Académique Documents
Professionnel Documents
Culture Documents
M. Tech Dessertation
MASTER OF TECHNOLOGY
by
Tushar S. Kanikdale
(02301018)
Examiners
(Internal)
(External)
Chairman
Date:
i
Acknowledgement
I take this opportunity to express my gratitude towards my project guide Prof. A. G.
Marathe and co-guide Prof. R.K.Pant for showing the way and guiding me constantly at
all stages, while working on this project. Infact the sound understanding of the problems
and clearer communucation of ideas and sugestions by them can said to have done much
of this project work.
I also extend my sincere thanks to my colleague Mr. Piyush Tagde for giving me his
precious time to accomplish the optimization through DACE.
I personally feel grateful to Ms. Preeti Khot and Charu Patil for developing the GUI
relevant to the present work to assist and partially automate the entire optimization process
loop.
Tushar S. Kanikdale
(02301018)
June. 2004
ii
Abstract
This work aims at the arriving at the optimum shape of an airship envelope of fixed vol-
ume from aerodynamic and structural considerations, using the concept of multi-disciplinary
design optimization. The multi-disciplinary nature of the problem is discussed and the cor-
responding objective functions and the constraints are elucidated. The strategies adopted
are described along with the shape generation algorithm. Fluent CFD code has been used
for computational experimentation in conjunction with SIMANN, simulated annealing
code for optimization. The low fidelity model available in the literature was found to give
large deviations from the physical value of envelope drag. The mathematical model for
envelope drag estimation has been developed based on the detail study of results obtained
through CFD experimentation. It was observed that the shape optimized for minimum drag
shows a significant improvement over a reference GNVR shape. The multi-disciplinary
nature of the problem is addressed by devising a composite objective function and corre-
sponding results are presented.
Keywords: airship envelope shape, optimization, CFD, simulated annealing, mathemti-
cal models
iii
Contents
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii
Nomenclature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
1 Introduction 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Objective and Scope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2 Multi-Disciplinary Aspects 4
2.1 Multi- Disciplinary Optimization Problem . . . . . . . . . . . . . . . . . . 4
2.1.1 Influence of Disciplines . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.2 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Strategy for Solving Problem . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2.1 Description of Available Tools . . . . . . . . . . . . . . . . . . . . 7
iv
4.1.3 Validation of Shape Generation Program for Min. Surface Area . . 23
References 51
v
List of Figures
6.1 Comparison showing Optimized Shape for CDV and GNVR shape . . . . . 33
6.2 CDV Vs Surface Area for Fixed Thickness Ratio . . . . . . . . . . . . . . . 34
6.3 CDV Vs (Xymax %/l) for Fixed Thickness Ratio . . . . . . . . . . . . . . . . 35
6.4 Optimized Shape with Improved Fidelity Model for CDV Vs GNVR Shape . 36
6.5 Optimized Shapes for CDV with Fixed and Variable Length Configurations . 37
6.6 Variation of Parameter β with Thickness Ratio . . . . . . . . . . . . . . . . 39
6.7 Variation of CD α with α . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
vi
6.8 Variation of Coefficient A and B with d l . . . . . . . . . . . . . . . . . 41
6.9 Comparison Showing Optimized Shapes for Fixed and Variable Length
Configurations Vs GNVR Shape . . . . . . . . . . . . . . . . . . . . . . . 42
6.10 Optimized Shapes for ’S’ with Fixed and Variable Length Configurations
Vs GNVR Shape . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
6.11 Comparison Showing Optimized Shape with Composite Obj. Function Vs
GNVR Shape . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
vii
List of Tables
6.1 Comparison of Optimized Shape for CDV and GNVR Shape using Low
Fidelity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
6.2 Comparison between the Results Predicted by Low Fidelity and Fluent code 33
6.3 Results with Improved Fidelity . . . . . . . . . . . . . . . . . . . . . . . . 36
6.4 Comparison Showing the Results Predicted by Intermediate and High Fi-
delity Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
6.5 Comparison Showing the Results Predicted by Improved Fidelity and Flu-
ent Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
6.6 Results with Improved Fidelity . . . . . . . . . . . . . . . . . . . . . . . . 43
6.7 Comparison Showing the Results Predicted by Improved and Fluent code . 43
6.8 Comparison for Min.’S’ Shapes for Fixed and Variable Length Configura-
tions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
6.9 Multi-Disciplinary Optimization using Composite Objective Function . . . 46
viii
Nomenclature
a Major radius of ellipse (A-B) in figure3.5
b Minor radius of ellipse (A-B) in figure3.5
CD Coefficient of drag
CDV Volumetric coefficient of drag
Cp Coefficient of pressure
d Maximum diameter of envelope
fb Bending stress in membrane
fh Hoop stress
g Acceleration due to gravity
I Membrane second moment of inertia
L Lift
l Length of envelope
M bending moment
p internal pressue of airship
N Number of grid points
R Radius of leading spherical envelope cap
R1 Major Radius of second ellipse (B-C) in figure 3.5
R2 Minor Radius of second ellipse (B-C) in figure 3.5
Re Reynold’s number
r radius at at any cross-section of airship
S Total surface area
T Thrust
t Thickness of airship membrane
V Free stream velocity
w1 w2 w3 Weight functions
W0 Total weight of airship
x1 X-coordinate of the start point of spline-I
x2 X-coordinate of the start point of spline-II
x3 X-coordinate of the end point of spline-II
Xymax X-coordinate of the position of maximum diameter x2
y1 Y-coordinate of the start point of spline-I
y2 Y-coordinate of the start point of spline-II
y3 Y-coordinate of the end point of spline-II
ix
Symbols
ρa Density of air
ρhe Density of helium
σmax Maximum hoop stress per unit thickness
Acronym
DACE Design and Analysis of Computer Experiments
SPACE Stochastic Process Analysis for Computer Experiments
x
Chapter 1
Introduction
An airship is a lighter than air vehicle which produces the significant lift due aerostatic
effect or buoyancy force. It differs from the conventional aircraft in terms of lift producing
mechanism, and thereby derive its associated features. The potential of the airship can be
realized in terms of less fuel consumption, high endurance, and ability to hover. There
performance is curtailed due to low speed operation, high sensitivity to wind gusts, lower
payload capacity, and operataion at relatively lower altitude due to density gradient.
The envelope, which generates lift, is the most important component of an airship. The
envelope volume is determined from the required payload and maximum operating altitude,
whereas the shape is usually decided from aerodynamic considerations, i.e. minimum drag.
This work aims at the arriving at the optimum shape of an airship envelope of fixed volume
from aerodynamic and structural considerations, using the concept of multi-disciplinary
design optimization. The aerodynamic considerations are reflected through minimum vol-
umetric drag coefficient, whereas the structural considerations are brought in through min-
imum surface area and hoop stress, both of which directly result in lowest envelope weight.
The manufacturing constraints are implicitly taken care, in the form of lowest rate of change
of curvature of the envelope along the length.
1.1 Background
The airship is a air vehicle which usually operates at lower altitude and lower speed not
beyond 180 Km/hr. The free body diagram of airship moving in air is shown in Fig.1.1
The airship is comprised of a bare hull along with the control surfaces and lower portion
gondola attached to it. It is filled by a lighter than air gas which is generally Hydrogen or
1
CHAPTER 1. INTRODUCTION 2
Bouyant Force
Thrust Gondola
Helium, and hence get the lift due to Archemedes principle where net lift produced by the
airship is given by
L V ρa ρg g W0 (1.1)
1.2 Motivation
Airships are currently used for various applications like transport over inhospitable regions,
advertisement, survillieance, tourism, and many more. Though airship operation seems to
be simpler, its complete design requires the balance between the aerodynamic, structures
and fabrication aspects as improvement in one discipline may impair the performance in
other. As compared to conventional aircraft, airship design demands less design variables,
it is always practicable and effective to implement the design conceived through multi
disciplinary optimization (MDO) where participating disciplines would be mainly aerody-
namics, structures and manufacturing. Because of simplicity in operation and body config-
uration airship can actually be built at the local place at a very cheap price. Small airships
CHAPTER 1. INTRODUCTION 3
can easily be manufactured meant for survilience or advertisement. There design can be
conceived after preliminary analysis and thumb rules. D.Howe [1] has examined the feasi-
bility of large freight airships in terms of low operating cost, lower structural weight and
minimum fuel consumption at design speed. To built such large airship to carry passenger
or heavy payload, it requires the study of various aspects of airship and there interactions
to achieve risk less and efficient flight.
Multi-Disciplinary Aspects
This chapter addresses the multi-disciplinary nature of the problem and the influence of
each discipline on the envelope shape design. The basic strategy to attempt the MDO prob-
lem has been discussed. The description of tools required has been outlined.
4
CHAPTER 2. MULTI-DISCIPLINARY ASPECTS 5
STRUCTURES
MIN. MIN
HOOP STRESS BENDING STRESS
(Wc)
CD MIN AIRSHIP FABRICATION
AERODYNAMICS MAX PAYLOAD,MAX VELOCITY, MANUFACTURING
(Wa) EASINESS
MAX LENGTH ETC.. (Wb)
FINAL DESIGN
OBJECTIVE FUNCTION
F=FaWa+FbWb+FcWc
3 2
7 d l 3
(2.1)
Since the thrust required for the airship is dictated by the total drag ,it is preferable to
take drag as a objective function than CD . Hence the coefficient of drag based on volu-
metric area (CDV ) which gives direct indication of drag, is adopted as one of the objective
functions. This volumetric coefficient of drag (CDV ) can be converted to coefficient of drag
CHAPTER 2. MULTI-DISCIPLINARY ASPECTS 6
Volume 2 3
CDV
CD (2.2)
S
Structures : From structures point of view the aerodynamic shape should give mini-
mum bending and hoop stresses. Bending stress is the function of length and can be calcu-
lated using the formula given by Khoury and Gillett [3] as
fb 1 t M r t I
T p π r2 2 π r
(2.3)
p d
fh (2.4)
2 t
Hess [4] has elucidated the actual structural design of airship along with the various
structural configurations. Here it can be said that for the same max. diameter, increase
in length improves the aerodynamic performance but impairs the structural aspects of an
airship.
Manufacturing : The manufacturing of the airship should be as easy as possible.
Conventionally the airships are manufactured by stitching together number of petals made
up of fabric. The fabrication easiness can be expressed mathematically using rate of change
of curvature of the envelope shape and hence in turn poses the additional constraints.
The multi-disciplinary nature of the problem involves more than one objective functions
each corresponding to the participating disciplines. In order to attempt such MDO problem
the two usual approaches followed are
- Solve the optimization problem for the primary objective function and include addi-
tional objective functions in the form of constraints posed by rest of the disciplines.
- Construct the composite objective function which will have the weighted combination
of all the objective functions. The weight-age are usually assigned based on the require-
ments and past experience. Fig.2.1 further clarifies the idea.
and fixed length constraint has arisen due to the stability requirement where the envelope
length has strong influence over the control surface area. The total drag, envelope weight
and maximum stress induced in the membrane are considered as three possible objective
functions. To address the multi-disciplinary nature of the problem a composite objective
function is defined as follows
Fcomp w1 CDV
w2 S
w3 σmax (2.5)
The design vector XD is a set of design variables which completely defines the aerody-
namic shape and hence the objective function. The solution of the optimization problem is
the design vector XD for which given objective function gets minimized or maximized as
per the requirements.
DESIGN VECTOR
XD
SHAPE GENERATION
CODE
OPTIMIZER
MODIFIED SHAPE
E
A
DESIGN VARIABLE (X )
based methods. The basic algorithm of calculus based method for a minimization problem
is given as follows
1. Find the gradient at the current point. Gradient gives the direction of maximum
change of the objective function, in the design space.
2. Move a small distance in the direction of negative gradient to get the next design
vector
4. If the current objective function becomes gre greater than previous, stop. Otherwise
go to step no.2.
Furthermore there are direct and indirect methods to handle constrained optimization
problems depending upon whether the constraints are treated explicitly or not. Steepest
descent method is one of the widely accepted methods as stated by Arora [6]. Gradient or
calculus based methods does not ensure the global optima and are used where the region in
CHAPTER 2. MULTI-DISCIPLINARY ASPECTS 10
the design space where global optima is expected, is known for sure. These methods gives
the results quickly and hence computationally less time intensive.
The actual tool selected for flow analysis and its validation is discussed in the next
chapter. The primary approach followed to attempt the problem has also been presented.
Chapter 3
This chapter discusses the tools selected for flow analysis for the concerned problem of
aerodynamic shape optimization. The results to validate the CFD software are presented.
The initial strategy adopted to attempt the problem using DACE has also been discussed.
11
CHAPTER 3. CFD TOOL VALIDATION AND PRIMARY STRATEGY ADOPTED 12
tion is found to be 1000 m3 . Past studies [9] indicated that this shape corresponds to least
volumetric coefficient of drag CDV under the given operating conditions. One of the aims
of this work is to confirm this and to obtain a better shape if possible.
Fluent code required around 2.5 hrs to reach to steady state solution on a standard
Pentium III (650 Hz) machine. The coefficient of pressure (C p ) distribution over the air-
ship envelope surface is compared with reference values for GNVR shape calculated using
source panel method by Narayana [9] and is shown in Fig. 3.3. The source panel method
uses the potential theory and is used for incompressible and inviscid fluids. The details of
this method are given by Anderson [10]. The values closely match except near the trailing
edge because of the flow seperation. The influence of computational domain size has also
been studied and it is found that the solution obtained with semicircular computational do-
CHAPTER 3. CFD TOOL VALIDATION AND PRIMARY STRATEGY ADOPTED 14
Pressure 0.3
Density 1.0
Body Forces 1.0
Momentum 0.7
Modi f ied 0.8
turbulent viscosity
Turbulent viscosity 1.0
Energy 0.8
main with 10l radius exactly matches with that of 5l computational domain size. To study
the effect of turbulence model standard k-ε model is used in place of S-A model. It is found
that C p distribution for both the models matches with each other. S-A model has been pre-
ferred which many a times results in the computational economy being a one equation
model [11].
1
"Fluent Code"
"Source Panel Method"
0.8
0.6
0.4
Cp
0.2
-0.2
-0.4
0 0.2 0.4 0.6 0.8 1
x/l
Table 3.3 provides the other relevant information regarding the Fluent code steady state
solution.
CHAPTER 3. CFD TOOL VALIDATION AND PRIMARY STRATEGY ADOPTED 15
2 2 2 2
x y x y
2 + 2 =1
2 + 2 =1
a b 2a b
B
b
C
A O
a
Figure 3.4: The NPL Low Drag Airship Body Shape [3]
is again an ellipse only. In order to take the advantage of both NPL and GNVR shape a
generalized shape is defined and is shown in Fig.3.5.
CHAPTER 3. CFD TOOL VALIDATION AND PRIMARY STRATEGY ADOPTED 16
2 2 2 2
x y x (y−yo) 2
2 + 2 =1 + =1 y = c (xe+xp−x)
a b R1 2 R2 2
B
b C
A O (0,0) D
xp
a xe
x2 y2
1 (3.1)
a2 b2
Ellipse (B - C) :
x2 y y0 2
1 (3.2)
R21 R22
CHAPTER 3. CFD TOOL VALIDATION AND PRIMARY STRATEGY ADOPTED 17
Parabola (C - D) :
y2 c xe
x p x (3.3)
R1 b
y0
Again to maintain the continuity, slope at point C due to ellipse (B - C) = Slope at
point C due to parabola (C - D).
Coordinates of point C xe ye should satisfy the equation of ellipse (B-C) and parabola.
Mathematically it implies
x2e ye y0 2
1
R21 R22
y2e c xe
x p x
With the six above mentioned condition, out of nine unknowns the design variables
to dictate the shape becomes three. Moreover it was observed that the value of maximum
diameter b gets fixed because of geometric and specified constraints. Hence the size of the
design vector reduced to two . The two parameters selected were R1 Xe , for convenience.
The FORTRAN code was written to generate the geometry for the input of design vector
XD R1 Xe . The output of the program for various design vectors is shown in Fig.3.6
30 "coordinates21.out" u 1:2
"coordinates22.out" u 1:2
"coordinates23.out" u 1:2
"coordinates24.out" u 1:2
"coordinates25.out" u 1:2
25 "coordinates26.out" u 1:2
20
15
10
0
-10 -5 0 5 10 15 20
and Analysis of Computer Experiments (DACE) is a technique used typically for problems
with multidimensional space to predict the response of a dependent variables as a func-
tion of a independent design variables (or design vector XD ). The philosophy of DACE
lies in the fact that there is absolutely no change in the output with the given input for
computer experiments. Whereas there exists the error in the results obtained through the
physical experiment if repeated for the same input. The same fact is used by Response
Surface Methodology (RSM) allowing error at certain points in the response surface fit.
DACE ensures zero error at every point and thus differs from conventional Response Sur-
face Methodology.
Since the design vector size is small for the configuration with primary shapes, it be-
came economical to use the surrogate model technique like DACE.. The design space is
descretized using Orthogonal Latin Square method and the actual values for CDV corre-
sponding to the design points are calculated using Fluent code. The ranges for design vari-
CHAPTER 3. CFD TOOL VALIDATION AND PRIMARY STRATEGY ADOPTED 19
ables are selected such that the entire design space lies in feasible region. In the current
optimization problem DACE has been used to obtain the response surface for CDV . Mathe-
matically the surface represents the relation of CD with design vector XD (Xe ,R1 ) expressed
as
CD = F(XD )
Thus the generalized geometry composed of standard shapes like ellipse and parabola
results in GNVR shape as an optimum shape for minimum CDV . To explore the possibil-
ity of better shapes in terms of aerodynamics and other considerations like weight,stress,
it is needed to construct more generalized geometry. Hence the cubic splines along with
parabola and sphere were used to define a generalized configuration of the shape. The cor-
responding equations governing the shape in the form of geometrical requirements and
constraints are discussed in the next chapter.
Chapter 4
This chapter presents the description of better shape configuration, conceived to explore
the possibility of better shapes. The geometric elements constituting the complete envelope
shape are described along-with the governing equations. The validation of shape generation
program is discussed for the case of feasible minimum area shape.
21
CHAPTER 4. IMPROVED SHAPE GENERATION ALGORITHM 22
Spline II
y a2 x3
b2 x2
c2 x
d2 (4.3)
Parabola
y2 an l x (4.4)
The size of the design vector must be 6 to find out 11 unknowns for 5 given conditions.
The design vector for the concerned problem is selected as
XD = (x1 , y2 , x2 , x3 , y3 , l)
The linear equations were solved using Gauss Jordan method and the nonlinear equation
for volume is coupled using the technique of successive approximation. The various shapes
possible corresponding to the specified design vector are shown in Fig.4.2.
The convex envelope shapes are discarded by forcing negative radius of curvature ,
since the shapes with convexcity particularly at the trailing portion would suffer from heavy
pressure drag and such shapes are also considered to have greater penalty in terms of manu-
facturing. Moreover the low fidelity empirical formulae for CDV is applicable for streamline
bodies only. Mathematically the criterion imposed can be stated as
d 2y
dx2
0 (4.5)
tested for the case of minimum surface area as objective function where the constraint on
the length was completely relaxed to see whether the resulted shape lies closer to sphere of
volume 1000 m3 . Fig. 4.3 compares the resulted shape with the GNVR shape. The surface
area found to be 494.5m2 whereas the sphere having the same volume has surface area
value 483.6m2
This chapter outlines the basic SA algorithm and its application for the present problem.
The various models required by the optimizer to estimate the objective functions have been
described. The modified design vector description along with design variables ranges is
presented
From the equation it can be said that as the temperature comes down the probability
that the higher energy state is achieved decreases.
The analogy of annealing process can be extended to optimization method where the
25
CHAPTER 5. SIMULATED ANNEALING OPTIMIZATION METHOD 26
design variables are analogous to the molecules and the combination of the design variable
gives the minimum energy state i.e the minimum objective function point as the temperature
gradually falls down. The available simulated annealing code SIMANN was incorporated
with improved shape generation program. The information flow for the resulting code is
shown in Fig.5.1 in the form of block diagram for the case of minimization of objective
function.
The SIMANN main program supplies design vector XD to subroutine user which is
’shape generation program’ in the present application which accepts XD and estimate the
objective function as output. If this objective function is less than the previous one then it is
certainly accepted but if it is more, then the probability of its acceptance is decide using the
criterion developed by Metropolis et al. [16] which incorporates the Boltzman’s probabil-
ity distribution formula. A separate random number generation subroutine is required for
Metropolis criterion. In each iteration of the main program the objective function is eval-
uated for (n*ns*nt) times. Each design variable is perturbed for ’ns’ times with small step
CHAPTER 5. SIMULATED ANNEALING OPTIMIZATION METHOD 27
size and corresponding to perturbed design vector the objective function is evaluated. This
cycle is repeated for ’nt’times constituting the outermost loop. Since ’n’ is the number of
design variables, the total number of function evaluations become (n*ns*nt) per iteration.
The temperature gets subsequently reduced after each iteration.
A very high acceptance rate implies that the function domain is not being fully ex-
plored,while a very low acceptance rate means that the new trial points are being generated
too far away from the current optimum. Both of these imply that the algorithm is not pro-
gressing efficiently and involves wasting of computational effort. Since the probability of
acceptance of higher energy state point is high at higher temperatures, during the process
some uphill moves are also permitted. This feature renders simulated annealing the ability
to find the global optima.
During the iterations the information for optimal value is stored and once the same
optimal values is reached within numerical convergence the program is terminated. In the
current program the process terminates when same optimal value is reached twice. If con-
vergence criterion is not meet then the program terminates when the maximum number of
evaluations exceeds the specified value.
As with all general purpose optimization methods, some control parameters in SIMANN
had to be ”tuned” to suit the objective function, and to ensure that the optimizer performs
efficiently. A bad choice for these parameters can make the algorithm extremely inefficient
and may even result in failure to arrive at the global optimum. The eight such parameters
in SIMANN are Tinit , vi , NS , ci , NT , rt , Neps and eps. Values of 1.0, 20, 30, 0.85,2.0, 2
are assigned for vi , NS , ci , NT , rt , Neps respectively, as recommended by Corana et al [15],
and based on previous experience. The most suitable values of the remaining 3 parameters
viz. Tinit , are determined by numerical experimentation. For instance, Tinit is determined
as 15000 by trial runs, and eps is assigned a value 1.0E-5 for CDV and 1.0 for ’S’. A higher
penalty is forced on the objective function for the undesired shapes resulted through the
shape generation algorithm. Being the problem is of multi disciplinary optimization type,
various objective functions defined are as follows
Coefficient of drag based on volumetric area (CDV )
Total surface area (S)
Maximum hoop stress in the envelope material (σ )
Combination of CD S, and σ with the specified weight-age values.
CHAPTER 5. SIMULATED ANNEALING OPTIMIZATION METHOD 28
Khoury and Gillete [3] have given the formula for CDV as a function of thickness ratio d l
for the streamline bodies of revolution and is depicted in Eqn.5.2
0 172 l d
1 3 0 252 d l
12 1 032 d l
27
CDV 1 6 (5.2)
Re
Variation of CDV with d l is shown Fig.5.2. It can be seen that CDV value initially
decreases, and then monotonically increases with thickness ratio. The minimum CDV is
observed near thickness ratio 0.23.
The total drag on airship also depends upon cross-flow. Jones et.al.[12] has given the
correlation for drag in terms of axial and cross-flow drag coefficients. The effect of cross-
flow is however neglected in the present work.
Total surface area has been considered as one of the objective functions as it gives the direct
indication of the envelope weight. Total surface area for the envelope shape is estimated
analytically as the equations governing the shape are known.
CHAPTER 5. SIMULATED ANNEALING OPTIMIZATION METHOD 29
The maximum hoop stress per unit thickness is given by Eq. 5.3 and is depends mainly upon
differential aerodynamic and internal pressure with respect to the ambient pressure and
variation in hydrostatic pressure as depicted in Eq.5.4. Pant and gawale [13] have provided
a procedure for estimation of total internal pressure. The aerodynamic pressure arises due
to the motion of the body. The internal pressure inside the envelope is generally taken 15%
more than the dynamic pressure to ensure that the shape of the envelope is maintained dur-
ing forward motion. The value for coefficient of pressure C p required for ∆P aerodynamic
calculation is generally taken as 0.33 as suggested by Khoury and Gillett[3]. The differen-
tial aerodynamic and internal pressure are calculated using Eq.5.5 and Eq.5.6 respectively.
∆Pd
σmax (5.3)
2
ρa ρhe gd
∆p ∆P aerodynamic
∆P internal
(5.4)
2
Where
∆P aerodynamic 1 2ρaV 2C p (5.5)
x2 x1 (5.7)
x3 x2 (5.8)
Hence the design vector was modified by replacing x2 and x3 by δ1 and δ2 such that
x2 x1
δ1 (5.9)
x3 x2
δ2 (5.10)
Thus the modified design vector becomes XD = (x1 , y2 , δ1 , δ2 ,y3 , l). Since many con-
straints are being handled implicitly by shape generation program, the ranges for design
variables are identified such that resulted shape lies in the feasible region. The constraint
of constant length is taken care by assigning a small range to length. The design variables
are normalized between (0 - 1) using the respective ranges . The ranges assigned to various
design variables are specified in the Table 5.1
The results obtained using the optimizer are discussed in the next chapter. The opti-
mized shapes for various objective functions, along with the comparison with reference
GNVR shape are also presented.
CHAPTER 5. SIMULATED ANNEALING OPTIMIZATION METHOD 31
This chapter discuss the results obtained using simulated annealing code SIMANN used
in conjunction with available and developed models particularly for CDV . The basis and
need for development the model for CDV has been emphasized. The results obtained with
existing and developed models are presented and discussed in details.
Table 6.1: Comparison of Optimized Shape for CDV and GNVR Shape using Low Fidelity
32
CHAPTER 6. RESULTS AND DISCUSSION 33
Figure 6.1: Comparison showing Optimized Shape for CDV and GNVR shape
Table 6.2: Comparison between the Results Predicted by Low Fidelity and Fluent code
0.016
0.014
0.012
CDV
0.01
0.008
0.006
0.004
570 575 580 585 590 595 600
S
Similar study is made to see the effect of position of maximum diameter on volumetric
coefficient of drag CDV . Figure 6.3 shows the variation of CDV with respect to (Xymax l ).
It is seen from the graph that as (Xymax l ) value increases CDV decreases as a consequence
of reduced pressure drag. 80% variation in (Xymax l) causes CDV to vary by 42%. So the
sensitivity of CDV with respect to surface area and (Xymax l ) is nearly same.
Though for the same thickness ratio surface area and (Xymax l ) seem to affect CDV value
equally, it can be argued that with the increase in maximum diameter for the same length,
will induce more pressure drag as the projected area increases in proportion to d 2 . Moreover
the absolute variation in CDV for the shape configuration with constant (d l) considered is
CHAPTER 6. RESULTS AND DISCUSSION 35
5.5
"Xymax.plot"
4.5
CDV
4
3.5
2.5
10 20 30 40 50 60 70
Xymax / l (%)
very high (80%) for the variation in the parameter (Xymax l). Since the original expression
for CDV is not very much able to take the effect of pressure drag, current low fidelity model
for CDV is updated using CDV variation with (Xymax l). To retain the feature of original
equation, the difference between CDV obtained using Fluent code and that using original
low fidelity analysis is considered for curve fitting. A third degree polynomial is fitted by
the least square method. The new expression arrived for CDV is presented in the form of
Eqn.6.1 where CDV is calculated using Eqn.5.2
Where X = (Xymax l)
Figure 6.4: Optimized Shape with Improved Fidelity Model for CDV Vs GNVR Shape
Table 6.4: Comparison Showing the Results Predicted by Intermediate and High Fidelity
Models
Figure 6.5: Optimized Shapes for CDV with Fixed and Variable Length Configurations
CHAPTER 6. RESULTS AND DISCUSSION 38
α1 =
d
l Xymax
These three parameters and Xymax are combined together to arrive at a single geometric
parameter ’α’ given as
β
α α1 α2 α3 Xymax (6.2)
The slope (dy dx) for the concerned shape configuration at x 0 and x l are thereot-
ically infinite and is taken care numerically by omitting the point on leading and trailing
edge in the calculations of the parameters α2 and α3 . The actual computation of the slopes
starts at a distance of the order of 0.05 away from both the edges. N indicates the number
of grid points considered across the initial 10% length for α1 estimation and trailing 90%
length for α2 estimation. The parameter β is calculated as a function of thickness ratio.
It becomes clear from the observed variation of β with d l shown in Fig. 6.6, that the
curve follows the linear nature from d l - 0.16 onwards and is essentially invariant below
this threshold. The second or higher degree fit for β to account for all the points would
result in appreciably higher error particularly for the shapes with lower thickness ratios in
this case. In the present approach β has been estimated conditionally as
214 d l 31 26 β 3
β (6.3)
3 Otherwise
CHAPTER 6. RESULTS AND DISCUSSION 39
14
12
10
8
β
6
0
0.14 0.15 0.16 0.17 0.18 0.19 0.2 0.21 0.22
(d / l)
The conditional estimation along with the linear fit depicted in Eqn.6.3 improves the
accuracy by inducing the lower order errors in the curve fit model.
The coefficient of drag based on surface area CD has been used in the model rather than
CDV as a consequence of smoother variation of CD with α. Figure 6.7 shows the variation
of CD with α for various thickness ratios. Similar graphs have been resulted for rest of
the thickness ratios. The linear fit for these graphs has resulted in R2 value between (0.95-
0.997) showing a maximum error of 4.5% in the design space considered.
CD A α
B E 04 (6.4)
The coefficients A and B are again obtained as the functions of thickness ratio. Figure
6.8 shows the variation of A and B with thickness ratio respectively. The similar approach
as that for variable β has been adopted for conditional estimation of coefficient ’A’ and is
given by Eq.6.5. The coefficient ’B’ is estimated using linear equation given in the form of
Eq.6.6 as also obvious from Fig.6.8.
12 02 d l
2 068 A 0
A (6.5)
0 Otherwise
The required objective function CDV can readily be estimated from CD using Eq.2.2
CHAPTER 6. RESULTS AND DISCUSSION 40
0.0047 0.0062
0.006
0.0046
0.0058
0.0045 0.0056
0.0054
0.0044
CDα
CDα
0.0052
0.0043
0.005
0.0042 0.0048
0.0046
0.0041
0.0044
0.004 0.0042
5 10 15 20 25 30 35 0 20 40 60 80 100 120 140
α α
(a) (b)
0.0066 0.0058
0.0064
0.0057
0.0062
0.006 0.0056
0.0058
0.0056 0.0055
CDα
CDα
0.0054
0.0054
0.0052
0.005 0.0053
0.0048
0.0052
0.0046
0.0044 0.0051
0 100 200 300 400 500 600 0 2000 4000 6000 8000 10000 12000
α α
(c) (d)
0.4 60
0.35
55
0.3
50
0.25
Coeff A
Coeff B
0.2 45
0.15
40
0.1
35
0.05
0 30
0.14 0.15 0.16 0.17 0.18 0.19 0.2 0.21 0.22 0.14 0.15 0.16 0.17 0.18 0.19 0.2 0.21 0.22
(d / l) (d / l)
(a) Variation of Coeff. ’A’ with d l (b) Variation of Coeff. ’B’ with d l
Table 6.5: Comparison Showing the Results Predicted by Improved Fidelity and Fluent
Code
Figure 6.9: Comparison Showing Optimized Shapes for Fixed and Variable Length Con-
figurations Vs GNVR Shape
small and hence can be attributed to the numerical error arising out of computations and
approximation through curve fitting. This fact can be made use of to infer that the resulted
shape is the optimized shape for the fixed length configuration. This can be further backed
by observing Fig. 4.1 that the exact GNVR shape is not possible with considered shape
configuration.
Table 6.7: Comparison Showing the Results Predicted by Improved and Fluent code
CDV value predicted using improved fidelity for variable length configuration differs by
3.5 % to that resulted using Fluent code as seen in Table 6.7. The shape shows significant
improvement in the drag of around 16% as predicted by the improved fidelity where the
actual improvement was observed to be around 12% using Fluent code. This proves at
first hand the superiority of the currently developed model compared to earlier two models
discussed in this report.
CHAPTER 6. RESULTS AND DISCUSSION 44
Figure 6.10: Optimized Shapes for ’S’ with Fixed and Variable Length Configurations Vs
GNVR Shape
Table 6.8: Comparison for Min.’S’ Shapes for Fixed and Variable Length Configurations
CDV S σmax
Fcomp w1 !
w2 #
w3 ! (6.7)
CDV GNV R " S GNV R " σmax GNV R "
Where w1 ,w2 and w3 are weight functions. For the present case all the weight functions
have been assigned a value unity. Figure 6.11 shows the optimal shape and GNVR shape.
The numerical values of three objective functions are presented and compared to GNVR
shape in Table 6.9.
Figure 6.11: Comparison Showing Optimized Shape with Composite Obj. Function Vs
GNVR Shape
The actual CDV value predicted by Fluent code is found to be closer to that predicted
using improved fidelity model proving the reliability of the result presented in Table 6.9 to
an appreciable extent. The resulted shape can be considered as the optimized shape for the
constraint of fixed volume and manufacturability avoiding convex shapes directly in the
CHAPTER 6. RESULTS AND DISCUSSION 46
shape generation algorithm. The value for composite objective function was found to be
slightly lower than that of GNVR shape.
Chapter 7
An airship is a lighter than air vehicle which obtains its lift due to buoyancy. The airship
has its own areas of research and applications. Since airship requires less number of param-
eters for design, it is advisable to design the airship using the concept of multi-disciplinary
optimization. The participating disciplines would be mainly aerodynamics, structures, and
manufacturing.
The three disciplines influences the airship design with their own constraints and re-
quirements. The aerodynamics of airship has been studied by many authors and based on
experimental observations empirical relation for coefficient of drag CD has been proposed.
It is shown that CD for axisymmetric streamline bodies is a function of thickness ratio (d /
l). As reported in the literature ,for the same maximum diameter of envelope an increase in
length improves the aerodynamic performance of airship but impairs the structural aspects.
This illustrates one of the possible interaction among various disciplines which has to be
handled during the optimization process. Most of the constraints can largely be satisfied by
aerodynamic shape of an airship envelope.
The specific problem of airship envelope shape optimization aims at arriving at the
optimum envelope shape using the concept of multi-disciplinary design optimization. The
constraints of fixed volume (1000m3 ) and fixed length (26.8m) were posed and the objective
function CDV is to be minimized. Among all objective function the calculation of flow
parameters particularly drag needs the primary attention as the calculatin of drag requires
the use of flow analyzer. The strategy to solve the problem would be to use computational
fluid dynamics tool in conjunction with the optimizer. It was decided to select Fluent which
is commercial state of the art CFD software as an flow analysis tool. A reference GNVR
shape has been used to validate Fluent for the present application. The same shape has
47
CHAPTER 7. SUMMARY AND CONCLUSION 48
further been used for the comparison with optimized shapes resulted through the optimizer.
An axisymmetric structured grid was built around the upper half of the GNVR body using
commercial package Gambit 6.0. The coefficient of pressure C p distribution over the airship
envelope surface was found to be closely matching with reference values for GNVR shape
calculated using source panel method given by NAL Banglore . The values closely matches
except near the trailing edge because of the flow seperation. Thus it validates Fluent for the
airship application.The influence of computational domain size has also been studied.
As a part of primary strategy to attempt the problem a generalized geometry was defined
for the constraints of fixed volume and length. The geometry represents the combination of
two elliptical shapes and one parabola and is being parametrized using two design variables.
Since the size of design vector is small, it became economical to use the surrogate model
technique like Design and Analysis of Computer Experiments DACE which resulted in
response surface for CDV . The design space was descretized using orthogonal latin square
method and the actual values for CDV corresponding to the design points were calculated
using Fluent code. Slapping technique has been used which gives the convergence much
faster than the conventional CFD technique. The resulted optimized shape for minimum
CDV was found to be closely matching with GNVR shape.
To explore the possibility of better shapes in view of multi disciplinary aspect of opti-
mization, the generalized configuration for an envelope shape has been improved. The im-
proved general configuration comprises of two cubic splines in series with spherical portion
and parabola at the leading and trailing edge respectively. The geometry was defined using
the six dimensional design vector for the constraints of fixed volume and length and other
geometrical requirements. The shape generation program was developed using FORTRAN
language. Other than CDV objective functions considered were surface area and maximum
stress. To start with low fidelity model to estimate CDV ssuggested in the literature was used
to calculate CDV .
Simulated Annealing optimization code SIMANN was implemented with the shape
generation code with low fidelity and was fine tuned for the parameters like initial temper-
ature. The results were obtained for the indivisual objective functions. It was observed that
the values obtained for CDV using the low fidelity model differs largely and are in contrast
to that predicted by Fluent code. To make the results more reliable, an intermediate fidelity
was devised using the data obtained through CFD experimentation for the bodies having
same thickness ratio as that of GNVR shape. The intermediate fidelity model was observed
to give better results than low fidelity but still the results are in contradiction to that of high
CHAPTER 7. SUMMARY AND CONCLUSION 49
fidelity. The constraint of the fixed length was relaxed to see the possibility of other aero-
dynamically better shape. It was found that for variable length configuration the optimized
shape for minimum CDV matches closely with that for fixed length.
To capture the flow physics more accurately, sufficient number of experiments were
carried our using Fluent code for the bodies with varying thickness ratio for the specified
constraints. After the detail study of the shapes an improved model for CDV eestimation has
been arrived at in terms of the predominant geometric parameters which directly influences
the aerodynamic characteristics of an envelope shape. For the fixed length constraint the
results obtained are in slight contradiction to that predicted by Fluent code. But the dif-
ference is small to enough to account for the numerical errors arising due to curve fitting
and mathematical computations. The optimized shape resulted for the variable length con-
figuration shows a significant improvement of 12% in CDV over GNVR shape. The same
exercise was carried out for surface area as objective function where it was observed that
shape tries to acquire the minimum length for minimum surface area.
To address the multi-disciplinary nature of the problem a composite objective function
consisting of CDV S, along-with the respective weight-ages has been devised. The optimized
shape obtained for the minimum composite objective function shows slight improvement
over the GNVR shape.
Conclusion :
$ Coefficient of pressure C p distribution around a reference GNVR shape obtained us-
ing Fluent code found to be matching to that arrived using source panel method except near
the trailing edge due to flow seperation.
$ The change in the computational size from 5l to 10l radius doesnot change the solu-
tion.
$ The slapping technique for CFD solution used found to give the convergence much
faster saving up to 40% of computational time per geometry.
$ It was observed that the resulted shape using DACE closely matches to the GNVR
shape for the shape configuration with primary shapes.
$ CDV as a function of d l suggested in the literature found to give the results largely
varying from predicted by Fluent code.
$ CDV found to be a function of Xymax and surface area other than thickness ratio.
$ Intermediate fidelity model developed using the parameter Xymax found to give better
CHAPTER 7. SUMMARY AND CONCLUSION 50
results than low fidelity model. But the results obtained are in contradiction to that predicted
by Fluent code.
$ Dominant parameters affecting the aerodynamic behavior of an envelope were found
to be certain geometric parameters which are mainly the function of the slopes at specific
locations along the envelope shape.
$ The dominant geometric parameters have been inducted in the CDV model. The vali-
dation of the improved fidelity shows the absolute error well within 5.0%.
$ The resulted shape for variable length configuration for minimum CDV observed to
offer lesser drag than GNVR shape showing a remarkable improvement of 11.8% .
$ The resulted value of CDV for the cases of fixed length and variabl e length configu-
ration found to be closely matching to that given by Fluent code.
$ The variable length configuration for the minimum envelope surface area case achieves
the minimum length allowed.
$ The objective function value for the minimum composite objective function was found
to be lesser than GNVR shape.
References
[1] Howe D.,“ The feasibility of the LArge Freight Airship”, Crankfield Report Aero ,
No.5m, March 1971.
[3] Gillett J.D., “Airship Technology”, Cambridge Aerospace Series, 1999, pp. 181.
[4] Hess T.E., “Structures Technology For Lighter Than Air Vehicles ”, Air Vehicle Tech-
nology Department, NADC, Pennsylvania, Technical Memeor
[5] Sophia Lefantzi and Doyle D. Knight, “ Automated Design Optimisation of a Three-
Dimensional S-Shaped Subsonic Diffuser” , Journal of Propulsion and Power, Vol.18,
No.4, July-August 2002, pp. 913-921. andum No. VT-TM-1391.
[6] Arora Jasbir, “Introduction to Optimum Design”, Mc Graw Hill Publication, 1999,
pp.309-313.
[8] Rachid Younsi, Ismail El-Batanony, Jeur-Bernard Tritsch, Hassan Naji and Bernard
Landjeri.,”Dynamic study of a Wind Turbine Blade with Horizontal Axis”, European
Journal of Mechanics - A/Solids,Volume 20, March 2001, pp. 241-252
[9] Narayana C.L., and Srilatha K.R., “Analysis of Aerostata Configuraton By Panel
Method,” PD CF 0010, July 2000, NAL, Bangalore.
51
REFERENCES 52
[11] Wilcox D.C., “Turbulenece Modelling for CFD”,DCW Industries , 1997, pp.79
[13] Pant, R., Gawale, A., ”Design, Fabrication And Flight Testing Of Remotely Con-
trolled Airship”,National Conference on LTA Technologies, Aerial Delivery R&D Es-
tablishment, Agra,India, October 2002, pp.4-6
[14] Goffe, W. L., G. D. Ferrier, J. Rogers. (1994). Global Optimization of Statistical Func-
tions with Simulated Annealing, Journal of Econometrics, 60, 65-100.
[15] Corana, A., Marchesi, M., Martini, C. M. and Ridella, S. (1987) Minimizing mul-
timodal functions of continuous variables with the Simulated Annealing algorithm.
ACM Transactions on Mathematical Software, 13, 262-280.
[16] Metropolis, N., Rosenbluth, A., Rosenbluth, M., Teller, A. and Teller, E. (1953) Equa-
tion of state calculations by fast computing machines. Journal of Chemical Physics,
21, pp. 1087-1090.