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

net/publication/225988383

Aerodynamic optimization of turbomachinery blades using evolutionary

methods and ANN-based surrogate models

Article  in  Optimization and Engineering · September 2008

DOI: 10.1007/s11081-007-9031-1

CITATIONS READS

42 296

2 authors:

Temesgen Mengistu Wahid Ghaly

10 PUBLICATIONS   100 CITATIONS   
Concordia University Montreal
54 PUBLICATIONS   486 CITATIONS   
SEE PROFILE
SEE PROFILE

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

turbomachinery design and optimization View project

Aerodynamic Shape Design of Turbomachinery Blading in 3D Viscous Flow View project

All content following this page was uploaded by Wahid Ghaly on 13 June 2014.

The user has requested enhancement of the downloaded file.

Optim Eng (2008) 9: 239–255
DOI 10.1007/s11081-007-9031-1

Aerodynamic optimization of turbomachinery blades

using evolutionary methods and ANN-based surrogate
models

Temesgen Mengistu · Wahid Ghaly

Received: 26 April 2006 / Accepted: 23 October 2007 / Published online: 1 December 2007
© Springer Science+Business Media, LLC 2007

Abstract A fast, flexible, and robust simulation-based optimization scheme using an

ANN-surrogate model was developed, implemented, and validated. The optimization
method uses Genetic Algorithm (GA), which is coupled with an Artificial Neural
Network (ANN) that uses a back propagation algorithm. The developed optimization
scheme was successfully applied to single-point aerodynamic optimization of a tran-
sonic turbine stator and multi-point optimization of a NACA65 subsonic compressor
rotor in two-dimensional flow, both were represented by 2D linear cascades. High
fidelity CFD flow simulations, which solve the Reynolds-Averaged Navier-Stokes
equations, were used in generating the data base used in building the ANN low fi-
delity model. The optimization objective is a weighted sum of the performance ob-
jectives and is penalized with the constraints; it was constructed so as to achieve a
better aerodynamic performance at the design point or over the full operating range
by reshaping the blade profile. The latter is represented using NURBS functions,
whose coefficients are used as the design variables. Parallelizing the CFD flow sim-
ulations reduced the turn-around computation time at close to 100% efficiency. The
ANN model was able to approximate the objective function rather accurately and to
reduce the optimization computing time by ten folds. The chosen objective function
and optimization methodology result in a significant and consistent improvement in
blade performance.

Keywords Artificial neural networks · Genetic algorithm · Computational fluid

dynamics · Aerodynamic design · Global optimization · NURBS · Response surface
approximation

T. Mengistu · W. Ghaly ()

Concordia University, Montreal, Quebec, Canada
e-mail: ghaly@encs.concordia.ca
Present address:
T. Mengistu
CENAERO, Gosselies, Belgium
240 T. Mengistu, W. Ghaly

1 Introduction

Aerodynamic optimization methods are becoming very attractive in today’s compet-

itive environment as they can reduce the design cycle time by automating the design
process. Until recently, designers were relying mostly on manual optimization. As
Computational Fluid Dynamics (CFD) matured over the last decade and as comput-
ing technology has greatly improved and has become more affordable, simulation-
based optimization is becoming affordable and more popular than ever. This is due to
the fact that optimization techniques give direct control on performance parameters,
even if the computational cost is at least one order of magnitude larger than the cost
of an analysis calculation.
Aerodynamic shape optimization allows the designer to automate the exploration
of the design space to achieve a given objective. One possible design objective is
to minimize flow losses, which can be measured by e.g. the total pressure loss (or
entropy generation), through proper reshaping of the blade profile. Automated aero-
dynamic design is accomplished by coupling a CFD flow simulation code with nu-
merical optimization methods. As the aerodynamic shape optimization problem is
a complex one with possibly many local minima, gradient-based methods can be
trapped in a local optimum, unless the initial guess is close to the global mini-
mum. For this reason, heuristic/evolutionary global algorithms such as Genetic Al-
gorithm (GA) and Simulated Annealing (SA), although more computation inten-
sive compared with gradient-based methods, are used to ensure reaching close to
the global minimum. These algorithms have been recently applied in turbomachin-
ery design problems; examples of such algorithms are given in (Dennis et al. 1999;
Wang and Damodaran 2000; Oyama et al. 2002).
In the work of Dennis et al. (1999), a combination of genetic algorithm and Se-
quential Quadratic Programming (SQP) algorithms were used to optimize a two-
dimensional turbine cascade. The GA followed by SQP implementation was intro-
duced to reduce the total number of required function calls, keeping the global ex-
ploration behavior of GA. The optimization scheme required from 220 to 675 calls
to the flow analysis code.
Oyama et al. (2002) worked on 3D blade shape optimization and included the
mass flow rate and pressure ratio as constraints in the objective function.
The above references represent a sample of aerodynamic cascade optimization,
and can be used as a useful tool to design blade cascades however they require large
computation resources. The effort now is focused on finding a way of reducing the
prohibitive computation time without compromising the solution accuracy.
Since aerodynamic design optimization problems are multi-modal and discon-
tinuous in nature, gradient-based numerical optimization algorithms risk of getting
trapped in local minima (Lai and Yuan 2002) or run into an infeasible design for
which the flow simulation does not converge. Therefore exploratory algorithms such
as Genetic Algorithm (GA) and Simulated Annealing (SA) are more appealing for
global exploration of the design space however, GA and SA can involve a pro-
hibitively high computational cost where a large number of costly CFD simulations
are needed, which makes exploratory algorithms less appealing than gradient-based
optimization algorithms (Lai and Yuan 2002).
Aerodynamic optimization of turbomachinery blades 241

In order to avoid this prohibitive cost of GA/SA, a low fidelity approximation of

the objective function, using a Response Surface Approximation (RSA), can be used
so as to reduce to a minimum the number of required CFD analyses. Moreover us-
ing a RSA can eliminate some of the noise in the objective function as it creates a
smooth response surface thereby improving the convergence of the optimization al-
gorithms. Examples of such an approximation are Artificial Neural Networks (ANN)
and quadratic polynomial response surface models, which have been successfully
used in aerodynamic shape optimization as well as in other fields (Pierret et al. 2000).
In the present work, an optimization method of the evolutionary type, namely GA,
and an approximation method, ANN, are first presented. They are then implemented
in the aerodynamic shape optimization of 2D gas turbine blades. The optimization
objective is to improve the blade performance, subject to some constraints, by mod-
ifying the blade shape which is approximated using NURBS, whose coefficients are
used as the design variables. CFD is used to generate the high fidelity data set that
is used in training and testing the ANN and in validating the optimum profile. The
optimization methodology as well as the approximation (ANN) are demonstrated by
redesigning a transonic impulse turbine stator and a subsonic axial compressor ro-
tor.

2 The design methodology

2.1 Problem definition and objective function

In any optimization problem, the choice of the objective function affects the opti-
mization process as well as the results. Thus a careful and well-studied identification
and formulation of that function is crucial. For example, the overall aerodynamic
performance of a compressor rotor is determined by its adiabatic efficiency, η, and/or
total pressure loss coefficient, ω, at design and off-design conditions, therefore one
can choose to optimize the efficiency or the total pressure loss or both; this can be
done either at the design point only or on the full operating range. With this design
strategy in mind, the objective function is constructed as a weighted sum of individ-
ual objectives and is penalized with the constraints such that it can serve for single
point or multi-point optimization, and can be defined as follows:
 
n 
n  n
Fobj (X) = Min C1 (1 − φi ) + C2 |φj − φi | + P T (1)
i=0 i=0 j =0

where φ stands for either adiabatic efficiency, η, or (1 − ω), where ω is the total pres-
sure loss coefficient. X is the vector of design variables, which include the backpres-
sure and the shape parameters that control the blade profile. Varying the backpressure
in the pre-determined range from the choke limit to the stall limit while fixing the
rotor speed allows for tracing a speed line, i.e. it allows for design and off-design
calculations that correspond to different mass flow rates with varying efficiency and
pressure ratio.
242 T. Mengistu, W. Ghaly

The first term in the objective function, (1), attempts to maximize the efficiency
(or minimize the total pressure loss coefficient) at the design and off-design points,
while the second term would eliminate large difference in efficiency (or total pressure
loss) between the design and off-design points which would tend to keep it constant
and optimum over the entire operating range. The last term in the objective function
is a Penalty Term (PT) that accounts for the aerodynamic, mechanical and geometric
constraints imposed on the optimization process. The aerodynamic constraints could
include the exit flow angle, the spacing to chord ratio, and the stall margin.
The summation is carried out over n pre-selected points, these pre-selected points
are the design point and off-design points. In this work, two cascade optimization
cases were considered in Sect. 5; one of them is a single-point optimization, i.e. n = 1,
while the second case is a multi-point optimization where n = 4. The weights Ck ,
where k = 1, 2, are prescribed by the designer, they are determined such that the
different components of the objective function have the desired influence on the opti-
mization process. Note that the current choice of the objective function, given in (1),
allows for different design options depending on the values given to the Ck coeffi-
cients: e.g., single or multi-point optimization for maximum efficiency or minimum
total pressure loss.

2.2 NURBS representation

The geometric representation of the profiles is an important part in the aerodynamic

shape optimization procedure. The parameters in the geometric representation of
the blades are used as design variables in the aerodynamic optimization process.
At present, the geometry is parameterized using Non-Uniform Rational B-splines
(NURBS). A clear advantage of using NURBS is that one can adjust the profile lo-
cally on specific regions of the blade, by modifying the weights and/or the x- and
y-coordinates of the NURBS control points, without necessarily affecting the whole
blade profile. The NURBS definition and formulation are very well described in
(Piegl and Tiller 1995).
An accurate geometric representation of 2D turbomachinery cascades was ob-
tained using NURBS with a minimum possible number of control points ranging
from 9 to 19 depending on the type of blade cascade, see (Ghaly and Mengistu 2003)
for more detail.
In the present work, the blade profile was defined by the mean camber line and
a thickness distribution; each of which was parameterized by a NURBS function
with a number of control points ranging from nine to eleven. The position of these
control points and the weights can be taken as the design variables in the optimization
process, see the results given in Sect. 5.

2.3 Flow simulation method

The two-dimensional turbulent transonic flow in a linear cascade is simulated using

a second-order accurate cell-vertex finite volume space discretization method on an
unstructured triangular mesh. The steady state solution is reached by pseudo-time
marching the Reynolds-averaged Navier-Stokes equations using an explicit five-stage
Aerodynamic optimization of turbomachinery blades 243

Runge-Kutta scheme. Local time stepping and implicit residual smoothing were used
to accelerate the convergence. A non-linear blend of second and fourth order artificial
viscosity was used in capturing shocks and eliminating pressure-velocity decoupling
with minimal numerical diffusion. The method of characteristics was used to impose
inflow and outflow boundary conditions. Turbulence is modeled using the Baldwin-
Lomax model in the discretization of the Reynolds-averaged Navier-Stokes equations
(Daneshkhah 2006).
The flow at the inlet and exit planes, which are placed at about one chord upstream
and downstream of the cascade, is subsonic. The boundary conditions at inlet are
given by the total pressure, total temperature and the tangential velocity. At the exit
plane, the ratio of exit static to inlet total pressure (referred to as the backpressure) is
specified. This set of boundary conditions allows for computing the flow at any point
on a given speed line. The latter is characterized by a fixed rotor angular speed, which
is represented by a fixed tangential velocity in the rotor relative frame and varying the
backpressure will result in changing the mass flow rate and the pressure ratio across
the rotor, hence will allow for moving along the speed line.

3 Optimization algorithm and surrogate model

The aerodynamic shape optimization problem is reduced to solving a constrained op-

timization problem, which is transformed into an unconstrained one by penalizing
the objective function with the constraints. The optimization problem is solved us-
ing Genetic Algorithm (GA) and is coupled with an Artificial Neural Network, as a
surrogate model, to reduce the optimization computing time. Both GA and ANN are
described in this section.

3.1 Genetic algorithm

Genetic algorithms are general-purpose search algorithms based upon the principles
of evolution observed in nature. Genetic algorithms combine selection, crossover,
mutation, and elitism operators with the goal of finding the best solution to a problem.
They search for this optimal solution until a specified termination criterion is met
(Goldberg 1985; Gen and Cheng 1997).
In the present work, the variables for the GA algorithm are real coded, where
an individual is characterized by a vector of real numbers. Two kinds of crossover
operations are included in the real-coded GA developed in this work namely, arith-
metic and heuristic crossover operators. Arithmetic crossover operator combines lin-
early two parent chromosome vectors to produce two new offsprings while heuristic
crossover operator uses the fitness values of the two parent chromosomes to deter-
mine the search direction and creates the new offsprings. In addition GA is naturally
a population-based parallel algorithm that is best suited in a parallel computation
algorithm. A population of 32 individuals with a crossover probability of 0.80, mu-
tation probability of 0.15 and elitism of 2 has been used for each generation. The
implementation of GA is detailed in (Mengistu and Ghaly 2003).
244 T. Mengistu, W. Ghaly

Table 1 GA validation

Test case Global minimum

Rastrigin Function
Present work: 40 design variables 10−14
(Deb and Joshi 2002): 20 design variables between 10 and 20
(Fogel and Beyer 1995): 30 design variables greater than 10

Welded-beam: 4 design variables and 5 constraints

Present work 2.197
(Ray and Liew 2003) 2.3809
(Reklaitis et al. 1983) 2.38116

Speed-reducer: 7 design variables and 11 constraints

Present work 2994.36
(Luo 2004) 2994.36
(Ray and Liew 2003) 2994.47

3.1.1 GA validation

The GA developed and used in this work was validated using the Rastrigin function
(Deb and Joshi 2002), the welded-beam problem and the speed reducer problem (Ray
and Liew 2003). These test functions vary in difficulty, in number of local minima,
and in number of design variables X and constraints. They have a global extremum
that is hidden among many local extrema. The search range that was chosen for each
function includes several local minima.
The Rastrigin function was tested with forty design variables, xi , that are defined
in the range ±100.12. Table 1 gives a summary of the results obtained for the three
test problems.
The welded beam problem is about the design of a welded beam for minimum cost
and maximum rigidity when it must carry a certain load. The problem is described and
solved by several authors to test optimization algorithms for engineering problems
(Reklaitis et al. 1983; Deb 2001; Ray and Liew 2003) who reported different results
at different times; which implies that the problem has several local optima and needs
to be solved for the global optimum. The GA implemented in this work found an
optimum cost of 2.197, satisfying all constraints. Table 1 summarizes the comparison
of the results with the literature mentioned above.
The speed reducer problem is another engineering problem that was investigated
by several authors for example, Rao (1996), Deb (2001), Ray (2003), Luo (2004). The
objective of this problem is to find the minimum weight, subject to 11 constraints.
There are seven design variables. The best-known feasible solution to this problem
is 2994.36. The GA implemented in the present work has found exactly the same
optimum solution with all constraints satisfied. Table 1 compares the literature results
with the result obtained by the GA developed in the present work.
Aerodynamic optimization of turbomachinery blades 245

Fig. 1 A typical neural networks architecture

3.2 Artificial neural networks

ANN is used as a low order RSA to approximate the objective function at a relatively
low computing cost and results in reducing the computing effort by a factor of ten.
A set of test cases is generated using the high fidelity CFD simulations and is used
in training and testing the ANN model, the latter is then used in the aerodynamic
optimization.
The present model is composed of a multi-layer feed-forward network with back-
propagation. It is composed of three layers, an input layer, one hidden layer having
41 nodes and an output layer, see Fig. 1. A sigmoid function is taken as the transfer
function between the nodes, and the weights in each connection in the network are
arbitrarily initialized to one. They are updated using an optimization algorithm to
minimize the error between the network output and the given training data set output.
The training strategy is enhanced using genetic algorithm and simulated annealing
algorithm in the initial stage of the training which is then followed by a gradient-
based method (Mengistu 2005). Results of the training and testing of this ANN model
for the aerodynamic optimization problem is given in the context of optimizing the
NACA compressor presented in the results section, see Fig. 10.

4 Numerical implementation

The aerodynamic design optimization involves four basic components: shape para-
meterization using NURBS, numerical optimization using GA, response surface ap-
proximation using ANN for the low fidelity calculation of the objective functions
246 T. Mengistu, W. Ghaly

and constraints and flow simulation using Reynolds Average Navier-Stokes solver
(RANS) for the high fidelity calculation of the objective functions and constraints.
The flow of calculations and information involved in building the ANN model and in
carrying out the aerodynamic design optimization is displayed in Fig. 2.
As the ANN training and testing requires a pool of high fidelity flow simulations
that are obtained by solving the RANS equations, parallel computations have been
implemented when producing these flow simulations. This parallelization, which
is 100% efficient, results in reducing the wall clock time required for optimization by
a factor that is equal to the number of processors used. Moreover, response surface
approximation used to evaluate the objectives and constraints, required negligible
time. The construction of the response surface needed a database of flow solutions,
which was obtained using high-fidelity flow simulation.
The overall computation time includes the selection of blade geometry candidates,
generation of CFD solution for the candidates, post-processing the solution, building
the database for response surface approximation, ANN training (Response surface
model construction) and optimization process using ANN. 75% of the total develop-
ment time is taken by the CFD solver while building the ANN response surface takes
about 17% of the time, but it should be emphasized that selection of blade geometry
candidates must be done with utmost care so that the database contains feasible blade
profiles, that cover reasonably well the design space, this requires about 7% of the
time. All the rest including the aerodynamic optimization takes 1% of the develop-
ment time.

Fig. 2 The ANN-based aerodynamic design optimization process

Aerodynamic optimization of turbomachinery blades 247

5 Results and discussion

In this section, two design cases are presented. The first case given in Sect. 5.1 is
intended to assess the aerodynamic optimization process while the second case is
intended to assess the ANN-based aerodynamic optimization.

5.1 Redesign of a transonic impulse turbine cascade

An impulse turbine cascade is redesigned to minimize the total pressure loss coeffi-
cient ω at a given operating point. For this cascade, the spacing to chord ratio is 0.526;
the camber and thickness distributions assume a parabolic profile with maximum
thickness to chord ratio of 21.45% and maximum camber to chord ratio of 21.45%,
both occurring at mid-chord. The inlet flow angle is 40.63° and the ratio of exit sta-
tic to inlet total pressure is 0.833. The flow is assumed inviscid and is simulated by
solving the Euler equations, which are a subset of the RANS equations described in
Sect. 2.3. Figure 3 shows a flow chart of the aerodynamic optimization process.
The objective function for this case is given by:

Fobj (X) = Min[ω + K1 (ṁ) + K2 (β)] (2)

where Fobj is the objective function, X is the vector of design variables, ω is the total
pressure loss coefficient, ṁ and β are the difference between the computed and
the target mass flow rates and exit flow angles (in degree). The weights K1 and K2
are user-specified penalty coefficients; they are chosen so as to have equal penalizing
effect on the objective function. In this case, they take the following values:

K1 = 1000 when |ṁ| > 0.01, and 0 otherwise,

(3)
K2 = 10 when |β| > 1◦ , and 0 otherwise.

The flow over this blade is transonic, where a shock is present on the blade suction
side. To reduce the total pressure loss, this shock should be weakened or eliminated,
which can be accomplished by reshaping the blade profile. The latter is described

Fig. 3 The aerodynamic optimization process

248 T. Mengistu, W. Ghaly

Fig. 4 Convergence history for

the impulse turbine cascade

by its camber line and its tangential thickness distribution, each is represented by
NURBS with 9 control points and weights; the control points and the corresponding
weights are determined as described in (Ghaly and Mengistu 2003). In this case,
the thickness distribution is fixed and the camber line is allowed to change. Hence
the design variables are taken to be the y-coordinates of the control points and the
corresponding weights along the camber line except those at the leading edge point;
this implies a total of 16 design variables. The y-coordinates of the control points are
allowed to vary from the original profile by ±15%, and the weights vary between 0.5
and 2.5; these ranges ensure a good coverage of the design space.
The optimization history is given in Fig. 4 where the best candidate in each GA
generation is given. The total pressure loss is reduced from 0.0043 to 0.0026 in five
generations; each generation consists of 32 individuals. A single flow field analysis
took approximately two minutes of CPU time (when starting from a converged solu-
tion obtained for a given candidate of the previous generation) with a CFL number of
4 on a mesh with 2800 points.
Figure 5 shows that the changes in shape between the original and the re-
designed blades are relatively small, which would be difficult to achieve by man-
ually changing the blade shape. However, the normal shock prevailing in the orig-
inal blade around mid chord has been eliminated in the redesigned blade, as can
be seen in Figs. 6 and 7 that show the Mach number contours and the Mach
number distribution along the original and redesigned blades. Note that the op-
timal design has a reversed curvature particularly along the suction side to al-
low the flow to compress reversibly. Other researchers (Ahmadi and Ghaly 1998;
Dang 1995) have also observed similar behavior.

5.2 NACA 65 subsonic compressor rotor

The optimization algorithm and the surrogate model presented earlier, were used to
redesign a well-documented subsonic compressor rotor (Emery et al. 1958) for best
sustained efficiency over the full operating speed line; this is accomplished by modi-
fying the rotor blade profile. The original profile is that of a NACA 65 airfoil and the
Aerodynamic optimization of turbomachinery blades 249

Fig. 5 The impulse turbine

cascade

Fig. 6 Isentropic Mach

contours

flow is assumed to be turbulent, the Reynolds number being 2.45 × 105 . The perfor-
mance measure φ that is included in the objective function, given in (1), is taken to
be the adiabatic efficiency, η, and the weights C1 and C2 are set to 1 and 0.5, respec-
tively. The Penalty Terms (PT) appearing in (1) include the optimization constraints,
namely the reduced mass flow rate, the inlet and exit flow angles, while the remain-
ing constraints are satisfied exactly through the imposition of the inlet and outlet flow
boundary conditions for the CFD flow simulation.
The original blade profile is defined in terms of a mean camber line and a thick-
ness distribution, both are parameterized using NURBS with 11 and 9 control points,
respectively. The design variables controlling the blade shape are taken to be the y-
coordinates of the camber line profile and the thickness distribution, excluding the
thickness at the LE and TE points and the camber line point at the LE, which gives
250 T. Mengistu, W. Ghaly

Fig. 7 Isentropic Mach along

the blade

a total of 17 geometric design parameters. The weights and the x-coordinates of the
control points are fixed during the optimization. The family of airfoils used in training
and testing the ANN was obtained by perturbing the above 17 geometric design pa-
rameters in a preset range and, to ensure a homogeneous representation of the design
space, the candidate airfoils were selected based on the Latin-Hypercube space-filling
experiments (McKay et al. 1979). To account for the full range of the operating speed
line, the flow was simulated at 4 values of backpressure that include the two limiting
points namely the stall and choke points, and two other points.
The ANN response surface model was constructed using a family of 50 blade pro-
files, 35 of these profiles were used in training the ANN model and 15 were used in
testing it. For each of these profile, the flow was simulated at 4 values of backpres-
sure, thus the total number of CFD runs amounted to 200 simulations that covered a
reasonably large range of the design space, as can be seen in Figs. 8 and 9. The CFD
flow simulations were carried out in parallel (one CFD simulation per node) on four
nodes of an SGI 2000 server using Message Passing Interface (MPI), which required
about 60 hours of wall-clock time.
The ANN model was constructed with one hidden layer containing 41 nodes, the
input layer with 18 design parameters and the output layer with 4 output variables
(efficiency, reduced mass flow rate, inlet flow angle and exit flow angle) used in com-
puting the individual terms appearing in the objective function, see (1). The ANN
training took about five hours on a Pentium IV PC; the errors in the ANN training
and testing are given in Fig. 10. The errors in the ANN approximation are given in
more detail in Fig. 11 which shows that these errors are less than 2% for about 45%
of the test set.
Unfortunately, there is no rule for determining the number of nodes in the hidden
layer. However, a good initial guess is the average of the number of input and output
nodes. This number is then increased to create an optimum-trained network. If the
number of nodes in the hidden layer is too small underfitting occurs, where high
training error and high generalization error occur. If the number of hidden layer nodes
is too large overfitting occurs, where low training error but high generalization error
occurs.
Aerodynamic optimization of turbomachinery blades 251

Fig. 8 The range of geometry

explored for the design space

Fig. 9 Speed lines of blade

profiles used in training/testing
the ANN

Figure 9 also shows that the optimal blade is 4% more efficient than the original
one. Moreover, the optimum blade shows a better performance in terms of pressure
ratio (although the latter is not part of the optimization function) over the full range of
operation; this behavior was also observed by Oyama et al. (2002). Figure 12 shows
that the optimized blade camber has significantly changed near the trailing edge, and
the maximum thickness has increased by 3.7% and has slightly moved downstream.
The high camber near the trailing edge resulted in a higher flow turning at the cost of
an increase in profile loss. Figures 13 and 14, where the efficiency is plotted vs. mass
flow rate and vs. pressure ratio, show a 7% improvement in efficiency and about 1%
increase in total pressure ratio for the same blade speed and mass flow rate.
252 T. Mengistu, W. Ghaly

Fig. 10 The convergence

history of ANN training/testing

Fig. 11 The ANN model

validation

The performance of the optimal rotor blade, which is given in Figs. 13, 14 and 15,
was evaluated by simulating the flow over the redesigned compressor blade at values
of backpressure that are different from those used in generating the surrogate model.
The fact that the performance of the optimal blade is smooth over the full operating
range reflects the robustness and the validity of the optimization strategy, and the
accuracy of the surrogate model.

6 Conclusion

A fast, flexible, and robust simulation-based optimization scheme using an ANN-

surrogate model was developed, implemented and used in the aerodynamic shape
optimization of turbomachinery blade cascades in two-dimensional flow. The opti-
mization method uses Genetic Algorithm (GA) and is combined with an Artificial
Neural Network (ANN) that uses a back propagation algorithm. GA was validated
Aerodynamic optimization of turbomachinery blades 253

Fig. 12 The original and

optimized blade profiles

Fig. 13 Efficiency vs. mass

flow rate for the NACA 65
subsonic compressor (Points 1,
3, 5, and 7 correspond to
preselected backpressure used in
the optimization while the others
are not)

against benchmark cases and ANN was assessed for aerodynamic shape optimiza-
tion. The ANN-surrogate model, used in building a low fidelity model to approximate
the optimization objective and constraints, was found to reduce the computing time
by a factor of ten.
The developed optimization scheme was successfully applied to single-point opti-
mization of transonic inviscid flow in an impulse turbine where the simulation-based
optimization scheme was assessed (ANN was not used); it was also applied to multi-
point optimization of turbulent flow in a NACA65 compressor where the ANN-based
optimization scheme was assessed. The choice of objective function and constraints
was found to be crucial for the success of the optimization process particularly for
254 T. Mengistu, W. Ghaly

Fig. 14 Efficiency vs. pressure

ratio for the NACA 65 subsonic
compressor (Points 1, 3, 5, and 7
correspond to preselected
backpressure used in the
optimization while the others are
not)

Fig. 15 Pressure ratio vs. mass

flow rate for the NACA 65
subsonic compressor (Points 1,
3, 5, and 7 correspond to
preselected backpressure used in
the optimization while the others
are not)

the multi-point optimization case; it was constructed so as to achieve a better aerody-

namic performance over the full operating range by reshaping the blade profile.
The test cases that were presented showed that the developed optimization algo-
rithm is general and capable of handling various aerodynamic design objectives and
constraints. They also showed that the present methodology was able to reach the op-
timization objective of improving the blade performance over the full operating range
while simultaneously satisfying the design constraints.
Aerodynamic optimization of turbomachinery blades 255

References

Ahmadi M, Ghaly W (1998) Aerodynamic inverse design of turbomachinery cascades using a finite vol-
ume method of unstructured meshes. Inverse Probl Eng 6:281–298
Daneshkhah K (2006) Aerodynamic inverse design of turbomachinery blading in two-dimensional viscous
flow. PhD thesis, Concordia University, Mechanical and Industrial Engineering
Dang T (1995) Inverse methods for turbomachine blades using shock-capturing techniques. AIAA paper
1995-2465
Deb K (2001) Multi-objective optimization using evolutionary algorithms. Wiley, New York
Deb KA, Joshi D (2002) A computationally efficient evolutionary algorithm for real-parameter optimiza-
tion. KanGAL, KanGAL Report No. 2002003
Dennis BH, Dulikravich GS, Han Z-X (1999) Constrained shape optimization of airfoil cascades using a
Navier–Stokes solver and a genetic/sqp algorithm. ASME paper 99-GT-441
Emery JC, Herrig LJ, Erwin JR, Felix AR (1958) Systematic two-dimensional tests of NACA 65-series
compressor blades at low speeds. NACA, NACA Report 1368
Fogel DB, Beyer H-G (1995) A note on the empirical evaluation of intermediate recombination. Evol
Comput J 491–495
Gen M, Cheng R (1997) Genetic algorithm and engineering design. Wiley, New York
Ghaly WS, Mengistu T (2003) Optimal geometric representations of turbomachinery cascades using nurbs.
Inverse Probl Eng 11(5):359–373
Goldberg DE (1985) Genetic algorithms in search, optimization and machine learning. Addison-Wesley,
Reading
Lai YY, Yuan X (2002) Blade design with three-dimensional viscous analysis and hybrid optimization
approach. AIAA paper 2002-5658
McKay M, Bechman R, Conover W (1979) A comparison of three methods for selecting values of input
variables in the analysis of output from a computer code. Technometrics 21(2):239–245
Mengistu T (2005) Aerodynamic design and optimization of turbomachinery. PhD thesis, Concordia Uni-
versity, Mechanical and Industrial Engineering
Mengistu TT, Ghaly W (2003) On the use of global optimization methods in the aerodynamic optimization
of transonic cascades. In: 11th annual CFD society of Canada, May 2003, vol 1, pp 238–247
Oyama A, Liou M-S, Obayashi S (2002) Transonic axial-flow blade shape optimization using evolutionary
algorithm and three dimensional Navier–Stokes solver. AIAA paper 2002-5642
Piegl L, Tiller W (1995) The NURBS book. Springer, Berlin
Pierret S, Demeulenaere A, Gouverneur B, Hirsch C, den Braembussche RV (2000) Designing turboma-
chinery blades with the function approximation concept and the Navier–Stokes equation. AIAA paper
2000-4879, Propulsion and Energetics Panel
Rao SS (1996) Engineering optimization: theory and practice, 3rd edn. Wiley, New York
Ray T (2003) Golinskis speed reducer problem revisited. AIAA 41:556–558
Ray T, Liew KM (2003) An optimization algorithm based on the simulation of social behaviour. IEEE
Trans Evol Comput 7(4):386–396
Reklaitis GV, Ravindran A, Ragsdell KM (1983) Engineering optimization methods and applications.
Wiley, New York
Wang X, Damodaran M (2000) Aerodynamic shape optimization using computational fluid dynamics and
parallel simulated annealing algorithms. AIAA paper 2000-4847
Luo Y-Z (2004) Parallel simulated annealing using simplex. AIAA paper 2004-4584

View publication stats