You are on page 1of 6

Engineering Optimization IV Rodrigues et al.

(Eds)
2015 Taylor & Francis Group, London, ISBN 978-1-138-02725-1

An improved methodology for airfoil shape optimization using surrogate


based design optimization
D. Rajaram
Manipal Institute of Technology, Manipal, Karnataka, India

R.S. Pant
Indian Institute of Technology Bombay, Mumbai, Maharashtra, India

ABSTRACT: This paper describes a Surrogate Based Design Optimization (SBDO) technique for airfoil
shape optimization. The airfoil shape is parameterized in terms of B-Splines whose control points are used as
design variables. Constraints are imposed on Maximum Camber and Thickness-to-Chord ratio of the airfoil.
Optimum Latin Hypercube sampling is used to evaluate max. Lift/Drag (L/D) ratio for candidate airfoils, using
XFOIL, an open-source aerodynamic analysis program. Using Kriging, a surrogate surface is created to obtain
an approximate value of L/D for any set of design variables. Efficient global optimization algorithm is used to
arrive at the optimum shape of airfoil for maximizing L/D at a specific angle of attack. The methodology resulted
in profiles with 20% to 30% higher L/D than two baseline airfoils viz., Wortmann FX74 mod and Selig 1210,
with several orders of magnitude lesser calls to XFOIL, as compared to a previous study, using teaching-learning
based optimization technique.

INTRODUCTION

Airfoil characteristics are of paramount importance


in aircraft design as they directly influence the performance and behavior of an aircraft. Selection of
appropriate airfoil shape is one of the key tasks
undertaken by an aircraft designer.
Airfoil shape optimization is an old problem that
goes back nearly four decades ago (Hicks et al. 1974).
In recent times, new emerging evolutionary techniques have been successfully applied in this field,
e.g., Genetic Algorithms (Liu et al. 2006) and Particle Swarm Optimization (Upali et al. 2010). But these
methods are notorious for requiring a very large number of function evaluations. In practical aircraft design
problems, such a large number of calls to analysis software are unacceptable, due to the high computational
costs involved.
In a recent study by Rajaram et al. (in prep.), the
airfoil shape optimization problem was solved using
a heuristic technique called the Teaching-Learning
based optimization algorithm suggested by Rao et al.
(2011). The aim was to obtain an airfoil profile having
highest value of Lift/Drag (L/D) ratio, operating at a
fixed angle of attack. Starting from a baseline airfoil
shape, several candidate airfoil shapes were generated. L/D for each shape was obtained by carrying
out its aerodynamic analysis using XFOIL ; an opensource aerodynamic analysis package developed by
Drela (2001), and later validated by Liu et al. (2006).

Airfoil profiles with around 30% higher L/D compared


to the baseline shape were obtained, but the exercise
required about 20,000 calls to XFOIL .
In this study, the same problem is attempted using
surrogate based design optimization techniques, to
explore if same or better solutions could be obtained
with much lesser number of calls to XFOIL .
2

OVERVIEW OF SURROGATE METHODS

Surrogates can be considered as models of models


(or meta-models), which mimic the performance of a
computationally expensive function while themselves
being very inexpensive to evaluate. They are formed by
evaluating responses of expensive functions at intelligently chosen points in the design space, followed by
fitting a surface using interpolation or regression techniques. Thereafter, new points for analysis and design
may be identified, added and reiterated, to improve the
quality of the fit. Determination of a surrogate function can be considered as a non-inverse problem, in
which we determine a continuous function of a certain set of design variables with a sparse amount of
available data.
Queipo et al. (2005) and Forrester & Keane (2009)
have reviewed the state of the art in SBDO in general;
and Ahmed & Qin (2009) have presented a review
of literature related to its application in aerodynamic
design optimization in particular.

147

The data on an airfoil shape is essentially a set of


points on a 2D plane. These points are fitted to the
parametric b-spline representation listed in Eq. 1.

Where,
C(t) is the parametric curve for knot parameter t,
n is the degree of Spline interpolation,
Pi is the control points in the 2D plane, and,
Ni,n (t) is a function in a given knot interval in the knot
vector.

Figure 1. Surrogate based design optimization (Viana,


2011).

The knot vector is a non-decreasing sequence of


scalars. An open uniform knot sequence has been used
in this study, where the knots are equally spaced, and
the first and last (n + 1) knots are same for the sake of
simplicity. The value of Ni.n (t) dictates the participation and/or sensitivity of control points in a given knot
interval.
The curve fitting procedure has been coded in
MATLAB , in which the data points, knot vector and
degree are fed as input, and the control points are
obtained as an output. Curve fitting in itself is an optimization process where the length of the foot point of
the mapped curve on the original curve is successively
minimized. It is, therefore, evident that accuracy of the
fit should directly depend upon the number of control
points and degree of interpolation.

Many open-source tools are available for SBDO,


e.g., Surrogates Toolbox Version 3.0 developed by
Viana (2011), which was used in the present study.
This toolbox consists of a suite of several tools for
the Design of Experiments (DOE), fitting of surrogate models, checking the cross validation errors, and
carrying out the optimization. The basic flowchart of
the surrogate based design optimization suggested by
Viana et al. (2011) is shown in Fig. 1.

3 AIRFOIL SHAPE PARAMETERIZATION


Changing the shape of an airfoil for better performance
is a rather non-intuitive exercise. If one follows a freeform parameterization, like expressing the airfoil in
terms of a free curve passing through some points,
it will result in lack of intuitive control and difficulties in the aerodynamic analyses and optimization.
Hence, a robust method yet meaningful scheme for
parameterizing the shape is essential.
In airfoil shape parameterization, design space
restrictions have been recommended to result in meaningful changes in the geometry. The most common
approach is incorporation of constraints on Thickness
Ratio (i.e., the distance between the upper and lower
surfaces at any given location), and Camber (i.e., the
height of the mid-point of the ordinates).
An intuitive parameterization scheme for airfoil
shapes was proposed by Sobieczky (1998), using 11
airfoil related parameters. NURBS (Lepine, 2000) and
point co-ordinates to fit airfoil shapes using B-splines
(Song & Keane, 2004) have also been suggested,
which are more accurate and simpler, but result in
higher computational time. Giammichele et al. (2007)
introduced the concept of multi-resolution decomposition using B-spline interpolation to study the effects
of number of constraints, increased global control
at various decomposition levels, and the associated
advantages.
In the present study, we chose the direct B-spline
interpolation method for shape generation, since the
computational effort required to fit co-ordinates is
incurred only to fit the base airfoil and not in the
optimization routine.

3.1

Constraints on shape

Airfoil co-ordinates are typically normalized with


chord length. As defined by Giammichele (2007) et al.
we use Thickness Ratio and Camber as constraints.
Application of such constraints requires only either of
the surfaces ordinates. However, choice of the deciding ordinate may cause difference in analysis. The
change in design variables (i.e., control points of the
B-Splines) as a function of the mapped parametric
variable causing the curve to remain between input
bounds of camber or thickness at any location can be
solved for by direct inversion of the basis matrix, using
Eq. (2) as:

148

C(t), which is the displacement allowed in the coordinates so as to constrain the airfoil within given limits
of Thickness ratio and Camber at any location can be
obtained by Eq. (3) as:

It may be noted that this corresponds to a change


in the position of the control points as given by the
RHS. A solution with minimum change in variables is
obtained by minimizing || P||.

Table 1.

Constraints imposed on the airfoils.

Parameter

Figure 2. Design of experiments. (a) Classical design. (b)


Space filling design (Viana et al. 2011).

Minimum

Maximum

Thickness Ratio
Camber

Wortmann FX74 mod


10 %
8%

14%
11%

Thickness Ratio
Camber

Selig 1210
9%
6%

12%
8%

The change in design variables can be estimated


using Eq. (4) as:

The LHS of Eq. (4) yields the displacement in the


control points so as to maintain the curve between
given limits of Thickness Ratio and Camber. The
number of design variables is decided by the values
obtained in P. All the control points that have a nonzero value in P automatically become the design
variables.
Having obtained the change in design variables
imposed by constraints, we proceed to find the best
solution in the feasible design space using SBDO.
4

DESIGN OF EXPERIMENTS

Design of Experiments (DOE) is a vital part of SBDO,


since it can have a direct impact on the efficacy of the
surrogate in terms of the quality of fit.To fit a surrogate
model, the objective function has to be evaluated at
some points in the design domain. These trial points
should be well distributed in the domain, to capture
the behavior well, but should also be small in number
to keep the computational cost low. As a thumb-rule,
it is suggested that the number of trial points should
be at least ten times the number of design variables.
Keeping in mind such trade-offs, various designs
such as full-factorial design (FFD), partial factorial
design (PFD), face-centered cubic (FCC), central
composite design (CCD) and D-optimal design have
been suggested. Although they are easy to implement, the number of points increase as the factor and
levels go higher. To counter these problems, space filling designs are used where deterministic errors are
expected. Latin Hypercube Sampling (LHS) (McKay
et al., 1979) and Orthogonal Arrays (OA) (Hedayat
et al., 1999) are some examples (Fig. 2).
In this study, an Optimum Latin Hypercube Sampling (OLHS) design consisting of 60 training points
was generated, which were then used for fitting a
surrogate surface.
5 PROBLEM SETUP AND COMPUTATIONAL
FORMULATION
In this study, our main aim was to improve performance of a small unmanned aircraft in cruise, by

Figure 3. Scheme for coupling MATLAB & XFOIL .

maximizing the L/D of the wings airfoil while operating at an angle of attack of 2.5 degrees. The study was
carried out for two baseline airfoils, viz., Wortmann
FX74 mod and Selig 1210 operating at Reynolds number of 0.3 Million and Mach number of 0.2, which are
typical for such aircraft.
Table 1 Lists the constraints imposed on the values of design variables, ensuring that Thickness Ratio
and Camber of the airfoils were within some specified
limits.
The flow analysis is carried out in XFOIL , and
the optimization is carried out in MATLAB , hence an
interface between these two is created, to automate the
procedure.
XFOIL can be directed from MATLAB by creating an input file which is essentially a text file that
contains all the commands that define the problem.
This is followed by the creation of a batch file. Batch
file is an executable file that runs XFOIL and loads
the input file into it. Finally, data importing facilities
of MATLAB need to be exploited to import the result
(of analysis) from XFOIL into MATLAB . The entire
process was written as a function in MATLAB which
returned the L/D value of a specified set of design
variables. Fig. 3 shows the framework described above.

SURROGATE FITTING AND SELECTION

The points were uniformly generated using the OLHS


method, and surrogate surfaces were created using
three techniques, viz., Kriging (Krige, 1951), Polynomial Response Surface (Myers et al. 1986), and
Radial Basis Functions (Broomhead & Lowe, 1988).
The goodness of fit of each of these methods was

149

Table 2. Kriging parameters for both the airfoils.


Selig 1210 Airfoil
Regression Function
Correlation Function

2
Cross validation Error

Constant
Gaussian
[0.4094 5 0.2087 0.1448 0.2530
0.2733 0.0670 0.1315]
0.6837
399.4172
17.80

Wortmann FX74 mod Airfoil


Regression Function
Constant
Correlation Function
Gaussian
Theta
[0.2219 0.1048 3.1688 0.02707
0.0119 0.0100]
Beta
0.1211
25.0734
2
Cross validation Error
1.3092

Figure 4. Graphical representation of probability of


improvement (Forrester et al. 2009).

evaluated by carrying out cross validation analysis.


In this method, the available information is divided
into k subsets. A surrogate model is fitted k times,
omitting one set each time. The RMS error for each
surrogate is averaged to get an unbiased generalization
error metric.
Finally, Kriging method was chosen because it
provides an error estimate at each prediction point.
Kriging relies on general cues such as points that lie
closer to each other have a higher correlation, to make
predictions. The prediction at a new point involves
a general trend in the observed data captured by the
mean and the correlation matrix which captures spatial correlation. Therefore, the prediction at any given
point keeps getting better as new points are added.
Table 2 lists the key parameters of Kriging surrogate. and are parameters that capture sensitivities
in each dimension of the landscape, and 2 is the
variance.
It may be noticed that the Cross-validation error
is high for both the surrogates. However, this shortcoming is handled by the optimizer, as explained
ahead.

7 THE OPTIMIZATION ALGORITHM


The simplest and most intuitive way to find an optimum solution is to locate such a point in the surrogate.
The obvious problems that arise in doing so are failure
to address poor sampling, and a bad surrogate fit, there
is also a possibility of getting trapped in local minima.
The Efficient Global Optimization (EGO) algorithm proposed by Jones & Schonlau (1998) provides
a solution to this problem, by introducing a new
parameter called the Expected Value of Improvement
(E[I (x)]), which is the expected value of Probability
of Improvement P[I (x)], which is shown as the shaded
region in Fig. 4.
This algorithm circumvents the aforementioned
issues by adding new points for fitting at locations

Figure 5. Predicted response and E(I) (Jones, 1998).

where E(I ) is high, viz., near the minimum of the predictor (local exploitation) and where there is a high
predictor error (global search) (Refer Fig. 5).
E(I ) is calculated in the entire design space, after
which the area with a high probability of improvement
is further investigated by adding an additional point.
The algorithm has to be terminated once there is no
further improvement in the solution in a pre-defined
number of iterations (which was 10 in our case).
A useful feature of EGO is that it works quite well
even though inaccurate surrogates (with high cross validation errors) are used. This is because EGO relies
on gradient information and infill points to converge;
hence it is possible to overcome the problem of a high
cross validation error at the expense of a higher computational time, which will occur because of the error
in the replicated landscape.

150

RESULTS

8.1 Test runs on baseline airfoils


In order to determine the number of iterations for EGO,
the code was run several times, and it was found that
the solution never increased after about 30 iterations. It
is to be noted that this number depends on the number
of initial members chosen to fit the surrogate using
the OLHS method, which were 60 for Wortmann FX74
mod airfoil and 80 for the Selig 1210 airfoil.

Figure 8. Comparison of profiles.

Figure 6. Comparison of profiles.

Figure 9. Comparison of L/D.


Baseline and Optimized Selig 1210 Airfoil

Figure 7. Comparison of L/D.


Baseline and Optimized Wortmann FX74 mod Airfoil

The EGO algorithm significantly decreased the


number of calls made to XFOIL to obtain the optimal solution. The total number of calls to XFOIL
was reduced to 90 for Wortmann FX74 mod airfoil,
and 110 for Selig 1210 airfoil, respectively. However,
the reduction in computational time for the problem
was not much less, since EGO required calculation of
additional parameters, e.g., probability densities.
Fig. 6 & 7 compare the shape and L/D of the baseline
Wortmann FX74 mod with those of the optimized airfoil shape using TLBO in the previous study (Rajaram
et al. in prep.), and using SBDO in the present study.
Fig. 6 shows that the profiles obtained using TLBO
and SBDO are identical, and have a lower Camber and
Thickness ratio compared to the baseline Wortmann
FX74 mod airfoil.
Fig. 7 shows that an angle of attack of 2.5 degrees,
the optimized airfoil using TLBO and SBDO has an
L/D of 110, compared to 88 for the baseline Wortmann
FX74 mod airfoil, which represents an improvement
of more than 20%. It is noteworthy however, that the
optimized airfoil has lower or comparable L/D than
the baseline Wortmann FX74 mod airfoil at angles of
attack higher than 8 degrees.
Fig. 8 & 9 compare the shape and L/D of the baseline
Selig 1210 airfoil with those of the optimized airfoil
shape usingTLBO in the previous study (Rajaram et al.
in prep.), and using SBDO in the present study.
Fig. 8 shows that the profiles obtained using TLBO
and SBDO are slightly different, and have a higher
Camber and lower Thickness ratio compared to the
baseline airfoil.

Fig. 9 shows that an angle of attack of 2.5 degrees,


the optimized airfoils using TLBO and SBDO have
an L/D of 122, compared to 91 for the baseline Selig
1210 airfoil, which represents an improvement of more
than 30%. It is noteworthy however, that the optimized
airfoils have lower or comparable L/D than the baseline
Selig 1210 airfoil at angles of attack higher than 5
degrees.

CONCLUSIONS

This study outlines a framework for optimizing the


shape of a 2D airfoil using Surrogate Based Design
Optimization (SBDO) techniques. The optimum L/D
of the airfoil using this scheme matched with that
obtained in a previous study using TLBO optimization algorithm for both the baseline airfoils. However,
SBDO needed much lesser number of calls to the
XFOIL analysis software, which was lower by two
orders of magnitude. The benefit of reduction in number of calls to the analysis software is likely to be much
larger in problems involving high-fidelity analyses,
where the computational time and effort in running
the analyses is much higher.
REFERENCES
Ahmed, M. Y. M. & Qin, N. 2009. Surrogate-Based Aerodynamic Design Optimization: Use of Surrogates in Aerodynamic Design Optimization, 13th International Conference on Aerospace Sciences & Aviation Technology:
ASAT- 13-AE-14: pp. 126.

151

Broomhead, D.S., and Lowe, D. 1988. Radial Basis Functions, Multivariable Functional Interpolation and Adaptive Networks, Memo 4148, Royal Signals and Radar
Establishment Memorandum, Worcestershire, UK.
Drela, Mark. 2001. MIT AERO & Astro, Harold Youngren,
Aerocraft, Inc., XFOIL 6.9 User Primer.
Forrester, A. I. J., & Keane, A. J. 2009. Recent advances
in surrogate based optimization, Progress in Aerospace
Sciences: Vol. 45: pp. 5079.
Giammichele, N., Trepanier, J. & Tribes, C. 2007. Airfoil Generation and Optimization using Multiresolution
B-spline Control with Geometrical Constraints, 48th
AIAA/ASME/ASCE/AHS/ASC Structures, 15th Structural Dynamics, and Materials Conference: Honolulu,
Hawaii.
Hedayat A., Sloane, N., & Stufken, J. 1999. Orthogonal arrays: theory and applications, Springer, Series in
Statistics: Berlin.
Jones, D. & Schonlau, M. 1998. Expensive global optimization of expensive black-box functions, Journal of Global
Optimization: Vol. 13: pp. 455492.
Krige, D.G. 1951. A statistical approach to some mine valuations and allied problems at the Witwatersrand, Masters
thesis: University of Witwatersrand.
Lepine, J., Trepanier, J.-Y. & Pepin, F. 2000. Wing aerodynamic Design Using an Optimized NURBS Geometrical Representation, AIAA Paper 00-0669, 38th AIAA
Aerospace Sciences Meeting and Exhibit: Reno, NV.
Liu, Q, Li, J & Zhou, Z. 2006. Low Reynolds Number HighLift Airfoil Design for HALE Concept UAV, 24th Applied
Aerodynamics Conference: San Francisco, California.
MATLABAcademic Research, Release 6.5, ftp://ftp.iitb.ac.
in/IITB/software/MATLAB/MATLAB6.5.
McKay, M., Conover, W., & Beckman, R. 1979. A comparison of three methods for selecting values of input
variables in the analysis of output from a computer code,
Technometrics: Vol. 21: pp. 239245.

Myers, R.H., Khuri, A.I., & Carter, W.H. 1986. Response


Surface Methodology: 1966-1986, Technical Report 273,
The Office of Naval Research, Arlington, VA, USA.
Queipo, N. V., Haftka, R. T., Shyy, W., Goel, T., Vidyanathan,
R. & Tucker, P. K. 2005. Surrogate based analysis and
optimization, Progress in Aerospace Sciences: Vol. 41: pp.
128.
Rajaram, D.,Akhria, H. & Omkar, S. N. in prep.Airfoil Topology Optimization using Teaching-Learning based Optimization, International Journal of Applied Metaheuristic
Computing.
Rao, R.V., Savsani, V.J. & Vakharia, D.P. 2011. Teaching
learning-based optimization: A novel method for constrained mechanical design optimization problems,
Computer-Aided Design: Volume 43: Issue 3: Pages
303315.
Sobieczky, H. 1998. Parametric Airfoils and Wings, Notes on
Numerical Fluid Mechanics: edited by K. Fujii and G.S.
Dulikravich: Vol. 68: Vieweg Verlag: pp. 7188.
Song, Wenbin, Keane, Andrew. 2004. A Study of Shape
Parameterisation Methods for Airfoil Optimisation, 10th
AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference: Albany, New York.
Upali K. Wickramasinghe, Carrese, Robert & Xiaodong Li.
2010. Designing Airfoils using a Reference Point based
Evolutionary Many-objective Particle Swarm Optimization Algorithm, WCCI 2010 IEEE World Congress on
Computational Intelligence: Barcelona, Spain.
Viana, F.A.C. 2011. SURROGATES Toolbox, Gainesville,
FL, USA, Ver.3.0, available at website https://sites.google.
com/site/srgtstoolbox/.

152

You might also like