Modeling and Multi-Objective Optimization of 4-Digit NACA Airfoils Using Genetic

Algorithms

A. Khalkhali 1, H. Safikhani 2, A. Nourbakhsh2, N. Nariman-Zadeh3

1

Department of Mechanical Engineering, Faculty of Engineering, Islamic Azad University- East Tehran Branch, Tehran, Iran,
akhalkhali@qdiau.ac.ir

Hydraulic Machinery Research Institute, School of Mechanical Engineering, College of Engineering, University of Tehran,
Tehran, Iran, h.safikhani@me.ut.ac.ir, anour@ut.ac.ir
3

Department of Mechanical Engineering, Faculty of Engineering, the University of Guilan, Rasht, Iran, nnzadeh@guilan.ac.ir

Abstract
In the present study, multi-objective optimization of 4-digit NACA airfoils is performed at three steps. At the first
step, lift (CL) and drag (CD) coefficient in a set of 4-digit NACA airfoils are numerically investigated using
commercial software NUMECA. Two meta-models based on the evolved Group Method of Data Handling (GMDH)
type neural networks are obtained, at the second step, for modeling of CL and CD with respect to geometrical design
variables and the angle of attack. Finally, using obtained polynomial neural networks, multi-objective genetic
algorithms are used for Pareto based optimization of 4-digit NACA airfoils considering two conflicting objectives,
CL and CD. It is shown that some interesting and important relationships as useful optimal design principles involved
in the performance of 4-digit NACA airfoils can be discovered by Pareto based multi-objective optimization of the
obtained polynomial meta-models.
Keywords: 4-Digit NACA airfoils; Multi-objective optimization; GMDH; Genetic Algorithm.

1. Introduction
Aerodynamic optimal design to increasing the lift and decreasing the drag has been an interesting topic for a
long time. By integrating numerical computations and computational fluid dynamics, applications of numerical
optimization methods in aerodynamic designs have been improved. Oyama et al. [1] extracted the useful design
information in airfoils using Pareto optimization method at a constant angle of attack. They used a B-spline curve to
parameterize the airfoil shape and finally determined three optimum airfoils with the maximum CL and L/D and
minimum CD respectively. Razaghi et al. [2] investigated a multi-objective optimization process to NACA 0015
airfoil with a synthetic jet and at a constant angle of attack. They tried to maximize CL and minimize CD and finally
presented 5 set of optimum design variables which the designer can select each of them. Both of the above
mentioned researches have been performed at a constant angle of attack, but angle of attack is one of the design
variables in the present study.
Optimization of airfoils is indeed a multi-objective optimization problem rather than a single objective
optimization problem that has been considered so far in the literature. Both the CL and the CD in airfoils are important
objective functions to be optimized simultaneously in such a real world complex multi-objective optimization
problem. These objective functions are either obtained from experiments or computed using very timely and highcost computer fluid dynamic (CFD) approaches, which cannot be used in an iterative optimization task unless a
simple but effective meta-model is constructed over the response surface from the numerical or experimental data.
Therefore, modeling and optimization of the parameters is investigated in the present study, by using GMDH-type
neural networks and multi-objective genetic algorithms in order to maximize the lift and minimize the drag
coefficient.
System identification and modeling of complex processes using input-output data have always attracted many
research efforts. In fact, system identification techniques are applied in many fields in order to model and predict the
behaviors of unknown and/or very complex systems based on given input-output data [3]. In this way, softcomputing methods [4], which concern computation in an imprecise environment, have gained significant attention.

The main components of soft computing, namely, fuzzy logic, neural network, and evolutionary algorithms have
shown great ability in solving complex non-linear system identification and control problems. Many research efforts
have been expended to use of evolutionary methods as effective tools for system identification [5]. Among these
methodologies, Group Method of Data Handling (GMDH) algorithm is a self-organizing approach by which
gradually complicated models are generated based on the evaluation of their performances on a set of multi-inputsingle-output data pairs ( X i , y i ) (i=1, 2, , M). The GMDH was first developed by Ivakhnenko [6] as a
multivariate analysis method for complex systems modeling and identification, which can be used to model complex
systems without having specific knowledge of the systems. The main idea of GMDH is to build an analytical
function in a feed forward network based on a quadratic node transfer function [7] whose coefficients are obtained
using regression technique. In recent years, however, the use of such self-organizing networks leads to successful
application of the GMDH-type algorithm in a broad range of areas in engineering, science, and economics [6].
Moreover, there have been many efforts in recent years to deploy Genetic Algorithms (GAs) to design artificial
neural networks since such evolutionary algorithms are particularly useful for dealing with complex problems having
large search spaces with many local optima [9]. In this way, GAs has been used in a feed forward GMDH-type
neural network for each neuron searching its optimal set of connection with the preceding layer [8]. In the former
reference, authors have proposed a hybrid genetic algorithm for a simplified GMDH-type neural network in which
the connection of neurons are restricted to adjacent layers. Moreover a multi-objective genetic algorithm has also
been recently used by some of authors to design GMDH-type neural networks considering some conflicting
objectives [10, 13, 14, and 15].
In this paper CL and CD in a set of 4-digit NACA airfoils are numerically investigated using CFD techniques.
The validations of results are achieved by comparison of the results obtained in this research versus experimental
data of Abbott and Vondoenhoff [16]. Next, genetically optimized GMDH type neural networks are used to obtain
polynomial models for the effects of geometrical parameters of the airfoils, and the angle of attack on both CL and
CD. Such an approach of meta-modeling of those numerical results allows for iterative optimization techniques to
design optimally the airfoils computationally affordably. The obtained simple polynomial models are then used in a
Pareto based multi-objective optimization approach to find the best possible combinations of CL and CD, known as
the Pareto front.

2. Numerical Simulation of Airfoils

2.1. Numerical Scheme
The governing equations of incompressible flow are as follows:
Continuity equation
Vi
=0
xi
Reynolds averaged momentum equation
2
DVi
Vi
1 p
=
+
Dt
xi
xj xj

(1)

xj

uiu j

(2)

Standard kD model

V
k2
k
Dk
[(C k
] ui u j i
+ )
=
xi
xj
Dt
xj

(3)
2
Vi
k2
D
C2
[(C k
] C 1 ui u j
+ )
=
xj
k
xj
k
Dt
xj
The simulations are performed using NUMECA software. Firstly one airfoil is modeled in Auto blade 3.6 and
then the Design 2D environment of NUMECA can automatically generate the database with different design
variables. For CFD grid generation, the Auto Grid environment of Numeca is coupled with the Auto Blade
environment. A structured C-type grid system is used for calculation of the flow field around the airfoils. The
computational domain is shown in figure (1).

Table1. Details of 3 grid types used in grid independency test around NACA 4412
Grid Type
a
b
c

Pressure Side
210
255
298

Suction Side
210
268
272

Normal to the Wall

120
160
190

Total No. of Cells

25490
37432
64320

Figure 1. Schematic representation of computational grid.

To test for grid independency, three grid types (named a, b, and c) with increasing grid density are studied and
their details are listed in table (1). The computational results of 3 grid types for different angles of attack are
compared in table (2). As seen the maximum difference between the results is less than 5% so the grid type (a) is
used for all computations in present study. On the outer boundary, the uniform flow boundary conditions are
imposed, at the upstream boundary and the right (outflow) boundary condition is set to a zero velocity gradient
condition [2]. A no-slip wall boundary condition is taken on the airfoil surface. The simulations are performed under
Reynolds number 2 10 .
Table2. Comparison of CL and CD for 3 grid types around NACA 4412
6

AoA (deg)
0
4
8
12
16
20

a
0.0056
0.0073
0.0099
0.0157
0.0244
0.0366

b
0.0058
0.0072
0.0101
0.0159
0.0247
0.0368

CD
c
0.0058
0.0075
0.0096
0.0159
0.0247
0.0367

max diff (%)

3.44
4.00
4.95
1.25
1.21
0.543

a
0.397
0.866
1.246
1.543
1.694
1.412

b
0.405
0.874
1.308
1.599
1.720
1.439

CL
c
0.404
0.876
1.309
1.589
1.717
1.441

max diff (%)

1.97
1.14
4.88
3.51
1.51
2.01

2.2. Definition of the Design Variables

The design variables in the present study are the maximum camber height as percentage of chord length (x1), the
maximum camber location as percentage of chord length (x2) and the angle of attack (x3). The design variables are
shown in figure (2). In this paper the thickness of airfoils are constant and equal to 12% of chord length, in fact we

study the xx12 NACA airfoils. By changing the design variables various designs will be generated and evaluated by
CFD. Consequently, some meta-models can be optimally constructed using the GMDH-type neural networks, which
will be further used for multi-objective Pareto based design of such airfoils.

Figure 2. Definition of design variables

2.3. Validation of the CFD Results
The number of 324 various CFD runs have been performed due to those different design geometrics. Samples of
numerical results, using CFD are shown in table (3).
Table 3. Samples of numerical results using CFD

Input Data

Output Data

Num
1
2
3
4
5
6
7
8
9
10

323
324

x1

x2

x3(deg)

CL

CD

6
9
3
7
5
5
1
5
9
5

2
2
4
3
1
6
3
8
2
6

0
0
4
4
8
8
12
12
16
16

0.687
1.024
0.874
1.321
1.395
1.491
1.541
1.539
1.341
1.661

0.0061
0.0066
0.0072
0.0069
0.0094
0.0115
0.0141
0.0212
0.0278
0.0271

9
4

8
8

20
20

1.451
1.495

0.0567
0.0461

To attain the confidence about the simulation it is necessary to compare the CFD results with the experimental
data. Figure (3) shows the computed lift coefficient versus angle of attack compared with the experimental data. As
seen from this figure, the computed result is reasonably close to the graph obtained experimentally. The stall angle is
overestimated by 2o and the maximum CL by 3 %. This is so because, in general, there exists a difficulty for
numerical approaches to match the lift coefficient for angles of attack above the separation angle [17].
The results obtained in such CFD analysis can now be used to build the response surface of both the CD and the
CL for those different 324 geometries using GMDH-type polynomial neural networks. Such meta-models will, in
turn, be used for the Pareto-based multi-objective optimization of the 4-digit NACA airfoils. A post analysis using

the CFD is also performed to verify the optimum results using the meta-modeling approach. Finally, the solutions
obtained by the approach of this paper exhibit some important trade-offs among those objective functions which can
be simply used by a designer to optimally compromise among the obtained solutions.

Figure 3. Comparison of numerical and experimental result [16] for CL versus angle of attack around NACA 4412.
3. Modeling of CL and CD Using GMDH-Type Neural Network
By means of GMDH algorithm a model can be represented as set of neurons in which different pairs of them in
each layer are connected through a quadratic polynomial and thus produce new neurons in the next layer. Such
representation can be used in modeling to map inputs to outputs. The formal definition of the identification problem
is to find a function f so that can be approximately used instead of actual one, f in order to predict output y for a
given input vector X = ( x1 , x 2 , x3 ,..., x n ) as close as possible to its actual output y . Therefore, given M
observation of multi-input-single-output data pairs so that

y = f ( x , x , x ,..., x )
i
i1 i 2 i3
in

(i=1, 2 M),

(4)

It is now possible to train a GMDH-type neural network to predict the output values y for any given input vector
i

X = ( x , x , x ,..., x ) , that is
i1

i2

i3

in

y = f ( x , x , x ,..., x )
i
i1 i 2 i 3
in

(i=1, 2 M),

(5)

The problem is now to determine a GMDH-type neural network so that the square of difference between the actual
output and the predicted one is minimized, that is

M
i =1

[ f ( x , x , x ,..., x )
i1 i 2 i 3
in

y ]
i

min

(6)

General connection between inputs and output variables can be expressed by a complicated discrete form of the
Volterra functional series in the form of

y = a0 +

n n
n n n
a i xi +
a ij xi x j +
a ijk x i x j x k +... ,
i =1
i =1 j =1
i =1 j =1 k =1

(7)

Where is known as the Kolmogorov-Gabor polynomial[7]. This full form of mathematical description can be
represented by a system of partial quadratic polynomials consisting of only two variables (neurons) in the form of

y = G ( x i , x j ) = a 0 + a1 x i + a 2 x j + a 3 x i x j + a 4 x i

2
+ a5 x j .

(8)

There are two main concepts involved within GMDH-type neural networks design, namely, the parametric and
the structural identification problems. In this way, some authors presented a hybrid GA and singular value
decomposition (SVD) method to optimally design such polynomial neural networks [8]. The methodology in these
references has been successfully used in this paper to obtain the polynomial models of the CL and CD. The obtained
GMDH-type polynomial models have shown very good prediction ability of unforeseen data pairs during the training
process which will be presented in the following sections.
The inputoutput data pairs used in such modeling involve two different data tables obtained from CFD
simulation discussed in Section 2. Both of the tables consists of three variables as inputs, namely, the geometrical
parameters of the airfoils, maximum camber (x1), location of maximum camber (x2) and angle of attack (x3) and
outputs, which are CL and CD. The tables consist of a total of 324 patterns, which have been obtained from the
numerical solutions to train and test such GMDH type neural networks. However, in order to demonstrate the
prediction ability of the evolved GMDH type neural networks, the data in both inputoutput data tables have been
divided into two different sets, namely, training and testing sets. The training set, which consists of 304 out of the
324 inputoutput data pairs for CD and CL, is used for training the neural network models. The testing set, which
consists of 20 unforeseen inputoutput data samples for CD and CL during the training process, is merely used for
testing to show the prediction ability of such evolved GMDH type neural network models. The GMDH type neural
networks are now used for such inputoutput data to find the polynomial models of CD and CL with respect to their
effective input parameters. In order to design genetically such GMDH type neural networks described in the previous
section, a population of 10 individuals with a crossover probability (Pc) of 0.7 and mutation probability (Pm) 0.07
has been used in 500 generations for CD and CL. The corresponding polynomial representation for CD is as follows:
2

Y1 = 0.018 + 0.0001 x1 - 0.0017 x 2 - 4.7e - 6 x1 + 0.0002 x 2 + 0.0001x1 x2

2

C D = 0.013 - 0.95Y1 - 0.001 x3 + 30.45Y1 + 9.81e - 005 x3 + 0.072 Y1 x3

(9a)
(9b)

Similarly, the corresponding polynomial representation of the model for CL is in the form of
2

Y1 = 0.715 + 0.0819 x1 + 0.167 x2 - 0.0054 x1 - 0.0178 x2 + 0.0034 x1 x2

2

Y2 = 0.178 + 0.212 x 2 + 0.1454 x3 - 0.0173 x 2 - 0.00441 x3 - 0.0021 x 2 x3

2

Y3 = 0.006 + 0.169 x1 + 0.163 x3 - 0.0066 x1 - 0.0043 x3 - 0.0062 x1 x3

Y4 = -5.770 + 12.50Y1 - 0.572`x1 - 5.6882Y

(10a)
(10b)
(10c)

- 0.0165 x1 + 0.5824 x1Y1

(10d)

Y5 = 0.1930 - 0.5041Y2 + 1.232Y3 - 0.3342Y 2 2 - 1.0680Y3 2 + 1.507Y3 Y2

(10e)

C L = -1.661 + 2.015Y4 + 0.973Y5 - 0.424Y4 + 0.1647Y5 2 - 0.2950 Y4 Y5

(10f)

The very good behavior of such GMDH type neural network model for CD is also depicted in figure (4), both for the
training and testing data.

Figure 4. Variation of drag coefficient with input data.

Such behavior has also been shown for the training and testing data of CL in figure (5). It is evident that the evolved
GMDH type neural network in terms of simple polynomial equations successfully model and predict the outputs of
the testing data that have not been used during the training process. The models obtained in this section can now be
utilized in a Pareto multi-objective optimization of the airfoil considering both CD and CL as conflicting objectives.
Such study may unveil some interesting and important optimal design principles that would not have been obtained
without the use of a multi-objective optimization approach.

Figure 5. Variation of lift coefficient with input data.

4. Multi-Objective Optimization of 4-Digit NACA Airfoils Using Polynomial Neural Network Models
Multi-objective optimization, which is also called multi criteria optimization or vector optimization, has been
defined as finding a vector of decision variables satisfying constraints to give acceptable values to all objective
functions [11]. In these problems, there are several objective or cost functions (a vector of objectives) to be
optimized (minimized or maximized) simultaneously. These objectives often conflict with each other so that
improving one of them will deteriorate another. Therefore, there is no single optimal solution as the best with respect
to all the objective functions. Instead, there is a set of optimal solutions, known as Pareto optimal solutions or Pareto
front [12] for multi-objective optimization problems.
In order to investigate the optimal performance of the 4-digit NACA airfoils in different conditions, the
polynomial neural network models obtained in the previous sections are now employed in a multi-objective
optimization procedure. The two conflicting objectives in this study are CL and CD that are to be simultaneously
optimized with respect to the design variables, x1, x2 and x3 (figure2). The multi-objective optimization problem can
be formulated in the following form:
Maximize CL= f1 (x1, x2, x3)
Minimize CD= f 2(x1, x2, x3)
1<x1<9
Subject to:

(11)

1<x2<8
0o<x3<20o

The evolutionary process of Pareto multi-objective optimization is accomplished by using the recently developed
NSGA II algorithm, namely the K-elimination diversity algorithm [8, 10] where a population size of 60 has been
chosen in all runs with crossover probability Pc and mutation probability Pm as 0.7 and 0.07, respectively.
Figure (6) depicts the obtained non-dominated optimum design points as a Pareto front of those two objective
functions. There are five optimum design points, namely, A, B, C, D and E whose corresponding design variables and
objective functions are shown in table (4).

Figure 6. Pareto front of lift and drag coefficients for 4-digit NACA airfoils.

These points clearly demonstrate tradeoffs in objective functions CD and CL from which an appropriate design
can be compromisingly chosen. It is clear from figure (6) that all the optimum design points in the Pareto front are
non-dominated and could be chosen by a designer as optimum airfoil. Evidently, choosing a better value for any
objective function in the Pareto front would cause a worse value for another objective. The corresponding decision
variables of the Pareto front shown in figure (6) are the best possible design points so that if any other set of decision
variables is chosen, the corresponding values of the pair of objectives will locate a point inferior to this Pareto front.
Such inferior area in the space of the two objectives is in fact bottom/right side of figure (6).
Table 4. Design variables and objective functions values of Pareto points
Design Points
A
B
C
D
E

x1
2
6
7
8
8

x2
3
3
4
4
5

x3 (deg)
2.75
3.51
4.01
9.04
13.84

CD
0.0056
0.0062
0.0068
0.0121
0.0229

CL
0.7188
1.2066
1.3256
1.7058
1.8692

In figure (6), the design points A and E stand for the best CD and CL respectively. Moreover, the other optimum
design points, B and D can be simply recognized from figure (6). The design point, B exhibit important optimal
design concepts. In fact, optimum design point B obtained in this paper exhibits an increase in CD (about 3.46%) in
comparison with that of point A whilst its CL improves about 42.4%, similarly optimum design point D exhibits a
decrease in CL (about 14.25%) in comparison with that of point E whilst its CD improves about 62.6% in comparison
with that of point E. It is now desired to find a trade-off optimum design points compromising both objective
functions. This can be achieved by the method employed in this paper, namely, the mapping method. In this method,
the values of objective functions of all non-dominated points are mapped into interval 0 and 1.Using the sum of these
values for each non-dominated point, the trade-off point simply is one having the minimum sum of those values.
Consequently, optimum design point C is the trade-off points which have been obtained from the mapping method.
The Pareto front obtained from the GMDH-type neural network model (figure 6) has been superimposed with
the corresponding CFD simulation results in figure (7). It can be clearly seen from this figure that such obtained
Pareto front lies on the best possible combination of the objective values of CFD data, which demonstrate the
effectiveness of this paper, both in deriving the model and in obtaining the Pareto front.

Figure 7. Overlay graph of the obtained optimal Pareto front with the numerical data.

5. Conclusion
Genetic algorithms have been successfully used both for optimal design of generalized GMDH type neural
networks models of CL and CD in 4-digit NACA airfoils and for multi-objective Pareto based optimization of such
processes. Two different polynomial relations for CL and CD have been found by evolved GS-GMDH type neural
networks using some experimentally validated CFD simulations for inputoutput data of the airfoils. The derived
polynomial models have been then used in an evolutionary multi-objective Pareto based optimization process so that
some interesting and informative optimum design aspects have been revealed for airfoils with respect to the control
variables of airfoils geometrical parameters of x1, x2 and angle of attack (x3). Consequently, some very important
tradeoffs in the optimum design of airfoils have been obtained and proposed based on the Pareto front of two
conflicting objective functions. Such combined application of GMDH type neural network modeling of inputoutput
data and subsequent non-dominated Pareto optimization process of the obtained models is very promising in
discovering useful and interesting design relationships.
References
[1] A. Oyama, T. Nonomura and K. Fujii, Data Mining of Pareto-Optimal Transonic Airfoil Shapes Using
Proper Orthogonal Decomposition, AIAA Journal: 2009-4000.
[2] R. Razaghi, N. Amanifard and N. Nariman-Zadeh, Modeling and Multi-Objective Optimization of Stall
Control on NACA 0015 Airfoil with a Synthetic Jet Using GMDH Type Neural Networks and Genetic
Algorithms, IJE Vol.22 (2008).
[3] K.J. Astrom, K.J. and P. Eykhoff, System identification, a survey. Automatica 7-123-62 (1971).
[4] E. Sanchez, T. Shibata and Zadeh, L.A. Genetic Algorithms and Fuzzy Logic Systems. World Scientific,
Riveredge, NJ, (1997).
[5] K. Kristinson and G. Dumont, System identification and control using genetic algorithms. IEEE Trans. On
Sys. Man, and Cybern, Vol. 22, No. 5, (1992), pp. 1033-1046.
[6] A.G. Ivakhnenko, Polynomial Theory of Complex Systems. IEEE Trans. Syst. Man & Cybern, SMC-1,
364-378 (1971).
[7] S.J. Farlow, Self-organizing Method in Modeling: GMDH type algorithm. Marcel Dekker Inc (1984).
[8] N. Amanifard, N. Nariman-Zadeh, M. Borji, A. Khalkhali, A. Habibdoust, Modeling and Pareto
optimization of heat transfer and flow coefficients in micro channels using GMDH type neural networks and
genetic algorithms. Energy Conversion and Management 49 (2008) 311-325.
[9] V.W. Porto, Evolutionary computation approaches to solving problems in neural computation. In Handbook
of Evolutionary Computation Back, T., Fogel, D. B., and Michalewicz, Z. (Eds). Institute of Physics
Publishing and New York: Oxford University Press, pp D1.2:1-D1.2:6 (1997).
[10] K. Atashkari, N. Nariman-Zadeh, M. Golcu, A. Khalkhali, A. Jamali, Modeling and multi-objective
optimization of a variable valve timing spark-ignition engine using polynomial neural networks and
evolutionary algorithms. Energy Conversion and Management 2007; 48(3):102941.
[11] C.A. Coello Coello and A.D. Christiansen, (2000) Multi objective optimization of trusses using genetic
algorithms. Computers & Structures, 75, pp 647-660.
[12] V. Pareto, Cours deconomic ploitique. Lausanne, Switzerland, Rouge (1896).
[13] D.E. Goldberg, Genetic Algorithms in Search, Optimization, and Machine Learning. Addison-Wesley, New
York (1989).
[14] C.M. Fonseca and P.J. Fleming, Genetic algorithms for multi-objective optimization: Formulation,
discussion and generalization. Proc. Of the Fifth Int. Conf. On genetic Algorithms, Forrest S. (Ed.), San
Mateo, CA, Morgan Kaufmann, pp 416-423 (1993).
[15] A. Toffolo and E. Benini, Genetic Diversity as an Objective in Multi-objective evolutionary Algorithms.
Evolutionary Computation 11(2):151-167 (2003), MIT Press.
[16] H. Abbott, A.E. Von Doenhoff, Theory of Wing Sections, Dover Publications, Inc. New York, pp 61 And
62 (1958).
[17] W.K. Anderson, Thomas, J.L. and Rumsey, C.L., Application of Thin-Layer Navier Stokes Equations near
Maximum Lift, AIAA Journal: 0049-0057, (1984).

10