An Aerostructural Perspective on Winglets

Shahriar Khosravi⇤ and David W. Zingg†

Institute for Aerospace Studies, University of Toronto

4925 Du↵erin Street, Toronto, Ontario, M3H 5T6, Canada

This paper presents an aerostructural perspective on the potential benefits of wingletted

wings in comparison to planar wings of the same projected span. There is no consensus in

the current literature on the efficiency gains possible from winglets. Conclusions made in

the past vary significantly depending on the design problem considered and the fidelity of

the tools used. The present paper takes a step further towards understanding the tradeo↵s

in the design of wingletted wings using high-fidelity numerical optimization based on both

purely aerodynamic and fully-coupled aerostructural analysis. The high-fidelity analysis in

both cases uses the Euler equations to model the flow along with a friction drag estimate

based on the wetted surface area. Three configurations are considered: winglet-up, winglet-

down, and planar. The results show that winglets oriented downward produce a greater

drag reduction than winglets oriented upward for two reasons. First, the winglet-down

configuration moves the tip vortex further away from the wing from a purely aerodynamic

standpoint. Second, the winglet-down configuration has a higher projected span at the

deflected state due to the structural deflections. This indicates that fully-coupled high-

fidelity aerostructural optimization is required to quantify the benefits of winglets properly.

We present results for two variants of the Boeing 737NG aircraft: B737-600 and B737-900.

The winglet-down configuration can reduce the total drag by up to 2% at the same total

weight as the optimal planar counterpart. These conclusions are applicable to new wing

designs, as opposed to retrofits; the potential efficiency improvements o↵ered by retrofitted

winglets may be di↵erent.

⇤ Ph.D. Candidate, Corresponding Author, shahriar.khosravi@mail.utoronto.ca
† Professor and Director, J. Armand Bombardier Foundation Chair in Aerospace Flight

1 of 37
 Nomenclature
b Projected wing span
` Location of wing-winglet junction
b⇤ Projected span at the deflected state
M Freestream Mach number
CL Lift coefficient
q1 Freestream dynamic pressure
Cp Pressure coefficient
W Wing weight
D Total drag
W0 Wing weight of the initial design
Dinviscid Inviscid drag
WMTO Maximum takeo↵ weight
D0 Inviscid drag of the initial design
x, y, z Streamwise, spanwise, and vertical coordinates
e Span efficiency factor
Objective function parameter
g Gravitational acceleration constant
✓ Winglet cant angle
J Optimization objective function
Dihedral angle
L Lift

I. Introduction

The environmental impact of carbon emissions from commercial aviation raises concerns about the future

of the airline industry. To sustain the current growth of the industry, commercial airplanes must become

highly efficient in terms of fuel consumption. As a result, drag reduction is an important area of research in

aviation because aircraft fuel consumption is directly related to drag.

Lift-induced drag (or simply induced drag) is typically 40% of the total drag of a commercial aircraft in

cruise [1]. It is therefore worthwhile to explore concepts that reduce the induced drag. This has been the

primary motivation for researchers to study nonplanar wings and, in particular, winglets. There have been

many studies in the past decades on the possible fuel efficiency improvements provided by winglets. We

mention a few important examples in this paper; a more thorough review is provided by Kroo [1].

The term winglet was first used by Whitcomb [2] at NASA. He showed that winglets can provide a

significant improvement in the lift-to-drag ratio over planar wingtip extensions at the same level of root

bending moment. While the winglets and wingtip extensions were designed by a combination of available

theory and wind tunnel tests, the wing itself was not redesigned. At the same time, another study done

by NASA considered the relative advantages of winglets and wingtip extensions [3]. It concluded that at

the same level of root bending moment, winglets provide a greater induced drag reduction than a wingtip

extension. Later, another study by Jones and Lasinsky[4] concluded that when optimized wing shapes are

considered, similar reductions of induced drag may be achieved either by extending the wingtip or by having

a vertical winglet.

Asai [5] studied the relative advantages of planar and nonplanar wings and concluded that the tradeo↵

between the induced drag and wing root bending moment alone is not enough to determine the e↵ectiveness

2 of 37
of winglets. The e↵ects of the viscous drag penalty incurred by winglets must also be taken into account. He

further suggested that if both the root bending moment and viscous drag are kept constant, it is possible to

design a planar wing with lower total drag than any nonplanar wing. Van Dam [6] considered planar wings

that produce a nonplanar wake at a nonzero angle of attack and suggested that these geometries can provide

considerable induced drag reductions. However, much of the improvement was later attributed to numerical

integration inaccuracies [7].

More recently, Takenaka and Hatanaka [8] performed a multidisciplinary design exploration for a winglet

using high-fidelity computational fluid dynamics (CFD) and computational structural mechanics. They

considered a sample of 32 winglets having various root chord lengths, taper ratios, sweep angles, spans, cant

angles, and toe angles for multidisciplinary design optimization based on the Kriging model. Their optimized

winglet provides a reduction in total drag of approximately 22 drag counts while increasing the wing root

bending moment by 5.3%. Furthermore, they demonstrated that a conventional winglet, one that is a vertical

extension of the wingtip geometry, can only provide a reduction of 17 drag counts while increasing the root

bending moment by 3.5% in comparison to the baseline planar wing. The total drag reductions provided by

the wings with winglets were also validated using wind tunnel tests. However, it is important to note that

only the winglets were optimized in this study. The shape of the wing was not optimized.

Another notable numerical study, conducted by Verstraeten and Slingerland [9], focused on the drag

characteristics of optimally loaded planar wings, wings with winglets, and c-wings using a low-fidelity model

of the aerodynamics and weight. They concluded that when a span constraint exists, a wingletted wing with

a height-to-span ratio of 28% provides a total drag reduction of 5.4% in comparison to a planar wing at

identical wing root bending moments. This study also demonstrated that winglets can be used to provide

induced drag reductions when there is a constraint on the aspect ratio of the wing. Ning and Kroo [10]

did a similar study, but included the area-dependent weight in their calculations too. They also took into

account the e↵ects of a critical structural load factor on the tradeo↵s in the design of wings with winglets.

Another notable di↵erence in this study was the inclusion of a stall speed constraint. They demonstrated

that whether a winglet performs better than a wingtip extension depends on the ratio of the maneuver lift

coefficient to the cruise lift coefficient. This is due to the fact that the area dependent weight is a function

of the ratio of the maneuver and cruise lift coefficients in the proposed weight model. When this ratio is

equal to unity, a wingtip extension is slightly more advantageous while the winglet performs marginally

better when this ratio is equal to 2.5. This trend held true for both retrofits and new wing designs. Using a

medium-fidelity aerostructural optimization approach, Jansen et al. [11] showed that a wing with a winglet

is the globally optimal design when a span constraint exists. Furthermore, when the span is unconstrained,

the optimal design is a raked wingtip.

3 of 37
 Despite the considerable research e↵ort on this subject, there is still no clear consensus on the ability of

nonplanar wingtips to provide significant fuel efficiency improvements, especially for a new wing design, and

the conditions under which the potential benefits are seen. Conclusions made in the past vary depending

on the specific design problem considered and the level of physical detail that the models used are able to

capture. This means that more work needs to be done in order to rigorously quantify the possible efficiency

gains from winglets. Furthermore, fully-coupled high-fidelity aerostructural optimization is essential in order

to capture the interdisciplinary and nonlinear e↵ects that are important in the study of winglets. The present

study contributes to this objective by conducting aerostructural optimization based on the Euler equations

with a friction drag estimate based on the wetted surface area during the post-optimality calculations. It

is important to note that our goal is not to introduce a new methodology for performing the numerical

optimization itself. Instead, we aim to focus on valuable design insights that may not be captured using low-

and medium-fidelity optimization tools.

In the study of nonplanar wingtip devices, it is important to make a distinction between retrofitted

winglets and new wing designs. A retrofitted winglet is intended as an after-market addition to an existing

wing design in an attempt to improve the aerodynamic performance of the wing. There are many examples of

retrofitted winglets on today’s modern transport aircraft. These include the B737NG, B747-400, B767-400,

MD-11, and KC-135. In all of these examples, the addition of a nonplanar wingtip device increases the span

of the wing [12]. This is an important consideration when discussing the potential benefits of winglets. It is

difficult to rigorously quantify the efficiency improvements of winglets when the projected span of the wing

is increased because the induced drag is reduced quadratically with an increase in the span. Therefore, the

fact that these retrofitted winglets improve the performance of existing wings may not necessarily mean that

new wing designs with winglets will also outperform their planar counterparts of the same projected span.

Our present conclusions apply in the context of new wing designs, as opposed to retrofits.

Aerostructural optimization based on the Euler equations is sufficient for studying the main trends

involved in the design of wingletted wings. At the end of the optimization, the post-optimality friction drag

estimate will ensure that the increase in the wetted surface area as a result of having a winglet is taken

into account in calculating the total drag of a wingletted wing. Using this strategy leads to similar trends

obtained from high-fidelity aerodynamic shape optimization based on the Reynolds-averaged Navier-Stokes

(RANS) equations [13]. Thus, it is not necessary to model the flow based on the RANS equations for the

purpose of the current study.

Most of the test cases presented in this paper involve analyzing the lift and drag performance of wings at

a single cruise point. As a result, the individual wings obtained from the optimizations will not perform well

in other on- and o↵-design conditions. Conducting multipoint optimization would address this challenge,

4 of 37
but it requires a large amount of computational resources given the high-fidelity nature of the aerostructural

analysis. However, a significant amount of insight about the general trends in the relative performance of

wingletted and planar wings can still be learned from the present studies.

In the present work, we size the structures based on the von Mises failure criterion at a 2.5g load condi-

tion. At this stage, we do not consider structural constraints such as buckling and flutter. Although these

are important considerations in practical wing design, we believe that neglecting them does not significantly

a↵ect our conclusions. Our results could overestimate the benefits of winglets, but are unlikely to under-

estimate them. Furthermore, we do not consider additional critical structural load conditions in sizing the

structures. However, this does not adversely a↵ect our main conclusions because our approach leads to the

same structural sizing trends as the ones obtained from studies that include more load conditions [14, 15].

It is important to note that our objective in the present work is to study the main trends, not to perform

detailed wing design. The present structural sizing strategy is sufficient for this purpose.

The present study uses a step-by-step approach to studying winglets where purely aerodynamic shape

optimization is conducted first before performing fully-coupled aerostructural optimization. This approach

is central to understanding the arguments presented in this paper. Therefore, it is worthwhile to provide a

brief description of what each individual step is intended to achieve. In Section III.B, we conduct purely

aerodynamic shape optimization in order to establish that the numerical tools work as intended and recover

the expected trends. Moreover, this section demonstrates that a winglet oriented downward is more e↵ective

than a winglet oriented upward, even for purely aerodynamic shape optimization with no deflections. Section

III.C presents results from fully-coupled aerostructural optimization for planar and wingletted wings of the

same projected span. The main objective is to see what the aerostructural optimizer does given the freedom

to choose between a winglet-up and winglet-down configuration starting from an initially planar wing. In

Section III.D, we aim to determine what the optimizer chooses to do given the freedom to either keep the

wing planar or create a winglet for each configuration. Finally, Section III.E will explore the possibility that

a wingletted wing may provide a higher efficiency improvement in high-lift, low-speed conditions where the

ratio of the induced to total drag is higher than in cruise. Sections III.B and III.C use the Boeing 737-600

wing as the baseline configuration, and Sections III.D and III.E consider the larger Boeing 737-900 as the

baseline design. This will help to gain a better understanding of the sensitivity of the conclusions with

respect to di↵erent baseline aircraft. The B737-600 and B737-900 aircraft are part of the same family known

as the B737NG and have identical planforms. However, they di↵er in terms of maximum takeo↵ weight,

range, and length of the fuselage. The higher weight of the B737-900 variant leads to a larger induced drag

in cruise and may a↵ect the tradeo↵ between weight and drag.

Our choice of B737NG for the baseline aircraft is justified by the fact that this family represents a signif-

5 of 37
icant share of today’s commercial flight operations. For instance, data from the United States Department

of Transportation shows that the B737NG aircraft alone accounted for 21.1% of the total amount of fuel

burned in 2015. In fact, the B737NG aircraft burned more fuel than any other commercial aircraft family in

the United States. If we consider the similar Airbus A320 family in addition to the B737NG, the fuel burn

percentage rises to 36.5%. This means that any fuel efficiency improvement from winglets for this type of

commercial aircraft can make a significant contribution to reducing the environmental footprint of aviation.

II. Methodology

The aerostructural optimization framework used in this work consists of six main components: 1) a

multiblock Newton-Krylov-Schur flow solver for the Euler and Reynolds-averaged-Navier-Stokes equations

[16, 17], 2) a finite-element structural solver for the analysis and optimization of the structure [18], 3) a

mesh movement technique based on the linear elasticity equations for moving the aerodynamic grid during

aerostructural analysis and optimization [19], 4) a surface-based free-form deformation (FFD) technique

for moving the structures mesh during optimization [20, 21], 5) a B-spline parameterization method for

geometry control which is coupled with the linear elasticity mesh movement technique [19], and 6) the

gradient-based optimizer SNOPT [22] with gradients calculated using the discrete-adjoint method for the

coupled multidisciplinary system. Since the two discipline solvers are written in di↵erent programming

languages, we use Python to provide an interface for the solvers [23]. Zhang et al. [20, 21] provide a detailed

description of the framework in addition to extensive validation and verification studies that establish the

validity of the methods used.

The flow solver is based on an efficient parallel multiblock finite-di↵erence methodology that makes use

of summation-by-parts operators for the spatial discretization and simultaneous approximation terms for the

imposition of boundary and block interface conditions. The solution to the discrete equations at each Newton

step is computed using the generalized minimum residual scheme with approximate Schur preconditioning.

The one-equation Spalart-Allmaras turbulence model is used to model turbulent flows. All the optimization

results presented in this paper are based on the Euler equations. Hicken and Zingg [16] and Osusky and

Zingg [17] provide comprehensive details of the flow solver.

The structural analysis is performed by a parallel finite-element code called the Toolkit for the Analysis

of Composite Structures (TACS) [18]. It is capable of performing either linear or geometrically nonlinear

analysis. However, only linear analysis is considered in this study. TACS is able to handle all of the required

design variables for the thickness values of structural components inside the wing box. It is written in C++

and provides a Python interface for straightforward coupling to solvers from other disciplines. Additionally,

it has the built-in capability to provide sensitivities with respect to the design variables.

6 of 37
 Performing aerostructural analysis and optimization requires transferring the loads and displacements

between the aerodynamic and structural solvers. To accomplish this goal, rigid links are constructed between

the aerodynamic surface nodes, where the aerodynamic forces are calculated, and the closest points on the

structures finite-element mesh [24]. The displacements are then extrapolated from the structures mesh to

the aerodynamic surface grid.

The aerodynamic volume grid is parameterized by B-spline tensor volumes. The B-spline control points

on the surface of the geometry provide an efficient means for the optimizer to accomplish shape changes

during optimization. This geometry parameterization technique is integrated with a linear elasticity mesh

movement algorithm [16]. As the B-spline control points that define the surface geometry move, each B-spline

tensor volume is treated as a linearly elastic solid, for which a finite-element solution is found to represent

its new shape. The updated aerodynamic grid is then recovered from the new B-spline tensor volumes. This

integrated geometry parameterization and mesh movement technique parameterizes both the undeflected

aerodynamic geometry and the deflected flying shape.

When the optimizer changes the aerodynamic geometry, the internal structure of the wing must also move

in order to reflect the new shape. This framework uses a surface-based FFD method that parameterizes

the space enclosed between the B-spline surfaces [20, 21]. The internal structures of the wing are then

embedded inside this space between the upper and lower surfaces of the aerodynamic geometry. Thus, as

the aerodynamic geometry is changed by the optimizer, the internal structures move with it in a consistent

manner.

A gradient-based optimizer is used for optimization because gradient-based optimizers typically require

fewer function evaluations than genetic algorithms [25]. Consequently, all optimized geometries presented

represent local minima which may not be the global minimum [26]. Although we have not conducted a

formal investigation on the degree of multimodality in fully-coupled high-fidelity aerostructural optimization

problems, we have tried a few cases using multiple initial geometries. These cases suggest that the design

spaces are unimodal. The gradients of the objective and constraints are computed using the coupled adjoint

method outlined by Martins et al. [27]. The adjoint method is efficient for problems with many more design

variables than constraints because the cost of computing the gradient is nearly independent of the number

of design variables. We use the gradient-based sequential quadratic programming optimization algorithm

SNOPT, which allows for the solution of large-scale constrained problems [22]. PyOpt [23] is used to provide

a Python interface for SNOPT.

7 of 37
Figure 1: The planform of the baseline planar configuration is based on the Boeing 737NG. Only the wing
is considered for shape optimization.

III. Results and Discussion

III.A. Baseline Geometry

The baseline wing geometry for this study is loosely based on the planform of the Boeing 737NG wing shown

in Figure 1 with the RAE 2822 supercritical airfoil. All of the wingletted and planar configurations have the

same projected span. There are two main reasons for constraining the projected span. First, we are assuming

that there is an airport gate constraint that prohibits any increase in the span relative to the baseline wing.

Second, if the wingletted wings grow in span, then it will not be clear whether any aerodynamic benefit is

due to the increased span or the nonplanar feature. This is due to the fact that, based on linear aerodynamic

theory, the induced drag is reduced in a quadratic fashion with any increase in wing span [28].

The Boeing 737NG aircraft with winglets have a span that is 4.5% larger than the span of the Boeing

737NG without winglets [29]. According to linear aerodynamic theory, the total drag reduction in cruise as a

result of the same span increase is approximately 4%. The Boeing Commercial Airplanes company claims a

3% to 4% total cruise drag reduction for the wingletted wings in comparison to the baseline planar wing [12].

As a result, it is not clear how much of this improvement is due to the winglet alone. This is one of the main

reasons that we do not allow our wingletted wings to grow in span relative to the planar configuration. We

constrain the undeflected projected wing span in all cases to a value of 106 ft. Thus, we should be cautious

in comparing the potential aerodynamic performance improvement from this investigation with the values

published for the Boeing 737NG blended winglets since the latter involve an increase in span.

8 of 37
Figure 2: The geometric parameterization and de- Figure 3: The geometric parameterization and de-
sign variables for the winglet-up configuration. sign variables for the winglet-down configuration.

III.B. Aerodynamic Shape Optimization

Although the primary focus of the present paper is on using high-fidelity aerostructural analysis and op-

timization, it is insightful to begin this investigation on winglets by conducting purely aerodynamic shape

optimization. This step-by-step approach, where the optimization studies are first performed using high-

fidelity aerodynamic analysis based on the Euler equations, will lead to a solid understanding of the best-case

scenario for wingletted wings. It also helps to establish the methodology used by reproducing similar trends

observed in previous studies [13, 30, 31]. In this section, we will show that the benefits of winglets are smaller

for transonic wings of fixed span than might appear from studies of wings at lower speeds. Furthermore, we

will provide evidence that the winglet-down configuration o↵ers a larger drag reduction than the winglet-up

design in the context of purely aerodynamic shape optimization.

There are three configurations considered in this study: winglet-up, winglet-down, and planar. All three

configurations have identical projected spans. Figures 2 and 3 show the geometric parameterization and

design variables for the winglet-up and winglet-down cases, respectively. Each surface patch on the geometry

is parameterized by 6 ⇥ 6 B-spline control points. There are a total of 252 section control points of which

224 are design variables. The coordinates of the remaining control points are a function of the neighboring

design variables such that slope continuity is maintained over the entire surface of the wing. The optimizer

has the freedom to change the sectional shape of the wing by manipulating the z-coordinates of the control

points at all spanwise stations across the wing and winglet. The winglet cant angle is constrained to +71 for

the winglet-up, and 71 for the winglet-down configuration. (The cant angle of the Boeing 737NG blended

winglets is equal to +71 .) The planar configuration has a cant angle of zero. In each case, the wing sweep

and winglet cant angles are fixed. As a result, only the sectional shape and twist distribution of the wings

and winglets are free to vary over the course of the optimization. The thickness-to-chord ratio of the wing

and winglet cannot reduce by more than 30%. However, the volume of the wing is constrained to the initial

9 of 37
 100 1.3
Feasibility
Optimality 1.2
10−2

Merit Function
1.1

10−4 1.0

0.9
10 −6
0.8

10−8 0.7
0 200 400 600 0 200 400 600
Iteration Number Iteration Number
Figure 4: The convergence of optimality and fea- Figure 5: The convergence of the Lagrangian merit
sibility measures for the winglet-up aerodynamic function for the winglet-up aerodynamic optimiza-
optimization case. tion case.

value to avoid unrealistic designs.

The objective function is the inviscid drag of the wing in cruise. The cruise Mach number is M = 0.785

at an altitude of 35, 000 ft. In all three cases, the lift constraint is set to CL = 0.486. This lift constraint

is based on performing a fully-coupled aerostructural optimization for a planar wing. The aerodynamic

optimizations are performed on a grid with 653, 184 nodes. This grid does not have sufficiently fine grid

spacings for accurate prediction of drag, but is able to capture the dependence of drag on the geometry. The

final optimal geometry is analyzed using a grid with 36, 793, 008 nodes to provide a more accurate estimate of

the final objective function values during the post-optimality analysis. Figures 4 and 5 show the optimization

convergence for the winglet-up case. Feasibility is a measure of the largest nonlinear constraint violation,

and optimality is a measure of the gradient. The merit function is equal to the objective function when

all nonlinear constraints are satisfied to machine precision. For a well-converged optimum, both measures

must be as low as possible. It is clear from Figures 4 and 5 that the optimization problem has converged

to an optimal design. An important secondary conclusion is that the optimizer achieved the majority of the

reduction in the merit function over the first 200 iterations. This provides some practical justification for

ending the optimization once the majority of the merit function reduction is achieved, especially in high-

fidelity optimization where the cost of a function evaluation and the corresponding gradient calculation is

high.

Figures 6, 7, and 8 show the optimal lift distributions obtained for the planar, winglet-up, and winglet-

down configurations, respectively. The lift distribution for the planar wing closely follows the elliptical load

suggesting that the optimizer has minimized the induced drag. The optimal spanwise lift distributions for

the winglet-up and winglet-down configurations have a higher tip loading in comparison to the planar case.

10 of 37
 1.2 1.2
Cruise Cruise
Normalized Lift 1.0 Elliptical 1.0 Elliptical

Normalized Lift
0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0.0 0.0
0 5 10 15 20 0 5 10 15 20
y (m) y (m)
Figure 6: The aerodynamically optimal lift distri- Figure 7: The aerodynamically optimal lift distri-
bution for the planar configuration. bution for the winglet-up configuration.

1.2
Cruise
1.0 Elliptical
Normalized Lift

0.8

0.6

0.4

0.2

0.0
0 5 10 15 20
y (m)
Figure 8: The aerodynamically optimal lift distribution for the winglet-down configuration.

This is consistent with previous studies that show aerodynamically optimal wingtip devices tend to increase

the tip loading [12, 30, 31]. This increase in tip loading has important implications in terms of wing weight

that are ignored by the aerodynamic shape optimizer. A higher tip loading tends to increase the weight

of the wing by requiring a heavier root structure to support the greater moment. Figure 9 shows plots of

pressure coefficient along the span for the optimized winglet-up and winglet-down configurations.

Figure 10 provides a comparison of the total drag in cruise between the wingletted and planar wings.

A post-optimality friction drag estimate of the entire aircraft is included in each case based on the wetted

surface area in order to account for the increased friction drag of the wingletted configurations as a result of

the larger surface area in comparison to the planar wing. For the baseline aircraft, we use the Vehicle Sketch

Pad modeling tool developed at NASA Langley Research Center [32] to estimate that the friction drag at

cruise is equal to 192 drag counts. The larger friction drag for the wingletted wings is taken into account

11 of 37
Figure 9: Plots of pressure coefficient in cruise condition for the optimized winglet-up and winglet-down
configurations.

1.0

Pressure
0.8 Friction
Normalized Drag

63% 64% 65%

0.6

0.4

0.2 37% 36% 35%

Planar Up Down
Figure 10: Comparison of total drag in cruise for
Figure 11: Contours of x-vorticity behind the trail-
the planar and wingletted wings, including the pro-
ing edge of the wing for all three cases.
portions of pressure and friction drag.

based on the increase in the wetted surface area relative to the baseline case [33]. Figure 10 indicates that

the winglet-up configuration provides a 1.1% total drag reduction in comparison to the planar wing of the

same projected span. The reader is reminded that, unlike the B737NG wings with blended winglets, these

wingletted wings have the same projected span as the planar wing. Hence the benefit of the winglet is smaller

than that achieved when the winglet leads to a span increase. It is also noteworthy that the optimal winglet

oriented downward produces a drag reduction of 2.6% relative to the planar wing, more than twice that of

the optimal winglet oriented upward. In order to examine why the winglet-down configuration produces a

larger benefit, it is insightful to plot contours of x-vorticity behind the trailing edge of the wing in each case.

As shown in Figure 11, these provide some insight into the physical mechanism by which induced drag is

generated. Figure 11 shows that the winglet-down configuration is able to push the tip vortex further away

12 of 37
 Figure 12: Contours of x-vorticity at the deflected state for all three cases.

from the wake of the wing in the positive y-direction, as indicated by the dashed line. This means that the

induced velocity (downwash) on the wing is minimized because it has an inverse relation with the distance

from the center of the vortex core [28]. This phenomenon has also been observed in other optimization

studies using the Euler [34] and Reynolds-averaged Navier-Stokes equations [13], where the winglet-down

configuration is shown to provide a larger drag reduction in comparison to an optimal winglet-up design.

The results shown in Figure 11 are obtained from purely aerodynamic shape optimization, ignoring the

e↵ects of structural deflections. Before we turn to fully-coupled aerostructural optimization, it is insightful

to examine how the x-vorticity distribution changes in each case when the structural deflections are taken

into account. This is shown in Figure 12. In the case of the winglet-up configuration, the inboard deflection

of the winglet feature brings the tip vortex closer to the wing relative to the purely aerodynamic case. This

can reduce the e↵ectiveness of this configuration from an aerostructural standpoint. On the other hand, the

winglet-down configuration has an outboard deflection and pushes the tip vortex even further away from the

wing relative to the purely aerodynamic case. Therefore, capturing the structural deflections is essential as

it may have a significant e↵ect on our conclusions on the possible efficiency gains from winglets.

III.C. Aerostructural Optimization with Variable Winglet Cant Angle

This section presents results obtained from fully-coupled high-fidelity aerostructural optimization, where the

e↵ects of weight and structural deflections are taken into account in addition to drag. The baseline geometry

is still based on the Boeing 737NG planform shown in Figure 1 with the RAE 2822 airfoil section. We

consider the same three configurations as before: winglet-up, winglet-down, and planar. The projected span

is constrained in all cases. The optimizer is free to vary the wing sweep and winglet cant angles in addition

to the section shape and twist distribution of the wing along the span, as shown in Figure 13. Note that in

13 of 37
Figure 13: The geometric parameterization and de- Figure 14: Three possible wing shapes permitted by
sign variables for the exploratory winglet optimiza- the parameterization. (1) and (2) are winglet-down,
tion case. and (3) is planar.

this case, the optimizer is free to choose the optimal wingtip configuration by varying the cant angle design

variable. For a winglet configuration to form from the initially planar wing, the cant angle is allowed to

vary between 90 and +90 . The structure moves consistently with the changes in the winglet cant angle

by the means of the surface-based FFD technique for structural mesh movement [20, 21]. A positive cant

angle corresponds to a winglet-up and a negative cant angle to a winglet-down configuration. The optimizer

is also free to keep the wing planar by forcing the cant angle to be equal to zero. The vertical height of

the wingletted feature can increase by up to 6% of the wing span. To further clarify this, Figure 14 shows

a simplified front view of three possible wing shapes permitted by this particular parameterization. The

parameter ✓ represents the cant angle. In all cases, the projected span of the wing at the undeflected state

remains constant even though the optimizer is free to move the location of wing-winglet junction along the

span. In Figure 14, the wing-winglet junction is depicted by a red circle, and its physical location along the

span is measured by the parameter `.

The choice of the objective function significantly influences the final optimized design. For practical

design of wings, the objective is carefully chosen based on the operating requirements for a particular

aircraft. However, our main goal in this study is not to discover the best wing for a particular commercial

airplane. Instead, we are interested in the fundamental tradeo↵s between drag and weight that are involved

in the design of wings with winglets. For this reason, we choose an objective function of the form

Dinviscid W
J= + (1 ) (1)
D0 W0

where is a parameter between zero and unity, Dinviscid is the inviscid drag of the wing in cruise, W is the

calculated weight of the wing satisfying the structural failure constraints at the 2.5g load condition, and D0

and W0 are the respective initial values. As we vary from zero to unity, we place more emphasis on drag

and less on weight. This allows us to focus on the most relevant aspects of aerostructural design for various

14 of 37
geometries and gain insight into the fundamental tradeo↵ between weight and drag. Two values for have

been chosen: 0.5 and 1.0. The same friction drag estimate as the one used in Section III.B is included in the

post-optimality calculations [32, 33].

Table 1 lists the constraints for each optimization test case considered in this study. There are two lift

constraints; one corresponds to the cruise load condition, the other to a 2.5g load condition. The cruise

condition is M = 0.785 at an altitude of 35, 000 ft, while the 2.5g load condition is M = 0.798 at an altitude

of 12, 000 ft. Since the weight of the wing is a function of the structural thickness values, it changes over the

course of the optimization. The total weight of the aircraft is assumed to be equal to the computed weight

of the wing plus a fixed weight of approximately 650, 000 N. This fixed weight is estimated based on the

maximum takeo↵ weight of a Boeing 737-600 discounted by the approximate wing weight. The wing weight

is approximately equal to 8.5% of the maximum takeo↵ weight.

In practical wing design, the structures are sized based on many critical structural load conditions in

order to ensure the structural integrity of the wing. The structural sizing has a profound e↵ect on the

aerodynamic performance of the wing. By considering a single 2.5g load condition, we aim to capture some

of the e↵ects of structural sizing on the tradeo↵ between drag and weight [11, 24, 35]. This means that

we must constrain the calculated stresses on the structures at the 2.5g load condition to prevent structural

failure. The finite-element model of the primary wing structures used for these optimization test cases has

30, 473 elements. We do not constrain each individual element’s stress value because this would require

the solution of the corresponding number of coupled adjoint systems. Thus, the stresses are aggregated

using the Kreisselmeier-Steinhauser (KS) technique [36, 37, 38]. This stress aggregation approach leads to

a conservative structural design. Three KS constraints with a weighting parameter of 30 are used at the

maneuver condition: one for the ribs and spars, one for the top skin, and one for the bottom skin. We use a

material based on 7075 Aluminum with a Poisson’s ratio ⌫ = 0.33 and Young’s modulus E = 70 GPa. The

yield stress is YS = 434 MPa, and a safety factor of 2.0 is applied. In practical design of aircraft, a safety

factor of 1.5 is often applied to this specific load condition. However, since we are only considering a single

structural load case, it is appropriate to apply a higher safety factor in order to better capture the correct

trends in structural sizing of the wing.

Table 2 provides a list of the design variables used in each case. These cases have a total of 389 design

variables that control the angle of attack, geometric shape, and structural thickness distribution of the wing.

There are two angle of attack design variables: one for cruise, the other for the 2.5g load condition.

Figure 15 shows the layout of the primary structural components in relation to the outer mold line of the

wing. Skin elements are not shown for clarity. The root section is at the intersection with the fuselage. The

structural layout does not include the leading and trailing edges because the current model cannot accurately

15 of 37
 Table 1: Nonlinear constraints used for optimiza- Table 2: Optimization design variables for all
tion in all cases cases
Constraint Description Design Variable Quantity
Cruise L WMTO = 0.0 Twist Angle 3
Maneuver L 2.5WMTO = 0.0 Sweep Angle 3
Cant Angle 1
Top Skin KS  1.0 Section Shape 224
Bottom Skin KS  1.0 Angle of Attack 2
Rib/Spar KS  1.0 Skin Thickness 60
Spar Thickness 66
Wing Span b = 106 ft Rib Thickness 30
Total 6 Total 389

Y
X

Figure 15: Primary structural layout of the wing Figure 16: Each colored surface represents a struc-
in relation to the outer mold line. The wingbox tural component, the thickness of which is a design
ends at approximately 98% span. variable.

represent them. This does not adversely a↵ect our conclusions because these secondary wing structures do

not carry a significant amount of load in comparison to the primary structural components. Furthermore,

their weight is mostly dependent on the wing projected area [10], which remains constant in our study. The

primary components consist of 30 ribs, 3 spars, and 60 skin patches. The structural design variables are

the thickness values of these components. Figure 16 shows every component for which there is a thickness

design variable. The thickness of each component can vary between 5 mm and 50 mm. The displacements

and rotations of the structural model at the root section are fixed, as shown in Figure 17.

These large aerostructural optimization cases are challenging to converge due to the presence of hundreds

of design variables and many nonlinear constraints [39]. In order to reduce the difficulties associated with

convergence, we first perform the optimizations on a coarse computational fluid dynamics (CFD) mesh

with 193, 536 nodes. After 240 optimization iterations on the coarse grid, the merit function plateaus and

the optimization is continued on a finer mesh with 653, 184 nodes in 112 blocks. This aerodynamic grid

16 of 37
 Z

Figure 18: Grid resolution of the surface and sym-

Figure 17: Displacement and rotation boundary metry plane for the fine optimization mesh.
conditions for the structures finite element mesh.

100
14

10−2 12
Merit Function

10
10−4
8

10−6 6

4
10−8
Feasibility 2
Optimality
10−10 0
0 50 100 150 200 250 300 350 0 50 100 150 200 250 300 350
Iteration Number Iteration Number
Figure 19: Convergence of optimality and feasibil- Figure 20: Convergence of the Lagrangian merit
ity measures for the winglet-down aerostructural function for the winglet-down aerostructural opti-
optimization case. mization case.

does not have sufficiently fine spacings for very accurate estimations of drag. Thus, we analyze each final

optimized design using a more refined CFD mesh with 37 million nodes in order to obtain more accurate

drag estimations for the tradeo↵ curves. It is important to evaluate the final optimized designs on a fine CFD

mesh because we would like to ensure that any relative performance benefit obtained on the optimization

grid translates in the same way to the fine mesh. Performing the optimizations using this aerodynamic

grid is appropriate because, although the grid is too coarse to predict drag accurately, we have verified that

the dependence of the drag on the geometry is accurately captured [30]. Figure 18 shows shows the grid

resolution of the surface and symmetry plane for the fine optimization mesh with 653, 184 nodes.

Figures 19 and 20 show the aerostructural optimization convergence trends for the winglet-down config-

17 of 37
Figure 21: The initial and optimized wing shapes Figure 22: The initial and optimized wing shapes
for the = 0.5 case. for the = 1.0 case.

uration with = 1.0 using the coarse and fine optimization grids. The dashed line marks the beginning of

the optimization using the fine grid. At the end of the optimization using the coarse grid, the optimized

geometry is used as the initial design for the subsequent fine optimization. On the finer aerodynamic grid,

the optimizer achieves a further merit function reduction of around 2.2% in 103 function evaluations and
6
satisfies the nonlinear constraints to a tolerance of 10 . The reduction in optimality on the fine grid is

around an order of magnitude, but the fact that the merit function has plateaued means that the majority

of the reduction in the merit function is already achieved. These trends hold true for the rest of the cases as

well.

Figures 21 and 22 show the optimal winglet configurations for the = 0.5 and = 1.0 cases, respectively.

In both cases, the optimization leads to a winglet-down configuration. It is important to note that the

optimizer does have the freedom to vary the cant angle and choose between a winglet-up, winglet-down,

or planar configuration. Thus, the winglet-down configuration has better performance in comparison with

planar and winglet-up configurations of the same projected span regardless of the value of . The cant angle

of the optimized winglet is equal to 90 in both cases. As an example, Figure 23 shows the contours of the

von Mises failure criterion along the span for the winglet-down configuration with = 0.5 at the 2.5g load

condition.

Figure 24 shows plots of the pressure coefficient along the span for the initial and optimized wings on

the fine optimization grid. The initial geometry is not shock-free as indicated by the rapid pressure drop at

half-chord over a large region of the upper surface of the wing. The initial pressure distribution is produced

by the starting geometry at an angle of attack of 2 . Note that the initial design simply serves as a starting

point for the optimization and is not meant to yield good lift and drag performance. The optimizer has

18 of 37
Figure 23: Contours of the von Mises stress criterion at the 2.5g load condition for the optimized winglet-
down configuration with = 0.5.

Figure 24: Plots of the pressure coefficient in cruise condition for the optimal planar wing with = 1.0.

eliminated the associated wave drag by eliminating the shocks. The rapid pressure recovery near the trailing

edge is typical for inviscid optimizations that ignore separation [40]. Ignoring viscous e↵ects does not a↵ect

our main conclusions because if wingletted wings do not provide significant drag reductions in inviscid flow,

then it is unlikely that their potential improvements will increase in the presence of viscous e↵ects.

Figures 25 and 26 show the optimal lift distributions for the winglet-down configuration for both values

of . With = 0.5, i.e. when there is equal emphasis on both weight and drag of the wing in the objective

function, the optimal cruise lift distribution is more triangular than the elliptical load. At the 2.5g load

condition, the spanwise lift distribution is much more triangular with a significant amount of tip load relief.

In fact, the tip loading at the maneuver condition is lower than at cruise. When the objective function is

equal to the inviscid drag, i.e. = 1.0, the optimal cruise lift distribution more closely follows the elliptical

load than the = 0.5 case, which is consistent with the expected trend. The optimizer still reduces the

tip loading at the 2.5g load condition down to the cruise level by introducing washout near the tip at the

deflected state. These plots indicate that the optimizer is actively taking advantage of aeroelastic tailoring.

Furthermore, it establishes the fact that the methodology used is able to capture the expected trends.

19 of 37
 3.5 3.5
Cruise Cruise
3.0 3.0
2.5g Load 2.5g Load
2.5 Elliptical 2.5 Elliptical
Normalized Lift

Normalized Lift
2.0 2.0

1.5 1.5

1.0 1.0

0.5 0.5

0.0 0.0

-0.5 -0.5
0 5 10 15 0 5 10 15
Span (m) Span (m)
Figure 25: Lift distributions for the optimized Figure 26: Lift distributions for the optimized
winglet-down configuration with = 0.5. winglet-down configuration with = 1.0.

Figure 27: Plots of the pressure coefficient in cruise condition for the optimal winglet-down configurations.

It is insightful to see how the optimizer achieves di↵erences in optimal span loading through aeroelastic

tailoring for di↵erent values of . Figure 27 shows plots of the pressure coefficient along the span for both

cases. With = 0.5, the pressure coefficient curves for the upper and lower surfaces near the wing-winglet

junction cross over. This indicates that the force near the leading edge is downward, but it switches direction

after the cross-over point. This results in more washout near the wingtip than the = 1.0 case. We can

see evidence of the di↵erences in the corresponding twist distributions along the span as shown in Figure

28. It is clear that the optimizer is introducing greater washout towards the tip for the = 0.5 case. This

is consistent with our expectation because with = 0.5, there is equal emphasis in the objective function

on weight and drag. As a result, the optimizer should reduce the tip loading and shift the load inboard

in order to maintain a lighter structure than the = 1.0 case. Note that the twist angle at the tip of the

winglet in Figure 28 is equivalent to the toe angle and partly controls the loading on the winglet, which in

turn determines the structural deflection (bending) of the winglet. We will later on demonstrate that the

20 of 37
 0
β = 0.5
-1 β = 1.0

-2

Twist (deg)
-3

-4

-5

-6
0 5 10 15 20
y (m)
Figure 28: The wing twist distributions in cruise condition for the optimal winglet-down configurations.

1.02 Planar
Winglet-Down

Normalized Total Drag

1.00

0.98

0.96

0.94

0.92

0.90
β = 0.5 β = 1.0
Figure 30: Comparison of total drag between the
Figure 29: The initial and optimized planform
optimized planar and the optimized winglet-down
shapes for the aerostructural optimization cases with
configuration.
= 0.5 and = 1.0.

deflection of the winglet is important to consider for induced drag comparisons because it influences the tip

vortex location.

It is clear from Figure 29 that the optimizer takes advantage of the geometric freedom in terms of planform

in both cases. With = 0.5, the overall sweep angle is lower than the case with = 1.0. As the emphasis

on drag in the objective function is increased, the optimizer increases the sweep angle in order to remove

the shocks that are present in the = 0.5 case. This is consistent with the expected trend because with

increasing , the weight growth as a result of the larger sweep angle has a smaller influence on the objective

function.

Figure 30 provides a comparison of drag between optimized planar wings with a fixed cant angle of zero

and the optimized winglet-down configurations with a variable winglet cant angle for both values of . The

drag improvement relative to the planar wing is up to 1.7%. It is clear that in both cases, the winglet-down

configuration has lower drag than the optimized planar counterparts. Table 3 provides a summary of the

21 of 37
Table 3: Summary of the cruise point performance of the planar and wingletted wings from aerostructural
optimization evaluated on the fine grid
Planar Winglet-Down
Parameter = 0.5 = 1.0 = 0.5 = 1.0
CL 0.472 0.478 0.473 0.488
Total drag (counts) 321 301 317 296
D 0.0% 0.0% -1.3% -1.7%
L/D 14.7 15.9 14.9 16.5
b⇤ 0.0% 0.0% 1.2% 1.8%
W (⇥105 N) 2.49 3.96 2.61 5.07
e 0.814 0.984 0.881 1.138

Figure 31: View of the undeflected spans for the Figure 32: View of the deflected spans for the planar
planar and winglet-down configurations. and winglet-down configurations.

most important results of this section. The drag di↵erence relative to the planar wing for each value of

is indicated by D. The inviscid span efficiency factor e = L2 /(Dinviscid b2 q1 ⇡) is also provided, where L

is the lift, Dinviscid is the inviscid drag, b is the wingspan, and q1 is the freestream dynamic pressure. The

inviscid drag does not include the friction drag estimate. It is noteworthy that although the span efficiency

factor for the wingletted wing with = 1.0 is significantly higher than the planar counterpart, it does not

translate to a total drag reduction of the same magnitude.

We saw in Section III.B that, from a purely aerodynamic standpoint, the winglet-down configuration is

able to push the tip vortex further away from the wing than the winglet-up counterpart. When aerostructural

optimization is considered, the winglet-down design pushes the tip vortex even further away from the wing

due to the outboard deflection of the winglet-down feature. Figures 31 and 32 demonstrate how the winglet-

down configuration has a slightly higher span at the deflected state under the aerodynamic loads. This e↵ect

is subtle and is only captured if the structural deflections are taken into account. It is important to note that

although the increase in span as a result of the deflections is small, its e↵ect is larger due to the quadratic

relation between span and induced drag [28]. This explains why the aerostructural optimizer produces a

winglet-down given the freedom to do so. Di↵erences in the projected span values at the deflected state

22 of 37
under the cruise loads relative to the optimized planar cases are indicated by b⇤ in Table 3.

There is an important point about the aerostructural optimization results that we must mention. Al-

though the fully-coupled aerostructural optimization results include more physics than the purely aerody-

namic optimization counterparts, we are still largely ignoring the implication of additional critical structural

load cases, buckling, and flutter. Consequently, the performance improvements obtained from these results

still represent a best-case scenario in the context of practical aircraft design. For instance, adding a winglet

may lead to a reduction of the flutter speed by reducing the torsional rigidity of the wing [41, 42]. Similarly,

a buckling constraint at the wing-winglet junction of a winglet configuration may overshadow the potential

aerodynamic benefit. As a result, if the wingletted wings do not produce a considerable benefit in the ab-

sence of such prohibitive constraints, then one can argue that they would not provide greater improvements

if we were to take these additional considerations into account. Thus, these results could overestimate the

benefits of winglets, but are unlikely to underestimate them.

We have so far limited our discussion of the aerostructural optimization cases to the winglet-down con-

figuration due to the fact that the optimizer seems to favor it over the planar and winglet-up designs.

However, there is an important yet subtle point about the winglet-up configuration. The inboard deflection

of the winglet-up would bring the tip vortex even closer to the wing relative to the aerodynamically optimal

counterpart. This can potentially make the winglet-up configuration less desirable from an aerostructural

standpoint. For this reason, it will be interesting to see what the optimizer chooses to do given the freedom

to either keep the wing planar or create a winglet-up configuration. Section III.D explores this in more

detail.

III.D. Further Investigation Using Aerostructural Optimization

Results from Section III.C indicate that when the aerostructural optimizer is given the freedom to choose

the optimal winglet shape at the tip of the wing, it produces a winglet-down configuration. An alternative

optimization strategy is to limit the choice of the cant angle in such a way that the optimizer is forced to either

produce a nonplanar feature or keep the wing planar for both winglet-up and winglet-down configurations.

If the optimizer does not create a winglet in either case, it could be argued that the planar wing is optimal.

In this section, we are particularly interested to see if the aerostructural optimizer creates a winglet-up

configuration from an initially planar wing for any value of , especially in light of the fact that this

configuration may bring the tip vortex even closer to the wing at the deflected state relative to the purely

aerodynamic optimization case.

For the purpose of this investigation we choose the Boeing 737-900 as the baseline aircraft, which is

heavier and has a longer fuselage than the one used in the previous sections. This is done in part to gain a

23 of 37
 Table 4: Winglet-up and winglet-down optimization design variables
Design Variable Quantity
Sweep Angle 3
Twist Angle 2
Cant Angle 1
Section Shape 84
Angle of Attack 2
Skin Thickness 60
Spar Thickness 66
Rib Thickness 30
Total 248

better understanding of whether our conclusions are sensitive to the choice of the baseline design. The larger

weight leads to a higher lift-induced drag in cruise and could a↵ect the tradeo↵ between weight and drag.

The planform of the initial wing and the corresponding structural wingbox are shown in Figure 15. All of

the winglet configurations that are considered have the same undeflected span as the baseline wing and, as

a result, have the same planform area as the baseline wing. The projected span of the wing, however, may

change due to the structural deflection of the wing. We consider the same cruise and 2.5g load conditions

as before. The objective function is the same as the one used in Section III.C and is shown in Equation 1.

Three values of have been chosen: 0.5, 0.75, and 1.0. The optimizations are conducted in a single stage

using the grid with 653, 184 nodes in 112 blocks. Based on our investigations in the previous sections, there

is practical justification for ending the optimization cases after approximately 150 design iteration as the

majority of the reduction in the merit function will be achieved.

The nonlinear constraints used in this investigation are the same as the ones listed in Table 1. However,

the total weight of the aircraft is assumed to be equal to the computed weight of the wing plus a fixed weight

of 785, 000 N. This fixed weight is estimated based on the maximum takeo↵ weight of a Boeing 737-900

excluding the approximate wing weight. The structural sizing methodology is the same as before. Table 4

provides a list of the design variables used for winglet-up and winglet-down optimization cases. These cases

have a total of 248 design variables. The geometric parameterization is slightly di↵erent from the previous

sections in that the sectional shape is controlled at fewer spanwise stations to reduce the number of design

variables. Figure 33 shows the geometric design variables that control the shape of the wing and winglet.

The optimizer has the freedom to change the sectional shape of the wing by manipulating the z-coordinates

of the control points at 6 spanwise stations. The airfoil shapes are interpolated between these stations using

the remaining control points. For a winglet-up configuration to form from an initially planar wing, the cant

angle at the wingtip is allowed to vary between 0 and +90 . Similarly, for a winglet-down configuration, the

optimizer is free to vary the cant angle between 0 and 90 . In both cases, the optimizer is free to keep the

24 of 37
Figure 33: Geometric parameterization and design Figure 34: The optimized winglet-down configura-
variables. tion with = 0.5 along with the initial wing.

0
10
Feasibility
Optimality
1.02

1.00 10
−2

0.98
Merit Function

−4
0.96 10

0.94

−6
0.92 10

0.90
−8
0.88 10
0 20 40 60 80 100 120 140 160 0 20 40 60 80 100 120 140 160
Design Iterations Design Iterations

Figure 35: Merit function convergence history for Figure 36: Feasibility and optimality convergence
the winglet-down configuration with = 1.0. histories for the winglet-down configuration with
= 1.0.

wing planar. The maximum height of the nonplanar feature at the wingtip is constrained to 6% of the wing

span. For the planar wing optimization case, the wingtip cant angle design variable is inactive and hence the

optimizer does not have the freedom to create a nonplanar feature at the tip of the wing. As an example,

Figure 34 shows the optimized winglet-down configuration with = 0.5 along with the initial wing.

Figure 35 shows the merit function history for the winglet-down configuration with = 1.0. Figure 36

shows the corresponding convergence histories for optimality and feasibility. These aerostructural optimiza-

tion cases tend to converge rather slowly [43]. Nevertheless, Figure 35 indicates that we have obtained the

majority of the merit function improvement. The optimizer completed 174 design iterations in 6 days of

walltime on 240 processors. These general convergence trends hold true for the rest of the optimization cases

as well.

Figure 37 shows plots of the pressure coefficient on the top surface for the initial geometry evaluated on

the fine mesh. Plots of the pressure coefficient for the optimized winglet-down configuration with = 0.75

are shown in Figure 38. The initial geometry does not satisfy the nonlinear lift constraint using the initial

25 of 37
 Figure 37: Contours of pressure coefficient on the upper and lower surfaces for the initial geometry.

Figure 38: Pressure coefficient contours for the optimized winglet-down configuration with = 0.75.

Figure 39: Skin thickness values in millimeters for the winglet-down configuration with = 0.75.

value of the angle of attack and leads to a shock that is present over a large portion of the upper surface.

The optimizer has eliminated this shock on the upper surface of the optimized design while satisfying all
6
nonlinear constraints to a tolerance of 10 .

The optimized skin thickness values for the winglet-down configuration are shown in Figure 39. Note

that the aerostructural optimization has thickened the skin inboard. This is somewhat expected because the

failure criterion at the 2.5g load condition tends to be closer to the critical value at the root of the wing and

near the crank, as shown in Figure 40. The optimized thickness distributions for the ribs and spars follow

26 of 37
Figure 40: Plot of the von Mises failure criterion at the 2.5g load condition for the winglet-down configuration
with = 0.75. A value of unity indicates structural failure.

3.5 3.0
Cruise Cruise
Cruise Elliptical Cruise Elliptical
2.5g Load 2.5g Load
3.0 2.5g Elliptical 2.5g Elliptical
2.5

2.5

2.0
2.0

Lift
Lift

1.5 1.5

1.0
1.0

0.5

0.5
0.0

−0.5 0.0
0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16
y [m] y [m]
Figure 41: Spanwise lift distributions at cruise Figure 42: Spanwise lift distributions at cruise
and 2.5g load conditions for the winglet-down con- and 2.5g load conditions for the winglet-down con-
figuration with = 0.5. figuration with = 1.0.

a similar pattern. Although we have not considered all of the necessary critical load conditions that are

required in the context of wing design, this result shows that we have captured at least some of the correct

trends dictated by the structural sizing of the wing.

The weight of the baseline aircraft considered in this section is significantly larger than the one in Section

III.C. In order to show that the optimizer can recover the fundamental tradeo↵ between weight and drag

correctly regardless of the baseline maximum takeo↵ weight, we examine the optimized lift distributions as

well as the planform shapes. Figure 41 shows the spanwise lift distributions for the cruise and 2.5g load

conditions corresponding to the winglet-down configuration with = 0.5. These have been normalized

with respect to the lift at the root for an elliptical lift distribution at cruise. The optimizer is aeroelastically

tailoring the wing in order to significantly reduce the tip loading at the 2.5g load condition while maintaining

a lift distribution closer to elliptical for the wing at the cruise condition. Figure 42 shows the spanwise lift

distributions for the winglet-down configuration with = 1.0. In this case, the cruise lift distribution more

closely follows the elliptical lift distribution than the = 0.5 case. Furthermore, the tip loading at the

2.5g load condition is greater in comparison to the = 0.5 result. In other words, the optimizer is further

27 of 37
 Figure 43: Top-down view of the optimal winglet-down configurations.

reducing the cruise drag at the cost of increasing the wing weight. Thus, the optimizer has captured the

expected tradeo↵ between weight and drag.

Figure 43 shows the top-down view of optimal designs for the winglet-down configuration. The values

of are also provided. As the value of increases from 0.5 to 1.0, i.e. as we place increasing emphasis on

the drag in the objective function, the sweep angles of the wings also increase. This is the expected trend

because wing sweep reduces wave drag. The same trend exists for the planar configurations.

Figure 44 shows the tradeo↵ curves of the optimal solutions obtained for all of the configurations con-

sidered. The optimized planar wing with = 0.5 has a normalized drag value equal to unity. Similarly,

the optimized planar wing with = 1.0 has a normalized weight value equal to unity. Note that for the

purpose of this plot, the weight values are obtained by adding the calculated wing weight to a fixed weight of

785, 000 N to get an approximate total aircraft weight. Similarly, the drag values are obtained by adding the

calculated inviscid drag based on the Euler equations to a friction drag estimate of the entire aircraft based

on the same methodology used in Section III.B. In this case, we estimate that the friction drag at cruise is

equal to 200 drag counts for the baseline planar configuration due to the longer fuselage length [32, 33]. In

all cases, as we place increasing emphasis on inviscid drag in the objective function, i.e. as we vary from

0.5 to 1.0, the drag decreases at the cost of increasing weight.

Figure 44 shows that the winglet-down configuration can provide a total drag reduction of up to 2% for

the same weight in comparison to the planar wings. Although the reduction in inviscid drag is as high as

8% at the same weight, the increase in the surface area due to the nonplanar feature at the tip reduces the

28 of 37
Figure 44: Tradeo↵ curves of optimal designs for all of the wingletted and planar configurations considered.

Table 5: Summary of the aerostructural optimization results obtained from the fine grid
Planar Winglet-Down
Parameter = 0.50 = 0.75 = 1.00 = 0.50 = 0.75 = 1.00
CL 0.576 0.587 0.596 0.581 0.594 0.603
Total drag (counts) 412 400 374 399 383 364
D 0.0% 0.0% 0.0% -3.2% -4.3% -2.7%
L/D 14.0 14.7 15.6 14.6 15.1 16.6
b⇤ 0.0% 0.0% 0.0% 1.1% 1.3% 1.5%
W (⇥105 N) 2.91 3.44 4.33 3.15 3.94 4.73
e 0.782 0.863 0.981 0.870 0.991 1.143

total drag improvement and the winglet adds weight as well, which increases the induced drag due to the

higher lift coefficient needed. Figure 44 does not include the winglet-up resultsa because the optimizer did

not create a winglet-up feature for any of the values even though it did have the geometric freedom to do

so. In other words, when the optimizer has the freedom to create a winglet-up feature, it chooses to keep the

wing planar regardless of the value of . On the other hand, the optimizer produces a winglet-down feature

for each value of when it is given the freedom to do so. This reaffirms our original findings presented in

Section III.C. Table 5 summarizes the results obtained from the fine mesh analysis.

Although the aerostructural optimizer does not create a winglet-up configuration given the freedom to

do so, it is important to eliminate the possibility that the winglet-up performs better than the planar wing

or the winglet-down configuration. In order to address this, we have run an optimization case where the
a The forced winglet-up case will be discussed in the following paragraph.

29 of 37
Figure 45: Contours of x-vorticity behind the trailing edge of the wing for the planar and winglet-down
cases. The dashed line marks the location of the tip vortex for the planar configuration.

cant angle is forced to be equal to +90 . This case is depicted by the red circle in Figure 44. It is clear that

this design is dominated by the planar and winglet-down configurations. This provides additional evidence

that the winglet-up configuration is not better than the planar wing or the winglet-down configuration. The

reader is reminded that even from a purely aerodynamic perspective, the winglet-up configuration reduces

the total drag only by 1.1% in comparison to an optimal planar wing of the same span, as shown in Section

III.B. From an aerostructural perspective, this configuration reduces the projected span at the deflected

state and brings the tip vortex closer to the wing. Furthermore, the addition of the winglet adds weight both

by increasing the structural span and modifying the tip loading. The combination of these e↵ects make the

winglet-up configuration have inferior performance relative to its planar counterpart.

The winglet-down configuration performs better than the planar and winglet-up counterparts for two

reasons. First, this design moves the tip vortex further away from the wing even when the deflections are

not taken into account. Second, the projected span of the winglet-down configuration at the deflected state is

in fact larger than its planar and winglet-up counterparts. The tip vortex moves even further away from the

wing in comparison to the purely aerodynamic optimization case. This highlights the importance of capturing

the deflected shape of the wing under the aerodynamic loads. We have reached the same conclusion even

when using di↵erent geometry parameterization and mesh movement schemes for the purpose of optimization

[44]. In the case of a winglet-up configuration, the structural deflection reduces the projected span and brings

the tip vortex closer to the wing. From an aerostructural perspective, this leads to a higher objective function

for the winglet-up configuration relative to the planar counterpart.

Figure 45 shows that the winglet-down configuration pushes the tip vortex further away from the wing in

the positive y-direction in comparison to the optimized planar wing of the same projected span. This means

that the induced downwash is reduced in the winglet-down case. As a result, the induced drag is lower.

Although the winglet-down configuration appears to be the most competitive design, its relative drag

benefit in comparison to the planar wing is smaller than those reported in past studies based on purely

aerodynamic shape optimization that considered low-speed flying conditions [30]. Furthermore, our results

30 of 37
 Figure 46: Primary structural layout of the ribs and spars used for the purpose of this investigation.

indicate that at least some of the induced drag advantage of the winglet-down configuration is due to the

increased span of the wing at the deflected state. This suggests that perhaps we should include other phases

of a commercial flight profile such as the steady climb, where the ratio of induced to total drag is higher

than in cruise. We will explore this in Section III.E.

III.E. Optimal Winglets for Cruise and Climb Conditions

The main objective of this section is to explore the possibility that wingletted wings may provide a larger

benefit in high-lift, low-speed conditions. The induced drag constitutes up to 80% of the total drag of the

aircraft in high-lift, low-speed conditions such as takeo↵ [1]. This is largely due to the fact that the coefficient

of lift is larger than in cruise. The aerostructural analysis and optimization framework used in this work

does not have the capability to analyze a takeo↵ condition, but we are indeed able to study other low-speed,

high-lift conditions including steady climb.

The optimization problem formulation in terms of the initial geometry and geometric design variables

and parameterization is the same as the one in Section III.C, but the structural layout is slightly simpler,

as shown in Figure 46. There is now an additional representative steady climb condition where the Mach

number is equal to 0.40 at an altitude of 10, 000 ft. The objective function is the sum of the total drag

in cruise and climb conditions. This choice of objective function may not be practical, but is e↵ective in

studying the most important trends that we are interested in.

Table 6 shows the comparison of the total drag reduction from the optimal wingletted wing in comparison

to the planar configuration with the climb condition included in the optimization as a design point. The

nonplanar optimization produces a winglet-down configuration when the optimizer is given the freedom to

31 of 37
 Table 6: Comparison of total drag reduction in Table 7: The ratio of induced to total drag for the
cruise and climb conditions for the optimal winglet- optimal winglet-down configuration in the cruise and
down configuration climb conditions
Flying Condition Total Drag Reduction Flying Condition Dinduced /Dtotal
Cruise 2.5% Cruise 0.35
Climb 3.7% Climb 0.45

Mach Mach
40,000 40,000
0.8

35,000 35,000 0.8

0.7

30,000 30,000 0.7

0.6

25,000 25,000
0.5 0.6
Altitude [ft]

Altitude [ft]
20,000 20,000
0.4
0.5
15,000 0.3 15,000
0.4
10,000 0.2 10,000

0.3
5,000 0.1 5,000

0 0 0.2
0 50 100 150 200 250 300 0 5 10 15 20 25 30 35 40 45 50
Time [min] Time [min]

Figure 47: The flight profile of a B737NG aircraft Figure 48: The flight profile of a B737NG aircraft
from Toronto to Vancouver. from Toronto to Montreal.

do so. It is clear that the nonplanar wing is providing a higher drag reduction in climb than in cruise. This

is due to the fact that the ratio of the induced to total drag is higher in climb than in cruise, as shown in

Table 7. Therefore, the friction drag penalty associated with the increased surface area of the wingletted

design is lower in comparison to cruise. In other words, any induced drag reduction provided by the winglet

has a greater influence on the total drag value.

The most important conclusion from our climb optimization study is that a wingletted wing may provide

a larger total drag reduction in high-lift, low-speed conditions such as steady climb. Thus, whether or not

a particular winglet configuration can help reduce the total drag of an aircraft significantly may depend on

its intended mission profile in addition to the choice of the objective function it is optimized for. To further

illustrate this point, Figure 47 shows the flight profile of a B737NG aircraft from Toronto to Vancouver,

where the distance flown is approximately 3, 600 kmb . It is clear that the time spent during the climb and

descent segments are significantly shorter in comparison to cruise. This suggests that perhaps a wingletted

wing may not provide a significant drag reduction for this flight profile. However, the same aircraft is also

used on shorter routes such as the one shown in Figure 48, where the distance flown is just under 600 km. In

this case, the aircraft spends a significant amount of time in low-speed, high-lift conditions relative to cruise

and as a result, a wingletted wing may be more useful.

b Flight information obtained from publicly available data at https://flightaware.com/

32 of 37
 IV. Conclusions

This paper provides an aerostructural perspective on the potential efficiency gains from wingletted wings.

Three configurations are considered: winglet-up, winglet-down, and planar. A step-by-step approach is

adopted where these configurations are first assessed using purely aerodynamic shape optimization based

on the Euler equations along with a friction drag estimate based on the wetted surface area. Fully-coupled

aerostructural optimization is subsequently performed to take into account the e↵ects of structural weight

and deflection in addition to drag. The most important conclusions are listed below.

1) From a purely aerodynamic perspective, winglets oriented downward produce a larger drag reduction

than winglets oriented upward by moving the tip vortex further away from the wing in the spanwise direction.

2) When fully-coupled aerostructural optimization is considered, the results indicate that the winglet-

down configuration is able to push the tip vortex even further away from the wing relative to the purely

aerodynamic case by increasing the projected span at the deflected state. This configuration reduces the

drag at the same total weight by 2% in comparison to the planar counterpart.

3) Purely aerodynamic shape optimization results show that the winglet-up configuration reduces the total

drag by 1.1% in comparison to an optimized planar wing of the same projected span. From an aerostructural

perspective, however, this configuration brings the tip vortex closer to the wing due to the inboard deflection

of the winglet feature. Furthermore, the addition of the winglet adds weight to the wing by changing the

spanwise loading and increasing the structural span. The combination of these e↵ects make the winglet-up

design inferior to an optimized planar wing of the same span.

4) Our results indicate that the potential benefits of wingletted wings are larger in low-speed, high-lift

conditions such as the steady climb because the ratio of induced to total drag is higher than in cruise.

Acknowledgments

The authors acknowledge Prof. Joaquim R. R. A. Martins of the University of Michigan, Ann Arbor for

sharing his framework for the purpose of constructing our methodology. Furthermore, the authors gratefully

acknowledge the financial support provided by the Ontario Graduate Scholarship program, the National

Science and Engineering Research Council, and the University of Toronto. Computations were performed

on the GPC supercomputer at the SciNet HPC Consortium. SciNet is funded by the Canada Foundation

for Innovation under the auspices of Compute Canada, the Government of Ontario, Ontario Research Fund

- Research Excellence, and the University of Toronto.

33 of 37
 References

1
Kroo, I., “Drag Due to Lift: Concepts for Prediction and Reduction,” Annual Review of Fluid Mechanics,

Vol. 33, No. 1, 2001, pp. 587–617.

2
Whitcomb, R., “A Design Approach and Selected Wind Tunnel Results at High Subsonic Speeds for

Wing-Tip Mounted Winglets,” Tech. Rep. TN-D-8260, NASA, 1976.

3
Heyson, H. H., Riebe, G. D., and Fulton, C. L., “Theoretical Parametric Study of the Relative Advantages

of Winglets and Wing-Tip Extensions,” Tech. Rep. TP-1020, NASA, 1977.

4
Jones, R. T. and Lasinsky, T. A., “E↵ects of Winglets on the Induced Drag of Ideal Wing Shapes,” Tech.

Rep. TM-81230, NASA, 1980.

5
Asai, K., “Theoretical Considerations in in Aerodynamic E↵ectiveness of Winglets,” Journal of Aircraft,

Vol. 22, No. 7, 1985, pp. 635–637.

6
van Dam, C. P., “Induced-Drag Characteristics of Crescent-Moon-Shaped Wings,” Journal of Aircraft,

Vol. 24, No. 2, 1987, pp. 115–119.

7
Smith, S. C. and Kroo, I., “Computation of Induced Drag for Elliptical and Crescent-Shaped Wings,”

Journal of Aircraft, Vol. 30, No. 4, 1993, pp. 446–452.

8
Takenaka, K. and Hatanaka, K., “Multidisciplinary Design Exploration for a Winglet,” Journal of Air-

craft, Vol. 45, No. 5, 2008, pp. 1601–1611.

9
Verstraeten, J. G. and Slingerland, R., “Drag Characteristics for Optimally Span-Loaded Planar, Winglet-

ted, and C Wings,” Journal of Aircraft, Vol. 46, No. 3, 2009, pp. 962–971.

10
Ning, S. A. and Kroo, I., “Multidisciplinary Considerations in the Design of Wings and Wing Tip Devices,”

Journal of Aircraft, Vol. 47, No. 2, 2010, pp. 534–543.

11
Jansen, P., Perez, R. E., and Martins, J. R. R. A., “Aerostructural Optimization of Nonplanar Lifting

Surfaces,” Journal of Aircraft, Vol. 47, No. 5, 2010, pp. 1491–1503.

12
Faye, R., Laprete, R., and Winter, M., “Blended Winglets for Improved Airplance Performance,”

Tech. rep., Boeing Commercial Airplanes, http://www.boeing.com/commercial/aeromagazine/aero_

17/winglets.pdf.

13
Koo, D. and Zingg, D. W., “Progress in Aerodynamic Shape Optimization Based on the Reynolds-

Averaged Navier-Stokes Equations,” The AIAA Science and Technology Forum and Exposition (AIAA

SciTech), San Diego, California, January 2016.

34 of 37
14
Kennedy, G. J. and Martins, J. R. R. A., “A Comparison of Metallic and Composite Aircraft Wings Using

Aerostructural Design Optimization,” 14th AIAA/ISSMO Multidisciplinary Analysis and Optimization

Conference, Indianapolis, IN, sep 2012, AIAA-2012-5475.

15
Kenway, G. K. W., Kennedy, G. J., and Martins, J. R. R. A., “Aerostructural optimization of the Com-

mon Research Model configuration,” 15th AIAA/ISSMO Multidisciplinary Analysis and Optimization

Conference, Atlanta, Georgia, June 2014.

16
Hicken, J. E. and Zingg, D. W., “Parallel Newton-Krylov Solver for the Euler Equations Discretized Using

Simultaneous-Approximation Terms,” AIAA Journal , Vol. 46, No. 11, 2008, pp. 2773–2786.

17
Osusky, M. and Zingg, D. W., “Parallel Newton-Krylov-Schur Solver for the Navier-Stokes Equations

Discretized Using Summation-By-Parts Operators,” AIAA Journal , Vol. 51, No. 12, 2013, pp. 2833–

2851.

18
Kennedy, G. J. and Martins, J. R. R. A., “A parallel finite-element framework for large-scale gradient-

based design optimization of high-performance structures,” Finite Elements in Analysis and Design,

Vol. 87, September 2014, pp. 56–73.

19
Hicken, J. E. and Zingg, D. W., “Aerodynamic Optimization Algorithm with Integrated Geometry Pa-

rameterization and Mesh Movement,” AIAA Journal , Vol. 48, No. 2, 2010, pp. 400–413.

20
Zhang, Z. J., Khosravi, S., and Zingg, D. W., “High-Fidelity Aerostructural Optimization with Integrated

Geometry Parameterization and Mesh Movement,” 56th AIAA/ASCE/AHS/ASC Structures, Structural

Dynamics, and Materials Conference, No. AIAA 2015-1132, Kissimmee, Florida, January 2015.

21
Zhang, Z. J., Khosravi, S., and Zingg, D. W., “High-Fidelity Aerostructural Optimization with Integrated

Geometry Parameterization and Mesh Movement,” Structural and Multidisciplinary Optimization, 2016,

Accepted.

22
Gill, P. E., Murray, W., Michael, and Saunders, M. A., “SNOPT: An SQP Algorithm For Large-Scale

Constrained Optimization,” SIAM Journal on Optimization, Vol. 12, 1997, pp. 979–1006.

23
Perez, R. E., Jansen, P. W., and Martins, J. R. R. A., “pyOpt: A Python-Based Object-Oriented Frame-

work for Nonlinear Constrained Optimization,” Structural and Multidisciplinary Optimization, Vol. 45,

No. 1, 2012, pp. 101–118.

24
Kennedy, G. J. and Martins, J. R. R. A., “Parallel Solution Methods for Aerostructural Analysis and De-

sign Optimization,” 13th AIAA/ISSMO Multidisciplinary Analysis Optimization Conference, No. AIAA-

2010-9308, Fort Worth, Texas, September 2010.

35 of 37
25
Zingg, D. W., Nemec, M., and Pulliam, T. H., “A comparative evaluation of genetic and gradient-based

algorithms applied to aerodynamic optimization,” Revue européenne de mécanique numérique, Vol. 17,

No. 1-2, March 2008, pp. 103–126.

26
Chernukhin, O. and Zingg, D. W., “Multimodality and Global Optimization in Aerodynamic Design,”

AIAA Journal , Vol. 51, No. 6, 2013, pp. 1342–1354.

27
Martins, J. R. R. A., Alonso, J. J., and Reuther, J. J., “A Coupled-Adjoint Sensitivity Analysis Method

for High-Fidelity Aero-Structural Design,” Optimization and Engineering, Vol. 6, 2005, pp. 33–62.

28
Anderson, J. D., Fundamentals of Aerodynamics, McGraw-Hill, 1221 Avenue of the Americas, New York,

NY 10020, United States of America, 4th ed., 2007.

29
“737 Airplane Characteristics for Airport Planning,” Tech. rep., Boeing Commercial Airplanes, September

2013, http://www.boeing.com/assets/pdf/commercial/airports/acaps/737.pdf.

30
Hicken, J. E. and Zingg, D. W., “Induced-Drag Minimization of Nonplanar Geometries Based on the

Euler Equations,” AIAA Journal , Vol. 48, No. 11, 2010, pp. 2564–2575.

31
Gagnon, H. and Zingg, D. W., “Two-Level Free-Form and Axial Deformation for Exploratory Aerody-

namic Shape Optimization,” AIAA Journal , Vol. 53, No. 7, 2015, pp. 2015–2026.

32
Fredericks, W., Costa, K. A. G., Deshpande, N., Moore, M., Miguel, E. S., and Snyder, A., “Aircraft

Conceptual Design Using Vehicle Sketch Pad,” 48th AIAA Aerospace Sciences Meeting Including the New

Horizons Forum and Aerospace Exposition, No. AIAA 2010-658, Orlando, Florida, January 2010.

33
Gur, O., Mason, W. H., and Schetz, J. A., “Full-Configuration Drag Estimation,” Journal of Aircraft,

Vol. 47, No. 4, 2010, pp. 1356–1367.

34
Bourdin, P., “NUMERICAL PREDICTIONS OF WING-TIP EFFECTS ON LIFT-INDUCED DRAG,”

International Council of Aeronautical Sciences (ICAS) Congress, No. 2002-2.2.3, Toronto, Ontario,

September 2002.

35
Kenway, G. K. W., Kennedy, G. J., and Martins, J. R. R. A., “A CAD-Free Approach to High-Fidelity

Aerostructural Optimization,” Proceedings of the 13th AIAA/ISSMO Multidisciplinary Analysis Opti-

mization Conference, Fort Worth,Texas, September 2010, AIAA 2010-9231.

36
Wrenn, G. A., “An Indirect Method for Numerical Optimization Using the Kreisselmeier–Steinhauser

Function,” Tech. Rep. CR-4220, NASA, 1989.

36 of 37
37
Akgün, M. A., Haftka, R. T., Wu, K. C., Walsh, J. L., and Garcelon, J. H., “Efficient Structural Op-

timization for Multiple Load Cases Using Adjoint Sensitivities,” AIAA Journal , Vol. 39, No. 3, 2001,

pp. 511–516.

38
Martins, J. R. R. A., Alonso, J. J., and Reuther, J. J., “High-Fidelity Aerostructural Design Optimization

of a Supersonic Business Jet,” Journal of Aircraft, Vol. 41, No. 3, 2004, pp. 523–530.

39
Kennedy, G. J., Aerostructural Analysis and Design Optimization of Composite Aircraft, Ph.D. thesis,

University of Toronto, September 2012.

40
Osusky, L., Buckley, H., Reist, T., and Zingg, D. W., “Drag Minimization Based on the Navier-Stokes

Equations Using a Newton-Krylov Approach,” AIAA Journal , Vol. 53, No. 6, June 2015, pp. 1555–1577.

41
Cui, P. and Han, J., “Prediction of flutter characteristics for a transport wing with wingtip devices,”

Aerospace Science and Technology, Vol. 23, No. 1, 2012, pp. 461–468.

42
Snyder, M. P. and Weisshaar, T. A., “Flutter and Directional Stability of Aircraft with Wing-Tip Fins:

Conflicts and Compromises,” Journal of Aircraft, Vol. 50, No. 2, 2013, pp. 615–625.

43
Kenway, G. K. W., Kennedy, G. J., and Martins, J. R. R. A., “Multipoint High-Fidelity Aerostructural

Optimization of a Transport Aircraft Configuration,” Journal of Aircraft, Vol. 51, 2014, pp. 144–160.

44
Khosravi, S. and Zingg, D. W., “A Numerical Optimization Study on Winglets,” 15th AIAA/ISSMO

Multidisciplinary Analysis and Optimization Conference, No. AIAA 2014-2173, Atlanta, Georgia, June

2014.

37 of 37