Optim Eng

DOI 10.1007/s11081-008-9078-7

AESOP—a numerical platform for aerodynamic shape

optimization

Olivier Amoignon

Received: 14 April 2008 / Accepted: 30 December 2008

© Springer Science+Business Media, LLC 2009

Abstract Aerodynamic shape optimization based on Computational Fluid Dynamics

can automatically improve the design of aircraft components. In order to obtain the
best computational efficiency, the adjoint method is applied on the complete map-
ping, from the parameters of design to the evaluation of the cost function or con-
straints. The mapping considered here includes the parameterization, the mesh defor-
mation, the primal-to-dual mesh transformation and the flow equations solved by the
unstructured flow solver Edge distributed by FOI. The numerical platform AESOP
integrates the flow and adjoint flow solver, mesh deformation schemes, algorithms of
shape parameterization and algorithms for gradient-based optimization. The result is
a portable and efficient implementation for large scale aerodynamic shape optimiza-
tion and future applications in multidisciplinary shape optimization. The structure
of the program is outlined and examples of applications are presented. The method
of shape parameterization using Radial Basis Functions is discussed in more details
because it is expected to play a major role in the development of multidisciplinary
optimization.

Keywords Adjoint sensitivity analysis · Finite volumes · Median dual grid · Radial
basis functions

1 Introduction

In aeronautic industry, the prospect of aerodynamic shape optimization is to speed

up the design of aircraft components. The continuous development of Computational
Fluid Dynamics (CFD) plays a major role in the development of this activity because
it improves the reliability of the predicted aerodynamic forces and moments that are

O. Amoignon ()
FOI (Swedish Defence Research Agency), 164 90 Stockholm, Sweden
e-mail: [email protected]
 O. Amoignon

used in problems of shape optimization. However, the accuracy of the CFD predic-
tions around aircrafts comes at a cost that is incomparably higher than for solving,
say, the aircraft structural deformations in the framework of linear elasticity. In aero-
dynamic optimization this cost is multiplied by the number of flow solutions required
for solving the (nonlinear) problem of aerodynamic optimization. This number of
CFD simulations is closely related to the algorithm of optimization, for a given prob-
lem to be solved. To date, the algorithms based on the derivatives of the cost function
and constraints, with respect to the design variables (gradients), are the most efficient,
provided that the derivatives can be calculated accurately and efficiently. The cost of
the classical computation of gradients by finite differences is proportional to the num-
ber of design parameters n, which penalizes this approach when n is larger than the
number of constraints. The finite difference method may also be inaccurate because
it requires small perturbations of the design parameters whereas shape perturbations
smaller than a certain limit may not be resolved by the CFD, thus leading to round-
off errors due to the difference scheme. In an optimal control approach, the gradient
of one function can be calculated by solving once an adjoint of the flow equations.
Regarding the cost of optimization, the optimal control approach, or adjoint sensi-
tivities analysis, has the advantage on the finite difference method for a number of
design parameters larger than the number of constraints. Nonetheless, the accuracy
of the adjoint approach does not rely on a perturbation parameter, which is another
advantage.
The numerical platform AESOP1 has been developed for the realization of
an optimal control approach based on the adjoint flow equations implemented in
Edge (Amoignon and Berggren 2006), FOI’s software for CFD (Eliasson 2001).
Design optimization using an optimal control approach has become popular in the
last two decades in all fields of engineering. For a detailed presentation of techniques
of optimal shape design for systems governed by elliptic equations, the reader may
refer to a book by Pironneau (1984). For an introduction focusing on aerodynamic
shape optimization, the reader may refer to Giles and Pierce (2000).
Natural Laminar Flow design based on the adjoint approach is one of the first
applications that AESOP was developed for (Amoignon et al. 2006). This requires
additional solvers, and adjoint solvers, for the laminar boundary layer equations and
for the stability equations in the boundary layer. The procedure will not be described
here but an example of application is given in Sect. 4.2 and the interested reader can
find a detailed presentation of the method in Amoignon et al. (2006).
The general structure of the optimization loop is presented in Sect. 2. The parame-
terization of shape deformations based on Radial Basis Functions (RBF) is presented
in Sect. 3. It has been recently investigated because it can give rise to a general strat-
egy for the optimization of various shapes in both two and three dimensional appli-
cations. This method is directly inspired from a previous work on RBF-based mesh
deformation (Jakobsson and Amoignon 2007). The examples are presented in Sect. 4.

1 Stands here for AErodynamic Shape OPtimization (Aesop is the name of a legendary Greek fabulist).
AESOP—a numerical platform for aerodynamic shape optimization

2 Gradient-based optimization in AESOP

Concerning optimization based on Computational Fluid Dynamics (CFD), the use

of adjoint equations has become a common approach in recent years (Anderson
and Bonhaus 1999; Baysal and Ghayour 2001; Beux and Dervieux 1991; Bugeda
and Oñate 1999; Burgreen and Baysal 1996; Elliot and Peraire 1996; Enoksson
and Weinerfelt 1999; Jameson and Kim 2003; Mohammadi 1997; Nadarajah 2003;
Reuther et al. 1999; Soemarwoto 1996; Sung and Kwon 2000).
There are however different views on the application of the theory of optimal
control in CFD (Gunzburger 2002). In the so-called continuous approach, the ad-
joint equations and the expression of the gradient are derived from the Partial Dif-
ferential Equations (PDEs) that model the flow and from the exact cost function,
see (Baysal and Ghayour 2001; Enoksson and Weinerfelt 1999; Jameson and Kim
2003; Soemarwoto 1996; Sung and Kwon 2000). The resulting expressions are then
discretized. In the so-called discrete approach, the adjoint equations and gradient
expression are obtained from the discretized flow equations and cost function (An-
derson and Bonhaus 1999; Burgreen and Baysal 1996; Elliot and Peraire 1996;
Nadarajah 2003).
The discrete approach is preferred here because it can provide the exact gradient
of the cost function and constraints being optimized. Otherwise, lack of accuracy can
cause a failure of the optimization algorithm in finding a descent direction. There are
though possible simplifications in the discrete approach, as shown in one example in
paragraph Sect. 2.5 or as discussed in Soto and Löhner (2001), Mohammadi (1999).
To become more familiar with the adjoint approach applied here consider a simpler
problem in the spirit of Giles and Pierce (2000). Let a cost function J be linear with
respect to the vector of state variables w,

J (w) = gT w, (1)

with g ∈ Rm given and w ∈ Rm subject to the state equation

Aw = Na, (2)

where a ∈ Rn is the vector of design variables, A ∈ Rm×m , and N ∈ Rm×n .

Assume that A is nonsingular. The reduced gradient of J , that is, the gradient of
the mapping a → J (w(a)), denoted ∇Ja , may be obtained by solving the sensitivity
equations of the state: given a variation of the control variable δa, a corresponding
variation of the state δw is defined as the solution to the sensitivity equations

A δw = N δa, (3)

which enables us to express the variation of the function J

δJ = gT δw ≡ gT A−1 N δa. (4)

Therefore, solving the sensitivity equations, once for each component of the vector a,
yields the gradient ∇Ja , component by component.
 O. Amoignon

However, rewriting (4) as

δJ = (NT (AT )−1 g)T δa, (5)

reveals that replacing (AT )−1 g in (5) by the adjoint state w∗ , defined as the solution
to
AT w∗ = g, (6)
gives an expression for ∇Ja ,
∇Ja = NT w∗ . (7)
The cost for computing the gradient by expression (7) is one costate solution (6) and
a matrix-vector product (7), instead of m solutions of the sensitivity equations (3)
when expression (4) is used.
The generalization to nonlinear state equations and nonlinear functions J is
straightforward:
– g is the vector of partial derivatives of the function J with respect to the vector of
the state variables w.
– A is the Jacobian matrix of the system of discretized state equations with re-
spect to the state variables. The efficient assembly of the products between the
transpose of this matrix and costate vectors (w∗ ) is crucial for solving efficiently
equation (6). These operations are presented in details in Amoignon and Berggren
(2006) when (3) is the system of equations obtained by linearization of the Euler
equations discretized by a median-dual finite volume scheme (Eliasson 2001).
– N is the Jacobian matrix of the system of discretized state equations with respect to
the parameters of optimization. In our CFD applications (Amoignon and Berggren
2006), this involves:
– The pre-processing of the mesh, which is the nodes-to-edges transformation,
also described as primal-to-dual transformation; the transposed Jacobian maps
gradients with respect to the dual data (the control surfaces attributed to an edge)
to gradients with respect to the nodal data (nodal coordinates).
– The deformation of the mesh, for a given deformation of the shape.
– The parameterization of the shape deformation, for given parameters of opti-
mization.
Section 2.2 presents in more details the operations involved in (7).

2.1 Optimization algorithms

The optimization algorithms implemented so far in the numerical platform (see

Fig. 1) are textbook gradient-based methods supplemented by a line search algorithm
using the Goldstein condition as globalization strategy (Nocedal and Wright 1999):
– conjugate gradient from Polak-Ribière;
– quasi-Newton method Broyden-Fletcher-Goldfarb-Shanno (BFGS) with Hessian
updates;
– exact augmented Lagrangian method.
AESOP—a numerical platform for aerodynamic shape optimization

Fig. 1 Iterative gradient-based

aerodynamic shape optimization
as implemented in the numerical
platform AESOP (“deform.”
stands for “deformation”)

The Lagrangian method handles constraints whereas the conjugate gradient and
quasi-Newton methods are used for unconstrained optimization. Note that the La-
grangian method approaches the solution of constrained problems of optimization
through a sequence of unconstrained optimization problems. The unconstrained prob-
lems are approximately solved using either the conjugate gradient or the quasi-
Newton algorithm above.

2.2 Functions and gradients computation

The following describes the algorithms or equations involved in the computation of

the cost function, or of a constraint, depending on just one flow solution. Such a
function is composed of five mappings:
– Parameterization of shape deformations: ah → dh
It maps a vector of design parameters ah into a vector of nodal displacements.
Consider the mesh node with index i on the shape ∂w . Its coordinates, denoted
x i , are related to the initial shape (∂0w ), through the nodal displacements dh =
{d i }1≤i≤Nw as
x i = x 0i + d i , for i ∈ V(∂w ), (8)
where V(∂w ) denotes the set of indexes of the mesh nodes on ∂w . The mapping
of the design parameters ah = {ai }1≤i≤N into displacements dh is expressed as

Sh (dh , ah ) = 0. (9)
 O. Amoignon

There are many strategies for the parameterization of shapes, or shapes deforma-
tions. Two kinds of parameterizations used here are presented below (see Sect. 3).
– Mesh deformation: dh → Xh
To retain mesh quality, the coordinates of all mesh nodes Xh must be adapted
to the displacements dh of the nodes on the part of the boundary being opti-
mized. This is accomplished here by a mesh deformation algorithm, a smoother
as in Amoignon and Berggren (2006) or a method of interpolation based on RBF
as in Jakobsson and Amoignon (2007):

Mh (Xh , dh ) = 0. (10)

– Re-calculation of the control volumes/surface vectors: Xh → nh

By deforming the mesh we conserve the connectivities between the nodes, but
the control volumes of the dual grid and their surface normals nh (the vector of all
control surface normals) need to be re-computed. Therefore, nh is a function of the
mesh coordinates, which we denote

nh ≡ nh (Xh ) . (11)

General expressions for the surface vectors of the dual mesh control volumes nh
are given in 2D and 3D in Amoignon and Berggren (2006).
– Solution of the discretized Euler equations: nh → wh
As mentioned previously, the program Edge (Eliasson 2001) solves the Euler
equations on a dual mesh. The discretized Euler equation in steady state yields a
system of equations in residual form where the discrete flow state wh (the den-
sity, velocity and pressure at all nodes) is the unknown and the dual mesh data nh
parameterize this system:2

Rh (wh , nh ) = 0 (12)

– Function evaluation: Xh , nh , wh → J
For the sake of the presentation, let us consider a function which minimization
involves a reduction of the drag coefficient (CD ) and penalized changes in lift (CL )
and pitching moment (Cm ):

1  2
J (wh , nh , Xh ) = μD CD (wh , nh ) + μL CL (wh , nh ) − CL0
2
1  2
+ μM Cm (wh , nh , Xh ) − Cm
0
, (13)
2

where μ are positive scalars and the upperscript 0 denotes the values at initial
design. The aerodynamic coefficients (CD , CL , Cm ), also used in the examples

2 The discretized Reynolds Averaged Navier–Stokes equations also depend explicitly on the mesh nodal
coordinates Xh .
AESOP—a numerical platform for aerodynamic shape optimization

(Sect. 4), are defined here as:

 pi ni · d D
CD (wh , nh ) = 1
,
2
i ∈ V(∂w ) 2 ρ∞ v ∞ Sref

 pi ni · d L
CL (wh , nh ) = ,
1 2 (14)
i ∈ V(∂w ) 2 ρ∞ v ∞ Sref
 
 pi d M · x i − O ref. × ni
Cm (wh , nh , Xh ) = 1
,
2
i ∈ V(∂w ) 2 ρ∞ v ∞ Sref Lref

were d D = v ∞ /|v ∞ |, v ∞ is the far-field air velocity, d L is an upward oriented unit

vector orthogonal to d D , d M is a unit vector orthogonal to d D and d L , pi is the
pressure at node i, ni , an element of nh , is the outward-oriented surface normal at
the boundary node i, ρ∞ denotes the far-field air density, Sref a reference surface,
Lref a reference length.
The gradient of J (13) with respect to ah , subject to the constraints (9)–(12), also
called the reduced gradient of J , is denoted ∇Ja . All functions in expressions (9)–
(13) being assumed continuously differentiable3 , the computation of ∇Ja can be for-
mulated as:
– For a given design ah and corresponding mesh Xh (9)–(10), dual grid data nh (11)
and discrete flow solution wh (12), compute the adjoint flow solution w∗h by solving
(see Amoignon and Berggren 2006)
 ∗  T
∂Rh ∂J
w∗h = . (15)
∂wh ∂wh

– Calculate the gradient with respect to the dual grid data ∇Jn and the gradient with
respect to the nodal coordinates ∇JX as follows:
  
∂Rh ∗ ∗ ∂J T
∇Jn = − wh + , (16)
∂nh ∂nh
   
dnh T ∂J T
∇JX = ∇Jn + . (17)
dXh ∂Xh

– Solve the adjoint mesh deformation equation, see (Amoignon and Berggren 2006;
Jakobsson and Amoignon 2007) for details:
 ∗
∂Mh
X∗h = −∇JX . (18)
∂Xh

3 Neglecting the artificial diffusivities used in the solution procedure of the flow equations (Amoignon and
Berggren 2006).
 O. Amoignon

– Calculate the gradient with respect to the shape deformations ∇Jd :

 
∂Mh ∗ ∗
∇Jd = Xh . (19)
∂dh
– Solve the adjoint parameterization equation, details are given below for two exam-
ples of parameterization (Sect. 3):
 
∂Sh ∗ ∗
dh = −∇Jd . (20)
∂dh
– Finally, calculate the reduced gradient with respect to the parameters of design:
 
∂Sh ∗ ∗
∇Ja = dh , (21)
∂ah

where ∗ denotes the adjoint of an operator or the unknown in an adjoint equation.

2.3 Example of geometrical constraint

Problems of shape optimization usually include constraints on the geometry. We give

here the expressions used in the examples (Sect. 4) for computing the thickness of
an airfoil section or its volume. Let us define the thickness of an airfoil: suppose,
for the sake of the presentation, that the leading and trailing edges are on the x-axis
as in Fig. 2, we call f (x, ah ) the distance, measured perpendicularly to the x-axis,
between the upper and lower sides of the airfoil and define the thickness of the airfoil
as the maximum of this function:

T () = max f (x, ah ) . (22)

x0 ≤x≤x1

Applying definition (22) on the discretization of an airfoil, or a section of a wing

is straightforward. However, it is not a differentiable function with respect to the
parameters of design.

Fig. 2 Sketch showing the principle of the streamwise discretization used for computing the thickness of
airfoils or streamwise sections of wings. Here m = 4, |S2u | = 4, |S2l | = 5, but in the applications Sect. 4
m = 20
AESOP—a numerical platform for aerodynamic shape optimization

As an alternative we can approximate the maximum norm by the p-norm (p < ∞):

x1 1
p

f
p = |f (x, ah ) | dx
p
(23)
x0

because:

f
∞ = lim
f
p . (24)
p→∞

The norm (23) is differentiable with respect to the shape even if the shape is dis-
cretized. The integral in (23) is approximated, after discretization of the x-axis in
m intervals (Fig. 2), here using the trapezoidal quadrature:
 1
1 p  p 1 p
m−1 p
Th = hf + hfk + hfm , (25)
2 1 2
k=2

where fk approximates f (xk , ah ):

fk =
uk − l k
, (26)

with uk and l k are the mass centers of the points on the upper or lower side of the
airfoil. Denoting by Sku , respectively, Skl , the set of indices of the nodes in the interval
k on the upper side, respectively, on the lower side, then:
1  1 
uk = xi , and l k = xi . (27)
|Sku | u |Skl |
i∈Sk i∈Skl

For p = 1 expression (23) is the cross-section area and expressions (25)–(27) give an
approximation of this area.

Gradient of the thickness constraint

We detail here the expressions that are necessary in order to calculate explicitly the
gradient of the thickness. Suppose that the coordinates of node i on the airfoil is
displaced by δx i . We derive the first variation of the thickness (25) due to the pertur-
bation δx i :

(1−p) p−1
hT f δf for i ∈ Sku or l with 2 ≤ k ≤ m − 1,
δTh = 1 h (1−p)k p−1 k (28)
2 hTh fk δfk for i ∈ Sku or l with k = 1 or m

and the first perturbation of the local thickness δfk is for a node i on the upper side:
1 uk − l k
δfk = · δx i (29)
|Sku | fk
or for a node i on the lower side:
1 uk − l k
δfk = − · δx i (30)
|Skl | fk
 O. Amoignon

which gives expressions for calculating ∂Th /∂x i for all nodes on the airfoil. Using
relations (20)–(21) gives the gradient of Th with respect to all design variables.

2.4 Aspects of parallel computations

The numerical platform AESOP is parallelized in two ways. Either the loop is par-
allelized so that the flow and adjoint flow solution are computed “sequentially”, but
several at a time. Or, the flow and adjoint flow solutions are obtained on a partitioned
domain, that is solving in parallel one flow (12) or adjoint flow problem (15) at a time.
The first option enables to obtain a speed up even for small problems where there is
no advantage in partitioning the computational domain. The second approach is of
course advantageous for large computational meshes. The computation of ∇JX (16)–
(17) is performed in Edge following the scheme described in Amoignon and Berggren
(2006), which is also parallelized.

2.5 Discrete against continuous adjoint sensitivities

The purpose of this example is to show that the discrete approach for calculating
gradients of the cost function or constraints (15)–(21) can sometimes be simplified
without important change in the results of the optimization.
For instance the calculation of gradients requires adjoints of the mesh deforma-
tion (18) (one adjoint per cost function and constraint that depend on mesh defor-
mations). If the shape optimization is based on inviscid flow equations, skipping the
solution of equation (18) may not affect the optimization as Fig. 3 shows the re-
sults of optimization for the M6 wing obtained by a complete gradient calculation
‘Optim1-1’ compared to the same optimization where the adjoint mesh equation is
not solved ‘Optim1-2’. The first optimization required a total of 155 flow equivalent
solutions4 against 163 for the second optimization, which indicates that the loss of
accuracy did not destroy the convergence of the numerical optimization. Regarding
the computer cost there is no savings here because the CPU time due to solving the
adjoint mesh equations is less than the cost of the additional flow computations re-
quired in ‘Optim1-2’. This could be different in optimization involving the RANS
equations because the cost of deforming the RANS mesh can be considerably higher
than deforming inviscid meshes (Jakobsson and Amoignon 2007) and the cost of
solving the adjoint of the mesh deformation equation has a similar amplitude.

3 RBF parameterization of shape deformations

The choice of the RBF method for the parameterization of shapes in aerodynamic
shape optimization can be motivated by the following properties:

4 Flow equivalent solution means here one flow or adjoint flow solution. The cost for one flow solution in
this example is 1520 seconds CPU, and 24 seconds CPU for the solution of one adjoint mesh deformation
equation.
AESOP—a numerical platform for aerodynamic shape optimization

Fig. 3 M6 wing optimization with complete gradient calculation ‘Optim1-1’ against a gradient calculation
without adjoint mesh solution ‘Optim1-2’

1. Scattered data interpolation. Shapes are generally described by clouds of points

without topological information. The primal idea of the parameterization defined
below (Sect. 3.1) is to interpolate the displacements of control points scattered in
R2 or R3 in order to define displacements everywhere.
2. Regularization. In gradient-based optimization, the parameters are updated given
shape gradients of the aerodynamic coefficients, which have no known regularity,
an example is given in Fig. 4. The parameters in our RBF parameterization are
the displacements of control points. When the RBF is such that it can resolve the
high frequencies of the shape gradient, the interpolated displacements would in-
evitably create wavy shape deformations. Therefore, a regularized approximation
as defined in Sect. 3.2 is in this case preferred to an interpolant, see Fig. 5.
3. Extrapolation. Aeroelastic shape optimization requires deforming the structure
model according to changes in the parameters that describe the shape. The same
RBF can be used for the parameterization of the wetted surface deformations and
for the deformation of the inner structure model.
4. ‘Arbitrary’ dimensions. The RBF formulation involves a distance between the lo-
cation of data, which is invariant with respect to the dimension.
The RBF interpolation and approximation (regularized RBF) are presented in the
next two sections. This type of representation of shape deformations often requires
to set boundaries to the region where the deformations are defined (Jakobsson and
 O. Amoignon

Fig. 4 Gradient of the drag coefficient with respect to the nodal coordinates on the airfoil, including flow
derivatives: x-coordinates (left) and y-coordinates (right)

Fig. 5 Comparison of inviscid

optimization of airfoils using
RBF parameterization: with
regularization (case: 1) and
without regularization (case: 2).
Pressure coefficients (−Cp ) on
the left and geometries on the
right

Amoignon 2007), for example defining cut-off functions as in Sect. 3.3. Finally, the
computation of gradients, the steps (20)–(21) above, is described in Sect. 3.4.

3.1 RBF interpolant

For the sake of the presentation, the parameters of design used here {ai ∈ R, 1 ≤
i ≤ n} are the displacements in y-direction of n control points. Denoting by x =
[x, y, z, . . .] the coordinates in Rd of a material point, and the superscript 0 denot-
ing its position at rest, a displacement mapping d h (x) : Rd → Rd is defined here,
generalizing the definition (8) to all points in Rd :
 
x = x0 + dh x0 . (31)

The mapping is further defined by imposing the following interpolation condition at

the control points:

dy (x i ) = ai , 1 ≤ i ≤ n, (32)
AESOP—a numerical platform for aerodynamic shape optimization

where dy (x) is the y-component of d h . The displacements dy (x), represented by a

RBF such as the one derived in Jakobsson and Amoignon (2007), are expressed as:


n 
d+1
dy (x) = cy,i φ (x − x i ) + by,k Qk (x) , (33)
i=1 k=1

where the radial function φ (r) is defined by a real valued function (t) (see Table 1
for examples) via
φ (r) = (
r
) , (34)
and
.
denotes here the Euclidean norm,  being the shape factor, a strictly positive
real number. For 3D applications (d = 3) the monomials Q are:

Q1 (x) = 1,
Q2 (x) = x,
(35)
Q3 (x) = y,
Q4 (x) = z.

The coefficients of the RBF (33), {ch ; bh } = {cy,i ∈ R, 1 ≤ i ≤ n ; by,k ∈ R, 1 ≤ k ≤

d + 1}, need to fulfill the system of n + d + 1 equations:
  
A P cy a
T = h , (36)
P 0 b y 0

where A is the interpolation matrix:

 
A = Aij 1≤i≤n ,
1≤j ≤n
  (37)
Aij = φ x i − x j , 1 ≤ i ≤ n, 1 ≤ j ≤ n,

and P is the matrix of constraints

P = {Pik } 1≤i≤n ,
1≤k≤d+1
(38)
Pik = Qk (x i ) , 1 ≤ i ≤ n, 1 ≤ k ≤ d + 1.

The system (36) is not singular if the coordinates of the control points x i are distinct
and if A defined by (37) is not singular. The last condition is fulfilled for certain
functions , see Wendland (2005) and the examples in Table 1. The d + 1 constraints
are imposed in Jakobsson and Amoignon (2007) in order to obtain interpolations
that are translation and rotation invariant, in addition to fulfilling the interpolation
conditions (32).

3.2 Regularized RBF

The interpolation presented above may induce wiggles when increasing the number
of control points n in (32)–(33). Even small oscillations on the geometry, as in the
 O. Amoignon

Table 1 Example of “RBF” functions

Basis function (t) where t ≥ 0

Wendland (W3,1 ) (4t + 1) × (1 − t)4 , if 0 ≤ t ≤ 1, 0 otherwise

Inverse multi-quadric (IMQ) √1
 1+t 2
Multi-quadric (MQ) 1 + t2
2
Gaussian (GS) e−t
Thin Plate Splines (TPS2) t 2 log(t), if t > 0, 0 otherwise
Thin Plate Splines (TPS4) t 4 log(t), if t > 0, 0 otherwise

example shown in Fig. 5, affect the pressure in a way that could trigger flow separa-
tion or (laminar to turbulent) transition, for example. A possible cure is to penalize a
norm of the RBF (33) while approximating the displacements of the control points.
By this approach the relation (32) does not necessarily hold. The standard approach,
in order to define an approximation dy of the form (33), is to observe that the solution
of the system of equations (36) above is solution of the least-square approximation
problem with constraints:
⎧
⎪ n
 2
⎪
⎪ dy (x i ) − ai
⎪
⎪
min
⎨ cy ∈Rn ,by ∈Rd+1 i=1
(39)
⎪
⎪ n
⎪
⎪
⎪
⎩ subject to cy,i q (x i ) = 0,
i=1

where dy (x) is defined by (33) and q(x) is any first degree polynomial.
Adding a penalty on the (native space) norm of the RBF yields the following
problem of approximation:
⎧
⎪ n
 2 n  n
 
⎪
⎪ min d (x ) − a + β cy,i cy,j φ x i − x j
⎪
⎪ y i i
⎨ cy ∈Rn ,by ∈Rd+1 i=1 i=1 j =1
(40)
⎪
⎪ 
n
⎪
⎪
⎪
⎩ subject to cy,i q (x i ) = 0,
i=1

where dy and q are defined as in problem (39) and β ∈ R+ . If the matrix A defined
by (37) is not singular, the solution of the problem (40) is given by the solution of the
system:
  
A + βI P cy a
= h . (41)
PT 0 by 0
The choice of the parameter β is of course determining the “quality” of the ap-
proximation in terms of the pointwise error:

ey,i = dy (x i ) − ai , 1≤i≤n (42)

AESOP—a numerical platform for aerodynamic shape optimization

and in terms of the oscillations of dy . Small values of β give small pointwise errors
|ey,i |, with a risk of overshooting, whereas large values of β potentially eliminate
overshooting at the cost of increased pointwise errors. When approximating function-
als, there are no known optimal values of β but an attempt can be made to reduce the
influence of “noise” and therefore reduce the integrated error (based on the L2 norm
for example) by cross-validation, as found in Rudholm and Wojciechowski (2007).

3.3 Boundary conditions

In certain applications, such as the 3D examples presented in Sect. 4.3 or in Jakobsson

and Amoignon (2007), it is necessary to limit the support of the deformation defined
by RBF (33). Given a plane P (defined by a point x P and a normal vector nP ), the
displacements in the half-space P − (defined by x P − nP ∈ P − ) should be zero. We
define a cut-off function ϕP (x) that is 0 for points in P − and 1 away from the plane
P by:
  
ϕP (x) = 1 − W P (x−x P )·nP

nP
/ P −,
, for x ∈
(43)
ϕP (x) = 0, for x ∈ P −

where W is the Wendland function from Table 1. A displacement dyP , null in P − ,

would for example be used instead of dy , the latter being defined by (39) or (40):

dyP (x) = ϕP (x) dy (x) . (44)

The same approach can be used in order to impose pointwise constraints.

3.4 Computation of shape gradients

The following gives the details of the operations (20)–(21) in cases where the para-
meterization (9) is defined by the RBF approach described above. In order to make
the example more general suppose that regularization is used (Sect. 3.2) as well as
boundary conditions (Sect. 3.3). Suppose that we are given the gradient of J with
respect to the y-coordinates displacements on the shape dJ /ddy , calculating the gra-
dient of J with respect to the RBF coefficients cy and by requires the following
operations:

∂J 
Nw
    dJ
y = ϕP x j ψi x j , 1 ≤ i ≤ n + d + 1, (45)
∂λi ddy,j
j =1

where we included eventual cut-off functions in ϕP (x j ) as defined in Sect. 3.3 by

relation (43). The following notations were used in (45):
y  
λi = cy,i , and ψi (x) = φ x − x j , 1 ≤ i ≤ n,
y
(46)
λn+i = by,i , and ψn+i (x) = Qi (x) , 1 ≤ i ≤ d + 1.
 O. Amoignon

Calling dJ /dλy the vector of all partial derivatives of J with respect to the RBF
coefficients (45)–(46), we can calculate the gradient of J with respect to the vector
of all control points (y-)displacements ah by solving the system:
  
A + βI P dJ /dah dJ /dλy
= . (47)
PT 0 0 0

4 Examples

All examples below are solved in AESOP using the augmented Lagrangian al-
gorithm (see Sect. 2.1). The accuracy of the gradients computed by adjoint flow
equations as exposed above is an important aspect of these optimizations. An
estimation of the accuracy can be done comparing “adjoint gradients” with fi-
nite difference approximations (see for example Amoignon and Berggren 2006;
Amoignon et al. 2006). In Fig. 6 the gradients of the drag, lift and pitch, used in
the multipoint optimization test case below (see Sect. 4.1), are computed by the pro-
cedure above and by finite differences approximations where the perturbations of the
design variables vary between 10−4 and 10−7 . The gradients obtained by the two
methods agree within 1–2% at design point 1, and within 0.5% at design point 2, at
least for perturbation parameters smaller or equal than 10−5 . The difference of “ac-
curacy” between the two design points probably depends on the difference in shock
strength. The shock is stringer at design point 1 than at design point 2, which involves
that the artificial viscosity has a more important role at design point 1 thus leading to
larger errors because the artificial viscosity was not differentiated when deriving the
adjoint flow equations used here (see Amoignon and Berggren 2006).

4.1 Inviscid multipoint optimization

This is an example of airfoil design at cruise for Mach number 0.716, a study per-
formed in the European project CESAR. The goal was here to minimize the drag due

Fig. 6 Comparison of the

gradients at first iteration
obtained by adjoint method vs
finite differences with different
perturbation of the design
variables. The subscript
indicates the design point
(1 or 2)
AESOP—a numerical platform for aerodynamic shape optimization

Fig. 7 Multipoint inviscid

airfoil optimization at
Mach = 0.716. Pressure
coefficients −Cp at design point
1 (left) and design point 2
(right), unscaled airfoils
(bottom), where the dashed blue
line is for the initial design and
the solid red line is for the final
design

to shocks, in contrast to the next example (see Sect. 4.2) where the main objective
is to reduce the viscous friction. It is therefore sufficient to use the Euler equations
in order to model the flow. Note, however, that the performance of the final design
was verified in the project, as it is customary, solving the RANS equations around
the initial and final airfoils, but these results are not presented here. Constraints are
imposed on the thickness, the pitching moment and the lift, whereas the goal is to
minimize the drag. The multipoint approach consists here in defining a problem of
optimization that depends on more than one flight condition, for instance two angles
of attack {α1 , α2 }. Here α2 gives the required lift coefficient at Mach = 0.716, but the
wave drag is only 20 drag counts. Choosing to minimize the drag at a higher angle of
attack was the strategy chosen by the author in order to have better sensitivities of the
cost function with respect to changes in the geometry. The cost function is thus the
wave drag at α1 , an angle of attack at which the initial drag is about 80 drag counts.
The lift coefficients at both design points are imposed to be larger or equal than the
values for the baseline. Regarding the pitching moments, the project required values
larger than those of the baseline, so that a lower bound was finally chosen at −0.12.
Note that the pitching moments based only on the pressure can be very different from
the values that include the effect of viscosity, so that the main objective is to increase
the pitching moments (the convention here is negative pitch for nose down moment),
not to achieve precise values.
To summarize, the optimization problem is formulated as:
⎧
⎪
⎪
CD 1
⎪
⎪ min subject to
⎪
⎪  CD 01
⎪
⎪
⎪
⎪
⎪
⎪ CD 2 ≤ CD 02 ,
⎪
⎪
⎪
⎪
⎪
⎪
⎨ CL1 ≥ CL 1 ,
0

(48)
⎪
⎪ C ≥ CL 02 ,
⎪ L2
⎪
⎪
⎪
⎪
⎪ Cm 1 ≥ Cm min
1 ,
⎪
⎪
⎪
⎪
⎪
⎪ Cm2 ≥ Cm min
⎪
⎪ 2 ,
⎪
⎪
⎩
t ≥ t min ,
 O. Amoignon

where the superscript 0 indicates the values at initial design (baseline), min indi-
cates values chosen that are different from the baseline: t min = 9%, Cm min
1 = −0.12,
Cm 2 = −0.12. The parameterization of the airfoil , or its deformations, is the reg-
min

ularized RBF method described in Sect. 3.2. The radial function φ in expression (33)
is the Gaussian (see Table 1) with a shape factor  = 2.5, a regularization parame-
ter β = 1.1 in (40)–(41), and 11 control points distributed on the surface. Note that
imposing the constraints on the pitching moment, in the multipoint approach, has al-
lowed here to achieve a better design at both design points {α1 , α2 } than the designs
that could be obtained by performing only single point optimization at α1 and α2 ,
with less constraining conditions on the pitching moment. Not only the wave drag is
reduced of 70 drag counts at α1 and 12 drag counts at α2 , the lift at α2 was increased
as well as the pitching moments at both design points, all other constraints being sat-
isfied (Fig. 8). A possible explanation is that nonlinear problems of optimization of
this kind have local minima and imposing constraints away from the values at initial
design may help moving away from local optimum solutions.

4.2 Natural laminar flow design optimization

This is an example of laminar airfoil design at cruise conditions, the Mach num-
ber is 0.372 and the Reynolds number is 12E6. This work was also performed for
the European project CESAR. In contrast to the previous example the goal is to re-
duce the viscous drag. The common strategy in the area, and the one applied here,
is to increase the laminar portion of the flow. For this purpose the flow equations
(Euler in this example) are complemented with a model of the propagation of distur-
bances in the laminar part of the boundary layer. The reason is that the mechanism
of laminar-to-turbulent transition is generally governed by the growth of such dis-
turbances. Therefore, damping the growth of those disturbances is expected to delay
downstream the location of this transition, increasing the laminar portion of the flow
where the friction coefficient is much lower than in the turbulent area. The strategy
used here has been implemented for the design of laminar airfoils using adjoint sen-
sitivities in Amoignon et al. (2006). For this purpose the platform AESOP communi-
cates with additional solvers for the boundary layer flow and stability equations (and
their adjoints). Following this strategy the cost function is the energy of a selected
disturbance (E) propagated in the boundary layer on the suction side. Constraints are
imposed on the thickness, the pitching moment and the lift. In order to avoid increas-
ing the wave drag, around 10 drag counts for the baseline, a penalty term is added to
the energy in the cost function. This is preferred to a usual constraint because a de-
sign with a slightly higher wave drag could be acceptable, if the laminar region of the
flow is sufficiently increased. The lift and pitching moment coefficients are imposed
to be larger or equal than the values for the baseline, the angle of attack being such
that the initial lift is the minimum required lift.
AESOP—a numerical platform for aerodynamic shape optimization

Fig. 8 Multipoint inviscid airfoil optimization at M = 0.716. Merit function and gradient norm (first row
left), drag and gradient at design point 1 (first row right), lift at design point 1 (second row left), lift at
design point 2 (second row right), pitching moment at design point 1 (third row left), pitching moment at
design point 2 (third row right), drag at design point 2 (bottom left), thickness (bottom right). The gradients
are represented at initial (dashed blue lines) and final design (solid red lines)
 O. Amoignon

Fig. 9 NLF airfoil optimization

at Mach = 0.312, Re = 12E6.
Pressure coefficients −Cp (left)
and unscaled airfoils (right),
where the dashed blue line is for
the initial design and the solid
red line is for the final design

To summarize, the optimization problem is formulated as:

⎧ CD
⎪
⎪ log (E) + 0.1 0
⎪ min
⎪
subject to
⎪
⎪
 CD
⎪
⎨
CL ≥ C L , 0
(49)
⎪
⎪
⎪
⎪ C ≥ Cm 0 ,
⎪
⎪ m
⎪
⎩
t ≥ t min

where the superscript 0 indicates the values at initial design (baseline) and min in-
dicates a different reference value than that of the baseline: t min = 12%. The para-
meterization of the airfoil  deformations is the regularized RBF method described
in Sect. 3.2. The radial function φ in expression (33) is the Gaussian (see Table 1)
with a shape factor  = 1.7, a regularization parameters β = 0.1 in (40)–(41), and
8 control points distributed on the surface. As explained earlier, the computation of
the energy of a disturbance involves here the solution of the flow in the laminar part
of the boundary layer and the solution of the transport equations for the disturbance
in the boundary layer (here the Parabolized Stability Equations). In order to calcu-
late efficiently and accurately the gradients, the adjoint of the boundary layer and of
the stability equations are solved, in addition to the adjoint of the Euler flow equa-
tions (Amoignon et al. 2006). The reduction of the energy function (Fig. 10) is due to
the favorable pressure gradient on the suction side, which damps the growth of this
disturbance (see Fig. 9). Note that the use of the logarithm of the energy is necessary
due to variations of several orders of magnitude of the function E from the baseline
design to the final design (Fig. 10).

4.3 Inviscid transonic wing optimization

The ONERA M6 wing, described in the AGARD report (Schmitt and Charpin 1979),
is optimized here at Mach number 0.8395 and angle of attack α = 3.06◦ using the
Euler equations as flow model. The parameterization of the deformations of the wing
 is the interpolation RBF method described in Sect. 3.1 (without regularization).
A similar optimization was carried out in Amoignon and Berggren (2006) using a
parameterization of the twist, camber and thickness along the spanwise direction.
Another difference with those previous results is that the lift and pitching moment
AESOP—a numerical platform for aerodynamic shape optimization

Fig. 10 NLF airfoil optimization at Mach = 0.312, Re = 12E6. Merit function and gradient norm (first
row left), disturbance energy and gradient (first row right), wave drag (2nd row left), lift (2nd row right),
pitching moment (bottom left), thickness (bottom right). The gradients are represented at initial (dashed
blue lines) and final design (solid red lines)

coefficients are here considered as constraints, which was not the case in Amoignon
and Berggren (2006) where changes in these coefficients were only penalized. The
radial functions φ in Table 1 are compared: W3,1 (‘Optim1’), GS (‘Optim2’), TPS2
(‘Optim3’) and TPS4 (‘Optim4’). The shape factor is  = 0.17 for a chord that varies
between 10m at the root to 5.5m at the tip (see Fig. 11) and there are 52 control points
distributed on the surface. Concerning the boundary conditions of the flow domain,
a symmetry plane (x-z) is placed at the root of the wing. The root section geometry
 O. Amoignon

Fig. 11 Distribution of the

control points on the M6 wing
(on both sides)

is fixed through a cut-off function as described in Sect. 3.3 where nP = (0, 1, 0)

(orthogonal to the plane x-z) and P = 0.33.
To summarize, the optimization problem is formulated as:
⎧ CD
⎪
⎪ min subject to
⎪
⎪ CD 0
⎪
⎪

⎪
⎪
⎪
⎪
⎪
⎪ C ≥ CL 0 ,
⎨ L
Cm ≥ Cm 0 , (50)
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪ t20% ≥ t20%
0
,
⎪
⎪
⎪
⎪
⎩ t50% ≥ t 0
50%

where the superscript 0 indicates the values at initial design (baseline), t20% and t50%
indicate the thicknesses of the streamwise wing sections at 20% and 50% of span.
The trend for all four cases is visible in Figs. 12–13: the drag is reduced due
to the weakening of the shock, while the lift and pitching moment coefficients are
kept to their initial values. The new shapes perform even better than in Amoignon
and Berggren (2006) without reducing the thickness of the wing. Unless for the
Optim2 case, where the drag at final design is larger than in the three other cases,
there is no large difference between the aerodynamic performances of the designs
obtained by the three parameterizations used in Optim1, Optim3 and Optim4 (see
Merit function, drag, lift and pitch in Fig. 12). Comparing the airfoils from Optim1
(Fig. 14) and Optim2 (Fig. 15) shows important differences, for example in trail-
ing edges thickness and leading edges radius, depending on the methods of para-
meterization (the Optim3 and Optim4 airfoils, not shown here, show similar trends
as Optim1). These results are very different from the ones obtained in Amoignon
and Berggren (2006) because the reduction in drag is here around 40 drag counts,
without violation of the constraints whereas the reduction obtained in Amoignon and
Berggren (2006) was at best 30 drag counts with up to 1% violation of the lift. This
AESOP—a numerical platform for aerodynamic shape optimization

Fig. 12 Optimization. Influence of the parameterization of the M6 wing deformations: Wendland (Op-
tim1), Gauss (Optim2), TPS2 (Optim3), TPS3 (Optim4)

improvement is due to the shape parameterization used here. The use of RBF en-
ables to change the shape of the “airfoils” without being limited to changes in twist
and camber. Note that the drag induced by the lift for an elliptic lift distribution is
CL2 /πAR, where AR is the aspect ratio of the M6 wing (3.85), which is 90 drag
counts for CL = 0.33. The wings obtained by optimization (Optim1, Optim3 and
Optim4) are thus quite closed to the theoretical optimal wing with the same aspect
ratio.
 O. Amoignon

Fig. 13 Pressure coefficients and iso-Mach lines of the M6 wing (left) and of the optimized ‘Optim1’
(right), at Mach = 0.84, angle of attack of 3 degrees

5 Summary and outlook

AESOP is a platform for large scale aerodynamic shape optimization and for the de-
velopment of multidisciplinary design optimization (MDO). Large scale refers here
to problems of optimization with many design parameters or with computationally
intensive functions evaluations. The main advantage of the program is to take fully
advantage of the adjoint approach developed in recent years in CFD, for instance via
the coupling with the unstructured flow and adjoint flow solver in Edge (Amoignon
and Berggren 2006). Several developments can improve the performance or broaden
the fields of applications. The performance could for example be improved using
other optimization algorithms, for example the Method of Moving Asymptotes from
Svanberg (1987) and NLPQLP (a Sequential Quadratic Programming implementa-
tion) from Schittkowski (2001), which is the goal of current investigations.
Typical problems of aerodynamic shape optimization in aeronautics concern the
design or re-design of airfoils, wings, turbine blades, canals, pylons, fairing of sur-
faces intersections, among others. Devising a strategy of parameterization that gives
satisfying results for the optimization of all these shapes is not straightforward but it
seems that the properties of the Radial Basis Functions (RBF) presented above could
help building a general strategy.
An adjoint RANS solver is also being developed in Edge (Gardberg and Amoignon
2008). The capability to take into account viscous turbulent flows in the optimization
will enable the optimization of systems where the viscous effects are dominant, such
as high-lift systems. Another important aspect is that the use of the RANS equations
instead of Euler will improve the reliability of the results of optimization. For in-
stance, the pressure distribution around an optimized shape obtained with Euler may
well lead to flow separation although the aerodynamic coefficients computed by the
inviscid flow model can indicate that the shape has a low drag coefficient and the
right lift.
AESOP—a numerical platform for aerodynamic shape optimization

Fig. 14 Optimization. Comparison of the M6 wing shape and Cp with Optim1 wing at 0, 43 and 93
percent of span

Finally, aeroelasticity is being integrated in the aerodynamic shape optimization

framework presented here, starting by superposing static deformations due to loads
and shape deformations due to optimization.

Acknowledgements The author wants to thank Martin Berggren, professor at the university of Umeå, as
well as Jan Pralits, Ardeshir Hanifi and Mattias Chevalier from FOI, for the work done together on shape
optimization. Finally, part of this work received fundings from PSCI (VINNOVA, the Royal Institute of
Technology in Stockholm, Uppsala University and a consortium of Swedish industries) and the European
research projects in aeronautics: Aeroshape, Eurolift II, SUPERTRAC, HISAC, CESAR and NACRE.
 O. Amoignon

Fig. 15 Optimization. Comparison of the M6 wing shape and Cp with Optim2 wing at 0, 43 and 93
percent of span

References

Amoignon O, Berggren M (2006) Adjoint of a median-dual finite-volume scheme: application to transonic

aerodynamic shape optimization. Technical Report 2006-013, Department of Information Technol-
ogy, Uppsala University, Uppsala, Sweden
Amoignon OG, Pralits JO, Hanifi A, Berggren M, Henningson DS (2006) Shape optimization for delay of
laminar-turbulent transition. AIAA J 44(5):1009–1024
Anderson WK, Bonhaus DL (1999) Airfoil design on unstructured grids for turbulent flows. AIAA J
37(2):185–191
Baysal O, Ghayour K (2001) Continuous adjoint sensitivities for optimization with general cost functionals
on unstructured meshes. AIAA J 39(1):48–55
AESOP—a numerical platform for aerodynamic shape optimization

Beux F, Dervieux A (1991) Exact-gradient shape optimization of a 2D Euler flow. Technical Report RR-
1540, INRIA Sophia Antipolis, 2004 Route des Lucioles, 06360 Valbonne, France
Bugeda G, Oñate E (1999) Optimum aerodynamic shape design for fluid flow problems including mesh
adaptivity. Int J Numer Methods Fluids 30:161–178
Burgreen GW, Baysal O (1996) Three-dimensional aerodynamic shape optimization using discrete sensi-
tivity analysis. AIAA J 34(9):1761–1170
Eliasson P (2001) Edge, a Navier–Stokes solver, for unstructured grids. Technical Report FOI-R–0298–SE,
Swedish Defence Research Agency, Stockholm, November 2001
Elliot J, Peraire J (1996) Aerodynamic design using unstructured meshes. AIAA Paper (96-1941)
Enoksson O, Weinerfelt P (1999) Numerical methods for aerodynamic optimization. In: Proceedings to
the 8th international symposium on computational fluid dynamics, Bremen, 5–10 September 1999
Gardberg N, Amoignon O (2008) Development of the discrete adjoint RANS for a median-dual finite
volume scheme—application in aerodynamic shape optimization. Unpublished preprint, FOI
Giles MB, Pierce NA (2000) An introduction to the adjoint approach to design. Flow Turbul Combust
65:393–415
Gunzburger MD (2002) Perspectives in flow control and optimization. Society for Industrial and Applied
Mathematics, Philadelphia
Jakobsson S, Amoignon O (2007) Mesh deformation using radial basis functions for gradient-based aero-
dynamic shape optimization. Comput Fluids 36(6):1119–11136
Jameson A, Kim S (2003) Reduction of the adjoint gradient formula for aerodynamic shape optimization
problems. AIAA J, 41(11)
Mohammadi B (1997) A new optimal shape procedure for inviscid and viscous turbulent flows. Int J Numer
Methods Fluids 25:183–203
Mohammadi B (1999) Dynamical approaches and incomplete gradients for shape optimization and flow
control. AIAA J (99-3374)
Nadarajah SK (2003) The discrete adjoint approach to aerodynamic shape optimization. PhD thesis, Aero-
nautics and Astronautics Department, University of Stanford
Nocedal J, Wright S (1999) Numerical optimization. Springer series in operations research
Pironneau O (1984) Optimal shape design for elliptic systems. Springer, Berlin
Reuther JJ, Jameson A, Alonso JJ, Rimlinger MJ, Saunders D (1999) Constrained multipoint aerodynamic
shape optimization using an adjoint formulation and parallel computers, part 1. J Aircraft 36(1):51–
60
Rudholm J, Wojciechowski A (2007) A method for simulation based optimization using radial basis func-
tions. Technical report, Department of Mathematical Sciences, Division of Mathematics, Chalmers
University of Technology, Göteborg University, Göteborg, Sweden, August 2007
Schittkowski K (2001) NLPQLP: a new Fortran implementation of a sequential quadratic program-
ming algorithm for parallel computing. Technical report, Department of Mathematics, University
of Bayreuth
Schmitt V, Charpin F (1979) Pressure distributions on the ONERA-M6-WING at transonic Mach numbers.
In: Experimental data base for computer program assessment, p B1-1–B1-44. AGARD-AR-138, May
1979
Soemarwoto BI (1996) Multi-point aerodynamic design by optimization. PhD thesis, Delft University of
Technology, Faculty of Aerospace Engineering, PO Box 5058, 2600 GB Delft, Netherlands
Soto O, Löhner R (2001) Cfd shape optimization using an incomplete-gradient adjoint formulation. Int J
Numer Methods Eng 51:735–753
Sung C, Kwon JH (2000) Accurate aerodynamic sensitivity analysis using adjoint equations. AIAA J
38(2):243–250
Svanberg K (1987) The method of moving asymptotes—a new method for structural optimization. Int J
Numer Methods Eng 24:359–373
Wendland H (2005) Scattered data approximation. Cambridge monographs on applied and computational
mathematics