Available online at www.sciencedirect.

com

ScienceDirect

Comput. Methods Appl. Mech. Engrg. 305 (2016) 512–536

www.elsevier.com/locate/cma

Low-fidelity 2D isogeometric aeroelastic analysis and optimization

method with application to a morphing airfoil
E. Gillebaart ∗ , R. De Breuker
Department of Aerospace Structures and Computational Mechanics, Faculty of Aerospace Engineering,
Delft University of Technology, The Netherlands

Received 9 November 2015; received in revised form 5 March 2016; accepted 9 March 2016
Available online 17 March 2016

Abstract

Low-fidelity isogeometric aeroelastic analysis has not received much attention since the introduction of the isogeometric
analysis (IGA) concept, while the combination of IGA and the boundary element method in the form of the potential flow
theory shows great potential. This paper presents a two-dimensional low-fidelity aeroelastic analysis and optimization framework
consisting of a closely coupled isogeometric potential flow model and isogeometric curved Timoshenko beam model combined
with a boundary layer model. Application of the framework to the optimization of the landing performance for an active morphing
airfoil demonstrates the potential of the isogeometric aeroelastic framework.
⃝c 2016 Elsevier B.V. All rights reserved.

Keywords: Isogeometric analysis; Aeroelasticity; Morphing airfoil; Optimization

1. Introduction

Isogeometric analysis (IGA) has captured the attention of many researchers in the past decade. The principle of
isogeometric analysis (IGA), introduced in 2005 [1], is to apply basis functions that are typically used in computer
aided design (CAD) to the analysis as well, instead of the conventional polynomial basis functions. The main benefit
of this approach is that the step from CAD geometry to an analysis suitable geometry is largely circumvented, since
the CAD geometry can directly be used in the analysis. This also greatly simplifies mesh refinement and degree
elevation, as the coarsest mesh is already the exact geometry. For an overview of IGA and its application, see the work
by Nguyen et al. [2].
The application of IGA in the field of aeroelasticity and fluid–structure interaction (FSI) has been limited so
far and is usually focused on high-fidelity modeling using, for example, fluid models based on the Navier–Stokes
equations and non-linear structural shell models. Most of the contributions come from Bazilevs and his team. Two
main subjects are FSI of wind turbines and arterial blood flows. The work on wind turbines is based on incompressible
Navier–Stokes and non-linear Kirchhoff–Love shell elements supported by bending strips, using either matching [3] or

∗ Correspondence to: Kluyverweg 1, 2629 HS Delft, The Netherlands.

E-mail address: [email protected] (E. Gillebaart).

http://dx.doi.org/10.1016/j.cma.2016.03.014
0045-7825/⃝ c 2016 Elsevier B.V. All rights reserved.
 E. Gillebaart, R. De Breuker / Comput. Methods Appl. Mech. Engrg. 305 (2016) 512–536 513

non-matching [4,5] discretizations. The non-matching discretization consists of, for example, a T-spline shell model
and a conventional finite element fluid model. The work on arterial blood flows is also based on incompressible
Navier–Stokes, but uses non-linear solids to model the arterial walls [6,7]. The high computational cost related to
these fluid and structural models prohibits their use in optimization routines.
Low-fidelity models, although having a somewhat lower accuracy, can enable the use of extensive optimization in
early design stages. Low-fidelity aeroelastic analysis has been almost untouched to the authors’ knowledge, while, for
example, aerodynamic analysis in the form of the boundary element method (BEM) has been shown to be a suitable
candidate for the implementation of IGA. The work of Politis et al. showed the application of IGA to exterior planar
Neumann potential problems and they demonstrated superior convergence properties compared to a conventional
BEM implementation [8]. Takahashi and Matsumoto demonstrated the application of the fast multipole method to the
same type of problem, reducing the computational cost from quadratic to linear with respect to the number of control
points [9]. One of the most important challenges in BEM, computing the singular integrals, was tackled by Heltai et al.
by desingularizing the boundary integral equation. The application to 3D Stokes flow problems showed the accuracy
of the method [10].
The fact that the BEM only requires a boundary description of the problem creates a perfect match with CAD,
since the output of such software is only a boundary discretization. Furthermore, analyzing the exact geometry is
beneficial for aerodynamic analysis, since small perturbations can have significant effects. Also in other fields it has
been shown that the benefits of IGA work well in a BEM framework. Simpson et al. applied IGA to 2D elastostatic
analysis [11] and later to acoustics, using T-splines instead of non-uniform rational B-splines (NURBS) [12]. The
elastostatic analysis was expanded to multipatch 3D analysis in the work of Lian et al. [13]. Scott et al. used T-spline
instead of NURBS in 3D elastostatic analysis [14].
Another advantage of IGA is that it is very suitable for shape optimization problems, especially in a BEM
framework, as has been shown for structural shape optimization in 2D, using a desingularized BEM formulation,
by Lian et al. [15] and in 3D by Li and Qian [16]. The IGA implementation delivers very accurate shape sensitivities,
resulting in precise optimization results. Furthermore, the discretization can be used directly in optimization, removing
the need for a separate shape parametrization as demonstrated by Cho and Ha [17].
All these benefits of IGA have lead to the idea of implementing the concept into a low-fidelity aeroelastic
framework with the purpose of performing simultaneous shape and stiffness optimization of composite wings during
the preliminary design phase. The performance increase resulting from this simultaneous optimization has been shown
already using high-fidelity aeroelastic frameworks [18], but these are computationally expensive, making them less
attractive for the early design stages.
As a first step, this paper presents a two-dimensional low-fidelity aeroelastic framework based on the IGA concept.
An IGA potential flow solver is closely coupled to an IGA curved Timoshenko beam solver formulated in a global
reference system. Subsequently, a one-way coupling is formed between the aeroelastic model and a boundary
layer model. A gradient-based optimizer is used in combination with sensitivities computed using algorithmic
differentiation. To show the capabilities of the framework, it is applied to the aeroelastic optimization of an active
morphing airfoil.
The paper is built up as follows. In Section 2 a short introduction into IGA is given, and the aerodynamic, structural
and boundary layer models are described. The coupling of the models is also explained. The next section presents
the design case of the morphing airfoil, together with the optimization specifications. The results are presented and
discussed in Section 4 and the paper is concluded in Section 5.

2. Methodology

The low-fidelity aeroelastic framework consists of an isogeometric potential flow aerodynamic model and an
isogeometric curved Timoshenko beam structural model. This section will describe the theory and implementation
of both these models. The coupling between these models and the loads and displacements transfer method will also
be presented. First a short introduction into NURBS and isogeometric analysis is given for completeness.

2.1. NURBS curves and basis functions

In the present work NURBS basis functions are used, because these are the most common in CAD software. A
NURBS curve is fully defined by three items:
514 E. Gillebaart, R. De Breuker / Comput. Methods Appl. Mech. Engrg. 305 (2016) 512–536

0.5

-0.5

-1
0 1 2 3 4 5

Fig. 1. Example of a NURBS curve with control polygon and knot location.

• The control points Pi ∈ Rd , 1 ≤ i ≤ n, where d indicates the dimensionality of the problem.

• The degree p of the basis functions.
• The knot vector Ξ = [ξ1 , ξ2 , . . . , ξn+ p+1 ].
An example of a two-dimensional NURBS curve of degree 3 is shown in Fig. 1. From this example it can be
seen that the control point may or may not be on the curve. Only so-called open knot vectors are used in the present
work, meaning that the first and last knots are repeated p + 1 times. As a consequence the NURBS curve is always
interpolatory at the start and end of the curve. The order of continuity of the rest of the curve can be modified by
repeating the other knots. The order of continuity at a knot is equal to p + 1 − r , where r is the multiplicity of this
knot.
Mathematically a NURBS curve can be represented as follows:
n

C(ξ ) = Ri, p (ξ )Pi (1)
i=1

where C(ξ ) is the vector with Cartesian coordinates of the point described by the parametric point ξ1 ≤ ξ ≤ ξn+ p+1 ,
and Ri, p (ξ ) is the ith rational basis functions of degree p. These basis functions are given by
Ni, p (ξ )wi
Ri, p (ξ ) = n (2)
N j, p (ξ )w j

j=1

where Ni, p (ξ ) is a B-spline basis function of order p, and wi is the weight factor corresponding to the ith control
point. The B-spline basis functions of degree 0 are defined as

1 if ξi ≤ ξ ≤ ξi+1

Na,0 (ξ ) = (3)
0 otherwise
and for higher degrees
ξ − ξi ξi+ p+1 − ξ
Ni, p (ξ ) = Ni, p−1 (ξ ) + Ni+1, p−1 (ξ ). (4)
ξi+ p − ξi ξi+ p+1 − ξi+1
The recursive character results in higher cost for computing the basis functions. However, several efficient algorithms
exist to speed up the computations, of which the Cox–de-Boor algorithm [19] is most popular. This algorithm was
also used in this work.
NURBS basis functions (as well as B-spline basis functions) possess some properties that are favorable for their
implementation into analysis models. The basis functions possess the local support property, meaning that in every
knot span [ξi , ξi+1 ) at most p + 1 non-zero basis functions exist, namely Ni− p, p , . . . , Ni, p . Furthermore, they are
 E. Gillebaart, R. De Breuker / Comput. Methods Appl. Mech. Engrg. 305 (2016) 512–536 515

non-negative and form a partition of unity at each parametric location. In the interior of the knot span all derivatives
of the non-zero basis functions exist and at a knot itself the functions are p − r times continuously differentiable.
The principle of IGA, as was mentioned before, is to use the NURBS basis functions both for describing the
geometry and to approximate the unknowns. For the aerodynamic model the unknowns are the doublet strength, as
will be explained in Section 2.2, and for the structural model the unknowns are the displacements, as will be explained
in Section 2.3. The power in this approach lies in the fact that the geometry and its boundary discretization can be
completely done in CAD software and this geometry can immediately be analyzed in the IGA analysis software.
Even though the exact geometry is provided by the CAD software, the discretization might not be detailed enough
to enable the analysis to provide accurate results. This problem can be easily solved by either knot refinement or
degree elevation, or a combination of both [20]. The computational algorithms to perform such operations on NURBS
functions are well documented in literature [19]. The knot refinement and degree elevation operations have no effect
on the geometry itself, making these operations efficient and simple.

2.2. Aerodynamic model

The aerodynamic model is based on the potential flow model, which is characterized by Laplace’s equation:

∇ 2Φ∗ = 0 (5)
where Φ∗ is the total velocity potential. The boundary conditions for the flow over an airfoil are the flow tangency
condition at the boundary of the airfoil and that the disturbance of the flow should go to zero far away from the airfoil:
∂Φ ∗ (x)
= 0 for x ∈ Γb (6)
∂n
lim ∇Φ = 0 (7)
r →∞

where n is the outward unit normal vector to the boundary of the body Γb , r is the distance from the body, and Φ
is the perturbation velocity potential. This problem is reduced to solving the following boundary integral equation
(BIE) [21]:
∂Φ(x) ∂G(x0 , x) ∂G(x0 , x)
   
C(x0 )Φ(x0 ) = G(x0 , x) − Φ(x) d S(x) − ∆Φ(x) d S(x) (8)
Γb ∂n(x) ∂n(x) Γw ∂n(x)
where x0 is the point under consideration, x the point of integration on the surface of the body and the wake, C the
jump term, Γw the wake surface, and G the fundamental solution as is given in Eq. (9).
 
1 1
G(x0 , x) = log . (9)
2π |x0 − x|
The normal derivative of the fundamental solution is given by
∂G (x0 − x) · n(x)
(x0 , x) = . (10)
∂n 2π |x0 − x|2
Setting the normal derivative of the perturbation potential to zero and assuming the notation C = C(x0 ), Φ0 =
Φ(x0 ), Φ = Φ(x), G = G(x0 , x), S = S(x), n = n(x), gives the following solution:
∂G ∂G
 
CΦ0 = − Φ dS − ∆Φ d S. (11)
Γb ∂n Γw ∂n
When the point x0 approaches point x on the boundary of the body, a singularity will be encountered. To
desingularize Eq. (11), the potential inside the body is investigated. The flow tangency boundary condition in Eq. (6)
implies that the potential inside the body is constant. Furthermore, the normal vector is opposite compared to that of the
external problem and the jump term becomes 1 − C [22]. The following equation thus holds for the interior problem:
∂G

1−C =− d S. (12)
Γb ∂n
516 E. Gillebaart, R. De Breuker / Comput. Methods Appl. Mech. Engrg. 305 (2016) 512–536

This relation is fully dependent on the geometry of the problem. Multiplying the equation with the total potential at
point x0 and subtracting it from Eq. (11) gives
∂G

Φ0 = (Φ − Φ0 ) dS (13)
Γb ∂n
which is no longer singular. Combining Eqs. (11) and (13) and introducing the free stream potential Φ∞ , gives the
equation for the total velocity potential:
∂G ∂G
 
Φ0∗ = − (Φ − Φ0 ) dS − ∆Φ d S + Φ∞ . (14)
Γb ∂n Γw ∂n
For the internal problem, the constant velocity potential is set to zero, which results in the final form of the problem
that needs to be solved to find the velocity potential along the boundary:
∂G ∂G
 
− (Φ − Φ0 ) dS − ∆Φ d S + Φ∞ = 0. (15)
Γb ∂n Γw ∂n
The velocity potential jump in the wake is related to the unknown velocity potential distribution on the body through
the Kutta condition [23].
The geometry is now discretized using a NURBS curve and the unknown doublet strength is approximated using
the same basis functions.
n
 n

x(ξ ) = Ri, p (ξ )Pi ; Φ(ξ ) ≈ Ri, p (ξ )Φi . (16)
i=1 i=1

Using the dot to indicate derivatives with respect to ξ to simplify notation, the Jacobian of the transformation from
the Cartesian coordinate system to the parametric space becomes

J (ξ ) = ẋ12 + ẋ22 . (17)

The non-zero knot spans in the knot vector of the curve can be considered as the elements that build up the geometry.
The collocation method is used to set up a linear system of n equations for the n + 1 unknown velocity potentials. The
final equation is provided by the Kutta condition. The resulting system is as follows:
    1
A1,1 · · · A1,n A1,n+1 Φ1 Φ∞
 .. . .. .
.. . .
..   ..   ... 
 .
   
= (18)

  
 An,1 · · · An,n An,n+1   Φn  Φ n 
∞
−1 0 1 −1 ∆Φ 0
where Ai, j are the aerodynamic influence coefficients. The wake is modeled as a single element with a constant value,
so its influence can be computed analytically. The other, non-singular, influence coefficients are computed numerically
in the parametric domain. The integrals, as given in Eq. (19), are computed using standard Gaussian quadrature. An
adaptive scheme is used for the number of integration points to ensure accurate values for near-singular integrals. The
number of integration points is increased as a collocation point is closer to the element over which the integral has to
be computed.
1 ∂G(ξi , ξ )

Ai, j = R j, p (ξ ) J (ξ )dξ. (19)
0 ∂n
The desingularized contributions are computed by summing up the non-singular contributions and multiply it with the
basis function values in the collocation point.
To find the pressure distribution over the airfoil the flow velocity tangential to the airfoil boundary has to be found.
This is found by taking the derivative of the doublet strength in the tangential direction. In conventional panel methods
the derivative has to be approximated using a finite difference scheme, but since the NURBS curve representing the
 E. Gillebaart, R. De Breuker / Comput. Methods Appl. Mech. Engrg. 305 (2016) 512–536 517

Fig. 2. Comparison of results of the present aerodynamic model and XFOIL for the pressure distribution on a NACA 2412 airfoil at an angle of
attack of 3◦ .

doublet strength is easily differentiable this can be done exactly:

dΦ 1 dΦ
Qt = = . (20)
ds J dξ
The pressure distribution is now found using Eq. (21)
ρ 2 
p=− Q t − Q 2∞ (21)
2
where p is the pressure, Q t the tangential velocity, and Q ∞ the free stream velocity. To find the force acting in the
control points of the aerodynamic mesh, the pressure is integrated over the element e after multiplication with the
corresponding non-zero basis functions in the set χ (e) :

fχ(e)
n = p(ξ )Rχ(e), p (ξ )J (ξ )dξ. (22)
Γe

The aerodynamic model is verified by comparing the results to those obtained using XFOIL [24]. XFOIL is also
based on potential flow theory, but uses the traditional discretization into linear panels. When using the inviscid solver
within XFOIL, the same results as the model described before should be found. Fig. 2 shows the pressure distribution
obtained using the present model and XFOIL for a NURBS approximation of the NACA2412 airfoil at an angle of
attack of 3 degrees. The control points and weights for the coarse mesh of the airfoil can be found in Appendix A.
A good match is observed between the two distributions over the entire chord length. The lift coefficient of the same
airfoil for an angle of attack ranging from −8 to 8◦ is compared in Fig. 3. Again a good match is found. Comparison
to the experimental results from Abbot and Doenhoff (1959) [25] shows that the inviscid results from both the present
model and XFOIL have a slightly higher slope.

2.3. Structural model

The structural model is based on linear curved Timoshenko beams using an isogeometric formulation. Existing
work in this field makes use of the local Frenet coordinate system to formulate the beam model [26,27]. In the present
framework, however, a global coordinate system was required to enable the transfer of loads from the aerodynamic
model to the structural model and the displacements the other way around. For clarity first the formulation in
the Frenet coordinate system is described, followed by the derivation of the formulation in the global coordinate
system.
518 E. Gillebaart, R. De Breuker / Comput. Methods Appl. Mech. Engrg. 305 (2016) 512–536

Fig. 3. Comparison of results of the present aerodynamic model, XFOIL and experimental data for the lift coefficient of a NACA 2412 at various
angles of attack.

2.3.1. Local formulation

In the Frenet coordinate system, defined by the tangential, normal and bi-normal unit vectors (t, n, b), the
membrane, transverse shear and bending strain components of a planar curved beam are defined as

du t (s) u n (s)
εm (s) = − (23)
ds R(s)
u t (s) du n (s)
γs (s) = + − θb (s) (24)
R(s) ds
dθb (s)
χb (s) = (25)
ds
where u t , u n and θb are the mid line tangential and normal displacements, and the cross-section rotation, respectively.
They are all expressed in the curvilinear coordinate s, as shown in Fig. 4.
The total potential energy of the planar Timoshenko beam is the sum of the elastic strain energy U , and the potential
energy of applied external forces V :

Π =U +V (26)
1 L
 
U= E Aεm 2 + G Aγs2 + E I θb2 ds (27)
2 0
 L
V =− (f · u) ds − F · u (28)
0

where E and G are the Young’s modulus and shear modulus respectively, A is the cross-sectional area and I is the
moment of inertia along b. The vectors F, f and u are the concentrated loads, distributed loads, and displacements and
rotations, respectively.
The IGA implementation starts with the representation of the geometry using NURBS curves. For simplicity the
explanation will first be given for a geometry built up out of a single NURBS curve (single patch). The required
modifications for a multi-patch geometry will be described later. The geometry is described by Eq. (29), where the
control points Pi contain the x and y coordinates.
n

x(ξ ) = Ri, p (ξ )Pi . (29)
i=1
 E. Gillebaart, R. De Breuker / Comput. Methods Appl. Mech. Engrg. 305 (2016) 512–536 519

Fig. 4. Curved beam coordinate systems.

Using the dot to indicate derivatives with respect to ξ to simplify notation, the Jacobian of the transformation from
the Cartesian coordinate system to the parametric space becomes

J (ξ ) = ẋ12 + ẋ22 (30)
and the radius of the beam
J3
R(ξ ) = . (31)
|ẋ1 ẍ2 − ẍ1 ẋ2 |
The unknown displacements and rotations are approximated using the same basis functions as those that are used to
describe the geometry:
n
 n
 n

u t (ξ ) ≈ Ri, p (ξ )u it ; u n (ξ ) ≈ Ri, p (ξ )u in ; θb (ξ ) ≈ Ri, p (ξ )θbi . (32)
i=1 i=1 i=1

In the parametric space the strains now become

u̇ t un
εm (ξ ) = − (33)
J R
ut u̇ n
γs (ξ ) = + − θb (34)
R J
θ̇b
χb (ξ ) = . (35)
J
Substituting Eqs. (32) to (35) in total potential energy equation and taking the second derivative of the strain energy
U and the first derivative of the potential energy V with respect to the displacement and rotation control points,
will result in the stiffness matrix, K, and force vector, f, respectively. The integrals are computed using Gaussian
quadrature at the knot span level. The unknown displacements and rotations, u, are found by simply solving Ku = f
and substituting the control point values in Eq. (32).
When the geometry is more complex and cannot be described by a single NURBS patch, multiple patches have to
be created and connected to each other. All the patches are interpolatory at the start and end of the curve, simplifying
the process of connecting them. In the present work a master–slave technique is used to make sure the patches are
connected properly and the redundant slave degrees of freedom are eliminated. The approach also allows connecting
patches through hinges by only matching the displacement degrees of freedom between two patches and leaving the
rotation degree of freedom as it is.

2.3.2. Global formulation

To obtain a structural model in a global frame of reference some straightforward changes have to be made to the
model described in the previous section. The tangential and normal displacements from the local coordinate system
are expressed in the global x and y displacements:
u t (ξ ) = ug (ξ ) · t(ξ ); u n (ξ ) = ug (ξ ) · n(ξ ) (36)
520 E. Gillebaart, R. De Breuker / Comput. Methods Appl. Mech. Engrg. 305 (2016) 512–536

where ug is a vector with x and y displacements u and w, and t and n are the local tangential and normal vectors,
respectively. These unit vectors are derived from the NURBS curve describing the geometry:
 ẋ 
1  
J −t2
t =  ; n= . (37)
ẋ2 t1
J
Substituting Eq. (36) into the strain Eqs. (33)–(35), results in the strains in a global coordinate system:

u̇g · t + ug · ṫ ug · n
εm (ξ ) = − (38)
J R
ug · t u̇g · n + ug · ṅ
γs (ξ ) = + − θb (39)
R J
θ̇b
χb (ξ ) = . (40)
J
The derivatives of the unit tangent and normal vectors are equal to

ẋ1 J˙
 
ẍ1
−  
J J2  −t˙2
ṫ =  ; ṅ = (41)
 ẍ
2 ẋ2 J˙  t˙1
− 2
J J
where the derivative of the Jacobian can be found with Eq. (42):
ẋ1 ẍ2 + ẍ1 ẋ2
J˙ = . (42)
J
The global displacements are again approximated by the NURBS basis functions:
n
 n

u(ξ ) ≈ Ri, p (ξ )u i ; w(ξ ) ≈ Ri, p (ξ )wi . (43)
i=1 i=1

The stiffness matrix and force vector in the global coordinate system can now be found by taking the second derivative
with respect to the displacements and rotations of the strain energy and the first derivative with respect to the
displacements and rotations of the potential energy of the externally applied forces.
Verification of the structural model is performed by comparing its results to some analytical solutions obtained by
Cazzani et al. (2014) [27]. The single patch implementation is tested using the vertically loaded cantilever arch shown
in Fig. 5a. The radius R is 2 m, the Young’s modulus is 80 GPa, and the Poisson’s ratio is 0.2. The cross-section of the
arch is rectangular with the thickness h equal to 0.01 m and the depth equal to 0.2 m. A vertical unit load P is applied
at the tip of the arch. The vertical displacement at the tip of the arch is computed using the described structural model
for an increasing number of elements. The convergence plot in Fig. 5b shows that the displacement converges nicely
to the analytical value.
The multi-patch implementation of the structural model is verified using the vertically loaded incomplete ring
shown in Fig. 6a. The single patch making up the geometry is split at point A to create two separate patches. The
radius R is 2.935′′ , the Young’s modulus 1.05 · 107 psi, and the Poisson’s ratio 0.3. The cross-section of the beam is
rectangular with a thickness h of 0.125′′ and a depth of 1.2′′ .
The vertical displacement in point A is computed for an increasing number of elements and the convergence plot
is shown in Fig. 6b. Again a nice convergence towards the analytical result is observed.

2.4. Coupling

The aerodynamic and structural models are closely coupled, to enable fast computation of the converged solution.
The coupling scheme is based on the conservation of energy, meaning that the virtual work, δW , done by the
 E. Gillebaart, R. De Breuker / Comput. Methods Appl. Mech. Engrg. 305 (2016) 512–536 521

(a) Geometry sketch. (b) Convergence plot.

Fig. 5. Single patch vertically loaded cantilever arch structural verification case.

(a) Geometry sketch. (b) Convergence plot.

Fig. 6. Multi-patch vertically loaded incomplete ring structural verification case.

aerodynamic loads on the aerodynamic mesh should be equal to the work done by the equivalent set of loads on
the structural mesh [28]:

δW = δuT · faero = δuaT · faaero (44)

where δu and δua are the virtual displacements of the structural and aerodynamic control points and faero and faaero
are the aerodynamic forces acting on the structural and aerodynamic control points. To achieve a close coupling, a
coupling matrix H is required which couples the aerodynamic displacements to the structural displacements:

ua = H · u. (45)

Inserting Eq. (45) into Eq. (44) shows that in that case the following relation also holds:

faero = HT · faaero . (46)

522 E. Gillebaart, R. De Breuker / Comput. Methods Appl. Mech. Engrg. 305 (2016) 512–536

(a) Initial coarse mesh.

(b) Aerodynamic mesh derived from initial mesh in Fig. 7a. (c) Structural mesh derived from initial mesh in Fig. 7a.

Fig. 7. Example of the meshes used for the aeroelastic model.

Conventional methods for finding this coupling matrix, such as radial basis function interpolation, fail in the current
work due to the fact that the control points of the IGA mesh are not always located on the boundary. In the present
work both the structural mesh and aerodynamic mesh originate from a coarse initial mesh, as is illustrated in Fig. 7.
To obtain the required coupling matrix the sensitivity matrices for the structural and aerodynamic mesh with respect
to the coarse mesh are used. These sensitivity matrices describe how the control points in either the aerodynamic or
structural mesh move as the control points in the coarse mesh are moved and can thus be used to link the aerodynamic
displacements, ua , to the structural displacements, u:

dPstruct −1
 
dPaero dPaero
ua = H · u = ·u= · · u. (47)
dPstruct dPcoarse dPcoarse
Both the aerodynamic and structural model are linear, so a single Newton–Raphson iteration gives the converged
displacement field:
Ks − Kas · u = faer
s
 
o,0 + f0
s T
Ka = H · Ka · H (48)
s T
faer o,0 = H · faer o,0
where K are the stiffness matrices and f0 any external force besides the aerodynamic load. The superscript s indicates
that the values are in terms of the structural degrees of freedom and the subscript 0 indicates that the forces are those
acting on the undeformed geometry. The converged aerodynamic forces are computed using the displacement field
obtained by solving Eq. (48):
faero = faer o,0 + Ka · H · u. (49)
From these converged aerodynamic forces the lift coefficient can be determined, by taking into account the angle of
attack.
2.5. Boundary layer model

To gain some insight in the viscous effects, such as drag and flow separation, a boundary layer model is coupled to
the aeroelastic model described in the previous subsections. The input into this boundary layer model is the deformed
geometry obtained from the aeroelastic analysis. A flow-chart of the system is given in Fig. 8.
 E. Gillebaart, R. De Breuker / Comput. Methods Appl. Mech. Engrg. 305 (2016) 512–536 523

Fig. 8. Flowchart of the analysis module.

Fig. 9. Comparison of results of the present boundary layer model and XFOIL for the drag coefficient of a NACA2412 airfoil for different angles
of attack at a Reynolds number of one million.

The deformed geometry is used to obtain the tangential velocity distribution along the boundary of the airfoil.
Using a boundary layer model based on the one described by Moran [29], which is based upon several semi-empirical
methods, the growth of the boundary layer along the airfoil is computed. The laminar part of the boundary layer
is described by Thwaites’ method and the turbulent part by Head’s method. The transition between the two different
flow states is determined using Michel’s method. The tangential velocity distribution is discretized into small elements
along the boundary and using these methods the development of the boundary layer is computed step-by-step from
the front stagnation point to the trailing edge.
The main modifications compared to the original implementation are made in the transition and separation steps. To
be able to use the drag coefficient in a gradient-based optimization process the response surface should be continuous.
Therefore, instead of only determining the element in which transition or separation occurs and assuming that the
full element is turbulent or separated, an interpolation is done to find the location of transition or separation more
accurately. This prevents any jumps in the response when the location switches from one element to the next.
The profile drag coefficient is subsequently determined using the Squire–Young formula, which takes into account
the momentum thickness and displacement thickness at the upper and lower side of the trailing edge of the airfoil [30].
For verification of the boundary layer model, its results are compared to those obtained using XFOIL. The drag
coefficient of a NACA2412 airfoil at an angle of attack ranging from 0◦ to 10◦ at a Reynolds number of one million
is computed using the presented boundary layer model and XFOIL. The results are shown in Fig. 9.
For this range of interest the results match quite well, especially when realizing that whereas XFOIL makes use
of a two-way coupling of the aerodynamic and boundary layer model and a more sophisticated drag prediction, the
present framework only uses a one-way coupling in combination with the Squire–Young formula. The deviation at
higher angles of attack is probably caused by flow separation, as the results from the Squire–Young formula become
inaccurate as soon as separation takes place. Most important is that the trend matches the XFOIL results, making the
present implementation suitable for optimization.
524 E. Gillebaart, R. De Breuker / Comput. Methods Appl. Mech. Engrg. 305 (2016) 512–536

Table 1
Flight conditions for UAV in cruise and landing flight.

Cruise Landing
Speed [km/h] 110 56.02
Altitude [m] 304.8 304.8
Angle of attack [deg] – 6.373
Lift coefficient [–] 0.2425 –

Fig. 10. Geometry and internal structure of the morphing airfoil.

3. Morphing airfoil optimization

The aeroelastic framework described in the previous section is applied to the optimization of an active morphing
airfoil. The goal is to morph the baseline airfoil shape, designed for optimal performance during cruise flight, into
an airfoil shape more suitable for landing conditions. This means that the lift coefficient of the airfoil at a certain
angle of attack needs to be maximized resulting in a reduced minimum landing speed and thus increased landing
performance. In Section 3.1, a description is given for the design case, including the unmanned aerial vehicle (UAV)
under consideration and the flight conditions of interest to the optimization problem. Section 3.2 covers the setup of
the optimization problem. The objective, constraints and design variables will be presented.

3.1. Design case

The UAV considered for the optimization has a mass of 25 kg and its wings are rectangular with the NACA2412
airfoil. The wingspan is 3.0 m and the chord length is 0.6 m. The cruise and landing flight conditions are listed in
Table 1.
The internal structure of the airfoil consists of a stiff wing box in the leading edge, with a single spar at 25.00% of
the chord. The actuators are located at 43.75%, 62.50% and 81.25% of the chord and have a vertical orientation in the
undeformed NACA2412 airfoil. The actuators are connected to the skin through hinges. Throughout all the analyses
and optimization runs the front spar of the airfoil is fully clamped. An illustration of the structure is shown in Fig. 10.

3.2. Optimization setup

The main objective of the optimization is to maximize the lift coefficient in landing flight conditions, while
the NACA2412 geometry is maintained during other flight conditions. Without taking into account viscous effect,
however, this would lead to infeasible results where flow separation would cause bad performance in landing
conditions. The viscous effects are introduced into the optimization by including the drag coefficient into the objective
function. Instead of simply optimizing the lift coefficient, the ratio of the lift and drag coefficient is optimized. The
 E. Gillebaart, R. De Breuker / Comput. Methods Appl. Mech. Engrg. 305 (2016) 512–536 525

Table 2
Material properties of carbon fiber AS4/8773.

AS4/8773
E1 [N/m2 ] 1.198 × 1011
E2 [N/m2 ] 9.08 × 109
G 12 [N/m2 ] 5.29 × 109
ν [–] 0.32
t ply [m] 1.83 × 10−4

emphasis is put on the lift coefficient by raising it to a certain power. In the present work a power of 1.5 is used, which
leads to realistic results where the influence of the lift coefficient is still dominating. This approach requires a two-step
optimization to avoid conservative results due to local minima present in the objective function, as is demonstrated in
the next section.
Two separate analysis runs are done. One at landing conditions with the actuators enabled and one at 2.5 g flight
conditions with the actuators locked. Additional constraints are added to ensure a feasible design is found:
• Strain constraints are used for the airfoil skin in the morphing section. The IGA framework allows easy strain
evaluations at any point on the geometry. In the present work the strain at 100 evenly spaced points were taken into
account as constraints. The maximum allowed strain is 4000 microstrain.
• Constraints on the maximum actuation forces and maximum actuator strokes are implemented to avoid infeasible
actuator requirements. The maximum value for the actuation force is set to 500 N and the actuator stroke may be
at most 40% of the initial length.
• When lamination parameters are included as design variables a set of constraint equations is implemented to ensure
feasible laminate designs.
Besides the lamination parameters, the thickness and actuation forces are also included as design variables. The
lamination parameters, however, deviate from the standard lamination parameters, because the current framework
only covers two-dimensional problems. A modified set of lamination parameters is used to scale the membrane and
bending stiffness of the laminate [31]:

wh 3 Er e f
EI = α (50)
12
E A = βwh Er e f (51)
where h and w are the thickness and width of the laminate, and Er e f is the reference Young’s modulus of the material.
In the current work the modulus in the fiber direction of the material used in the optimization, carbon fiber AS4/8773,
was selected as the reference Young’s modulus. The specifications of the material are given in Table 2.
A set of boundaries for the feasible design space was found by simply checking many different laminates and take
the outer boundaries of the dataset. For unidirectional pre-impregnated carbon fiber in an epoxy material AS4/8773
the bounding box is created by the following equations:

−βi + 0.474αi + 0.0368 ≤ 0, for i = 1 . . . Np

βi − 1.3013αi + 0.5656 ≤ 0, for i = 1 . . . Np
βi − 9.4808αi + 8.8408 ≤ 0, for i = 1 . . . Np
(52)
−βi + 0.4694αi + 0.5306 ≤ 0, for i = 1 . . . Np
−βi + 1.2842αi + 0.2476 ≤ 0, for i = 1 . . . Np
βi − 8.9981αi + 0.5595 ≤ 0, for i = 1 . . . N p.
The design space of a laminate with fewer layers is conservative with respect to the design spaces of a laminate with
more layers, so a minimum thickness can be chosen and the design space for this thickness can be safely used for
thicker lamina as well.
The globally convergent method of moving asymptotes (GCMMA) [32] is used to perform the gradient-based
optimization. The objective, constraints and design variables are all normalized or scaled such that their magnitudes are
similar. The direct sensitivities are computed using algorithmic differentiation in all three models, ensuring efficient
526 E. Gillebaart, R. De Breuker / Comput. Methods Appl. Mech. Engrg. 305 (2016) 512–536

Table 3
Settings for the four optimization cases.

# Skin thickness Actuation forces Lamination parameters

1 Baseline 3.66 mm Free a = b = 0.5
2 Thickness Free Free a = b = 0.5
3 Laminate 3.66 mm Free Free
4 Thickness + laminate Free Free Free

computation and accurate sensitivities [33]. The Karush–Kuhn–Tucker optimality conditions are used as convergence
criterion.
Four different optimization cases are studied in the present work.
• First of all, a baseline optimization is done where only the actuation forces are included as design variables. The
thickness for the baseline is chosen such that the cruise geometry can still be closely maintained during the 2.5 g
load case and is equal to 20 plies, or 3.66 mm. The laminate is set to a quasi-isotropic one by setting the lamination
parameters to 0.5.
• The second optimization also includes the thickness of the skin as design variable besides the actuation forces.
• The third case uses the actuation forces and lamination parameters to obtain an optimal result.
• The final optimization includes all the design variables and should theoretically give the best result.
Table 3 gives an overview of the settings per optimization case.

4. Results

As was mentioned in the previous section, the drag coefficient was included in the objective function to ensure
a feasible result without flow separation. To show the influence of the objective function formulation on the final
morphed airfoil, both single step and double step optimization approaches will be discussed. In the first subsection the
single step optimization approaches, using the inverse of the lift coefficient or the fraction of drag and lift coefficient
as objective function, are presented. Subsequently, the final two-step optimization formulation result is shown and the
results for the four optimization cases are presented and discussed.

4.1. Single step optimization

The single step optimization approaches are only demonstrated for the fourth optimization case, which includes
both thickness and lamination parameters as design variables. The effect of the objective function will be most
pronounced for this case. The results from the optimization based on only the lift coefficient is presented first.
Subsequently, the results from the optimization using both drag and lift coefficient are presented.

4.1.1. 1/Cl minimization

In Fig. 11 the deformed airfoil is shown. At the trailing edge a large deformation can be seen. This deformation
causes the flow to separate, which is unfavorable during any flight condition. The lift coefficient for the deformed
airfoil is 1.56 and the drag coefficient is equal to 0.0181.
The solution of this optimization approach is insensitive with respect to the initial design vector. In Fig. 12 the
optimization history of the objective function is shown for three different starting points. The solid line represents the
starting point as mentioned in the previous section, with the design variables in the middle of their ranges. The dashed
and dashed–dotted lines result from starting close at the maximum and minimum of the variable ranges, respectively.
The figure shows that all optimizations converge to the same objective function value and on inspection of the resulting
geometries it can indeed be concluded that the same solution has been found in all three cases.

4.1.2. Cd /Cl1.5 minimization

Taking the objective function equal to Cd /Cl1.5 results in a much more conservative solution. The increased
complexity of the objective functions introduces more local minima, so the chances of the optimizer ending up in
one of these minima is increased. The resulting deformation in Fig. 13 indeed shows that the deformation is less
pronounced compared to Fig. 11. A lift coefficient of 1.148 and a drag coefficient of 0.0106 is found for this airfoil.
 E. Gillebaart, R. De Breuker / Comput. Methods Appl. Mech. Engrg. 305 (2016) 512–536 527

Fig. 11. Deformed airfoil for lift coefficient maximization for optimization case 4.

Fig. 12. Objective function optimization history for three different initial design variable sets for optimization case 4.

Fig. 13. Deformed airfoil for Cl1.5 /Cd maximization for optimization case 4.
528 E. Gillebaart, R. De Breuker / Comput. Methods Appl. Mech. Engrg. 305 (2016) 512–536

Table 4
Objective function and lift and drag coefficient for the NACA2412 and the four optimization cases.

Objective function [·10−3 ] Lift coefficient [–] Drag coefficient [–]

NACA2412 10 1.028 0.0105
Baseline 9.58 1.074 0.0107
Thickness 9.54 1.318 0.0144
Laminate 9.46 1.259 0.0134
Thickness + laminate 8.80 1.459 0.0155

(a) Step 1. (b) Step 2.

Fig. 14. Objective function optimization histories for the two-step optimization for the four cases normalized with respect to the undeformed
NACA2412 performance.

4.2. Two-step optimization results

Neither of the morphing airfoils found in the previous two subsections satisfy the requirements. The first airfoil
experiences too much deformation, causing flow separation. The second airfoil is still far away from the possible
increase in lift as demonstrated by the first airfoil. From these results the idea was formulated to create a hybrid
two-step optimization approach. The two steps are as follows:
1. The maximum possible lift coefficient is found by using the inverse of the lift coefficient as objective function.
2. From that point in the design space, the next optimization step is started with the drag coefficient included in the
objective function and the lift coefficient raised to the power 1.5.
The benefit of this approach is that it will be insensitive to the initial design vector, as was shown in Section 4.1.1.
The second step is more sensitive to the starting design, but as this point is always the same this is not a problem.
The optimizer will thus find a solution close to the solution with the maximum lift coefficient, but a little bit more
conservative to avoid flow separation.
This two-step optimization approach is used for the four optimization cases described in previous section. The
results are presented in the remainder of this subsection. A summary of the optimization results is given in Table 4. An
increase in the lift and drag coefficient is observed for all designs compared to the NACA2412 airfoil. The optimization
histories for the objective function in the first and second optimization step are collected in Fig. 14.
In the first step the objective function converges faster as less design variables are involved. In the second step only
the case with both the thickness and lamination parameters included changes significantly. The other cases were not
able to achieve a solution with flow separation in the first optimization step due to the bounds on the design variables.
The minimum in the second step thus coincided with that of the first step. The combined effect of the thickness and
lamination parameters enables a more extreme deformation and thus results in flow separation and a large increase in
drag in the first step.
 E. Gillebaart, R. De Breuker / Comput. Methods Appl. Mech. Engrg. 305 (2016) 512–536 529

Fig. 15. Baseline deformed and undeformed shape in landing flight conditions.

In the following subsections more details will be given per design. First the baseline result is presented to
understand what the performance of a non-optimized structure would be. The results of the other optimization cases
are compared with the baseline, but also with each other. Finally, the absence of flow separation is investigated.

4.2.1. Baseline
The undeformed NACA2412 airfoil in landing conditions has a lift coefficient of 1.028 and a drag coefficient of
0.0105, resulting in a value of 0.01 for the objective function. For the baseline configuration only the actuation forces
are optimized, as was explained before.
The resulting deformed shape with the constant skin thickness and quasi-isotropic layup is shown in Fig. 15,
together with the undeformed shape and actuator connection points.
It is visible that the thickness of the trailing edge is decreased and the camber is slightly increased, which indeed
results in an increase in lift coefficient. The objective function for the baseline case is equal to 9.58 · 10−3 , which
translates into a lift coefficient of 1.074 and a drag coefficient of 0.0107. A slight increase in both lift and drag is thus
achieved compared to the NACA2412 airfoil.
Because the thickness and lay-up is uniform throughout the skin, the deformation is almost symmetric. The
actuation force in this case limits the deformation. All three actuation force design variables are hitting their upper
limit of 500 N, preventing the optimizer to move towards more extreme camber lines. None of the other constraints
are active. The cruise shape is maintained during the 2.5 g load case. The strain in the skin is shown in Fig. 16, which
shows that it does not reach the limit anywhere. Similar as the deformation, the strain is also nearly symmetric on the
upper and lower surface.

4.2.2. Skin thickness optimization

Fig. 17a shows the deformed shape resulting from the skin thickness optimization. The actual skin thickness
distribution is shown in Fig. 17b, where also the skin thickness of the baseline structure is shown for comparison.
The variation in skin thickness allows the optimizer to introduce more camber compared to the baseline result. This
results in an increased lift coefficient of 1.318, which is significantly higher than that of the NACA2412 and baseline
airfoils.
The thickness distribution shows that most of the thickness design variables are at their limit values. The thickness
of the skin connected to the spar at the bottom side has the lowest thickness possible to introduce a compliant hinge.
At the top the opposite is observed. The thickness is increased to the maximum value in order to maintain a stiff
connection in this location. This allows for more camber to be introduced in the structure.
The middle and last actuators are at their maximum stroke and force, and the strain approaches the strain limit in
most of the thin sections, as is shown in Fig. 18. The cruise shape is again maintained during the 2.5 g load case.
530 E. Gillebaart, R. De Breuker / Comput. Methods Appl. Mech. Engrg. 305 (2016) 512–536

Fig. 16. Strain in the skin of the baseline configuration.

(a) Skin deformation. (b) Thickness distribution.

Fig. 17. Deformation and design variables resulting from the thickness optimization case.

4.2.3. Laminate optimization

The skin deformation and lamination parameters for the laminate optimization case are shown in Fig. 19. The
deformation is again larger compared to the baseline, but less severe compared to the thickness optimization case. This
is also reflected in the resulting lift coefficient of 1.259, which is in between the baseline and thickness optimization
values.
The lamination parameters both show similar behavior as the thickness distribution in the previous case. The
stiffness is strongly reduced and increased in same sections as where the thickness was reduced and increased in
the previous results. The thickness, however, has a larger influence on the bending stiffness, explaining the larger
displacement for the skin thickness optimization.
Similar as to the previous results, the middle and last actuator reach their maximum stroke and force. Fig. 20 shows
that the strain limit is now reached in three out of four flexible sections. The cruise shape is maintained during the
2.5 g load case.
 E. Gillebaart, R. De Breuker / Comput. Methods Appl. Mech. Engrg. 305 (2016) 512–536 531

Fig. 18. Strain in the skin of the thickness optimized configuration.

4.2.4. Skin thickness and laminate optimization

Fig. 21 shows the deformation and design variable distributions for the optimization case with both the thickness
and lamination parameters included as variables. The increased design freedom results in the highest lift coefficient of
1.459, achieved by the bigger displacement and larger curvature of the trailing edge compared to the previous results.
Contrary to the previous results, which showed very similar trends, this design case shows different behavior near
the trailing edge. In the previous results both the upper and lower skin at the trailing edge were stiff, but in this case
all design variables go the to minimum value on the upper side. When comparing the design variables within the
optimization case, it can be seen that the results are very similar in terms of stiff and compliant sections.
The actuator closest to the trailing edge reaches maximum force and stroke, while the one furthest away from the
trailing edge reaches an actuation force of zero. The middle actuator exerts a force of approximately 215 N. The strain
in the skin close to the trailing edge reaches the limit, as well as the strain at the bottom of the spar. This is illustrated
in Fig. 22.

4.3. Attached flow

The drag coefficient was added to the objective function to make sure the optimized results would be separation
free, as was mentioned before. In Fig. 23 the skin friction coefficient on the upper and lower surface of the last airfoil
is shown, as computed by XFOIL. The curve for the upper surface clearly shows the transition point at around 14%
of the chord, because the friction coefficient rises sharply at that point. On the bottom surface the same phenomenon
can be observed at the transition point at around 69% chord. At the trailing edge the friction coefficient on the upper
surface decreases towards zero, but is still slightly above. When the friction coefficient is equal to or lower than zero
the flow has separated, so this shows that indeed no flow separation is taking place.

5. Conclusions

A two-dimensional low-fidelity isogeometric aeroelastic optimization framework was presented and applied to the
optimization of an active morphing airfoil. The framework consists of an isogeometric BEM potential flow solver and
an isogeometric curved Timoshenko beam solver, which are closely coupled. The coupling matrix is derived from the
sensitivity matrices of the structural and aerodynamic control points with respect to a shared coarse control polygon.
Both solvers make use of NURBS basis functions to describe the geometry and approximate the unknown variables.
A boundary layer model is included in the framework to enable the computation of viscous effects.
The integrals for the aerodynamic model are desingularized, enabling easy calculation of the originally singular
contributions to the aerodynamic influence coefficients. The multi-patch beam model is formulated in a global frame
of reference to simplify the coupling to the aerodynamic model. The patches are connected through a master–slave
532 E. Gillebaart, R. De Breuker / Comput. Methods Appl. Mech. Engrg. 305 (2016) 512–536

(a) Skin deformation.

(b) α distribution. (c) β distribution.

Fig. 19. Deformation and design variables resulting from the laminate optimization case.

technique, enabling both rigid connections and hinged connections. The boundary layer model is based on existing
semi-empirical integral methods improved with interpolation techniques to accurately determine points of flow
transition or separation.
The aeroelastic optimization framework was used to optimize the landing performance of a 25 kg UAV. The airfoil
shape for cruise conditions was set to a NACA2412 and the goal was to maximize the lift coefficient during landing
flight conditions by adding three actuators in the trailing edge of the airfoil and optimizing the skin thickness and
lamination parameters. Four cases were optimized with varying design variables. To ensure feasible results without
significant flow separation the drag coefficient was added to the objective function.
The results showed that higher lift coefficients were achieved, as more design variables were included, as was
expected. The skin thickness was shown to be the most important design variable in the presented design case, because
the structure is predominantly loaded in bending. The inviscid results showed an increase in lift coefficient of up to
42% for the case with both skin thickness and lamination parameter as design variables. Analysis of the airfoil in
Xfoil showed that indeed no flow separation was taking place, caused by the use of the drag coefficient in the objective
function.
 E. Gillebaart, R. De Breuker / Comput. Methods Appl. Mech. Engrg. 305 (2016) 512–536 533

Fig. 20. Strain in the skin of the laminate optimized configuration.

(a) Skin deformation. (b) Thickness distribution.

(c) α distribution. (d) β distribution.

Fig. 21. Deformation and design variables resulting from the thickness and laminate optimization case.
534 E. Gillebaart, R. De Breuker / Comput. Methods Appl. Mech. Engrg. 305 (2016) 512–536

Fig. 22. Strain in the skin of the thickness and laminate optimized configuration.

Fig. 23. Development of skin friction coefficient in the boundary layer on the upper and lower surface of the airfoil with optimized skin thickness
and laminate.

Appendix

See Table 5.

Table 5
The coarse mesh control points and
weights for the NACA2412 airfoil.

x [m] y [m] Weight

0.6 0 1
0.596 −0.0003 1
0.5878 −0.0009 1
0.5738 −0.002 1
0.5564 −0.0033 1
0.5346 −0.0049 1
0.5088 −0.0067 1
0.4793 −0.0088 1
(continued on next page)
 E. Gillebaart, R. De Breuker / Comput. Methods Appl. Mech. Engrg. 305 (2016) 512–536 535

Table 5 (continued)

x [m] y [m] Weight

0.4467 −0.011 1
0.4115 −0.0133 1
0.3743 −0.0157 1
0.3358 −0.0181 1
0.2967 −0.0203 1
0.2575 −0.0222 1
0.2189 −0.0236 1
0.1817 −0.0248 1
0.1465 −0.0255 1
0.1138 −0.0254 1
0.0842 −0.0245 1
0.0583 −0.0226 1
0.0367 −0.0197 1
0.0197 −0.0157 1
0.0081 −0.0109 1
0.0015 −0.0053 1
−0.0003 0 1
0.0005 0.0056 1
0.006 0.0124 1
0.0167 0.0193 1
0.033 0.0263 1
0.0542 0.0327 1
0.0799 0.0382 1
0.1096 0.0427 1
0.1426 0.0458 1
0.1784 0.0475 1
0.2161 0.0477 1
0.2552 0.0463 1
0.295 0.0439 1
0.3346 0.0406 1
0.3736 0.0367 1
0.4111 0.0322 1
0.4465 0.0273 1
0.4793 0.0224 1
0.5089 0.0175 1
0.5348 0.0129 1
0.5566 0.0088 1
0.5739 0.0054 1
0.5879 0.0026 1
0.5961 0.0008 1
0.6 0 1

References

[1] T.J.R. Hughes, J.A. Cottrell, Y. Bazilevs, Isogeometric analysis: CAD, finite elements, NURBS, exact geometry and mesh refinement, Comput.
Methods Appl. Mech. Engrg. 194 (39–41) (2005) 4135–4195. http://dx.doi.org/10.1016/j.cma.2004.10.008.
[2] V.P. Nguyen, C. Anitescu, S.P.A. Bordas, T. Rabczuk, Isogeometric analysis: An overview and computer implementation aspects, Math.
Comput. Simulation 117 (2015) 89–116. http://dx.doi.org/10.1016/j.matcom.2015.05.008.
[3] Y. Bazilevs, M.-C. Hsu, J. Kiendl, R. Wüchner, K.-U. Bletzinger, 3D simulation of wind turbine rotors at full scale. Part II: Fluid–structure
interaction modeling with composite blades, Internat. J. Numer. Methods Fluids 65 (1–3) (2011) 236–253. http://dx.doi.org/10.1002/fld.2454.
[4] Y. Bazilevs, M.C. Hsu, M.A. Scott, Isogeometric fluid–structure interaction analysis with emphasis on non-matching discretizations, and with
application to wind turbines, Comput. Methods Appl. Mech. Engrg. 249–252 (2012) 28–41. http://dx.doi.org/10.1016/j.cma.2012.03.028.
[5] A. Korobenko, M.-C. Hsu, I. Akkerman, J. Tippmann, Y. Bazilevs, Structural mechanics modeling and FSI simulation of wind turbines, Math.
Models Methods Appl. Sci. 23 (02) (2012) 249–272. http://dx.doi.org/10.1016/j.cma.2012.03.028.
[6] Y. Bazilevs, V.M. Calo, Y. Zhang, T.J.R. Hughes, Isogeometric fluid–structure interaction analysis with applications to arterial blood flow,
Comput. Mech. 38 (4–5) (2006) 310–322. http://dx.doi.org/10.1007/s00466-006-0084-3.
536 E. Gillebaart, R. De Breuker / Comput. Methods Appl. Mech. Engrg. 305 (2016) 512–536

[7] Y. Bazilevs, J.R. Gohean, T.J.R. Hughes, R.D. Moser, Y. Zhang, Patient-specific isogeometric fluid–structure interaction analysis of thoracic
aortic blood flow due to implantation of the Jarvik 2000 left ventricular assist device, Comput. Methods Appl. Mech. Engrg. 198 (45–46)
(2009) 3534–3550. http://dx.doi.org/10.1016/j.cma.2009.04.015.
[8] C. Politis, A.I. Ginnis, P.D. Kaklis, K. Belibassakis, C. Feurer, An isogeometric BEM for exterior potential-flow problems in the plane,
in: 2009 SIAM/ACM Joint Conference on Geometric and Physical Modeling, SPM’09, ACM, New York, NY, USA, 2009, pp. 349–354.
http://dx.doi.org/10.1145/1629255.1629302.
[9] T. Takahashi, T. Matsumoto, An application of fast multipole method to isogeometric boundary element method for Laplace equation in two
dimensions, Eng. Anal. Bound. Elem. 36 (12) (2012) 1766–1775. http://dx.doi.org/10.1016/j.enganabound.2012.06.004.
[10] L. Heltai, M. Arroyo, A. DeSimone, Nonsingular isogeometric boundary element method for Stokes flows in 3D, Comput. Methods Appl.
Mech. Engrg. 268 (2014) 514–539. http://dx.doi.org/10.1016/j.cma.2013.09.017.
[11] R.N. Simpson, S.P.A. Bordas, J. Trevelyan, T. Rabczuk, A two-dimensional isogeometric boundary element method for elastostatic analysis,
Comput. Methods Appl. Mech. Engrg. 209–212 (2012) 87–100. http://dx.doi.org/10.1016/j.cma.2011.08.008.
[12] R.N. Simpson, M.A. Scott, M. Taus, D.C. Thomas, H. Lian, Acoustic isogeometric boundary element analysis, Comput. Methods Appl. Mech.
Engrg. 269 (2014) 265–290. http://dx.doi.org/10.1016/j.cma.2013.10.026.
[13] H. Lian, R. Simpson, S. Bordas, Stress analysis without meshing: isogeometric boundary element method, Proc. ICE - Eng. Comput. Mech.
166 (2) (2013) 88–99.
[14] M.A. Scott, R.N. Simpson, J.A. Evans, S. Lipton, S.P.A. Bordas, T.J.R. Hughes, T.W. Sederberg, Isogeometric boundary element analysis
using unstructured T-splines, Comput. Methods Appl. Mech. Engrg. 254 (2013) 197–221. http://dx.doi.org/10.1016/j.cma.2012.11.001.
[15] H. Lian, P. Kerfriden, S.P.A. Bordas, Implementation of regularized isogeometric boundary element methods for gradient-based shape
optimization in two-dimensional linear elasticity, Comput.-Aided Des. (2015) http://dx.doi.org/10.1002/nme.5149.
[16] K. Li, X. Qian, Isogeometric analysis and shape optimization via boundary integral, Computer-Aided Design 43 (11) (2011) 1427–1437.
http://dx.doi.org/10.1016/j.cad.2011.08.031.
[17] S. Cho, S.-H. Ha, Isogeometric shape design optimization: exact geometry and enhanced sensitivity, Struct. Multidiscip. Optim. 38 (1) (2008)
53–70. http://dx.doi.org/10.1007/s00158-008-0266-z.
[18] G.K.W. Kenway, J.R.R.A. Martins, Multipoint high-fidelity aerostructural optimization of a transport aircraft configuration, J. Aircr. 51 (1)
(2014) 144–160. http://dx.doi.org/10.2514/1.C032150.
[19] L. Piegl, W. Tiller, The NURBS Book, Springer Science & Business Media, 2012.
[20] J.A. Cottrell, T.J.R. Hughes, Y. Bazilevs, Isogeometric Analysis: Toward Integration of CAD and FEA, John Wiley & Sons, 2009.
[21] L. Morino, Boundary integral equations in aerodynamics, Appl. Mech. Rev. 46 (8) (1993) 445–466. http://dx.doi.org/10.1115/1.3120373.
[22] E. Klaseboer, C.R. Fernandez, B.C. Khoo, A note on true desingularisation of boundary integral methods for three-dimensional potential
problems, Eng. Anal. Bound. Elem. 33 (6) (2009) 796–801. http://dx.doi.org/10.1016/j.enganabound.2008.12.002.
[23] J. Katz, A. Plotkin, Low-speed Aerodynamics, Cambridge University Press, 2001.
[24] M. Drela, XFOIL: An analysis and design system for low Reynolds number airfoils, in: T.J. Mueller (Ed.), Low Reynolds Number
Aerodynamics, in: Lecture Notes in Engineering, vol. 54, Springer, Berlin, Heidelberg, 1989, pp. 1–12.
[25] I.H. Abbott, A.E.V. Doenhoff, Theory of Wing Sections, Including a Summary of Airfoil Data, Courier Corporation, 1959.
[26] R. Bouclier, T. Elguedj, A. Combescure, Locking free isogeometric formulations of curved thick beams, Comput. Methods Appl. Mech.
Engrg. 245–246 (2012) 144–162. http://dx.doi.org/10.1016/j.cma.2012.06.008.
[27] A. Cazzani, M. Malagù, E. Turco, Isogeometric analysis of plane-curved beams, Mathematics and Mechanics of Solids (2014)
1081286514531265. http://dx.doi.org/10.1177/1081286514531265.
[28] A. Beckert, Coupling fluid (CFD) and structural (FE) models using finite interpolation elements, Aerosp. Sci. Technol. 4 (1) (2000) 13–22.
http://dx.doi.org/10.1016/S1270-9638(00)00111-5.
[29] J. Moran, An Introduction to Theoretical and Computational Aerodynamics, John Wiley & Sons, 1984.
[30] H. Squire, A. Young, Air Ministry. Aeronautical Research Committee, The Calculation of the Profile Drag of Aerofoils, Aeronautical Research
Committee Reports and Memoranda, 1838, ARC Technical Report, HMSO, London, 1937.
[31] G.A.A. Thuwis, Stiffness and layout tailoring of a morphing high-lift system with aeroelastic loads (Ph.D. thesis), Delft University of
Technology, Delft, 2012, June.
[32] K. Svanberg, A class of globally convergent optimization methods based on conservative convex separable approximations, SIAM J. Optim.
12 (2) (2002) 555–573. http://dx.doi.org/10.1137/S1052623499362822.
[33] L.B. Rall, G. Goos, J. Hartmanis, W. Brauer, P. Brinch Hansen, D. Gries, C. Moler, G. Seegmüller, J. Stoer, N. Wirth (Eds.), Automatic
Differentiation: Techniques and Applications, in: Lecture Notes in Computer Science, vol. 120, Springer, Berlin, Heidelberg, 1981.