Professional Documents
Culture Documents
net/publication/242741331
CITATIONS READS
15 651
3 authors, including:
All content following this page was uploaded by Hadi Winarto on 22 February 2017.
Nomenclature
λ = Design Variable fi = Solution Fitness
( x / c) i = Chord Length Interval f i (x) = Shape Functions (Polynomials)
[λ1 , K , λ n ] = Cost / Fitness of Each Design Variable ƒtarget = Target Airfoil
ak = Scalar Step Length k = Iteration
c = Chord Length m = Swarm Population Size
c1, c2 = Learning Factors n = Design Variable Population
cd = Profile Drag Coefficient rand = Uniform Random Value [0,1]
cf = Coefficient of Friction vi = Velocity of Particle
cl = Profile Lift Coefficient w = Inertia Weight
cp = Coefficient of Pressure xi = Current Position of Particle
ƒapprox. = Approximated Airfoil ∆ƒ = Airfoil Geometrical Difference
(Objective Function)
1
PhD Candidate, The Sir Lawrence Wackett Aerospace Centre, 850 Lorimer Street, Port Melbourne, VIC 3000,
Australia - AIAA Student Member
2
Associate Professor, School of Aerospace, Mechanical and Manufacturing Engineering, GPO Box 2476V, 3001,
VIC, Australia - AIAA Member
3
Director Aerospace and Aviation, The Sir Lawrence Wackett Aerospace Centre, 850 Lorimer Street, Port
Melbourne, VIC 3000, Australia - AIAA Member
1
American Institute of Aeronautics and Astronautics
I. Introduction
T HE Sir Lawrence Wackett Aerospace Centre pioneered a conceptual design of Re-Configurable Multi-Mission
Unmanned Aerial Vehicle (RC-MM-UAV) for futuristic civil and military operational needs. Uni mission
UAVs presently are under various stages of design, development and trials as an alternative over manned systems to
address dull, dirty and dangerous missions. Future operational doctrines merit the development of a Multi-Mission
UAV to cover a wide mission spectrum. Though uni-mission designs address a certain section of the requirements,
the performance is limited; in regards to range/endurance and speeds. A uni-mission UAV with fixed wing geometry
design addresses a specific performance requirement of speed and endurance. RC-MM-UAV concept will
encompass long endurance/range requirements of Intelligence, Surveillance and Reconnaissance (ISR) and the
flexibility of speed for Suppression of Enemy Air Defence (SEAD) sorties in its mission profile – a typical futuristic
operational scenario. Multi-Mission profile requires variations in speeds, altitude of operation and maneuverability
to address disparate performance requirements.
UAV operations in Australia have been restricted to atmospheric monitoring and aerial photography with
mineral exploration emerging as a new sector. A detail market survey identified the requirement of a Tier of UAVs
based on future operational requirements (Table 1).2 The Australian Defence Force’s – AIR 7000 Request-for-
Proposal calls for the development of a Multi-Role UAV for introduction within the current fleet, and the JP129
project for Tactical Unmanned Aerial Vehicle for Intelligence, Surveillance, Target Acquisition and Reconnaissance
Capabilities. The Australian program on UAV acquisition covers the request of a High-Altitude Long Endurance
UAV for Reconnaissance and Surveillance, thus the development and introduction of an intelligent UAV platform is
identified as a key operational asset.
Table 1. UAV Tier Classification based on Operational Capabilities
UAV Category Designation Max. Alt [ft] Range [km] Endurance [hrs]
Tier I Interim Medium Altitude, Up to 15,000 Up to 250 5-24
Endurance
Tier II Medium Altitude, 3,000 - 25,000 900 More than 24 hrs
Endurance
Tier II Plus High Altitude, Endurance 65,000 Up to 5,000 Up to 42 hrs
Tier III Minus Low Observable – High 45,000 - 65,000 800 Up to 12 hrs
Altitude, Endurance
The pioneering RC-MM-UAV design is a viable concept to address the identified requirements in the market
survey. The re-configurable modular concept of wings and payload considered a total of 52 missions and covered
high to medium altitude long endurance of 25-42 hours and range in excess of 4,000 nm at Mach numbers 0.32-0.60.
A major section of missions (total of 32) shared commonality in flight performance and mission specific payloads
with a maximum gross take-off weight and wing area envelope of 4,000-10,000 kg and 49-68m2 respectively.
The investigation on wing extension options and its design and operational merits resulted in the morphing
concept being considered for further detailed study. Mid-flight wing planform flexibility through the use of smart
actuators / materials will provide multi-mission capabilities in the mission profile from slow speed ISR roles to
quick dash segments for SEAD operations.
In this paper a Direct Numerical Optimization (DNO) methodology to facilitate the design of mission segment
based airfoils for RC-MM-UAV is proposed. The methodology (Fig. 1) comprises of the following: a) Mathematical
shape function to represent potential airfoils; b) Flow solver validation for aerodynamic computation; and c)
Intelligent search agent for overall optimisation.
3
American Institute of Aeronautics and Astronautics
points to generate a smooth geometry.6 Undulating sections will be present during the optimisation search, leading
to additional computational time.
Conformal mapping techniques have been previously analyzed including Joukowski and Kármán-Trefftz
transformations for airfoil development. This mapping technique requires fewer design variables to represent a
relatively large set of airfoils.3 The Joukowski mapping through two design variables controls camber and thickness-
to-chord ratio. The Kármán-Trefftz transformations are introduced with an additional variable to control the trailing
edge angle.3 The conformal mapping technique is unsuitable for multi-objective optimization problems, due to
geometrical limitations within the design space.8 The transformations are unable to control airfoil nose radius and
will not produce divergent trailing-edge airfoils, which provides improved airfoil effectiveness in terms of increased
thickness with enhanced lift capability at a lower drag performance.9 Additionally, it is limited in the chordwise
flexibility for the placement of maximum thickness and camber locations. Thus, the advantage of a lower
computation time is not justified against the limitations present in controlling important airfoil features.
Wu et al. examined the PARSEC, Hicks-Henne and mesh-point geometrical functions to represent cascade
blades through an inverse design process. A base airfoil was modified through the different shape functions to match
the pressure distribution with the benchmark pressure coefficient of a target airfoil.10 The investigation confirmed
that the selection of a method is governed by the design requirements and its subsequent application.10
B. Test Methodology
A geometrical shape parameterization investigation with an analytical approach is needed for the design of long
endurance airfoils. Methods examined include the Hicks-Henne,11 Wagner,12 Legendre,* Bernstein,† and NACA
normal modes13, 14. The polynomials represent a series of sinusoidal curves that are generated across a specified
chord length interval. The technique operates by adding a finite sum of closed shape functions to an initially
specified airfoil shape to generate a target section. The design variables ' λi ' , are multipliers to shape functions and
determine the contribution of each function to the final shape.4 Mathematically it is represented as follows:
n
y ( x / c, λi ) = y ( x / c ) initial +
airfoil
∑ λ f ( x / c)
i =1
i i (1)
Where:
λi = Design Variable f i ( x / c) = Shape Functions
n = Design Variable Population Size ( x / c) i = Chord Length Interval
A symmetrical NACA 0015 profile is used as the
Airfoil Profiles for Geometrical Shape Parameterization
benchmark. Three target sections for the shape
Base: NACA 0015
convergence test, were considered with low speed and Target 1: NASA LRN(1)-1007
0.15 Target 2: NASA LS(1)-0417Mod
long endurance, being a critical requirement for RC- Target 3: NASA NLF(1)-1015
target airfoil. This is mathematically represented as Figure 2. Airfoils for Shape Parameterization
follows: Optimization.
⎡ ⎤
∆ ƒ min ∑ ⎢( y / c )i − ( y / c )i ⎥ (2)
⎣ approx t arg et ⎦
*
http://mathworld.wolfram.com/LegendrePolynomial.html [cited 2 May 2007]
†
http://mathworld.wolfram.com/BernsteinPolynomial.html [cited 2 May 2007]
4
American Institute of Aeronautics and Astronautics
Where:
∆f = Objective Function ( y / c )iapprox = Approximated Airfoil Profile ( y / c )it arg et = Target Airfoil Profile
Two search algorithms are used to minimize the objective function (Eq. 2). A conjugate gradient line search
method14 and a “probabilistic global seeking particle swarm optimizer”15-18 were applied to test and measure the
following: a) Flexibility of the shape functions by minimizing the objective function across the three target sections;
and b) Computational efficiency of the optimization models by recording the number of design iterations and clock
time required for solution convergence.
The pressure and suction surfaces are independently simulated over a design variable population size ‘n’, from 2
to 10 and then coupled to form the final shape. The testing process requires the following inputs for evaluation:
Base airfoil profile : (x / c )Base and ( y / c )Base ;
Airfoil Airfoil
Target section profile : ( y / c )T arg et in line with (x / c )Base distribution; and
Airfoil Airfoil
Design variable size :‘n’
The Hicks-Henne methodology requires additional inputs to adjust the peak magnitudes (aH-H) and location (pii)
of each contour that are prior-adjusted.6 The proposed test methodology is presented in Fig. 3.
( )
i
1 ⎛ d ⎞ 2 i
Legendre f i ( x) = ⎜ ⎟ x − 1 ; where i = 1, n Design Variable Population n
2 i i! ⎝ dx ⎠ Magnitude of Design Variable/s [λ1 , K , λ n ]
⎛n⎞
f n,i ( x ) = ⎜⎜ ⎟⎟ x i (1 − x) n −i ;
⎝i ⎠ Design Variable Population n
Bernstein
⎛n⎞
where ⎜⎜ ⎟⎟ =
n!
, i = 0, n ; where i = 1, n
Magnitude of Design Variable/s [λ1 , K , λ n ]
⎝ i ⎠ i!(n − i)!
f1 ( x) = x − x
Design Variable Population n
f 2 ( x) = x(1 − x)
NACA Magnitude of Design Variable/s [λ1 , K , λ n ]
f i +1 ( x) = x i (1 − x), for i = 2,3,4,5
5
American Institute of Aeronautics and Astronautics
C. Airfoil Shape Optimization
The two optimization models proposed operate with unique search strategies. Gradient methods are
computationally efficient and provide rapid solution convergence if the starting region is well defined. When a
function requires adjustment of multiple variables, local methods settle into a local minima. Hence, evolutionary
programming model needs to be adopted.
A stand-alone PSO model is used for shape convergence across the five functions proposed and convergence
performance was measured against the line search method for Wagner, Legendre, Bernstein and NACA normal
modes (Table 3). As the Hicks-Henne method involves computing multiple functions (Table 2), a PSO model is
used to obtain a starting point. The results are used as an input to the line search and a hybrid methodology is
proposed (Table 3). The test evaluates which polynomial is most appropriate across the three target sections and the
effectiveness of the gradient, global, and the combination of both global and local search agents based on
geometrical and aerodynamic convergence is evaluated.
Table 3. Airfoil Shape Optimization Tools
Polynomial Optimizer Utilized Type
PSO Global
Hicks-Henne
Hybrid – PSO / Line Search Global / Local
PSO Global
Wagner
Line Search Local
PSO Global
Legendre
Line Search Local
PSO Global
Bernstein
Line Search Local
PSO Global
NACA
Line Search Local
∇f ( x * ) = Hx * + c ≈ 0 (4)
Where:
x ∗ = Optimal Solution c = Constant Vector H = − x* c
−1
The gradient of the objective function through the perturbation of design variables is obtained with a finite
differences method, to refine the search towards a minimum solution. The line search algorithm uses a pre-defined
scalar step length a k , to direct the search p k , with a decreasing objective function, after each iteration x k . The
iterative process is based on Wolfe’s method which operates along the line x k + a k p k , until the convergence criteria
is established.14 The line search algorithm and the search mode are expressed as follows:
x k +1 = x k + a k p k (5)
6
American Institute of Aeronautics and Astronautics
2) Particle Swarm Optimization
The PSO methodology is based on the evolution of social behavior where the solutions are represented by a set
of particles that heuristically navigate through a design space.15 Unlike Genetic Algorithms, the effectiveness of a
potential solution is not dependent on crossovers and mutations. It is updated based on an information sharing
methodology with the population of particles.
Each particle in swarm of size m, is a potential solution of the objective function (Eq. 2) and is represented by
the relative position xi , and velocity vi . The optimization in initiated through the random dispersion of the particles
with each variable n, set within a domain of minimum and maximum limit (Fig. 4). The
position [ xi = ( xi1, xi 2,..., xin ), i = 1,2,..., m] and velocity [vi = (v i1, vi 2,..., vin ), i = 1,2,..., m] of the individual particles is
recorded and solution fitness evaluated (Eq. 2) and stored [ Pgbest = ( Pg1, Pg 2,..., Pgn )] (Fig. 4). A particle with the
‘best’ global solution, ' Pgbest ' , from the population, is recorded and the remaining particles update their position and
velocity to follow Pgbest over the subsequent iterations until solution convergence (Fig. 4).19
At each iteration ' k ' , the search direction is refined by updating the position, ' xi (k + 1)' , and velocity, ' vi (k + 1)' ,
(Eqs. 7-8); where the velocity rate of change is a function of user pre-defined learning factors as follows: a)
Cognitive c1 and b) Social c 2 parameters that influence local and global search patterns;17 and c) Constriction weight
factor (Eq. 9) for solution convergence, by neutralizing a balance between the local and global search regions (Fig.
4).15,18
xi (k + 1) = x i (k ) + v i (k + 1) (7)
2
w= ; where ϕ = c1 + c 2 (9)
2 − ϕ − ϕ 2 − 4ϕ
The solution dimensional space is governed by the constraints of the design variables. The velocity of the
particles are controlled, through a scalar multiplier (Eq. 8) to limit the maximum velocity and avoid particles
overshooting the computational domain.19 Despite the velocity restrictions, particles conjugate outside the design
domain and the application of boundary conditions is needed to re-direct the particles within the search space.19 The
absorbing, random and reflecting conditions are used for such purposes. A global search pattern is preferred to cover
a wide search region. The absorbing technique offers higher local search capabilities, as the particles are imposed
into the boundary of the computational domain. It will offer rapid convergence for instances where the solution is
within the wall region. Conversely, if a local minimum is present, the particles may be trapped within the boundary,
and will ultimately diverge over a progression of iterations, leading to an increase in computational requirements.
The random initialization condition, disperses solution space exceeded particles within the search domain, thus re-
instating a heuristic search process. The reflecting wall technique will promote a linear search pattern, as the
particles are reflected in the opposite direction to the original search path. This method offers rapid convergence if
the global minimum lies in the original search path and is ignored due to the overshooting effect of the particles and
high velocity. The random and reflecting conditions were independently simulated and tested in this paper, to
determine which method is appropriate for airfoil shape convergence.
The inputs to the PSO algorithm define the solution search behavior and together with the imposed wall
boundary conditions, the robustness of the model is defined (Table 4). The boundary wall condition test is simulated
over a single shape function and covered in section IV of this paper.
7
American Institute of Aeronautics and Astronautics
Table 4. PSO Inputs and Constraints
Input / Parameters Value
Fixed Variables: Minimum Maximum
Cognitive - c1 ---- 25 ----
Social - c2 ---- 10 ----
Swarm Population - m ---- 2,000 ----
Velocity Scalar Multiplier ---- 0.1 ----
Continuous Variables:
Shape Function Population Size - n 2 10
Shape Function Design Variables Limit - λn -0.05 0.05
Hicks-Henne Geometrical Features
αH-H 0 5
pii 0 1
Algorithm Set-Up
Boundary Conditions: a) Random Initialization
b) Reflected Wall
Convergence Criteria:
Constant Gradient Run 15
Degree of Precision 10-6
Maximum Iterations 400
The optimization definition in Table 4, is presented in the following flowchart (Fig. 4). The cognitive c1 and
social c2 parameters, together with swarm size are user defined and required to initiate the search process. Swarm
population effects convergence rate, with high values resulting in greater computation efforts, while may lead to
improved solution fitness. A valid combination of these inputs, allows for an overall effective search algorithm.
* For Bernstein, Legendre, NACA and Wagner polynomials, the final PSO solution is stored in array Pbestg, as
[λ1 , K , λ n ] . The solution to the Hicks-Henne function is stored in Pbestg as [(α H − H1 , L , α H − H n ), ( p ii1 , L , p iin )] and
used as an input into the line search algorithm for an initial estimate to derive the solution [λ1 , K , λ n ]
8
American Institute of Aeronautics and Astronautics
IV. Geometrical Shape Optimization Results and Discussion
This section covers convergence data by the PSO algorithm with random and reflected boundary wall conditions,
line search and hybrid method (Table 3). The total cost presented, is related to the fitness of the objective function
(Eq. 2), and is the geometrical difference between target and approximated planform. The cost magnitude
incorporates modeling upper and lower airfoil surfaces over a variable size of 2-10. A complete airfoil will
incorporate twice the original variables of 4-20.
0.09 0.1
0.09
0.09
0.08 0.09
0.6 0.08
0.07
0.08 0.25 0.08
0.07
0.07
0.06
0.07
0.06
0.06
0.05
0.5
y/c
0.06
y/c
0.05 0.05
0.04
y/c
0.2
y/c
0.05
Objective Function(∆ f)
0.04
Objective Function(∆ f)
0.03 0.04
0.03 0.04
0.02 0.03
0.4 0.02 0.03
0.01 0.02
Target Target
0.01
0
Approx.
0.15 0.01
0.02
Approx.
Target Target
0
-0.01 Approx. 0.01 Approx.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.3 -0.01 x/c
0
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 0 x/c
x/c 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x/c
0.1
Random: Pressure Surface
0.2
Random: Suction Surface Random: Pressure Surface
Random: Suction Surface
Reflected: Pressure Surface
Reflected: Pressure Surface
Reflected: Suction Surface 0.05
0.1 Reflected: Suction Surface
0 0
0 20 40 60 80 100 120 140 160 0 50 100 150 200 250 300 350 400 450
Iteration Iteration
Figure 5. NASA LRN(1)-1007: Effect of Boundary Figure 6. NASA LRN(1)-1007: Effect of Boundary
Condition on Objective Function; n = 4 Condition on Objective Function; n = 20
The convergence evolution of the three target sections in Fig. 2, over maximum and minimum solution set, n is
presented in Table 5 through the Hicks-Henne shape functions, with boundary wall condition variations. The shaded
regions indicate lower costs between the two conditions examined. In a solution set with 4 variables, the reflected
condition is superior for all three airfoils over the suction side and provides lower cost distribution on the pressure
side for only one of the three airfoils. Hence, the reflected condition indicates superior convergence over the random
methodology for this test case. The search process is not compromised with local search pattern, an attribute of the
reflecting methodology due to the restricted search space. The method is computationally inefficient across the
examined solution space, due to the possible oscillation of particles between the inner and outer boundaries which
only settle when the convergence criteria is satisfied. For large search domain (n = 20), the random methodology
exhibits superior performance in all cases (Table 5). A global search process is required for multi-dimensional
solution domains, and the random initialization of particles, promotes this search pattern. A linear/local search
9
American Institute of Aeronautics and Astronautics
process through the reflecting condition, was ineffective in this case, thus suggesting the requirement of
implementing the random methodology, for optimization studies with large population of n.
Table 5. Effect of Boundary Condition on Hicks-Henne Geometrical Convergence
∆f Random ∆f Re flected ∆f Random ∆f Re flected k Random k Re flected k Random k Re flected
Airfoil Suction Suction Pr essure Pr essure Suction Suction Pr essure Pr essure
The Hicks-Henne provides the lowest fitness for the Figure 7. Geometrical Cost Magnitude – PSO
three airfoils over the balance parameterization Random Initialization Methodology
methods. The Legendre was omitted from the analysis due to the computational expense attributed to the derivation
of a factorial and derivative term per iteration. This included small design variable population size of n =2, for
which total convergence time of 8 and 14 minutes (NASA LRN(1)-1007 and NASA LS(1)-0417 Mod airfoil) was
recorded respectively on a CPU running at 2.20GHz with 1.0 GB of RAM.
10
American Institute of Aeronautics and Astronautics
Hicks-Henne Polynomials: Effect of Optimization Methodology on Shape Convergence
1.4
NASA LRN(1)-1007
of the initial solution starting the search phase within a local NASA LS(1)-0417 Mod
NASA NLF(1)-1015
1.2
region is possible, and is influenced by the location of initial
guess. To address this, the PSO through the validated random 1
estimate of αH-H and pii. These parameters were the inputs into the
0.6
line search from Fig. 4, to reduce the function from multi-to-
single variable, hence deriving the magnitude of the design 0.4
variables. The PSO provides a valid staring point for the line 0.2
The analysis has shown that Bernstein, NACA and Wagner Figure 9. Comparison of Uni & Hybrid Based
functions encounter shape convergence issues (Fig. 7-8). The Optimization Technique for Hicks-Henne Function
Legendre polynomials were computationally inefficient, within
NASA LRN(1)-1007: Cost Function History through PSO and Hyrbid Methodology
the line search model, hence was not considered in the exhaustive 1
10
PSO: Suction Surface
search process of a PSO algorithm. The Hicks-Henne model, PSO / Line Search: Suction Surface
PSO: Pressure Surface
provided the lowest fitness of all the shape polynomials tested PSO / Line Search: Pressure Surface
Objective Function ( ∆ f)
model and a hybrid technique indicate performance improvement
with the hybrid technique due to a lower cost distribution (Fig. 9).
This provides the avenue for further analysis. -1
10
hybrid technique with the Hicks-Henne model, over the three 0 50 100 150 200 250
Iteration
benchmark airfoils is presented in Fig. 10-12. Convergence of the Figure 10. Hicks-Henne Fitness Convergence through the
NASA LRN(1)-1007 airfoil (Fig. 10), indicates the hybrid PSO & Line Search Model: NASA LRN(1)-1007
technique as computationally efficient with a lower cost and
NASA LS(1)-0417 Mod: Cost Function History through PSO and Hybrid Methodology
iteration count over the suction side. The hybrid technique 10
1
provides a 3% improvement in fitness and 70% reduction in total PSO / Line Search: Suction Surface
PSO: Pressure Surface
PSO / Line Search: Pressure Surface
iteration count on the upper surface. On the lower surface, a
fitness reduction of 27% is observed with the hybrid technique, at 0
O bjective Function ( ∆ f)
10
a convergence rate approximately 20% greater than the standard
PSO run. The pressure side converges at a slow rate and is
attributed to strong geometric variations within the trailing edge.
The geometrical functions encounter complications in providing 10
-1
method populating at local minimum. This is interrupted by the Figure 11. Hicks-Henne Fitness Convergence through the
PSO model that works to re-direct the search towards a global PSO & Line Search Model: NASA LS(1)-0417 Mod
solution, resulting in a fluctuating phase. Hence, the hybrid model
NASA NLF(1)-1015: Cost Function History through PSO and Hyrbid Methodology
is efficient in terms of fitness and iteration time for the NASA 2
10
PSO: Suction Surface
LRN section in comparison to the standard PSO model. PSO / Line Search: Suction Surface
PSO: Pressure Surface
The solution plot for the NASA LS(1)-0417 Mod airfoil (Fig. 10
1
PSO / Line Search: Pressure Surface
distribution of the airfoil is considerably lower than the 0 50 100 150 200
Iteration
250 300 350 400
magnitude observed over the low Reynolds number airfoil (Fig. Figure 12. Hicks-Henne Fitness Convergence through the
PSO & Line Search Model: NASA NLF(1)-1016
11
American Institute of Aeronautics and Astronautics
10). This is due to convergence evolution over the trailing edge region. Thus, the hybrid technique indicated lower
fitness in modeling the NASA LS airfoil. Iteration count between the two optimizers was also negligible.
The hybrid technique provides accurate geometrical convergence on the suction side of the NLF section (Fig.
12). A 25% reduction in fitness and 78% less computation iteration time is required with the hybrid method over the
PSO run in modeling the suction side (Fig. 12). Similarly, the pressure side is also modeled with fewer iterations
(37%), and a fitness reduction of 20% in comparison to a PSO model. Iteration fluctuations are similar to the trend
in Fig. 10, due to the presence of local minima’s within the solution set. Thus, the hybrid optimizer produces higher
fitness and iteration count in comparison to the analysis presented in Fig. 10-11.
The results presented in this section provides the following:
A PSO model with random initialization of particles boundary condition, provided accurate and rapid shape
convergence over the reflecting method (Fig. 5-6);
In single variable based functions (Wagner, Legendre, Bernstein, and NACA normal modes), gradient and global
search agents provided similar geometrical convergence (Fig. 7-8), with the line search methodology proving to
be computationally efficient;
The Hicks-Henne method provided the lowest fitness of all the shape functions tested, regardless of the
optimization technique utilized (Fig. 7-8);
A hybrid technique with a PSO model was used to obtain an initial estimate of αH-H and pii within the Hicks-
Henne function (Table 2). The parameters are then integrated into a line search model, that provided lower
fitness in comparison to a standard PSO algorithm, over the three target airfoils (Fig. 9); and
The benefit of the hybrid optimization technique, with the Hicks-Henne model, is shown in Fig. 10-12 across
the three benchmark airfoils, with lower fitness and rapid convergence over both upper and lower surfaces.
V. Aerodynamic Convergence
In this section the second unit of the DNO process is presented. It involves selection and validation of a suitable
flow solver. A high fidelity CFD solver was implemented in this study. The pre-processing analysis is in GAMBIT
and the computation process is in FLUENT. Both pre and post processing systems require a balance between
computational efficiency and solution accuracy. It is a compromise between coarse and fine mesh and choice of
turbulence modeling to achieve the required results.
12
American Institute of Aeronautics and Astronautics
the PRESTO interpolation method applied to calculate the face pressure within the segregated solver.20 Turbulence
modeling with the Spalart-Allmaras (S-A)21 and Menter’s Shear Stress Transport (SST) κ-ω22, 23 with enhanced wall
treatment, were independently applied in this study. These models were previously tested with success in predicting
the aerodynamics at pre-stall angles of attack, and the SST models were applied to compute post-stall trends.24-26
The velocity inlet boundary condition was defined with a turbulence intensity and viscosity ratio of 0.5% and 5
respectively, and the pressure outlet specified with a backflow turbulent intensity of 2% and a viscosity ratio of 20.
L
C
a) Flow solver is used to compute and compare the aerodynamic 1
Manually tripping of the boundary layer provides an opportunity Figure 14. Lift Curve Slope Validation;
to test the suitability of adopting RANS within the overall DNO R e = 6.0 × 10 6 ; Mach = 0.32
process of airfoil design for a RC-MM-UAV. Drag Polar Curve Validation
The four grids generated were used in the validation process 2
over three separate case studies as follows: a) The κ-ω SST 1.8
model, with a fully turbulent flow condition; b) The κ-ω SST and 1.6
Transition - κ-ω
Data presented in Fig. 14-16, is based on grid that provided 1 Fully Turbulent(κ-ω)
modeling, approximates lift that is within 3% of experimental Figure 15. Drag Polar Curve Validation;
data. The fully turbulent model exhibits lower lift and the onset of R e = 6.0 × 10 6 ; Mach = 0.32
flow separation is at higher angles of attack. The effectiveness of Effect of Turbulence Modelling on Coefficient of Pressure Convergence
-4
the three case studies was established from the drag polar (Fig. Exp.
-3.5 Transition: Spalart-Allmaras
15), where drag was within 4% of experimental solutions for the Fully Turbulent: κ-ω
κ-ω SST flow transition methodology over the testing angle-of- -3 Transition: κ-ω
attack range. The full turbulent condition over estimates drag by -2.5
-1.5
C
13
American Institute of Aeronautics and Astronautics
identical results and in line with experimental solutions. Downstream of x / c = 0.4 and towards the trailing edge, the
tripping models provide acceptable solution agreement in comparison to the fully turbulent condition which
underestimates the aft loading of the section.
Based on the analysis over the three case studies, the κ-ω SST transition modeling scheme is most effective
through grid 4 as shaded in Table 6. The lift coefficient was underestimated across the four grids with the magnitude
of percentage difference remaining negligible, while the evolution of c d was within acceptable agreement. The total
average percentage difference over the testing angle-of-attack domain for lift and
drag, %∆clα [−1,0,K,12] and %∆c dα [−1,0,K,12] further indicated similar cl trend over the four grids, and c d difference was
minimal.
Table 6. κ-ω Grid Lift and Drag Coefficient Evaluation
Total Cell %∆ Cl %∆ Cd
Grid y+ Cl [α=10°] Cd [α=10°]
Count [α -1,0,…,12] [α -1,0,…,12]
Exp.1 ---- ---- 1.67 ---- 0.0155 ----
1 31,580 ≈ 200-450 1.59 (≈5% ↓) 2.4 0.0147 (≈5% ↓) 5.8
2 48,120 ≈ 200-450 1.60 (≈4% ↓) 2.3 0.0143 (≈8% ↓) 7.1
3 55,160 ≈ 50-160 1.58 (≈5% ↓) 2.6 0.0163 (≈5% ↑) 6.4
4 96,200 ≈ 150-275 1.57 (≈6% ↓) 2.7 0.0157 (≈1% ↑) 4.3
0.01
C
fully turbulent model exhibits a very sharp and sudden rise 0.004
boundary layer was manually tripped at x / c = 0.04575. 1 Figure 17. Suction Side Skin Friction Coefficient
Following the laminar region, the onset of turbulent flow is evident with the sudden increase in skin friction, where
the peak of the S-A model is in-line with the fully turbulent condition. The κ-ω SST model indicates an increase of
c f due to the onset of turbulent flow, although the gradient to the peak is considerably low and delayed. A decrease
in c f indicates a region of flow separation and further reattachment downstream, hence the delay. Upon
reattachment, c f rises to its maximum which is considerably lower than the two models presented.
The c f evolution of the three conditions is in-line with the computed lift and drag coefficients (Table 6). The
fully turbulent condition provides discrepancy in lift and drag based on experimental solution, while the tripping
term within S-A induces slight improvement due to the implementation of a laminar region (Table. 7). Drag
prediction with S-A remains high and is related to the c f evolution, which is similar to the fully turbulent condition
(Fig. 17).
Table 7. Effect of Turbulence Modeling on Lift and Drag Coefficients; α=10°
Turbulence Model Cl Cd
Exp.1 1.67 0.0155
Fully Turbulent: κ-ω SST 1.50 (≈10% ↓) 0.0214 (≈38% ↑)
Tripping: Spalart-Allmaras 1.59 (≈5% ↓) 0.0169 (≈9% ↑)
Tripping: κ-ω SST 1.57 (≈6% ↓) 0.0157 (≈1% ↑)
14
American Institute of Aeronautics and Astronautics
Wall Normal Velocity Magnitude
Local normal velocity component v / U ∞ , at four wall 2
x/c=0.03
stations x / c = 0.03, 0.05, 0.3, and 0.5 for the κ-ω SST model 1.8 x/c=0.05
x/c=0.3
x/c=0.5
is presented (Fig. 18) to show the effect of flow transition 1.6
y/c
1
15
American Institute of Aeronautics and Astronautics
c p , fluctuations are evident that indicate the geometrical function has not converged. The performance mismatch for
the NASA LRN section (Fig. 22), was attributed to the inadequacy of the Hicks-Henne in providing sufficient
geometrical resolution. The performance fluctuations in this case (Fig. 24) is attributed to an issue within the flow
solver as both benchmark and final sections undergo a similar oscillation phase. A fitness magnitude of 0.0175 is
almost identical to 0.0179 for the low Reynolds number airfoil (Fig. 19), though the aerodynamic convergence
between the two differs considerably in Fig. 22 and Fig. 24 respectively. This further validates that the cost
magnitude is sensitive in the design of low Reynolds number airfoils.
NASA LRN(1)-1007 Shape Convergence NASA LRN(1)-1007 Coefficient of Pressure Convergence
0.1
-1 Initial
Benchmark
0.08 Final (Hybrid: Hicks-Henne)
-0.8
0.06 -0.6
0.04 -0.4
-0.2
0.02
y/c
CP
Initial 0
0 Benchmark
BP(∆f)=0.0527 0.2
-0.02 Hicks-Henne(∆f)=0.0179
Legendre(∆f)=0.1123 0.4
-0.04 NACA(∆f)=0.0274
0.6
Wagner(∆f)=0.0197
-0.06
0.8
-0.08 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
x/c x/c
Figure 19. NASA LRN(1)-1007 Airfoil Shape Figure 22. NASA LRN(1)-1007 Cp Convergence
Function Convergence
NASA LS(1)-0417 Mod Shape Convergence NASA LS(1)0417 Mod Coefficient of Pressure Convergence
0.15
-1.5 Initial
Benchmark
Final (Hybrid: Hicks-Henne)
0.1
-1
0.05 Initial
-0.5
Benchmark
y/c
BP(∆f)=0.0246
C
Hicks-Henne(∆f)=0.0238
0 Legendre(∆f)=0.025 0
NACA(∆f)=0.0412
Wagner(∆f)=0.032
-0.05 0.5
-0.1 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
x/c x/c
Figure 20. NASA LS(1)-0417 Mod Airfoil Shape Figure 23. NASA LS(1)-0417 Mod Cp
Function Convergence Convergence
NASA NLF(1)-1015 Shape Convergence NASA NLF(1)-1015 Coefficient of Pressure Convergence
0.12 Initial
-1 Benchmark
Final (Hybrid: Hicks-Henne)
0.1
-0.8
0.08
-0.6
0.06
Initial -0.4
0.04 Benchmark
BP(∆f)=0.0297 -0.2
y/c
CP
0.02 Hicks-Henne(∆f)=0.0175
Legendre(∆f)=0.0558 0
0 NACA(∆f)=0.0271
0.2
Wagner(∆f)=0.0197
-0.02
0.4
-0.04
0.6
-0.06 0.8
-0.08 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
x/c x/c
Figure 21. NASA NLF(1)-1015 Airfoil Shape Figure 24. NASA NLF(1)-1015 Cp Convergence
Function Convergence
16
American Institute of Aeronautics and Astronautics
VI. Concluding Remarks
The DNO test methodology developed a validated architecture for the design of RC-MM-UAV airfoils. A shape
parameterization study examined the analytical approach by measuring the accuracy and flexibility of five geometric
shape functions in converging to a set of benchmark airfoils. The line search and PSO optimizers were
independently simulated over the three target airfoils, with the exception of Hicks-Henne that required integration of
the two optimizers; as multiple variables were to be resolved. It was concluded, that a multi-variant Hicks-Henne
method was superior over single function based models due to the lower geometrical cost and quicker solution
convergence. A hybrid optimization model was needed to address multi-variant functions. The PSO was used to
reduce the design space from multi to single based function and the line search used this initial estimate to guide the
search towards a minimum state. Aerodynamic convergence study indicated that low Reynolds number airfoils,
primarily used in the design of micro UAVs are highly sensitive to surface perturbations and the adopted test
methodology was not adequate in achieving sufficient convergence. Acceptable geometric and aerodynamic
convergence was established for airfoils operating at Reynolds number in excess of 2.0 million. This test case
presented its suitability for application in the design of RC-MM-UAV project.
Further work is needed to develop a relationship between design variable population size and the aerodynamic
coefficients through a statistical analysis to quantitatively present the relationship between the two terms. A data-
mining study through self-organizing maps similar to the PARSEC representation,30 will further illustrate the
relationship between each design variable and the overall aerodynamics. This will provide the avenue to eliminate
the variables that have marginal or no effect on the aerodynamics.
Within the overall DNO process, a flow solver validation study provided adequate lift and drag convergence in
comparison to experimental data. The study was limited since the location of the tripping point was manually
specified within the solver. A RANS solution through a two-equation turbulence model was seen to provide
acceptable agreement for a linear angle-of-attack range. Further studies will require establishing a finer grid with
y + ≈ 1 to enhance the accuracy of the computation process. The issue of boundary layer transition and modeling of
high angles-of-attack aerodynamics, is critical for the overall DNO process. The fully turbulent flow simulation
provided results that were not comparable to experimental solutions, thus illustrating the importance of
incorporating a flow transition model. Present research focuses on testing the effectiveness of a Detached Eddy
Simulation to predict high-lift flows. The boundary layer transition calculations based on the e n method are also an
area of active investigation for eventual implementation within the flow solver.
Acknowledgment
Research candidate Omar Ilaya – The Sir Lawrence Wackett Aerospace Centre, for his technical guidance with
the Particle Swarm Optimizer.
17
American Institute of Aeronautics and Astronautics
References
1
McGhee, R. J. and Beasley, W. D., "Wind-Tunnel Results for a Modified 17-Percent-Thick Low-Speed Airfoil Section,"
NASA, 1981.
2
Wong, D. K. C., "Aerospace Industry Opportunities in Australia - Unmanned Aerial Vehicles (UAVs) - Are They Ready
This Time? Are We?," 1997, pp. 13.
3
Khurana, M., Sinha, A. and Winarto, H., "Multi-Mission Re-Configurable UAV - Airfoil Analysis through Shape
Transformation and Computational Fluid Dynamics," Twelfth Australian International Aerospace Congress - Second
Australasian Unmanned Air Vehicles Conference, 2007, pp. 1-22.
4
Samareh, J. A., "A Survey of Shape Parameterization Techniques," CEAS/AIAA/ICASE/NASA Langley International Forum
on Aeroelasticity and Structural Dynamics, 1999, pp. 333-343.
5
Keane, A. J. and Nair, P. B., Computational Approaches for Aerospace Design: The Pursuit of Excellence, Wiley, John &
Sons, Incorporated, 2005.
6
Namgoog, H., "Airfoil Optimization of Morphing Aircraft," PhD, Aerospace Engineering, Purdue, Indiana, 2005.
7
Song, W. and Keane, A. J., "A Study of Shape Paramaterisation Methods for Airfoil Optimisation," AIAA/ISSMO
Multidisciplinary Analysis and Optimization Conference, 2004.
8
Khurana, M., Sinha, A. and Winarto, H., "Multi-Mission Re-Configurable UAV - Airfoil Shape Parameterisation Study,"
22nd International Unmanned Air Vehicle Systems Conference, 2007, pp. 1-15.
9
Henne, P. A. and Gregg, I., Robert D, McDonnell Douglas Corporation (Long Beach, CA), USA, Patent Application for a
"Divergent Trailing-Edge Airfoil," Docket No. 056250, filed 22 August 1989
10
Wu, H.-Y., Yang, S. and Liu, F., "Comparison of Three Geometric Representations of Airfoils for Aerodynamic
Optimisation," 16th AIAA Computational Fluid Dynamics Conference, 2003.
11
Hicks, R. M. and Henne, P. A., "Wing Design By Numerical Optimisation," AIAA Aircraft Systems & Technology Meeting,
1977, pp. pp 8.
12
Ramamoorthy, P., Dwarakanath, G. S. and Narayana, C. L., "Wagner Functions," National Aeronautical Laboratory, 1969.
13
Chang, I.-C., Torres, F. J. and Tung, C., "Geometric Analysis of Wing Sections," NASA Technical Memorandum, 1995.
14
Mathworks, "Optimization Toolbox User's Guide," 2005.
15
Eberhart, R. and Kennedy, J., "A New Optimizer Using Particle Swarm Theory," International Symposium on Micro
Machine and Human Science, 1995.
16
Eberhart, R. C. and Shi, Y., "Comparing Inertia Weights and Constriction Factors in Particle Swarm Optimization,"
Congress on Evolutionary Computation, Vol. 1, 2000, pp. 84-88.
17
Shi, Y. and Eberhart, R., "A Modified Particle Swarm Optimizer," IEEE World Congress on Computation Intelligence,
1998, pp. 69-73.
18
Yong-ling, Ma, L.-h., Zhang, L.-y. and Qian, J.-x., "Empirical Study of Particle Swarm Optimizer with an Increasing
Inertia Weight," 2003.
19
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.
20
FLUENT, "Introductory Fluent Notes," FLUENT, 2005.
21
Spalart, P. R. and Allmaras, S. R., "A one-equation turbulence model for aerodynamic flows," 30th AIAA Aerospace
Sciences Meeting & Exhibit, 1992,
22
Menter, F. R., "Performance of Popular Turbulence Models for Attached and Separated Adverse Pressure Gradient Flows,"
AIAA Journal, Vol. 30, No. 8, 1992, pp. 2066-2072.
23
Menter, F. R., "Two Equations Eddy-Viscosity Turbulence Models for Engineering Applications," AIAA Journal, Vol. 32,
No. 8, 1994, pp. 1598-1605.
24
Kern, S., "Evaluation of Turbulence Models for High-Lift Military Airfoil Flowfields," 34th Aerospace Sciences Meeting &
Exhibit, 1996.
25
Fuhrmann, H., "Design Optimisation of a Class of Low Reynolds, High Mach Number Airfoils For Use in the Martian
Atmosphere," 23rd AIAA Applied Aerodynamics Conference, 2005.
26
Kotapati-Apparao, R. B., Squires, K. D. and Forsythe, J. R., "Prediction of the Flow over an Airfoil at Maximum Lift,"
42nd AIAA Aerospace Sciences Meeting and Exhibit, 2004, pp. 1-13.
27
Evangelista, R., Pfenninger, W., Mangalam, S. M. and Bar-Sever, A., "Design and Wind Tunnel Test of a High
Performance Low Reynolds Number Airfoil," AIAA Applied Aerodynamics Conference, Vol. 5th, 1987, pp. 175-185.
28
Somers, D. S., "Design and Experimental Results for a Natural-Laminar-Flow Airfoil for General Aviation Applications,"
Langley Research Center, 1981.
29
Maughmer, M. and Somers, D. M., "An Airfoil Designed for a High-Altitude, Long Endurance Remotely Piloted Vehicle,"
AIAA Applied Aerodynamics Conference, Vol. 5th, 1987, pp. 539-545.
30
Jeong, S., Chiba, K. and Obayashi, S., "Data Mining for Aerodynamic Design Space," Journal of Aerospace Computing,
Information and Communication, Vol. 2, 2005, pp. 452-469.
18
American Institute of Aeronautics and Astronautics