Bevan, R., Poole, D., Allen, C., & Rendall, T. (2016).

Optimization of
Vane-Type Vortex Generators for Tiltrotor Wings using Computational
Fluid Dynamics. In 54th AIAA Aerospace Sciences Meeting Article
AIAA 2016-1304 American Institute of Aeronautics and Astronautics
Inc. (AIAA). [Link]

Peer reviewed version

Link to published version (if available):

10.2514/6.2016-1304

Link to publication record on the Bristol Research Portal

PDF-document

This is the author accepted manuscript (AAM). The final published version (version of record) is available online
via AIAA at [Link] Please refer to any applicable terms of use of the
publisher.

University of Bristol – Bristol Research Portal

General rights

This document is made available in accordance with publisher policies. Please cite only the
published version using the reference above. Full terms of use are available:
[Link]
 Optimization of Vane-Type Vortex Generators
for Tiltrotor Wings using Computational Fluid
Dynamics

R.L.T. Bevan∗, D.J. Poole†, C.B. Allen‡, T.C.S. Rendall§

Department of Aerospace Engineering, University of Bristol, Bristol, BS8 1TR, U.K.

Tiltrotor wings are thick and highly loaded airfoils, and are thus particularly susceptible
to separation and early onset buffet and vortex generators are commonly used to alleviate
these issues. Hence. the design of counter-rotating vortex generators on tiltrotor wings to
control separation is considered. A representative tiltrotor wing airfoil (a modified NACA
64(4)-421) has been tested at a Reynolds number of 7.5 million (M = 0.24) and CL = 1.3 to
demonstrate that a large separation is problematic in this flow case. Five parameters have
been introduced to quantify the vortex generator design space (length, aspect ratio, angle,
chord location and spacings ratio) and Latin Hypercube sampling of this design space has
been undertaken using high fidelity CFD to construct a radial basis function based surrogate
model. This model has been interrogated and refined using an efficient adaptive sampling
method which incorporates both space-filling and local refinement properties in order to
achieve an optimum solution. To determine this optimum, a global search algorithm based
on differential evolution was used for optimization of the model. The results obtained from
the optimization eliminated the separation experienced by the clean geometry (from 11.1%
to 0%), whilst reducing the drag obtained from the best Latin Hypercube sample by 7%.

I. Introduction
Tiltrotor aircraft combine the vertical lift capability of a helicopter with the speed and endurance of
a conventional fixed-wing aircraft. In forward flight, the aircraft lift is provided by the wings. However,
due primarily to structural design considerations, these are very stiff, low-aspect ratio wings, which inhibit
efficient aerodynamic performance. Due to a short span and a typical cruise Mach between 0.4 and 0.5, these
wings experience a high loading; a typical cruise lift coefficient for a tiltrotor being above 1.0. Structural
requirements also result in a thick aerofoil. The Bell XV-15 utilised a NACA 64A223 in this regard.1 These
design factors therefore make tiltrotor aircraft particularly susceptible to stall and early onset buffet, so
boundary layer control is critical and upper surface flow control devices are an important consideration.
Flow control devices are adopted regularly in aerodynamic design, to control boundary layer behaviour,
and are often retro-fitted where problems are identified. Occasionally, complex active systems have been
implemented to control boundary layer growth, including boundary layer suction, blowing, and bleeding, but
by far the most common methods are passive schemes, with vortex generators (VGs) the most conventional
approach. These involve simple small flat plates normal to the surface, usually rectangular in shape, inclined
to the freestream flow. VGs were utilised on the Bell XV-15 tiltrotor to eliminate premature stall and retain
attached flow up to an angle of attack of 15 degrees.2 The V-22 Osprey tiltrotor also employs VGs; at
mid to high angles of attack, experimental data shows that wing-mounted VGs suppress the early onset of
separation.3 The majority of VGs protrude into the external flow, i.e. above the boundary layer, but there
have also been immersed, or sub-layer, VGs adopted.4 The role of a VG is to induce vortical flow via the
flow separation from its sharp upper edge. The vortex which subsequently propagates downstream entrains
∗ Researcher. Email: [Link]@[Link]
† Researcher. Email: [Link]@[Link]
‡ Professor of Computational Aerodynamics. Email: [Link]@[Link]
§ Lecturer. Email: [Link]@[Link]

1 of 19

American Institute of Aeronautics and Astronautics

higher energy, higher momentum flow from outside the boundary layer into the lower energy boundary layer.
This re-energises the boundary layer, suppressing separation.
It is necessary to understand the physics that the VGs are introducing into the flow to be able to design
them to mitigate against unwanted flow phenomena (e.g. separation), and this is particularly important on
tiltrotor wings due to the susceptibility to separation or buffet. To understand and quantify the effects of VGs,
computational fluids dynamics (CFD) approaches are often employed.5, 6 However, the modelling of VGs by
conventional CFD approaches poses difficult issues in capturing the vortex and its convection downstream.
This often requires fine numerical meshes on geometries where the size of the flow control device is orders of
magnitude smaller than the global object. Furthermore, a sufficiently high fidelity numerical scheme must
be selected, often requiring Reynold’s averaged Navier-Stokes (RANS) solutions with turbulence models.
In order to determine an optimized VG geometry, the high fidelity simulation approach can either be
coupled directly to an optimization routine or alternatively, it can be incorporated within a surrogate-
based analysis. The latter approach uses a limited number of high fidelity runs to train a surrogate model.
This model then provides interpolation of the known sample points to obtain a continuous design space
approximation away from the data sites. Using this surrogate-based approach allows the relatively cheap
design of the VG compared to the directly coupled approach. Design space interrogation methods can then
locate an optimum VG design without the need for large numbers of simulations. This surrogate modelling
approach for design of VGs has received only limited coverage in previous publications,7–9 including by the
use of lower-fidelity simulation approaches.10
The surrogate-based approach relies on limiting the number high fidelity sample points within the design
space, and thus efficient design space interrogation methods are desired. Two sampling strategies for con-
structing surrogate models were examined in Mackman et al 11 for construction of aerodynamic models. The
two methods employed were adaptive, i.e. information from successive iterations is used to reconstruct the
surrogate model between iterations. It was determined that both methods were suitable for aerodynamic
application, and both provided better than traditional one-stage or sequential space filling algorithms. An
alternative adaptive sampling method for surrogate-based modelling of multidimensional design space, com-
bines both space-filling properties and local refinement.12
A surrogate modelling approach was applied to tilt-rotor aircraft wings in Bevan et al.13 This earlier
work introduced a framework of statistical sampling and surrogate modelling, together with a preliminary
examination to highlight the potential of the approach. This initial examination utilised a simplified design
space, which was optimised for a single design point. The work presented in this paper refines the prior
framework and broadens the design space through both increasing variable range and also improving and
expanding on the parameters considered. Finally, adaptive sampling is introduced in order to augment
interrogation of the surrogate model and further optimise the resulting VG configuration.

II. Computational Modelling of Vortex Generators

The physics of a flowfield when a vortex generator (VG) is placed into the flow differs considerably from
a clean geometry. The VG produces a vortex which encourages mixing of turbulent flow downstream of the
device, often used to avoid unwanted separation. The exact physics of flow control using VGs is complicated
and often requires fine computational meshes with high fidelity flow simulations. The detail in this section
gives examples of the type of simulations used for VG modelling.

II.A. Mesh Generation

The precise physical effects and the topology of VGs makes for difficult modelling, both in set-up and in
solution methods. Multi-block structured meshes are difficult to produce for an aerofoil with a VG due
to the large differences in scale that exist between the parent aerofoil and the flow control device, and the
sudden changes in curvature when progressing over the surface of the aerofoil. Furthermore, to effectively
capture and convect downstream the vortex produced by the device, fine resolution grids are often required
around the VG and in its wake, resulting in the exponential growth of mesh sizes when compared to a clean
wing counterpart. Despite the difficulties in producing structured meshes, this meshing approach is generally
considered to produce the highest quality meshes that are most suitable to model flow around VGs and so
many examples can be found.14–19
To alleviate some of the issues with structured mesh generation, unstructured meshes can instead be

2 of 19

American Institute of Aeronautics and Astronautics

used. The advantage of unstructured meshes is that the algorithms for producing these meshes typically
use space-filling or marching approaches and so only require surface and farfield geometry, making their
production considerably simpler than multiblock structured meshes. However, to capture the physics of the
VG in the flow, a structured prismatic layer must exist on the surface. Examples include.20–22

II.B. Flow Solver

The suitable selection of a CFD solver is also imperative to ensure the complete convection of the vortex
downstream from its source without too much dissipation being added. The lowest fidelity simulation of
VGs that has been done is by Kerho and Kramer,23 who used XFOIL,24, 25 with a modification to the stress-
transport equation at the VG location to mimic the effect of a VG. This was demonstrated to reduce the
drag values near the separated region of an aerofoil.
In general, full flowfield analysis of a three-dimensional geometry must be performed to fully capture
the physics of the vortex generator in the flow, and this must be done by high fidelity RANS, or even LES
and DNS solutions. If using a RANS simulation approach then it is of paramount importance to select an
appropriate turbulence model to ensure detailed capture of the vortex. Two common turbulence models are
Spalart-Almaras (SA)26 and k-ω SST,27 which have both been used to model flow around a VG. Both tend
to under-predict streamwise vorticity of the vortex, though the SST model generally matches the experiment
better than SA.15, 16 This can be alleviated somewhat by including the strain-vorticity modification28 to the
SA model. Yaras and Grosvenor29 also compared both models with the k-ǫ and standard k-ω, and showed
that both SA and SST outperformed the the k-ǫ and k-ω models, but that SA perhaps demonstrated the best
balance between prediction, robustness and computational efficiency. The work of Tai5 utilised a modified
Baldwin-Lomax algebraic model, as well as Menter’s SST model as a supplement to ensure the physical flow
features were captured, on a V-22 Osprey simulation with VGs.
A lower-fidelity approach using embedded models within a CFD framework is also a popular approach.
Models such as the vortex-source30, 31 and lifting-force10, 32–34 have been shown to produce results that mimic
the effect of a vortex generator placed in the flow without the need to mesh and solve the full geometry.
To fully capture all the flow-physics associated with the VG, high-fidelity, real-time, simulations via
large eddy simulation (LES) or direct numerical simulation (DNS) methods are required, though these are
prohibitively expensive, particularly for design. As such, historically it has been difficult to perform such
simulations for large Reynolds numbers, however, with the continual increase in computational resources
available these are becoming more common.20, 35, 36 The Lattice-Boltzmann method37 has also been used to
produce time-dependent flow simulations around VGs.6

III. Simulation Approach

As outlined above, the modelling of the vortex generator and resulting flow is not a trivial task; the
suitable capture of the physics requires high fidelity solvers and high density numerical meshes. The primary
purpose of the research presented in this paper is the simulation of numerous VG configurations, in order to
assess the optimal design considerations. A solution method is firstly developed and presented in this section.
This involves obtaining high-fidelity viscous flows using OpenFOAMa with a bespoke mesh generation tool
to produce high quality structured viscous meshes around a representative aerofoil section with a vortex
generator.

III.A. Mesh Generation

All the meshes generated utilised the C-grid approach. A representative tiltrotor aerofoil, which is a modified
NACA 64(4)-421 section was used, with a 0.4% chord blunt trailing edge. A two-dimensional four-block
structured multiblock mesh was generated38, 39 initially, and three views are presented in figure 1; this is for
a mesh with 769 points on the aerofoil surface, 97 streamwise points on either side of the wake line, 65 points
across the blunt trailing edge, and 129 points in the normal direction.
To include the vortex generator, the two-dimensional mesh is extruded to the correct part-span position,
and a high quality mesh deformation scheme used, to rotate the mesh to the required vortex generator
a [Link]

3 of 19

American Institute of Aeronautics and Astronautics

incidence, and pull in mesh cells from upstream, downstream, above and either side of the generator to
match the required local spacings. This is achieved via several local radial basis function deformations.40, 41
Figure 2 shows the resulting wing surface mesh near the leading edge and the surface with starboard
plane, together with a closer view of the VG; this is for a sample point with 9.85% chord wide surface, with
rectangular vortex generator 1.88% × 1.31% chord length and height, with leading edge at 27.83% chord
from leading edge, 0.5% chord from port plane, at 23.75 degrees incidence to the freestream. The number
of spanwise planes extruded in dependent upon the spanwise domain width required. In this instance, 127
spanwise planes were extruded.

Figure 1: Four-block structured mesh

Figure 2: Surface mesh including vortex generator

The final stage in the mesh generation process reduces the cell count away from sections of interest.
This is undertaken through point removal and combining of hexahedra cells to form polyhedra. The mesh
reduction process typically involves a 50-60% cell count reduction.

III.B. CFD Solver

Once a suitable mesh has been produced, this is used to model the viscous flow around the wing with VGs
using OpenFOAM with a RANS model. OpenFOAM is an open source CFD software package, capable of
solving problems such as: complex fluid flows involving chemical reactions, turbulence and heat transfer,
solid dynamics and electromagnetics.
The flow simulations utilise a unstructured solver based upon the SIMPLE algorithm (Semi-Implicit
Method for Pressure-Linked Equations), an iterative procedure to solve the Navier-Stokes equations for
steady-state problems. This iterative procedure, as originally implemented in OpenFOAM, relies on basic
residual checking for steady state convergence criteria. After modification, steady state convergence is
assessed using the standard deviation of the force coefficients.
OpenFOAM has numerous turbulence models that can be utilised, although these models are not neces-
sarily considered standard implementations (e.g. SA-fv3). After modifications, this work utilises the standard
Spalart-Allmaras (SA) model as defined by NASA Langley’s Turbulence Modelling Resourceb. The standard
SA model implemented in this work uses the minimum limiter of 0.3 × Ω for Ŝ.
b [Link]

4 of 19

American Institute of Aeronautics and Astronautics

III.C. Example Solutions
The simulations were undertaken at a Reynolds number of 7.5 million, and a Mach number of 0.24. If
present, the VG was located at 11.25%. The distance between two counter-rotating pairs was 5.6% and
the intra-pair spacing was 0.5%. The VG height and aspect ratio (defined as length/height) were 1.16%
and 1.78 respectively, while the VG angle relative to freestream was 20.25 degrees. All percentages are
relative to the airfoil chord. No-slip conditions were applied to both the aerofoil and the VG. The farfield
was located at 100 chord lengths from the aerofoil. At the farfield, freestream boundary conditions were
imposed, together with a constant pressure at the downstream farfield. While the clean geometry (modelled
using a two dimensional mesh) utilises the two dimensional specific span-wise boundary condition within
OpenFOAM, the three dimension aerofoil and VG utilised slip conditions. Earlier work13 determined that
this span-wise boundary condition was suitable.
At a CL of 1.3 (9 degrees angle of attack for aerofoil with VG, 10 degrees for clean), the clean aerofoil
has a significant separated region, as shown in figure 3a.
-4 0.015
With VG With VG
Clean Clean
-3

0.01
-2
Cp

Cf
-1 0.005

0
1
0 0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8
x x

(a) Clean Geometry. Streamlines (b) Example VG. Cp Span Averaged (c) Example VG. Cf Span Averaged

Figure 3: Example solutions around modified NACA 64(4)-421

In contrast, the aerofoil with VG does not have a separation region. This separation is assessed by
examination of the x-direction friction coefficient (with x being the stream-wise direction), which is given
in figure 3c. From the figure, it is clear that the separation has been eliminated through introduction of
the VG, while the clean geometry has separation after 88% chord. The improved performance relative to
the clean geometry is also evident in the plot of the pressure coefficient (figure 3b), with the span-wise Cp
variation centred around the immediate downstream of the VG location and then diminishing.

IV. Vortex Generator Design Space

A high-fidelity simulation approach that utilises structured numerical meshes and viscous flow solutions
by OpenFOAM has been outlined and tested above for the simulation of flows around VGs on infinite wings.
The work in this paper is focussed on the process of performing effective design of the vortex generator (in
terms of its planform and location on the wing) and this process is outlined in this section. A full design
space of a number of design parameters is considered and sampled by the random Latin hypercube statistical
sampling method. These sample points are then simulated using the tools developed to allow investigation
of the effect of the design parameters on the flow solution around the VGs.

IV.A. Design Parameters

Five design parameters are investigated in this work. These are VG length (l), VG aspect ratio (AR), VG
setting angle relative to freestream (θ), VG chordwise location (c) and the spacing ratio (s). The VG aspect
ratio is defined as AR = l/h, and the spacing ratio is defined as the ratio between the intra-pair spacing
(x1 ) and the inter-pair spacing (x2 ), where the intra-pair spacing is fixed (x1 = 0.5%). The use of VG
aspect ratio, rather than VG height directly was based upon the work of Bevan et al.13 This earlier initial
investigation incorporated both VG height and length as individual parameters. Results obtained during
the study demonstrated that an AR < 1 produced undesirable flow with significant drag penalties. The five

5 of 19

American Institute of Aeronautics and Astronautics

parameters are given in figure 4.

(a) (b)

Figure 4: VG Configuration

Rectangular VGs are modelled in this work. The VGs are counter-rotating in nature. The five design
parameters were chosen to give a sufficient variation in the flow nature and final force coefficients to design
for a chosen objective function. The maximum and minimum values of the design parameters are shown
in table 1, and chosen to represent sensible coverage without including extremes of the design space. All
parameters are scaled such that the design space is [0, 1]5 .

Table 1: Design space range (% based on airfoil chord)

Parameter Minimum Maximum

h 0.2% 2%
AR 1 4
θ 5◦ 30◦
c 5% 50%
s 5 20

For tiltrotor aircraft wings, which are thick and highly loaded, early onset buffet can be particularly
problematic and is typically linked to separation. For this reason, buffet is one of the primary design
considerations in this work, and particularly separation at a chosen design condition. The loading on the
wings in cruise can be as high as CL = 1.0, however, it is unlikely that buffet in cruise is a problem. The
chosen design loading is therefore higher, and set to be CL = 1.3, which is realistic for a manoeuvre loading.
To test for possible buffet, a simulation around the clean aerofoil has been performed for this loading at
a representative flight condition (Re = 7.5 × 106 , M∞ = 0.24, CL = 1.3) and the flowfield is shown in figure
3a. As can be seen, separation at this flight condition is considerable and therefore it is probable that buffet
may also be an issue. The vortex generators are used to alleviate the buffet, which is assumed to be caused
by a large separation region. Therefore the objective function (equation 1), against which to design the
vortex generators, is to minimize separation subject to minimizing the drag penalty. This objective function
will balance drag induced by introduction of the VG and that produced via separation.

J = CD + kS (1)
J is the objective function to be minimized, S is the span-average separation chordwise location (as a
percentage), and k is a constant used to balance the effect of the two. To ensure equal magnitude of both
parameters, k is taken to be 0.0005 such that 1% separation equates to adding 5 drag counts to the problem.

6 of 19

American Institute of Aeronautics and Astronautics

IV.B. Latin Hypercube Sampling
To efficiently sample the design space, and reduce the number of function evaluations required from that of
brute force sampling, a Latin Hypercube42 (LHC) statistical sampling approach is used. A Latin Hypercube
sampling approach is a statistics-based approach designed to produce an optimum configuration of sample
points through the design space; the design space produced has each division in all dimensions sampled by
exactly one sample point. Corners of the sample space are added to the sampling as this is important for
interpolation, which is the basis of a surrogate modelling approach developed below. Whilst a random LHC
was utilised in,13 an optimised LHC offers an improvement in providing a more evenly sampled design space
for construction of the surrogate model. Numerous methodologies exist for construction of an optimised
LHC, however the detailing of these are beyond the scope of this paper. The approach chosen utilised the
Morris and Mitchell43 φp criteria
 1/p
m
X
φp (X) =  gj d−p
j
 (2)
j=1

where d represents the distance between two points in hypercube X and g represents the number of oc-
currences of this distance within the hypercube. d can be calculated either with the absolute norm or the
Euclidean norm. p is a positive integer. Minimizing φp improves the space-filling properties of the hyper-
cube. As recommended by Morris and Mitchell, this work minimizes phip for p = [1, 2, 5, 10, 20, 50, 100] then
the best of the resulting hypercubes is used. Since the possible combinations for a 5 dimension hypercube
with only 100 sample points is of the order of 10790 , an optimization algorithm is required. This work uses
the Enhanced Stochastic Evolution algorithm proposed by Jin et al.44 An example of the resulting Latin
Hypercube sample is shown in figure 5 for 100 sample points in two-dimensions plus the corner points.

1 1

0.8 0.8

0.6 0.6
y

0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x x

(a) Initial random LHC (b) Optimised LHC

Figure 5: Latin Hypercube sampled design space (plus corner points)

The Latin Hypercube used in this work consisted of 181 sample and corner points.

IV.C. Results
At all Latin Hypercube samples, a complete polar was run and interpolation used to obtain the angle, and
therefore drag, where the lift was the design lift coefficient (CL = 1.3). A brief statistical summary of the
results is shown in tables 2 and 3.
From table 2, the standard deviation relative to the mean is significant for both the separation (S) and
drag (CD ), although the objective function J has a marginally smaller standard deviation amongst the
points sampled. It is evident from table 3 that considering drag in isolation is not sufficient to reduce (and
eliminate) separation. Of the 181 samples, 24.5% have S < 0.1%, but only two (4.4%) of this reduced sample
have a CD < 0.0210. These two samples are spread throughout the design space, such that the distance

7 of 19

American Institute of Aeronautics and Astronautics

 Table 2: Design space results

Parameter Min Max Mean St. Dev

CD 0.0174 0.0784 0.0259 0.0098
S (%) 0.0 17.39 5.78 5.22
J 0.0204 0.0784 0.0288 0.0090

Table 3: Design space results

Parameter Value h AR θ c s
CD 0.0203
Min. J Separation 0.08% 0.82% 3.84 16.89◦ 55.6% 8.8
J 0.0204
CD 0.0784
Max. J Separation 0.00% 1.98% 2.98 27.00◦ 15.73.0% 10.17
J 0.0784
CD 0.0178
Min. CD Separation 10.39% 1.72% 2.58 22.75◦ 37.45% 18.05
J 0.0230
CD 0.0784
Max. CD Separation 0.00% 1.98% 2.98 27.00◦ 15.73% 10.17
J 0.0784
CD 0.0263
Max. S Separation 17.39% 1.89% 1.87 13.13◦ 59.45% 13.16
J 0.0349

between these points with the normalised [0, 1]5 design space is 0.85 which equates to 38% of the design
space range.

V. VG Design Space Modelling

A full design space investigation of a number of VG design parameters has been considered above using
high-fidelity viscous analysis. A Latin Hypercube sampled design space of the VG geometry and location
on the aerofoil has been performed. The next stage of this work is the design optimization of the VG based
on the design parameters that have been identified as being of significance. This is achieved by representing
the remainder of the design space (i.e. away from the already simulated design points) by an interpolation-
based surrogate model that is built around radial basis functions. This interpolation recovers the original
data and therefore preserves the high-fidelity nature of the solutions while simultaneously distributing that
information throughout the design space.

V.A. Surrogate Modelling

A surrogate model is an interpolation or approximation made from a small sample of data sites that give the
value of the design space at those know data sites. The model can be formulated in a number of different
ways, though the most common methods are by a polynomial, radial basis functions (RBFs) or Kriging. A
brief outline of the three methods is presented below, but reviews of surrogate modelling and surrogate-based
optimization have been presented,45–47 which the reader is guided to for more in-depth discussions on the
formulations and common uses for each approach.
The polynomial method48 involves fitting an order m polynomial through the given data, often by a least
squares fit. The resulting model is not guaranteed to recover any of the existing data, meaning it tends to
be more suited to fitting noisy data, where an underlying trend is the desired outcome instead of an exact

8 of 19

American Institute of Aeronautics and Astronautics

model. An RBF model is a linear combination of a number of functions that depend on the radial distance
from known data sites.49, 50 Unlike polynomial interpolation, RBF interpolation does provide exact recovery
at the known data sites, though interpolation of noisy data using RBFs can also be done using a regulation
parameter.51 Finally, the Kriging approach, pioneered by Krige52 and later used for surrogate modelling by
Sacks et al.53 and Jones et al.,54 is a special form of an RBF interpolation that uses a generalised Gaussian
equation as the basis function. The function allows variation in the response from variables to variable, and
is more statistically-based than a pure Gaussian function.
The surrogate model used in this work is developed around a multivariate interpolation using RBFs.
This interpolation method has the advantage of providing exact recovery of data at the known sites, thus
preserving the CFD data at the Latin Hypercube sampled points.
The general theory of RBFs is presented by Buhmann55 and Wendland,56 and the basis of the method
used here is described in detail by Rendall and Allen40 and surrogate specific detail in Mackman et al .11 Let
f (x) be the original function to be modelled, and fi be the scalar values at n discrete points xi , i = 1, ..., n,
where xi is the vector of inputs at the ith sample point in d-dimensional space xj , j = 1, ..., d. The set of
data points X = {x1 , ..., xn } is confined to a domain Ω in d-dimensional space. A RBF model is then a linear
combination of basis functions, whose argument is the Euclidean distance between the point x at which the
interpolation is made and the n points in the known data set. In other words, the interpolation at an untried
site is a sum of contributions from all the known function values, the influence of which is controlled by a
basis function that depends on the distance they are from the new site. If φ is the chosen basis function and
k · k is used to denote the Euclidean norm, then an interpolation model s has the form
n
X
s(x) = αi φ(kx − xi k) + p(x) (3)
i=1

where αi , i = 1, ..., n are model coefficients, and p(x) is an optional polynomial. The coefficients are found
by requiring exact recovery of the original data, sX = f , for all points in the training data set X . Hence
the model is an interpolant, and all original solution information is preserved. When the polynomial term is
included, the system is completed by the additional requirement
n
X
αi p(x) = 0 (4)
i=1

which is sometimes referred to as the side condition, for a polynomial that takes the form
d
X
p = γ0 + γj xj (5)
j=1

Setting up a global RBF interpolation then requires a solution to a linear system:

 
  γ0
0
γ 
 
.  1

 .. 
  !  ... 


0 0 P 
  
  = T
 γ d
 (6)
 f1  P M  
α
 
 ..  1
  
.  .. 
 
 . 
fn
αn
where
 
1 ··· 1  
 1 φ1,1 · · · φ1,n
x1 · · · x1n 

 . .. .. 
P =
 .. .. .  M =  ..
 . . 
 φi,j = φ(kxi − xj k)
. . .. 

φn,1 · · · φn,n
xd1 · · · xdn
Once the coefficients are known, the objective function at an interpolated point is then found by equation 3.

9 of 19

American Institute of Aeronautics and Astronautics

 Many choices are available for the basis function itself, but a radial conditionally positive definite function
is chosen to ensure the system has a unique solution. Basis functions are said to be compactly supported when
they decay to zero at a given distance from the centre, known as the support radius R. Compactly supported
functions have desirable numerical characteristics and modelling behaviour, hence the work presented here
concentrates on Wendland’s functions56 φd,k , which are compact functions of minimal degree for a stated
number of continuous derivatives C 2k in d dimensions. Wendland’s C 2 (φd,1 ) in up to three dimensions is
given by

φ3,1 = (1 − r)4+ (4r + 1), (7)

where r = k · k denotes the norm and (·)+ refers to the cut-off function
(
1 − r, 1 − r ≥ 0,
(1 − r)+ = (8)
0, 1 − r < 0.
The support radius used in the function also acts to control the region of influence of each of the centres.
A larger support radius allows each sample point to influence the interpolation at a given point from a
greater distance away, and in general leads to a smoother interpolation. For a compact function, it defines a
hypersphere around each sample point outside of which the value of φ is zero, and inside of which the point
has some influence. The support radius scales the basis function argument as follows:
r
rscaled =
, (9)
R
hence the distance between two points is then the scaled Euclidean distance
v
u d
1u X
kx − xi kscaled = t (xj − xj,i )2 . (10)
R j=1

V.B. VG Surrogate Model

The surrogate model is constructed using Wendland’s C2 function with a support radius sufficient to cover
the entire design space. Selected projections for the objective function J are presented in figure 6 for the
design space [0, 1]5 .

VI. VG Design Optimization

The use of a surrogate model allows very quick evaluation of an interpolation of an otherwise expensive,
high-fidelity design space. The high fidelity nature of the data used to produce the interpolation is, however,
preserved by building a surrogate model using RBFs. The next stage of this work is the optimization of the
VG design based on the five parameters outlined as being important to VG design.
One option for the VG design optimization is to use the simulation approach developed in this work
within an optimization framework, which is known as full-fidelity aerodynamic shape optimization. This
approach, however, is expensive as very large numbers of simulation evaluations are required during the
process. A less expensive approach, and one that exploits the use of the surrogate model, is surrogate-based
optimization and is the approach used in this work. This optimization approach substitutes a full simulation
evaluation to obtain an objective function value for a given set of inputs with a surrogate model evaluation
to obtain an interpolated approximation to the true value of the objective function for the given inputs. The
surrogate evaluation is simply given by equation 3, and represents a very cheap way of obtaining responses
to the output for the given inputs.
To interrogate the surrogate design space, an optimization algorithm is required. One choice of optimiza-
tion algorithm is a gradient-based method, however, these are prone to termination in a local minimum that
is not necessarily the global minimum solution. Gradient-based approaches are well suited to optimization
problems with expensive objective function evaluations, such as full-fidelity aerodynamic shape optimization,
for example.57–59 When multimodality (multiple local minima) is an issue, a global optimization algorithm
is often more effective at locating the globally optimal solution. These algorithms tend to be agent-based,
such as particle swarm60 or gravitational search,61 or can also be evolutionary-based, such as differential

10 of 19

American Institute of Aeronautics and Astronautics

 1 1

0.8 0.8

0.6 0.6
AR

AR
0.4 0.4

0 0
0.2 0.2
0.2 0.2

0 0.4 0 0.4
0 0
0.2 0.2
0.6 0.6
0.4 0.4

h 0.6
0.8
0.8 h 0.6
0.8
0.8

1 1 1 1

(a) Projection at c = 0, s = 0 (b) Projection at c = 1, s = 0

1 1

0.8 0.8

0.6 0.6
h

0.4 0.4

0 0
0.2 0.2
0.2 0.2

0 0.4 0 0.4
0 0
0.2 0.2
0.6 0.6
0.4 0.4

c 0.6
0.8
0.8 c 0.6
0.8
0.8

1 1 1 1

(c) Projection at AR = 0, s = 0 (d) Projection at AR = 1, s = 0

1 1

0.8 0.8

0.6 0.6
h

0.4 0.4

0 0
0.2 0.2
0.2 0.2

0 0.4 0 0.4
0 0
0.2 0.2
0.6 0.6
0.4 0.4

c 0.6
0.8
0.8 c 0.6
0.8
0.8

1 1 1 1

(e) Projection at AR = 0, s = 1 (f) Projection at AR = 1, s = 1

Figure 6: Surrogate model projections (blue represents lower objective, J)

11 of 19

American Institute of Aeronautics and Astronautics

evolution.62 These global methods usually require many more objective function evaluations compared to
gradient-based methods and are therefore suited to problems that have cheap objective function evaluations.
An objective function evaluation in this work is represented by a surrogate evaluation, which is very
cheap to evaluate. Furthermore, due to the unknown nature of the modality in the design space, a global
optimization algorithm has been developed for optimizing the VG surrogate design space. The global al-
gorithm that has been developed is based on the differential evolution algorithm presented by Storn and
Price,62 and is outlined in section VI.A.
In addition to searching the surrogate design space for an optimum solution directly, a complementary
approach interrogates the surrogate model in order to both refine regions of interest and ensure the accuracy
of the surrogate design space. Regions of interest in this instance being local minima within the design space.
An efficient method of improving surrogate-based accuracy and refinement is adaptive sampling. Adaptive
sampling relies on reconstructing/updating the surrogate model across a number of iterative cycles by updat-
ing the surrogate model with newly determined adaptive sample points. A kriging based adaptive strategy
was compared to a similar RBF adaptive strategy in Mackman et al.11 These utilise the primarily space
filling MaxMSE (largest value of error predicted by the mean squared error function, otherwise referred to as
entropy)63 and a statistical optimization function (expected improvement function). Expected improvement
balances local and global search criteria.54 An alternative RBF adaptive approach that incorporates both
local gradient-based refinement and space-filling properties has also been developed.12 This approach is
presented in section VI.B.

VI.A. Differential Evolution

Differential evolution (DE) is a swarm intelligence algorithm built around the concept of evolutionary me-
chanics. Like the other swarm intelligence algorithms DE takes a number of agents (parameter vectors)
which are often termed chromosomes, and follows three steps to advance the optimization algorithm:

1. Mutation: generate a mutated candidate solution by small, random changes

2. Crossover : ‘breeding’ of mutated and parent solutions to produce a child
3. Selection: accept the new child based on its fitness

The parameter vectors are vectors of length d, where each entry represents a position in each of the d
design parameters:

xn = {x1n , x2n , . . . , xdn }T (11)

The mutation stage involves the production of a new, candidate solution to introduce variability and
exploration into the algorithm. The candidate/mutated solution for the n-th parent/target solution is called
the donor vector, vn , and is produced by combining three existing parameter vectors:

vn = xr1n + F (xr2n − xr3n ) (12)

where r1 , r2 and r3 (r1 6= r2 6= r3 6= n) are random integers chosen from the range [1, N ], for a population
of N parameter vectors (or agents), for each parent vector, and F is a scalar that is typically in the interval
[0, 1] and is chosen to either maximize or minimize exploration away from the existing agents.
The crossover stage is used to enhance diversity in the population by combining aspects of the given
parent and donor vectors. For the d-th entry in the parent (xdn ) and donor vectors (vnd ), the d-th entry in
a trial/child vector (udn ) is produced where each entry is from the donor vector based on a probability CR,
otherwise it is from the parent vector. The ensure there is guaranteed variability and that un gets at least
one entry from vn , a random entry from vn in always introduced into un .
The final stage of the algorithm for each of the parameter vectors is to assess whether the new child is
of sufficient fitness to replace the given parameter vector. If the fitness of the child vector is better than its
parent then the child replaces the parent.
Multiple further additions have been presented for DE, and in the review by Neri and Tirronen,64 many
of these were tested for cost and performance comparison. A small addition that is simple to make to DE to
considerably improve its performance is population size reduction (PSR),65 which splits the overall number
of iterations of the optimizer into a number of stages where the number of parameter vectors are reduced by

12 of 19

American Institute of Aeronautics and Astronautics

a half at each stage. Only the parameter vectors with the best fitness are kept after each stage which has the
effect of managing the loading between exploration with many vectors at the beginning of the optimization,
and exploitation towards the end of the optimization.
The DE algorithm was run with constants commonly found in the literature; 50 particles, crossover
probability 0.3, mutation rate 0.7, 300,000 function evaluations and 4 stages for the PSR.

VI.B. Adaptive Sampling

An adaptive sampling method requires an initial set of samples to first build an interpolation, and then
updated sampling points are added based on a compromise between space-filling updates and local refinement
in non-linear regions. This non-linearity is determined by the Laplacian. A smooth separation function
quantifies the sample spacing in order to assess space-filling updates. A criterion, C, was proposed in12
which merges these two criteria:

C = (|∇2 J| + ǫ)(1 − h)2 0≤h≤1 (13)

where ǫ is a small number, sufficient to avoid C = 0 when |∇2 J| = 0. The variable h is the space-filling
function, such that 1 − h grows away from previously sampled locations. However, this criterion C was
proposed in order to improve information around both maxima and minima within the surrogate model
(hence the magnitude of the Laplacian). If only a minimum is desired (positive Laplacian) or maximum
(negative Laplacian) then the C function is adjusted appropriately. Larger values of C indicate new potential
sample locations. Either a grid evaluation can be conducted to determine maximum C, and hence the next
sample location, or a more robust and efficient strategy utilising the global optimization algorithm outlined
in the previous subsection is undertaken.
The space-filling (or separation) function, h, is defined as a smooth function, for blending with the local
refinement component of C. In this function, h = 1 at a sample point, and diminishes as distance increases
away from the sample point. This is produced using an RBF kernel with a compactly supported basis
function and small support radius. The support radius used in this work was determined as a function of
two distances, the minimum distance (d1 ) and suprenum minimum distance (d2 ) between sample points.
p
Srh = Rh (d1 d2 ) (14)
where Rh is a scaling factor, which in this work was Rh = 1. However, it was found that the adaptive
sampling approach outlined above was insufficient to progressively refine the solution without significant
adaptive sample points. Therefore the adaptive sampling approach was modified with the addition of an
exploitative term

Cb = (|∇2 J| + ǫ)(1 − h)2 b3 (15)

where b is a distance based function and 0 ≤ b ≤ 1. The b function is also smooth and produced via an RBF
kernel in the same manner as h. Instead of using all sampled points, the function merely utilises a subset of
pbest fitness values taken from the full known sample set, where p is the number of samples in the subset.
The support radius for the b function is the maximum distance between a subset sampled point to the full
sample set. The b function will be have a value of b = 1 at any subset sample point and will decay away from
the subset of p points to zero at the furthest reaches of the domain (relative to the subset points). Thus
the original adaptive sampling approach can be seen as primarily explorative (when b = 1 throughout the
design space), whilst the latter is exploitative (0 ≤ b ≤ 1). This can be demonstrated using two analytical
functions in the following subsections.

VI.B.1. Franke’s Function

Franke’s function is a bivariate function, consisting of two Gaussian peaks of differing heights and a smaller
trough. In this instance, the adaptive sampling is undertaken using a grid evaluation seeking only the
maxima location, with the trough refinement ignored. The function is defined using

13 of 19

American Institute of Aeronautics and Astronautics

 (9x1 − 2)2 (9x2 − 2)2 (9x1 + 1)2
   
9x2 + 1
f (x) = 0.75 exp − − + 0.75 exp − −
4 4 49 10
(16)
(9x1 − 7)2 (9x2 − 3)2
 
− 0.2 exp −(9x1 − 4)2 − (9x2 − 7)2


+0.5 exp − −
4 4
between xi ∈ [0, 1].
After an initial 9 sample points, 24 adaptive sample points are introduced (sequentially, rather than in
batches). The results for using both adaptive sampling techniques is given in figure 7.

(a) C Approach (2d) (b) C Approach (3d projection)

(c) Cb Approach (2d) (d) Cb Approach (3d projection)

Figure 7: Comparison of C and Cb adaptive sampling approaches using Franke’s Function

From the figure, it is clear that the Cb approach is highly concentrated around the primary peak. Less
obviously, it is also more rapid in identifying the exact peak location during the adaptive cycle.

VI.B.2. Rastrigin Function

The two-dimensional variant of the n-dimensional Rastrigin function is a more complex problem, with a
plethora of local minima and maxima. It is defined by

f (x) = −An − Σni=1 x2i − A cos (2πxi )

(17)

14 of 19

American Institute of Aeronautics and Astronautics

where A = 10, n = 2, xi ∈ [−5.12, 5.12]. The global maximum (f (x) = 0) is located at x = 0. After an
initial 16 sample points, 24 adaptive sample points are introduced sequentially. As shown in figure 8, both
adaptive approaches capture the overall global maximum, although the Cb function has again clearly covered
all potential local maxima within proximity.

-5 -5
-15 -15
-25 -25
-35 -35
-45 -45
-55 -55
-65 -65
-75 -75

Z
Y

(a) C Approach (2d) (b) C Approach (3d projection)

-5 -5
-15 -15
-25 -25
-35 -35
-45 -45
-55 -55
-65 -65
-75 -75

Z
Y

(c) Cb Approach (2d) (d) Cb Approach (3d projection)

Figure 8: Comparison of C and Cb adaptive sampling approaches using Rastrigin’s Function

Both adaptive strategies offer advantages, with C providing an initial adaptive coverage, and Cb refining
the final solution.

VI.C. VG Optimization Results

The results of optimizing the VG surrogate model for the objective function given by equation 1 are given
in table 4, along with the clean values and the values that are the lowest objective values from the latin
hypercube sample. To increase the adaptive sampling rate, a number of updates per iteration were undertaken
in parallel. This only requires updating the separation function before determination of the next update
location. The number of updates per iteration and the number of iterations was prescribed in advance. Two
rounds of explorative adaptive sampling using five samples in each round were followed by one round of
exploitative sampling.
The results clearly demonstrate that the optimized value is a considerable improvement in the separation
experienced on the aerofoil and therefore a reduction in buffet may be expected. The optimization process

15 of 19

American Institute of Aeronautics and Astronautics

via differential evolution has also improved the performance of the VG compared to just taking the best
sample tested in the latin hypercube, therefore vindicating the optimization process. Comparing to the
clean geometry, it is particularly impressive that the process developed here, of design space sampling,
surrogate-based optimization and adaptive sampling, has been able to eliminate separation at a highly
loaded condition and achieved a decrease in the drag on the aerofoil of 7% relative to best result achieved
by the Latin hypercube sampling alone.

Table 4: Design space results

J CD Separation h AR θ c s
Clean 0.0246 0.0190∗ 11.1%∗ - - - - -
◦
Min. LHC 0.0204 0.0203 0.08% 0.81% 3.84 16.89 55.6% 8.8

Optimized (Surrogate) 0.0189 – –
0.85% 3.64 8.42◦ 21.5% 5.8
Optimized (CFD) 0.0189 0.0189 0.0%
∗
oscillatory result: time-averaged

Streamlines proximal to the VG and trailing edge of the airfoil are given in figure 9. From the figure,
the large separation region present in the clean geometry has been entirely eliminated using the optimum
VG configuration; see figure 3a for the clean geometry streamlines. Figure 9c shows the vortex extent, as
defined by the λ2 = 0 criterion between the minimum LHC sample and the optimum solution. This criterion
denotes the intermediate eigenvalue of the symmetric tensor S2 + Ω2 where Ω = 12 [∇u − (∇u)T ] is the
vorticity tensor and S = 12 [∇u + (∇u)T ] the rate of strain tensor. Negative values of the λ2 criterion denote
the vortex. The induced vortex of the optimum design has eliminated separation at the trailing edge of the
airfoil, as shown in figure 9b, although a small region is still present in the minimum LHC sample.

(a) Opt VG streamlines (b) Opt TE streamlines (c) Min. LHC and Opt, λ2 = 0

Figure 9: Induced vortex impact downstream of VG

The high performance of the optimization algorithm can be demonstrated by considering the value of
the gradients of the objective with respect to changes in the design variables at the optimized point. An
optimum solution for this type of problem is one where the gradient values approach a very small number.
Table 5 shows these gradient values and demonstrates the high performance of the optimization algorithm,
producing a solution that is optimum to within a very small tolerance, and an improvement on the minimum
sample from the Latin Hypercube, which is not at an optimum solution as defined by the gradient.

16 of 19

American Institute of Aeronautics and Astronautics

 Table 5: Gradients at minimum LHC sample and optimum solution

Min. LHC Optimized

∂J/∂h 6.2×10−3 -2.3×10−10
∂J/∂AR -1.3×10−2 1.2×10−9
∂J/∂θ 3.0×10−3 -3.8×10−11
∂J/∂c 3.2×10−3 7.1×10−10
∂J/∂s -4.2×10−3 1.5×10−9

VII. Conclusions
The quantification of the effects of flow around a tiltrotor-like aerofoil with vortex generators (VGs) and
subsequent design of these VGs has been considered. The highly loaded and thick wing of a tiltrotor is
particularly susceptible to early onset buffet, so boundary layer control is critical and, hence, upper surface
flow control devices are an important consideration. An analysis package has been presented that uses
the open source library, OpenFOAM, to obtain high-fidelity viscous flow simulation, and a custom mesh
generator to autonomously obtain high quality structured meshes.
Design of counter rotating VGs has been considered using, firstly, a Latin Hypercube sampling approach
to obtain design space samples for five variables; VG height, VG aspect ratio, VG angle, chord-wise loca-
tion, and inter- and intra-pair spacing ratio. A surrogate model that uses radial basis functions has been
developed to produce an approximation to a high-fidelity design space but still preserves the important flow
physics required for design. An optimization algorithm based on differential evolution, which is a global
search algorithm, was used to fully interrogate the surrogate model design space. With the initial sample
points defined via an optimised Latin Hypercube, adaptive sampling using a combined objective function
Laplacian/space-filling criterion was then employed in order to locate a global optimum VG design for the
high loading manoeuvring condition. The objective function blends the requirement for drag minimization
with reduction of separation.
The best Latin Hypercube result obtained reduced the separation from 11.1% to 0.08%, but increased
the drag by 6.8% relative to the clean geometry. Following the surrogate-based design optimization, the
optimum result obtained has eliminated separation entirely at CL = 1.3 whilst decreasing the drag by 7%
relative to the minimum Latin Hypercube sample and with a marginal decrease in drag relative to the clean
geometry.

Acknowledgements
The authors would like to thank David Tring and Nigel Scrase of AgustaWestland, for their support of
and contribution to this work. The financial support, via the HiPerTilt project, of AgustaWestland and
Innovate UK (Technology Strategy Board) is also gratefully acknowledged. Finally, the authors would like
to acknowledge that this work was carried out using the computational facilities of the Advanced Computing
Research Centre, University of Bristol.

References
1 M.D. Maisel, D.J. Giulianetti, and D.C. Dugan. The history of the XV-15 tilt rotor research aircraft: From concept to

flight. Technical Report SP 2000-4517, NASA, 2000.

2 M.D. Weiberg, J.A. Maisel. Wind-tunnel tests of the XV-15 tilt rotor aircraft. Technical Report TM 81177, NASA.
3 J. Abras and R. Narducci. Analysis of cfd modeling techniques over the mv-22 tiltrotor. In 66th American Helicopter

Society International Annual Forum, Phoenix,Arizona, 2010.

4 C-S. Yao, J. C. Lin, and B. G. Allan. Flow-field measurements of device-induced embedded streamwise vortex on a flat

plate. In 1st AIAA Flow Control Conference, St. Louis, Missouri, 2002. AIAA Paper 2002–3162.
5 T.C. Tai. Effect of midwing vortex generators on v-22 aircraft forward-flight aerodynamics. Journal of Aircraft, 40(4):623–

630, 2003.
6 B. Konig, E. Fares, and S. Nolting. Fully-resolved lattice-boltzmann simulation of vane-type vortex generators. In 7th

AIAA Flow Control Conference, Alanta, Georgia, 2014. AIAA Paper 2014–2795.
7 A. Jirasek. Design of vortex generator flow control in inlets. Journal of Aircraft, 43(6):1886–1892, 2006.

17 of 19

American Institute of Aeronautics and Astronautics

 8 Y. Kuya, K. Takeda, X. Zhang, and A. I. J. Forrester. Multifidelity surrogate modeling of experimental and computational

aerodynamic data sets. AIAA Journal, 49(2):289–298, 2011.

9 N Namura, S. Obayashi, and S. Jeong. Efficient global optimization of vortex generators on a super critical infinite-wing

using kriging-based surrogate models. In 52nd Aerospace Sciences Meeting and Exhibit, National Harbor, Maryland, 2014.
AIAA Paper 2014–904.
10 JS. Yi, C. Kim, and B. J. Lee. Adjoint-based design optimization of vortex generator in an S-shaped subsonic inlet.

AIAA Journal, 50(11):2492–2507, 2012.

11 T. J. Mackman, C. B. Allen, M. Ghoreyshi, and K. J. Badcock. Comparison of adaptive sampling methods for generation

of surrogate aerodynamic models. AIAA Journal, 51(4):797–808, 2013.

12 T. J. Mackman and C. B. Allen. Investigation of an adaptive sampling method for data interpolation using radial basis

functions. International Journal for Numerical Methods in Engineering, 83:915–938, 2010.

13 R. L. T. Bevan, D. J. Poole, C. B. Allen, and T. C. S. Rendall. Simulation and surrogate-based design of rectangular

vortex generators for tiltrotor aircraft wings. In 33rd AIAA Applied Aerodynamics Conference, Dallas, Texas, 2014.
14 J. W. Hamstra, D. N. Miller, and P. P. Truax. Active inlet flow control technology demonstration. In 22nd International

Congress of Aeronautical Sciences, pages Harrogate, UK, 2000. ICAS-2000-6.11.2.

15 B. G. Allan, C-S. Yao, and J. C. Lin. Numerical simulations of vortex generator vanes and jets on a flat plate. In 1st

AIAA Flow Control Conference, St Louis, Missouri, 2002. AIAA Paper 2002–3160.
16 E. Wik and S. T. Shaw. Numerical simulation of micro vortex generators. In 2nd AIAA Flow Control Conference,

Portland, Oregon, 2004. AIAA Paper 2004–2697.

17 J. Dandois, V. Brunet, P. Molton, J.-C. Abart, and A. Lepage. Buffet control by means of mechanical and fluidic vortex

generators. In 5th AIAA Flow Control Conference, Chicago, Illinois, 2010. AIAA Paper 2010–4975.
18 F. K. Lu, A. J. Pierce, Y. Shih, C. Lin, and Q. Li. Experimental and numerical study of flow topology past micro vortex

generators. In 40th AIAA Fluid Dynamics Conference and Exhibit, Chicago, Illinois, 2010. AIAA Paper 2010–4463.
19 Q. Li and C. Liu. Declining angle effects of the trailing edge of a microramp vortex generator. Journal of Aircraft,

47(6):2086–2095, 2010.
20 C. Memory, D. O. Snyder, and J. Bons. Numerical simulation of vortex generating jets in zero and adverse pressure

gradients. In 46th AIAA Aerospace Sciences Meeting and Exhibit, Reno, Nevada, 2008. AIAA Paper 2008–558.
21 Y. Ito, M. Murayama, and K. Yamamoto. High-quality unstructured hybrid mesh generation for capturing effects of

vortex generators. In 51st AIAA Aerospace Sciences Meeting including the New Horizons Forum and Aerospace Exposition,
Grapevine, Texas, 2013. AIAA Paper 2013–554.
22 K. J. Forster and T. R. White. Numerical investigation into vortex generators on heavily cambered wings. AIAA Journal,

52(5):1059–1071, 2014.
23 M. Kerho and B. Kramer. Enhanced airfoil design incorporating boundary layer mixing devices. In 41st AIAA Aerospace

Sciences Meeting and Exhibit, Reno, Nevada, 2003. AIAA Paper 2003–211.
24 M. Drela. Xfoil: An analysis and design system for low reynolds number airfoils. In T. J. Mueller, editor, Low Reynolds

Number Aerodynamics, pages 1–12. Springer Berlin Heidelberg, 1989.

25 M. Drela and M.B. Giles. Viscous-inviscid analysis of transonic and low reynolds number airfoils. AIAA Journal,

25(10):1347–1355, 1987.
26 P. R. Spalart and S. R. Allmaras. A one-equation turbulence model for aerodynamic flows. Recherche Aerospatiale,

1:5–21, 1994.
27 F. R. Menter. Two-equation eddy-viscosity turbulence model for engineering applications. AIAA Journal, 32(8):1598–

1605, 1994.
28 J. Dacles-Mariani, G. G. Zilliac, J. S. Chow, and P. Bradshaw. Numerical/experimental study of a wingtip vortex in the

near field. AIAA Journal, 33(9):1561–1568, 1995.

29 M. I. Yaras and A. D. Grosvenor. Evaluation of one- and two-equation low-Re turbulence models. part II–vortex-generator

jet and diffusing s-duct flows. International Journal for Numerical Methods in Fluids, 42(12):1321–1343, 2003.
30 W. G. Kunik. Application of a computational model for vortex generators in subsonic internal flows. In
AIAA/ASME/SAE/ASEE 22nd Joint Propulsion Conference, Huntsville, Alabama, 1986. AIAA Paper 1986–1458.
31 N. E. May. A new vortex generator model for use in complex configuration CFD solvers. In 19th AIAA Applied

Aerodynamic Conference, Anaheim, California, 2001. AIAA Paper 2001–2434.

32 E. E. Bender, B. H. Anderson, and P. J. Yagle. Vortex generator modeling for Navier-Stokes codes. In American Society

of Mechanical Engineers, 1999. FEDSM 99-6929.

33 R. V. Chima. Computational modeling of vortex generators for turbomachinery. In ASME TURBO EXPO, Amsterdam,

Netherlands, 2002. GT-2002-30677.

34 A. Jirasek. Vortex-generator model and its application to flow control. Journal of Aircraft, 42(6):1486–1491, 2005.
35 S. Lee, E. Loth, and H. Babinsky. Normal shock boundary layer control with various vortex generator geometries.

Computers and Fluids, 49(1):233–246, 2011.

36 H. Shan, L. Jiang, C. Liu, M. Love, and B. Maines. Numerical study of passive and active flow separation control over a

naca0012 airfoil. Computers and Fluids, 37(8):975–992, 2008.

37 H. Chen, S. Kandasamy, S. Orszag, R. Shock, S. Succi, and V. Yakhot. Extended boltzmann kinetic equation for turbulent

flows. Science, 301(5633):633–636, 2003.

38 C. B. Allen. CHIMERA volume grid generation within the EROS code. Proceedings of the Institution of Mechanical

Engineers, Part G: Journal of Aerospace Engineering, 214(3):125–141, 2000.

39 C. B. Allen. Towards automatic structured multiblock mesh generation using improved transfinite interpolation. Inter-

national Journal for Numerical Methods in Engineering, 74(5):697–733, 2008.

18 of 19

American Institute of Aeronautics and Astronautics

 40 T. C. S. Rendall and C. B. Allen. Unified fluid–structure interpolation and mesh motion using radial basis functions.

International Journal for Numerical Methods in Engineering, 74(10):1519–1559, 2008.

41 T. C. S. Rendall and C. B. Allen. Efficient mesh motion using radial basis functions with data reduction algorithms.

Journal of Computational Physics, 228(17):6231–6249, 2009.

42 M. D. McKay, R. J. Beckman, and W. J. Conover. A comparison of three methods for selecting values of input variables

in the analysis of output from a computer code. Technometrics, 21(2):239–245, 1979.

43 M. D. Morris and T. J. Mitchell. Exploratory designs for computational experiments. Journal of Statistical Planning

and Inference, 43:381–402, 1995.

44 R. Jin, W. Chen, and A. Sudjianto. An efficient algorithm for constructing optimal design of computer experiments.

Journal of Statistical Planning and Inference, 134:268–287, 2005.

45 D. R. Jones. A taxonomy of global optimization methods based on response surfaces. Journal of Global Optimization,

21(4):345–383, 2001.
46 N. V. Queipo, R. T. Haftka, W. Shyy, T. Goel, R. Vaidyanathan, and P. K. Tucker. Surrogate-based analysis and

optimization. Progress in Aerospace Sciences, 41:1–28, 2005.

47 A. I. J. Forrester and A. J. Keane. Recent advances in surrogate-based optimization. Progress in Aerospace Sciences,

45:50–79, 2009.
48 G. E. P. Box and N. R. Draper. Empirical Model-Building and Response Surfaces. Wiley, 1987.
49 D. S. Broomhead and D. Lowe. Multivariable function interpolation and adaptive networks. Complex Systems, 2(3):321–

355, 1988.
50 A. Sobester, S. J. Leary, and A. J. Keane. On the design of optimization strategies based on global response surface

approximation models. Journal of Global Optimization, 33(1):31–59, 2005.

51 T. Poggio and F. Girosi. Regularization algorithms for learning that are equivalent to multilayer networks. Science,

247(4945):978–982, 1990.
52 D. G. Krige. A statistical approach to some basic mine valuation problems on the witwatersrand. Journal of the Chemical,

Metallurgical and Mining Society of South Africa, 53(5):159–162, 1952.

53 J. Sacks, S. B. Schiller, and W. J. Welch. Designs for computer experiments. Technometrics, 31(1):41–47, 1989.
54 D. R. Jones, M. Schonlau, and W. J. Welch. Efficient global optimization of expensive black-box functions. Journal of

Global Optimization, 13(4):455–492, 1998.

55 M. Buhmann. Radial Basis Functions. Cambridge University Press, 1st edition, 2005.
56 H. Wendland. Scattered Data Approximation. Cambridge University Press, 1st edition, 2005.
57 J. E. Hicken and D. W. Zingg. Aerodynamic optimization algorithm with integrated geometry parameterization and

mesh movement. AIAA Journal, 48(2):400–413, 2010.

58 F. Bisson, S. K. Nadarajah, and D. Shi-Dong. Adjoint-based aerodynamic optimization framework. In 52nd AIAA

Aerospace Sciences Meeting, National Harbor, Maryland, 2014. AIAA Paper 2014–0412.
59 G. K. W. Kenway and J. R. R. A. Martins. Multi-point high-fidelity aerostructural optimization of a transport aircraft

configuration. Journal of Aircraft, 51(1):144–160, 2014.

60 J. Kennedy and R. Eberhart. Particle swarm optimization. In 1995 IEEE International Conference on Neural Networks,

Perth, Australia, 1995.

61 E. Rashedi, H. Nezamabadi-pour, and S. Saryazdi. GSA: A gravitational search algorithm. Information Sciences,

179:2232–2248, 2009.
62 R. Storn and K. Price. Differential evolution - a simple and efficient adaptive scheme for global optimization over

continuous spaces. Technical report, ICSI, UC Berkeley, 1995. TR-95-012.

63 M. C. Shewry and H. P. Wynn. Maximum entropy sampling. Journal of Applied Statistics, 14(2):165–170, 1987.
64 F. Neri and V. Tirronen. Recent advances in differential evolution: a survey and experimental analysis. Artificial

Intelligence Review, 33:61–106, 2010.

65 J. Brest and M. S. Maucec. Population size reduction for the differential evolution algorithm. Applied Intelligence,

29(3):228–247, 2008.

19 of 19

American Institute of Aeronautics and Astronautics