .
shifti ng separatio n line downstream on suction surface near tip of the blade . It is
3- 11 , doi : 10. 10 I 6/0 142-727X(82)90036-4 .
also found that the optimized rotor blades can reduce the Mach number of a
[3] Mcnally, W.D. and Sockol, P.M ., 1985, " Review - Computationa l Methods
bow shock at the span where the tip leakage vortex is located.
for Interna l Flows with Emphasis on Turbomac hinery," ASM E
The blade thickness and stacking modification with PBA surrogate mode l, it is Transactions, Journa l of Fluids Engineering, Vol. 107, no . I , pp. 6-22.
found that when adiabatic efficiency is considered as objective, adi abatic [4] Casey, M.V., 1994, "Computationa l Methods for Pre liminary Design and
effici ency is increased by 1.68% while total press ure ratio is reduced by 2.68%. Geometry Definition in Turbomac hinery," AGARD, Turbomachinery
And, when total pressure is considered as objective, the tota l pressure rati o is Design Using CFD, (N95-14127 03-34).
increased by 0.24% a nd adiabati c effic iency is increased by 0.60%. The multi - [5] Danton, J. and Dawes, W. , 1999, "Computational Fluid Dynamics for
obj ective optimizati o n combining these two objectives produced enhanced Turbomachinery Design," Proceedings of the Institution of Mechanical
effici ency with less total pressure ratio decrement. Engineers, Part C : Journal of Mechanical Engineering Science, Vol.
213 , No.2, pp. 107-124(1 8).
For efficienc y optimized blade, the sepa ration and attac hment lines moves
[6] Horlock, 1. H., and Denton, J. D., 2005 , "A Review of Some Early Des ign
towards trai ling edge of blade enhancing the blade performance. The blade
Practice Using Co mputational Fluid Dynamics a nd a C urrent Perspective,"
maximum thickness is reduced and mo ved towards the trailing edge increasing
Journal of Turbo machinery, Vol. 127, Issue I , pp. 5-13.
the fl at portion of the blade front area. The blade max imum thickness is
[7] Molinari, M. and Dawes, W.N., 2006, " Review of Evo lution of Compresso r
increased and mo ved towards the tra iling edge for pressure optimized blade.
Design Process and Future Perspecti ves," Proceedings of the Institution of
Blade is leaned towards the rotatio na l direction in both the pressure and
Mechanical Engi neers, Part C: Journal of Mechanical Engi neering Science ,
effi c iency optimized blades.
Vol. 220, No.6, pp . 76 1-771.
[8] Hirsch, c., 1994, "CFD methodol ogy and va lidation for Turbomachinery
flo ws," AGARD, Turbomachinery Design Using CFD (N95- 14I 27 03-34).

11 8 119
[9] Myers, R.H. and Montgomery, D.C., 1995, " Response Surface [18] Watanabe, H., and Zangeneh, M. , 2003, " Design of the Blade Geometry of
Methodology-Process and Product Optimization Using Designed Swept Transonic Fans by 3D Inverse Design," Proceedings of ASME Turbo
Experiments," John Wiley & Sons , Inc : New York. Expo, GT-2003- 38770.
[10] Martin, 1. D. and Simpson T. W., 2005 , "Use of Kriging models to [19] Denton, 1. D., and Xu, L., 2002, "The Effects of Lean and Sweep on
approximate deterministic computer models," AIAA Journal, Vol. 43, No. Transonic Fan Performance," Proceedings of ASME Turbo Expo, GT-2002-
4,pp. 853-863. 30327.
[II] Orr, M.J.L. , 1996, " Introduction to Radial Basis Neural Networks" Center [20] Cai, N. , Xu, 1., and Benaissa, A., 2003, "Aerodynamic and Aeroacoustic
for cognitive science, Edinburgh University, Scotland, UK. Performance of a Skewed Rotor," Proceedings of ASME Turbo Expo, GT-
http://anc.ed.ac.uklrbf/. 2003 -38592.
[12] Goel, T. , Haftka, R., Shyy, W. and Queipo, N., 2007, "Ensemble of [21] Fischer, A., Riess, W., and Seume, J. , 2003, "Performance of Strongly
Surrogates,"Structural and Multidisciplinary Optimization, Vol. 33, No.3 , Bowed Stators in a 4-Stage High Speed Compressor," Proceedings of
pp . 199-216( 18). ASME Turbo Expo, GT-2003-38392.
[13] Reid, L., and Moore, R . D. , 1978, "Design and Overall Performance of [22] Jang, C.M., Li, P. and Kim, K.Y., 2005, "Optimization of Blade Sweep in a
Four Highly-Loaded, High-Speed Inlet Stages for an Advanced, High- Transonic Axial Compressor Rotor," JSME International Journal-Series B,
Pressure-Ratio Core Compressor," NASA TP-1337. Vol. 48, No.4, pp.793 -801.
[14] Dunham, J., 1998, "CFD Validation for Propulsion System Components," [23]Jang, C.M. and Kim, K.Y., 2005, "Optimization of a Stator Blade Using
AGARD Advisory Report 355, ISBN 92-836-1 075-X . Response Surface Method in a Single-Stage Transonic Axial Compressor,"
[15] Gallimore, S. J., Bolger J. J., and Cumpsty, N. A. , 2002, "The Use of Sweep Proceedings of The Institution of Mechanical Engineers, Part A-Journal of
and Diahedral in Multistage Axial Flow Compressor Blading, Part I: Power and Energy, Vol. 2 19, No.8, pp.595-603.
University Research and Methods Development," Proceedings of ASME [24] Ahn, C.S. and Kim, K.Y., 2003, "Aerodynamic Design Optimization of A
GT-2002-30328. Compressor Rotor with Navier-Stokes Analysis," Proceedings of The
[16] Bliss, D. B., 1976, "Method of and Apparatus for Preventing Leading edge Institution of Mechanical Engineers, Part A-Journal of Power and Energy,
Shock Related Noise in Transonic and Supersonic Blades," US Patent Vol. 217, No.2, pp . 179-184.
3989406. [25] Lee, S. Y. and Kim, K. Y., 2000, "Design Optimization of Axial Flow
[17] Hah, C., Puterbaugh, S. L., and Wadia, A. R. , 1999, "Control of Shock Compressor Blades with Three-Dimensional Navier-Stokes Solver," KSME
Structure and Secondary Flow Field Inside Transonic Compressor Rotor International Journal, Vol. 14, No.9, pp. 1005-1012.
Through Aerodynamic Sweep," Proceedings of ASME 99-GT-561. [26] Choi, J. H., Kim, K. Y. and Chung, D. S., 1997, "Numerical Optimization
for Design of an Automotive Cooling Fan," Journal of Passenger Cars -
SAE 1997 Transactions, Vol. 106, Section 6, Part I , pp. 1485-1489
120 121
[27] Kammerer, S., Mayer, J. F., Paffrath, M., We ve r, U. and Jung, A. R., 2003 , [3S] Li, H.D., He, L., Li, Y.S ., and Wells, R., 2006, "Blading Aerodynamic
"Three Dimensional Optimi zation of TurbomachineryBladings Using Desi gn Optimization wi th Mechanical a nd Aeromec hanical Constraints",
Sensitivity Analysis," Proceedings of ASME Turbo expo 2003,Power for ASME Turbo Ex po 2006, Barcelona, Spain, GT2006-90S03 .
Land, Sea and Air, June 16-19, At lanta Georgia, USA, GT2003-38037. [36] Buche, D., Guidati ,G., and Sto ll, P. , 2003, "Automated Design
[28]Cai, N . and Xu, J., 2001 , "Aerodynamic-Aeroacoustic Performance of Optimization of Compressor Blades for Stationary Large Scale
parametric Effects for Skewed-Swept Rotor," Proceedings of ASME turbo Turbomachi nery, " ASME Turbo Expo 2003, Georgia, USA, GT2003-
expo 200 1, June 4 -7,2001, New Orleans Louisiana, USA, 200 1-GT-03S4. 3842 l.
[29] Gummer, V., Wenger, U. and Kau, H. P., 2001, "Usi ng Sweep and [37] Benini, E., and Bio llo, R., 2006, "On the Aerodynamics of sweep and
Dihedral to Control Three Dimensional Flow in Transonic Stator of Axi al Leaned Transonic Co mpressor Rotors," ASME Turbo Expo 2006,
Compressor", Transactions of ASME, Vol 123 , pp 40-48. Barce lona, Spa in, GT2006-90S47.
[30] Sonoda, T., Yamaguchi, Y., Arima, T ., Olhofer, M., Sendho ff, B. and [38] Jun, L. , Guojun, L. , Zhenp ing, F., and Lijun, L. , 200S , "Multiobjective
Schreiber, H.A., 2003, "Advanced Hi gh Turning Co mpressor Aerofo ils for Optimization Approach to Turbomachinery Blades Design," ASME Turbo
Low Reynolds Number Condition, Part I: Des ign and Optimization," Expo 200S, Reno-Tahoe, Nevada, USA, GT200S-68303.
Proceed ings of ASME Turbo Expo 2003, June 16-19, 2003, Georgia, USA, [39] Kammerer, S., Maye r, J. F. , Stetter, H., Paffrath, M. , Wever U. , and Jung,
GT2003 -3 84S 8. A. R. 2004, "Deve lopme nt of a Three Dimensional Geo metry Optimization
[3 1] Papi la. N, Shyy, W ., Grif, L., and Dorney, 0.1. , 2002, "Shape Optimization Method for Turbomachinery App lications," Inte rnational Journa l of Rotary
of Supe rsonic Turbines Using Global Approximation Methods," Journal of Mac hinery, 10(S): 373 -38S .
Propuls ion and Power, Vol. 18, No.3, pp . S09-S 18. [40] Lampart, P., and Yershov, S., 2001, "3D Shape Optimization of
[3 2] Ama no, R.S., a nd Xu, c., 2003, "Aerodynamic Blade Des ign of Turbomachinery Bladin g," CF D for Turbomac hinery Applications, Gdansk,
Turbo mac hinery," Proceedings of the Internationa l Gas Turbine Congress Po land CFD Turbo 2001-C30.
2003, Tokyo, Japan, TS-029. [41] Burguburu, S., Toussaint, c., Bonhomme, C., and Leroy, G. , 2004,
[33] Keskin, A., Dutta, A. K., and Bestle, D., 2006, "Modern Compressor "Numerical Optimization of TurbomachineryBladings," Journa l of
Aerodynamic Blading Process us ing Multi-Objective Optimization," Turbomachinery, January 2004, Vol 126, pp 91-100 .
ASME Turbo Expo 2006, Barce lona, Spain, GT2006-90206. [42] Oyama, A., Liou, M.S. , Obayashi , S., 2004, "Transonic Axia l-F low Blade
[34] Vob, C., Aulich, M., Kaplan, B., and Nicke, E. , 2006, "A utomated Multi- Optimization: Evo lutionary Algorithms/Three-Dimensional Navier-Stoke
Objective Optimization in Axial Compressor Blade Design," ASME Turbo Solver," Journal of Propuls ion and Power, Vol. 20, No .4, pp. 6 12-6 19.
Expo 2006, Barce lona, Spain, GT2006-90420 . [43] Lian, Y., and Kim, N. H., 2006, "Reliability Based Design Optimization of
a Transonic Compressor", AIAA Journal , Vol. 44, No .2, 2006, pp 368-37S.

122 123
[44] Song, L., Feng, Z., and Li, 1., 2005, "Shape Opti miza tion o f Turbine Stage 1 and a 3 D Nav ier-Stokes Solver," ASME Turbo Expo 2006, Spa in, GT2006-
Us ing Adaptive Range Differential Evolution and Three-Dime nsional 90659.
Navie r-Stokes Solver," ASME T urbo Expo 2005, Nevada, USA, GT2005 - [5 3] Xing, X . Q. a nd Da moda ran, M. , 2003 , "Optimal Tra nso nic Ae rody na mic
68280. Shape Design Using Simulta neo us Pe rturbation Stochastic Approximation
[45] Yi, W., Hua ng, H., and Han, W., 2006, "Design Optimi zatio n of Tra nsonic Method Co upled with Globa l and Local Optimization Methods, " 21 sl
Compressor Rotor Us ing CFD a nd Genetic A lgorithm," ASME Turbo Expo App lied Aerodynamics Co nfere nce, AIAA 2003-3 786 .
2006, Barce lona, Spain, GT2006-90155. [54] Li , H. D. a nd He, L. , 2005, "Toward Intra-Row G ap Optimization for one
[46] C hung, K. N ., Kim, Y. I. , Sung, 1. H., Sung, I.H., C hung, I.H. , a nd Shin, S. and Half Stage Transo nic Compressor," Journal of Turbomachinery, Vol.
H. , 2005 , "A Study of Blade Section Shape for a Steam Turbine," ASME 127, pp . 589-598.
Fluid Engineering Divi s ion Summer Mee ting and Exhibition, 2005 , [55] Bara nkiewicz, W. S. and Hathaway, M. D., 1997, "Effect of Stator
Houston, TX, USA, FEDSM2005-77385. Indexing on Performance in a Low Speed Multistage Axial Compressor,"
[47] Oksuz, 0 ., and Akmandor, I. S., 2005, "Turbine Blade Shape Aerodynamic NASA TM 113113.
Desi g n Using Artificial Intelligence," ASME Turbo Expo 2005 , Nevada, [56] Chen, N.X., Zhang, H. W ., Du, H., Xu, Y. 1. and Huang, W. G. , 2005 ,
USA, GT2005-68094. "Effect of Maximum Camber Location on Aerodynamics Performance of
[48] Lampart, P., 2004, "N umerica l Shape Optimization of a High Pressure Tra nso nic Compressor Blades," ASME Turbo Expo 2005 , Nevada, USA,
Stea m Turbine Stage," Journal of Computational and Applied Mechanics, GT2005-68541 .
Vol. 5, No.2, pp. 311-32l. [57] Dennis , B. H. , Egorov, I. N. , Ha n, Z . X. , Dulikravich, G. S., and Polini, C.,
[49] Oyama, A., Liou, M. S., and Obayas hi, S., 2002, "Transonic Axial-Flow 2000, " Multi-Objective Optimization of Turbomachinery Cascades for
Blade Shape Optimization Usin g Evolutionary Algorithm and Three- Minimum Loss, Maximum Loading, and Maximum Gap-to-Chord Ratio,"
Dimensional Navier-Stokes Solver," AIAA 2002-5642 . 8 1h AIAAlNAS A/U SAF/ISSMO Symposium on Multidisciplinary Analysis
[50] Pie rret, S. , Coelho, R. F. , and Kato, H., 2007 , " Multidisciplinary and and Optimization, California, USA, AIAA2000-4876.
Multiple Operating Points Shape Optimization of Three-Dime ns ional [58] Kelner, V., Grondin, G. , Leona rd, o. and Moreau, S., 2005 , "Multi
Co mpressor Blades," Structural a nd Multidisc iplinary Optimi zation, Vol. Objective Optimiza tion of a Fan Blade by Coupling a Genetic Algorithm
33, No. I, pp. 1615-1488. and a Pa rametric Flow Solver," Evolution and Deterministic Methods for
[51] Papila, N. U., 200 1, "Neural Netwo rk a nd Po ly nomi a l Based Response Design, Optimi zatio n a nd Co ntrol with Application to Industri a l and
Surface Tec hnique for Supersonic Turbine Design Optimizat ion ", Ph D Societal Proble ms, EU ROGEN 2005 , Munich, Germany.
Thes is, Unive rsi ty o f Florida. [59] C laric h, A., Mosetti, G. , Pediroda, V . a nd Polini, C., 2005 , "Application of
[52] Lofii , 0. , Te ixera, 1. A. , Ivey, P. c., Kin g horn, I.R . a nd Sheard, A. G. , Evo lutio nery Algorithms a nd Statistical Analysis in the Numerical
2006, "Shape Optimi zation of Axial Fan Blades Using Genetic Algorithms
124 125
 Optimization of an Axial Compressor," Inte rnational Journal of Rotating [68] Shyy, W. , Papila, N. , Vaidya natha n, R. and Tucker, K. , 2001 , '·G lobal
Mac hinery, Vol. 2, pp. 143-151. Design Optimization for Aerodynamics and Rocket Propulsion
[60]Jang, C.M., and Kim, K.Y., 2007, " Applications of Numerical Components," Progress in Aerospace Science, Vol. 37, pp . 59-118.
Optimization Techniques to Design of Axi al Compresso r Blades," Journal [69] Zerpa, L., Queipo, N.V. , Pintos, S., and Salager, J., 2005 , " An Optimization
of Aerospace Power, Vol. 22, No.4, pp . 29-3 6. Methodology of Alkaline-Surfactant-Polymer Flooding Processes Usi ng
[6 1] Yang, L., Ouyang, H. , and Hui, D. Z., "Optimization Des ign and Field Scale Numerical Simu lation and Multiple Surrogates," Journal of
Experimental Study of Low-Pressure Axial Fan with Forward-Skewed Petroleum Science and Engineering, Vol. 47, pp . 197-208 .
Blades," International Journal of Rotary Machinery, Vol. 2007 , Article 10 [70] Goel, T. , Zhao, J., Thakur, S. , Haftka, R.T. and Shyy, w., 2006, " Surrogate
85275 , 10 pages, doi : 10 .1155/2007 /85275 . Model-Based Strategy for Cryogenic Cavitation Model Validation and
[62] Chen, N., Zhang, H., Xu, Y. , and Huang, W., 2007 , " Blade Sensitivity Evaluation," 42nd AlAN ASMEI SAEI ASEE Joint Propulsion
Parameterization and Aerodynamic Des ign Optimization for a 3D Conference and Exhibit, Sacramento, USA, AIAA-2006-5047 .
Transonic Compressor Rotor," Journal of Thermal Science, Vol. 16, No. 2, [71] Samad, A. , Shin D. Y., Kim, K. Y., Goel , T. and Haftka, R. T. , 2007,
pp . 105-114. "S urrogate Modeling for Optimization of a Dimpled Channel to Enhance
[63] Benini, E., 2004, "Three-Dimensional Multi-Objective Desi gn Heat Transfer Performance. AIAA Journal of Thermophysics and Heat
Optimization of a Transonic Compressor Rotor," Journal of Propulsion and Transfer," Vol. 21 , Number 3, pp. 667-670.
Power, Vol. 20, No.3, pp. 559-565. [72] Messac, A., 1996, "Physical Programming: Effective Optimization for
[64] Keskin, A. and Bestle, D., 2006, "Application of Multi-Objective Computational Design", AlAAjournal, Vol. 34, No. I, pp. 149-158.
Optimization to Axial Compressor Preliminary Design," Aerospace Science [73] Chankong, V., and Haimes, Y. Y., 1983 , "M ultiobj ective Decision Making
and Technology 10, pp. 581 - 589. Theory and Methodology," New York : Elsevier Science.
[65] Sanger, N. L. , 1983, " The Use of Optimization Techniques to Des ign- [74] Sen, P. and Yang, lB ., 1998, Multiple Criteria Decis ion Support III

Controlled Diffus ion Co mpressor Blading," Journal of Engineering fo r Engineering Des ign, London : Springer Verlag.
power, Vol. 105, pp. 256-264. [75] Collette, Y., and Siarry, P., 2003, Multiobjecti ve Optimization: Principles
[66] Queipo, N. V.,Haftka, R.T., Shyy, W., Goel, T. , Vaidyanathan, R. and and Case Studies, New York, Springer.
Tucker, P. K., 2005, " Surrogate-Based Ana lys is and Optimi zation," [76] Deb K. , 200 I, Multi-objective optimization us ing evolutionary algorithms,
Progress in Aerospace Sciences, Vol. 41, pp . 1-28. 1st ed., John Wiley & Sons Inc.
[67] Li , W. and Padula, S., 2004, "Approximation methods for conceptual des ign [77] Kicinger, R., Arciszewski, T., and Jong, K. D. , 2005 , "Evo lutionary
of complex systems," Eleventh International Conference on Approximation co mputation and structural design: A survey of the state-of-the-art,"
Theory (eds. Chui , c., Neaumtu, M., Schumaker, L.) . Computers and Structures 83, pp . 1943- 1978 .

126 127
[78] Marler, R.T., and Arora, 1. S., 2004, " Survey of multi-objective
AUTHORS PU BliCA nONS
optimization methods for engineering," Structural and Multidisciplinary
Optimization, Vol. 26, No.6, pp. 369- 395 .
International journal publications
[79] Baldwin, B. S., and Lomax, H. , 1978, "Thin Layer Approximation and
I. AbdusSamadandKwang-Yong Kim, "S hape Optimization of an Ax ial
Algebraic Model for Separated Turbulent Flow," AIAA Paper No. 78~257 .
Compressor Blade by Multiobjective Genetic Algorithm," Proceedings of
[80]Jameso n, A., Schmidt, w., and Turkel, E., 1981 , "Numerical Solutions of
The Institution of Mechanical Engineers, Part A-Journal of Power and
the Euler Equation by Finite Volume Methods Using Run ge-Kutta Time
Energy (Accepted).
Stepping Schemes," AlAA Paper No. 8 1 ~ 1 259.

[81] Ansys CFX-II.O, 2006, Ansys Inc. 2. AbdusSamad, Kwang-Yong Kim, TusharGoel, Raphael T. Haftka and Wei

[82] Beheshti, B. H., Teixeira, 1. A. , Ivey, P.C., Ghorbanian, K., and Farhanieh, Shyy, "Multiple Surrogate Modeling for Axia l Compressor Blade Shape

B., 2004, " Parametric Study of Tip Clearance Cas ing Treatment on Optimization," AIAA Journal of Propulsion and Power, Vol. 24, No .2,

Performance and Stability of a Transonic Axia l Compressor," Journa l of 2008, pp. 302-310.

turbomachinery, Vol. 126, Issue 4, pp . 527-535. 3. AbdusSamad, Dong-Yoon Shin, Kwang-Yong Kim, TusharGoel , and
[83] JMP® 5.1, 2004, SAS Institute, Inc. Raphael T. Haftka, "S urrogate Modeling for Optimization of a Dimpled
[84] McKay, M.D. , Beckman, R.J ., Conover, W.J. , 1979. "A comparison of Channe l to Enhance Heat Transfer Performance," AIAA Journal of
three methods for selecting values of input variables in the analysis of Thermophysics a nd Heat Transfer, Volume 2 1, Number 3, 2007, pp.667-
output from a computer code," Technometrics. 21, pp. 239-245. 670.
[85] Sacks, 1., Welch, W.J., Mitchell, T.J . and Wynn, H.P., 1989, " Design and 4. AbdusSamad and Kwang-Yong Kim, "Multi-objective Optimization of an
analysis of computer experiments," Statistical Science. 4, pp. 409-435.
Axial Compressor Blade," Journal of Mechanical Science and Techno logy,
[86] Tailor, J.S., and Cristianini , N. , 2000, Support Vector Machines and other
22(2008), pp. 999-1007.
kernel-based learning methods, Cambridge University Press.
5. AbdusSamad, Ki-Don Lee and Kwang-Yong Kim, "M ulti-Objective
[87] MATLAB®, 2004, The Language of Technical Computing, Release 14,The
Optimi zation of a Dimpled Channel for Heat Transfer Augmentation," Heat
MathWorks Inc.
and Mass Transfer (On review).

6. AbdusSamad, Ki-Sang Lee, Kwang- Yong Kim, "Multi-Objecti ve Shape

Optimization of an Axial Fan blade," International Journal of Air-
Conditioning a nd Refri geration, Vol. 16, No. I, March 2008, pp. 147-154.

7. Ki-Sang Lee, Kwang-Yong Kim, and AbdusSamad, "Des ign Optimization

of Low-Speed Axial Flow Fan Bl ade with Three-Dimensional RANS
128
129
 Analysis," Journal of Mechanical Science and Technology (Accepted) . Proceedings of ASME urbo Expo 2006: Power fo r Land, Sea and Air, May

8. AbdusSamad, Kwang-Yong Kim, and Dong-Yoon Shin, "Optimization of 2006, Barcelona, Spa in, GT2006-90384.

Cooling Channel Roughened by Dimples ," Journal of Fluid Science and 3. AbdusSamad, Kwang-Yong Kim, TusharGoel, R.T. Haft ka and Wei Shyy,
Technology, Vol. 3, No.7, September 2008. "S hape Optimi zatio n of Turbomachine ry Blade Using Multiple Surrogate
Models", ASME Joint-U.S. -European Fluids Engi nee ring Summer Meeting,

Domestic (Korean) journal publications Jul y 2006, Miami, FL, USA, FEDSM2006-98368 .

I. Choon-Man Jang, AbdusSamad and Kwang-Yong Kim, " Optimal Design 4. AbdusSamad and Kwa ng-Yong Kim, "Effects of Surface Ro ughness on

for Stacking Line of Rotor Blade in a Si ngle-Stage Transonic Ax ial Performance of Axial Compressor Blade with Sweep and Lean," The 9th

Compressor," Journal of Fluid Machinery, 3rd November 2006, pp 7 -13 . Asian Internatio nal Conference on Fluid Machinery, AICFM-9, Jej u, Korea,
October 16-1 9, 2007, AICFM9-021.
2. AbdusSamad and Kwang-Yong Kim, "Effect of surface Roughness o n
Performance of Axial Compressor Blade," Journal of Fluid Machinery, June 5. AbdusSamad, Dong-Yoon Shin, Kwang-Yong Kim and Raphae l T. Haftka,

2007, pp. 9-16. " Application of Multiple-Surrogate Model to Optimization of a Dimpled

Channel," 7th World Congress on Structural and Multidisc iplinary
3. Choon-M an Jang, AbdusSamadandKwang-Yong Kim, "Optimization of
Optimization, WCS M07 , May 21- 25 , 2007, Seoul, Korea, A0069.
Rotor Blade Stacking Line Using Three Different Surrogate Models ,"
(Nominated fo r ISSMO/Springer Prize and included in best 5 finalists:
Journa l of Fluid Machinery, April 2007, pp .22-31.
http://www.issmo.net/about/prize.asp ).
4. AbdusSamad, Kwang-Yong Kim and Jae-Ho Choi, " Blade Shape
6. AbdusSamad and Kwang- Yong Kim, "Multi-objective Optimization of
Optimization of a Transonic Co mpressor by SutTogate Model," Journal of
Turbomachinery Blade," The 9th Asian International Conference on Fluid
KSME Transactions, 32(4), April 2008 , pp. 317-326.
Machinery, AICFM-9, Jej u, Korea, October 16-19, 2007, AICFM9-294 .

7. AbdusSamad, Kwang-Yo ng Kim and Ki-Sang Lee, "M ulti-objective

International conference publications
Optimization of a Turbomachinery Blade Using NSGA-II ," 5th Joint
I. Choon-Man Jang, AbdusSamad&Kwang- Yong Kim, "Enhancement of
ASME/JSME Fluids Engineering Co nfe rence, Jul y 30 - August 2, 2007,
Effic ie ncy by Blade Optimization in a Single-Stage Transo nic Axial
CA, USA, FEDSM2007-37434.
Compressor", Proceedings of the Asian Congress on Gas Turbines,
8. AbdusSamad, Dong-Yoo n Shin and Kwang-Yong Kim, "Optimization of
November 2005 , Seoul, Korea, ACGT 2005-005.
Cooling channe l Roughened by Dimples," The 9th Asia n International
2. Choo n-Man Jang, AbdusSamad&Kwang-Yong Kim, Optimal Design Of
Conference on Fluid Machinery, AICFM-9, Jeju Island, Korea, October 16-
Swept, Leaned a nd Skewed Blades in a Transonic Axial Compressor,
19, 2007, AICFM9-025 .
130 131
9. Ki-Sang Le , K ang-Yong Ki m and AbdusSamad, "Design Optimization hape Optimization of an Axial Compressor Rotor," AIAAlISSMO
of Low Speed Axial Flow Fan Blade with Three Dimensiona l RANS Multidisc iplinary Analysis and Optimization, British Columbia, Canada,
Ana lys is" The 9 th Asian Internationa l Conference on Fluid Mac hinery, September 10-12, 2008 .
AICFM9-026, Jeju, Korea, October 16-19, 2007.

10. AbdusSamadand Kwang-Yong Kim, " Stacking and Thickness

Domestic (Korean) conference publications
Optimization o f a Compressor Blade Using Weighted Average Surroga te
I. Choon-Man Jang, A bdusSamadand Kwang-Yong Kim, "Optimizat ion of
Model," Proceedings of ASME Turbo Expo 2008 , June 9-13 , 2008 , Berlin,
Stacking Line in Rotor and Stator Blades of Single-Stage Transonic
Germany, GT2008-50262 .
Axia l Compressor", Korean Soc iety of Mec han ical Engineers, 2005,
II. AbdusSamad, Ki-Don Lee and Kwang-Yong Kim, "Optimization of a
Yonpyong, Korea, pp.S25-S32.
Channe l Roughened by Dimples on Opposite Surfaces for Heat Transfer
2. Choon-Man Jang, AbdusSamad&Kwang-Yong Kim, Shape Optimization
Enhancement," Proceedings of ASM E Turbo Expo 2008, June 9- 13, 2008,
of Swept Leaned and Skewed Blades in a Transonic Axia l Compressor
Berlin, Germany, GT2008-S0264.
for Enhancing Rotor Effic iency, Korean Fluid Machinery Associatio n
12. AbdusSamad, Ki-Don Lee, Kwang-Yong Kim and Jin-Hyuk Kim,
Conference, December 2005, Changwon, Korea, pp .52S-532 .
"Optimum Design of A Chan ne l Roughened By Dimples to Improve
3. Choon-Man Jang, AbdusSamad a nd Kwang-Yong Kim, "Evaluation of
Cooling Performance" The 2nd In ternational Symposi um on Jet Propul sio n
Efficiency by App lying Different Optimizati on Method for Axia l
and Power Engineering, September 22-26, 2008, Guilin, C hina . 2008-
Compressor" , Proceedings of The Fourth National Congress on Fluids
ISJPPE-20 II.
Engineering, August 23-25,2006, Kyungju, Korea, pp. 1097-1100.
13 . Ki-S ang Lee, Kwang-Yong Kim, AbdusSamad and Sang-Ho Jung,
4 . AbdusSamadand Kwang-Yong Kim, " Performance of Surrogate Models in
"Optimization of a n Axia l Flow Fan Blade for Efficiency Enhancement
Optimiz ing Heat Transfer Augmentation Devices", 4th National Congress of
Usi ng Neural Network Techniques" The 2nd In ternational Symposium on
Fluids Engineering, August 23-25, 2006, Kyungju, Rep ublic of Korea, pp.
Jet Propulsion and Power Engineering, September 22 -26, 2008 , Guilin,
111 3-1116.
China. 2008-[ SJPPE-40 II.
5. AbdusSamadand Kwang- Yong Kim, "Evaluation of Surrogate Models for
14. AbdusSamad, Ki-Sang Lee andKwa ng-Yong Kim, " Optimization of
Shape Optimization of Compressor Blades", 4th National Congress of
Stacking Line and Blade Profile for Design of Axial Fan Blade," 2008
Fluids Engineering, August 23-25 , 2006, Kyungju, Repub lic of Korea, pp.
ASME Fluids Engineering Conference, FL, USA, August 10-14, 2008,
367-370.
FEDSM2008-553 17.
6. AbdusSa madand Kwang-Yong Kim, "Heat Transfer Enhanceme nt by
[S. Abd usSamadandKwa ng-Yong Kim, "Multidi scip linary and Multiobjective
Shape Optimization of Dimpled C hanne l" , KSME fall conference,
132 133
 ove mber 1-3, 2006 , Muju, Korea . pp. 21 68- 2 173 15. bdus Samad, Ki -Sang Lee, Sang- Ho l ung and Kwang-Yong Kim,

7. AbdusSamadand Kwa ng-Yong Kim, "S hape Optimization of Compressor "Optimization of Stack ing Line and Blade Profi le fo r Des ign of Ax ia l Flow

Blade Us ing Surrogate Mode ls", KSME fa ll co nfere nce, November 2006, Fa n Blade," The 5 th National Co nfe rence on Fluids Engineering, March 26 -

Muju, Korea. pp . 2 120-2 125. 28, 2008, l ej u, Korea.

8. AbdusSamad and Kwang-Yo ng Kim, "Effects of Surface Roughness on

Tu rbomac hinery Blade Performance ," KFMA Annua l Confe re nce,
Nove mber 2006, Kwa ngj u, Korea.

9. AbdusSamad, Kwa ng-Yong Kim and Ki -Sang Lee, " Multi-Obj ecti ve
Optimization of a Fan Blade Us ing NSG A- II," KSME confere nce May
2007, Korea.

10. AbdusSamad, Do ng-Yoo n Shin, and Kwang-Yong Kim, " Surrogate

Mode ling for Optimizati on of a Dimpl ed C hanne l to Enhance Heat Transfer
Performance," KSM E confe rence May 2007, Korea.

II. Ki -Sang Lee, Kwa ng-Yo ng Kim and AbdusSamad , " Design Optimization
of an A xia l Flow Fan Blade Considerin g Airfo il Shape and Stacking Line,"
KSME Co nfere nce, October 2007, Korea.

12. Ki-Don Lee, AbdusSamad, and Kwa ng-Yo ng Kim, "Numeri cal
Optimization of Stagge red Dimpled C hannel Us ing Weighted- Average
Surrogate Mode l," KSM E Co nference, October 2007, Korea.

13. Ki-Sa ng Lee, Kwang-Yong Kim and Abd usSamad, "Optimum Des ign o f
an Ax ia l Flow Fan Us ing Ne ural Network," KFM A Co nfe re nce, Daejeon,
Nove mber 29-30, 2007. Ko rea

14. Ki-Don Lee, AbdusSamad, and Kwa ng-Yong Ki m, "M ult i-Objective
Optimization of Dimpled C hannel usi ng NSGA II ," T he 5 th Natio na l
Confere nce on Fluids Engineering, March 26-28, 2008 , l eju, Korea.

134 135
 BIOGRAPHICAL SKETC H

AbdusSamad was born o n Marc h 5, 1977 in India and completed B.Tech.

andM.Tech. in Mec hanical Enginee rin g fro m A ligarh Muslim Uni ve rsity, India.
He jo ined Inha Uni versi ty fo r a doctoral course in 2004. He worked with
specializatio n in Thermo dynamics and Fluid Mechanics in Mechanica l
Eng ineering Departme nt. During his PhD period he has pub lished a good
number of papers in nationa l and international journa ls and conferences.

136