Navier- Stokes Optimization Using Genetic Algorithm and a Flexible

Parametric Airfoil Method

Ava Shahrokhi*, Alireza Jahangirian, Nematolah Fouladi
Aerospace Engineering Department, Amirkabir University of Technology
PO. Box 15875-4413, Tehran, IRAN
Email: [email protected]

A new method for airfoil shape parameterization is presented. This method is then applied
to Genetic Algorithm for airfoil shape optimization. Considered objective function is L/D at
transonic speed and viscous Navier-Stokes equations are solved by double time implicit
method. This method profits by the merits of explicit methods to accelerate the
convergence by solving the residual equation explicitly in a pseudo time, however it solves
the original equations implicitly and therefore it has not the limitations of explicit methods.
With the aid of this new method and spring analogy for grid movement, the computation of
objective function and optimization process is decreased. Results from usual PARSEC and
new parameterization method for airfoil shape parameterization are compared to show the
efficiency of the method with respect to the PARSEC method.

Key Words: Aerodynamic Shape Optimization, Genetic Algorithm, Airfoil Parameterization

1. INTRODUCTION
Aerodynamic shape optimization is typically a difficult problem to solve. Aerodynamic
objective functions are often multimodal and rough. In traditional deterministic gradient-based
methods, the design is updated iteratively in the direction of the steepest descent from the initial
design guess [1]. However these methods only search one part of the design space which is
suspected to converge to local optimum. On the other hand, because of non-linear behavior of
the flow equations especially in the presence of viscous effects, CFD methods for solving these
equations are very time consuming. One more problem of the gradient based methods is that
they are limited to the cases in which the number of design variables and the interactions
between parameters are low [2]. These all confirm the necessity of using another method which
is robust and suitable for aerodynamic optimization field. Among different methods for
optimization, genetic algorithms (GA) are known to be the best way for such problems. One of
the key features of GAs is that they search from a population of points and not from one special
point resulting in a global optimized point. Another advantage of using GA is that it only uses
the objective function not its derivatives which decreases the demand for calling flow solver and
high dependency of the computation time to the number of design variables. These features and
some other features, cause GA to be a robust and accurate method for aerodynamic shape
optimization aims which is widely used in the recent researches [3], [4], [5], [6].
Aircraft wings are the main subject for optimization efforts in aerodynamic design field.
Airfoils are the basic element of the wing geometry. They determine a large share of wing flow
phenomena although they are two dimensional section of the wing.
In the present study, genetic algorithm is coupled with a Navier-Stokes flow solver to optimize
an airfoil in transonic speed. To obtain a good transonic airfoil shape such as supercritical
airfoils, airfoil definition with a large degree of freedom is necessary [7]. One of the most usual
methods for airfoil representation is PARSEC method, which has been used widely in airfoil
design optimization [8]. A remarkable point of this technique is that it has been developed
Figure1- PARSEC method for airfoil parameterization Figure2- Sobieczky method for trailing edge modeling

aiming to control important aerodynamic features effectively by selecting the finite

design parameters. These parameters, which are changed during the optimization process,
define the design variables of each individual. However the need for the most efficient
configuration at transonic flow conditions calls for a more flexible airfoil optimization
especially in the vicinity of the TE.

2. AIRFOIL SHAPE PARAMETERIZATION

As mentioned above, PARSEC is one of the most usual and most effective methods for airfoil
representation in design optimization field. Figure (1) illustrates 11 basic parameters for
PARSEC method which are leading edge radius ( rLE ) , upper and lower crest
location ( X UP , Z UP , X LO , Z LO ) and curvature ( Z xxUP , Z xxLO ) , trialing edge coordinate ( Z TE ) ,
direction (α TE ) , trailing edge wedge angel ( β TE ) and thickness (∆Z TE ) . Using these
parameters, one can control the maximum curvature of the upper and lower surfaces effectively
that is very useful in reducing the shock wave strength or delaying its occurrence. A linear
combination of shape functions is used to present the airfoil shape in this method:

6 n −1
Z= ∑ an X 2 (1)
n =1

Coefficients an are determined from described geometric parameters. Despite its benefits on
controlling the important parameters on the upper and lower surfaces, PARSEC does not
provides enough control over the trailing edge shape where definitely affects the flow
phenomena and therefore aerodynamic efficiency at the ending part of a viscous flow. On the
other hand, increasing the curvature quite close to the trialing edge can reduce the boundary
layer de-cambering effect. The practical consequence for physically relevant airfoils which do
not have negative thickness or too thin TE is a concave surface shaping with curvature
increasing toward the TE at both upper and lower surfaces [8] such airfoils are called Divergent
Trailing Edge (DTE). This method for trialing edge modeling was proposed by Sobieczky [8]
which is mainly based on viscous flow control near the trailing edge that may extremely
influence aerodynamic efficiency. Figure (2) illustrates the Sobiczky method for trailing edge
modeling. In the simplest case parameters ∆α , L1 , L2 control the increment in trialing edge
thickness ∆Z added to airfoil surface to become a Divergent Trailing Edge. The function
considered for ∆Z is:
 L.Tan∆α
∆Z = .[1. − µ .ξ n − (1. − ξ n ) µ ] (2)
µ. n
Different values are possible for parameters ξ 1 , ξ 2 , µ and n. In the present study the
considered values for µ and n are 1.3 and 6 respectively.
In the present investigation, airfoil shapes are represented by a combination of PARSEC and
Sobieczky method for trailing edge modeling increasing the total number of design variables to
12 which include leading edge radius (rLE ) , upper and lower crest
location ( X UP , Z UP , X LO , Z LO ) and curvature ( Z xxUP , Z xxLO ) trialing edge direction (α TE ) and
wedge angel ( β TE ) form PARSEC method and ∆α TE , L1, L2 form Sobieczky method. This
results in having more freedom to decrease the thickness a few percent upstream of the TE.
Increasing curvature quite close to the trialing edge can create a flow in the vicinity of TE which
has a favorable pressure gradient on the airfoil surface that compensates the probable decrease
in lift mainly due to decreased camber on the upper surface. Therefore as will be shown an
optimum shape with increased lift and without any need for additional constrains is more
reasonable.

3. FLOW SOLVER

Since about 90% of the computational time required for optimization process is consumed by
flow solver, CFD solver which drives the optimization process must possess high efficiency and
convergence rate. To decrease the huge computational time required for viscous solver, an
implicit method is used. This method which is a double time implicit method was developed by
Jahangirian and doolabi for unstructured grids and has the advantages of reducing the
computational time due to utilizing convergence acceleration tools of explicit method [9].
Navier-Stokes equations are discretized using central difference technique. Unstructured grid
was used to discretize the field. The primary mesh generated around initial airfoil is moved to
be fitted to the new generated airfoils using spring analogy. To avoid interfere between
boundary layer meshes during movement, these meshes are moved rigidly with the airfoil
boundary and then the rest mesh is moved using spring analogy. Then in the optimization
process, flow properties of each cell in the primary flow field are considered as initial guess for
the new generated airfoils. This technique, also, increases the convergence rate of the flow field
solution.

4. OPTIMIZATION USING GENETIC ALGOROTHM

Among optimization algorithms, gradient-based methods are well-know optimization

algorithms which probe the optimum by calculating the local gradient information. Although
gradient-based methods are superior to other optimization algorithms in the local search, the
optimum obtained from these methods may not be a global one, especially in the aerodynamic
optimization problems [10]. Despite its simplicity, Genetic Algorithm is capable of finding a
global optimum therefore it is attractive for aerodynamic design optimization. Genetic
Algorithm is a search algorithm based on natural selection and genetics. It utilizes three
operators of reproduction, cross over and mutation. The fundamental aspects of Genetic
Algorithm are described in reference [11].
In the present study a real coded Genetic algorithm is applies to the optimization of an airfoil.
Design parameters as was mentioned are a combination of PARSEC and Sobieczky divergent
trialing edge. 12 design variables including leading edge radius (rLE ) , upper and lower crest
location ( X UP , Z UP , X LO , Z LO ) and curvature ( Z xxUP , Z xxLO ) , trailing edge direction (α TE ) , trailing
edge wedge angel ( β TE ) form PARSEC method and ∆α , L1 , L2 from Sobieczky DTE are used
to represent the airfoil. Total population is set to 20 and design parameters are bounded to create
reasonable shapes.

5. RESULTS

Using the combined method which was mentioned above, 12 airfoil shape parameters for each
member are introduced to Genetic Algorithm and changed until the optimum parameters are
obtained. The best member of each generation is directly copied to the next generation as one of
its members without any recombination or mutation. Objective function is considered as Cl/Cd
which is computed using Navier-Stokes equations at transonic Mach number of 0.73 and fully
turbulent flow of Re=6.5 million. The incidence angle is considered as 2.0 Degree. The
thickness of the airfoil is limited between 0.122 and 0.100 to avoid impractical results.
Considering RAE-2822 airfoil as initial airfoil, optimum airfoil gained by new the combined

Figure3- Initial airfoil and design shapes Figure4-Airfoil pressure coefficient distribution
for initial and designed shapes

Figure5- Mach Contours Around Initial airfoil Figure6- Mach Contours Around Designed PARSEC airfoil
method is compared with PARSEC method. Initial RAE-2822, and both designed airfoils are
shown in figure (3). This figure proves the above idea about the flexibility of airfoils obtained
from Sobieczky method in changing the curvature of the ending part of airfoil. Pressure
coefficient distributions for designed and initial airfoils are plotted in figure (4). According to
the figure, there is a considerable strong shock near the middle part of the initial airfoil.
However this shock is nearly damped after optimization with both methods. As expected,
moving toward the ending part of airfoil, airfoil curvature is increased in both upper and lower
surfaces that leads to increase the potential to produce more lift in this part of airfoil in Soieczky
method in comparison with PARSEC. Figure (4) proves this idea by showing that how Cp from
Sobieczky method distributes better than PARSEC method at the ending part of airfoil. On the
other hand, when using PARSEC method, the optimization process tends mainly to increase Cl
rather than change both Cl and Cd , so we had to apply additional constrain to limit increasing Cl
to force the optimization process to eliminate the shock by decreasing Cd. Otherwise we had to
change the objective function, for instance use pressure drag in the objective function instead of
total drag which is not very suitable because we somehow omit the viscous effects [7]. The
additional constrain also causes extra computation of objective function that increases the total

Figure7- Mach Contours Around Designed Combined Airfoil Figure8- Maximum Objective Function

Table1- characteristics of initial and designed airfoils from two methods

Airfoil Cl Cd Cl/Cd
RAE-2822 0.642 0.01340 47.90
Design (PARSEC) 0.663 0.01121 59.09
Design (Combined) 0.694 0.01093 61.13

time of optimization but using the combined method, we can eliminate such difficulties. Mach
contours for initial and two optimum airfoils are shown in figures (5) - (7). These figures show
the occurrence and weakening of the shock on the initial and design airfoils respectively.
Convergence rate of the maximum objective function is shown in figure (8). According to this
figure maximum objective function obtained from combined method is higher than that of
PARSEC. Initial value of objective function and its final values from both methods are
illustrated in table (1). According to the obtained results, objective function using the PARSEC
method is increased about 23.36 percent; however the increase in objective function from
combined method is about 27.62 percent. Similarly, both resulting Cl and Cd coefficients are
more optimum in the combined method.
 6. CONCLUSION

Using Genetic Algorithm, airfoil shape of a viscous transonic airfoil was optimized to create the
maximum Cl/Cd at a specified flight condition. In order to decrease the total computational time,
some tricks where used to speed up the objective function evaluation. Firstly, a new double time
implicit flow solver was used which has a high convergence rate. Unstructured mesh was used
to descretize the flow field. Primary mesh was moved to fit to new airfoil boundaries by spring
analogy technique. This technique has two advantages. Firstly there is no need to reproduce the
mesh for new shapes and secondly this movement provides us the chance to map the grids of the
new mesh to the old one. Therefore the new mesh can carry its superior flow characteristics as a
primary value which in tern increases the convergence time. A new combined method was used
to introduce necessary parameters of the airfoil to the optimization process. The new method
provides more flexibility in defining airfoil resulting in a more optimum shape.

7. REFERENCES

[1] Orozco,C.E., Ghattas, O.N., 'Infeasible Path Optimal Design Methods with Applications to
Aerodynamic Shape Optimization', AIAA Journal, vol.34, No.2, Ferbruary 1996, pp. 217-224.
[2] Kim, H.J., Nakahashi, K., 'Aerodynamic Design Optimization Using Unstructured Navier-
Stokes Adjoint Method', 24th International Congress of the Aeronautical Sciences(ICAS2004).
[3] Marco, N., Lanteri, S.,'A Two-Level Parallelization Strategy for Genetic Algorithm
Applied to Optimum Shape Design', Parallel Computing 26, 2000, pp. 377-397.
[4] Sasaki, D., Obayashi, Sh., Nakahashi, K., 'Navier-Stokes Optimization of Supersonic
Wings with Four Objectives Using Evolutionary Algorithm', Journal of Aircraft, vol.39, No.4,
July-August 2002, pp.621-629
[5] Wang, J.F., Periaux, J., Sefrioui, M., 'Parallel Evolutionary Algorithms for Optimization
Problems in Aerospace Engineering', Journal of Computational and Applied Mathematics
149, 2002, pp. 155-169.
[6] Quagliarella, D., Cioppa, A.D., 'Genetic Algorithm Applied to the Aerodynamic Design of
Transonic Airfoils', Journal of Aircraft, vol.32. No.4, 1995, pp. 889-891.
[7] Oyama, A., Obayashi, Sh., Nakahashi, K., 'Real-Coded Adaptive Range Genetic
Algorithm and Its Application to Aerodynamic Design', Lecture Notes in Computer Science,
Vol. 1917, Springer-Verlag, Newyork,2000, pp. 712-721.
[8] Sobieczky, H., 'Parametric Airfoils and Wings', Notes on Numerical Fluid Mechanics,
pp.71-88, Vieweg, 1998.
[9] Jahangirian, A., Hadidoolabi, M., 'An Implicit Central Difference Method for Solution of
three Dimensional Unsteady Aerodynamic on Unstructured Moving Grids', 23rd AIAA
Applied Aerodynamic Conference, Toronto, June 2005.
[10] Oyama, A., Obayashi, Sh., Nakahashi, K., 'Real-Coded Adaptive Range Genetic
Algorithm and Its Application to Aerodynamic Design', Lecture Notes in Computer
Science,Vol. 1917, Springer-Verlag,Newyork, 2000, pp. 712-721.

[11] Goldberg, D.E., 'Genetic Algorithm in Search, Optimization and Machine Learning',
Massachusett, Addison-Wesley, 1989.