50th AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference<br>17th AIAA 2009-2219

4 - 7 May 2009, Palm Springs, California

Multidisciplinary Design Optimization of a UAV Wing using

Kriging based Multi-Objective Genetic Algorithm
S. Rajagopal*
Aeronautical Development Establishment, DRDO, Bangalore-560093, India

Ranjan Ganguli†
Department of Aerospace Engineering, Indian Institute of Science, Bangalore-560012, India

This paper investigates the preliminary wing design of Unmanned Aerial Vehicle (UAV)
using a two step optimization approach. The first step is a single objective aerodynamic
optimization whereas the second step is a coupled dual objective aerodynamic and structural
optimization. In the single objective case, airfoil geometry is optimized to get maximum
endurance parameter at a 2D level with maximum thickness to chord ratio and maximum
camber as design variables. Constraints are imposed on the leading edge curvature, trailing
edge radius, zero lift drag coefficient and zero lift moment coefficient. After arriving at the
optimized airfoil geometry, the wing planform parameters are optimized with minimization
of wing weight and maximization of endurance parameter corresponding to the wing and
four more design variables from the aerodynamics discipline namely taper ratio, aspect
ratio, wing loading and wing twist are added in the second step. Also, four more design
variables from the structures discipline namely the upper and lower skin thicknesses at root
and tip of the wing are added with stall speed, maximum speed, rate of climb, strength and
stiffness as constraints. The 2D airfoil and 3D wing aerodynamic analysis is performed by
the XFLR5 code and the structural analysis is performed by the MSC-NASTRAN software.
In the optimization process, a relatively newly developed multi-objective evolutionary
algorithm named NSGA-II (non-dominated sorting genetic algorithm) is used to capture the
full Pareto front for the dual objective problem. In the second step, in order to reduce the
time of computation, the analysis tools are replaced by a Kriging meta-model. For this dual
objective design optimization problem, numerical results show that several useful Pareto
optimal designs exist for the preliminary design of the UAV wing.

Nomenclature
UAV = Unmanned Aerial Vehicle
AR = Wing Aspect Ratio
CL = 3D Lift coefficient for wing
CD = 3D drag coefficient for wing
Cl = 2D Lift coefficient for airfoil
Cd = 2D drag coefficient for airfoil
g = Inequality constraint
HCR = Loiter Altitude in m
ROC = Rate of Climb in m/s
t = Endurance in hrs
TC = Thickness to Chord ratio of wing in percentage
TR = Wing Taper Ratio
UAV = Unmanned Aerial Vehicle
VMAX = Maximum Speed in kmph

*
Scientist; E-mail: [email protected]; Address: Aero Division, ADE, CV Raman Nagar, Bangalore-560093,
India; Phone: +918025058138; Fax: +918025283188.
†
Associate Professor; Associate Fellow AIAA; Email: [email protected]; Phone: +918022933017; Fax:
+918023600134.

1
American Institute of Aeronautics and Astronautics

Copyright © 2009 by S. Rajagopal and Ranjan Ganguli. Published by the American Institute of Aeronautics and Astronautics, Inc., with permission.
VS = Stall Speed in kmph
WL = Wing Loading in kg/m2
WW = Wing Weight in kg
x = Design variable

I. Introduction
Aircraft design is an ideal candidate for Multidisciplinary Design Optimization (MDO). This is because of the fact
that aircraft design is governed by more than one discipline like aerodynamics, structure, control and propulsion. A
conventional aircraft design process starts with aerodynamic design to satisfy the performance requirements and is
followed by design iterations and checks to satisfy requirements from other disciplines. In contrast, a good optimal
design handles all the inputs from various disciplines and performs an interdisciplinary trade off. Aircraft designers
are looking for such an optimal solution through MDO.
In the aircraft design field, many optimization works have been carried out over the last 30 years. These works
primarily focus on obtaining the best aerodynamic or structural design. Over the last decade, designers have applied
MDO to aircraft design1,2,3 . The survey of developments by Sobieski and Haftka1 reports that MDO methodology
has transcended its structural optimization roots and is growing in scope and depth toward encompassing the
complete set of disciplines required by applications. According to that survey, the two major obstacles in realizing
the full potential of MDO technology appear to be the high computational demands and complexities arising from
organization of the MDO task.
Bartholomew2 provides a definition of MDO incorporating state of the art analysis tools and discusses the
function of MDO as a key tool in the context of concurrent engineering. He also says that MDO permits the
constraints of a diverse range of disciplines to be addressed from an early stage of the design process. Kroo3
highlights some important aspects of MDO applications in the preliminary design phase and summarizes how the
field is evolving. Kroo also covers the evolution of computational tools, strategies and challenges. Most of the
researchers focus mainly on applying MDO to the complete aircraft design4-5 during its conceptual design stage and
to the design of wing6-14 during the preliminary design phase. Sobester and Keane4 studied and constructed a
multidisciplinary analysis for UAV airframes. They consider a blended wing body design and illustrate optimization
of the geometry using a constraint analysis. Rajagopal et al5 formulated the conceptual design of an UAV as an
optimization problem and performed the initial aircraft sizing through the optimization approach. During the
conceptual design phase, the different disciplines involve low fidelity analysis tools like empirical relations and
hence the MDO approach does not demand large computing power and time. On the other hand, during the
preliminary design phase, the design is maturing and hence the involved disciplines deploy high fidelity analysis
tools like Computational Fluid Dynamics (CFD), Finite Element Method (FEM) etc. that demand enormous
computing power and time. Also, during this stage, the coupling between the disciplines increases thereby
increasing the time and power of computation.
Much research in the field of aircraft design using the MDO approach has focused on applying MDO for
conventional commercial transport aircraft and current generation fighters. Grossman et al6 integrate aerodynamic
and structural design of a subsonic transport wing for minimum weight subject to a constraint on range. They
recommend two methods to alleviate the computational burden. One is to reduce the cost of sensitivity derivatives
that is called as a modular sensitivity method which allows the usage of black box disciplinary software packages.
They have shown in the study that derivatives of the aeroelastic response and divergence speed can be calculated
without the costly computation of derivatives of aerodynamic influence coefficient and structural stiffness matrices.
Apart from this, in order to reduce the computational cost, a sequential approximate optimization is used. Dovi and
Wrenn7 provide a new technique KSOPT, which is an envelope function formulation that converts a constrained
optimization problem into an unconstrained one. The primary benefit from this new method for multi-objective
optimization is the elimination of separate optimization for each objective, which is required by some optimization
methods. A typical wide body transport aircraft is used for comparison studies. This method is compared with the
other two classic multi-objective optimization methods namely the Penalty function method and Global Criterion
method. Wakayama et al8 presents the basic results from wing planform optimization for minimum drag with
constraints on structural weight and maximum lift. The study gives the basic influence of drag, weight and
maximum lift on optimal wing planform. It clearly brings out that induced drag and structural considerations
strongly favor highly tapered wings to attain large spans while parasite and compressibility drag have limited effect
on wing taper, making maximum lift constraints necessary for generating realistic tip chords. Martins et al9 focus on

2
American Institute of Aeronautics and Astronautics
demonstrating a new integrated aerodynamic structural design method for aerospace vehicles. They employ high
fidelity models for both aerodynamic and structural disciplines and also use a high fidelity coupling procedure. They
employ Euler equations for aerodynamic analysis and a detailed FEM model for the primary structure. Carrier10
describes the MDO system implemented at ONERA. It contains different optimization algorithms including a
gradient based optimizer and a GA. Here the two disciplines of aerodynamics and structure are analyzed with high
fidelity methods whereas the other disciplines such as engine performance and flight mechanics are evaluated with
simpler methods. This system is applied for optimizing the performance of a high-speed civil transport aircraft. The
overall objective is to maximize the aircraft range while multiple design constraints are considered. Kumano et al11
describe the MDO system for a small jet aircraft design by integrating the CFD codes and NASTRAN based
aeroelastic structural interface code. They employ a kriging model to save computational time of objective function
evaluation in the multi-objective genetic algorithm (MOGA). Several non-dominated solutions indicating the trade
off among the drag, structural weight, drag divergence and pitching moment are found by Obayashi and his co-
workers11. Kim et al12 provide an aerodynamic/structural multidisciplinary design with multiple objectives for a
supersonic fighter wing using response surface methodology. Nine wing and airfoil parameters were chosen for the
aerodynamic design variables and four structural variables were added to determine the wing skin thickness. To
consider various flight conditions, multipoint design optimization was performed on the three representative design
points.
In recent years, UAV’s have gained the attention of aerospace engineers due to their possibility in
reconnaissance roles related to counter terrorism. Some unique opportunities are provided by UAV design as
compared to the conventional manned aircraft design. UAV’s are not governed by the strict airworthiness
requirements of manned aircraft and provide more flexibility in the selection of wing design parameters. Some
recent research activity has focused on UAV applications13-14. Gonzalez et al13 discuss the use of evolutionary
algorithms (EA) for a single and multi-objective airfoil optimization. They bring out the demerits of applying the
gradient based approach for problems involving multi-objective, multi-modal and non-differentiable functions. Also,
they show that EA’s have the capability to find global optima and can be executed in parallel by adapting to
arbitrary solver codes. In another work, Gonzalez et al14 highlight the difficulties in the design of UAV’s arising
due to the varied and non-intuitive nature of the configurations and missions that can be performed by these
vehicles. An MDO framework14 is applied and two case studies are performed using high fidelity analysis codes.
The first case study involves dual objective UAV airfoil section optimization. Detailed design of a single element
airfoil for a small UAV application similar to RQ-7A Shadow 200 tactical UAV is performed with the two fitness
functions defined as minimization of drag at two different flight conditions. Three constraints for maximum
thickness, maximum thickness location and pitching moment are used. In the second case study, multi-criteria wing
design optimization for a UAV with the two fitness functions defined as minimization of wave drag and
minimization of the spanwise cap weight is performed. Constraints are imposed on minimum thickness and position
of maximum thickness.
We see that some works have been done on the design optimization of UAV’s. However, the use of MDO in UAV
designs is much less compared to its use in conventional manned aircraft. This is especially true for the application
of evolutionary algorithm to UAV design. Also, most works use simple GA for the optimization problem and have
not exploited the power of evolving multi-objective GAs. The advantages of evolutionary methods over classical
algorithms in single and multi-objective optimization problems are well highlighted by Goldberg15 and Deb16,17. This
paper investigates the preliminary design of an UAV wing as an optimization problem by taking into account the
aerodynamics and structural concepts. Evolutionary algorithms, which capture the full Pareto Front for multi-
objective problems are used in conjunction with a Kriging meta-model of the analysis. An example of preliminary
design of a low speed, long endurance UAV is illustrated.

II. Problem Formulation

The preliminary design of the UAV under consideration focuses on achieving its main goal i.e. the long endurance
and minimum structural weight. The fuselage design is usually governed by the amount of the fuel to be carried and
the volume of systems like payload and equipment. Therefore, only the wing design is considered for the
optimization problem. In the optimization process, the design aims at maximizing the endurance, which is an
aerodynamic aspect and minimizing the wing weight, which is a structural aspect. The optimization problem
involves objective functions, design variables and constraints.

3
American Institute of Aeronautics and Astronautics
A. Objective functions
The choice of the objective function in any aircraft optimization problem is dictated by the design mission of the
aircraft. Since the design mission of the UAV under consideration is long endurance, the main objective is chosen as
endurance while formulating the optimization problem. Also, achieving minimum structural weight is a general
challenge for aircraft designers. Therefore, minimization of wing weight is also considered as another objective for
the optimization problem. Since the aerodynamic analysis of the wing is performed in two steps, the optimization
methodology is also performed in two parts namely the 2D airfoil optimization and 3D wing optimization. In the 2D
airfoil optimization, the airfoil geometry is optimized with a single objective of maximizing the endurance
parameter ( C3/2
l / Cd ). The lift and drag characteristics correspond to the 2D characteristics of the airfoil. In the 3D
wing optimization, the wing planform parameters are optimized with dual objectives. The first objective function is
the maximization of the endurance parameter ( C3/2 L / CD ), which reflects the aerodynamic discipline. Here the lift
and drag characteristics corresponds to that of the wing. The second objective function is minimization of the wing
weight that reflects the structures discipline.

B. Design Variables
The design space is chosen to reflect the effect of aerodynamic and structural discipline. In this study, the wing
parameters related to planform, the airfoil shapes and the structural skin thicknesses are identified as the design
variables, as summarized in Table 1. The first two parameters described in Table 1 namely the wing thickness to
chord ratio and the maximum camber are the design variables for the 2D airfoil optimization. The rest of the
parameters are additional design variables included for the 3D wing optimization. The 2D design optimization uses
2 design variables while the 3D design optimization uses 8 additional design variables.
Table 1. Design Variables – Upper and Lower Bounds
No Design Variables Notation Usage Lower Bound Upper Bound
1 Wing Thickness to chord Ratio TCR 2D 0.727 1
2 Maximum camber MCR 2D 0 1
3 Wing Aspect Ratio AR 3D 0.727 1
4 Wing Loading WS 3D 0.6 1
5 Wing Taper Ratio TR 3D 0.25 1
6 Wing twist angle θ 3D 0 1
7 Upper skin thickness at root ∆t1 3D 0 1
8 Upper skin thickness at tip ∆t2 3D 0 1
9 Lower skin thickness at root ∆t3 3D 0 1
10 Lower skin thickness at tip ∆t4 3D 0 1

Throughout this paper, the design variables are normalized and presented as ratios with respect to the upper
bound value.

C. Constraints
Aerodynamic constraints are imposed on the performance parameters of the UAV and on the airfoil shape. For
the 2D airfoil optimization, the aerodynamic constraints are imposed on the leading edge curvature (LEC), trailing
edge radius (TER), zero drag and zero moment coefficient. Moreover, for the 3D wing optimization, the
aerodynamic constraints are imposed on (1) rate of climb (ROC) at that altitude, (2) stall speed (VS) and (3)
maximum speed (VMAX) at sea level condition, which arise from the requirements. The structural constraints
imposed are based on the strength and stiffness of the wing.

4
American Institute of Aeronautics and Astronautics
D. Mathematical Representation of 2D airfoil Design Problem
The 2D airfoil design problem can be written as a standard optimization problem:

 C3 / 2 
Maximize f(x) = Maximize  l 
 C 
 d 
Subject to

g (x) =  LEC(x) − LEC*  ≤ 0

1  
g (x) =  −TER(x) + TER*  ≤ 0
2  
 
g (x) = C (x) − C *  ≤ 0
3 d d
 0 0 
 
g (x) = C (x) − C *  ≤ 0
4 m m
 0 0 
x ≤ x ≤ xu
l
T
x = [TC MC ]
(1)

where Cl is the 2D lift coefficient of the airfoil, Cd is the 2D drag coefficient of the airfoil, LEC* is the maximum
allowable leading edge curvature in the airfoil geometry, TER is the minimum trailing edge angle allowed, C * is
d
0
the maximum allowable zero lift drag coefficient and C * is the maximum allowable zero lift moment coefficient.
m
0
The 2D optimization problem has one objective function, four constraints and two design variables. Only the
aerodynamics discipline is involved in this problem.

E. Mathematical Representation of 2D airfoil Design Problem

The 3D wing design problem can be written as a standard optimization problem:
 C3 / 2 
 L 
Maximize f(x) = Maximize  C 
 D 
 −WW(x)

5
American Institute of Aeronautics and Astronautics
 Subject to

g (x) =  − ROC(x) + ROC*  ≤ 0

1  

g (x) = VS(x) − VS*  ≤ 0

2  

g (x) =  −VMAX(x) + VMAX *  ≤ 0

3  
g (x) =  LEC(x) − LEC*  ≤ 0
4  

g (x) =  −TER(x) + TER*  ≤ 0

5  
g (x) = σ (x) − σ *  ≤ 0
6  

g7 (x) = δ (x) − δ *  ≤ 0
 
x ≤ x ≤ xu
l
T
x = [ AR WS TR θ ∆t1 ∆t 2 ∆t 3 ∆t 4]
(2)

where CL is the 3D lift coefficient of the wing, CD is the 3D drag coefficient of the wing, WW is the wing
weight, σ is the maximum stress, δ is the maximum deflection of the wing, ROC* is the minimum allowable rate of
climb at that altitude, VS* is the maximum allowable stall speed, VMAX* is the minimum allowable maximum
speed at sea level, σ * is the maximum allowable stress and δ * is the maximum allowable deflection of the wing.
The 3D optimization problem has two objective function, seven constraints and eight design variables. It involves
multi-objective optimization. Both aerodynamic and structural disciplines are involved in this problem.

III. Analysis
As the optimization problem involves analysis in the aerodynamics and structures discipline, the mathematical
model consists of two key components namely (i) Aerodynamic analysis and (ii) Structural analysis. These analysis
procedures are briefly described below.

A. Aerodynamic Analysis
The panel method code XFLR519 is considered for the aerodynamic analysis. Basically, XFLR5 is a user friendly
interface for the XFOIL20 code. The XFOIL code uses a higher order panel method with coupled integral boundary
layer. The algorithms for foil analysis implemented in XFLR5 are exactly the same as those of the original XFOIL
code, except for the translation from FORTRAN to C++.
The inviscid formulation of XFOIL is a simple linear-vorticity stream function panel method. A finite
trailing edge base thickness is modeled with a source panel. The equations are closed with an explicit Kutta
condition. A Karman-Tsien compressibility correction is incorporated, allowing good compressible predictions all
the way to sonic conditions. The theoretical foundation of the Karman-Tsien correction breaks down in supersonic
flow, and as a result accuracy rapidly degrades as the transonic regime is entered. Of course, shocked flows cannot
be predicted with any certainty.
As far as the viscous formulation is concerned, the boundary layers and wake are described with a two-
equation lagged dissipation integral BL formulation and an envelope en transition criterion, both taken from the
transonic analysis/design ISES code. The entire viscous solution (boundary layers and wake) is strongly interacted
with the incompressible potential flow via the surface transpiration model (the alternative displacement body model
is used in ISES). This permits proper calculation of limited separation regions. The drag is determined from the
wake momentum thickness far downstream. A special treatment is used for a blunt trailing edge which fairly
accurately accounts for base drag. The total velocity at each point on the airfoil surface and wake, with contributions
from the freestream, the airfoil surface vorticity, and the equivalent viscous source distribution, is obtained from the

6
American Institute of Aeronautics and Astronautics
panel solution with the Karman-Tsien correction added. This is incorporated into the viscous equations, yielding a
nonlinear elliptic system which is readily solved by a full-Newton method as in the ISES code.
If lift is specified, then the wake trajectory for a viscous calculation is taken from an inviscid solution at the
specified lift. If alpha is specified, then the wake trajectory is taken from an inviscid solution at that alpha. This is
not strictly correct, since viscous effects will in general decrease lift and change the trajectory. This secondary
correction is not performed, since a new source influence matrix would have to be calculated each time the wake
trajectory is changed. This would result in unreasonably long calculation times. The effect of this approximation on
the overall accuracy is small, and will be felt mainly near or past stall, where accuracy tends to degrade anyway. In
attached cases, the effect of the incorrect wake trajectory is imperceptible.
Also, wing analysis capabilities using the Vortex Lattice Method (VLM) are added in the XFLR5 code. The
wing is defined as a set of panels. Each panel is defined by its length, its root and tip chords, by its dihedral angle
and by its mesh for VLM analysis. Twist is processed as a modification of the angle of attack19. As discussed in Ref
19, the principle of a VLM is to assimilate the perturbation generated by the wing to that of a sum of vortices
distributed over the wing’s planform. The strength of each vortex is calculated to meet the appropriate boundary
conditions, i.e. non-penetration conditions on the surface of the panels. The induced drag is calculated by integration
of surface forces at the ¾ point of the VLM panels. The viscous drag is estimated by interpolation of XFOIL pre-
generated polars from the Cl value resulting from the linear VLM analysis.
The aerodynamic analysis is performed in two parts namely the 2D airfoil analysis and the 3D wing
planform analysis. Both these analysis are performed using the XFLR5 code where the 2D airfoil analysis is exactly
same as that of the XFOIL code and the 3D wing analysis is done using the VLM method.
The performance constraints are evaluated using the commercially available and well-validated aircraft design
software RDS based on the book by D. Raymer21 on the aircraft design, which is introduced after the XFLR5 code.
The sizing and synthesis analysis of this software is well-validated for many aircraft conceptual design stage
applications. The aerodynamic characteristics evaluated by the XFLR5 are used by the performance module of RDS
that evaluates the performance constraints.

B. Structural Analysis
In this research, the commercially available and well validated FEM code MSC-NASTRAN22 is used for the
structural analysis of the wing. This software is used to perform the analysis and estimate the wing weight. For
carrying out the structural analysis using the FEM code MSC-NASTRAN, the geometry is modeled using
commercially available and well validated CAD software CATIA23. This software is used to generate the model
geometry in the CAD environment and this geometric model is read by the commercially available and well
validated mesh generation code MSC-PATRAN24. The mesh generated by this software is analyzed using the FEM
solver MSC-NASTRAN.
The structural model is developed by dividing the entire structure into a number of discrete elements. The basic
steps required to perform the structural analysis using FEM solver MSC-NASTRAN are as follows,
a) The continuous structure is represented as a collection of nodal points connected by discrete elements
b) From the given element properties, material properties and geometry, the elemental stiffness matrices are
formulated
c) The global stiffness matrix corresponding to the full structure is assembled from the elemental stiffness
matrices
d) The boundary conditions are applied to constrain the model and the load vectors are generated
e) The static equilibrium equation {f}=[K]{d} where K is the system stiffness, f is the load vector and d is the
nodal displacement vector is solved. The unknowns are the nodal displacements which are evaluated by
inverting the stiffness matrix and multiplying by the force vector.
f) Other required outputs like strains and stresses can be derived from the nodal displacements.
The geometry, loads and boundary conditions, material properties are captured in the pre-processor MSC-
PATRAN. The meshing is also carried out in the pre-processor itself.

IV. Optimization
The optimization process is performed in two steps. First the airfoil is finalized through a single objective
optimization problem. Once the airfoil is frozen, the wing planform parameters are arrived at through a dual

7
American Institute of Aeronautics and Astronautics
objective problem. The optimization model is illustrated in Fig. 1 and is composed of different modules namely the
mesh module, analysis module, meta-model module and optimization module. For the first step, the analysis module
includes only the aerodynamic analysis which is performed by the XFLR5 code. For the second step, the analysis
module includes both the aerodynamics and structural analysis.

Figure 1. Optimization Model

Large computer time remains the biggest challenge in solving any MDO problem. The computation time
increases because of the strong interactions between the disciplines and the use of high fidelity codes for
aerodynamic and structural analysis. Such codes typically solve partial differential equations through discretization
methods in a procedure which is computationally intensive. Fortunately, the computation time can be drastically
reduced if accurate approximate models to replace the analysis tools can be created. Such models of models, or
meta-models, are very useful in optimization. For example, the response surface method12 and Kriging model5,11
have been successfully applied for MDO problems. While response surface methods use polynomial expressions
which are locally valid, Kriging is suited to approximate highly nonlinear functions and can be used to create
globally valid meta-models. In the present study, Kriging model is employed for design of the UAV wing. The
meta-model module generates the Kriging model and validates for sample points.

A. Kriging Model
Kriging model has its original roots in the field of geostatistics which is a mixed discipline of mining,
engineering, geology, mathematics and statistics. In this field, this meta-model is used to predict temporal and
spatial correlated data. A wide range of correlation functions can be chosen for building the meta-model thereby
making the Kriging meta-models extremely flexible.

8
American Institute of Aeronautics and Astronautics
 The analysis tool is replaced with the Kriging model in the objective function evaluation process of MOGA. As
the Kriging model introduces uncertainty at the prediction point, the biggest advantage of GA in obtaining the global
optimization may be lost25. In order to retain this advantage, both the prediction value and its uncertainty are to be
considered at the same time25. This is captured by updating the Kriging model during the optimization. If the
optimization algorithm is not converged, additional random designs are initialized and the Kriging model is
reconstructed. The Gaussian correlation function is used in the meta-model interpolates all data points exactly.
Kriging model takes a combination of a polynomial model and a statistical function and is written as follows,
y(x) = f(x) + z(x)
where y(x) is the unknown function, f(x) is the known polynomial function and z(x) is the function from a
stochastic process with mean zero, variance and non zero covariance. The polynomial function f(x) approximates
the design space globally and the localized deviations are created by the function z(x) . The polynomial function is
taken as a constant term for this study. The covariance matrix of z(x) that is responsible for the local deviations is
given below,
Cov[z(x i ), z(x j )] = σ 2 R ([R(x i , x j )])
where R is the correlation matrix, and R(x i , x j ) is the correlation function between any two of the sample data
i j
points x and x . The correlation matrix considered in this paper is the common Gaussian correlation function as
described below,
2
exp ( −θk d k )
n
∏ k =1

where n is the number of design variables, θ k is the unknown correlation parameters used to fit the model and
dk is the distance between the kth component of sample points x i and x j .
The optimization algorithm in this approach is as follows,
a) Initially some random designs are initialized. The procedure to identify these designs is to randomly choose the
design points available within 50% on both sides of the initial population.
b) The analysis code is run to construct the Kriging model for the two objective functions.
c) The MOGA is performed on the Kriging model.
d) If the convergence is not achieved, additional random designs are chosen and the optimization process is
repeated till convergence.
Further details of Kriging model are available in Ref 26.

B. NSGA II
An important aspect of this paper is the use of a multi-objective genetic algorithm NSGA-II18. This
algorithm can give the full Pareto front for multi-objective problems. The merits and demerits of the classical
gradient based approach and evolutionary approach for a given optimization problem are extensively discussed in
literature27-28. In a nutshell, the difficulties encountered by the classical optimization algorithm for general
optimization problem occur in situations with (1) non-smooth variables, (2) nonlinear and discontinuous constraints,
(3) noisy functions and (4) multiple minima. However, there are disadvantages in evolutionary optimization such as
(1) no clearly defined convergence criteria, (2) parameter tuning mostly by trial and error, (3) computationally
expensive population-based approach and (4) slow convergence to optimum. The main difference between classical
optimization and evolutionary algorithm (EA) is that EA uses a population of solutions in each of the iterations,
instead of a single solution. Since a population of solutions is processed in each of the iterations, the outcome of an
EA is also a population of solutions.
When the optimization problem involves more than one objective function, the task of finding one or more
optimum solutions is termed as multi-objective optimization16. If the objective functions are conflicting in nature,
each objective corresponds to a different optimal solution. Thus, there exist a set of optimal solutions where a gain
in one objective calls for a sacrifice in some other objective. For a designer, knowing a number of optimal solutions
becomes important and it also gives considerable insight into the design thereby providing several feasible and
useful design solutions. Therefore, as far as the designer is concerned, the ideal multi-objective optimization
procedure is to find the multiple trade-off optimal solutions with a wide range of values for the objectives and

9
American Institute of Aeronautics and Astronautics
choose one of the obtained solutions using higher-level information. Often this higher-level information would be
non-technical, qualitative and experience driven. Most of the multi-objective optimization algorithms use the
concept of domination. This concept of domination is described in detail in Ref 5 and Ref 16.
It is well known that there exist multiple Pareto optimal solutions in a problem only if the objectives are
conflicting with each other. If the objectives are not conflicting with each other, the cardinality of the Pareto optimal
set is one as shown in an example described in Ref. 16. We need to find out whether the dual objectives considered
in the conceptual design of UAV are conflicting in nature. To check for this conflict, a design space exploration is
done by performing a parametric study. Each of the design variables is allowed to vary and the variation in both the
objectives and constraints is observed. This exercise is already performed in the conceptual design of the UAV
under consideration. The results of that work are given in Ref 5.
The optimization model workflow is created in such a way that for a given starting value of design variables, it
calculates the aerodynamic characteristics namely the lift, drag and moment coefficient. Using these coefficients, it
maximizes the endurance of the UAV. Simultaneously, it generates the wing weight and checks for the constraint
violation of strength and stiffness.
Most of the multi-objective optimization algorithms use the concept of domination. In these algorithms, two
solutions are compared on the basis of whether one dominates the other solution or not. A solution x
(1) is said to
dominate the other solutions x (in other words, x is non-dominated by x ), if both the following
( 2) (1) ( 2)
conditions are true17,
(1) The solution x is no worse than x in all objectives
(1) ( 2)
(2) The solution x is strictly better than x in at least one objective
(1) ( 2)
The non-dominated set of the entire feasible search space S is the globally Pareto optimal set and these solutions
are known as non-dominated solutions (Pareto optimal solutions). The rest are called dominated solutions. Since
none of the solutions in the non-dominated set is absolutely better than any other, any one of them is an acceptable
solution. Since the NSGA-II is based on sorting the non-dominated solutions, it is referred to as non-dominated
sorting GA. There are three approaches for determining the non-dominated set17 namely naive and slow approach,
continuously updated approach and Kung et al’s efficient approach.
The steps involved in the non-dominated sorting of a population are listed below17,
- Step 1: Set all non-dominated sets P , ( j = 1,2,...) as empty sets. Set non-domination level counter j = 1
j

'
- Step 2: Use any one approaches to find the non-dominated set P of population P

' P
- Step 3: Update P = P and P =
j
P'
- Step 4: If P ≠ 0 increment j by one and go to Step 2. Otherwise, stop and declare all non-dominated sets
P for i = 1,2,..j
j
The binary coded GAs has the following difficulties:
(1) Search space becomes discrete,
(2) Hamming cliffs problem,
(3) Search restricted with variable boundaries,
(4) Arbitrary precision impossible due to fixed length coding.
To address these difficulties with binary GA, the NSGA-II uses the real coded GAs and the decision variables
are coded directly. The idea behind the NSGA-II algorithm is that a ranking selection method is used to emphasize

10
American Institute of Aeronautics and Astronautics
good points and a niche method is used to maintain stable sub populations of good points. NSGA-II is different from
the simple GA only in the way selection operation works. The mutation operations for NSGA-II are similar to
simple GA. The “Simulated Binary Crossover (SBX)” operator used in NSGA-II is slightly different from the cross
over operator of the simple GA and details are given in Ref. 22. The NSGA-II procedure is outlined as follows,
- Step 1: Combine parent P and offspring Q populations and create R = P ∪ Q . Perform a non
t t t t t
dominated sorting to R and identify different fronts: F ,i = 1,2.... etc
t i

- Step 2: Set new population P = 0 . Set a counter i = 1 . Until P + F < N , perform

t +1 t +1 i
P =P ∪ F and i = i + 1 .
t +1 t +1 i
- Step 3: Perform the crowding sort procedure and include most widely spread solutions by using the
crowding distance values in the sorted F to P
i t +1
- Step 4: Create offspring population Q from P by using the crowded tournament selection, cross over
t +1 t +1
and mutation operators.
The NSGA-II algorithm for finding the multiple Pareto optimal solution in a multi objective optimization
problem has the three key features as discussed in Ref. 22: (a) use of an elitist principle (b) an explicit diversity
preserving mechanism and (c) emphasis on non dominated solution in a population. The NSGA-II requires a non-

( )
dominated sorting of a population of size 2N . This requires at most O MN 2 computations, where N is the
number of solutions and M is the number of function evaluations. Thus the computational complexity of this

( )
algorithm is at most O MN 2 . The greatest advantage of NSGA-II is that the diversity among the non-dominated
solutions is introduced by using the crowding comparison procedure, which is used with the tournament selection
and during the population reduction phase. Since the solutions compete with their crowding distances, no extra
niching parameter is required here. The elitism mechanism does not allow an already found Pareto optimal solution
to be deleted.

V. Results and Discussions

The low speed UAV under consideration has to loiter over the target area for many hours at a medium altitude.
From the preliminary weight estimation, the empty weight fraction is 0.53 and the useful weight (payload weight
and fuel weight) fraction is 0.47. For this study, the following parameters are set for simulations in the genetic
algorithm: Population size = 50, Number of generations = 100, Crossover probability = 0.9, Crossover distribution
index = 20, Mutation probability = 0.1, Mutation distribution index = 100. For the kriging model construction, the
number of random designs chosen is 20 and the number of additional random designs is 10. The baseline airfoil for
aerodynamic optimization is the NASA/LANGLEY LS (1)-0417 (GA (W)-1). The baseline airfoil has the following
characteristics
Maximum thickness to chord ratio = 17%
Maximum thickness position = 30.2%c
Maximum camber = 2.33%c
Maximum camber position = 20.10%c
Trailing edge gap = 0.73%c
The airfoil is discretized into 100 panels in the XFLR5 code.

As far as the 2d airfoil design optimization problem is concerned, Fig 2 shows the comparison of the airfoil
geometry for the baseline and optimized design. Figs. 3-5 present the lift, drag and endurance parameter. The
physical properties of the designed airfoil are compared with the original GA (W)-1. The thickness of the airfoil is
increased from 17% to 18.9%. The optimized airfoil is having slightly better characteristics in terms of area enclosed
by the airfoil, which is a measure of the total internal volume available for fuel storage between the spars. Optimized
airfoil maintains almost same camber of the seed airfoil as 2.51% while the leading edge curvature is reduced from
28.94% to 18.42%. The trailing edge angle is increased slightly from 3.19º to 4.62º in order to reduce the constraint
imposed to structural stiffness of the airfoil. The optimized airfoil has the following characteristics

11
American Institute of Aeronautics and Astronautics
Maximum thickness to chord ratio = 18.9%
Maximum thickness position = 26.7%c
Maximum camber = 2.51%c
Maximum camber position = 21.3%c
Trailing edge gap = 0.67%c

0.15

OPTIMIZED AIRFOIL

0.1
BASELINE AIRFOIL

0.05
y/c

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

-0.05

-0.1
x/c

Figure 2. Geometry of baseline and optimized airfoil

OPTIMIZED AIRFOIL

BASELINE AIRFOIL
Cl

a (degrees)

Figure 3. 2D Lift characteristics for baseline and optimized airfoil

12
American Institute of Aeronautics and Astronautics
 OPTIMIZED AIRFOIL

BASELINE AIRFOIL

Cl

Cd

Figure 4. 2D Drag characteristics for baseline and optimized airfoil

OPTIMIZED AIRFOIL

BASELINE AIRFOIL
Endurance Parameter

a (degrees)

Figure 5. Endurance parameter for baseline and optimized airfoil

13
American Institute of Aeronautics and Astronautics
 Table 2. Comparison of characteristics of baseline and optimized airfoil
PARAMETER Baseline airfoil Optimized airfoil
Clmax 1.85 2.1
Cl0 0.55 0.55
Alpha – max in deg 19 17
Alpha 0 in deg -4 -4
Cl-alpha (Linear range) per deg 0.1125 0.1175
Cd0 at alpha = 0 0.005 0.005
Cm0 -0.12 -0.09
Max Endurance Parameter 108 155

It can be clearly seen from Table 2 that the aerodynamic characteristics of the optimized airfoil is far
superior compared to the baseline airfoil. The Clmax has increased by 13.5% and the maximum endurance parameter
has increased by 43.5%.
For the baseline model in the structural analysis, the wing is made of three parts in the chordwise direction
namely the leading edge box, center section and the trailing edge box. For the optimization problem, only the center
box is considered. The center box consists of upper and lower skin and front and rear spars. The upper and lower
skin thicknesses at root and tip are considered as the design variables. The structural model is created in MSC-
PATRAN with loads and boundary conditions. The element chosen for modeling the wing structure is the general
shell element. Only one half of the wing is modeled with symmetric boundary conditions at the root of the wing. All
the six degrees of freedom of the nodes at the root of the wing takes the values as UZ=0, ROTX=0, ROTY=0. Also,
in order to arrest the rigid body motion of the wing, only node at the wing root takes the boundary condition UX=0,
UY=0. The air loads are given in each element of the wing. The self weight of the wing structure is also taken into
consideration which gives a weight (inertia) relief for the structure. The wing is made up of CFRP skin with
Rohacell foam core. The material properties are defined in MSC-PATRAN. The material properties considered for
this optimization problem is the quasi-isotropic properties of the composite material under consideration. Fig 6 gives
the FEM model created in the MSC-PATRAN.

Figure 6. FEM Model of the 3D wing

14
American Institute of Aeronautics and Astronautics
 The structural analysis results are presented in Fig 7-8. The displacement and stress contour of the optimized
wing is shown in Fig 7 and Fig 8 respectively.

Figure 7. Displacement contour of the optimized 3D wing

Figure 8. Stress contour of the optimized 3D wing

15
American Institute of Aeronautics and Astronautics
 As far as the 3D wing design optimization problem is concerned, a graph between the two objectives
namely the wing weight and endurance parameter are plotted to obtain the Pareto front. The Pareto front is shown in
Fig 9. It is a typical Min-Max Pareto front. Five design solutions are possible with optimum solutions lying in the
Pareto front. These Pareto points are designated with PP1 to PP5. The design variables corresponding to the two
extreme Pareto points PP1 and PP5 are described in Table 3.

0.9
E n d u r a n c e P a ra m e te r ra ti o

0.8

0.7

0.6

0.5

0.4
0.06 0.065 0.07 0.075 0.08 0.085 0.09 0.095 0.1 0.105 0.11
Wing weight ratio

Figure 9. Pareto Front

Table 3. Design Variables corresponding to Pareto points PP1 and PP5

Design Variables Pareto Point PP1 Pareto Point PP5
Wing Aspect Ratio 0.82 0.98
Wing Loading 0.99 0.91
Wing Taper Ratio 0.3125 0.45
Wing twist angle 0.4 0.93
Upper skin thickness at root 0.73 0.92
Upper skin thickness at tip 0.37 0.44
Lower skin thickness at root 0.73 0.92
Lower skin thickness at tip 0.37 0.44
Wing Weight 0.067 0.105
Endurance parameter 0.4 0.92

VI. Conclusion
An evolutionary optimization algorithm capable of finding the full Pareto front of multi-objective optimization
problems is applied for preliminary UAV wing design. The preliminary design problem is formulated as a
mathematical optimization problem. A single objective problem of maximizing the endurance parameter and a dual
objective problem of maximizing endurance parameter and minimizing wing weight is considered. The formulation
and workflow is done in two steps, initially a 2D airfoil optimization where the airfoil geometry is optimized to give
a maximum endurance parameter with constraints on geometrical parameters. In the second step, the 3D wing
planform is optimized with objective functions from aerodynamics and structures. In the second step, the Kriging

16
American Institute of Aeronautics and Astronautics
approximation model, replaces both the aerodynamic and structural analysis code. Employing the Kriging model
reduces the computation time. The robust evolutionary algorithm NSGA-II used to capture the Pareto front is
capable of identifying the trade-off between the conflicting objectives thereby providing alternative useful designs
for the designer. It is found that the five Pareto designs obtained offer various possibilities to the designer in terms of
higher-level requirement and may offer novel and non-traditional solutions to the design problem. The Kriging
model employed reduces the computation time without losing the Pareto points. The optimization results confirm
the feasibility of the Kriging based MOGA for UAV wing design optimization.

Acknowledgments
The authors gratefully acknowledge the support given by all colleagues who have helped directly and indirectly
in carrying out this work and Director, ADE (DRDO) for permitting the paper to be published.

References
1
Sobieski and Haftka, R.T., “Multidisciplinary Aerospace Design Optimization: Survey of Recent
Development”, AIAA Paper No. 96-0711, 1996
2
Bartholomew, P., “The Role of Aerospace Design and Progress Towards an MDO Capability”, AIAA Paper
No. 98-4705, 1998
3
Kroo, I., “Multidisciplinary Optimization Applications in Preliminary Design”, 38th
AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics and Materials Conference, Kissimmee, Florida, USA,
7-10 April, 1997
4
Sobester, A., and Keane, A.J., “Multidisciplinary Design Optimization of UAV Airframes”, AIAA 2006-1612.
5
Rajagopal, S., Ganguli, R., Pillai, A.C.R., and Lurdharaj, A., “Conceptual Design of Medium Altitude Long
Endurance UAV using Multi Objective Genetic Algorithm”, AIAA 2007-1885.
6
Grossman, B., Haftka, R.T., Kao, P.J., Polen, D.M., and Rais-Rohani, M., “Integrated Aerodynamic-Structural
Design of a Transport wing”, Journal of Aircraft, Vol. 27, No 12, Dec 1990, pp. 1050-1056.
7
Dovi, A.R., and Wrenn, G.A., “Aircraft Design for Mission Performance using Nonlinear Multi-objective
Optimization Methods”, Journal of Aircraft, Vol. 27, No 12, Dec 1990, pp. 1043-1049.
8
Wakayama, S., and Kroo, I., “Subsonic wing planform design using multidisciplinary optimization”, AIAA
Journal, Vol. 32, No 4, Aug 1995, pp. 746-753.
9
Martins, J.R.R.A., and Alonso, J.J., “Complete configuration aero-structural optimization using coupled
sensitivity analysis method”, AIAA 2002-5402.
10
Carrier, G., “Multidisciplinary Design Optimization of a Supersonic Transport Aircraft Wing Planform”,
ECCOMAS 2004, 24-28 July 2004.
11
Kumano, T., Jeong, S., Obayashi, S., “Multidisciplinary Design Optimization of Wing Shape for a small Jet
Aircraft Using Kriging Model”, AIAA 2006-932.
12
Kim, Y., Jeon, Y.H., and Lee, D.H., “Multi-objective and Multidisciplinary Design Optimization of
Supersonic Fighter Wing”, Journal of Aircraft, Vol 43, No 3, May-June 2006, pp. 817-824.
13
Gonzalez, L.F., Lee, D.S., Srinivas, K., and Wong, K.C., “Single and Multi-objective UAV Aerofoil
Optimization via Hierarchical Asynchronous Parallel Evolutionary Algorithm”, The Aeronautical Journal, Vol 110,
No 1112, 2006, pp. 659-672.
14
Gonzalez, L.F., Periaux, J., Srinivas, K., Whitney, E.J., “A Generic Framework for Design Optimization of
Multidisciplinary UAV Intelligent Systems using Evolutionary Computing”, 44th AIAA Aerospace Sciences Meeting
and Exhibit, Reno, NV, USA, 9-12 Jan 2006.
15
Goldberg, D.E., Genetic Algorithms in Search, Optimization and Machine Learning, Addison Wesley Pub
Co., Reading, MA, 1989.
16
Deb, K., Multi Objective Optimization using Evolutionary Algorithms, Chichester, UK:Wiley, 2001
17
Deb, K., and Srinivas, N., “Multi Objective Optimization using Non Dominated Sorting Genetic Algorithm”,
Journal of Evolutionary Computation, Vol 2, No 3, 1994, pp. 221-248.
18
Deb, K., Agrawal, S., Pratap, A., and Meyarivan, T., “A Fast and Elitist Multi Objective Genetic Algorithm:
NSGA-II”, IEEE Transactions on Evolutionary Computation, Vol. 6, No 2, 2002, pp 182-197.
19
Guidelines for XFLR5 V2.00 Rev 5.
20
Drela, M., XFOIL 6.94 User guide, 2001, MIT Department of Aeronautics and Astronautics.
21
Raymer, D.P., Aircraft Design: A Conceptual Approach, AIAA Education series, AIAA, Washington DC,
1992.

17
American Institute of Aeronautics and Astronautics
 22
MSC Software Inc., “MSC Nastran User Guide and Reference manuals”, URL:
http://www.mscsoftware.com/.
23
MSC Software Inc., “MSC Patran User Guide and Reference manuals”, URL: http://www.mscsoftware.com/.
24
Dassult Systems, “CATIA V5R17 Documentation”.
25
Jeong, S., Murayama, M. and Yamamoto, K. Efficient optimization design method using Kriging model,
Journal of Aircraft, Mar-Apr 2005, 12, (2), pp. 413-420,.
26
Timothy, W.S., Timothy, M.M., John, J.K. and Farrokh, M., Comparison of response surface and kriging
models for MDO, AIAA Paper No. 98-4755, Sept 1998.
27
Lu, X. and Zhenghong, G., The investigation of multi-disciplinary and multi-objective optimization method
for the aircraft configuration design, ICAS Paper No 2002-811.1
28
Arora, J.S. and Marler, R.T., Survey of multi-objective optimization methods for engineering, Structural
Multidisciplinary Optimization, 2004, 26, pp. 369-395.

18
American Institute of Aeronautics and Astronautics