Journal of Computational Science 7 (2015) 112

Contents lists available at ScienceDirect

Journal of Computational Science

journal homepage: www.elsevier.com/locate/jocs

Simulation-driven design of low-speed wind tunnel contraction

Leifur Leifsson a,b, , Slawomir Koziel a
a
Engineering Optimization & Modeling Center, Reykjavik University, Menntavegur 1, 101 Reykjavik, Iceland
b
Iowa State University, Ames, IA 50011, USA

a r t i c l e i n f o a b s t r a c t

Article history: A low-speed wind tunnel is developed for conducting research on the ow past micro air vehicles. The tun-
Received 9 June 2014 nel is of open suction type and is composed of a square inlet with a honeycomb and turbulence screens,
Received in revised form settling chamber, contraction, experimental section housing, diffuser, and axial fan. In this paper, we
15 December 2014
describe the details of the design optimization procedure of the contraction, which is key to getting a high
Accepted 29 December 2014
quality ow in the experimental section. A high-delity computational uid dynamic (CFD) ow solver is
Available online 8 January 2015
used to capture the nonlinear ow physics. Due to the high computational expense of the CFD simulations,
surrogate-based optimization (SBO) is used to accelerate the design process. The SBO approach replaces
Keywords:
Simulation-driven design
direct optimization of the high-delity (accurate but computationally expensive) model by iterative opti-
Numerical optimization mization of a properly corrected low-delity model. Here, we exploit variabledelity CFD simulations,
Surrogate-based modeling as well as a simple multiplicative response correction technique to construct the surrogate model of the
Low-speed wind tunnel wind tunnel contraction, allowing us to optimize its shape at a low computational cost. To our knowl-
Experimental verication edge, it is the rst application of variabledelity surrogate modeling to wind tunnel contraction design.
The optimum nozzle design is veried using a high-delity CFD simulation, as well as by experimental
measurements of the fabricated wind tunnel. Experimental validation conrms the correctness of the
numerical optimization procedures utilized to design the contraction.
2015 Elsevier B.V. All rights reserved.

1. Introduction resolution video cameras, or a particle image velocimetry (PIV)

A low-speed wind tunnel, intended for uid dynamics research,
has been developed at the authors institution. The utilization of
the facility will involve research on micro air vehicles (MAVs), both
with xed- and apping-wings. The work conducted in the tunnel
will involve analysis of various aspects of MAVs, such as the wing
shape and apping parameters, and control effectiveness. High ow
quality is essential for the research as the facility is expected to
provide capability for the characterization of the aerodynamic per-
formance of MAVs at low Reynolds numbers (Re < 100,000). The
required ow speed range is from 0 to at least 20 m/s.
The wind tunnel is of an open suction type, as shown in Fig. 1,
and is composed of a square inlet with a honeycomb and turbu-
lence screens, settling chamber, contraction, experimental section
housing, diffuser, and axial fan. The experimental section is located
in a large enclosure which houses the experimental equipment
(the MAV) and the instrumentation (such as a force balance, high-
Corresponding author at: Iowa State University, Ames, IA 50011, USA.
Tel.: +1 515 294 6549.
E-mail address: leifur@iastate.edu (L. Leifsson).
system).
The contraction accelerates and aligns the ow into the test sec-
tion. The contraction is, therefore, a critical component of the wind
tunnel for the provision of high quality test ow in the working sec-
tion [1]. The size and shape of the contraction controls the quality
(such as ow angularity and uniformity, as well as the turbulence
intensity levels) of the ow at the outlet [2]. The contraction should
be short to minimize boundary layer growth and the ow leaving
the contraction should be uniform and steady [3,4]. Moreover, ow
separation, due to streamline curvature in the contraction, needs
to be avoided at any cost [5].
As the effects of the shape on the ow are highly nonlinear, it
is necessary to use computational uid dynamics (CFD) to make
design decisions about key features of the contraction, such as the
length, contraction ratio, and, in particular, the wall shape. Exam-
ples of such work can be found in Mathew [3] and Mathew et al.
[4], Watmuff [6], Su [7], and Doolan and Morgans [8]. The rst three
are similar as they all employ potential ow analysis and perform
parametric studies to nd an optimized design. The last one uses
numerical optimization techniques.
Watmuff [6] designed an axisymmetric low-speed wind tun-
nel contraction using potential ow analysis. The wall shape was

http://dx.doi.org/10.1016/j.jocs.2014.12.004
1877-7503/ 2015 Elsevier B.V. All rights reserved.
2 L. Leifsson, S. Koziel / Journal of Computational Science 7 (2015) 112

three-dimensional potential ow analysis and Stratfords criterion

Nomenclature for separation for a two-dimensional boundary layer are employed
in the parametric studies. The optimized design is validated by solv-
B Bzier curve ing the steady, turbulent, three-dimensional, Reynolds-averaged
f High-delity model NavierStokes equations using Menters SST k turbulence model.
H Contraction inlet height or an objective function Doolan and Morgans [8] employ a potential ow solver, but
N Number of mesh points along an axis they couple it with a laminar viscous solver. They move away from
n Order of a Bzier curve performing parametric studies and integrate their computational
P Control point of a Bzier curve model into an automated contraction design software tool which
s Surrogate model uses numerical optimization techniques to search for the opti-
V Mean cross-sectional ow speed mized design. Three different optimization techniques are used to
v Local ow speed optimize a contraction shape. Sequential quadratic programming
x Design variable vector [9], a gradient-based method, solved the problem efciently, but
x, y, z Cartesian axes not robustly. DIRECT [10], a gradient-free method, provided robust
global optimization at the expense of a large number of function
Greek letters
evaluations. Efcient global optimization [11], a surrogate-based
Local ow angle
method, was robust and always gave acceptable results, but its
Penalty factor
efciency depended on the initial random sampling required to
Trust-region radius
construct the functional surrogate model (Kriging [12] was used
 Variation
in this case). Further work on contraction physics and design can
 Gain ratio
be found in Refs. [3033].
Free-stream
All of the aforementioned work use traditional uid ow
solvers (coupled potential and viscous uid ow solvers), which
Subscripts
are computationally cheap, as well as conventional optimization
c Coarse model
techniques. The likely reason for not using three-dimensional high-
d Diagonal
delity CFD models during design is that the simulations are
f High-delity model
computationally expensive and numerical optimization requires a
h Horizontal
large number of function calls, resulting in a prohibitively large
s Surrogate model
design time.
max Maximum
The objective of this work is to design the shape of the contrac-
x, y, z Cartesian axes
tion so that the ow in the experimental section is of high quality.
It is necessary to use high-delity CFD simulations to capture
Superscripts
accurately the nonlinear ow physics throughout the optimization
i Design iteration
process. The coupled potential and viscous ow solvers, as used in
previous work, will not capture these effects with sufcient accu-
racy. As a matter of fact, high-delity simulations are essential. As
demonstrated in Section 5, the contraction design obtained based
parameterized by polynomials with a single design parameter.
on low-delity CFD analyses does not ensure the high-quality ow.
The design criterion was to minimize the adverse pressure gra-
In this work, we describe the design of the wind tunnel
dient near the inlet while not signicantly increasing the adverse
contraction using high-delity CFD simulation and surrogate-
pressure gradient near the exit. Su [7] considers rectangular wind
based optimization (SBO). SBO replaces direct optimization of the
tunnel contractions with incompressible potential ow analysis. A
high-delity (accurate but computationally expensive) model by
comparative parametric study is conducted based on seven geo-
iterative optimization of a properly corrected low-delity model.
metric parameters and ve criteria (including pressure extrema,
This reduces the number of evaluations of the high-delity model
ow non-uniformity, and crossow features). Mathew [3] designed
and, consequently, accelerates the optimization process. In partic-
a low-speed wind tunnel for aeroacoustic measurements. He per-
ular, we adopt the variabledelity approach with the high-delity
forms a parametric study of the geometry parameters (contraction
model based on the steady Reynolds-averaged NavierStokes
length, contraction ratio, outlet aspect ratio, and a single parameter
(RANS) equations, and the low-delity model based on the same
governing the wall shape (described by polynomials)) to minimize
equations, but with coarse discretization and relaxed convergence
the non-uniformities in the axial ow and the magnitudes of the
criteria. We use a simple response correction technique to create
other ow components, as well as checking for separated ow. A

Fig. 1. A schematic of the wind-tunnel from the side, as well as a cross-sectional view of the test section.
 L. Leifsson, S. Koziel / Journal of Computational Science 7 (2015) 112
Fig. 2. (a) One fourth of a three-dimensional contraction shape; (b) view of the upper half of the xz symmetry plane; (c) control directions in the outlet plane.
a surrogate that is subsequently used as a prediction tool to yield
an approximate high-delity model optimum. On completion of
the contraction design optimization, the entire wind tunnel was
fabricated and assembled. Physical measurements of the ow qual-
ity in the experimental section have been performed to validate
the nal wind tunnel design. In particular, we mapped the ow
speed within the cross-section of the experimental test section.
The experimental results and comparisons with our computational
model are presented.
2. Problem formulation
In this section, we describe the design problem and dene the
design objectives and constraints. We also formulate our solution
approach and describe the shape parameterization method for the
three-dimensional contraction.
2.1. Wind tunnel sizing
The wind tunnel overall dimensions were set prior to designing
the contraction prole due to housing size limitations. The inlet is
square shaped with a dimension of 2 m by 2 m. The settling chamber
length is 1.5 m. The square contraction follows the settling cham-
ber with a length of 2 m and, following Mathew [3], an area ratio
of 8 to 1. The contraction outlet is, therefore, a 71 cm square. The
axial length of the experimental section is 1.8 m. The opening to the
diffuser section is 80 cm square and transitions to a 1 m diameter
circle over a 2 m axial distance, where it connects to the axial-ow
fan. The overall length of the wind tunnel is around 8 m. To achieve
a ow speed in the experimental test section above 20 m/s, the
volumetric ow rate has to be at least 40,000 m3 /h.
2.2. Design objectives
The design objective is to determine a contraction shape which
provides uniform ow to the experimental test section. A sketch
of a typical square three-dimensional contraction shape is shown
in Fig. 2(a) and (b). At the outlet, the ow has to be uniform,
and the boundary layer needs to be thin to maximize the usable
cross-sectional area. The overall goal is thus to maximize the ow
uniformity at the outlet in terms of speed and angularity. More
specically, we want to minimize the variation of the speed (v)
and angularity () at the horizontal and diagonal control directions
in the outlet plane (Fig. 2(c)). The vertical control direction is not
needed due to symmetry.
3
The above objectives are formalized in Section 4.1 so that the
design process can be executed using numerical optimization pro-
cedures.
2.3. Design variables
The variables controlling the contraction shape are the inlet size,
the contraction area ratio, the length, and the wall shape. To sim-
plify our problem, we will x the inlet size, the contraction area
ratio, and length. This is necessary in this particular work, since
the size of the laboratory that houses the wind tunnel is a lim-
iting factor. Given these parameters, the design variables are the
parameters controlling the wall shape.
The contraction wall shape is dened by Bzier curves on the
symmetry planes. The Bzier curves, of order n, are dened as (see,
e.g., Ref. [13])

m 
n
n! ni i
B(t) = (1 t(k)) t(k) P(i), (1)
i!(n i)!
k=1 i=0
where Pi , i = 0,1,. . .n, are control points, and t is an 1 m array with
entries from 0 to 1. Any number of control points can be used to
represent the wall shape. Here, we use 5 control points, as shown
in Fig. 3. The rst control point is xed to the inlet height H. Control
points 2 and 4 are xed in the horizontal direction. The third control
point is free, i.e., it can be moved both horizontally and vertically.
The fth control point is xed to the outlet height. In total, this
gives 4 design variables. The design variable vector can be written
as x = [x2 x3 z3 x4 ]T .
3. Computational uid dynamic model
A single CFD simulation consists, in general, of four steps: the
geometry generation (described in Section 2.3), meshing of the
solution domain, numerical solution of the governing uid ow
equations, and post-processing of the ow results, which involves,
in the case of numerical optimization, calculating the objectives
and constraints. In this section we describe the high-delity CFD
model setup.
3.1. Governing equations
The ow is assumed to be steady, three-dimensional,
incompressible, and turbulent. The steady Reynolds-averaged
NavierStokes equations are taken as the governing uid ow
equations. The uid medium is air, assumed to be an ideal gas,
4 L. Leifsson, S. Koziel / Journal of Computational Science 7 (2015) 112

1
2
Wall Shape
1
Control Line
0.8
3

0.6
z/(H/2)

0.4
4

5
0.2

0
0 0.2 0.4 0.6 0.8 1 Fig. 5. A view of the computational mesh (coarsened for visualization) at a few
x/L planes and on the sides.

Fig. 3. Bzier curve of the contraction wall shape (solid line) in the symmetry plane,
and the control line (dashed line) with 5 control points (dots).
with the dynamic viscosity described by Sutherlands formula. Tur-
bulence is modeled by Menters Shear Stress Transport (SST) k
model [14]. The SST k turbulence model is suitable for complex
boundary layer ows under adverse pressure gradient and separa-
tion.
3.2. Computational mesh
The contraction is symmetric about the xy- and xz-planes. There-
fore, the computational domain is one quarter of the contraction
geometry, see Fig. 2(a). The mesh is body-tted structured mesh
with quadrilateral elements. The number of elements are parame-
terized along the x-, y-, and z-axes with Nx , Ny , and Nz , respectively,
as shown in Fig. 4. Here, we set Ny = Nz . The mesh is made denser in
the near the wall to capture the boundary layer effects. An exam-
ple mesh is shown in Fig. 5 showing the structure and topology of
the mesh. The distance to the rst node point above the wall was
set such that the non-dimensional wall distance y+ is less than 1.
The mesh is generated using the computer code ICEM CFD [15].
A parametric study was used to determine the appropriate mesh
parameters. The results are presented in Section 3.4.
3.3. Flow solver
Numerical simulations are performed using a nite volume
approach and a pressure-based formulation with FLUENT [16].
Asymptotic convergence to a steady state solution is obtained in
each case. Iterative convergence of each solution is examined by
monitoring the overall residual, which is the sum (over all the cells
in the computational domain) of the L2 norm of all the governing
equations solved in each cell. The convergence criteria is the rst
occurring of the following: a drop in the residuals by six orders of
magnitude, or a maximum number of iterations of 1000.
A velocity boundary condition is prescribed at the inlet, with
the initial velocity set to the cross-sectional mean velocity for a
given volumetric ow rate and a cross-sectional area (here we are
assuming that the boundary layer thickness at the contraction inlet
is zero). A pressure boundary condition is prescribed at the outlet
with the pressure set to the ambient pressure. Symmetry bound-
ary conditions are prescribed at the xy- and xz-planes. A no-slip
boundary condition is prescribed at the contraction surface.
3.4. Mesh independence study
A relatively good contraction shape is selected for the para-
metric study (the initial shape used in the optimization studies
described in Section 5.2). The volumetric ow rate is set to
40,000 m3 /h (as noted in Section 2.1). Four meshes with differ-
ent level of density were congured and simulation performed for
each of them. The coarse mesh has Nx = 15 and Nz = 15 (2800 cells),
the medium mesh Nx = 25 and Nz = 40 (37,000 cells), the ne mesh
Nx = 50 and Nz = 80 (306,000 cells), and the extra ne mesh Nx = 100
and Nz = 160 (2,503,000 cells).
The results of the parametric study are shown in Fig. 6, com-
paring the velocity and angularity distributions of the ow at the
outlet. In terms of the velocity, the ne and extra ne meshes are
similar and yield different results than both the medium and coarse
meshes. The ne and the extra ne meshes capture a sharp decrease
in speed near the surface, especially near the corner (as seen by the

Nx

Nz Ny

z Nz z Nz
x Nx y

Longit udinal cross-sect ion Inl et/outl et c ross-sect ion

Fig. 4. Mesh parameterization.

 L. Leifsson, S. Koziel / Journal of Computational Science 7 (2015) 112
Fig. 6. Results of a parametric study showing velocity and angularity proles at the outlet for different levels of mesh density.
diagonal velocity distribution in Fig. 6(b)), whereas the medium and
coarse do not. The medium mesh velocity distributions are signi-
cantly closer to the ne mesh results than the coarse mesh results.
The angularity distributions compare much better between the dif-
ferent meshes. Although, the medium and ne meshes results are
closer to each other than the coarse mesh results. Overall, the dif-
ference between the medium and ne mesh results are small, aside
from the velocity distribution near the end of the diagonal control
direction. The ne and the extra ne mesh results are very similar. It
can, therefore, be concluded that the ne mesh is sufciently dense
to adequately capture the ow physics and to be used to predict the
contraction performance in the design optimization study.
4. Optimization methodology
In this section, we describe our approach to optimizing the
nozzle shape. As explained in Section 2.1, the goal is to obtain as uni-
form velocity eld v as possible, where v is a function of four shape
parameters x = [x2 x3 z3 x4 ]T (cf. Section 2.3). In the following, we
describe the formulation of the objective function, the surrogate-
based optimization approach that is adopted here to design the
nozzle shape, as well as the construction of the surrogate model
which is based on the low-delity CFD simulations.
4.1. Objective function
As explained in Section 2.1, we are interested in minimizing the
variation of the velocity magnitude as well as the velocity angle
for the two control directions, horizontal and diagonal, denoted
as vh.f (x,t), h.f (x,t), vd.f (x,t), d.f (x,t), with 0 t 1 (0 corresponds
5
to the nozzle center, 1 corresponds to the nozzle wall). The sub-
script f indicates that the models are associated with high-delity
CFD simulations. For the horizontal direction we are interested
in minimizing vh.f (x,t) and h.f (x,t) for 0 t 0.9 (avoiding the
boundary layer region), for the diagonal direction we want to
minimize vd.f (x,t) and d.f (x,t) for 0 t 0.85 (avoiding the bound-
ary layer region). Because our two objectives (velocity magnitude
and angle) are non-commensurable, only one of them (angle) is
treated directly, the other one (magnitude) is handled through
penalty function. This a priori preference articulation allows us
to turn the intrinsically multi-objective problem into a single-
objective one, which is much simpler to handle. In general, having
two (potentially conicting) objectives would require genuine (and
computationally much more expensive) multi-objective approach
and extracting the most suitable design from the Pareto front [17].
The objective function is dened as:
Hf (x) = max{h.f (x), d.f (x)}
2
+[(max{vh.f (x)/vh.f.max , vd.f (x)/vd.f.max } vmax )/vmax ] , (2)
where vh.f (x) = max{0 t 0.9: |vh.f (x,t)|} min{0 t 0.9:
|vh.f (x,t)|}, vd.f (x) = max{0 t 0.85:
|vd.f (x,t)|} ( min{0 t 0.85: |vd.f (x,t)|}, vh.f. max = max{0 t 0.9:
|vh.f (x,t)|}, vd.f. max = max{0 t 0.85: |vd.f (x,t)|}. The denitions of
h.f (x) and d.f (x) are similar. In our numerical experiments
we use = 1000. vmax is the maximum allowable variation of the
velocity magnitude. Here, we set vmax = 0.01.
6 L. Leifsson, S. Koziel / Journal of Computational Science 7 (2015) 112
Fig. 7. High-delity, low-delity and surrogate models at the design x(i) and at another design x: (a) vh.c (x(i) ,t) (- - -) and vh.f (x(i) ,t) (); (b) vh.c (x,t) ( ), vh.s (i) (x,t) (- - -) and
vh.f (x,t) (); (c) vd.c (x(i) ,t) (- - -) and vd.f (x(i) ,t) (); (d) vd.c (x,t) ( ), vd.s (i) (x,t) (- - -) and vd.f (x,t) (); (e)(h): analogous plots for velocity angle.
4.2. Surrogate-based optimization
In order to reduce the design cost, we exploit surrogate-based
optimization (SBO) principle [1821]. More specically, the direct
minimization of the high-delity objective function (2) is replaced
by an iterative procedure [19]
(i)
x(i=1) = argminHs (x),
x
that generates a sequence of approximate solutions x(i) , i = 0, 1,. . .,
to the original problem argmin{x: Hf (x)}. Each x(i+1) is the opti-
mal design of the surrogate model Hs (i) , i = 0, 1,. . ., Hs (i) is assumed
to be a computationally cheap and sufciently reliable represen-
tation of Hf , particularly in the neighborhood of x(i) . Under these
assumptions, the Algorithm (3) is likely to produce a sequence of
designs that quickly approach xf . Usually, the high-delity model
is only evaluated once per iteration (at every new design x(i+1) ) for
verication purposes and to obtain the data necessary to update
the surrogate model. Because of the low computational cost of the
surrogate model, the cost of each iteration of the Algorithm (3) is
substantially lower than the cost of solving the original problem
with the high-delity objective function. Moreover, the number of
iterations for a well performing surrogate-based algorithm is sub-
stantially smaller than for any direct optimization method (e.g.,
gradient-based or pattern-search), which results in substantial
time savings. In this work, the surrogate model is optimized using
the pattern search algorithm [12].
The surrogate model Hs (i) is dened as follows:
(i) (i) (i)
Hs (x) = max{h.s (x), d.s (x)}
2
(i) (i)
+[(max{vh.s (x), vd.s (x)} vmax )/vmax ] , (4)
where the denitions of vh .s (i) (x), vd.s (i) (x), h.f (i) (x), and
d.f (i) (x) are based on the surrogate functions vh.s (i) (x,t), vd.s (i) (x,t),
(3) h.s (i) (x,t), and d.s (i) (x,t), and analogous to that for the high-delity
model.
4.3. Surrogate model construction
The surrogate functions can be constructed either from sam-
pled high-delity model data using an appropriate approximation
technique [22] or by utilizing a physically-based low-delity model
[18]. While surrogates based on approximation of high-delity data
are denitely more popular, they are not particularly suitable for
ad-hoc (particularly local) design of a given structure or a system
because of a high-computational cost of setting up such models.
Even for relatively small number of designable parameters, the
number of training samples necessary to ensure sufcient design
(typically, less than 5% of relative least-square error [23]) may be
a few hundred, which is not practical. Here, we exploit the latter
approach as we have a reliable low-delity model at our disposal
(see Section IV.D). The primary reason is that reliable physically-
based surrogates can be constructed using a fraction of high-delity
model data necessary to build accurate approximation models [20].
In the context of surrogate-based optimization, construction of the
 L. Leifsson, S. Koziel / Journal of Computational Science 7 (2015) 112 7

Table 1
Details of the four optimization studies.

Case Details

1 Direct optimization of the low-delity model

2 Direct optimization of the high-delity model
3 Surrogate-based optimization by the methodology of Section 4 embedded in the trust-region framework [28]
4 Surrogate-based optimization using the methodology of Section 4 without trust region

physics-based surrogate requires, in many cases, a single high- 5. Contraction shape optimization
delity model evaluation per design iteration [20].
There are several methods of constructing the surrogate from a 5.1. Setup
physically-based low-delity model. They include, among others,
space mapping (SM) [18], various response correction techniques The volumetric ow rate is xed at 40,000 cubic meters per
[24], manifold mapping [25], and shape-preserving response pre- hour (as noted in Section 2.1), which should yield an average uni-
diction [26]. Here, the surrogate model is created using a simple form speed at the outlet of 22 m/s. The wind tunnel contraction
multiplicative response correction, which turns out to be sufcient shape is optimized using the methodology described in Section 4.
for our purposes. An advantage of such an approach is that the sur- For the sake of comparison and verication, we carried out four
rogate is constructed using a single high-delity model evaluation, optimization runs using the approaches given in Table 1.
and it is easy to implement. In the trust-region approach, the surrogate-based optimization
At iteration i, the surrogate vh.s (i) (x,t) of the velocity magnitude Scheme (3) is replaced by:
vh.f (x,t) is constructed using the multiplicative response correction (i)
of the form: x(i=1) = arg min Hs (x), (6)
x,xx(i) (i)

where optimization of the surrogate model is constrained

vh.f (x(i) , t) to the vicinity of the current iteration point, determined
v(i)
h.s
(x, t) = vh.c (x, t) , (5) by the trust-region radius (i) . The value of the radius is
vh.c (x(i) , t)
updated in each iteration based on the so-called gain ratio
 = [Hf (x(i+1) ) Hf (x(i) )]/[Hs (i) (x(i+1) ) Hs (i) (x(i) )], which compares
the actual improvement of the high-delity objective function
where vh.c (x,t) is the low-delity velocity magnitude function. Simi-
to the improvement predicted by the surrogate. The radius is
lar denition holds for the remaining surrogate functions vd.s (i) (x,t),
increased when  is sufciently large and decreased if it is too
h.s (i) (x,t), and d.s (i) (x,t). Note that the Formulation (5) ensures
small. In case of no improvement, i.e., Hf (x(i+1) ) Hf (x(i) ), the new
zero-order consistency [27] between the surrogate and the high-
design is rejected and the iteration is restarted from the previous
delity model, i.e., vh.s (i) (x(i) ,t) = vh.f (x(i) ,t). Whilein principlethis
design with smaller (i) . Assuming that the surrogate is rst-order
is not sufcient to ensure the convergence of the Algorithm (2) to
consistent with the high-delity model, i.e., Hs (i) (x(i) ) = Hf (x(i) ),
an optimum of the high-delity model, as demonstrated in Section
sufcient reduction of the trust-region radius always leads to pos-
5, the algorithm performs very well. The primary reason is that
itive gain ratios (provided the functions are sufciently smooth),
the low-delity model is physically-based so that the changes of
which guarantees convergence of (6) to at least local optimum of
vh.f (x,t) and vh.c (x,t) are correlated under the change of the shape
the high-delity model [25]. In our case, rst-order consistency
parameters x. Fig. 7 illustrates the surrogate model construction
does not hold, however, because of using physics-based surrogate
and its generalization performance. The plots in the left panel (i.e.,
it is satised approximately [26], which usually allows for obtain-
(a), (c), (e), and (g)) illustrate the low- and high-delity model
ing better design (particularly in terms of satisfying constraints)
responses at certain design, whereas the plots in the right panel (i.e.,
and ensures convergence of the algorithm. However, this is usually
(b), (d), (f), and (h)) show the low- and high-delity, as well as the
obtained at a slightly higher computational cost as compared to (3).
surrogate responses at some other design. It can be observed that
the alignment between the surrogate and the high-delity model
at that other design is signicantly better than between the low- 5.2. Numerical results
and high-delity model.
In all cases, the optimization engine was a pattern search algo-
rithm [29]. The numerical results are presented in Table 2. It

Table 2
4.4. Low-delity CFD model Wind tunnel contraction shape design: numerical results.

The low-delity model is based on the same CFD model as the Variable Initial Case 1 Case 2 Case 3 Case 4

high-delity one (cf. Section 3). However, as the corrected low- x1 0.25 0.3333 0.5000 0.3925 0.4977
delity model will replace the high-delity one in the optimization x2 0.50 0.6333 0.5038 0.6236 0.6840
x3 0.70 0.6667 0.5556 0.5545 0.2691
process, it needs to be faster in order to reduce the computa-
x4 0.70 0.6327 0.5025 0.5262 0.5305
tional cost of the design. The simulation time can be substantially  [deg] 4.62 3.8 1.6 2.0 0.7
reduced by making the grid coarser. Here, we used the medium v [%] 1.1 1.0 1.0 1.0 1.2
mesh obtained from the mesh independence study (see Section Nf N/A 0 134 9 4
3.4). The simulation time is reduced further by reducing the number Nc N/A 120 0 900 400
Total costa N/A <0.5 134 <13 <6
of iterations. We, therefore, relax the convergence criteria for the
a
low-delity model by setting the maximum number of iterations to The total optimization cost is expressed in terms of the equivalent number of
high-delity model evaluations, based on the number of high-delity evaluations
15 (based on parametric studies for a given initial design). The ratio
(Nf ), the number of low-delity evaluations (Nc ), and the ratio of the high-delity
of simulation time of the high-delity model to the low-delity model evaluation time to the corrected low-delity model evaluation time, which
model is around 250. is around 250.
8 L. Leifsson, S. Koziel / Journal of Computational Science 7 (2015) 112
Fig. 8. Optimized nozzle shape obtained using four different design approaches:
initial shape (), the shape resulting from direct optimization of the low-delity
model (), the shape resulting from the direct optimization of the high-delity
model (), the shape obtained by surrogate-based optimization within trust-
region framework ( ), the shape obtained by surrogate-based optimization (no
trust region) (- - - -).
can be observed that the shape obtained by optimizing the low-
delity model does not bring much of the improvement in terms
of reducing the velocity angle variation. As shown in Fig. 8, the cor-
responding nozzle shape is actually quite similar to the initial one.
This indicates that simply replacing the high-delity model by the
low-delity one in the design process is not a good option.
Direct optimization of the high-delity model substantially
reduces  (from 4.6 to 1.6 ), however, the design cost is quite sub-
stantial, 134 high-delity model evaluations. The surrogate based
optimization enhanced by the trust-region framework yields the
design with slightly worse reduction of  (2.0 ), however, the
design cost is much lower only 13 equivalent high-delity model
evaluations.
Fig. 9. Velocity contours of the optimized design (Case 4) in (a) the longitudinal cross-section and (b) the outlet.
Fig. 10. Flow properties at the wall of the cross-section of the optimized design (Case 4): (a) pressure coefcient, and (b) skin friction coefcient.
The surrogate-based optimization working without trust region
yields much better design with respect to the velocity angle varia-
tion,  of only 0.7 and at a very low cost of less than 6 equivalent
high-delity model evaluations. However, the v constraint is
slightly violated in this case, which is a result of not using the trust-
region approach (TR enforces satisfaction of v constraint in Case
3). The nozzle shapes obtained in Cases 2 through 4 are rather sim-
ilar to each other. Note that the lowest value of  corresponds to
the attest design of the nozzle outlet (cf. Fig. 8).
5.3. Discussion
The velocity contours of the cross-section of the optimized
design, shown in Fig. 9(a), indicate how the speed is smoothly
increasing along the contraction. The velocity contours at the outlet,
shown in Fig. 9(b), indicate a uniform ow speed in majority of the
area, with reduction in speed in the thin boundary layer adjacent
to the wall, as well as in the corner. The pressure coefcient (Cp )
distribution along the corner surface, shown in Fig. 10(a), exhibits
a smooth reduction in pressure from the inlet, but a local minima
half-a-meter from the exit, indicating an adverse pressure gradi-
ent. However, the skin friction (Cf ) coefcient distribution, shown
in Fig. 10(b), does not reach zero, indicating that no ow separation
occurs.
6. Experimental validation
In this section, we describe the measurements conducted in the
wind tunnel after fabrication. In particular, we present measure-
ments of the ow speed uniformity and uctuating levels in the
experimental test section.
 L. Leifsson, S. Koziel / Journal of Computational Science 7 (2015) 112 9

Fig. 11. Wind tunnel conguration and experimental setup: (a) a side view showing (from the left) the inlet (includes the bell mouth and turbulence screens), settling
chamber, contraction, experimental section, diffuser, and axial fan (the overall length is 8 m), (b) a view inside the test section housing showing the single-component
constant-temperature hot-wire probe on a stand.

the downstream end of the test section there is 80 cm square open-

ing to the diffuser section. The diffuser section, made from sheet
metal, transitions from a square cross-section to a 1 m diameter
circle over a 2 m axial distance. The axial fan was manufactured by
FlktWoods and is driven by an 18 kW motor which is controlled
by a variable frequency drive inverter. The maximum volumetric
ow rate of the fan is 50,000 m3 /h; yielding a maximum ow speed
of approximately 30 m/s in the test section.

6.2. Experimental setup and calibration

Fig. 12. Locations of the experimental measurements within the test section (H = 71
cm, L = 1.8 m). The ow is along the x-direction.
6.1. Wind tunnel description
The low-speed wind tunnel is of open suction type and is
composed of a square inlet with turbulence screens, settling cham-
ber, contraction, test section housing, diffuser, and an axial fan
(Fig. 11(a)). The inlet has a cross-sectional area of 2 m by 2 m. There
are 3 turbulence screens, spaced apart by 100 mm, to break up any
coherent large scale motions. The mesh density gradually increases
from 12 per inch, then 16 per inch, and lastly 33 per inch. The set-
tling chamber is 1.5 m long and allows for any small scale organized
motions generated by the ow conditioning section to decay. The
square contraction follows the settling chamber. The contraction is
2 m long with an area ratio of 8 to 1 that results in a 71 cm square
entrance to the test section. The contraction contour was numer-
ically optimized as described in the previous section (cf. Case 4
in Section 2.4). The inlet, settling chamber and contraction are all
made from sheet metal. The enclosure surrounding the open jet
test section has a volume of 13.5 m3 and an axial length of 1.8 m. At
A single-component constant-temperature hot-wire anemome-
ter system was used to investigate the ow uniformity, as well as
the ow speed uctuating levels, in the test section. The sensing
element of the hot-wire probe was made from tungsten and had a
diameter of 5 microns and length of nominally 1.2 mm. The sensing
element was connected to the MiniCTA bridge. The output of the
bridge was then digitized with a National Instruments data acqui-
sition card with a 16-bit resolution. The sampling time was 6.5 s at
10 kHz. For the low order statistics presented here, we estimate the
uncertainty in the hot-wire measurements to be less than 1%.
Calibration of the hot-wire probe was accomplished by placing a
pitotstatic probe in close proximity to the hot-wire and collecting
data for a series of wind tunnel settings that extended over the full
range of the VFD, i.e., from 0 to 50 Hz. For each of the free-stream
velocity settings, mean values of both the hot-wire voltage and
speed (calculated from the pitotstatic probe) were determined,
and a curve t was used to relate the two. The error of the cal-
ibration curve t of the mean velocities was less than 1% for all
velocities.
In the test section, the hot-wire probe was placed on a stand fac-
ing the ow coming from the contraction exit as shown in Fig. 11(b).

(a) y = 0 (b) 10 H Z

Fig. 13. Average ow speeds within the test section: (a) along the streamwise direction, and (b) normalized cross-sectional ow speed proles.
10 L. Leifsson, S. Koziel / Journal of Computational Science 7 (2015) 112

Fig. 14. Flow uniformity and uctuation levels at 25% of the test section length from the contraction exit. Data is shown for motor settings of 10, 20, and 30 Hz, or, mean
free-stream speeds (V ) of 5.4, 11.1, and 16.5 m/s, respectively. Flow speed is given in m/s and turbulence intensity in %. The area of interest in the design optimization study
(cf. Section 2) is shown here with the dashed lines (85% value of the contraction exit size). The contraction exit boundary is shown with solid lines.

Fig. 12 gives the layout and coordinate system of the experimen-
tal setup. Three different experimental investigations were carried
out. In the rst one, we measured the averaged ow speed in the
ow direction on the centerline of the test section (y = 0, x/L = 0%,
25%, 50%, and 75%) at three settings of the variable frequency drive
(VFD), i.e., at 10, 20, and 30 Hz. Then we measured the average ow
speed at 11 equally spaced locations in the cross-section direction
(along the y direction) at a VFD setting of 10 Hz at three different
stations (x/L = 25%, 50%, and 75%). Finally, by traversing the probe,
measurements were taken in one cross-sectional plane (located at
x/L = 25%). These measurements were taken at nine equally spaces
locations, separated by 10 cm, in each direction, yielding a total
of 81 measurements. The cross-sectional area where the measure-
ments were taken was 80 cm by 80 cm and extended well outside
the contraction exit area (71 cm by 71 cm). The angularity distribu-
tions in the ow eld were not measured.
6.3. Experimental results: ow uniformity and uctuation levels
The average ow speed along the centerline (y = 0) of the test
section at four locations along the streamwise direction (x/L = 0%,
25%, 50%, and 75%) are shown in Fig. 13(a). The results show that,
 L. Leifsson, S. Koziel / Journal of Computational Science 7 (2015) 112
at the centerline, the average ow speeds are nearly constant along
the streamwise direction at 5.4, 11.1, and 16.5 m/s for 10, 20, and
30 Hz VFD settings, respectively.
The velocity prole across the test section, shown in Fig. 13(b),
is quite uniform within 85% of the test section width at x/L = 25%,
and 50%. At x/L = 75%, the ow speed drops signicantly near the
85% width.
These measurements indicate that the ow speed distribution is
quite similar at x/L = 0%, 25%, and 50%, which is where the placement
of the experimental equipment is intended to be located.
Contour maps of the mean velocity eld and the turbulence
intensity at x/L = 25%, calculated from the hot-wire probe measure-
ments, are shown in Fig. 14 for three different free-stream velocity
settings. The mean velocity eld contours indicate that the ow
is uniform within the region of interest, the 85% area of the con-
traction exit, and the ow speed within the cross-section matches
well with the free-stream velocity. However, there is a slight varia-
tion at the lowest free-stream velocity, where the speed is reduced
to 5.3 m/s in the upper left hand corner (Fig. 14(a)). This variation
of the ow speed is not within the measurement uncertainty. It
points to an unsymmetrical inow at the inlet due to a perturbation
in the experimental setup. This will be rectied by clearing these
objects.
Overall, the experimental measurements indicate that the ow
eld within the experimental section is quite uniform within the
region of interest, i.e., for 0.85 y/(H/2) 0.85 and from x/L = 0%
to 50%. However, the turbulence intensity contours indicate that
the ow unsteadiness is rather high, ranging from 1 to 5% within
the region of interest. A desired level of turbulence intensity is less
than 1% in the whole cross-section. The turbulence intensity levels
are controlled by the turbulence screens, with each new screen
the level can be halved [1]. Currently, the wind tunnel has three
turbulence screens; therefore, three additional screens are needed
to reach the desired intensity levels.
7. Conclusions
We present an efcient design optimization of the wind tun-
nel contraction shape. The shape is parameterized using Bzier
curves with ve control points. The optimized design is obtained
using surrogate-based optimization with the response-corrected
low-delity CFD model exploited as a fast surrogate model that
allows us to predict the optimum control parameters at a low com-
putational cost. To our knowledge, it is the rst application of SBO
to wind tunnel contraction design. The design is veried through
high-delity CFD simulations, as well as by experimental measure-
ments. The experimental measurements show that the desired ow
speed uniformity has been achieved in the region of interest for
wind tunnel testing. The ow angularity was not measured. The
turbulence intensity levels are still rather high, but this indepen-
dent of the contraction shape and can be rectied by adding three
more turbulence screens. Although, the optimization methodol-
ogy presented here is applied to a specic design problem, it is
general and can be applied to other internal ow problems such
as the design of arbitrary shape contractions, nozzles, or engine
inlets, as well as external ow problems such as axisymmetric
bodies.
Acknowledgements
The authors like to acknowledge the help of the undergraduate
students, Fannar Andrason and Kristjan Magnusson, assisting with
experimental measurements.
11
References
[1] J.B. Barlow, W.H. Rae, A. Pope, Low-Speed Wind Tunnel Testing, 3rd ed., John
Wiley & Sons, 1999.
[2] G.I. Derbunivich, A.S. Zemskaya, E.U. Repik, Y.P. Sosedko, Effect of Flow
Contraction on the Level of Turbulence, Izvestiya Akademii Nauk SSSr,
Mekkhanika Zhidkosti I Gaza, No. 2, pp. 146152, MarchApril 1987.
[3] J. Mathew, Design, fabrication and characterization of an anechoic wind
tunnel facility, in: PhD Dissertation, Dept. of Mech. and Aerospace
Engineering, University of Florida, 2005.
[4] J. Mathew, C. Bahr, B. Carroll, M. Sheplak, L. Cattafesta, Design, fabrication, and
characterization of an anechoic wind tunnel facility, in: 11th AIAA/CEAS
Aeroacoustics Conference, Monterey, California, May 2325, 2005.
[5] L. Cattafesta, C.J. Bahr Mathew, Fundamentals of wind-tunnel design, in:
Encyclopedia of Aerospace Engineering, John Wiley and Sons, Inc., 2010.
[6] J. Watmuff, Wind tunnel contraction design, in: 9th Australasian Fluid
Mechanics Conference, Auckland, New Zealand, Dec. 812, 1986.
[7] Y. Su, Flow analysis and design of three-dimensional wind tunnel
contractions, AIAA J. 29 (1991) 19121920.
[8] C.J. Doolan, R.C. Morgans, Numerical evaluation and optimization of low
speed wind tunnel contractions, in: 18th AIAA Computational Fluid Dynamics
Conference, Miami Florida, June 2528, 2007.
[9] K. Schittkowski, NLPQL: a FORTRAN subroutine solving constrained nonlinear
programming problems, Ann. Oper. Res. 5 (1985) 485500.
[10] D. Jones, C. Perttunen, B. Stuckman, Lipschitzan optimisation without the
Lipschitz constant, J. Optim. Theory App. 79 (1993) 157181.
[11] D.R. Jones, M. Schonlau, W.J. Welch, Efcient global optmization of expensive
black-box functions, J. Global Optim. 13 (1998) 455492.
[12] N.V. Queipo, R.T. Haftka, W. Shyy, T. Goel, R. Vaidynathan, P.K. Tucker,
Surrogate based analysis and optimization, Prog. Aerosp. Sci. 41 (1) (2005)
128.
[13] J. Lepine, F. Guibault, J.-Y. Trepanier, F. Pepin, Optimized nonuniform rational
B-spline geometrical representation for aerodynamic design of wings, AIAA J.
39 (11) (2001) 20332041.
[14] F. Menter, Two-equation eddy-viscosity turbulence models for engineering
applications, AIAA J. 32 (1994) 15981605.
[15] ICEM CFD, ver. 14.0, ANSYS Inc. Southpointe, 275 Technology Drive,
Canonsburg, PA 15317, 2011.
[16] FLUENT, ver. 14.0, ANSYS Inc. Southpointe, 275 Technology Drive,
Canonsburg, PA 15317, 2011.
[17] C.A. Coello Coello, Evolutionary Algorithms for Solving Multi-Objective
Problems, 2nd ed., Springer, 2013.
[18] J.W. Bandler, Q.S. Cheng, S.A. Dakroury, A.S. Mohamed, M.H. Bakr, K. Madsen,
J. Sndergaard, Space mapping: the state of the art, IEEE Trans. Microwave
Theory Tech. 52 (1) (2004) 337361.
[19] N.V. Queipo, R.T. Haftka, W. Shyy, T. Goel, R. Vaidynathan, P.K. Tucker,
Surrogate-based analysis and optimization, Prog. Aerosp. Sci. 41 (1) (2005)
128.
[20] S. Koziel, D. Echeverra-Ciaurri, L. Leifsson, Surrogate-based methods, in: S.
Koziel, X.S. Yang (Eds.), Computational Optimization, Methods and
Algorithms, Series: Studies in Computational Intelligence, Springer-Verlag,
2011, pp. 3360.
[21] A.I.J. Forrester, A.J. Keane, Recent advances in surrogate-based optimization,
Prog. Aerosp. Sci. 45 (2009) 5079.
[22] T.W. Simpson, J. Peplinski, P.N. Koch, J.K. Allen, Metamodels for
computer-based engineering design: survey and recommendations, Eng.
Comput. 17 (2001) 129150.
[23] I. Couckuyt, Forward and inverse surrogate modeling of computationally
expensive problems, in: PhD Thesis, Ghent University, 2013, 2015.
[24] J. Sndergaard, Optimization using surrogate models by the space mapping
technique, in: PhD Thesis, Informatics and Mathematical Modelling, Technical
University of Denmark, Lyngby, 2003.
[25] D. Echeverra, P.W. Hemker, Manifold mapping: a two-level optimization
technique, Comput. Visual. Sci. 11 (2008) 193206.
[26] S. Koziel, Shape-preserving response prediction for microwave design
optimization, IEEE Trans. Microwave Theory Tech. 58 (11) (2010)
28292837.
[27] N.M. Alexandrov, R.M. Lewis, An overview of rst-order model management
for engineering optimization, Optim. Eng. 2 (2001) 413430.
[28] A.R. Conn, N.I.M. Gould, P.L. Toint, Trust Region Methods, MPS-SIAM Series on
Optimization, 2000.
[29] S. Koziel, Multi-delity multi-grid design optimization of planar microwave
structures with Sonnet, in: International Review of Progress in Applied
Computational Electromagnetics, Tampere, Finland, April 2629, 2010, pp.
719724.
[30] J.H. Bell, R.D. Mehta, 1988. Contraction Design for Small Low Speed Wind
Tunnels, NASA, Contractor, Rep. No. NASA-CR-177488.
[31] B. Brassard, M. Ferchichi, Transformation of a polynomial for a contraction
wall prole, J. Fluids Eng. 127 (2005) 183185.
[32] R.D. Mehta, P. Bradshaw, Design rules for small low speed wind tunnels,
Aeronaut. J. 73 (1979) 443449.
[33] C.J. Doolan, Numerical evaluation of contemporary low-speed wind tunnel
contraction designs, J. Fluids Eng. 129 (2007) 12411244.
12 L. Leifsson, S. Koziel / Journal of Computational Science 7 (2015) 112
Leifur Leifsson received a Ph.D. degree in Aerospace Engi-
neering from Virginia Tech, USA, in 2006. He is currently
an Associate Professor with the School of Science and
Engineering at Reykjavik University (RU), Iceland. Leifur
is the director of the Laboratory for Unmanned Vehicles at
RU. His research interests include applied aerodynamics,
surrogate-based modelling and optimization, multidis-
ciplinary design optimization, and unmanned vehicle
design. Prior coming to RU, Leifur held positions at Airbus
UK Ltd. and Hafmynd Ltd.
Slawomir Koziel received the M.Sc. and Ph.D. degrees in
electronic engineering from Gdansk University of Tech-
nology, Poland, in 1995 and 2000, respectively. He also
received the M.Sc. degrees in theoretical physics and in
mathematics, in 2000 and 2002, respectively, as well as
the Ph.D. in mathematics in 2003, from the University of
Gdansk, Poland. He is currently a Professor with the School
of Science and Engineering, Reykjavik University (RU),
Iceland. He is the director of the Engineering Optimiza-
tion & Modeling Center, RU. His research interests include
CAD of microwave circuits, surrogate-based modeling and
optimization, circuit theory, and numerical analysis.