See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/269062082

Airfoil Optimisation by Swarm Algorithm with Mutation and Artificial Neural

Networks

Conference Paper · January 2009

DOI: 10.2514/6.2009-1278

CITATIONS READS

32 2,062

3 authors, including:

Manujit Khurana Hadi Winarto

Chandigarh University RMIT University
10 PUBLICATIONS   136 CITATIONS    17 PUBLICATIONS   243 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Ph D Thesis for Robert Caresse on aerodynamic optimization View project

All content following this page was uploaded by Hadi Winarto on 22 February 2017.

The user has requested enhancement of the downloaded file.

 Airfoil Optimization by Swarm Algorithm with
Mutation and Artificial Neural Networks

Manas S. Khurana1, Hadi Winarto2 and Arvind K. Sinha3

The Sir Lawrence Wackett Aerospace Centre – RMIT University, Melbourne, VIC, 3000

The process of aerodynamic shape optimisation requires the development of

intelligent models to address the stipulated design goals. The Direct Numeric
Optimisation (DNO) approach is examined in this paper, which analyses the
feasibility of a shape, in iteration until convergence based on defined objectives and
constraints. The method is computationally intensive hence the components of the
DNO architecture are defined, validated and modified to generate an efficient
search optimisation model. Efficiency is enhanced by mapping the solution space for
High-Altitude Long Endurance (HALE) airfoil design problem, through an inverse
mapping of PARSEC airfoil shape variables over a series of benchmark profiles.
Solution regions with aerodynamically infeasible shapes are identified and
eliminated from the search process, to reduce computational time. A single-point
airfoil optimisation with Gradient-Based method, over the defined search space is
examined. Variations in base airfoils confirmed the solution space is highly
multimodal and gradient methods merely locate the local optima. A Particle Swarm
Optimisation (PSO) algorithm incorporating a double-mutation operator to mitigate
local optimal solutions, for highly multimodal solution topologies was defined and
validated. The swarm algorithm for airfoil shape optimisation confirmed the limited
search flexibility of gradient methods, by establishing a global solution with a 16%
reduction in drag. The swarm algorithm is computationally intense for shape
optimisation. An Artificial Neural Network (ANN) is developed and validated with a
relationship between the mapped PARSEC solution space and the aerodynamic
coefficients of lift and drag established. A network sensitivity study indicated a
double-layered network with 30 neurons for lift and 20 for drag is required to
establish the aerodynamic coefficients with acceptable accuracy. The surrogate
model is used for airfoil shape optimisation by replacing the flow solver from the
DNO loop. Time savings are established with the aerodynamic performance of the
output solution in line with the results of the direct PSO-Flow solver combination.
Neural network simulations for fitness function approximation are prone to errors.
Hence, future research will focus on developing a hybrid search methodology by
integrating the flow solver and ANN in the DNO approach.

1
PhD Candidate, The Sir Lawrence Wackett Aerospace Centre – RMIT University, 850 Lorimer Street,
Port Melbourne, VIC 3000, Australia – AIAA Student Member
2
Associate Professor, School of Aerospace, Mechanical and Manufacturing Engineering – RMIT
University, GPO Box 2476V, 3001, VIC, AIAA Member
3
Director Aerospace and Aviation, The Sir Lawrence Wackett Aerospace Centre – RMIT University, 850
Lorimer Street, Port Melbourne, VIC 3000, Australia - AIAA Member.

1
American Institute of Aeronautics and Astronautics
 I. Introduction

T HE development of intelligent search models are required for shape optimisation routines. The
methodology needs to be in the form of direct or inverse design principles. In the direct search
approach, based on design objectives and constraints, shapes are ‘intelligently’ examined by iteration for
convergence. In the inverse approach, the design requirements are transformed into a set of velocity profiles
and shape conforming to these requirements is established. In either of the approach, the complexity of the
search process is governed by the performance and shape constraints. To illustrate, a stipulated flight
performance for a designated structural layout, defines the geometrical constraints. In the design of a rotor
blade, high maximum static and dynamic lift coefficients is required for flight at high maneuver load
factors.3 Thus, operation at high mach numbers is investigated and an airfoil with high Mach divergence
number stipulates the design requirement3. Shape design iterations include aerodynamic and geometrical
features. Minimum wing volume at low drag performance over the flight envelope is a typical design
objective for shape optimisation and this may result in computationally intense search process. The
intensity is governed by the parameters of design requirements and constraints. Thus, the development of
an efficient shape optimisation architecture is critical for the design process.
The issues and challenges associated with the design principles of direct and inverse architectures needs
to be addressed. In the inverse approach, generation of un-realistic airfoil shapes for the specified flight
performance is a viable option.4 To illustrate, ‘fish-tail’ airfoils addresses the issue and requires pressure
coefficient re-specification, to establish an operating profile as an alternative.4 A combination of direct and
indirect approach is attempted to address this issue. A constraint on the trailing edge thickness is
introduced, to mitigate ‘cross-over’ airfoils.4 Thus, a small portion of the planform is directly specified,
with the remaining airfoil contour inversely designed through the specified pressure coefficient.4 The
inverse approach requires stringent aerodynamic and geometrical constraints to mitigate undesirable
features. The initial specified pressure coefficient is continuously iterated for the pressure profile to produce
an airfoil conforming to the required constraints. Depending on the degree-of-modification required to the
initially specified pressure coefficient to address these constraints, the operating requirements for the final
profile will differ to the specified. Indirect methods such as the hodograph approach are not recommended
due to the limited flexibility it provides in the application of aerodynamic and geometrical constraints.4
Thus, for single-point and multi-point airfoil optimization, the use of indirect design methods is
impractical.
Direct optimization approach as an alternate to the inverse design methodology has been extensively
applied.5-8 The method provides enhanced control over the application of geometrical and aerodynamic
constraints7, over the inverse methodology. Variables within the airfoil shape parameterizer impose
geometrical constraints and the shape design coefficients needs to be proportional to airfoil geometry. The
optimiser is further applied to impose aerodynamic constraints, in terms of minimum lift requirements. The
search process is computationally intensive. The optimiser initiates the search process through random
permutation of airfoil geometry variables to compute the aerodynamic features through a validated flow
solver. The probability of the initial population conforming to the stipulated aerodynamic constraints is
low, since the optimiser has no prior knowledge of viable design solutions. Hence, extensive computational
resources are applied for computation on shapes that are not aerodynamically feasible. This integration of
‘intelligent’ search agents dampens the effect over the iteration process. As the search progresses, the
search agents converge to a design solution region, characterised by viable design solutions. Thus, the
probability of shapes conforming to user defined objectives increases as the search progresses, but at
additional computing expense. To address the demerits of the direct method, the requirement of developing
an efficient optimisation algorithm coupled with a surrogate model that approximates the flow solver is
required.
The integration of Navier-Stokes solutions as the flow solver within the direct search approach, results
in computationally enhanced simulations. Gradient Methods (GM) have been used for airfoil shape
optimisation6, 9 and offer computational merits in comparison to evolutionary programming techniques.
Namgoong reported that the design space of transonic airfoils has several local minima and the final GM
solution is influenced by the initial starting point. Solution initialisation through variations in starting base
airfoils was investigated. The GM solution provided local solutions in comparison to the developed Genetic
Algorithm (GA) that reported global solutions, but at a higher computational cost.6
The prospect of reducing the computational time of global search models through the integration of a
surrogate model requires further analysis. An Artificial Neural Network has the potential of reducing the

2
American Institute of Aeronautics and Astronautics
computational efforts of direct search methods significantly, as it duplicates the role of the flow solver. An
ANN model is developed to approximate the aerodynamic coefficients which are integrated into an
Evolutionary Algorithm (EA) search method for shape optimisation. The proposed methodology was
attempted by Duvigneau et. al in the design of a single-point airfoil through the integration of a GA
model.10 Significant computational time savings were observed in comparison to a stand-alone GA
optimisation process with the design requirement of maximising lift and minimising drag performance.10 It
provides for an avenue for further research in the design development of ANN models for single and multi
point airfoils for Multi-Mission UAV platforms.
The paper is presented in six sections as follows:
Sec. II: The modules comprising the DNO approach are defined. The requirements of a geometrical
shape parameterizer, the optimisation strategy and the surrogate model are defined in the context of airfoil
shape optimisation;
Sec. III: The solution space is defined by mapping of PARSEC shape variables over similar long-
endurance airfoils;
Sec. IV: A Particle Swarm Optimiser which utilises a mutation operator that circumvents local
solutions is presented and validated over a series of test functions;
Sec V: The development and validation of a surrogate model to replace the flow solver, for reducing the
computational time associated with global search simulation in the direct design approach; and
Sec. VI: Results are presented and the research roadmap the design optimisation of airfoils for Multi-
Mission UAVs is discussed.

II. Problem Definition

The DNO comprises of three key modules of study as follows: a) Airfoil geometry parameterisation for
shape representation; b) Validated flow solver for aerodynamic analysis and; c) Search agent to establish
optimal shapes based on user defined objectives and constraints. The three components operate iteratively
for solution convergence. The modules require development and experimental validation prior to
integration in the DNO structure. Airfoil shape representation and the development of a validated search
agent are examined in the test development case. XFOIL is used as the flow solver for early results.

A. Airfoil Shape Function – Mapping Solution Space & Identifying Important Design Variables
Geometry parameterization was tested by measuring the flexibility and accuracy of several analytical
shape functions.11 A symmetrical airfoil was used as the starting section and required to geometrically
converge to a set of benchmark airfoils.11 The magnitude of the design coefficients in the analytical
function for shape convergence, were established with a swarm optimiser.11 The Hicks-Henne methodology
satisfied the stipulated design merits of flexibility and accuracy. Equating the design coefficients directly to
airfoil geometry was not viable hence, the PARSEC shape representation12 was examined.13 The function
indicated similar degree-of-flexibility and accuracy in comparison to the Hicks-Henne model.13 The method
provided added flexibility, with each geometrical design variable theoretically related to airfoil geometry.
This relationship was experimentally proven by examining the effect of perturbing each variable on airfoil
geometry and aerodynamics.14
The operating range of each design variable represents the solution space to the optimisation problem.
If the solution region is restricted, the real optimum will not be attained. Alternatively, if the solution space
is vast, aerodynamically infeasible shapes will be analysed, thus increasing computational time. The ability
of the neural network to generalise, with the presence of outliers (these solutions are regarded as statistical
outliers) in the training database, is a design issue. Thus, an appropriate definition of the solution space is
needed, without compromising the ability to establish an optimal shape.
Neural network development further benefits from the mapping of PARSEC variables. The distribution
of the training database is collated from a restricted design space, instead of an ill-defined search limit. Pre-
screening of design space, prior to network development is vital for overall network success. It provides
acceptable network generalisation performance. Thus, mapping of PARSEC shape coefficients addresses
one of the issue and challenge of neural network structure development (Sec. III).

B. Airfoil Shape Optimizer Type & Implementation of a Surrogate Model

The optimizer type significantly affects the search capabilities, computational efficiency and validity of
the final solution. If the algorithm is limited in the search process, then a sub-optimal solution is

3
American Institute of Aeronautics and Astronautics
established. Gradient-Methods provide rapid convergence but isolate the search to local regions. This issue
is addressed by correctly defining the initial starting point. Thus, the effect of varying solution initialisation
point with the use of long-endurance airfoils is examined with a GM algorithm (Sec. IV).
Global search algorithms are an alternate to GM tools. The exhaustive search methods are also sensitive
to local search regions, particularly for highly multimodal test problems, if not correctly tuned. This was
proven with a standard PSO model, when used for optimising a series of ten-dimensional test functions
(Sec. IVB). The search process of the swarm algorithm requires modification to mitigate this demerit.
Introduction of mutation operators to diversify the search agents is proposed and validated (Sec. IV). The
modified version is then used for airfoil shape optimisation with performance merits measured directly
against GM solutions.
The feasibility of replacing the flow solver in the optimisation cycle, by an ANN model is examined to
address the intense computation simulations of the swarm algorithm. The neural network architecture is
governed by the number of hidden layers and neurons and type of activation function between the layers.
The optimum configuration of the network is problem dependent thus an experimental study is required to
verify an acceptable structure. Variations in training population size and its effect on network generalisation
to data independent of training, is to be addressed (Sec. V).

III. Mapping of Solution Space

The PARSEC shape variables define the solution space to the problem. Hence, a methodology to map
an acceptable search region is required. The airfoil design optimisation focuses on the development of long-
endurance planforms, with operation at low lift coefficients. Comparable airfoils designed with similar
performance goals were identified as follows:

Natural Laminar Flow Airfoil (NLF): A series of airfoils developed by NASA to provide maximum
lift capability at low-speeds with low-drag performance associated with NACA 6-series of airfoils;15

NASAs 17-Percent-Thick Low-Speed Airfoil: Designed to provide higher lift-drag ratio during the
climb and higher maximum lift coefficient at low-speeds, without additional wing surfaces in
comparison NACA airfoils;16

Low-Reynolds Number Airfoil (LRN 1015): Implemented in the RQ-4A Global Hawk platform for
high-altitude, long-endurance performance for ISR (Intelligence, Surveillance and Reconnaissance)
roles;

LRT 17.5 Airfoil: Designed specifically for a HALE sensor-craft platform for high altitude (20kms)
and endurance performance in excess of 40 hours with loiter speeds of Mach 0.6 over a 24 hour limit;
and17

Thick-Airfoil for HALE UAV (TH 25816): Airfoil with 25% thickness designed for HALE operations
with acceptable lift and drag performance at Mach 0.30 and moderate-to-low Reynolds numbers18

The eleven design coefficients in the PARSEC methodology are manipulated to represent disparate
classes of airfoils (Fig. 1 and Table 1)

Table 1. PARSEC Airfoil Design Parameters

Description Symbol Description Symbol

Leading Edge Radius RLE Lower Crest Curvature ZXXLO
Upper Crest Abscissa XUP Trailing Edge Ordinate ZTE
Upper Crest Ordinate ZUP Trailing Edge Thickness ∆ZTE
Upper Crest Curvature ZXXUP Trailing Edge Direction αTE
Lower Crest Abscissa XLO Trailing Edge Wedge Angle βTE
Lower Crest Ordinate ZLO Figure 1. PARSEC Airfoil

An inverse-fitting of PARSEC shape variables is required to determine the magnitude of the geometry
design coefficients over disparate classes of airfoils. The test forms an optimisation problem, with the aim
of minimising the sum of geometrical difference at each ordinate i , between the theoretical ( y / c) it arg et and
approximated PARSEC airfoil ( y / c) iapprox (Eq. 1). A swarm algorithm (Sec. IV) was implemented to
determine the magnitude of the PARSEC coefficients with the following objective function:

4
American Institute of Aeronautics and Astronautics
  
∆ ƒ min ∑ ( y / c )i − ( y / c )i  (1)
 approx t arg et 

Where:
∆f = Objective Function; ( y / c )iapprox = Approximated Airfoil Profile; and ( y / c )it arg et = Target Airfoil Profile
In conjunction with the identified long-endurance airfoils, the scatter of PARSEC coefficients was
further established by performing geometry convergence of airfoils with disparate flight performance
categories as follows:

High-speed operations with Mach Number in excess of 0.70 – Sections examined include the RAE2822
and High-Speed Natural Laminar flow airfoils HSNLF(1)-2013;

Symmetrical NACA 0012 airfoil: Used in helicopter rotor blade applications; and

NACA 5-6 series airfoils: NACA 23015 exhibits acceptable maximum lift performance with NACA
632615 providing acceptable low-drag at moderate lift coefficients

The inclusion of airfoils under disparate performance goals were included with the intent of
categorising the PARSEC coefficients based on unique flight objectives. Consequently, it is hypothesized
that each geometry coefficient will be grouped in accordance to the stipulated flight performance goal. To
illustrate, airfoils designed for operations at transonic speeds will have pronounced lower and upper crest
curvature magnitudes (ZXXLO & ZXXUP) and categorised over a specific operating
tolerance ( Z XXLOmin − Z XXLO max & Z XXUPmin − Z XXUP max ) . The importance of setting the lower and upper
crest shape curvature, in the design of transonic airfoils was reported by Winnemöller et.al.19, with the
coefficients affecting shock development over the airfoil surface. Operations at low loiter-speeds will be
characterised by a unique set of dominating design coefficients at these operations. The design limits of the
lower and upper crest curvatures is characterised by different operating tolerances, as the role of controlling
the airfoil contour curvature is not significant at such speeds. Thus, the aim of the shape convergence study
is to group the design coefficients based on specific flight performance requirements. Visual inspection by
observing the scatter of different groups will assist in the formulation of search limits for each variable,
based on low-speed HALE flight performance requirements.
The PARSEC design coefficients, as a result of the inverse-fitting shape convergence, over the
identified test airfoils are presented in Table 2. At this stage, the trailing edge thickness was omitted from
the analysis and fixed at zero. The results provide the best achievable shape convergence while considering
limitations in the optimiser and the geometrical function. The PARSEC methodology is limited to a class of
airfoils that it can generate. Limitations in the optimiser coupled with the geometrical flexibility tolerance
of the PARSEC representation, result in select airfoils that do not exhibit ‘exact’ shape closure. The leading
edge radius design coefficients for the NLF(1)-0115 and NLF(1)-0414 airfoils (Table 2) reflect this
limitation, where acceptable closure about the leading edge radius is not achievable.
Table 2. Test Airfoils PARSEC Coefficients
PARSEC Coefficient NACA 0012 NLF(1)-0115 NLF(1)-0414 RAE2886
rle 0.0147 0.0095 0.0105 0.0083
XUP 0.3015 0.4468 0.4808 0.4312
ZUP 0.0599 0.1194 0.0957 0.0629
ZXXUP -0.4360 -0.7844 -0.7237 -0.4273
XLO 0.2996 0.2774 0.4368 0.3441
ZLO -0.06 -0.0341 -0.0477 -0.0588
ZXXLO 0.4406 0.1810 0.3859 0.7018
ZTE 0 0 -0.0015 0
αTE (deg) 0 -18.82 -9.60 -6.86
βTE (deg) 14.67 9.74 3.5 8.08

The disparity in PARSEC design coefficients in the four test airfoils (Table 2) is related to the
operating performance of the airfoils. The leading edge radius spreads between the four planforms (Table
2). The low-speed, natural laminar flow airfoils are characterised by larger leading edge radii and are within
10% of each other. In comparison, the higher speed RAE2822 airfoil is characterised by a lower leading
edge radius, attributed to maintaining flow attachment about the region at higher speeds. The upper crest

5
American Institute of Aeronautics and Astronautics
curvature ZXXUP, for the two natural laminar flow airfoils is comparable. Alternately, the NACA 0012
profile, which encounters local shocks in rotary wing applications, illustrates a significantly larger
magnitude. The size is comparable to the RAE2822 planform operating at similar speeds, where the effect
of ZXXUP mitigates the adverse effects of wave drag as a result of shock development.
Airfoils sharing common flight goals and performances are clustered in similar groups, thus the spread
of PARSEC coefficients are assumed to be normally distributed about the mean profile for a specific flight
condition. In the case of designing long endurance airfoils, the NLF(1)-0414 planform is used as the
benchmark section in this analysis. The airfoil is designed with a performance metric that closely matches
the design goals of the project; maximising lift-to-drag ratio under loiter conditions at low Mach and
moderate Reynolds numbers. By setting the NLF(1)-0414 section as the mean and computing the standard
deviation of the tested airfoils, a normal distribution about the mean provides an estimation of the search
limits of the PARSEC variables.
Leading Edge Radius Solution Mapping Trailing Edge Wedge Angle Solution Mapping
50 3.5
µ = NLF(0)414
45
3
40 µ = NLF(0)414

35 Upper Limit: r 2.5 Lower Limit: tew

Lower Limit: r le
le
30 ϑµ,σ2(tew)
ϑµ,σ2(rle)

2
µ - 1σ µ + 1σ
25 34.1% 34.1%
µ - 1σ µ + 1.64σ
20 1.5
34.1% 45%
HSNLF(1)-2013; Upper Limit: tew
15 NACA 23015; TH 25816 1
10 µ - 2σ µ + 2σ NACA 23015;
13.6% 13.6% µ - 2σ µ + 2σ
0.5 TH 25816
5 13.6% 13.6%
0 0
-0.005 0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
rle tew
Figure 2. Normal Distribution about PARSEC Figure 3. Normal Distribution about PARSEC
Leading Edge Radius Trailing Edge Wedge Angle
The distribution of the PARSEC leading edge radius and trailing edge wedge angle over the tested
airfoils is presented in Figures 2 and 3 respectively. The cluster of airfoils, either side of the mean in the
leading edge radius distribution map (Fig. 2), are airfoils belonging to Natural Laminar Flow database. One
standard deviation about the mean is sufficient to cover test required airfoils. Two standard deviations less
than the mean, will result in airfoils with negative leading edge radius; a physically impractical solution.
Conversely, two standard deviations added to the mean will include airfoils which are not directly related to
the stipulated performance requirements of the project. This includes airfoils designed for operation at high
Mach numbers to a planform with acceptable maximum lift coverage but at higher drag count in terms of
the NACA 23015 airfoil. Though the TH 25816 airfoil is designed specifically for HALE operations, a
maximum thickness in excess of 25 percent of the airfoil is not in the stipulated design requirement of the
project thus, the solution space is not required. The design limits for rle, are bounded by the region
characterised by one standard deviation from the mean airfoil. This accounts for approximately 68% of the
airfoil population set.
The trailing edge wedge angle was similarly examined (Fig. 3) with HALE airfoils located either side of
the mean planform. On the lower side of the design spectrum, one standard deviation away from the mean
accounts for the remaining HALE airfoils in the cluster. Two standard deviations provide an insufficiently
large design space and the mapping indicates that no airfoil is clustered in this region. Thus, one standard
deviation is adequate in the definition of the lower limit of tew. The upper limit spectrum is set to the
threshold at the point coinciding with the NACA 0012 planform. Two standard deviations, will expand the
search space, but will not provide airfoils conforming to HALE requirements. This is further supported by
the data indicating that two standard deviations, does not provide sufficient wedge angle design coverage
for the NACA 23015 and TH 25816 airfoils. Similar to the findings reported in the leading edge radius
analysis (Fig. 2), the aerodynamics goals of the NACA 23015 do not match directly the HALE goals of this
project. Conversely, the TH 25816 is geometrically not a viable option due to its increased thickness.
Each PARSEC design variable was mapped following the procedure outlined in this section (Fig. 2 &
3) and is presented at Table 3. There are no set benchmarks which govern a set of procedures to effectively
map the design space for airfoil optimisation. Instead, user intuition is required to establish a sound
judgment based on the distribution of the variables in the cluster (Fig. 2 & 3). This was detailed in the case

6
American Institute of Aeronautics and Astronautics
of the leading edge radius, where two standard deviations were producing negative radii sections. Selection
of a different airfoil as the mean and taking the standard deviation about this section, will produce different
design limits. In this case, an airfoil which best matched the design goals was identified as the benchmark
planform. User intuition will best address the suitability of a given airfoil as the mean section based on
defined objectives. The analysis outlined in this section adequately addresses the requirement of defining
the solution space for airfoil shape optimisation (Sec. V). The practice of leaving the design variables open,
without pre-screening the solution space will result in an inefficient computation process due to the
examination of airfoils that do not conform to user defined objectives.

Table 3. PARSEC Airfoil Search Region

αTE βTE
rle XUP ZUP ZXXUP XLO ZLO ZXXLO ZTE
(deg) (deg)
Lower
0.0055 0.3573 0.088 -1.0294 0.336 -0.071 -0.0486 -0.02 -20.51 1.15
Bound
Upper
0.0215 0.6043 0.1194 -0.418 0.5376 -0.057 0.8204 0.02 1.32 14.73
Bound

IV. Airfoil Optimisation using Local and Global Search Models

In this section, optimisation algorithms for shape optimisation are defined. The search process governed
by the proposed methods are characterised by local and global search patterns. Gradient-Methods yield
rapid convergence but limit the search to local regions. The effect of initialising the solution with disparate
airfoils is adverse and this is further investigated. The PSO algorithm theoretically mitigates local sub-
optima solutions, at high computation expense. Prior to applying the swarm algorithm for airfoil shape
optimisation, validation tests are executed over a series of mathematical test functions. The verification
process provides an avenue to identify and rectify issues in the PSO search process. Based on the validation
process results, the code is modified to address the identified issues. Thus, an overall efficient algorithm
results in the use of DNO approach for airfoil shape optimisation.

A. Gradient-Based Optimisation
A Gradient-Search optimisation is performed using MATLABs optimisation toolbox. The Sequential
Quadratic Programming (SQP) method is used as the search algorithm. At each iteration, the Quadratic
Programming (QP) function is solved, with an estimate of the Hessian of the Lagrangian updated using the
BFGS formula at each iteration.20
Similar to the analysis performed by Namgoong in the design optimisation of transonic airfoils6, the
effect of varying the initial airfoil, for HALE planforms at subsonic speeds is investigated. The spread of
results with disparate starting points will provide an indication of the topology for solution space and the
presence of local minima.
A single-point optimisation problem is formulated, with the objective of minimising drag at cruise. The
problem is defined as follows:
min( F ) = min[(C L − C LT ) 2 + C D ] (2)
α = 2 ° ; Re = 3.0 × 10 6 ; Mach = 0.35; and C LT = 0.40
The base airfoils used in the test are as follows: a) Symmetrical subsonic NACA 0012 profile; b)
Natural Laminar Flow section designed for low-speed operations, NLF(1)0115; and c) RAE 2822 cambered
airfoil designed to reduce wave drag at transonic speeds.
The airfoil shape and fitness function iterations are presented in Figures 4-5 for the independent test
airfoils. The evolution of airfoil shape during the search is superimposed to illustrate shape changes during
the iteration phase.
Airfoil optimisation using the symmetrical section is illustrated in Figure 4. Changes in geometry are
observed between initial and final planforms (Fig. 4a). Significant fitness reduction is evident during
iterations 1-4 (Fig. 4b). This is attributed to the optimiser establishing the lift coefficient of the airfoil
towards the specified target. Drag performance improvements attribute to further fitness reduction after the
fourth iteration.

7
American Institute of Aeronautics and Astronautics
 Airfoil Shape Evolution using Gradient-Based Optimiser with NACA 0012 Gradient Method Optimisaton Fitness Convergence with NACA 0012
Base Section Base Section
1
10
NACA 0012
0.1
Iteration 1
Iteration 2

Airfoil Shape Optimisation (f)

0
Final Iteration 10
0.05

-1
y/c

0 10

-0.05 -2
10

-0.1
-3
10
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 2 4 6 8 10 12 14
x/c Iteration
(a) (b)
Figure 4. Airfoil Shape Design Using Gradient Method with NACA 0012 Base Airfoil
The use of the NLF(1)-0115 airfoil as the base section, results in changes in airfoil shape over the
iteration process that are negligible (Fig. 5a) in comparison to NACA 0012 (Fig. 4a). The fitness
convergence further shows minimal activity (Fig. 5b) in comparison to the symmetrical profile (Fig. 4b).
The NLF airfoils are designed specifically for the conditions outlined by the objective function (Eq. 2). The
PARSEC variables mapping (Sec. III) identified NLF airfoils, as benchmark sections in the design
optimisation of HALE planforms; sections which exhibit favorable aerodynamic performance for the
conditions specified by the objective function (Eq. 2). Theoretically, the NLF profile is a solution to the
design problem and the results provided by the GM analysis (Fig. 5) supports this. The NACA 0012 airfoil
is designed for different operating conditions, thus the optimiser indicated greater search activity (Fig. 4) in
comparison. Consequently, the improvements provided by the gradient-based method include matching the
lift coefficient of the profile as required with minimal drag reduction.
Airfoil Shape Evolution using Gradient-Based Optimiser with NLF(1)-0115 Gradient-Based Optimisation Fitness Convergence with NLF(1)-0115
Base Section Base Section

0.1 NLF(1)-0115
Airfoil Shape Optimisation Fitness (f)

Iteration 3
Final Iteration -2
10
0.05
y/c

-0.05

-0.1
-3
10
0 0.2 0.4 0.6 0.8 1 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6
x/c Iteration
(a) (b)
Figure 5. Airfoil Shape Design Using Gradient Method with NLF(1)-0115 Base Airfoil

Solution using the RAE 2822 airfoil as the base section is presented in Figure 6. The contour features of
the airfoil are similar between initial and final sections (Fig. 6a). The end airfoil is thicker in comparison,
with the remaining shape features resembling the initial profile (Fig. 6a). Despite the notable difference of
shape between initial and iteration 1, the profile exhibits similar shape contour till convergence (Fig. 6a).
The magnitude of the fitness reduction at the commencement of the search process is attributed to the shape
change forced by the optimiser to conform to user specified objectives (Eq. 2). The overall magnitude of
shape change between initial and final airfoil is significant in comparison to the symmetrical (Fig. 4a) and
NLF profile (Fig. 5a). This is reasonable, as the RAE 2822 airfoil is designed for conditions which are on
the extreme-side of the operating envelope in terms of speed. The objective defined in this study proposes a
section operating at low sub-sonic speeds. Thus, the optimiser exhibits significant changes to the airfoil
profile in an attempt to transpose a high-speed airfoil to a low-subsonic section.

8
American Institute of Aeronautics and Astronautics
 Airfoil Shape Evolution using Gradient-Based Optimiser with RAE 2822 Gradient-Based Optimisation Fitness Convergence with RAE 2822
Base Section Base Section
0
10
RAE 2822

Airfoil Shape Optimisation Fitness (f)

Iteration 1
0.1 Iteration 4
-1
Final Iteration 10

0.05
y/c

-2
10
0

-3
10
-0.05

-4
-0.1 10
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 2 4 6 8 10 12 14 16 18
x/c Iteration
(a) (b)
Figure 6. Airfoil Shape Design Using Gradient Method with RAE 2822 Base Airfoil
B. Particle Swarm Based Optimisation
An Adaptive-PSO (APSO) algorithm developed by Qin. et.al21 was validated, by examining the effect
of varying particle velocity, on solution convergence over a series of test function.14 The solution
convergence was forced by setting a low maximum iteration count, as a function of dimension space.14 A
series of experiments were formulated to illustrate that the maximum velocity of the particles must be
limited to 0.1 × The length of search space, for effective convergence based on the defined termination
criteria. The results established were favorable in comparison to the data presented by Qin. et al,21 where
the maximum velocity was set to the size of the search space.
The validation process is extended to test the capability of the optimiser in converging to the theoretical
solution for a benchmark test problems. Two test functions, Schwefel and Michalewics are identified for
the validation test. The solution space is characterised with diverse topological features which are ideal to
verify the robustness and flexibility of the swarm algorithm (Eq. 3-4). The Schwefel function comprises of
several peaks and troughs and is characterised by a local minima that is located far from the global solution.
Consequently, it is reported that many algorithms become trapped in this region.2 The global solution is
located near the intersection of the hypersurfaces of different dimensions. Thus, locating the global minima,
at such an extremum is a design challenge. The Michalewics function is highly multimodal and is
characterised by the presence of local minima’s which are distributed evenly in the solution space. A
multimodal design topology promotes a sub-optimal local solution to the problem, as an output of the
optimisation process for models with limited search capabilities. Thus, the two test models provide an
acceptable test bed in the design, validation and verification of the swarm algorithm to mitigate local
solutions as a solution output.
A literature survey provided an indication for a suitable number of iterations required for convergence
based on algorithm testing over a series of test functions.1, 2, 22, 23 The domain size of the test models are set
to ten-dimensions n, to duplicate the population size of the PARSEC geometry variables, in the context of
airfoil shape optimisation. Thus, the maximum iteration count for the Schwefel and Michalewics function
was set to 10,000 and 20,000 respectively to provide the search agents adequate time for convergence to the
theoretical solution. Comparatively, iterations were fixed to 1,000 for a ten-dimensional test conducted for
the APSO model, where the aim was to establish the effect of particle velocity on the search capabilities of
the model over a limited search phase. This further reinforces the aim of this investigation; verify the true
search capabilities of the model over an extended time span.
Test information for the two functions is summarised in Table 4. The particles are initialized away from
the theoretical solution with intent, to examine the flexibility and robustness of the search process in
locating the theoretical solution of the model (Table 4). A swarm of 10 particles are used in the validation
tests.

9
American Institute of Aeronautics and Astronautics
 Schwefel Function
n
f ( x) = 418.9829n − ∑ ( x sin
i =1
i xi ); n = 10
(3)

Michalewics Function
2
f ( x1 , x2 ) = − ∑ sin( x )[sin( jx
j =1
j
2
j / π )]2 n ; n = 10 (4)

Table 4. PSO Set-Up for Test Function Evaluation and Summary of Theoretical Solution
Dimension Search Space Initialization Global Minima
Function
(i1,2,...,n) Range (Theoretical Solution)
Schwefel −500 ≤ x i ≤ 500, i = 1,2, … , n 250 ≤ x i ≤ 500 X * = (420.96871 , … ,420.968710 ), f ( X * ) = 0
Michalewics 0 ≤ xi ≤ π , i = 1,2, … , n 0 ≤ xi ≤ 0.1 X * = (2.201 , … , 1.5710 ), f ( X * ) = −9.66

Simulations using the APSO model indicated local optima results (Table 5). Thus, acceptable
convergence to the theoretical solution for the two test functions (Table 5) was not attainable. With a local
solution as the output, it was obvious that further design modifications are required to the APSO model.
Lack of search diversity on the part of the optimiser attributes to the establishment of sub-optimal solutions.
Mutation operators which are used successfully in genetic algorithms are introduced to overcome this
shortfall. A double-mutation operator is proposed and is set-up as follows:
1. Compute the Euclidean distance ds, thus fitness between the two extreme particles in swarm
2. First Mutation Operator: Compute Probability of Mutation (Pm) by:
ds
Pm = (5)
No. of Particles
3. Generate Matrix: No. of Particles × 1, of random numbers (randi), between 0 and 1
4. If Pm > randi, then mutate the particle/s by:
xik,mutate = xik + [ gaussian(σ )] (6)
Where:
xik,mutate = Mutated Position of Particle/s i at iteration k
xit = Current Position of Particle/s i at iteration k
σ = Random Number Generated from Gaussian Distribution
Mean zero & Standard Deviation of 0.1 × The Length of Search Space
5. Update personal Pbest ,i and global Pbest , g best of solution in the swarm
6. Second Mutation Operator: If Pbest , g stagnates for N successive iterations, then apply the following
mutation to Pbest,i of each particle:
xik, stagnation = Pbest ,i + [ gaussian(σ )]
mutate
Where:
xik, stagnation = Updated Mutated Position of Particles i at iteration k as a Result of Stagnant Pbest,g
mutate

Pbest,i = Personal Best of Each Particle

σ = Random Number Generated from Gaussian Distribution
Mean zero & Standard Deviation of 0.01 × The Length of Search Space
The simulation results of the proposed mutation PSO model, M-PSO is presented in Table 5 for the two
test functions. Significant convergence improvements are observed in comparison to APSO which does not
include particle re-initialisation through mutation. In both test functions, the proposed M-PSO model
converges directly to the theoretical solution. The analysis by Liang et. al1 and Chen et.al2 provided a
comprehensive comparative analysis between different PSO variants used in the optimisation of Schwefel

10
American Institute of Aeronautics and Astronautics
and Michalewics functions respectively. Models that consistently converge to the theoretical solution were
reported1, 2 and are identified in Table 5. For Schwefel, a Comprehensive Learning PSO, CLPSO1 was
identified and an Hybrid PSO with an External Optimiser EO with local-search capabilities2 was reported
for Michalewics function. The M-PSO, in the analysis of the Schwefel function, consistently provided
efficient convergence towards the theoretical solution with fewer iterations required in comparison to
CLPSO1 (10,000 versus 30,000 in Table 5). The performances of the M-PSO and PSO-EO2 for the
Michalewics function are similar with negligible differences between the search capabilities and efficiency
between the two models (Table 5).
Table 5. Test Functions Optimisation Results
Schwefel Function
Fitness Solution Iterations
APSO 690.90 (-302,521, 420.96872,…,-302.5310) 10,000
M-PSO 1.27 x 10-4 (420.96871,420.96872,..,420.968710) 10,000
CLPSO1 0 (420.96871,420.96872,..,420.968710) 30,000
Michalewics Function
APSO -9.35 (2.20,1.57,1.29.1.12,2.22,1.57,1.87,1.76,1.28,1.57) 20,000
M-PSO -9.66015 (2.20,1.57,1.29.1.92,1.72,1.57,1.45,1.76,1.66,1.57) 20,000
PSO-EO2 -9.66015 (2.20,1.57,1.29.1.92,1.72,1.57,1.45,1.76,1.66,1.57) 20,000

With the swarm algorithm validated, a single-point airfoil optimisation analysis is performed. In
practice the swarm methodology does not require a starting point, since the initial population is randomly
distributed in the solution space. In the analysis, the starting position of the 20 particles is represented about
the base airfoil. Consequently, a comparative analysis between the results of gradient based method and a
global search run is established. The swarm simulations are repeated for each base airfoil (Sec. IV. A) to
determine the influence of varying the initialisation region on the results. The termination criterion is set
when all the particles share a common consensus on the optimality of the final solution over 40 iterations.
The mutation condition associated with a stagnant run is active during this part of the search phase. This is
defined to be operational when the global best of the solution has not improved over 10 consecutive
iterations. Thus, if mutation does not result in fitness improvement of the best particle, together with the
particles in the swarm sharing a fixed personal best over 40 iterations, termination is set.
The results of the single-point airfoil optimisation (Eq. 2) are illustrated in Figure 7, corresponding to
the individual base airfoils used for swarm initialisation. Despite initialising the solution at different
regions, the airfoil shapes (Fig. 7a) and the corresponding coefficient of pressure (Cp), (Fig. 7b) share
similarities over the three independent simulations. The solution to the NLF(1)-0115 swarm, indicates a
larger upper crest ordinate ZUP, hence the Cp on the upper surface varies comparatively (Fig. 7b).
Examination of final airfoil aerodynamics and fitness (Table 6) confirms that the swarm algorithm is
approaching the same optimal solution and the search process is independent of the initial swarm. The
objective of maintaining a Cl of 0.40 at cruise is satisfied and similar drag performance is established with
exception of RAE 2822, which differs by ≈4% in comparison to the NACA and NLF solution.
Table 6. Comparison of Results between GM & PSO Based Optimiser
Base Airfoil Method Lift Drag % Drag Improvement Fitness
GM 0.3957 0.0054 --- 2.20 × 10-3
NACA 0012
M-PSO 0.4002 0.0045 16.67 4.52 × 10-4
GM 0.4159 0.0048 --- 3.60 × 10-3
NLF(1)-0115
M-PSO 0.4000 0.0045 6.25 4.55 × 10-4
GM 0.4000 0.0051 --- 5.00 × 10-4
RAE 2822
M-PSO 0.4002 0.0047 7.84 4.68 × 10-4

The results indicate that the swarm algorithm is superior to gradient based methods with significantly
lower drag performance hence, function fitness (Table 6). Despite the gradient-based methods conforming
to the required lift performance criteria, variations in drag for the independent base airfoils, indicates the

11
American Institute of Aeronautics and Astronautics
presence of local minima in the solution space. The subtle differences in the shape contours (Fig. 7)
reinforces that the solution space to single-point optimisation problem is highly multimodal and multiple
shapes exist with almost the same objective function, (Table 6) also confirmed by Ray.24 The developed
swarm algorithm provides the search flexibility to overcome this issue and exhibit an optimal solution as
attributed by the consistency in aerodynamics (Table 6) and final shape contours (Fig. 7). Gradient based
methods are sensitive to the initial starting point as attributed by the variances in drag performance (Table
6). Similarities between initial and final shapes (Fig. 4-6) further indicate that the gradient tool merely
locates the nearest local optima to the initial airfoil. Thus, the search tool is inadequate for highly
multimodal problems. The developed swarm algorithm addresses this shortfall.
Final Optimal Solution using PSO with Disparate Base Airfoils

Base Airfoil: NACA 0012

Base Airfoil: NLF(1)-0115
0.1 Base Airfoil: RAE 2822

0.05
y/c

-0.05

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

x/c
(a) Final Optimal Shapes (b) CP Distribution of Final Optimal Shapes
Figure 7. Airfoil Shape Design Using M-PSO with Disparate Base Airfoils

V. Artificial Neural Networks

The implementation of a swarm algorithm for airfoil shape optimisation in the DNO approach is
computationally demanding. Each particle represents a potential solution to the problem. Design
optimisation studies utilising the swarm algorithm indicate an acceptable particle population range of 10-30
agents for convergence.25, 26 Mutation operators provide the benefit of introducing diversity to the search
process. The demerit is the additional demand on the computational resources, as the re-initialised particles
require fitness evaluation. Integration of Reynolds Average Navier Stokes (RANS) solver as function
evaluators in the DNO methodology will result in a computationally cumbersome optimisation cycle.
Development of a surrogate model to duplicate the performance of the RANS solver will address this
issue. An ANN is developed to simulate the design space of the proposed problem. Hence, the DNO
components defined in this paper are utilised in the design, development and validation of an ANN
structure. This proposed structure uses the PARSEC shape coefficients as inputs, which are processed
through a structure of hidden-layers, to generate the aerodynamic coefficient of the RANS solver as an
output, with a degree-of-accuracy. The network is designed to output a single aerodynamic coefficient, as is
reported to yield precise modeling in comparison to a multiple-output network.27 The time required for
input-output process is negligible, thus potential time savings for shape optimisation with a validated
surrogate model is significant.

A. Neural Network Structure Development & Validation

A trial-and-error process is undertaken to determine network layout for the proposed test problem. The
simulations are performed using MATLABs Neural Networks Toolbox28 which facilitates powerful
surrogate algorithms for design analysis. The following network characteristics are addressed:

Source of training data (computational and/or experimental);

Size of training data;

Number of layers;

Number of neurons in the hidden layer/s;

Transfer function type; and

Generalisation performance of the network to new input data

12
American Institute of Aeronautics and Astronautics
 The fundamental design principle for surrogate modeling is to minimise the source of errors; the
difference between theoretical and network solution for a defined input. Ideally, validated experimental
data for training is an acceptable choice. Computer simulations are associated with a degree-of-uncertainty
and computational aerodynamics is prone to such errors. Experimental data for airfoils generated with
shape functions is not possible, thus computational simulations are used to generate the input data. For
early design analysis, low-fidelity solvers are acceptable and provide rapid computational turnover at the
expense of solution accuracy. Regardless of solver fidelity, the trend of airfoil aerodynamics does not
change. Hence, XFOIL is ideal to test the effectiveness of the proposed ANN model. An experiment into
varying training sample size and its effect on network generalisation is established. The high computation
time of RANS solver prohibits this experiment. The integration of RANS model for network development
is required only when the optimality of the network structure is established from the experimental study.
For computational experiments, the training parameters must be efficiently distributed to improve the
coverage of the input space. Hence, the design of experiments (DoE), Latin Hypercube Sampling (LHS)29,
which uses a stratified scheme for data distribution is implemented. The methodology generates a
population of PARSEC airfoils from the identified solution space (Table 3) for network training.
The performance of the network is enhanced by normalizing the input and target data into a unit cube
between ±1. A feed-forward backpropagation network with Bayesian Regularization,30, 31 training
algorithm is used. Overfitting of data is mitigated by using an early stopping condition. Thus, the input
dataset is distributed into three subsets: 60% training, 20% validation and 20% generalisation. The training
process continues as long as the error on the distributed validation vectors decreases at each training
simulation. Training termination occurs when one of the following conditions are satisfied:

Mean-square-error of the training dataset is zero; or

An increase in error on the validation dataset for 50 iterations. The reported solution is the last iteration
before overfitting commenced.
A computational experiment was set-up to determine the influence of network structure on the
generalisation performance. Airfoil flow computations are at Reynolds Number of 3.0 Million, at Mach
0.35 and angle-of-attack of zero degrees. The size of the hidden layers was varied from one to two, neurons
10-50 in increments of ten for a training dataset of 4801, 8011 and 13,493 airfoils. With 60% of the original
data distributed for training, the input training vector is reduced to 2880, 4806 and 8095 airfoils
respectively in comparison to the original dataset. Generalisation is measured over 2,878 airfoils sampled
from the LHS methodology. The tan-sigmoid transfer function is used in all experiment simulations with a
linear function for model output. The percentage variance between network output and theoretical
aerodynamic coefficient of lift and drag of each airfoil is computed as a measure of generalisation
performance. The mean percentage error variance over the generalisation population is presented for lift
and drag in Figures 8-9 respectively.

(a) One Hidden Layer: Lift Coefficient (b) Two Hidden Layers: Lift Coefficient
Figure 8. Effect of Neural Network Structure on Lift Coefficient

The mean error carpet plots with one and two hidden layers for lift coefficient (Fig. 8a-b) indicate a
reduction in error variance with an increase in training size. The highest error bands are associated at the
lower spectrum of training sample population. The largest mean error variances are in excess of 25% for
the two models (Fig. 8a-b) for a network with ten neurons. With fewer neurons, the model is not

13
American Institute of Aeronautics and Astronautics
sufficiently complex to model the aerodynamic lift coefficient signals over new input data. Overfitting of
data in a single-layered network (Fig. 8a) is not present, despite an increase in neurons. Performance
improvements illustrate signs of stagnation hence, indicating the minimum error threshold is established.
Comparably, a double layered network (Fig. 8b), indicates signs of overfitting. The point of solution
instability is related to the size of the training data. A network with fewer training vectors encounters
overfitting earlier than a model with a larger training data with exact neuron sample size. Thus, a network
with 2,880 airfoils is overfitted after 20 neurons, compared to 30 for a training size of 4,806 and 40 for
8,095 (Fig. 8b). Thus, network complexity is proportional to the size of the training vector and the
underlying structure. The number of hidden layers has a strong influence on the over and under fitting of
data as a single-layered network indicated no signs of overfitting despite the larger neuron sample size. Yet,
a double-layered network indicated sings of instability. The optimal point at which the effects of under and
over fitting are suppressed is recorded. An error reading of 8% for single-layered model (Fig. 8a) compared
to 5% for a doubled-layered structure is observed (Fig. 8b).
The mean error drag variances indicate similar convergence properties in comparison to coefficient of
lift. Increases in training sample size result in significant reductions in mean errors. A single-layered
network with 2,880 training airfoils (Fig. 9a) indicates negligible drag prediction improvements as a
function of neuron size. The lack of performance improvement activity is related to the undersized training
vector with further increases in network complexity, leading to performance degradation through
overfitting. Training sample sizes comprising 4,806 and 8,095 PARSEC airfoils result in significant
performance improvements as function of neuron size. Overfitting of data, which will cause errors to
increase with additional neurons, is not evident. A flat plateau region is observed where errors are stagnant
about 4%. Thus, an optimal configuration for a single-layered network is established. The effect of
introducing an additional hidden layer is evident in Figure 9b. The mean error magnitudes are lower for
larger training sample sizes. Network performance for a training dataset of 2,880 airfoils is similar to a
single-layered network with negligible performance merits due to the addition of a second hidden layer.
Hence, this confirms that 2,880 airfoils are insufficient to accurately model the drag performance, despite
an increase in neurons (Fig. 9a) and with the addition of a second layer coupled with an increase in neuron
population (Fig. 9b). Significant performance improvements are observed with larger training vectors in
comparison to a single hidden layered network. A network with 4,806 airfoils illustrates consistent error
performance of 2.9% between 20-40 neurons, thus indicating a converged error tolerance region. With the
addition of ten neurons, error reading increase to 4.0% thus suggesting overfitting. An 8,095 airfoil training
configuration shows an inactive performance improvement region between 20-50 neurons, with an error of
2.16%. Overfitting is not observed which is attributed to the larger training size. This has an effect of
delaying the onset at which model complexity becomes excessive, to trigger overfitting, in comparison to a
model with fewer training vectors. If further neurons were added to the database comprising 8,095 airfoils,
model complexity will lead to instabilities as noted in the 4,806 airfoil database configuration.

(a) One Hidden Layer: Drag Coefficient (b) Two Hidden Layers: Drag Coefficient
Figure 9. Effect of Neural Network Structure on Drag Coefficient

Network with lowest mean error difference between theoretical and generalised sample is examined
(Table 7). The performance is measured with a regression analysis which outputs network response against
the corresponding target. A linear fit through the theoretical target against network output is used to

14
American Institute of Aeronautics and Astronautics
 measure the ‘goodness’ of fit and is equated by the r-correlation coefficient, r-value. From the 2,878
generalisation sample population, the minimum and maximum error magnitude of an airfoil is reported.
The standard deviation is computed to determine the scatter of percentage error data over the generalisation
sample.
Table 7. Optimal Neural Network Configuration for Lift and Drag Coefficient
Aerodynamic Training Hidden Mean Error Standard Min Max
Neurons r-value
Coefficient Size Layers Variance Deviation Error % Error %
Lift 8,095 2 30 4.93% 21.37 0 614.29 0.999
Drag 8,095 2 20 2.16% 3.58 0 53.73 0.995

A mean error of ≈5% of lift over the generalisation sample is acceptable when examined independently.
An r-correlation close to one indicates a strong positive relationship between network output and theory.
Due to the large disparity between minimum percentage error of zero and a maximum ≈615%, the presence
of outliers is evident. The standard deviation further suggests disperse error distribution. A histogram
analysis using the leverage on the X axis and the error residuals on the Y axis is used to observe the
coefficient of lift modeling error distribution (Fig. 10a). Each histogram error sub-group spans 12.3%. Of
the 2,878 generalisation airfoils, 2,679 sections are categorised in the lowest error histogram, with an error
range of 0% - 12.3%. The frequency of airfoils, with errors greater than 12.3% decrease substantially. A
total of 155 airfoils are approximated by 12.3% - 49.14% with the balance 44 airfoils equated with error
histograms of 49.14% - 614.29%. The frequencies of airfoils with errors in excess of 172% are isolated to
one airfoil per error sub-group, up to the maximum miss-match of 614.29%. Hence, despite the low mean
error variance of ≈5%, the histogram analysis illustrates regions of poor solution modeling.
The coefficient of drag approximation requires ten fewer neurons in comparison to the lift coefficient
network, for accurate solver modeling (Table 7). A mean error variance of ≈2% and a low standard
deviation indicates errors are not widely dispersed, in comparison to lift coefficient analysis. A minimum
error of zero is computed with a maximum error percentage of ≈54% observed (Table 7). An r-correlation
of 0.995 indicates a strong linear relationship between network output of drag and theory. The histogram of
drag errors is illustrated in Figure 10b. Each histogram bar is categorised by an error of 1.075%. In the first
error sub-group, 1,199 airfoils are classified with an error of 0% - 1.075%. From the population, 1,081
airfoils have an exact drag approximation with an error of zero percent. For error histograms of 1.075% -
5.373%, 1,456 airfoils are classified in this category with the 136 airfoils characterised with errors spanning
5.373% - 10.75%. Population of airfoils with errors in excess of 10% is limited to one section per error
histogram sub-group, with a maximum error of ≈54% observed.
Histogram of Coefficient of Lift Percentage Error for Histogram of Coefficient of Drag Percentage Error
Neural Network PARSEC Airfoil Simulation for Neural Network PARSEC Airfoil Simulation
1200
2500
1000

2000
Simulation Frequency
Simulation Frequency

800

1500
600

1000
400

500 200

0 0
0 100 200 300 400 500 600 0 5 10 15 20 25 30 35 40 45 50 55
Percentage Error Percentage Error

(a) Coefficient of Lift (b) Coefficient of Drag

Figure 10. Neural Network Aerodynamic Coefficient Modeling – Histogram of Errors

B. Airfoil Shape Optimisation using Developed Neural Network Structure

The proposed networks of lift and drag are used for single-point airfoil shape optimisation (Sec. IV Eq.
2). Considerable computation time benefits are hypothesized with the omission of the flow solver from the
DNO loop. Instead of generating the initial swarm about a base airfoil, (Sec. IV B) the LHS methodology is

15
American Institute of Aeronautics and Astronautics
 used to initiate the search process. This is to test the flexibility of the proposed PSO-ANN methodology in
establishing optimal shapes which are comparable to the direct PSO-XFOIL optimisation process (Sec. IV
B). The final optimal shape is illustrated in Figure 11a and is superimposed with the solutions of the direct
search process of Figure 7a for ease of comparison (Fig. 11b).
Airfoil Shape Optimisation with a PSO / Neural Network Hybrid Optimal Shapes with Direct Optimisation Approach
Optimisation Methodology & Fitness Evaluations using Surrogate Modeling

PSO / XFOIL - Base Airfoil: NACA 0012

0.1 0.15 PSO / XFOIL - Base Airfoil: NLF(1)-0115
PSO / XFOIL - Base Airfoil: RAE 2822
PSO / ANN Airfoil - LHS Base Airfoil
0.05 0.1

0.05
y/c

y/c
0

-0.05 0

-0.05
-0.1

-0.1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x/c x/c
(a) PSO / ANN (b) PSO / XFOIL & PSO / ANN
Figure 11. Airfoil Shape Optimisation using Surrogate Modeling & Direct Flow Solver for Fitness
Evaluation

The final shape of the PSO / ANN methodology (Fig. 11a) is similar to the results of the direct fitness
flow solver evaluations (Fig. 11b). Solution initialisation has a minor effect on the final airfoil shape.
Results of the direct fitness evaluations are about the base airfoils used for gradient-based analysis (Fig.
11b) and the LHS methodology for the ANN model (Fig. 11a). This suggests that the solution is converging
to a specified region in the design space. The aerodynamics of the final PSO / ANN model are presented in
Table 8.
Table 8. PSO / ANN Aerodynamics & Fitness Results of Optimal Planform
Base Method ANN ANN ANN Theoretical Theoretical Theoretical
Airfoil Lift Drag Fitness Lift Drag Fitness
0.3985 0.0046 6.58 × 10-4
LHS PSO / ANN 0.4000 0.0046 4.56 × 10-4
(↓0.375%) (0%) (↑≈44%)

The results of the surrogate model (Table 8) conform to the stipulated objective function (Eq. 2).
Minimum lift coefficient performance requirement is satisfied with drag performance comparable to the
direct PSO / Flow Solver methodology (Table 6). Similarities in airfoil shape (Fig. 11b) and aerodynamics
(Table 6 and 8) between fitness evaluation through a surrogate model and direct computation through
XFOIL, re-enforces that the solution is converging to a common design region. The aerodynamics of the
ANN profile are computed directly with XFOIL to establish the difference between theoretical and
simulation solutions. Drag is simulated with a zero percent error difference between XFOIL and ANN, with
lift coefficient underestimated by 0.375% (Table 8). The significance of achieving the required minimum
lift coefficient is established through the multiplication of a weight term to the objective function. Hence, a
penalty function is introduced, for profiles that exhibit lift coefficients lower that the minimum
requirement. Theoretical analysis of lift for the ANN profile yields a lift of 0.3985, which is lower than the
minimum requirement. The penalty function is activated thus, leading to a significantly higher fitness
evaluation.
The analysis presented indicates the process required for the development of an ANN model for airfoil
shape optimisation. Residual error distribution indicated the presence of extreme errors. Hence, the
feasibility of the final solution using the ANN model is affected due to the accumulation of errors in
aerodynamics simulations during the search process. Further design development and validations are
required to reduce error magnitudes. The analysis has shown that the architecture provides the design merit
of enhanced computation. The simulation presented required 25 seconds for convergence on a simulation
operating on a PC with 1.86GHz CPU and 2.0 GB of RAM.

16
American Institute of Aeronautics and Astronautics
 VI. Conclusion and Future Work
The DNO approach for shape optimisation is defined. The process of developing and validating a
geometrical shape function, optimisation algorithm for multimodal solution topologies and a surrogate
model were examined. The design goal focuses on developing an efficient search model, capable of
locating an optimal solution with acceptable accuracy without resorting to extensive computational
resources.
Definition of the solution space is one approach to address this requirement. The geometrical shape
function defines the class of airfoils that are examined during the optimisation process. The process of
establishing the search limits of the geometry variables for the shape function hence, the solution space was
presented. Airfoils used for high-speed operations, symmetrical sections for rotary platforms, to mission
goals stipulating HALE requirements were identified. The NASA NLF airfoil was identified as a
benchmark solution, as the profile performance closely matched the design intent. The PARSEC geometry
variables corresponding to the identified shapes were established through a shape convergence optimisation
study. The scatter of geometry variables is dependent on the intent design goals and was normally
distributed. High speed airfoils are characterised by smaller leading edge radii and larger upper and lower
crest curvatures to address shock wave development. The design limits hence, the solution space, was
established by measuring the magnitude of the standard deviation required about the mean airfoil until all
HALE airfoils were mapped.
The process of airfoil optimisation was undertaken in the defined solution space. The feasibility of
gradient method was examined by initiating the search process at different solution regions. Each
simulation resulted in a new shape, with the final profile contour matching the initial starting point. The
fitness function was dependent on the initial starting point, thus suggesting the solution space is highly
multimodal and gradient based models are indicating sub-optima solutions. A PSO model was introduced
and modified for multimodal problems. Two benchmark mathematical models were used in the validation
study and it was shown that a standard PSO model is incapable of locating the optimal solution over a ten-
dimensional problem. The diversity of the search agents through a double-mutation operator improved the
convergence performance significantly. The modified swarm algorithm was used for airfoil shape
optimisation with variation in initial starting points. Regardless of the initialisation region, the results of the
modified swarm algorithm with mutation were consistent. The fitness values converged to one value, with
slight variations in final shapes, reinforcing the multimodal nature of the design space. The global search
agent indicated drag improvements in excess of 16% in comparison to local gradient methods.
A surrogate model was developed to map the relationship between airfoil solution space and
aerodynamics. A sensitivity study to determine the influence of network structure on the generalisation
performance of the network was examined. Overfitting of data is not evident with single-layered networks,
but is an issue with double-layered networks for lift and drag. The errors reported for drag coefficient are
significantly lower that lift convergence errors. A residual histogram analysis indicated 93% of airfoils with
a lift approximation error of less than 12%. For drag coefficient modeling, 38% of airfoils were
approximated exactly with zero percentage error in comparison to theory. In total 96% of airfoils are
modeled with errors less than 10%. From the experimental study, network with the lowest mean errors over
the generalisation data were identified. Airfoil shape optimisation is established by coupling the PARSEC
shape function with the surrogate model, to duplicate the lift and drag performance of the solver. The
results of the optimisation process are comparable to the simulations comprising XFOIL as used directly
for fitness evaluations, with independent base airfoils for solution initialisation. The final shapes and
aerodynamics share similarities thus indicating that the solution is converging to a common design space
region. The drag of the final ANN solution is simulated with exact precision and lift 0.375% in comparison
to theory. Significant computational time benefits are reported, with single-point airfoil optimisation
simulation requiring 25 seconds for convergence.
Further design studies will focus on developing a direct relationship between geometry variables and
airfoil aerodynamics. A pre-screening process to determine the importance of variable on the objective
function will be attempted. Screening studies using multiple regression analysis and Morris’ method will be
used to develop an input-output relationship between geometry coefficients and objective function of drag.
Consequently, variables which have insignificant effect on the objective can be identified and eliminated to
further reduce the complexity of the solution space.
The integration of a RANS solver to enhance the accuracy of the aerodynamics, in the DNO process
will be initiated. The swarm algorithm will be modified to incorporate parallelization for analysis on

17
American Institute of Aeronautics and Astronautics
supercomputing clusters, to reduce the computational time required for convergence. The use of gradient
methods to finish the search process by using the end solution of the PSO as input for solution initialisation,
will be attempted to further reduce the drag of HALE airfoils.
The results of the surrogate model indicated the presence of extreme outliers, particularly for lift with
error in excess of 600% observed. The scope of the sensitivity study will be increased. The effect of
varying transfer functions and the addition of a third hidden layer, to further improve network performance
and reduce the error magnitude of the extreme outliers will be examined. A methodology to map solution
region where the network exhibits extreme simulation errors will be identified. The fitness function will be
computed directly from the flow solver for regions that exhibit extreme errors as outputs of the surrogate
model. This will mitigate solution inconsistencies of performing a search optimisation with false fitness
evaluations. The time benefits associated with optimisation through surrogate modeling will be extended to
Multi-Point-Multi-Objective airfoil design optimisation in the design of a HALE planform for a Multi-
Mission UAV.

References
1
Liang, J. J. and Qin, A. K., "Comprehensive Learning Particle Swarm Optimizer for Global Optimization of
Multimodal Functions," IEEE Transactions on Evolutionary Computation, Vol. 10, No. 3, 2006, pp. 281-295.
2
Chen, M.-R., Lu, Y.-Z. and Luo, Q., "A Novel Hybrid Algorithm with Marriage of Particle Swarm Optimization
and External Optimization," Shanghai Jiaotong University,
3
Prouty, R. W., Helicopter Performance, Stability and Control, Krieger Publishing Company, Inc., Malabar,
Florida, 1995.
4
Labrujere, T. E. and Slooff, J. W., "Computational Methods for the Aerodynamic Design of Aircraft
Components," Annual Reviews of Fluid Mechanics, Vol. 25, 1993, pp. 183-214.
5
Winnemoller, T. and Dam, C. P. V., "Design and Numerical Optimization of Thick Airfoils," 44th AIAA
Aerospace Sciences Meeting and Exhibit, 2006, pp. 1-14.
6
Namgoong, H., "Airfoil Optimization of Morphing Aircraft," PhD, Aerospace Engineering, Purdue, Indiana, 2005.
7
Quagliarella, D. and Cioppa, A. D., "Genetic Algorithms Applied to the Aerodynamic Design of Transonic
Airfoils," AIAA Applied Aerodynamics Conference, Colorado Springs, CO, 1994,
8
Holst, T. L. and Pulliam, T. H., "Aerodynamic Shape Optimization Using a Real-Number-Encoded Genetic
Algorithm," NASA, 2001.
9
Gallart, M. S., "Development of a Design Tool for Aerodynamic Shape Optimization of Airfoils," Master of
Applied Science, Department of Mechanical Engineering, University of Victoria, 2002.
10
Duvigneau, R. and Visonneau, M., "Hybrid Genetic Algorithms and Neural Networks For Fast CFD-Based
Design," 9th AIAA/ISSMO Symposium on Multidisciplinary Analysis and Optimization, Atlanta, Georgia, 2002, pp.
633-643.
11
Khurana, M., Winarto, H. and Sinha, A., "Airfoil Geometry Parameterization through Shape Optimizer and
Computational Fluid Dynamics," 46th AIAA Aerospace Sciences Meeting and Exhibit, Reno, Nevada, 2008, pp. 18.
12
Sobieczky, H., "Parametric Airfoils and Wings," Numerical Fluid Dynamics, Vol. 68, 1998, pp. 71-88.
13
Khurana, M., Sinha, A. K. and Winarto, H., "Multi-Mission Re-Configurable Unmanned Aerial Vehicle - Airfoil
Optimisation Architecture," International Conference on Engineering Technology, Kuala Lumpur, Malaysia, 2007, pp.
10.
14
Khurana, M., Winarto, H. and Sinha, A., "Application of Swarm Approach and Artificial Neural Networks for
Airfoil Shape Optimization," 12th AIAA/ISSMO Multidisciplinary Analysis and Optimisation, Victoria, British
Columbia, Canada, 2008,
15
Somers, D. S., "Design and Experimental Results for a Natural-Laminar-Flow Airfoil for General Aviation
Applications," Langley Research Center, 1981.
16
McGhee, R. J. and Beasley, W. D., "Wind-Tunnel Results for a Modified 17-Percent-Thick Low-Speed Airfoil
Section," NASA, 1981.
17
GORAJ, Z., et al., "High altitude long endurance unmanned aerial vehicle of a new generation – a design
challenge for a low cost, reliable and high performance aircraft," BULLETIN OF THE POLISH ACADEMY OF
SCIENCES, Vol. 52, No. 3, 2004, pp. 173-194.
18
Hall, D., "Thick Airfoil Designs for a HALE Vehicle," 8th AIAA Applied Aerodynamics Conference, Portland,
OR, 1990,
19
Winnemoller, T. and Dam, C. P. v., "Design and Numerical Optimization of Thick Airfoils Including Blunt
Trailing Edges," Journal of Aircraft, Vol. 44, No. 1, 2007, pp. 232-240.
20
Mathworks, "Optimization Toolbox User's Guide," 2005.
21
Qin, Z., Yu, F., Shi, Z. and Wang, Y., "Adaptive Inertia Weight Particle Swarm Optimization," ICAISC, 2006, pp.
450-459.
22
Andrews, P. S., "An Investigation into Mutation Operators for Particle Swarm Optimization," IEEE Congress on
Evolutionary Computation, Vancouver, BC, Canada, 2006, pp. 1044-1051.

18
American Institute of Aeronautics and Astronautics
 23
Ratnaweera, A., Halgamuge, S. K. and Watson, H. C., "Self-Organizing Hierarchical Particle Swarm Optimizer
with Time-Varying Acceleration Coefficients," IEEE Transactions on Evolutionary Computation, Vol. 8, No. 3, 2004,
pp. 240-255.
24
Ray, T. and Tsai, H. M., "Swarm Algorithm for Single-and Multiobjective Airfoil Design Optimization," AIAA
Journal, Vol. 42, No. No. 2, 2004, pp. 366-373.
25
Robinson, J. and Rahmat-Samii, Y., "Particle Swarm Optimization in Electromagnetics," IEEE Transactions on
Antennas and Propagation, Vol. 52, No. 2, 2004, pp. 397-407.
26
Carlisle, A. and Dozier, G., "An Off-The-Shelf PSO," in ProcWorkshop Particle Swarm Optimization,
Indianapolis, IN, 2001,
27
Greenman, R. M. and Roth, K. R., "Minimizing Computational Data Requirements for Multi-Element Airfoils
Using Neural Networks," Journal of Aircraft, Vol. 36, No. 5, 1999, pp. 777-784.
28
Mathworks, "Neural Network Toolbox," 1984-2008.
29
McKay, M. D., Beckman, R. J. and Conover, W. J., "A Comparison of Three Methods for Selection Values of
Input Variables in the Analysis of Output From a Computer Code," Technometrics, Vol. 21, No. 2, 1979, pp. 239-245.
30
MacKay, D. J. C., "Neural Computation: Bayesian Interpolation," Neural Comput., Vol. 4, No. 3, 1992, pp. 415-
447.
31
Foresee, F. D. and Hagan, M. T., "Gauss-Newton Approximation To Bayesian Learning," International
Conference on Neural Networks, Vol. 3, Houston, TX, 1997, pp. 1930-1935.

19
American Institute of Aeronautics and Astronautics

View publication stats