Download as pdf or txt
Download as pdf or txt
You are on page 1of 144

Structural sizing of post-buckled thermally stressed

stiffened panels

By

Walid Arsalane

Approved by:

Shanti Bhushan (Major Professor)


Eric Collins
Matthew Priddy
Adrian Sescu (Graduate Coordinator)
Jason M. Keith (Dean, Bagley College of Engineering)

A Dissertation
Submitted to the Faculty of
Mississippi State University
in Partial Fulfillment of the Requirements
for the Degree of Degree of Doctor of Philosophy
in Computational Science and Engineering
in the Bagely College of Engineering

Mississippi State, Mississippi

May 2022
Copyright by

Walid Arsalane

2022
Name: Walid Arsalane

Date of Degree: May 13, 2022

Institution: Mississippi State University

Major Field: Computational Science and Engineering

Major Professor: Shanti Bhushan

Director of Dissertation: Manav Bhatia

Title of Study: Structural sizing of post-buckled thermally stressed stiffened panels

Pages of Study: 127

Candidate for Degree of Degree of Doctor of Philosophy

Design of thermoelastic structures can be highly counterintuitive due to design-dependent

loading and impact of geometric nonlinearity on the structural response. Thermal loading generates

in-plane stresses in a restrained panel, but the presence of geometric nonlinearity creates an

extension-bending coupling that results in considerable transverse displacement and variation

in stiffness characteristics, and these affects are enhanced in post-bucking regimes. Herein a

methodology for structural sizing of thermally stressed post-buckled stiffened panels is proposed

and applied for optimization of the blade and hat stiffeners using a gradient-based optimizer. The

stiffened panels are subjected to uniform thermal loading and optimized for minimum mass while

satisfying stress and stability constraints. The stress constraints are used to avoid yielding of

the structure, whereas the stability constraints are used to ensure static stability. Corrugation of

the hat stiffeners is also studied through variation of its magnitude and position. A continuation

solver has been validated to tackle the highly nonlinear nature of the thermoelastic problem, and
formulations for the stability constraints have been derived and imposed to satisfy the static stability

of the structure. The study confirms that geometric nonlinearity is an important aspect of sizing

optimization and is needed for an accurate modeling of the structural behavior. The results also

show that modeling of geometric nonlinearity adds extra complexity to the thermoelastic problem

and requires a path-tracking solver. Finally, this work supports that corrugation enhances the

stability features of the panel but requires a blending function to reduce stresses at the panel

boundaries.

Key words: FEA, Thermoelasticity, Design optimization, Stiffened panels, Blade stiffeners,
Continuation solvers, Nonlinear elasticity
DEDICATION

To Manav

ii
ACKNOWLEDGEMENTS

First of all, I wish to express my profound appreciation and respect to my mentor, former

adviser and friend Dr. Manav Bhatia for his support and guidance throughout this research. I am

very grateful to him for the countless opportunities and exciting challenges that have significantly

improved my knowledge.

I am thankful to Dr. Shanti Bhushan who is serving as my co-adviser, for stepping in and

providing me with guidance and support in this difficult time. I would like to thank the committee

members Dr. Adrian Sescu, Dr. Matthew W. Priddy and Dr. Eric M. Collins for their time and

valuable inputs. Lastly, I would like to thank my parents and my girlfriend for their endless support

during my education at Mississippi State University.

iii
TABLE OF CONTENTS

DEDICATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii

ACKNOWLEDGEMENTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii

LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii

LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii

CHAPTER

I. INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1 Challenges in the design of aerospace vehicles . . . . . . . . . . . . . . 1


1.2 Stiffened panels in aerospace structures . . . . . . . . . . . . . . . . . . 5
1.3 Post-Buckling in Structural design . . . . . . . . . . . . . . . . . . . . . 6
1.4 Solvers for Nonlinear System of Equations . . . . . . . . . . . . . . . . . 8
1.5 Static Stability in Design Optimization . . . . . . . . . . . . . . . . . . . 10
1.6 Objective and approach . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.7 Organization of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . 16

II. PROBLEM FORMULATION . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

2.1 Stress-Strain and Strain-displacement relationships for plates . . . . . . . 18


2.2 Stress-Strain and Strain-displacement relationships for beams . . . . . . . 20
2.3 Derivation of the variational statement . . . . . . . . . . . . . . . . . . . 21
2.3.1 Static analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.3.2 Modal analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

III. OPTIMIZATION PROBLEM DESCRIPTION . . . . . . . . . . . . . . . . . . 26

3.1 Optimization statement . . . . . . . . . . . . . . . . . . . . . . . . . . . 26


3.2 Stress Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.3 Stability Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.3.1 Approach 1: Constraints at final temperature with backtracking . . 28
3.3.2 Approach 2: Aggregation . . . . . . . . . . . . . . . . . . . . . . 31

iv
3.3.2.1 Eigenvalues aggregation: . . . . . . . . . . . . . . . . . . 32
3.3.2.2 Constraints aggregation: . . . . . . . . . . . . . . . . . . . 33
3.3.3 Approach 3: constraint on temperature . . . . . . . . . . . . . . . 33

IV. IMPLEMENTATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

4.1 Continuation Solver . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37


4.1.1 Arc-length Predictor . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.1.2 Pseudo Arc-length . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.2 Globally Convergent Method of Moving Asymptotes (GCMMA) . . . . . 45
4.3 Verification and Validation . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.3.1 Verification of the sensitivity equations . . . . . . . . . . . . . . . 49

V. RESULTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

5.1 Problem setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53


5.2 Blade-stiffened panel . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
5.2.1 Structural sizing with increasing thermal loads . . . . . . . . . . . 60
5.2.2 Impact of geometric nonlinearities . . . . . . . . . . . . . . . . . 67
5.2.2.1 Comparison 1: comparing the feasibility of structures 1 and 2 68
5.2.2.2 Comparison 2: comparing the frequency behavior and the
panel deformation . . . . . . . . . . . . . . . . . . . . . . 68
5.3 Hat-stiffened panel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.3.1 Hat-stiffened panel without corrugation . . . . . . . . . . . . . . 74
5.3.2 Hat-stiffened panel with corrugation . . . . . . . . . . . . . . . . 80
5.3.2.1 Corrugated panel without a blending function . . . . . . . . 82
5.3.2.2 Impact of the blending function . . . . . . . . . . . . . . . 89

VI. CONCLUSION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96

6.0.1 Optimization of the blade-stiffened panel . . . . . . . . . . . . . . 97


6.0.2 Optimization of the hat-stiffened panel . . . . . . . . . . . . . . . 97
6.1 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
6.2 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
6.2.1 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
6.2.2 Disciplinary Extensions . . . . . . . . . . . . . . . . . . . . . . . 99
6.2.3 Computational Issues . . . . . . . . . . . . . . . . . . . . . . . . 99

REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

APPENDIX

A. SENSITIVITY ANALYSIS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

v
A.0.1 Direct sensitivity . . . . . . . . . . . . . . . . . . . . . . . . . . 111
A.0.2 Adjoint sensitivity . . . . . . . . . . . . . . . . . . . . . . . . . . 111
A.0.3 Direct vs Adjoint Sensitivity . . . . . . . . . . . . . . . . . . . . 112

B. THERMOELASTICITY IN BEAMS . . . . . . . . . . . . . . . . . . . . . . . . 113

vi
LIST OF TABLES

1.1 Summary of the test cases and simulation setups. . . . . . . . . . . . . . . . . . . 16

4.1 Pseudo Arc-length continuation solver parameters. . . . . . . . . . . . . . . . . . 42

4.2 GCMMA optimization parameters. . . . . . . . . . . . . . . . . . . . . . . . . . 48

4.3 Verification of the analytical sensitivities against finite difference approach. . . . . 52

5.1 Typical physical properties of Inconel-718 at room and elevated temperatures. . . . 58

5.2 Initial values of the design variables for the optimization problem. . . . . . . . . . 59

5.3 Optimization parameters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

5.4 Structural mass of the feasible optimized designs for ∆T = {120, 130, 140, 150, 160}◦ C. 61

5.5 Initial geometry of the hat-stiffened panel. . . . . . . . . . . . . . . . . . . . . . . 74

5.6 GCMMA optimization parameters. . . . . . . . . . . . . . . . . . . . . . . . . . 74

5.7 Structure mass of the feasible optimized design for ∆T = {120, 140, 160, 180, 200}◦ C. 76

5.8 Structure mass of the feasible optimized design for ∆T = {120180, 200}◦ C. . . . . 82

vii
LIST OF FIGURES

1.1 Photograph of Lockheed SR-71 Blackbird [27]. . . . . . . . . . . . . . . . . . . . 2

1.2 Photograph of the Aérospatiale/BAC Concorde supersonic commercial transport[22]. 3

1.3 Photograph of the North American X-15 experimental rocket-powered aircraft[50]. 4

1.4 Photograph of the Northrop B-2 Spirit stealth bomber [78, 106]. . . . . . . . . . . 4

1.5 The first natural frequency of the optimized stiffened panels for operating thermal

loads of ∆T = {120, 140, 160}◦ C. . . . . . . . . . . . . . . . . . . . . . . . . . . 13

3.1 A flowchart describing the backtracking algorithm within the optimization process. 30

3.2 The first natural frequency of the optimized stiffened panels for operating thermal

loads of ∆T = {120, 140, 160}◦ C. . . . . . . . . . . . . . . . . . . . . . . . . . . 35

4.1 Illustration of path continuation approach that iterates through points A → B →

C → D, such that D is on the curve R(X 0,T) = 0 at a distance ∆s from A. . . . . . 38

5.1 Percentage of the out-of-plane deformation at ∆T = 0◦ C with respect to the defor-

mation at ∆T = 160◦ C. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

5.2 3D representation of the blade-stiffened panel. . . . . . . . . . . . . . . . . . . . 57

5.3 y − z cross-section of the blade-stiffened panel. . . . . . . . . . . . . . . . . . . . 57

5.4 y − z cross-section of the blade-stiffened panel. . . . . . . . . . . . . . . . . . . . 60

viii
5.5 Convergence of the scaled mass and constraints in sizing optimization of a stiffened

panel for operating thermal loads of ∆T = {120, 140, 160, 180}◦ C. . . . . . . . . . 62

5.6 The first natural frequency of the optimized stiffened panels for operating thermal

loads of ∆T = {120, 140, 160}◦ C. . . . . . . . . . . . . . . . . . . . . . . . . . . 63

5.7 Stress distibution (Pa) of the optimized stiffened panels for operating thermal loads

of ∆T = {120, 140, 160}◦ C. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

5.8 Out-of-plane deformation (m) of the optimized stiffened panels for operating thermal

loads of ∆T = {120, 140, 160}◦ C. . . . . . . . . . . . . . . . . . . . . . . . . . . 65

5.9 Design variables distribution of the optimized stiffened panels for operating thermal

loads of ∆T = {120, 140, 160}◦ C. . . . . . . . . . . . . . . . . . . . . . . . . . . 66

5.10 Convergence of the scaled mass and maximum stress constraint in sizing optimiza-

tion of a stiffened panel for a thermal load of ∆T = 160◦ C using linear and nonlinear

strain. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

5.11 The fist eigenvalue λ1 of the optimized stiffened panels (structures 1 and 2). . . . . 69

5.12 Stress distibution (Pa) of structure 2 using a linear and nonlinear strain for a thermal

load of ∆T = 160◦ C. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

5.13 Out-of-plane deformation (m) of structure 2 using linear and nonlinear strain for a

thermal load of ∆T = 160◦ C. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

5.14 Schematic of a hat-stiffened panel with skin corrugation amplitude η. . . . . . . . 72

5.15 Geometry of the hat-stiffened panel with skin corrugation η. . . . . . . . . . . . . 72

5.16 Stations position in the stiffened panel. . . . . . . . . . . . . . . . . . . . . . . . . 73

ix
5.17 Convergence plot of the objective function and constraints subjected to thermal loads

of ∆T = {160, 180, 200, 220}◦ C for a hat-stiffened panel. . . . . . . . . . . . . . . 75

5.18 The first natural frequency of the optimized design of a hat-stiffened panel obtained

with thermal loads of ∆T = {160, 180, 200}◦ C. . . . . . . . . . . . . . . . . . . . 76

5.19 Out-of-plane deformation (m) of the final design of a hat-stiffened Inconel panel

obtained with thermal loads of ∆T = {160, 180, 200}◦ C. . . . . . . . . . . . . . . 77

5.20 Stress distibution (Pa) of the final design of a hat-stiffened Inconel panel obtained

with thermal loads of ∆T = {160, 180, 200}◦ C. . . . . . . . . . . . . . . . . . . . 78

5.21 Distribution of the stiffeners thickness for hat-stiffened Inconel panels subjected to

thermal loads of ∆T = {160, 180, 200}◦ C. . . . . . . . . . . . . . . . . . . . . . . 79

5.22 Illustration of hat-stiffened panels with and without the blending function. . . . . . 81

5.23 Convergence plot of the structural mass and stress and stability constraints subjected

to thermal loads of ∆T = {160, 200, 220}◦ C for a stiffened panel with corrugation

η = 1.5mm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

5.24 The first natural frequency of the final design of a corrugated stiffened panel obtained

with thermal loads of ∆T = {160, 200}◦ C. . . . . . . . . . . . . . . . . . . . . . . 84

5.25 Stress distibution (Pa) of the final design of stiffened panels obtained for operating

temperatures ∆T = {160, 200}◦ C with corrugations η = {1.5, −1.5}mm. . . . . . . 85

5.26 Out-of-plane deformation (m) of the final design of stiffened panels obtained for

operating temperatures ∆T = {160, 200}◦ C with corrugations η = {1.5, −1.5}mm. . 86

5.27 Design variables distribution for thermal loads of ∆T = {160, 200}◦ C and corruga-

tion η = {1.5, −1.5}mm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88

x
5.28 Blending function fbd in the range x/L ∈ [0, 0.1] and x/L ∈ [0.9, 1]. . . . . . . . . 91

5.29 Convergence plot of the objective and constraints for thermal loads ∆T = {200, 220, 250}◦ C

for a corrugated stiffened panel (η = 1.5mm) with x0 = 0 and x1 = 0.1. . . . . . . 92

5.30 The first natural frequency of the final design of corrugated stiffened panels (η =

{1.5, −1.5}mm) obtained with thermal loads of ∆T = {200, 220}◦ C using a blending

function with x0 = 0 and x1 = 0.1. . . . . . . . . . . . . . . . . . . . . . . . . . . 93

5.31 out-of-plane deformation (m) of the final design of corrugated stiffened panels

(η = {1.5, −1.5}mm) obtained with thermal loads of ∆T = {200, 220◦ }C using a

blending function with x0 = 0 and x1 = 0.1. . . . . . . . . . . . . . . . . . . . . . 94

5.32 Stress distibution (Pa) of the final design of corrugated stiffened panels (η =

{1.5, −1.5}mm) obtained with thermal loads of ∆T = {200, 220}◦ C using a blending

function with x0 = 0 and x1 = 0.1. . . . . . . . . . . . . . . . . . . . . . . . . . . 95

xi
Nomenclature

E = young’s modulus, (N/m2 )

ρ = density, (kg/m3 )

ν = Poisson’s ratio

G = shear modulus, (N/m2 )

hi = thickness of the panel at the i th station along x−axis, i = 1, 2, ..7, (m)

P = load per unit area (N/m2 )

u0 = in-plane displacement of the reference plane along x−axis, (m)

v0 = in-plane displacement of the reference plane along y−axis, (m)

w = the out-of-plane deformation of the reference plane, (m)

θx = structural rotation about the x−axis, (rad)

θy = structural rotation about the y−axis, (rad)

θz = structural rotation about the z−axis, (rad)

ωk = k th natural frequency of the structure, (rad/s)

Ni = the i th finite element shape function

Nelems = number of elements

xs = vector of variables used to describe the system

x s∆ = vector of small perturbations in the solution

α = coefficient of thermal expansion, (◦ C)

κ = shear correction factor (=5/6)

σcrit = maximum stress allowed in the stiffened panel, (GPa)

T = operating temperature, (◦ C)
xii
T0 = reference temperature, (◦ C)

T = temperature difference with-respect-to a reference value, (◦ C)

ε = structural strain vector in Voigt notation

ε0 = linear strain vector in Voigt notation

εvk = von Karman strain vector in Voigt notation

σ = structural stress vector in Voigt notation, (N/m2 )

σvm = von Mises stress, (N/m2 )

m = mass of the structure, (kg)

m0 = initial mass of the structure before the optimization, (kg)

R = vector of design variables

rj = the j th design variable in the vector R, (m)

ru = upper bound for the vector of design variables, (m)

rl = lower bound for the vector of design variables, (m)

t = time, (s)

X = vector of discrete coeffcients for the solution x

Ωp = domain analysis for the plate in the xy-plane

Ωb = domain analysis for the beam in the x-axis

xiii
CHAPTER I

INTRODUCTION

1.1 Challenges in the design of aerospace vehicles

The demand for performant, reliable, robust and low risk aerospace vehicles is ever-increasing.

The optimization of the design and manufacturing process of those systems for atmospheric flights,

supersonic and hypersonic regimes will lead to an important development and extension in the

current aerospace type of missions. When designing such structures for the purpose of supersonic

and hypersonic regimes many challenges are encountered and starts rising in multiple engineering

disciplines such as heat transfer [16, 18, 19, 29, 30, 31], stability and control [80], shock interactions

[35] and other disciplines [108, 110].

Example of heat and aircraft structure interaction manifests in the launch and re-entry of space

shuttles. During reentry, the vehicle flight speed reaches Mach 25 causing the windward (bottom)

surface of the shuttle to be exposed to a maximum heat flux load of about 0.5 MW/m2 , and the

leeward (upper) surface to about 0.002 MW/m2 [3, 56]. Other examples of critical interaction

between heat and structures have arisen in the development of sustained high-speed flight vehicles

such as the Aérospatiale/BAC Concorde supersonic commercial transport [22] (Fig. 1.2), the

Lockheed SR-71 supersonic reconnaissance aircraft [27] (Fig. 1.1), and the North American X-15

experimental rocket-powered aircraft [50] (Fig. 1.3).

1
Figure 1.1

Photograph of Lockheed SR-71 Blackbird [27].

During sustained high-speed flight the aircraft structure experiences significant aerodynamic

heating resulting in thermal expansion, material softening and high compressive stresses, which

can cause loss of static stability. These effects have often been mitigated by accommodation,

i.e. designing the airframe with margin that permits thermal expansion and relieves thermal

stresses [68, 100]. The SR-71 provides a famous example of accommodation, as the aircraft leaked

fuel while on the ground and in subsonic flight, only sealing in supersonic flight due to thermal

expansion of its structure [38]. Yet, in some cases, such as the B-2, boundary fixivity is required,

and accommodation is not an option. The designers of the Northrop B-2 Spirit stealth bomber had

to overcome the problem of heating in the exhaust-washed structure. The exhaust-washed structure

(EWS) is the structure embedded in the fuselage and exposed to the exhaust of the engines. The

2
Figure 1.2

Photograph of the Aérospatiale/BAC Concorde supersonic commercial transport[22].

positioning of the engines inside the fuselage enabled low observability of the aircraft [78, 106].

This heating resulted in an out-of-plane deformation in the interior of the structural panels which

led to a stress concentration at their boundaries. This challenging interactions of heat and structural

response was discussed in literature in [16, 18, 19, 29, 30, 31].

In the cases where accommodation is not an option, the structure must be designed to overcome

the associated challenges in other ways. This can be counterintuitive and difficult due to the design

dependent nature of the loading in thermoelastic applications. While increasing the thickness of the

panel leads to a stiffer structure, it causes a counterintuitive effect of higher thermoelastic loading

in the structure. The EWS is often modeled as stiffened panels [30, 62]. The stiffened panels and

their use in the aerospace field will be discussed in the following section.

3
Figure 1.3

Photograph of the North American X-15 experimental rocket-powered aircraft[50].

Figure 1.4

Photograph of the Northrop B-2 Spirit stealth bomber [78, 106].

4
1.2 Stiffened panels in aerospace structures

In order to decrease the operating cost and reach a broader range of aerospace types of missions,

the weight of the aerospace vehicle is one of the primary concerns. One way to alleviate this problem

is to use stiffened panels. Such structures are made of sheet materials often called panel skin and

supported with stiffeners which add stiffness and buckling resistance to the structure while being

importantly lightweight. In literature, there are multiple studies that investigate stiffened panels

through the type of stiffeners used or panel skin thickness [23, 41, 49, 81, 85, 104]. Those studies are

often conducted through parametric studies based on analytical models or finite element analysis.

A parametric study performed by Chen and Guedes Soares [23] showed that for a stiffened panel,

the panel skin thickness can increase the post-buckling compressive strength. Jain and Upadhyay

[49] studied the buckling behavior of laminated composite blade-, angle-, T-, and hat-stiffened

panels subjected to in-plane shear loading using finite element methods and developed guidelines

for better stiffener proportioning. Van Dung et al. [104] investigated stiffened cylindrical shells

under torsion and studied the impact of the number and dimension of stiffeners on buckling and

post-buckling. Guo et al. [41] conducted parametric and comparative studies for different plate

aspect ratios, plate thickness to length ratios, degrees of layer orthotropy, ply orientations, and

stiffener depth to plate thickness ratios. It was demonstrated in [41] that the interactive buckling

or the mode shape transition between laminated plate and stiffener plays an important role in

maintaining the buckling behavior of stiffened laminated plates. It was also reported in [41] that

a deeper stiffener leads to a higher buckling load only to a certain extent and concluded that the

lateral buckling of a deep stiffener must be monitored as a critical factor to the buckling behavior

of stiffened laminated plates. Pevzner et al. [81] investigated bending buckling, torsional buckling,

5
combined bending and torsion buckling, and local buckling of the stringers for stiffened panels

with blade type stiffeners, J-form stiffeners, and T-form stiffeners. Rahimi et al. [85] studied the

effect of the ribs profile on the buckling of shells under axial loading and reported that stiffening

the shells increased the buckling load from 10% to 36% and decreased the buckling load to weight

ratio by 52% compared to an unstiffened shell. As discussed, stiffened panels is a genuine way

of reducing the weight of the aerospace vehicles. Another way of reaching the same goal is to

design the structure to be used in the post-buckling regime where the properties of the structure

were shown to be better [62]. The discussion on post-buckling in structural design is presented in

the next section.

1.3 Post-Buckling in Structural design

Manufacturing aerospace vehicles with advanced alloys is a key aspect in the obtention of

lightweight vehicles. Significant savings in structural weight can also be obtained through post-

buckling design. There is undeniably an ongoing movement to shift the modeling and design

community perspective about structural instabilities as sources of catastrophic failure to opportu-

nities for functionality [86]. Since lightweight structures are susceptible to structural instabilities

and failure [47], the prevailing methodology in the design community is to avoid buckling or

at least make its effect less important. In contrast with bucking-resistant design, post-buckling

allows for buckling before the limit load, thereby allowing to take full advantage of the structural

properties in the post-buckling region and exploit, rather than avoid, elastic instabilities. Reis [86]

reviewed the research effort focusing on taking advantage of instabilities enabling the discovery

of new design possibilities, and provided a new angle on buckling and structural instabilities,

6
namely from "buckliphobia to buckliphilia". Such designs that would take advantage of the post-

buckling region can be in the form of optimized thin-walled structures that operate safely in the

post-buckling regime [9, 74, 109]; shape-morphing structures that snap between different configura-

tions [6, 7, 32, 33, 52, 60, 67, 99, 111], and other types of designs [14, 34, 70, 72, 76, 77, 84, 93, 94].

Design of panels for post-buckled response requires careful consideration of bifurcation character-

istics in order to prevent loss of static stability. Lee and Bhatia [62] demonstrated that corrugated

hat-stiffened panels exhibit primary and secondary bifurcation points and both were shown to be

sensitive to geometric features, such as the amplitude of corrugation. Particularly, it was shown

that primary bifurcation disappears in the presence of corrugations and secondary bifurcation

temperature increases by as much as 50%. Shallow arches are representative of thin bending

structures that exhibit rich bifurcation response. The snap-through response of shallow arches has

been extensively studied since the early work of von Karman and Tsien [51, 107] and Tsien [103].

Recent work of Pi and Bradford [82] and Harvey and Virgin [43] has identified the presence of

multiple unstable equilibrium branches for shallow pre-stressed arches. Stanciulescu et al. [95],

Moghaddasie and Stanciulescu [69] and Virgin et al. [105] found the bifurcation response of shal-

low arches to be highly sensitive to its thermal environment and identified that different parametric

combinations of temperature and arch height result in a structural response with 0 to multiple

bifurcation points. Cox et al. [26] noted that design of shallow arches based on linear analysis

is significantly different from the one based on nonlinear analysis, and that the post-buckled load

capacity can be greatly tailored through geometric tweaking. One of the challenges that rises

when investigating the post-buckling region of thermoelastic problems is the high nonlinearity of

7
the system of equations for the static equilibrium. The nonlinear solvers available in literate are

discussed in the next section.

1.4 Solvers for Nonlinear System of Equations

Geometric nonlinearity is related to the nonlinear behavior of deformable bodies, such as

beams, plates and shells, when the relationship between the extensional strains and shear strains,

on the one hand, and the displacement, on the other, is taken to be nonlinear, resulting in nonlinear

strain-displacement relations. As a consequence of this fact, the differential equations governing

this system will turn out to be nonlinear. This is true in spite of the fact that the relationship

between curvatures and displacements is assumed to be linear.

Nonlinearity critically influences the thermoelastic post-buckled response and stability of thin-

walled aerospace structures. It creates a coupling between in-plane and out-of-plane response,

and influences the stiffness and stability characteristics of the structure. Deaton and Grandhi [29]

demonstrated the significance of geometric nonlinearity in the response of thermally loaded struc-

tures with computational experiments comparing predictions from linear and nonlinear analysis

and assessing the dependency of these analyses on material properties, thickness of the structure,

and the curvature of the structure. Neiferd et al. [73] analyzed a hypersonic panel subjected to

both high temperatures and large temperature gradients and demonstrated that linear analyses of

the problem leads to inaccurate response prediction. Notwithstanding the challenges posed by

nonlinearity, recent experience suggests that it is more economical to design a thermally loaded

structure while allowing it to operate in the post-buckled region [15, 62, 73, 87], as opposed to

designing for linearized buckling. Cinoglu et al. [24] found that including geometric nonlinearity

8
and elastoplastic material response can expand the design space of a thermoelastic structure by a

factor of 2 to 4.

Computational analysis of nonlinear thermally stressed structures poses additional challenges.

Neiferd et al. [73] and Lee and Bhatia [62] documented lack of robustness in the application of

load-stepping for nonlinear thermoelastic analysis, even in the absence of turning-points. The static

solution tends to jump from one equilibrium branch to another during the load-stepping process

and these jumps were shown to be highly dependent on the value of temperature increments. This

occurs for simple structures (eg., beams or plates) and complex built-up structures (eg., stiffened

panels), and across different codes (Abaqus, MAST [1, 17] and other hand-written routines).

Temperature increments of increasingly small amplitude are required to ensure a smooth evolution

of the panel deformation, and this too is found to be problem dependent. Multiple simulations

reported by Lee and Bhatia [62] needed as many as 1000 load steps to prevent the panel from

jumping across equilibrium branches. It is emphasized that this is a computational artifact for

nonlinear thermoelastic analyses since none of the thermoelastic studies in [62, 73] exhibited limit

points. This issue can be overcome through the use of a continuation solver [20] and this work

uses the pseudo-arc-length continuation solver for all nonlinear analyses.

Most commercially available structural analysis software packages provide the ability to perform

a nonlinear thermoelastic analysis with supporting continuation solvers and stability analysis.

However, efficient gradient-based optimization of such nonlinear problems remains a research

topic, which requires accurate and efficient sensitivity analysis. Bhatia [16] and Bhatia and Livne

[18, 19] developed the formulation for sensitivity analysis, approximation concepts for repeated

analysis, and a design-oriented software capability for optimization of thermally loaded structures.

9
This development was merged into Multidisciplinary-design Adaptation and Sensitivity Toolkit

(MAST) [1, 17], which is an actively developed software within our research group. MAST is

an open-source computational framework written in object-oriented C++, and leverages state-of-

the-art open-source computational libraries [2, 4, 5, 11, 12, 44, 45, 55, 89] to provide a scalable

finite element analysis framework for design optimization of multi-physics simulations. MAST’s

capability to perform nonlinear thermoelastic analysis has been verified through benchmark analysis

with Abaqus [62, 73]. When design thermoelastic structures, the highly nonlinear system of

equation could be accuratly solved using a path-tracking solver as mentionned earlier. Another

aspect that needs to me given attention is the static stability of the equilibrium point. Optimization

problems with stability constrains have been around in literature for some time now. The topic of

static stability in design optimization will be discussed in the upcoming section.

1.5 Static Stability in Design Optimization

Research in structural stability analysis saw a resurgence in recent years. This new interest in

structural stability can be explained by a shift in the design philosophy regarding post-buckling

and its effects on instabilities. As discussed earlier, there is undeniably an ongoing movement to

shift the modeling and design community perspective about structural instabilities as sources of

catastrophic failure to opportunities for functionality. The newly structural designs can be modeled

as thin-walled structures optimized for the post-buckled region or designs with shape morphing

that takes advantage of multiple stability points between configurations and many other types of

designs. Therefore, structural designers are starting to utilize elastic instabilities for repeatable

well-behaved adaptations, instead of considering them a source for structural failure.

10
Structural design optimization with frequency constraints is remarkably convenient in order

to control the dynamic properties of a structure in multiple manners. In aeronautical design, the

torsional and bending properties of an aircraft wing dictate its aeroelastic features. The frequency

constraint is frequently enforced on the aircraft control surfaces and structural components to

prevent flutter [15]. For other space vehicles, designers often attempt to control frequencies of the

lower structural modes by bounding them by an upper and lower limit in order to avert coupling

with the control system. Similarly, structural response to dynamic excitation in low-frequency

vibration problems is usually a function of the structure’s natural frequencies and mode shapes.

For those types of problems, controlling the structure frequency can positively alter the properties

of the structure.

The stability constraints can be enforced through solving a vibration problem [42] about the

nonlinear equilibrium state and making sure all eigenvalues of the natural vibration eigenvalue

problem are above a specified lower limit. It was shown in [9, 53] that structural sizing with

stability constraints is affected by the type of modelization (linear or nonlinear) of the structure.

In the work by Khot [53], an optimality criterion method was shown for a weight minimization

problem under stability constraints. The optimal designs obtained were analyzed with a nonlinear

analysis where it was concluded that a nonlinear analysis should be considered when optimizing

structures subject to stability constraints. Bhatia and Beran [15] presented a methodology to

perform multidisciplinary design optimization with transonic flutter constraints which included the

nonlinear influence of thermal stresses.

Several papers exist on structural optimization with eigenvalue constraints [9, 39, 40, 61, 64,

75, 79, 92, 101, 102]. Grandhi et al. [40] presented a work that demonstrated the generalized

11
compound scaling algorithm in the design of plate structures with frequency constraints. Patnaik

and Maiti [79] designed stiffened cylindrical panels and waffle plates with constraints on static

instabilities, natural frequencies, and initially stressed vibration frequencies where they considered

the interaction between instability and the natural frequency of the structure. Torii and De Faria

[101] employed an approximation for the smallest magnitude eigenvalue in topology optimization

for the maximization of the first natural vibration frequency of plane stress structures. Other papers

on structural optimization with frequency constraints can be found in the survey paper by Grandhi

[39].

One of the problems in structural optimization with frequency constraints is mode switching

of the structure mode shapes. Structural designers might not be following the same mode shape,

whether it is a bending, torsion, or axial mode when the design variables change. This phenomenon

occurs during structural size modifications which cause the eigenmode associated with a specific

eigenvalue to change. The eigenvalue is not differentiable at the point where the eigenmode

switches [92] and causes convergence difficulties for the optimizer. This issue can be addressed by

using the bound formulation with modal assurance techniques [54, 65], or aggregation techniques

[61, 101].

Another problem that manifests is the repetitiveness of eigenvalues. The optimum designs

obtained might have repeated eigenvalues although the initial design did not have any. This

difficulty has been overcome by many authors through different algorithms. Seyranian et al. [92]

used an approach based on perturbation which has been applied in structural optimization. Ma

et al. [64] used a weighted sum of the eigenvalues. Bendsoe and Sigmund [13] suggested the

addition of constraints to force well-spaced eigenvalues. Torii and De Faria [101] used a smooth

12
p-norm approximation for the smallest magnitude eigenvalue where they assumed strictly positive

eigenvalues.

A typical behavior of the structure natural frequency when accounting for structural nonlinear-

ities is shown in Figure 5.6. It can be noticed that the natural frequency is nonmonotonic and its

behavior is highly nonlinear where the frequency decreases in the range of small temperatures then

increases until the final temperature is reached.

1,500

1,000
80
ω1 (rad/s)

60
500
40
20
0 0
30 40 50
0 50 100 150
temperature difference ∆T (◦ C)
∆T = 120◦ C ∆T = 140◦ C ∆T = 160◦ C

Figure 1.5
The first natural frequency of the optimized stiffened panels for operating thermal loads of
∆T = {120, 140, 160}◦ C.

1.6 Objective and approach

Review of the literature suggest that missions dealing with heat-structure interaction require

performant structures to enrich the flight envelope and extend the durability and performance of the

aerospace vehicles. In addition heat-structure interaction leads to important challenges that need to

be considered when designing reliable aerospace vehicles. Stiffened panels have been heavily used

in the engineering community and specially in the aerospace community where it reduces weight

without impacting the stiffness of the structure. In literature, the design optimization of stiffened

13
panels is mainly done in the context of pre-buckling where designers prevent the structure from

reaching a critical buckling load. The design optimization of stiffened panels in the post-buckling

region appears to be a gap in literature which we target through this research. By designing for

the post-buckling region, we can enhance the structure ability to withstand higher loads but on

the other hand, important challenges starts to rise. One of the challenge is the high nonlinearity

of the thermoelastic problem which requires advanced numerical methods to solve the system of

equations which was shown to suffer from numerical divergence and required load steps of the

order O(1000) when solved using the Newton-Raphson solver. Another challenge is the static

stability of the structure in the post-buckling region which requires careful consideration.

The objective of this work is to alleviate the challenges of heat-structure interaction through

the development of a structural sizing methodology for post-buckled thermally stressed stiffened

panels. In order to achieve this goal, the following approach is used

• We set the mathematical foundation for the sizing optimization defined as a mass minimiza-

tion problem subjected to stress and stability constraints which uses the deformation, stresses

and eigenvalues of the structure at each optimization step to seek an optimal light-weight

panel[9, 10].

• To alleviate the challenges rising from the nonlinearity of the problem, a path track-

ing/following solver is used for the thermoelastic system of equations.

• Different formulations for the stability constraints are derived and compared in order to

address the challenge of structural stability [8].

14
• The modal problem is solved about the nonlinear equilibrium state and analytic sensitivity

analysis is used.

• Temperature is modeled as a uniform load applied on the structure and for the given range of

temperatures used in this work, the material properties are assumed to be constant and are

set to their corresponding values at the final temperature.

• The structural sizing methodology is applied on blade and hat-stiffened panels where the

impact of each stiffener type is studied.

A summary of the test cases, analysis and sizing optimization simulations are shown in Ta-

ble. 1.1. The simulations shown are divided into two main parts, first parametric studies with

increased complexity and second sizing optimization of the stiffened panels which serves as our

application of the developed methodology. Linear and nonlinear strains are used in the parametric

studies and the simulations cover a range of thermal loads for structures of added complexity (plate,

blade-stiffened panel, hat-stiffened panel). The purpose of the parametric studies is to verify the

finite element code and develop an intuition about the impact of the stiffeners on the structural

and modal properties of the panels. The second part of table 1.1 contains the sizing optimization

studies of the thermally stresses stiffened panels. It also shows the range of thermal loads, initial

mechanical load and corrugation magnitude and position. The Finite element codes developed are

part of the Multidisciplinary-design Adaptation and Sensitivity Toolkit (MAST) and are available

for public use on GitHub.

15
Table 1.1
Summary of the test cases and simulation setups.
Structure Linear or Nonlinear Thermal Load (◦C) Mechanical load (Pa) Corrugation (mm) Corrugation position (%)
Linear 0 103 − 5 × 105 N/A N/A
Linear 50 − 200 300 N/A N/A
Plate
Nonlinear 0 103 − 5 × 105 N/A N/A
Nonlinear 50 − 200 300 N/A N/A
Linear 0 103 − 5 × 105 N/A N/A
Linear 50 − 200 300 N/A N/A
Blade-Stiffened panel
Nonlinear 0 103 − 5 × 105 N/A N/A
Nonlinear 50 − 200 300 N/A N/A
Nonlinear 50 − 200 300 0 0
Static and modal Analysis
Nonlinear 50 − 200 −300 0 0
Nonlinear 50 − 200 300 1.5 0
Nonlinear 50 − 200 300 1.5 0-10
Nonlinear 50 − 200 300 1.5 5-10
Hat-Stiffened panel
Nonlinear 50 − 200 300 1.5 10-20
Nonlinear 50 − 200 300 −1.5 0
Nonlinear 50 − 200 300 −1.5 0-10
Nonlinear 50 − 200 300 −1.5 5-15
Nonlinear 50 − 200 300 −1.5 10-20
Linear 100 − 200 300 N/A N/A
Blade-Stiffened panel
Nonlinear 100 − 200 300 N/A N/A
Nonlinear 100 − 250 300 0 0
Strctural Sizing Nonlinear 100 − 250 300 1.5 0
Hat-Stiffened panel
Nonlinear 100 − 250 300 −1.5 0
Nonlinear 100 − 250 300 1.5 0-10
Nonlinear 100 − 250 300 −1.5 0-10

1.7 Organization of the thesis

The remaining part of this dissertation is organized as follow. Chapter II presents the governing

equations and analysis formulation for the thermoelastic problem. In Chapter III, we define the

optimization problem and derive the sensitivity equations for the stress and stability constraints.

The code implementation is discussed in Chapter IV with primary focus on the continuation solver

and the optimizer. In Section 4.1 we define the continuation method, formulate its main equations

and provide a pseudo-code for the solver. In Section 4.2, we discuss the optimizer used in this work

and explain the main steps in the structural sizing process. The verification and validation steps

are shown in Section 4.3.1 The results are presented in Chapter V. The results section is divided

into two main subsection. The first section contains results for the blade-stiffened panel and the

second results for the hat-stiffened panel. Finally, Chapter VI, states the major contributions of this

research, and the directions for future work.

16
CHAPTER II

PROBLEM FORMULATION

In this chapter, we present the formulation for the static and modal analysis of the stiffned panels.

The panel skin and hat stiffeners are modeled as plates using Reissner-Mindlin theory which allows

for shear strain through the thickness. The blade stiffeners are modeled using Bernoulli beam

equations with offsets in the z−axis. We start the chapter by deriving the stress-strain and strain-

displacement relations for plates and beams. Those relations will be used in the derivation of the

displacement-based variational forms of the thermoelastic problem.

The vector of variables used to describe the stiffened panels (blade and hat) is given by

x = {u0, v0, w, θ x , θ y, θ z }T , where u0 and v0 are the in-plane displacement of the reference plane in

the x and y directions, respectively. w is the out-of-plane deformation and θ x , θ y and θ z are the

rotations about the x−,y− and z−axis, respectively.

17
2.1 Stress-Strain and Strain-displacement relationships for plates

For plates, the motion of the reference plane is defined by the in-plane displacement u0 (x, y, t)

and v0 (x, y, t), the out-of-plane deformation w(x, y, t) and the two in-plane rotations along the x−

and y−axis θ x (x, y, t) and θ y (x, y, t). The plate kinematic relations are expressed as

u = u(x, y, z, t) = u0 + zθ y, (2.1)

v = v(x, y, z, t) = v0 − zθ x , (2.2)

w = w(x, y, t), (2.3)

where z is the distance from the reference plane. The components of the strain ε are given

by ε(x) = {ε x x , ε yy, γ xy, γ xz γ yz }T . The strain ε can be defined through the strain-displacement

relations expressed as

ε(x) = ε0 (x) + εvk (x) (2.4)

where

ε0 (x) = Ds x (2.5)
 
∂ 0 0 0 z ∂∂x 
 ∂x
 
 
0 ∂ 0 −z ∂∂y 0 
 ∂y
 
Ds =  ∂ ∂ (2.6)
 
0 −z ∂∂x z ∂∂y 
 ∂y ∂x 
 

0 0 0 1 
 
 ∂x 
 

0 0 −1 0 
 
 ∂x 

18
and
  2
1 ∂w 

 
2 ∂x

 


 

2

   

1 ∂w 

 
2 ∂
 


 y 


 
εvk (x) = (2.7)

 

∂w ∂w

 ∂x ∂y 


 

0

 


 


 


 

0

 


 

 
Linearization of the strain is given by

ε ∆ (x, x ∆ ) = ε0∆ (x, x ∆ ) + εvk



(x, x ∆ ) (2.8)

where

ε0∆ (x, x ∆ ) = Ds x ∆ (2.9)


 
∂w ∂w ∆

 

∂x ∂x

 


 


 

∂w ∂w ∆

 

∂ ∂
 


 y y 


 
εvk (x, x ) = (2.10)
∆ ∆

 

∂w ∂w ∆ ∂w ∂w ∆
 ∂x ∂y + ∂y ∂x 

 

 
0

 


 


 


 

0

 


 

 
The stress components of σ(x) are expressed as σ(x) = {σx x , σyy, τx y, τxz τyz }T . The stress σ(x)

is defined through the stress-strain relations given by

σ(x) = C (ε(x) − α∆T : e) (2.11)

19
where e = {1, 1, 0, 0, 0}T and
 

 E
 1−ν2 1−ν 2
0 0 0 
 
 
 νE E
0 0 0 
 1−ν2 1−ν 2
 
C =  0 (2.12)
 
0 G 0 0 
 
 
 0 0 0 G 0 

 
 
 0 0 0 0 G
 
 

2.2 Stress-Strain and Strain-displacement relationships for beams

In the context of beams, displacement components of a point of coordinate x, y and z in the

small bending theory are defined by

u = u0 − (y + y0 )θ z + (z + z0 )θ y, (2.13)

v = v(x, t) = v0, (2.14)

w = w(x, t), (2.15)

where, y0 and z0 are the offsets from the reference plane of the beam. Bending is assumed in both

the xy and xz planes. Extension and bending of the beam is characterized by the axial deformation

of the neutral axis u0 (x, t), deformation in the y−direction v0 (x, t), deformation in the z−direction

w(x, t) and rotations about the y− and z-axis θ y and θ z . The strain vector ε(x) = {ε x x , γ xy, γ xz }T

is expressed through the strain-displacement relations given by

∂u0 ∂θ z ∂θ y 1 ∂v 2 1 ∂w 2
   
εx x = − (y + y0 ) + (z + z0 ) + + , (2.16)
∂x ∂x ∂x 2 ∂x 2 ∂x
∂v0
γ xy = − θ z, (2.17)
∂x
∂w
γ xz = + θy. (2.18)
∂x
20
Linearization of the strain vector is given by ε ∆ (x, x ∆ ) = {ε ∆x x , γ xy
∆ , γ }T which components are
xz

expressed as

∂u0∆ ∂θ ∆z ∂θ ∆y ∂v ∂v ∆ ∂w ∂w ∆
ε ∆x x = − (y + y0 ) + (z + z0 ) + + , (2.19)
∂x ∂x ∂x ∂x ∂x ∂x ∂x
∂v0∆
γ x∆y = − θ ∆z , (2.20)
∂x
∂w ∆
γ xz

= + θ ∆y . (2.21)
∂x

The stress vector σ(x) = {σx x , τxy, τxz }T is defined through the stress-strain relations given by

σ(x) = C (ε(x) − α∆T : e) , (2.22)

where e = {1, 0, 0}T and


 
E 0 0 
 
 
C =  0 G 0  . (2.23)
 

 
 
 0 0 G
 
 

2.3 Derivation of the variational statement

First, we start with the strain-displacement relation which is expressed as

ε(x) = ε0 (x) + εvk (x) (2.24)

εvk is the von Karman strain, a nonlinear strain that couples inplane and out-of-plane deformations

and is used to capture stretching of the structure. Linearization of the strain is given by

ε ∆ (x, x ∆ ) = ε0∆ (x, x ∆ ) + εvk



(x, x ∆ ) (2.25)

21
where x ∆ = {u0∆, v0∆, w ∆, θ ∆x , θ ∆y , θ ∆z }T is the vector of small perturbations in displacements. The

stress σ(x) is defined through the stress-strain relations given by

σ(x) = C (ε(x) − α∆T e) (2.26)

where e = {1, 1, 0, 0, 0}T for plates and e = {1, 0, 0}T for beams, with ∆T a temperature difference

with respect to a reference temperature T0 , i.e. ∆T = T − T0 . Linearization of the stress can be

written as

σ ∆ (x, x ∆ ) = Cε ∆ (x, x ∆ ) (2.27)

Based on the principal of virtual work,the displacement-based variational statement is expressed

as: Find x, such that for all δx

∂2 x
∫ ∫   ∫
δx ρ 2 + ε (x, δx) σ(x) dAdΩ =
T ∆ T
δw Pnz dΩ, (2.28)
Ω p ∪Ωb A ∂t Ωp

where, δx and ε ∆ (x, δx) are the virtual displacement and the virtual strain, respectively. P is the

load per unit area, nz is the normal unit vector to the mid-plane of the panel in the z direction.

Ω p is defined as a two dimensional domain in the xy−plane for the plate and Ωb is defined as

a one dimensional domain along the x−axis for beams. A is the cross-sectional area for the

plate and thickness for the beams. The finite element approach is used for spatial discretization

of the governing equations. Following the standard formulation, the solution is approximated

over the domain of each finite element using Lagrange shape functions and the corresponding

vector of discrete coefficients. For example, u0 = Ni (ξ, η)U0ei , i = 1, .., nshape , where ξ, η are
Í
i

the non-dimensional coordinates, Ni (ξ, η) is the i th shape functionand U0ei are the discrete nodal

coefficients for u0 . Note that (ξ, η) can be uniquely mapped to (x, y) and vice-a-versa. Variables

22
{u0, v0, w, θ x , θ y, θ z } can be described using their discrete coefficients {U0,V0, W, Θ x , Θ y, Θz } over the

analysis domain. The discrete vector for x can then be expressed as X = {U0,V0, W, Θ x , Θ y, Θz }T .

2.3.1 Static analysis

The transient variational statement in Eq. (2.29) is re-expressed here for convenience

∂2 x
∫ ∫   ∫
δx ρ 2 + ε (x, δx) σ(x) dAdΩ =
T ∆ T
δw Pnz dΩ, (2.29)
Ω p ∪Ωb A ∂t Ωp

∂2 x
The static form of the variational statement can be obtain by setting ∂t 2
= 0. The weak form or

variational statement can then be expressed as: find x 0 , such that for all δx

∫ ∫   ∫
0 0
ε (x , δx) σ(x ) dAdΩ =
∆ T
δw Pnz dΩ (2.30)
Ω p ∪Ωb A Ωp

The variational statement can be written in the matrix-vector form

R = FI (X 0 ) − FE (X 0 ) = 0, (2.31)

where R is the residual of the system and X 0 the vector of discrete coefficients used to discretize

x 0 . For the sake of simplicity and convenience, the discrete vector of coefficient previously defined

as X 0 will be referred to to as X where we drop the superscript (.)0 since this work deals with static

and modal analysis only. FI (X 0 ) and FE (X 0 ) relate to the variational form in Eq. (2.30) through

the following equations

∫ ∫  
0
δX FI (X ) =
T
ε ∆ (x 0, δx)T Cε(x 0 ) dAdΩ (2.32)
Ω p ∪Ωb A
∫ ∫   ∫
0 0
δX FE (X ) =
T
ε (x , δx) Cα∆T e dAdΩ +
∆ T
δw Pnz dΩ (2.33)
Ω p ∪Ωb A Ωp

23
The nonlinear system given in Eq. (2.31) can be solved using Newton-Raphson (N-R) solver, where

the Jacobian is given by


∫ ∫  
J= ε ∆ (x ∆, δx)T σ(x 0 ) + ε ∆ (x 0, δx)T σ ∆ (x 0, x ∆ ) dAdΩ (2.34)
Ω p ∪Ωb A

A conventional load-stepping procedure that uses Newton-Raphson iterations with exact Jacobian

J can fail to stay on the stable equilibrium branch for nonlinear thermoelastic analysis even in

the absence of limit points. Bhatia et al. [20] presented a continuation method able to follow the

deformation branch in a load-displacement curve [36, 90]. The continuation method was shown

to circumvent the challenges of exact N-R iterations. The continuation solver has been used in

the structural optimization context [9], and its application to nonlinear thermoelastic load-stepping

resulted in an increased robustness. For more details we point the reader to the following papers

[9, 36, 90]. The continuation method used in this work is discussed in section 4.1.

2.3.2 Modal analysis

Modal eigensolutions are derived from the small-disturbance homogeneous response of the

structure about a nonlinear equilibrium state such that x(t) = x 0 + ∆x(t). Harmonic motion is

assumed such that ∆x(t) = y k eiωk t , where ω k and y k are the k th eigenvalue and eigenvector,

respectively. Linearizing Eq. (2.29) about the steady state solution x 0 , the variational statement

can be expressed as: find the k th eigenpair {ω k , y k }, such that for all δx
∫ ∫ ∫ ∫  
ω2k δx ρy k dAdΩ =
T
ε ∆ (y k , δx)T σ(x 0 ) + ε ∆ (x 0, δx)T σ ∆ (x 0, y k ) dAdΩ
Ω p ∪Ωb A Ω p ∪Ωb A

(2.35)

Eq. (2.35) can be expressed in the matrix form as

ω2k MYk = JYk (2.36)


24
where M and J are the mass and stiffness matrices, respectively, and relate to the variational

statement in Eq. (2.35) through

∫ ∫
δX MYk =
T
δxT ρy k dAdΩ (2.37)
Ω p ∪Ωb A
∫ ∫  
δX JYk =
T
ε ∆ (y k , δx)T σ(x 0 ) + ε ∆ (x 0, δx)T σ ∆ (x 0, y k ) dAdΩ (2.38)
Ω p ∪Ωb A

25
CHAPTER III

OPTIMIZATION PROBLEM DESCRIPTION

3.1 Optimization statement

The structural optimization problem used in this work is a minimization of the structural mass

m subjected to stress and stability constraints. The stress constraints are used to prevent structural

failure and the stability constraints are used to ensure static stability of the optimized structure.

The optimization problem is formulated as

minimize m(R)
R

subjected to: g1i (X, R) ≤ 0, for i = 1, . . . , nelems (3.1)

g2k (X, R) ≤ 0, for k = 1, . . . , 20 (3.2)

rl ≤ r j ≤ ru (3.3)

where R denotes the vector of design variables with r j the j th entry in R which is bounded with a

lower and upper limits rl and ru , respectively. In the following sections, we will define the stress and

stability constraints by providing their expressions as well as their sensitivity equations. The stress

constraints use the von Mises stress and relies on aggregation of the stress values at the element

nodes in order to obtain one stress value per element. The formulation for the stability constraints

is derived using three different approaches where in each method we derive the corresponding

sensitivity equations.
26
3.2 Stress Constraints

The stress constraints in this work use the von Mises stress expressed as

1
 
2 2     12
2
σvm = σx x − σyy + σyy − σzz + (σzz − σx x ) + 3 τx2y + τxz
2 2
+ τyz (3.4)
2

The stress constraints prevent reaching a yield stress which may cause structural failure. The

von Mises stress is evaluated at each quadrature point on multiple locations corresponding to the

material limits of the cross-sectional dimensions. For a quadrature point on a beam this corresponds

to z = {−h/2, 0, h/2}. On a plate, this corresponds to the nine locations obtained from combination

of (y, z) = {−w/2, 0, w/2} ⊗ {−h/2, 0, h/2}. Values on each element are then aggregated through

a p−norm aggregation expression and one constraint is included in the optimization problem for

each element ∫ 1/p


(σ (Ω)) p dΩ
© Ωei vm
σaggi = ­ ® . (3.5)
ª

Ωei
dΩ
« ¬
This work uses a value of p = 2 for stress aggregation over an element i, i = 1, ..., Nelems . The

stress constraints are defined as

g1i (X, R) = σaggi ≤ σcrit, i = 1, 2..., nelems (3.6)

where σcrit is the critical stress. For a well-posed optimization problem, the stress constraints are

rewritten as
σaggi
g1i (X, R) = − 1 ≤ 0, i = 1, 2..., nelems (3.7)
σcrit

The stress constraints in Eq. (3.1) are enforced at each element of the structure. The sensitivity

equation [16, 42, 57] for the stress σaggi with respect to a design variable r j can be expressed as

∂σaggi ∂σaggi dX

dσaggi
= + . (3.8)
dr j ∂r j X ∂ X dr j
27
where direct sensitivity analysis is used due to the large number of constraint functions. More

details on the direct and adjoint sensitivity are provided in Appendix A. The sensitivity of the

vector of discrete solutions X needed in Eq. (3.8) is obtained through


 −1
∂R ∂R

dX
=− , (3.9)
dr j ∂ X X ∂r j
∂R
where, J = ∂X is the Jacobian at the equilibrium state X computed from the solution of the system

of equation Eq. (2.31) which is provide here for convenience

R = FI (X) − FE (X) = 0, (3.10)

3.3 Stability Constraints

The stability constraints are used to ensure static stability of the optimized structure. Stability of

the structure requires the Jacobian matrix J to be positive definite. The modal eigenvalue problem

λ k MYk = JYk serves as a surrogate for static stability of the structure such that a positive eigenvalue,

λ k > 0, indicates a statically stable structure. A structure designed for an operating temperature

T must remain stable for all temperatures in the range [0,T]. This requires solving the modal

eigenvalue problem and compute the eigenvalues at multiple temperatures in this range and enforce

a positivity constraint at each temperature. In this work, we will present three approaches for

enforcing the stability constraints, with different definitions of Eq. (3.2) derived for each approach.

3.3.1 Approach 1: Constraints at final temperature with backtracking

In this approach, the frequency constraints are enforced at the final temperature T through

Eq. (3.2) which can be expressed as

−λ k + λ0
g2k (X, R) = ≤ 0, k = 1, 2..., 20 (3.11)
106
28
where λ0 is a lower bound for the eigenvalues which we set to λ0 = 100 rad2 /s2 . At the initial

design point and at the reference temperature T0 = 20◦ C, the smallest eigenvalue of the structure is

of order 106 rad2 /s2 . Therefore, a factor of 106 is used to appropriately scale the stability constraint

function for a well-conditioned optimization problem. More details on proper conditioning of the

optimization problem are given in section 4.2. Mode switching is one of the common problem

in optimization with stability constraints. In this work we chose to track the first 20 eigenvalues

to prevent enforcement of the stability constraint on the wrong mode. The stability constraints

in Eq. (3.11) ensure that the lowest 20 eigenvalues are bounded with a minimum eigenvalue

λ0 . These constraints ensure a statically stable design at the final operating temperature T but

does not force the structure to be stable in the range [0,T]. In order to overcome this issue, the

lowest 20 eigenvalues of Eq. (2.36) are computed at each load-step of the continuation solver.

The optimization line search is modified to trigger a backtracking algorithm when an eigenvalue

violating the constraint λ k ≥ λ0 , k = 1, . . . , 20, is encountered during the continuation solver

steps. The flowchart describing the backtracking algorithm in the optimization context is shown in

Fig. 3.1.

When a negative eigenvalue is found at a specific load step, the optimizer changes the vector of

design variable using the following equation

Rl = Rl−1 + γ (Rl − Rl−1 ) (3.12)

where Rl and Rl−1 are the vectors of design variables at the current and previous design iteration,

respectively. γ is a parameter used to alternate the optimization line search and get closer to the

vector of design variable obtained in the previous design iteration Rl−1 . After numerous parametric

29
Figure 3.1
A flowchart describing the backtracking algorithm within the optimization process.

studies with varying γ, its value was set to 0.2 for the backtracking algorithm. The backtracking

continues along this line-search until a design point that satisfies the eigenvalue constraints for

all load-steps is obtained. If backtracking is triggered but still results in a violated constrain, the

parameter γ is reduced through γ = γ 2 in order to obtain a new set of design variables closer to

Rl−1 .

The sensitivity equations for the eigenvalues λ k with respect to a design variable r j is given by

 
dλ k dM dYk d(JYk )
MYk + λ k Yk + M = (3.13)
dr j dr j dr j dr j

Multiplying Eq. (3.13) with the transposed eigenvector YkT and re-arranging the equation gives

K) T ∂M ∂Yk
dλ k YkT d(JY
dr j − λ k Yk ∂r j Yk − λ k Yk M ∂r j
T
= . (3.14)
dr j YkT MYk
30
The sensitivity of the stiffness matrix is expressed in matrix form as

T ∂J ∂ JYk dX
 
d(JYk ) dYk
δX T
= δX Yk + +J (3.15)
dr j ∂r j ∂ X dr j dr j

The first term on the right-hand side in Eq. (3.15) accounts for the sensitivity of the integral form

on the parameter, evaluated at the reference solution, and the third term is cancelled by the like

term from the mass matrix. The second term accounts for the sensitivity of the reference solution

and is evaluated as

∫ " T
d x0
 0
∂ JYk dX
∫  
∆ dx
 
δX T = ε (y k , δx) σ
∆ T ∆
x 0s , +ε , δx σ ∆ x 0, y k
∂ X dr j Ω p ∪Ωb A dr j dr j
d x0
 T  
0
+ ε ∆
x , δx σ ∆
, yk dAdΩ (3.16)
dr j

3.3.2 Approach 2: Aggregation

In this approach, we use the discrete K-S aggregation technique [58, 59] to force the structure to

be stable in the range [0,T]. Since the continuation solver has both the temperature and deformation

as outputs at each iteration, we can not determine the number of iterations a priori. For this reason,

we choose to aggregate over the number of iterations in the continuation solver which leads to a

total of 20 inequality constraints, one for each frequency. We define the eigenvalue λ ki , k = 1, ..., 20

and i = 1, ..., niters the k th eigenvalue at the i th iteration where niters is the total number of iterations

in the continuation solver. In this approach, we present the formulation for the aggregation of two

important quantities: aggregation of the eigenvalues and aggregation of the constraints shown in

Eq. (3.11).

31
3.3.2.1 Eigenvalues aggregation:

For the eigenvalue aggregation over the range [0,T], the expression for the stability constraint

in Eq. (3.2) is written as


λ0 − λaggk
g2k (X, R) = ≤0 (3.17)
106

where the aggregated eigenvalues λaggi are defined as


"n  #
iters
1 Õ  
λaggk = min λ ki − ln exp −ρ λ ki − min λ ki (3.18)
 
i ρ i=1
i

As indicated in Approach 1, the smallest eigenvalue is of order 106 , hence the scaling factor 106

used for a well-defined optimization problem. The limit property of Eq. (3.17) is given by

1
lim g2k (X, R) = (λ0 − min λ ki ), for k = 1, ..., 20 and i = 1, ..., niters (3.19)
ρ→∞ 106 i

While increasing the parameter ρ reduces the aggregation error, it makes the optimization problem

highly nonlinear which makes the problem difficult to solve. A value of ρ = 100 is used in this

work. Sensitivity of the constraint expressed in Eq. (3.17) can be expressed as


Íniters    dλk 
dg2k (X, R) −1 j exp −ρ λ kj − min λ
j k j dr j
j

= 6 Íniters (3.20)
10 exp −ρ λ ki − mini λ ki

dr j i

The sensitivity of the eigenvalue λ ki with respect to a design variable r j is given in Eq. (3.14) and

recalled here for convenience

K) T ∂M ∂Yk
dλ k YkT d(JY
dr j − λ k Yk ∂r j Yk − λ k Yk M ∂r j
T
= . (3.21)
dr j YkT MYk

32
3.3.2.2 Constraints aggregation:

Another way of enforcing the stability constraints using aggregation would be to aggregate

the constraints cki (X, R) for k = 1, 2, . . . , 20 over the load-stepping iterations. This lead to the

aggregated constraints g2k (X, R) expressed as


"n  #
iters
1 Õ  
g2k (X, R) = max cki (X, R) + ln exp ρ cki (X, R) − max cki (X, R) ≤0 (3.22)
i ρ i=1
i

where
−λ ki + λ0
cki (X, (R)) = , k = 1, 2, . . . 20 and i = 1, 2, . . . niters (3.23)
106

The sensitivity of the eigenvalue λ ki with respect to a design variable r j is given in Eq. (3.14) and

recalled here for convenience

K) T ∂M ∂Yk
dλ k YkT d(JY
dr j − λ k Yk ∂r j Yk − λ k Yk M ∂r j
T
= . (3.24)
dr j YkT MYk

3.3.3 Approach 3: constraint on temperature

In this approach, we are constraining the temperature T such that T is less than a temperature

Tmin at which the lowest eigenvalue λmin equals the lower bound λ0 , i.e.

Tmin | λmin =λ0 ≥ T (3.25)

33
This allow us to reduce the number of constraints from 20 to 1 compared to Approach 1 and 2.

The optimization problem for this approach is reformulated and expressed as

minimize m(R)
R

subjected to: g1i (X, R) ≤ 0, for i = 1, . . . , nelems (3.26)

g2 (X, R) ≤ 0, (3.27)

rl ≤ r j ≤ ru (3.28)

where the inequality constraint g2 (X, R) is expressed as

Tmin
g2 (X, R) = 1 − ≤0 (3.29)
T

Tm in is calculated using an interpolation scheme when λ0 is reached. In the case where T is

reached before Tmin , a first order interpolation is used to obtain the value of Tmin which saves in

computational cost. The sensitivity of λmin with respect to a design variable r j is given by

dλmin ∂λmin ∂λmin dX ∂λmin dTmin ∂λmin ∂ X


   
dTmin
= + + + =0 (3.30)
dr j ∂r j ∂ X dr j T ∂r j dr j ∂r j ∂Tmin r j dr j

The sensitivity of Tmin with respect to a design variable r j can be expressed as


h i
∂λmin ∂λmin dX
dTmin − ∂r j − ∂X dr j T
= h i (3.31)
dr j ∂λmin ∂λmin ∂X
∂r j + ∂r j ∂Tmin rj

This constraint ensures that the all eigenvalues of the structure are greater than λ0 in the range

[0,T].

34
1,500

1,000
80

ω1 (rad/s)
60
500
40
20
0 0
30 40 50
0 50 100 150
temperature difference ∆T (◦ C)
∆T = 120◦ C ∆T = 140◦ C ∆T = 160◦ C

Figure 3.2
The first natural frequency of the optimized stiffened panels for operating thermal loads of
∆T = {120, 140, 160}◦ C.

A typical behavior of the lowest frequency ω1 versus T is shown in Fig. 5.6. In case the

temperature Tmin is reached before T, the constraint g2 (X, R) is violated. Otherwise, when T is

reached before Tmin the constraint is satisfied which guaranty all eigenvalues λ ki to be greater or

equal than the lower bound λ0 .

35
CHAPTER IV

IMPLEMENTATION

This work leverage the functionality in the Multidisciplinary-design Adaptation and Sensitivity

Toolkit (MAST) [1, 17], which is an actively developed software within our research group. MAST

is an open-source computational framework written in object-oriented C++, and leverages state-

of-the-art open-source computational libraries [2, 4, 5, 11, 12, 44, 45, 55, 89] to provide a scalable

finite element analysis framework for design optimization of multi-physics simulations. MAST’s

capability to perform nonlinear thermoelastic analysis has been verified through benchmark analysis

with Abaqus [62, 73]. This work uses Globally Convergent Method of Moving Asymptotes

(GCMMA) [96, 97, 98] as the optimizer for the structural sizing optimization studies.

The first step in the development of a new design methodology for these types of structures is

development of the appropriate design analysis capability. The analysis capabilities of commercial

software are well developed incorporating a wide variety of disciplines including some tightly

coupled multidisciplinary analysis capabilities and many sequentially coupled analyses. The

“design” analysis capabilities of such software is much weaker however due to the lack of sensitivity

analysis. While sensitivity analysis capabilities do exist in some software, it is often limited to

single disciplines or linear analyses. For those supporting nonlinear sensitivity analysis, it is often

achieved in a linear way using equivalent static loads (ESLs), which do not work well with design-

36
dependent loads. Nonlinear analyses of multidisciplinary structures and the sensitivity of their

responses can be challenging to obtain in an efficient manner.

This chapter contains two main sections. The first section is dedicated to the continuation solver

where we define and describe the solver used to solve the nonlinear system of equation defined

previously. In the second section, we discuss the optimizer GCMMA which is used to generate the

results provided in this work.

4.1 Continuation Solver

This section describes the continuation method used in this work. Consider the nonlinear

system previously defined in Sec 2.3.1 which expressed as

R(X,T) = 0, (4.1)

where, X ∈ R is the system state and T is the thermal load, a parameter that influences the system.

A fixed point of this system identifies an equilibrium state. Stability of a fixed point is typically
∂R(X,T)
identified using eigenvalues of the Jacobian, J = ∂X , such that a stable fixed point is one where

all eigenvalues of the Jacobian are on the negative real-axis, Re(λi ) ≤ 0, Im(λi ) = 0, i = 1, ..., 20

and λi is the ith eigenvalue of J. A fixed point is unstable when the eigenvalue lies on the positive

real-axis Re(λi ) > 0, Im(λi ) = 0. Since the system depends on T, changing this parameter leads

to a change in the system state X. Starting from one fixed point, the loci of all connected points

obtained with varying values of T is called a branch and the nature of stability evolves along each

branch. Two branches can intersect each other at a bifurcation point. Computational approaches

used to trace these branches in the {X,T } space are called continuation methods. Within the

contexts of limit-point analysis in structural systems the continuation method or path-following


37
of Crisfield [28] and Riks [88] are very popular. An extensive literature exists on the analysis

of nonlinear systems and on numerical methods to compute various properties of such systems

[37, 91]. The present work follows the development of a continuation solver described in Seydel

[91] and Govaerts [37]. The continuation approach falls under the category of predictor-corrector

methods. The superscript (.)T denotes the transpose of a vector whereas the superscript (.)(T) means

the quantity corresponding to the temperature T. At a point on the branch the tangent is defined as
T
t1 = {t1(X) , t1(T) }T , where t1(X) and t1(T) are the components of t1 corresponding to X and T. Then, a

predictor is obtained based on this search direction and a prescribed step-size ∆s. A corrector then

searches in a direction that is orthogonal to this predictor (see Fig. 4.1).

Figure 4.1
Illustration of path continuation approach that iterates through points A → B → C → D, such
that D is on the curve R(X 0,T) = 0 at a distance ∆s from A.

Various elements of the continuation solver implemented in the present work are discussed

next. A continuation system defines a system-of-equations to solve for both X and T. Numerical

38
conditioning of the system requires the variables to be suitably scaled. In this example, T the

thermal load (in ◦ C) is significantly different from the order-of-magnitude of displacements (in

m) and rotations (in rad). Consequently, the choice of these variables can significantly influence

the step-size computed for the system. The implementation uses factors νX and νT for X and T,

respectively to define a scaled space { X̃, T̃ }. The scaled coordinates are defined as

X̃ = νX X (4.2)

T̃ = νT T (4.3)

(4.4)

leading to the relation for infinitesimal changes

d X̃ = νX dX (4.5)

dT̃ = νT dT (4.6)

(4.7)

These are used to define the arc-length in the following discussion and the initialization of νX and νT

is discussed later in this section. A predictor is obtained from the tangent of the load-displacement

curve. Two typically approaches can be used to compute the tangent vector, an arc-length aproach

and a pseudo-arc length approach which are discussed in Sec. 4.1.1 and Sec. 4.1.2, respectively.

Starting from a point { X̃0T , T̃0 }T on the branch, the predictor step; defined using either of the two

approaches mentionned earlier; uses the tangent direction and the step-size ∆s̃ to define a point

{ X̃∗T , T̃∗ }T (see Fig. 4.1):


     
     ( X̃) 
 X̃∗   X̃0   t1 

  
  
 
= + ∆s̃ (4.8)
 
  
  

     (T̃) 
 T̃∗   T̃0   t1 

  
  
 
  
     
39
Since this predicted point may not lie on the branch, a corrector is introduced which ensures that

the correction is orthogonal to the tangent direction, which also serves as the constraint equation.

If the corrected point on the curve is { X̃ T , T̃ }T , then the constraint is defined as

    T  
     ( X̃) 
 X̃   X̃∗  t
     
©    ª® 1 

G(X,T, ∆s̃) = ­ =0 (4.9)
­     

− ®
­    ®  (T̃) 
 T̃   T̃∗   t1 

 
 
 
 
 

«   ¬  
Note that the constraint is defined in terms of {X,T } since the final set of equations will directly

solve for these variables as opposed to the scaled variables. Substituting from Eq. (4.8) into

Eq. (4.9) gives

      T  
     ( X̃)   ( X̃) 
X̃ X̃ t t
       
©   0  
1
 ª 
1 

G(X,T, ∆s̃) = 0 = ­ (4.10)

 
 
 
 
 
 
 
− − ∆s̃ i
­ ®
®
­     ( T̃)  ®  ( T̃) 
 T̃   T̃   t1  t 
       
«   0   ¬  1 
       
    T  
     ( X̃) 
X̃ X̃ t
     
©   0
  ª 1 

= ­ (4.11)
­    ®  

− ® − ∆s̃
­    ®  ( T̃) 
 T̃   T̃  t 
     
«   0 ¬  1 
     
  T  
( X̃) 
ν
  
(X − X ) t
   
© X 0  ª 1 

= ­ (4.12)
­ ®  

®
 (T̃) 
 νT (T − T0 ) 
­ ® 

 t1 

 
  
« ¬
  
where, the last term in Eq. (4.11) is simplified due to the condition t1T t1 = 1, and the scaling

factors defined in Eqs. 4.2 and 4.3 are used in Eq. (4.12). The continuation method appends the

nonlinear system in Eq. (4.1) with the constraint equation G(X,T, ∆s̃) = 0, such that the combined

system-of-equations becomes
   
0
   
 R(X,T)

 
 
 
= (4.13)
 
  

0
   
 G(X,T, ∆s̃) 

  
 
 
   

40
with X and T as the unknowns. Eq. (4.13) is solved using a Newton-Raphson method, where each

l th iteration is defined as

 ∂R(X,T) ∂R(X,T)
  
 
  
  
 X(l) − X(l−1)   R(X,T) 
   
 ∂X ∂T
  
= (4.14)
    

 
 ∂G(X,T,∆s̃) ∂G(X,T,∆s̃) 
   
  
 T(l) − T(l−1)   G(X,T, ∆s̃) 
  
 ∂X ∂T     
  (l−1)    
{·}(l−1) implies that the quantity is evaluated using solution at the (l − 1) iteration. The Jacobian

requires sensitivity of constraint with respect to X and T, which are readily obtained from Eq. (4.12)

as

∂G(X,T, ∆s̃)
= νX t1( X̃) (4.15)
∂X
∂G(X,T, ∆s̃)
= νT t1(T̃) (4.16)
∂T

The Newton-Raphson iterations are continued till a convergence criterion is satisfied.

Formulation of the constraint in Eq. (4.13) requires the step-size ∆s̃. This is computed prior to

initiating the continuation solver iterations. Assuming that the initial configuration is defined by the

pair (X0,T0 ), a load step of value ∆T is applied and the standard Newton-Raphson iterations are used

to compute the equilibrium solution for Eq. (4.1). Representing this solution as (X0 + ∆X,T0 + ∆T),

the step size is computed as ∆s̃ = ∆ X̃ T ∆ X̃ + ∆T̃ 2 . Given ∆X and ∆T, the scaling factors are

initialized as νX = 1/||∆X || and νT = 1/|∆T |. Then the step-length is obtained as

νX2 ∆X T ∆X + νT2 ∆T 2 = ∆s̃2 (4.17)

This value of ∆s̃ can be suitably scaled up/down during the continuation solves. Continuation of a

smooth system response in the absence of bifurcation and limit-points can be effectively handled

with large values of ∆s̃, otherwise smaller steps are needed when the response is strongly nonlinear.

41
An effective measure of the nonlinearity is the number of N-R iterations taken to solve Eq. (4.13).

If MD is the desired number of N-R iterations then the arc-length can be scaled for the i th step of

continuation solver as [28]


 b
MD
∆s̃i = ∆s̃i−1 (4.18)
Mi−1

where, Mi−1 is the number of N-R iterations to converge the k − 1 continuation step and b is a

predefined exponent, usually taken as 0.5. Upper and lower limits on ∆s̃ k are specified in the

present work as 0.005 and 100, respectively. The values of the continuation solver parameters used

in this work are shown in Table. 4.1.

Table 4.1
Pseudo Arc-length continuation solver parameters.

Continuation solver Parameters


Initial load ∆T0 0.01
Desired number of N-R iterations MD 5
Maximum number of N-R iterations 6
Minimum arc-length step min ∆s̃ 10−9
Maximum arc-length step max ∆s̃ 100
Maximum number iterations in the Continuation solver 105

A pseudo code for the continuation solver used in this work is shown in Algorithm. 1.

The next sections discusses the two approaches used to compute the tangent vector, the arc-

length aproach and the pseudo-arc length approach.

42
Algorithm 1 Continuation solver pseudo code
1: Compute the step length ∆ s̃
2: while Temperature < Final temperature do
3: while Continuation solver flag is true do
4: Solve X and T using Predictor-Corrector
5: if Residual norm < tolerance then
6: Continuation solver flag is false
7: end if
8: if max Newton-Raphson iterations is reached then
9: if ∆s̃ ≥ minimum step length then
10: reduce step size
11: else if ∆s̃ < minimum step length then
12: Quit the solver
13: end if
14: end if
15: end while
16: Adaptively scale the step length ∆s̃
17: Update the temperature
18: end while

4.1.1 Arc-length Predictor

This approach defines a tangent along the branch coordinate s̃ in the { X̃, T̃ } space as the vector

{d X̃ T /d s̃, dT̃/d s̃}T . It follows that t1( X̃) = d X̃ T /d s̃ and t1(T̃) = dT̃/d s̃. The following procedure

provides a mechanism to compute this tangent direction. The Euclidean norm of d s̃ is defined as

d X̃ T d X̃ + dT̃ 2 = d s̃2 (4.19)

Assuming differentiability along s̃ this can be rewritten as


 2
d X̃ T d X̃ dT̃
+ =1 (4.20)
s̃ d s̃ d s̃

which is used later in the definition of the constraint equation. Similarly, assuming differentiability

with respect to T̃, Eq. (4.19) is rewritten as


 2
d X̃ T d X̃ d s̃
+1= (4.21)
d s̃ d s̃ dT̃
43
which provides a relationship for dT̃/d s̃ as

  −1
dT̃ d s̃ 1 1
= =q =r (4.22)
d s̃ dT̃ d X̃ T d X̃ 2
+1
 
νX dX T
dT̃ dT̃ νT
dX
dT dT +1

Note that dX/dT is the sensitivity of state, X, with respect to the thermal load T and can be

computed from linearization of Eq. (4.1), which gives

dX R(X,T)
J =− (4.23)
dT dT

Once dX/dT is available from the solution of Eq. (4.23), d X̃/d s̃ can be computed from Eq. (4.22).

Then, d X̃/d s̃ is readily obtained from chain-rule

d X̃ d X̃ dT̃ νX dT̃ dX
= = (4.24)
d s̃ dT̃ d s̃ νT d s̃ dT

The results of Eqs. 4.22 and 4.24 provide the tangent direction, and t1T t1 = 1 due to Eq. (4.20).

4.1.2 Pseudo Arc-length

The arc-length predictor described in the previous section lacks robustness around limit points

where the Jacobian matrix, J, of the original nonlinear system, Eq. (4.1), becomes singular.

Since the search direction of the predictor step depends on solution of the sensitivity problem

in Eq. (4.23), singularity of J leads to an ill-posed problem. In order to overcome the challenge

outlined above, the computation of the search direction is modified by using the following approach.

It is recognized that tangency of the search direction, t1 , also implies orthogonality with respect

to the gradient of the branch in the { X̃, T̃ } space. In order to maintain directionality of the search

direction, it is required that its dot product with a previous direction, t0 , be a positive number, that

44
is t0T t1 = 1, where the value of 1 is an arbitrarily chosen positive number. These two conditions

can be stated in the form of a system-of-equations

 ∂R(X,T) ∂R(X,T) 
    
( X̃) 
0
  
 t1 

 
 
 
 ∂ X̃ ∂T̃ 
= (4.25)
   


    
 t ( X̃) t1(T̃) (T̃) 
 t1  1
  
 0    
    

where, t0X̃ and t0T̃ are the components of t0 corresponding to X̃ and T̃, respectively. Multiplying the

first row with the scaling factor νX results in

 ∂R(X,T) νX ∂R(X,T) 
    
( X̃) 
0
  
t
   
 ∂X νT ∂T  1 
  
 
 =  (4.26)
   

 

 t ( X̃) t1(T̃) (T̃) 
 t1  1
  
 0    
    
Solution of Eq. (4.26) provides the search direction at a given point on the curve, after which t1 is

rescaled to a unit norm,


t1
t1 = (4.27)
||t1 ||

It is noted that this work uses the pseudo arc-length approach for the calculation of the tangent

vector.

4.2 Globally Convergent Method of Moving Asymptotes (GCMMA)

The Globally Convergent Method of Moving Asymptotes, GCMMA, is an optimizer used to

solve the optimization problem stated in Sec. 3.1, which we re-write here for convenience.

minimize m(R)
R

subjected to: g1i (X, R) ≤ 0, for i = 1, . . . , nelems (4.28)

g2k (X, R) ≤ 0, for k = 1, . . . , 20 (4.29)

rl ≤ r j ≤ ru (4.30)
45
For the sake of simplicity, the optimization problem is rewritten in a compact form and expressed

as

minimize f0 (R)
R

subjected to: fi (X, R) ≤ 0, for i = 1, . . . , nconst (4.31)

where f0 is the objective function and fi is the i th constraint function, and nconst is the total number

of constraint functions. In order to describe the optimizer mechanism, we will define the main

components and function used by the optimizer. The optimizer relays on two main loops, an

outer one with index m and an inner one with index n. The double index {m, n} denotes the inner

iteration n at the outer iteration m. At iteration m = 1, we start with the initial vector of design

variables R 0 which satisfies the optimization problem listed earlier. In order to march in the outer

loop from an iteration m to an iteration m + 1, the optimizer generates and solves a subproblem.

For more details on the subproblem generated by the optimizer, we point the reader to [98]. In the

subproblem, the objective function f0 and constraint function fi are replace with convex functions

f˜im,0 , for i = 0, ..., nconst . The optimal solution of the subproblem at {m, 0} is given by X̂ m,0 . The

original objective function and constraint are then compared to the convex function to determine

if the obtained solution is conservative. This is done through the expression

f˜i(m,0) ( X̂ (m,0) ) ≥ fi ( X̂ (m,0) ) for all i = 0, ..., nconst (4.32)

If the condition is true, the outer loop goes through another iteration without any inner iteration

and the vector of design variables is updated such that If the condition is false, a new set of convex

functions f˜im,1 are generated. The functions f˜im,1 are more conservative than the previous convex

functions at the inner iteration n = 0 at the index i where the inequality is violated. This mean
46
that f˜im,1 (X (m) ) > f˜im,0 (X (m) ) when the inequality is violated and f˜im,1 (X (m) ) = f˜im,0 (X (m) ) when

the inequality is satisfied. The optimizer calculates the new optimal solution of the subproblem

denoted by X m,1 . Similarly, the original obejective function and constraints are compared to the

convex functions through the expression

f˜i(m,1) ( X̂ (m,1) ) ≥ fi ( X̂ (m,1) ) for all i = 0, ..., nconst (4.33)

If the inequality is satisfied, the inner loop is terminated with one iteration and we move to iteration

m + 1 in the outer loop. Otherwise, the optimizer undergo another inner iteration which generates

a new subproble, convex function and optimial solution. It should be mentioned that at the inner

iterations, there is no need to recalculate the gradients, since the vector of design variable have not

changed. The gradients of the functions are only calculated once in each outer iteration. This leads

to an important saving in computational cost since in most optimization problems the calculation

of the functions gradient is the most expensive operation. A pseudo code for the GCMMA code

is provided in Algorithm. 2. Relevant parameters for the optimization problems are listed in

Table. 5.6.

47
Table 4.2

GCMMA optimization parameters.

GCMMA Parameters

rel. step size 0.01

constraint penalty 105

asymptote reduction 0.7

asymptote expansion 1.2

max num. inner iters. 10

Algorithm 2 GCMMA pseudo code


1: Set Flag to terminate outer loop and inner loop to False
2: while Flag to terminate outer loop is False do
3: Compute the objective function and constraints
4: Compute the objective function gradient and the constraints gradients
5: while Flag to terminate inner loop is False do
6: Generate the sub-problem and solve it
7: Calculate the objective function and constraints
8: if maximum number of inner iteration is reached then
9: Change Flag to terminate inner loop to True
10: else if Solution is conservative then
11: Change Flag to terminate inner loop to True
12: else if solution is not conservative then
13: restart the inner loop
14: end if
15: end while
16: Update the solution
17: if maximum number of outer iteration reached then
18: Change Flag to terminate outer loop to True
19: else if Relative change in the objective function is within the tolerance then
20: Change Flag to terminate outer loop to True
21: else if Relative change in the objective function is bigger than tolerance then
22: Start new outer iteration
23: end if
24: end while
48
4.3 Verification and Validation

The computational tool developed in this work can be divided into two part: Structural analysis

and Sizing Optimization. While the structural analysis can be verified and validated, verification of

the sensitivity equations was performed through a comparison with finite difference. The validation

step of the optimization is nontrivial since the purpose of an optimization study is to obtain a non-

intuitive design. The verification step of the structural analysis has been conducted in [62] where

structural results for our model were benchmarked against Abaqus, and numerical validation of

MAST for future use in multidisciplinary nonlinear design methodologies was performed in [73].

4.3.1 Verification of the sensitivity equations

The design variables are the parameters that you are allowing the optimizer to change when

seeking a minimum or a maximum of a function depending on your optimization problem. The

optimization problems are typically formulated as minimize f (R) such that R is the vector of

design variables. For the blade-stiffened panels, the design variables are the thickness of the panel

skin and the width and length of the stiffeners. For the hat-stiffened panel, the design variables are

the thickness of the panel skin and thickness of the stiffeners.

In optimization and in the special case of gradient-based optimizers, the gradients, often called

sensitivity analysis, are necessary and needed by the optimizer to find a local minimum/maximum

of your function. Those gradients are the derivatives of the objective function with respect to the

design variables and the derivatives of the constraint functions with respect to the design variables.

These gradients/derivatives can be calculated analytically or Numerically. Analytic derivatives are

calculated by hand and then implemented. Often times, they are hard to compute specially when

49
we are dealing with shape optimization or topology optimization. Numerical and semi-analytic

derivatives are usually used when the derivative is hard to compute by hand. Some of the know

methods to compute those derivatives numerically are Finite Difference (FD) [48], Complex step

(CS) [66], and automatic differentiation (AD) [21]. Finite difference method provides insight into

approximate value of the derivative, it is a widely used technique due to its simplicity and ease of

implementation. One of the problems faced when estimating derivatives with finite difference is the

computational cost. Another problem is the dependency of the derivatives on the step size chosen

for FD. The complex step technique follows the same logic as finite difference except that it uses both

real and imaginary parts for the computation of the derivatives. While this technique alleviates

the challenges related to the step size, the issue with computational cost remains. Automatic

differentiation is a relatively newer technique which was shown to be computationally efficient and

bypasses the problems of the step size. It is currently implemented in C++ libraries and frameworks

such as ADEL[48] and its incorporation with MAST is currently ongoing. The semi-analytic as the

name explains is a combination of analytic derivation and numerical approximation. The choice of

using analytic derivatives is know self-explanatory, the derivative are exact and are computationally

efficient. The only step then required is verification. One of the methods to verify the gradients is

to compare them against finite difference. In this comparison, the only objective is to make sure

the derivatives are correct up to a certain tolerance.

In finite difference the step size is little tricky. Driving the error to machine precision zero

is impossible. When the step size is large enough the finite difference derivative suffers from

truncation error and when the step size is small enough or goes toward zero the derivative suffers

from round-off error. For the purpose of verification, we used a range of steps until we determine

50
that a step of 10−5 was the step size with minimal error. That step size was then used to verify

our analytic derivatives. Verification of the derivatives was performed whenever the constraint

functions would change, the structure would change or the type of analysis (linear/nonlinear) would

change.

In this work, we chose to use analytic derivative in order to save on computational cost and

avoid dependencies on the step size in the numerical techniques. Our verification of the derivatives

is performed against the finite difference approach where we used a central scheme with a step

size of 10−5 . Due to the large number of constraints (O(5000)), the derivatives presented are the

derivatives of the objective function, the derivative of the first stability constraint and the derivative

of the first stress constraint. It should be noted that all derivatives satisfied a relative tolerance set

to 10−3 with a step size of 10−5 .

51
Table 4.3

Verification of the analytical sensitivities against finite difference approach.

Objective function first stability constraint first stress constraint

Analytical Numerical (FD) Analytical Numerical (FD) Analytical Numerical (FD)

6.1889501953125 6.18895019531828 -8.96467903921057 -8.96467942260725 1.98732907635062 1.98732886254604

12.3960498046875 12.3960498046971 -3.16699790136018 -3.16699962789046 -5.61562677956741 -5.61562783559232

12.3960498046874 12.396049804686 14.5067757369616 14.5067754230066 2.3288580647616 2.32885772306068

6.1889501953125 6.18895019532383 11.8881426893017 11.8881458329278 1.29835719159657 1.29835722468829

0.2475580078125 0.247558007815396 -3.33436997537994 -3.33436985576751 0.004569450905028 0.004569450962055

0.4958419921875 0.49584199218744 -2.25675336705377 -2.2567549580188 -0.072395610630274 -0.072395610711817

0.4958419921875 0.49584199218744 -1.11635483082149 -1.11635636577123 0.022251396161398 0.022251396186723

0.2475580078125 0.247558007809845 -0.918573621100849 -0.918573501307307 0.045618302727194 0.045618302746409

0.12377900390625 0.123779003902147 -4.92385968387231 -4.92385945285445 0.002315332638813 0.002315332631175

0.24792099609375 0.24792099609372 -2.54575842795441 -2.54575737979045 -0.036202260660767 -0.036202260683682

0.24792099609375 0.247920996088169 -2.08359791413682 -2.08359930785606 0.011111183649784 0.011111183606527

0.12377900390625 0.123779003913249 -2.23811468054497 -2.23811175423804 0.022786744790083 0.022786744813263

0.12377900390625 0.123779003902147 -4.59423566856003 -4.594235618538 0.003203437106936 0.003203437121124

0.24792099609375 0.247920996099271 -1.92006642582367 -1.92006649111231 -0.010452156491712 -0.010452156490315

0.24792099609375 0.24792099609372 -0.75635555983757 -0.756355688458221 -0.000327314636409 -0.000327314642057

0.12377900390625 0.123779003902147 -1.2683443339279 -1.26834290402611 0.007543417543146 0.007543417518407

0.061889501953125 0.061889501956625 -7.4009711432348 -7.40097102418513 0.001596520671391 0.001596520654568

0.123960498046875 0.123960498055187 -3.42327154207339 -3.42327116183982 -0.005234137141557 -0.005234137145926

0.123960498046875 0.123960498049636 -2.87881771032956 -2.8788194032292 -0.000174145518302 -0.000174145542431

0.061889501953125 0.061889501956625 -3.3104843425584 -3.31048473898309 0.003753950172919 0.003753950178487

52
CHAPTER V

RESULTS

5.1 Problem setup

In this Section, we present results for the optimization problem described in Chapter. III which

we rewritten here for convenience.

minimize m(R)
R
σaggi
subjected to: g1i := − 1 ≤ 0, for i = 1, . . . , nelems (5.1)
σcrit
−λ k + λ0
g2k := ≤ 0, k = 1, 2..., 20 (5.2)
106
0.5mm ≤ r j ≤ 5cm (5.3)

The optimization study is conducted on two types of stiffened panels, a blade-stiffened panel

(section 5.2) and a hat-stiffened panel (section5.3). The physics modeled in this work include

a static and a modal probelm. The static problem aims to solve the thermoleastic response of

the structure under a uniformly disctributed thermal load. The modal problem is solved about the

nonlinear equilibrium state where the eigenvalues and eigenvectors are computed. Those parameter

are necessary to determine stability of the structure.

In section 5.2, we first conduct sizing optimization studies for thermally stressed blade-stiffened

panels with operating temperatures in the range ∆T = {120 − 180}◦ C, subjected to stress and

stability constraints. The results obtained include convergence of the objective function and
53
constraints, the frequency behavior of the optimized structures and the out-of-plane deformation

as well as the stress distribution in the panel. In the second part, we investigate the impact of

the nonlinear strain εvk in sizing optimization of thermally stressed stiffened panels. Finally, in

section 3.3 we inspect the effect of backtracking defined in section 3.3.1 on the stability constraints

and frequency behavior of the optimized design.

The results shown for the hat-stiffened panel in section 5.3 include sizing optimization studies of

corrugated and noncorrugated hat-stiffened panels. Each structure is studied where the frequency

of the structure is plotted and the stress distribution with the deformation plots are presented. The

corrugation aspect is extend further by studying the impact of its position on the structural stresses

and the stability properties.

For each structure, an initial deformation is applied on the panel skin to model panel imper-

fections. In [83], the authors stated that modelling of initial imperfections is of great importance

when evaluating panel stability behavior. In [71], The initial geometry is subsequently seeded

with an imperfection in the shape of the fundamental buckling mode. The magnitude of this

imperfection is 10% of the skin thickness, a value that is representative of typical imperfections

present in conventional riveted fuselage structures. The authors in [71] also stated that initial

overall imperfections can significantly affect the structural response predictions, with the failure

load being particularly sensitive. While such imperfections, in general, do not occur in actual

fuselage structures, they may occur in test specimens hence, from a test–analysis correlation per-

spective, any notable overall imperfection should be correctly reproduced in the analysis model. In

[46] , The detailed initial imperfections of the test specimens, referred to an orthogonal Cartesian

coordinate system, were measured with a Nardini-SZ25120T lathe machine before testing. The

54
imperfections of the finite element model were directly mapped from the test measurements. In

[63], the inclusion of the panel imperfection in the computational specimen was done in two steps.

First, the perfect mesh was distorted via a static analysis to obtain a nodal displacement similar to

the experimentally measured imperfection. Second, the mesh obtained from the first step is used

to perform the collapse analysis. In this work, we use the same procedure as [63], where in the

first step we apply the small load P on the structure and in the second step we thermally load the

structure. While we don’t have an experimentally measured imperfection, we choose to an initial

load magnitude P that generates a deformation less than 10% of the maximum deformation at the
max w|∆T =0◦ C
final temperature. In Figure 5.1, the ratio max w|∆T =160◦ C × 100 is plotted for multiple values of P,

where w is the out-of-plane deformation of the panel.

55
Figure 5.1
Percentage of the out-of-plane deformation at ∆T = 0◦ C with respect to the deformation at
∆T = 160◦ C.

5.2 Blade-stiffened panel

The stiffened panel used in this work is a square plate of length L = 0.3m connected to three

equally spaced blade stiffeners along the x−axis as shown in Fig. 5.14.

56
Figure 5.2
3D representation of the blade-stiffened panel.

Figure 5.3
y − z cross-section of the blade-stiffened panel.

57
Clamped boundary conditions with in-plane restraints (u0 = v0 = w = θ x = θ y = θ z = 0)

are enforced on all four edges of the plate and end-points of the stiffeners. Following a mesh

convergence study, the analysis domain used for the blade-stiffened panel is discretized with a

mesh of 64 × 64 elements for the panel skin and 64 elements along the x−axis for each blade

stiffener. The material used for the structure is Inconel−718 [25]. The material properties for

Inconel−718 with varying temperature are shown in Table 5.1.

Table 5.1
Typical physical properties of Inconel-718 at room and elevated temperatures.

∆T T E α yield strength
ν
(◦ C) (◦ C) (GPa) (×10−5 /◦ C ) (GPa)
0 20 204 0.294 1.41 1.03
60 80 200.25 0.2887 1.41
80 100 199 0.287 1.41 1.06
100 120 197.8 0.2856 1.41
120 140 196.6 0.2842 1.41
140 160 195.4 0.2828 1.41
160 180 194.2 0.2814 1.41
170 190 193.6 0.287 1.41
180 200 193 0.28 1.41 1.04
200 220 191.8 0.2786 1.41
220 240 190.6 0.2721 1.41
240 260 189.4 0.2757 1.41
250 270 188.8 0.275 1.41
260 280 188.2 0.2742 1.41
280 300 187 0.2728 1.42 1.02

While the material properties for Inconel−718 are known to be temperature dependent, the

present work assumes constant properties chosen at the operating temperature ∆T. The imperfection

in the panel is modeled through an initial mechanical load of P = 300Pa applied uniformly on the

panel skin in order to create a small geometric imperfection that induces transverse displacement.

58
The magnitude of P was carefully chosen so that its impact on the post-buckled response and stress

is negligible. The temperature difference ∆T is calculated with respect to a reference temperature

T0 = 20◦ C. The design variables include the panel skin thickness and the width and height of the

stiffeners. The initial magnitude of the design variables along with the upper and lower bounds are

shown in Table 5.5.

Table 5.2
Initial values of the design variables for the optimization problem.

Parameter Initial value Lower limit Upper limit


panel skin thickness 1 mm 0.5 mm 50 mm
height of the stiffeners 3 mm 0.5 mm 50 mm
width of the stiffeners 3 mm 0.5 mm 50 mm
m mass of the structure m0 = 0.80 kg

The design variables are defined at 7 equally spaced stations specified along the x−axis with a

linear variation between stations leading to a total of 48 design variables.

In order to save on computational cost, we choose to take advantage of the structure symmetry

by using the same design variables for stiffener 1 and 3. Furthermore, we use the same design

variables for stations 1, 2, 3 and stations 5, 6, 7, respectively. This symmetry assumptions enable us

to reduce the number of design variables by more than 50%, where the number of design variables

reduces from 48 to 20.

59
Figure 5.4
y − z cross-section of the blade-stiffened panel.

5.2.1 Structural sizing with increasing thermal loads

Separate sizing optimization studies are conducted for operating thermal loads

∆T = {120, 160, 170, 180}◦ C. The parameters used in the optimization study are defined in Ta-

ble. 5.6.

The convergence of the objective function and constraints is plotted in Fig. 5.5.

In Fig. 5.5(a), we can notice that the operating temperature ∆T dictates the optimal mass

of the structure, where a larger thermal load requires a heavier structure. In Fig. 5.5(b), the

convergence of the maximum stress constraint maxi {g1i } is plotted. This plot determines if the

stress constraints are feasible, active or violated. From Fig. 5.5(b), the stress constraints are satisfied

60
Table 5.3
Optimization parameters.

Parameter Value
critical stress σcrit 1.02GPa
design variables upper bound ru 50mm
formulation design variables lower bound rl 0.5mm
initial mass of the structure m0 0.81kg
lower bound for the stability constraints λ0 100rad2 /s2
relative step size 0.01
constraint penalty 105
GCMMA (optimizer) asymptote reduction 0.7
asymptote expansion 1.2
max number of inner iterations 10
initial load ∆T0 0.01
desired number of N-R iterations MD 5
maximum number of N-R iterations 6
Continuation solver
minimum arc-length step min ∆s̃ 10−9
maximum arc-length step max ∆s̃ 100
maximum iterations in the Continuation solver 105

for ∆T = {120, 140}◦ C, active for ∆T = 160◦ C and violated for ∆T = 180◦ C. The structural mass

for the optimized feasible structures is shown in Table. 5.7

Table 5.4
Structural mass of the feasible optimized designs for ∆T = {120, 130, 140, 150, 160}◦ C.

Thermal load ∆T (◦ C) optimized mass m (kg)


120 0.5508
130 0.568
140 0.62266
150 0.6483145
160 0.6989962

In Fig. 5.5(c), the convergence of the maximum stability constraint max k {g2k } is plotted and

we can notice that for all temperatures, the stability constraints are feasible. As a reminder, the

61
0.8
1
0.6
0.9
0.4
= 120◦ C = 120◦ C

maxi {g1i }
∆T ∆T
0.8 0.2
∆T = 140◦ C ∆T = 140◦ C
∆T = 160◦ C ∆T = 160◦ C
∆T = 180◦ C 0 ∆T = 180◦ C
0.7
−0.2
0 5 10 15 20 25 30 0 5 10 15 20 25 30
design iterations design iterations

(a) scaled mass m/m0 (b) maximum stress constraint

−1.5

−2 ∆T = 120◦ C
∆T = 140◦ C
∆T = 160◦ C
max k {g2k }

−2.5 ∆T = 180◦ C

−3
0 5 10 15 20 25 30
design iterations

(c) maximum stability constraint

Figure 5.5
Convergence of the scaled mass and constraints in sizing optimization of a stiffened panel for
operating thermal loads of ∆T = {120, 140, 160, 180}◦ C.

stability constraints shown in Fig. 5.5(c) are enforced at the final temperature ∆T only. The structure

stability in the loading process is shown in Fig. 5.6 where the first natural frequency ω1 of the

optimized designs is plotted.

62
1,500

1,000
80

ω1 (rad/s)
60
500
40
20
0 0
30 40 50
0 50 100 150
temperature difference ∆T (◦ C)
∆T = 120◦ C ∆T = 140◦ C ∆T = 160◦ C

Figure 5.6
The first natural frequency of the optimized stiffened panels for operating thermal loads of
∆T = {120, 140, 160}◦ C.

The non-monotonic dependency of the frequency on temperature is clearly evident from this

plot and is a considerable challenge in the optimization problem. For ∆T = {120, 140, 160}◦ C, the

natural frequency reduces to a value close to 10rad/s. This rapid decrease in frequency is indicative

of a loss of stability margin, although none of the designs become unstable. Backtracking is

triggered at multiple optimization iterations for each of these designs, implying that the stability

constraint plays a crucial role for each case.

In Fig. 5.7, the stress distribution is presented for the optimized structures with the operating

temperatures ∆T = {120, 140, 160}◦ C. The stress distribution in Fig. 5.7 shows the maximum

stress being located at the boundaries of each structure. The stress in those panels results from

the combination of compressive stresses caused by the temperature, the restrained edges, and the

bending stresses caused by the transverse deflection. In the absence of bending, the magnitude

of compressive thermal stresses will be lower than the maximum stress by order-of-magnitudes;

as seen by the stress value at the center of the panel; where the curvature and the slope of the

63
(a) ∆T = 120◦ C (b) ∆T = 140◦ C

(c) ∆T = 160◦ C

Figure 5.7
Stress distibution (Pa) of the optimized stiffened panels for operating thermal loads of
∆T = {120, 140, 160}◦ C.

deformed panel are nearly zero with almost no contribution to the bending stresses. The higher

stresses on the boundary results from the considerable contribution of bending superimposed with

the compressive stresses.

64
(a) ∆T = 120◦ C (b) ∆T = 140◦ C

(c) ∆T = 160◦ C

Figure 5.8
Out-of-plane deformation (m) of the optimized stiffened panels for operating thermal loads of
∆T = {120, 140, 160}◦ C.

Contour plots of the out-of-plane deformation w of the optimized panels are shown in Fig. 5.8

for ∆T = {120, 140, 160}◦ C. The maximum displacement is located at the center of the panel

65
and increases with the operating thermal load ∆T. It should also be noted that the nonlinear

strain-displacement relationship causes negative displacement at the corners of the panel skin.

·10−3 ·10−5
1.2

1
1.1
0.9
1
0.8
(m)

(m2 )
0.9
0.7
0.8
0.6

0.7
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x/L x/L
∆T = 120◦ C ∆T = 140◦ C ∆T = 120◦ C ∆T = 140◦ C
∆T = 160 C
◦ ∆T = 160 C

(a) thickness of the panel skin (b) cross-section area (stiffeners 1 and 3)

·10−5
1.2

1.1

1
(m2 )

0.9

0.8

0.7
0 0.2 0.4 0.6 0.8 1
x/L
∆T = 120◦ C ∆T = 140◦ C
∆T = 160 C

(c) cross-section area (stiffener 2)

Figure 5.9
Design variables distribution of the optimized stiffened panels for operating thermal loads of
∆T = {120, 140, 160}◦ C.

The thickness distribution of the panel skin is shown in Fig. 5.9(a) for ∆T = {120, 140, 160}◦ C.

The thickness distribution has a similar pattern with different magnitude for each temperature.

66
The maximum thickness is located at the edges (stations 1 and 7) and the minimum thickness is

located at approximately x/L = 0.2 and x = 0.8 (stations 2 and 6). The design of a blade-stiffened

panel requires a heavier structure when increasing the operating thermal load as can be seen in

where a thicker panel skin is required for larger temperatures. In Figs. 5.9(b) and 5.9(c), the

cross-sectional area is plotted for the three blade stiffeners. The cross-sectional area distribution

for stiffeners 1 and 3 is similar to the panel skin thickness distribution where the maximum is

located at the boundaries and the minimum is located at stations 2 and 6. It can also be noticed that

increasing the thermal load requires a larger cross section area in stiffeners 1 and 3. For stiffener

2, the cross-sectional area is relatively unchanged when increasing the thermal load compared to

stiffeners 1 and 3. It is important to note that thermoelastic analysis can be counter intuitive due to

the design dependent load. Increasing the stiffeners dimensions will provide both a higher stiffness

and a higher thermoelastic loading and unlike purely mechanical loads, the optimizer cannot simply

increase thickness to reduce the out-of-plane deformation.

5.2.2 Impact of geometric nonlinearities

In this section, we investigate the impact of using a nonlinear strain in sizing optimization. To

do so, we are conducting a sizing optimization study using a linear strain with an operating thermal

load of ∆T = 160◦ C. We then compare the optimized structure with the results obtained before.

Two sizing optimization studies are performed where the first uses a nonlinear strain and the second

uses a linear strain. The optimized structures obtained are referred to structure 1 and structure 2 for

the former and latter, respectively. We start the investigation by analyzing the convergence of the

objective function and constraints for both sizing optimization studies. We then perform a static

67
nonlinear analysis on structure 2 and compare the frequency behavior, stress distribution and panel

deformation in the two structures.

5.2.2.1 Comparison 1: comparing the feasibility of structures 1 and 2

1 0.6

0.4
0.8 0.2

maxi {g1i }
0
0.6
−0.2

−0.4
0 5 10 15 20 0 5 10 15 20
design iterations design iterations
structure 1 structure 1
structure 2 structure 2

(a) scaled mass m/m0 (b) maximum stress constraint

Figure 5.10
Convergence of the scaled mass and maximum stress constraint in sizing optimization of a
stiffened panel for a thermal load of ∆T = 160◦ C using linear and nonlinear strain.

The convergence history for the structural mass and the stress constraints are presented in

Fig. 5.10. The figure shows that the maximum stress constraint remains inactive for structure 2,

while being active for structure 1. In structure 2, the optimizer is able to reduce all design variables

to the lower bound which leads to a lighter structure compared to structure 1.

5.2.2.2 Comparison 2: comparing the frequency behavior and the panel deformation

We conduct a nonlinear static analysis on the optimized design obtained with linear strain

(structure 2). The objective of this analysis is to understand the properties of the obtained structure

68
when incorporating the nonlinear strain. The first eigenvalue of structures 1 and 2 is plotted in

Fig. 5.11.

·106

1.5

λ1 (rad2 /s2 ) 0.5

−0.5
0 50 100 150
temperature difference ∆T (◦ C)
structure 1
structure 2

Figure 5.11
The fist eigenvalue λ1 of the optimized stiffened panels (structures 1 and 2).

For structure 2, the panel becomes unstable with the eigenvalue reducing below 0 rad/s at

approximately 49◦ C (Fig. 5.11). Structure 1 experiences the lowest eigenvalue at the same tem-

perature 49◦ C but remains positive at all temperatures. The stress and displacement contours are

shown in Figs. 5.12 and 5.13, respectively.

In Fig. 5.12, we compare the stress distribution in structures 2 using a linear and a nonlinear

strain. The stress distribution with linear analysis in Fig. 5.12(a) remains nearly constant due to

low transverse displacement. However, when this design is analyzed with nonlinear von Karman

strain (Fig. 5.12(b)) the extension-bending coupling causes the stress to increase by a factor of 1.5.

The maximum stress in structure 2 is larger than the critical stress set in the optimization problem,

meaning that the linear strain does not capture the true magnitude of the stresses in the structure.

69
(a) linear strain (b) nonlinear strain

Figure 5.12
Stress distibution (Pa) of structure 2 using a linear and nonlinear strain for a thermal load of
∆T = 160◦ C.

While sizing optimization of the panel with a linear strain resulted in feasible stress constraints,

in reality the stresses are larger than the limit we set. The extension-bending coupling creates an

out-of-plane deformation in the negative z axis at the corners of the panel as shown in Fig. 5.13.

The lack of this coupling influences also the magnitude of the deformation where nonlinearity

cause the displacement to increase by approximately a factor of 7.

These results show that a linear analysis is not reliable and can not be used for structural sizing

of thermally stressed stiffened panel. Nonlinear analysis is needed for accurately capturing the

physics of the problem.

70
(a) linear strain (b) nonlinear strain

Figure 5.13
Out-of-plane deformation (m) of structure 2 using linear and nonlinear strain for a thermal load of
∆T = 160◦ C.

5.3 Hat-stiffened panel

In this section, studies are presented for hat-stiffened panels subject to increasing thermal loads.

The stiffened panel used is a square plate of length L = 0.3m connected to three equally spaced

hat stiffeners along the x−axis (see Fig. 5.14 and 5.15)

The panel is subjected to elevated temperatures ranging from ∆T = 160◦ C to ∆T = 250◦ C. An

initial mechanical pressure load of 300Pa is applied uniformly to the surface of the panel to create a

small geometric imperfection in the panel that induces transverse displacement. The material used

for the structure is Inconel-718[25] and while its material properties are known to be temperature

dependent, we assume constant properties chosen at the final operating temperature similar to the

blade-stiffened panel. Clamped boundary conditions with in-plane restraints (u0 = v0 = w = θ x =

θ y = θ z = 0) are enforced on all four edges of the plate. The axial deformation along the x−axis

71
Figure 5.14
Schematic of a hat-stiffened panel with skin corrugation amplitude η.

Figure 5.15
Geometry of the hat-stiffened panel with skin corrugation η.

is set to zero at the extremities of the stiffeners, i.e. u0 | x=0 = u0 | x=L = 0. Following a mesh

convergence study, the analysis domain is discretized with a mesh of 70 × 70 elements for the panel

skin and 70 × 14 elements for each stiffener.

The design variables include the thickness of the panel skin and the thickness of each stiffeners

(Table 5.5). The design variables are defined at 7 equally spaced stations specified along the x−axis

with a linear variation between stations leading to a total of 28 design variables (see Fig. 5.16). In

order to save on computational cost, we choose to take advantage of the structure symmetry when
72
solving the optimization problem. The design variables for stiffener 1 are used for stiffener 3 and

the design variables at stations {1, 2, 3} are used at stations {7, 6, 5}, respectively.

(a) Stations location on the panel skin (b) Stations location on the stiffeners

Figure 5.16
Stations position in the stiffened panel.

This symmetry assumptions enable us to reduce the number of design variables by more than

50%, where the number of design variables reduces from 28 to 12. It should be noted that this

symmetry assumption is used for the optimization model only and that the structural analysis is

using the full finite element model. The upper and lower bounds for all design variables are set to

50mm and 0.5mm, respectively.

Relevant parameters for the optimization problems are listed in Table. 5.6.

The maximum allowable stress in the structure is set to σcrit = 1.02GPa. Geometric nonlinear-

ities are taken into account in the optimization process where ε = ε0 + εvk .

73
Table 5.5
Initial geometry of the hat-stiffened panel.

Parameter Initial value Lower limit Upper limit


plate thickness 1 mm 0.5 mm 50 mm
stiffeners thickness 1 mm 0.5 mm 50 mm
m mass of the structure m0 = 1.27 kg

Table 5.6
GCMMA optimization parameters.

GCMMA Parameters
rel. step size 0.01
constraint penalty 105
asymptote reduction 0.7
asymptote expansion 1.2
max num. inner iters. 10

5.3.1 Hat-stiffened panel without corrugation

This section presents results for a sizing optimization study of the hat-stiffened panel described

earlier, subjected to multiple thermal loads ranging from ∆T = 160◦ C to ∆T = 220◦ C. Fig. 5.17

shows convergence of the objective function (mass of the structure) and constraints (stress and

stability).

We can notice from Fig. 5.17 that for ∆T = {160, 180, 200}◦ C, the optimizer reaches a feasible

design where both stress and stability constraints are active. For ∆T = 220◦ C, the optimizer

converges to an unfeasible design where the stress constraints are not satisfied, i.e. g1 > 0

(Fig. 5.17(b)). The final structural mass of the feasible optimized structures are presented in

Table. 5.7.

It can be noticed that increasing the thermal load results in a heavier structure which is required

to satisfy the stress and stability constraints. By increasing the structural mass, the panel is able
74
1.3

1.2

1.1

1 = 160◦ C

m (kg)
∆T
∆T = 180◦ C
0.9 ∆T = 200◦ C
∆T = 220◦ C
0.8

0 5 10 15 20 25 30 35
Design iterations

(a) Convergence of the structural mass m

0.6

0.4

= 160◦ C
maxi {g1i }

0.2 ∆T
∆T = 180◦ C
∆T = 200◦ C
0 ∆T = 220◦ C

0 5 10 15 20 25 30 35
Design iterations

(b) Convergence of the maximum stress constraint

−5
∆T = 160◦ C
∆T = 180◦ C
−10
∆T = 200◦ C
max k {g2k }

∆T = 220◦ C
−15

−20
0 5 10 15 20 25 30 35
Design iterations

(c) Convergence of the maximum satbility constraint

Figure 5.17
Convergence plot of the objective function and constraints subjected to thermal loads of
∆T = {160, 180, 200, 220}◦ C for a hat-stiffened panel.

to sustain the thermal stresses. While in elasticity, increasing the thickness of a structure adds

extra stiffening without influencing the stresses, in thermoelasticity, increasing the thickness of the

75
Table 5.7
Structure mass of the feasible optimized design for ∆T = {120, 140, 160, 180, 200}◦ C.

Thermal load ∆T (◦ C) optimized mass m (kg)


120 0.6475
140 0.6666
160 0.7781
180 0.8410
200 0.9245

structure results in an increase in the stiffness and thermal stresses. This makes the design process

a nontrivial task. While increasing the structure thickness for ∆T = {160, 180, 200}◦ C allowed the

stress to stay under the maximum allowable stress σcrit , for ∆T = 220◦ C increasing the thickness of

the structure led to higher thermal stresses, and as a consequence violating the stress constraints.

The first frequency of the structure ω1 is plotted in Fig. 5.18 for ∆T = {160, 180, 200}◦ C.

4,000

3,000

2,000
ω1 (rad/s)

1,000

0
0 50 100 150 200
∆T (◦ C)
∆T = 160◦ C ∆T = 180◦ C
∆T = 200◦ C

Figure 5.18
The first natural frequency of the optimized design of a hat-stiffened panel obtained with thermal
loads of ∆T = {160, 180, 200}◦ C.

For all thermal loads shown in Fig. 5.18, the frequency decreases until it reaches its first

buckling/bifurcation point at approximately ∆T = 20◦ C. It then increase monotonically until it

76
reaches ∆T = 120◦ C, then the frequency decreases to its all time minimum. The combination of

backtracking and frequency constraints at the final temperature results in optimized structures that

are statically stable throughout the load-stepping process and at the final temperature. The panels

out-of-plane deformation for ∆T = {160, 180, 200}◦ C is shown in Fig. 5.19.

(a) ∆T = 160◦ C (b) ∆T = 180◦ C

(c) ∆T = 200◦ C

Figure 5.19
Out-of-plane deformation (m) of the final design of a hat-stiffened Inconel panel obtained with
thermal loads of ∆T = {160, 180, 200}◦ C.

It should be noted that for all temperatures, the out-of-plane deformation of the panel skin

occurs towards the stiffeners (in the negative z direction). The positive out-of-plane deformation

in the panel skin can be noticed between the three stiffeners, more specifically to the left and right

77
of the stiffeners extremities. By increasing the thermal load the positive out-of-plane deformation

becomes more pronounced in the whole region between stiffener 1 and 2 and between stiffener

2 and 3. Similarly, the out-of-plane deformation in the area above the three stiffeners becomes

more pronounced when increasing the thermal load. The stress distribution in the panel for

∆T = {160, 180, 200}◦ C is shown in Fig. 5.20.

(a) ∆T = 160◦ C (b) ∆T = 180◦ C

(c) ∆T = 200◦ C

Figure 5.20
Stress distibution (Pa) of the final design of a hat-stiffened Inconel panel obtained with thermal
loads of ∆T = {160, 180, 200}◦ C.

The stress distribution in the three structures show a maximum stress located at the boundary of

the panels and stiffeners. By increasing the thermal load, the stress in the panel increases. This is

78
caused by the higher out-of-plane deformation in the negative z direction in the area above the three

stiffeners and the higher out-of-plane positive deformation between stiffener 1 and 2 and between

stiffener 2 and 3. Nonetheless, the stress concentration at the panel skin boundary is what needs

to be given attention in order to reach feasibility for higher thermal loads. The design variables

distribution for ∆T = {160, 180, 200}◦ C is plotted in Fig. 5.21.

·10−3 ·10−4
7
1.4
6.5
1.2
6
1
Thickness (m)

Thickness (m)
5.5
0.8

0.6 5
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x/L x/L
∆T = 160◦ C ∆T = 180◦ C ∆T = 160◦ C ∆T = 180◦ C
∆T = 200◦ C ∆T = 200◦ C

(a) Panel skin thickness (b) Stiffener thickness (stiffener 1 and 3)

·10−4
7

6.5

6
Thickness (m)

5.5

5
0 0.2 0.4 0.6 0.8 1
x/L
∆T = 160◦ C ∆T = 180◦ C
∆T = 200◦ C

(c) Stiffener thickness (stiffener 2)

Figure 5.21
Distribution of the stiffeners thickness for hat-stiffened Inconel panels subjected to thermal loads
of ∆T = {160, 180, 200}◦ C.

79
The thickness distribution in the panel skin shown in Fig. 5.21(a) increases with increasing

thermal loads. The maximum thickness for each thermal load is located at the extremities x/L = 0

and x/L = 1. Similarly for stiffeners 1 and 3, the maximum thickness is located at the extremities

of each stiffener. For stiffener 2, the thickness reaches the lower bound for ∆T = 160◦ C at all the

stations. The thickness of stiffener 2 for ∆T = 200◦ C is similar to the distribution of stiffener 1

and 3 where the maximum thickness is located at the extremities. For ∆T = 180◦ C, stiffener 2

thickness distribution is slightly different from the other temperatures where a peak in the thickness

is noticed at X/L = 1/2.

5.3.2 Hat-stiffened panel with corrugation

In this section, we investigate the impact of corrugation on sizing optimization of hat-stiffened

panels. This work uses two values for skin corrugation η = 1.5mm and η = −1.5mm. In

the work by Lee and Bhatia [62], the skin corrugation was shown to have a big impact on the

structure frequency. By introducing corrugation, the structure bypassed the first bifurcation point

and delayed the second. This means that the frequency remains at the same order of magnitude

as was shown in Fig. 5.18 instead of decreasing to the first bifurcation point, and it also increase

the temperature at which the second bifurcation occurs. Two types of structures are considered,

a corrugated panel without a blending function and a corrugated panel with a blending function.

Fig. 5.22 shows a representation of the two corrugated panels.

The skin corrugation previously defined is multiplied with a blending function fbd in the x−axis

with the purpose of setting a start and end point for the corrugation instead of starting and ending at

80
(a) w/o blending function (b) w/ blending function

Figure 5.22
Illustration of hat-stiffened panels with and without the blending function.

the extremities of the panel skin. Similar to the previous section, we conduct sizing optimization of

thermally stressed hat-stiffened panels for a range of thermal loads ∆T = {160, 200, 220, 250}◦ C.

81
5.3.2.1 Corrugated panel without a blending function

Convergence of the objective function and constraints for a hat-stiffened panel with corrugation

η = 1.5mm is shown in Fig. 5.23. For the sake of simplicity and compactness, we choose to include

only convergence hatstiifenedresults/plots for η = 1.5mm.

Similar to the previous section, it can be noticed in Fig. 5.23 that the optimizer reaches

feasibility for ∆T = {160, 200}◦ C. For ∆T = 220◦ C, a feasible design cannot be reached since the

stress constraints are violated. It should be noted that the same set of results were obtained when

using a corrugation of η = −1.5mm. The optimized mass of the stiffened panels with corrugation

is show in Table. 5.8, where it is compared to the optimized mass of non-corrugated structures

given in Table.5.7.

Table 5.8
Structure mass of the feasible optimized design for ∆T = {120180, 200}◦ C.

Thermal load ∆T (◦ C) optimized mass m (kg)


w/o corrugation w/ corrugation η = 1.5mm
160 0.7781 1.0504
180 0.8410 1.1748
200 0.9245 1.3216

When comparing the both structural masses, it can be noticed that adding skin corrugation

requires a heavier structure in order to reach a feasible design. The question then is what is the

purpose of adding corrugation if we cannot obtain lighter structures or feasible panels at higher

operating temperature. The impact of corrugation is visible in the frequency behavior of the

structure. In Fig. 5.24, the first frequency of the structure is plotted versus temperature for a

corrugation of η = 1.5mm and η = −1.5mm.


82
1.6

1.4

∆T = 160◦ C

m (kg)
∆T = 200◦ C
1.2
∆T = 220◦ C

1
0 10 20 30
Design iterations

(a) Convergence of the structural mass m

∆T = 160◦ C
maxi {g1i }

0.5
∆T = 200◦ C
∆T = 220◦ C

0
0 10 20 30
Design iterations

(b) Convergence of the maximum stress constraint

−10 ∆T = 160◦ C
∆T = 200◦ C
max k {g2k }

∆T = 220◦ C

−20

0 10 20 30
Design iterations

(c) Convergence of the maximum stability constraint

Figure 5.23
Convergence plot of the structural mass and stress and stability constraints subjected to thermal
loads of ∆T = {160, 200, 220}◦ C for a stiffened panel with corrugation η = 1.5mm.

As discussed earlier, introducing skin corrugation in the structure allows to bypass the first

bifurcation point. This is apparent in Figs. 5.24(a) and 5.24(b), where the first bifurcation point

83
5,000
5,000
4,000
4,000

3,000 3,000
ω1 (rad/s)

ω1 (rad/s)
2,000 2,000

1,000 1,000

0 50 100 150 200 0 50 100 150 200


∆T (◦ C) ∆T (◦ C)
∆T = 160◦ C ∆T = 200◦ C ∆T = 160◦ C ∆T = 200◦ C

(a) Corrugation of η = 1.5mm (b) Corrugation of η = −1.5mm

Figure 5.24
The first natural frequency of the final design of a corrugated stiffened panel obtained with
thermal loads of ∆T = {160, 200}◦ C.

of the structure previously present in Fig. 5.18 at approximately ∆T = 20◦ C is absent. Moreover,

both figure show a monotonic behavior in frequency until the final temperature is reached. Such

behavior is important, but the primary reason the optimizer is not able to reach a feasible design

for ∆T = 220◦ C is the stress constraint. While adding skin corrugation allows the structure to have

better stability performance, it adds additional stresses in the structure which justify the heavier

structures for corrugated panels in Table. 5.8 . Fig. 5.25 shows the stress distribution in the optimized

panels with η = 1.5mm and η = −1.5mm subjected to thermal loads of ∆T = {160, 200}◦ C.

Similar to the previous section, the maximum stress is located at the boundary of the panel skin..

The stress concentration is located at x/L = {1/4, 3/4}, this can explained by the deformation of

the panel shown in Fig. 5.26.

For ∆T = 160◦ C, the deformation of the panel at x/L = 1/4 and x/L = 3/4 is alternating

between positive and negative out-of-plane deformation. By increasing the thermal load, the panel

deformation between stiffeners increases. This leads to higher stresses in the region bounded by

84
(a) ∆T = 160◦ C with corrugation η = 1.5mm (b) ∆T = 200◦ C with corrugation η = 1.5mm

(c) ∆T = 160◦ C with corrugation η = −1.5mm (d) ∆T = 200◦ C with corrugation η = −1.5mm

Figure 5.25
Stress distibution (Pa) of the final design of stiffened panels obtained for operating temperatures
∆T = {160, 200}◦ C with corrugations η = {1.5, −1.5}mm.

x/L = 1/4 and x/L = 3/4 with the highest stress values located at the extremities of that area.

Moreover, we can notice from Fig. 5.26 that the value of corrugation influences the deformation

direction of the panel. In Fig. 5.25(a) and 5.25(b), the positively corrugated panel deforms in a

positive out-of-plane direction. The region above the stiffeners deforms in the positive z direction.

On the other hand, the negatively corrugated panel deforms in the negative z direction as can be seen

in Figs. 5.25(c) and 5.25(d). This means that the direction of corrugation affects the direction of the

panel deformation. Furthermore, the maximum deformation magnitude is larger for the negatively

corrugated panel. Thus, when designing a corrugated stiffened panel, the corrugation direction and
85
(a) ∆T = 160◦ C with corrugation η = 1.5mm (b) ∆T = 200◦ C with corrugation η = 1.5mm

(c) ∆T = 160◦ C with corrugation η = −1.5mm (d) ∆T = 200◦ C with corrugation η = −1.5mm

Figure 5.26
Out-of-plane deformation (m) of the final design of stiffened panels obtained for operating
temperatures ∆T = {160, 200}◦ C with corrugations η = {1.5, −1.5}mm.

magnitude needs to be carefully chosen to achieve the structural properties we seek. The design

variables distribution is shown next in Fig. 5.27 for the positively and negatively corrugated panels.

Fig. 5.27 shows the thickness distribution of the panel skin in Figs. 5.27(a) and 5.27(d) for

η = 1.5mm and η = −1.5mm, respectively. For η = 1.5mm, the thickness distribution of

stiffeners 1 and 3 is shown in Fig. 5.27(b), and shown in Fig. 5.27(c) for stiffener 2. Similarly, for

η = −1.5mm, the thickness distribution of stiffeners 1 and 3 is shown in Fig. 5.27(e), and shown

in Fig. 5.27(f) for stiffener 2. We can notice from Figs. 5.27(a) and 5.27(d) that the panel skin

thickness distribution is similar for both η = 1.5mm and η = −1.5mm. The highest thickness

86
is located at the edges x = 0 and x = L. Increasing the thermal load requires higher thickness

to compensate the additional stresses. For η = 1.5mm, the thickness distribution in stiffeners 1

and 3, show a spike at x/L = 1/2 for ∆T = 200◦ C. In stiffener 2, the lower bound is reached

for ∆T = 160◦ C, and a U shaped distribution is obtained for ∆T = 200◦ C where the maximum

thickness is located at the extremities. For η = −1.5mm, the spike at x/L = 1/2 occurs in stiffener

2 and the U shaped distribution is obtained for stiffeners 1 and 3.

87
·10−3 ·10−4
5

4 5.3

3 5.2
Thickness (m)

Thickness (m)
2
5.1

1
5
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x/L x/L
∆T = 160◦ C ∆T = 200◦ C ∆T = 160◦ C ∆T = 200◦ C

(a) Panel skin with η = 1.5mm (b) Stiffener 1 and 3 with η = 1.5mm

·10−4 ·10−3
5
7.5

7 4

6.5 3

6
Thickness (m)

Thickness (m)
2
5.5
1
5
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x/L x/L
∆T = 160◦ C ∆T = 200◦ C ∆T = 160◦ C ∆T = 200◦ C

(c) Stiffener 2 with η = 1.5mm (d) Panel skin with η = −1.5mm

·10−4 ·10−4
9
6.5
8

6
7
Thickness (m)

Thickness (m)

5.5 6

5 5
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x/L x/L
∆T = 160◦ C ∆T = 200◦ C ∆T = 160◦ C ∆T = 200◦ C

(e) Stiffener 1 and 3 with η = −1.5mm (f) Stiffener 2 with η = −1.5mm

Figure 5.27
Design variables distribution for thermal loads of ∆T = {160, 200}◦ C and corrugation
η = {1.5, −1.5}mm.

88
5.3.2.2 Impact of the blending function

Following the results obtained in the previous section, we are investigating the impact of the

corrugation positioning on the stresses in the structure. The skin corrugation previously defined is

multiplied with a blending function fbd (x), with the purpose of choosing a starting and end point

for the corrugation. The blending function fbd is a quintic Hermite interpolation function that is

based on the solution, first and second derivative at two given points. For the sake of simplicity, we
df (x) d 2 f (x)
will refer to the first and second derivative dx and dx as f 0(x) and f 00(x), respectively. The

general form of the blending function fbd (x) can be expressed as

1 00
fbd (x) = f (x0 ) + f 0(x0 )(x − x0 ) + f (x0 )(x − x0 )2 + (5.4)
2
f (x1 ) − f (x0 ) − f 0(x0 )(x1 − x0 ) − 12 f 00(x0 )(x1 − x0 )2
3
(x − x0 )3 +
(x1 − x0 )
3 f (x0 ) − 3 f (x1 ) + (2 f 0(x0 ) + f 0(x1 ))(x1 − x0 ) + 21 f 00(x0 )(x1 − x0 )2
(x − x0 )3 (x − x1 ) +
(x1 − x0 )4
6 f (x1 ) − 6 f (x0 ) − 3( f 0(x0 ) + f 0(x1 ))(x1 − x0 ) + 21 ( f 00(x1 ) − f 00(x0 ))(x1 − x0 )2
(x − x0 )3 (x − x1 )2
(x1 − x0 )5

The blending function allows the corrugation to start at point x0 that we choose, and reach its

maximum corrugation value η at chosen point x1 . Similarly, towards the end of the panel skin in

the x−direction, the blending function allows the corrugation in the panel skin to go from a value

89
of η at x2 to a value of zero at x3 . The blending function fbd (x) in the range x ∈ [0, L] can be

expressed as

fbd (x) = 0 x < x1









+
fbd (x) = fbd

(x) x0 ≤ x ≤ x1







(5.5)


 fbd (x) = 1 x1 < x < x2




 fbd (x) = fbd




 (x) x2 ≤ x ≤ x3




 fbd (x) = 0 x > x3




where the solution, first and second derivative used to define the blending function are expressed

as

+

(x0 ) = 0




 fbd



+

 fbd (x1 ) = 1



(5.6)

 0+ 0+
 fbd (x0 ) = fbd (x1 ) = 0







 00+ 00+
 fbd (x0 ) = fbd (x1 ) = 0





(x2 ) = 1




 fbd




 fbd (x3 ) = 0


 −

and (5.7)

 0− 0−
fbd (x0 ) = fbd (x1 ) = 0








 00− 00−
 fbd (x0 ) = fbd (x1 ) = 0



+ (x) and f − (x) are plotted in Figs. 5.28(a) and 5.28(b), respectively.
The functions fbd bd

Next, we conduct a sizing optimization study for a corrugated hat-stiffened panels using the

blending function fbd (x). Similar to earlier sections, we subject the structure to multiple thermal

loads. Both η = 1.5mm and η = −1.5mm are considered. The following results are using x0 = 0
90
1 1

0.8 0.8

0.6 0.6
fbd (x)

fbd (x)
0.4 0.4

0.2 0.2

0 0
0 2 · 10−2 4 · 10−2 6 · 10−2 8 · 10−2 0.1 0.9 0.92 0.94 0.96 0.98 1
x/L x/L
+ (x)
(a) fbd (x) = fbd (b) fbd (x) = fbd
− (x)

Figure 5.28
Blending function fbd in the range x/L ∈ [0, 0.1] and x/L ∈ [0.9, 1].

and x1 = 10%L. For symmetry, we use x2 = 90%L and x3 = L. Convergence of the objective

function and constraints are shown in Fig. 5.29 for η = 1.5mm.

It can be noticed from Fig. 5.29 that for ∆T = {200, 220}◦ C, the optimizer is able to reach a

feasible design where the stress and stability constraints are both satisfied. For ∆T = 250◦ C, the

optimizer is not able to reach a feasible design where the stress constraints are violated. Now,

comparing the obtained results with the previous section, the addition of the blending function

results in a reduction in the stresses and leads to a feasible design for ∆T = 200◦ C. Although

convergence hatstiifenedresults/plots for η = −1.5mm are not included, it should be noted that

the same results were obtained for the negatively corrugated panel. The first frequency is plotted

in Fig. 5.30 for the corrugated structures with η = 1.5mm and η = −1.5mm using the blending

function fbd (x).

It can be noticed that the addition of the blending function does not impact the frequency

behavior. For both η = 1.5mm and η = −1.5mm the first bifurcation point is still bypassed and the

91
1.4

1.3

∆T = 200◦ C

m (kg)
1.2 ∆T = 220◦ C
∆T = 250◦ C
1.1

0 10 20 30
Design iterations

(a) Convergence of the structural mass m

0.6

0.4
∆T = 200◦ C
maxi {g1i }

0.2 ∆T = 220◦ C
∆T = 250◦ C

0
0 10 20 30
Design iterations

(b) Convergence of the maximum stress constraint

−5
∆T = 200◦ C
−10 ∆T = 220◦ C
max k {g2k }

∆T = 250◦ C
−15

−20

0 10 20 30
Design iterations

(c) Convergence of the maximum stability constraint

Figure 5.29
Convergence plot of the objective and constraints for thermal loads ∆T = {200, 220, 250}◦ C for a
corrugated stiffened panel (η = 1.5mm) with x0 = 0 and x1 = 0.1.

frequency monotonically decreases until it reaches the vicinity of the final temperature. The panel

deformation for ∆T = {200, 220}◦ C is shown in Fig. 5.31.

92
6,000 6,000

4,000 4,000
ω1 (rad/s)

ω1 (rad/s)
2,000 2,000

0 0
0 50 100 150 200 0 50 100 150 200
∆T (◦ C) ∆T (◦ C)
∆T = 200◦ C ∆T = 220◦ C ∆T = 200◦ C ∆T = 220◦ C

(a) Corrugation of η = 1.5mm (b) Corrugation of η = −1.5mm

Figure 5.30
The first natural frequency of the final design of corrugated stiffened panels (η = {1.5, −1.5}mm)
obtained with thermal loads of ∆T = {200, 220}◦ C using a blending function with x0 = 0 and
x1 = 0.1.

It can be noticed from Fig. 5.31 that the addition of the blending function introduces an

oscillatory behavior in the panel deformation, as can be seen in Fig. 5.31(a) and 5.31(c). By

increasing the thermal load, the oscillation in the structure seems to reduce (Figs. 5.31(b) and

5.31(d)). The corrugation direction still dictates the direction of the out-of-plane deformation

where the positively corrugated panel deforms in the positive z direction and vice versa. In addition

to the oscillations, the out-of-plane deformation between stiffeners alternates between positive and

negative deformation. Fig. 5.32 shows the stress distribution in the corrugated hat-stiffened panels

with the blending function.

The addition of the blending function resulted in a reduction of the stresses in the structures.

The oscillation as well as the alternative deformation between stiffeners mentioned earlier results

in a scattered stress as can be seen in Fig. 5.32.

93
(a) ∆T = 200◦ C with corrugation η = 1.5mm (b) ∆T = 220◦ C with corrugation η = 1.5mm

(c) ∆T = 200◦ C with corrugation η = −1.5mm (d) ∆T = 220◦ C with corrugation η = −1.5mm

Figure 5.31
out-of-plane deformation (m) of the final design of corrugated stiffened panels
(η = {1.5, −1.5}mm) obtained with thermal loads of ∆T = {200, 220◦ }C using a blending
function with x0 = 0 and x1 = 0.1.

94
(a) ∆T = 200◦ C with corrugation η = 1.5mm (b) ∆T = 220◦ C with corrugation η = 1.5mm

(c) ∆T = 200◦ C with corrugation η = −1.5mm (d) ∆T = 220◦ C with corrugation η = −1.5mm

Figure 5.32
Stress distibution (Pa) of the final design of corrugated stiffened panels (η = {1.5, −1.5}mm)
obtained with thermal loads of ∆T = {200, 220}◦ C using a blending function with x0 = 0 and
x1 = 0.1.

95
CHAPTER VI

CONCLUSION

The objective of this research was to alleviate the challenging problem of heat-structure inter-

action rising in supersonic and hypersonic regimes. Our approach consisted of multiple steps to

achieve that goal mainly the development of a structural sizing methodology for post-buckled ther-

mally stressed stiffened panels, incorporation of geometric nonlinearities in the structural model

and the derivation and implementation of stability constraints.

The geometric nonlinearity was shown to be an important aspect of sizing optimization and was

needed for an accurate modeling of the structural behavior. When the assumption of linear strain

was used, the results showed an optimized structure which violated stress and stability constraints.

The modeling of geometric nonlinearity added extra complexity to the thermoelasticity problem

and required using the pseudo arc-length solver. Although the path-tracking solver needed a

parametric study to determine the appropriate values for the tuning parameters, the solver was

adapting properly with the difficulty of the problem and it was shown to converge in a reasonable

number of steps.

Three formulations for the stability constraints were derived where the first approach used a

backtracking algorithm, the second used aggregation and last one used an extrapolation technique

96
based on the final temperature. It was concluded that the first approach led to a statically stable

structure without the added computational cost attached with the second and third approach.

Conclusions drawn from the optimization of the stiffened panels are presented next.

6.0.1 Optimization of the blade-stiffened panel

In this optimization study, the design variables included the panel skin thickness along with

the height and width of the stiffeners. The panel weight reduction reached 50% compared to the

initial structure for ∆T = 120◦C and was shown to be primarily driven by the stability constraints.

The panels were not able to sustain thermal loads greater than 160◦ C, where a load of ∆T = 180◦ C

caused the stress constraint to be violated. This phenomenon is due to thermoelastic effects on the

structure where increasing the thickness of the panel increases the stiffness as well as the thermal

stresses in the structure.

The thickness distribution of the optimized panels was shown to be higher at both edges x/L = 0

and x/L = 1, which can be explained by the high concentration of stresses due to the bending and

extension coupling in those areas.

The change in the natural frequency versus temperature in a load stepping process was shown

to be highly non-monotonic. This confirms the need for a robust solver able to withstand these

narrow zones of nonlinearity and requires a properly implemented stability constraints to prevent

reaching negative eigenvalues.

6.0.2 Optimization of the hat-stiffened panel

For the model used in this study, sizing optimization of the non-corrugated panel led to a feasible

design for up to ∆T = 200◦ C, where larger operating temperature violated the stress constraints.
97
The stress distribution showed that the maximum stress is concentrated at the boundaries of the

panel skin and stiffeners. This was due to the thermoelasticity effects and the in-plane restraint

imposed through the boundary conditions.

The corrugation of the hat stiffeners allowed by-passing of the first buckling point which

enhanced the stability properties of the stiffened panel and resulted in less nonlinearity in the load-

deformation curve. On the other hand, the corrugation of the hat stiffeners led to an increase in

the stresses at the boundary compared to non-corrugated panels. The introduction of the blending

function to control the start and end point of the stiffeners corrugation reduced the stresses at the

boundary and created a more scattered distribution of the stresses.

6.1 Contributions

A generalized finite element tool was developed for structural sizing of thermally stressed

stiffened panels with application to blade and hat stiffeners. The code provides sensitivity of the

thermoelastic problem with geometric nonlinearities included. It is the first code to provide a

completely integrated sensitivity analysis for the nonlinear thermoelastic problem using analytic

derivations. Robust and reliable stability constraint were derived and imposed for this highly

nonlinear problem. Corrugation aspect of the hat-stiffened panels was investigated in the context

of sizing optimization.

6.2 Future work


6.2.1 Applications

One way to extend this work is to use the corrugation magnitude and direction as a parameter

for shape optimization. This will allow the proper choice of corrugation for an optimum weight

98
of the structure. Another path to extend this work would be to focus on the blending function and

introduce it as a parameter for shape optimization. This would allow the optimizer to find the most

efficient location for corrugation in order to minimize stresses.

6.2.2 Disciplinary Extensions

Extra fidelity to this work can be added by implementing a coupled heat-transfer and thermoe-

lastic problem. The material nonlinearities could be incorporated to get an idea about the impact

of temperature on the stiffening of the panel when conducting sizing optimization studies. In the

supersonic regime, we can be study the impact of the surrounding fluid on the optimization of

the stiffened panels. A coupled fluid-structure interaction problem can be created where the CFD

framework Loci/CHEM can be linked with the developed methodology in order to incorporate the

effects of the fluid on the optimized structures.

6.2.3 Computational Issues

For future work, the effect of stress aggregation and the use of adjoint sensitivity can be

explored. As consequence, the constraint functions will be importantly reduced, thus, high savings

in computational cost is expected. The continuation solver was shown to be serial which defeats

the purpose of using high-performance computing and MPI parallel programming. Solutions to

parallelize this computational solver should be studied, since it will lead to an important saving in

computational cost for structural sizing of nonlinear problems.

99
REFERENCES

[1] Mast. URL https://github.com/MASTmultiphysics/mast-multiphysics.

[2] Mumps users manual. Technical Report Version 5.0, 2015. URL
http://mumps.enseeiht.fr.

[3] Gong A., Quinn R., and Ko W. Reentry heating analysis of space shuttle with comparison of
flight data. In Computational Aspects of Heat Transfer in Structures, NASA CP-2216, page
271–294. 1982.

[4] Patrick R Amestoy, Iain S Duff, and J-Y L’Excellent. Multifrontal parallel distributed sym-
metric and unsymmetric solvers. Computer methods in applied mechanics and engineering,
184(2):501–520, 2000.

[5] Patrick R Amestoy, Iain S Duff, Jean-Yves L’Excellent, and Jacko Koster. A fully asyn-
chronous multifrontal solver using distributed dynamic scheduling. SIAM Journal on Matrix
Analysis and Applications, 23(1):15–41, 2001.

[6] Gaetano Arena, Rainer MJ Groh, Alex Brinkmeyer, Raf Theunissen, Paul M. Weaver, and
Alberto Pirrera. Adaptive compliant structures for flow regulation. Proceedings of the Royal
Society A: Mathematical, Physical and Engineering Sciences, 473(2204):20170334, 2017.

[7] Andres F Arrieta, Onur Bilgen, Michael I Friswell, and Paolo Ermanni. Modelling and
configuration control of wing-shaped bi-stable piezoelectric composites under aerodynamic
loads. Aerospace Science and Technology, 29(1):453–461, 2013.

[8] Walid Arsalane and Manav Bhatia. Formulation of stability constraints in structural sizing
of thermally stressed stiffened panels. In AIAA Sci-tech 2022 FORUM (submitted).

[9] Walid Arsalane and Manav Bhatia. Structural sizing of post-buckled ther-
mally stressed stiffened panel. 2020. doi: 10.2514/6.2020-3180. URL
https://arc.aiaa.org/doi/abs/10.2514/6.2020-3180.

[10] Walid Arsalane and Manav Bhatia. Structural sizing of thermally stressed hat-stiffened
panels. In AIAA AVIATION 2021 FORUM, page 3092, 2021.

[11] Satish Balay, William D. Gropp, Lois Curfman McInnes, and Barry F. Smith. Efficient
management of parallelism in object oriented numerical software libraries. In E. Arge, A. M.
Bruaset, and H. P. Langtangen, editors, Modern Software Tools in Scientific Computing,
pages 163–202. Birkhäuser Press, 1997.

100
[12] Satish Balay, Shrirang Abhyankar, Mark F. Adams, Jed Brown, Peter Brune, Kris Buschel-
man, Lisandro Dalcin, Victor Eijkhout, William D. Gropp, Dinesh Kaushik, Matthew G.
Knepley, Lois Curfman McInnes, Karl Rupp, Barry F. Smith, Stefano Zampini, Hong
Zhang, and Hong Zhang. PETSc users manual. Technical Report ANL-95/11 - Revision
3.7, Argonne National Laboratory, 2016. URL http://www.mcs.anl.gov/petsc.

[13] Martin Philip Bendsoe and Ole Sigmund. Topology optimization: theory, methods, and
applications. Springer Science & Business Media, 2013.

[14] Katia Bertoldi, Pedro M Reis, Stephen Willshaw, and Tom Mullin. Negative poisson’s ratio
behavior induced by an elastic instability. Advanced materials, 22(3):361–366, 2010.

[15] M Bhatia and P Beran. Design of thermally stressed panels subject to transonic flutter
constraints. Journal of Aircraft, 54(6):2340–2349, 2017.

[16] Manav Bhatia. Design-oriented thermoelastic analysis, sensitivities, and approximations


for shape optimization of aerospace vehicles. 2007.

[17] Manav Bhatia and Philip S Beran. Mast: An open-source computational framework for
design of multiphysics systems. In 2018 AIAA/ASCE/AHS/ASC Structures, Structural Dy-
namics, and Materials Conference, page 1650, 2018.

[18] Manav Bhatia and Eli Livne. Design-Oriented Thermostructural Analysis with External and
Internal Radiation, Part 1: Steady State. AIAA Journal, 46(3):578–590, March 2008.

[19] Manav Bhatia and Eli Livne. Design-Oriented Thermostructural Analysis with External
and Internal Radiation, Part 2: Transient Response. AIAA Journal, 47(5):1228–1240, May
2009.

[20] Manav Bhatia, David John Neiferd, and Josh Deaton. Robust load-stepping for nonlinear
thermoelastic analysis of thin-walled structures, submitted. 2019.

[21] HM Bucker and George F Corliss. A bibliography of automatic differentiation. LECTURE


NOTES IN COMPUTATIONAL SCIENCE AND ENGINEERING, 50:321, 2006.

[22] Sebastien Candel. Concorde and the future of supersonic transport. Journal of propulsion
and power, 20(1):59–68, 2004.

[23] Nian-Zhong Chen and C Guedes Soares. Progressive failure analysis for prediction of post-
buckling compressive strength of laminated composite plates and stiffened panels. Journal
of reinforced plastics and composites, 26(10):1021–1042, 2007.

[24] I. Soner Cinoglu, Matthew R. Begley, Joshua D. Deaton, Philip S. Beran, Robert M.
McMeeking, and Natasha Vermaak. Elastoplastic design of beam structures subjected to
cyclic thermomechanical loads. Thin-Walled Structures, 136:175–185, 2019.

[25] Special Materials Corporation. Inconel properties. 2007. URL


.
http://www.specialmetals.com/assets/smc/documents/inconel_alloy_718.pdf

101
[26] BS Cox, RMJ Groh, Daniele Avitabile, and A Pirrera. Modal nudging in nonlinear elasticity:
Tailoring the elastic post-buckling behaviour of engineering structures. Journal of the
Mechanics and Physics of Solids, 116:135–149, 2018.

[27] Paul F Crickmore. Lockheed SR-71 Blackbird. Bloomsbury Publishing, 2015.

[28] Michael A Crisfield. A fast incremental/iterative solution procedure that handles “snap-
through”. In Computational methods in nonlinear structural and solid mechanics, pages
55–62. Elsevier, 1981.

[29] J D Deaton and R V Grandhi. Significance of geometric nonlinearity in the design of


thermally loaded structures. Journal of Aircraft, 52(4):1226–1234, 2014.

[30] Joshua D Deaton. Design of Thermal Structures using Topology Optimization. PhD thesis,
Wright State University, Dayton, OH, 2009.

[31] Joshua D Deaton and Ramana V Grandhi. Stiffening of restrained thermal structures
via topology optimization. Structural and Multidisciplinary Optimization, 48(4):731–745,
2013.

[32] Cezar G Diaconu, Paul M Weaver, and Filippo Mattioni. Concepts for morphing airfoil
sections using bi-stable laminated composite structures. Thin-Walled Structures, 46(6):
689–701, 2008.

[33] Samir A Emam and Daniel J Inman. A review on bistable composite laminates for morphing
and energy harvesting. Applied Mechanics Reviews, 67(6), 2015.

[34] Bastiaan Florijn, Corentin Coulais, and Martin van Hecke. Programmable mechanical
metamaterials. Physical review letters, 113(17):175503, 2014.

[35] F Gnani, H Zare-Behtash, and K Kontis. Pseudo-shock waves and their interactions in
high-speed intakes. Progress in Aerospace Sciences, 82:36–56, 2016.

[36] W. J. F. Govaerts. Numerical Methods for Bifurcations of Dynamical Equilibria. SIAM,


Philadelphia, Pennsylvania, 2000.

[37] Willy JF Govaerts. Numerical methods for bifurcations of dynamical equilibria. SIAM,
2000.

[38] Richard H. Graham. SR-71 Revealed: The Inside Story. MBI Publishing Company, St. Paul,
Minnesota, 1996. ISBN 978-0-7603-0122-7.

[39] Ramana Grandhi. Structural optimization with frequency constraints-a review. AIAA journal,
31(12):2296–2303, 1993.

[40] RV Grandhi, G Bharatram, and VB Venkayya. Optimum design of plate structures with
multiple frequency constraints. Structural optimization, 5(1-2):100–107, 1992.

102
[41] Mei-Wen Guo, Issam E Harik, and Wei-Xin Ren. Buckling behavior of stiffened laminated
plates. International journal of solids and structures, 39(11):3039–3055, 2002.

[42] Raphael T Haftka and Zafer Gürdal. Elements of structural optimization, volume 11.
Springer Science & Business Media, 2012.

[43] P.S. Harvey and L.N. Virgin. Coexisting equilibria and stability of a shallow arch: Uni-
lateral displacement-control experiments and theory. International Journal of Solids and
Structures, 54:1–11, 2015.

[44] V. Hernandez, J. E. Roman, and V. Vidal. SLEPc: Scalable Library for Eigenvalue Problem
Computations. Lect. Notes Comput. Sci., 2565:377–391, 2003.

[45] Vicente Hernandez, Jose E. Roman, and Vicente Vidal. SLEPc: A scalable and flexible
toolkit for the solution of eigenvalue problems. ACM Trans. Math. Software, 31(3):351–362,
2005.

[46] SZ Hu and L Jiang. A finite element simulation of the test procedure of stiffened panels.
Marine structures, 11(3):75–99, 1998.

[47] Giles W Hunt. Imperfection-sensitivity of semi-symmetric branching. Proceedings of


the Royal Society of London. A. Mathematical and Physical Sciences, 357(1689):193–211,
1977.

[48] Jocelyn Iott, Raphael T Haftka, and Howard M Adelman. Selecting step sizes in sensitivity
analysis by finite differences, volume 86382. National Aeronautics and Space Administra-
tion, Scientific and Technical . . . , 1985.

[49] Hemendra Kumar Jain and Akhil Upadhyay. Buckling behavior of blade-, angle-, t-, and
hat-stiffened frp panels subjected to in-plane shear. Journal of reinforced plastics and
composites, 29(24):3614–3623, 2010.

[50] Dennis R Jenkins. Hypersonic: the story of the North American X-15. Specialty Press,
2008.

[51] Th Von Karman and H-S Tsien. The buckling of spherical shells by external pressure.
Journal of the Aeronautical Sciences, 7(2):43–50, 1939.

[52] E Kebadze, SD Guest, and S Pellegrino. Bistable prestressed shell structures. International
Journal of Solids and Structures, 41(11-12):2801–2820, 2004.

[53] NS Khot. Nonlinear analysis of optimized structure with constraints on systemstability.


AIAA Journal, 21(8):1181–1186, 1983.

[54] Tae Soo Kim and Yoon Young Kim. Mac-based mode-tracking in structural topology
optimization. Computers & Structures, 74(3):375–383, 2000.

103
[55] B. S. Kirk, J. W. Peterson, R. H. Stogner, and G. F. Carey. libMesh:
A C++ Library for Parallel Adaptive Mesh Refinement/Coarsening Sim-
ulations. Engineering with Computers, 22(3–4):237–254, 2006. URL
http://dx.doi.org/10.1007/s00366-006-0049-3.

[56] W. Ko, Quinn R., and Gong A. Reentry heat transfer analysis of the space shuttle orbiter.
In Computational Aspects of Heat Transfer in Structures, NASA CP-2216, page 295–326.
1982.

[57] Vadim Komkov, Kyung K Choi, and Edward J Haug. Design sensitivity analysis of structural
systems, volume 177. Academic press, 1986.

[58] G Kreisselmeier and R Steinhauser. Systematic control design by optimizing a vector


performance index. In Computer aided design of control systems, pages 113–117. Elsevier,
1980.

[59] Gerhard Kreisselmeier and Reinhold Steinhauser. Application of vector performance op-
timization to a robust control loop design for a fighter aircraft. International Journal of
Control, 37(2):251–284, 1983.

[60] Izabela K Kuder, Andres F Arrieta, Wolfram E Raither, and Paolo Ermanni. Variable
stiffness material and structural concepts for morphing applications. Progress in Aerospace
Sciences, 63:33–55, 2013.

[61] Mark K Leader, Ting Wei Chin, and Graeme J Kennedy. High-resolution topology opti-
mization with stress and natural frequency constraints. AIAA Journal, 57(8):3562–3578,
2019.

[62] Juhyeong Lee and Manav Bhatia. Impact of corrugations on bifurcation and thermoelastic
responses of hat-stiffened panels. Thin-Walled Structures, 140:209 – 221, 2019. ISSN
0263-8231. doi: https://doi.org/10.1016/j.tws.2019.03.027.

[63] C Lynch, A Murphy, M Price, and A Gibson. The computational post buckling analysis of
fuselage stiffened panels loaded in compression. Thin-walled structures, 42(10):1445–1464,
2004.

[64] Zheng-Dong Ma, Noboru Kikuchi, and Hsien-Chie Cheng. Topological design for vibrating
structures. Computer methods in applied mechanics and engineering, 121(1-4):259–280,
1995.

[65] Y Maeda, S Nishiwaki, K Izui, M Yoshimura, K Matsui, and K Terada. Structural topology
optimization of vibrating structures with specified eigenfrequencies and eigenmode shapes.
International Journal for Numerical Methods in Engineering, 67(5):597–628, 2006.

[66] Joaquim RRA Martins, Ilan Kroo, and Juan Alonso. An automated method for sensitivity
analysis using complex variables. In 38th Aerospace Sciences Meeting and Exhibit, page
689, 2000.
104
[67] F Mattioni, Paul M Weaver, and MI Friswell. Multistable composite plates with piecewise
variation of lay-up in the planform. International Journal of Solids and Structures, 46(1):
151–164, 2009.
[68] Peter Merlin. Design and Development of the Blackbird: Challenges and Lessons Learned.
2009. doi: 10.2514/6.2009-1522.
[69] Behrang Moghaddasie and Ilinca Stanciulescu. Equilibria and stability boundaries of shallow
arches under static loading in a thermal environment. International Journal of Non-Linear
Mechanics, 51:132–144, 2013.
[70] Tom Mullin, Stéphanie Deschanel, Katia Bertoldi, and Mary C Boyce. Pattern transformation
triggered by deformation. Physical review letters, 99(8):084301, 2007.
[71] Adrian Murphy, Mark Price, and A Gibson. Toward virtual testing of airframe stiffened
panels. In Royal Aeronautical Society, Structures and Materials Group Conference on
Virtual Testing, 2006.
[72] Neel Nadkarni, Chiara Daraio, and Dennis M Kochmann. Dynamics of periodic mechanical
structures containing bistable elastic elements: From elastic to solitary wave propagation.
Physical Review E, 90(2):023204, 2014.
[73] David J. Neiferd, Ramana V. Grandhi, Joshua D. Deaton, Philip S. Beran, and
Manav Bhatia. A Nonlinear Finite Element Analysis Capability for the Opti-
mization of Thermoelastic Structures. 2017. doi: 10.2514/6.2017-1302. URL
https://arc.aiaa.org/doi/abs/10.2514/6.2017-1302.
[74] Xin Ning and Sergio Pellegrino. Imperfection-insensitive axially loaded thin cylindrical
shells. International Journal of Solids and Structures, 62:39–51, 2015.
[75] M Ohsaki, Katsuki Fujisawa, N Katoh, and Y Kanno. Semi-definite programming for
topology optimization of trusses under multiple eigenvalue constraints. Computer Methods
in Applied Mechanics and Engineering, 180(1-2):203–217, 1999.
[76] Johannes TB Overvelde, Tamara Kloek, Jonas JA D’haen, and Katia Bertoldi. Amplifying
the response of soft actuators by harnessing snap-through instabilities. Proceedings of the
National Academy of Sciences, 112(35):10863–10868, 2015.
[77] Johannes Tesse Bastiaan Overvelde, Sicong Shan, and Katia Bertoldi. Compaction through
buckling in 2d periodic, soft and porous structures: effect of pore shape. Advanced Materials,
24(17):2337–2342, 2012.
[78] John Paterson. Overview of low observable technology and its effects on combat aircraft
survivability. Journal of Aircraft, 36(2):380–388, 1999. doi: 10.2514/2.2468.
[79] SN Patnaik and Manoranjan Maiti. Optimum design of stiffened structures with constraint
on the frequency in the presence of initial stresses. Computer Methods in Applied Mechanics
and Engineering, 7(3):303–322, 1976.
105
[80] Steven C Peters and Karl Iagnemma. Stability measurement of high-speed vehicles. Vehicle
System Dynamics, 47(6):701–720, 2009.

[81] P Pevzner, H Abramovich, and T Weller. Calculation of the collapse load of an axially
compressed laminated composite stringer-stiffened curved panel–an engineering approach.
Composite Structures, 83(4):341–353, 2008.

[82] Yong-Lin Pi and Mark Andrew Bradford. Multiple unstable equilibrium branches and non-
linear dynamic buckling of shallow arches. International Journal of Non-Linear Mechanics,
60:33–45, 2014.

[83] Damian Quinn, A Murphy, W McEwan, and F Lemaitre. Stiffened panel stability behaviour
and performance gains with plate prismatic sub-stiffening. Thin-Walled Structures, 47(12):
1457–1468, 2009.

[84] Ahmad Rafsanjani, Abdolhamid Akbarzadeh, and Damiano Pasini. Metamaterials: Snap-
ping mechanical metamaterials under tension (adv. mater. 39/2015). Advanced Materials,
27(39):5930–5930, 2015.

[85] GH Rahimi, M Zandi, and SF Rasouli. Analysis of the effect of stiffener profile on buckling
strength in composite isogrid stiffened shell under axial loading. Aerospace science and
technology, 24(1):198–203, 2013.

[86] Pedro M Reis. A perspective on the revival of structural (in) stability with novel opportunities
for function: from buckliphobia to buckliphilia. Journal of Applied Mechanics, 82(11), 2015.

[87] PM Reis. A perspective on the revival of structural (in) stability with novel opportunities
for function: from buckliphobia to buckliphilia. Journal of Applied Mechanics, 2015.

[88] E Riks. An incremental approach to the solution of snapping and buckling problems.
International journal of solids and structures, 15(7):529–551, 1979.

[89] J. E. Roman, C. Campos, E. Romero, and A. Tomas. SLEPc users manual. Technical Report
DSIC-II/24/02 - Revision 3.7, D. Sistemes Informàtics i Computació, Universitat Politècnica
de València, 2016.

[90] Rüdiger Seydel. Practical bifurcation and stability analysis, volume 5. Springer Science &
Business Media, 2009. ISBN 978-1-4419-1740-9.

[91] Rüdiger Seydel. Practical bifurcation and stability analysis, volume 5. Springer Science &
Business Media, 2009.

[92] Alexander P Seyranian, Erik Lund, and Niels Olhoff. Multiple eigenvalues in structural
optimization problems. Structural optimization, 8(4):207–227, 1994.

[93] Sicong Shan, Sung H Kang, Jordan R Raney, Pai Wang, Lichen Fang, Francisco Candido,
Jennifer A Lewis, and Katia Bertoldi. Multistable architected materials for trapping elastic
strain energy. Advanced Materials, 27(29):4296–4301, 2015.
106
[94] Jongmin Shim, Claude Perdigou, Elizabeth R Chen, Katia Bertoldi, and Pedro M Reis.
Buckling-induced encapsulation of structured elastic shells under pressure. Proceedings of
the National Academy of Sciences, 109(16):5978–5983, 2012.

[95] Ilinca Stanciulescu, Toby Mitchell, Yenny Chandra, Thomas Eason, and Michael
Spottswood. A lower bound on snap-through instability of curved beams under thermo-
mechanical loads. International Journal of Non-Linear Mechanics, 47(5):561–575, 2012.

[96] K Svanberg. A class of globally convergent optimization methods based on conservative


convex separable approximations. SIAM journal on optimization, 12(2):555–573, 2002.

[97] Krister Svanberg. The method of moving asymptotes- a new method for structural op-
timization. International journal for numerical methods in engineering, 24(2):359–373,
1987.

[98] Krister Svanberg. Mma and gcmma, versions september 2007. Optimization and Systems
Theory, 104, 2007.

[99] Denis Terwagne, Miha Brojan, and Pedro M Reis. Smart morphable surfaces for aerody-
namic drag control. Advanced materials, 26(38):6608–6611, 2014.

[100] E. A. Thornton. Thermal Structures for Aerospace Applications. AIAA Education Series,
AIAA, Reston, VA, 1996.

[101] André J Torii and Jairo R De Faria. Structural optimization considering smallest magnitude
eigenvalues: a smooth approximation. Journal of the Brazilian Society of Mechanical
Sciences and Engineering, 39(5):1745–1754, 2017.

[102] André J Torii, Rafael H Lopez, and Leandro FF Miguel. Modeling of global and lo-
cal stability in optimization of truss-like structures using frame elements. Structural and
Multidisciplinary Optimization, 51(6):1187–1198, 2015.

[103] Hsue-Shen Tsien. A theory for the buckling of thin shells. Journal of the Aeronautical
Sciences, 9(10):373–384, 1942.

[104] Dao Van Dung et al. Research on nonlinear torsional buckling and post-buckling of ec-
centrically stiffened functionally graded thin circular cylindrical shells. Composites Part B:
Engineering, 51:300–309, 2013.

[105] LN Virgin, R Wiebe, SM Spottswood, and TG Eason. Sensitivity in the structural behavior
of shallow arches. International Journal of Non-Linear Mechanics, 58:212–221, 2014.

[106] Ryan N Vogel. Structural-Acoustic Analysis and Optimization of Embedded Exhaust-Washed


Structures. PhD thesis, Wright State University, Dayton OH, 2013.

[107] Theodore von Karman and H-S Tsien. The buckling of thin cylindrical shells under axial
compression. Journal of the Aeronautical Sciences, 8:303–312, 1941.

107
[108] L Vu-Quoc and M Olsson. A computational procedure for interaction of high-speed vehicles
on flexible structures without assuming known vehicle nominal motion. Computer Methods
in Applied Mechanics and Engineering, 76(3):207–244, 1989.

[109] SC White and PM Weaver. Towards imperfection insensitive buckling response of shell
structures-shells with plate-like post-buckled responses. The Aeronautical Journal, 120
(1224):233, 2016.

[110] Xiao-Hui Zeng, Han Wu, Jiang Lai, and Hong-Zhi Sheng. Influences of aerodynamic loads
on hunting stability of high-speed railway vehicles and parameter studies. Acta Mechanica
Sinica, 30(6):889–900, 2014.

[111] Shannon A Zirbel, Kyler A Tolman, Brian P Trease, and Larry L Howell. Bistable mecha-
nisms for space applications. PloS one, 11(12):e0168218, 2016.

108
APPENDIX A

SENSITIVITY ANALYSIS

109
Define a system of equations of the form

R = KU − F = 0

X is used to compute a quantity of interest Q(U). An example can be solving for displacement

and use that displacement to compute stresses. In this example, U is the displacement of a given

structure and Q(U) are the stresses. An optimization problem can be formulated as:

minimize Q1 (U) (A.1)


xi

s.t. f ≤0 (A.2)

where xi are the design variables and f = Q2 (U) − Q˜3 is the constraint function with Q˜3 a given

constant. The sensitivity equations are expressed as

dfi ∂ fi ∂ fi dU
= + (A.3)
dx j ∂ x j ∂U dx j
∂K dU ∂F
U+K − = 0 (A.4)
∂ xj dx j ∂ x j
dU ∂F ∂K
K = − U (A.5)
dx j ∂ xj ∂ xj

In a more general way we can define the problem as

f = f (x, U(x)) (A.6)

R(x, U(x)) = 0 (A.7)

where x are the design variables, U(x) is the solution which depends on the design variables x and

R is the residual equation which is a function of the design variables and the solution. In order to

compute sensitivity of f with respect to a design variable x we calculate the following derivative

df ∂f ∂ f dU
= + (A.8)
dx ∂ x ∂U dx
110
A.0.1 Direct sensitivity

∂f ∂f
Starting from Equation A.8, ∂x and ∂U can be calculated using f . On the other hand, in order

to compute dU
dx we use the following steps

R(x, U(x)) = 0 (A.9)


dR ∂R ∂R dU
= + (A.10)
dx ∂ x ∂U dx

rearranging the previous equation

∂R dU ∂R
= − (A.11)
∂U dx ∂x
 −1
∂R ∂R

dU
=− (A.12)
dx ∂U ∂x

We can solve the previous equation for dU


dx and plug it back into Equation A.8.

A.0.2 Adjoint sensitivity

If we use the definition of dU


dx in Equation A.12 and substitute it in Equation A.8 we obtain

 −1
∂f ∂ f ∂R ∂R

df
= − (A.13)
dx ∂ x ∂U ∂U ∂x

we can define the parameter ψT as

 −1
∂ f ∂R

ψ =−
T
(A.14)
∂U ∂U

the Equation A.8 becomes


df ∂f ∂R
= + ψT (A.15)
dx ∂x ∂x

111
Going back to Equation A.14, we can solve for ψ

 −1
∂ f ∂R

ψ = −
T
(A.16)
∂U ∂U
 −T
∂R ∂f T

ψ = − (A.17)
∂U ∂U
T
∂R ∂f T

ψ = − (A.18)
∂U ∂U

A.0.3 Direct vs Adjoint Sensitivity

Direct Adjoint

∂R dU
 ∂R  T ∂f T
∂U dx = − ∂R
∂x ∂U ψ = − ∂U

∂f ∂ f dU ∂f
df
dx = ∂x + ∂U dx
df
dx = ∂x + ψT ∂R
∂x

When solving a convective equation; where the information propagates in one direction

∂R ∂R T
(U1 − > U2 − > ...Un ); the matrix ∂U will be a lower triangular matrix. Its transpose ∂U on the

other hand is an upper triangular matrix. The adjoint equations are therefore solving a problem in

the opposite direction of the direct equations.

For elliptic problems, specifically Poisson’s equation, the problem is self-adjoint therefore "no

reversal" of propagation, since there is no propagation in forward problem. Another benefit of

direct sensitivity is its relatively cheaper computational cost when the number of design variables

are very small compared to the number of the constraints. Adjoint sensitivity on the other hand is

used when the number of constraints is smaller than the number of design variables.

112
APPENDIX B

THERMOELASTICITY IN BEAMS

113
In this section, linear and nonlinear thermoelasticity problems are solved. First we go through

derivations of the main equations used for the implementation in matlab. The problem is defined

by a beam under a body force p and a thermal load ∆T. The beam is solved using Timoshenko

equations which allow shear deformation. The difference between linear and nonlinear is the

deformation assumption and the inclusion of the nonlinearstrain. In nonlinear thermoelasticity, the

Von Karman strain is used. First we start with a definition of the axial and transverse displacements

u and w and the rotation about the y axis β defined as

∂β
u = u x0 (x, t) + (z + z0 ) (B.1)
∂x
w = w(x, t) (B.2)
∂w
β = β(x, t) = −γ (B.3)
∂x

where γ is the shear strain. The von Karman strains are defined as

s = Ds xs + vk (xs ) (B.4)



 

u
 
    x0 
   2
∂ (z0 + z) ∂∂x  
 1 ∂w 
 

 
  
 
  
  ∂x    2 ∂x 
 
=   w + (B.5)
 
    

  

 0
γ   0 
  
−1  
      
∂x
  
  
 β
    
  
 
 
Linearization of the strain is given by

s∆ (xs, xs∆ ) = Ds xs∆ + vk



(xs, xs∆ ) (B.6)

 
 ∆
u

  x0 
     
∂ ∂  
(z0 + z) ∂ x  

 
 ∂w ∂w ∆ 

 ∂x    ∂x

∂x 

=   w∆ + (B.7)
    
  
0 ∂
0

−1  
   
∂x
  
 
 

β
  ∆

  

  
 
114
The stress-strain relation is expressed as
 
σ(xs ) = C s − α∆T {1, 0} T
(B.8)

 

u
 
    ©   
 x 0

   2   ª
∂ ∂ 1 ∂w
 σx x   1−ν2 0  ­­  ∂ x 0 (z0 + z) ∂ x  
E
1

 
   ­   
 
 
 
 
 
 ®
    2 ∂x 
    ®®
=   w + − α∆T
 
    
  
 ­ ®
 ­   

 τ   0 G ­­  0 ∂ x 0 0
     ®
−1  

       
 
 
 
 
 ®
      
 β
     



     ®
«   ¬
where E and G are the elastic and shear modulus, respectively. α is the coefficient of thermal

expansion and ν is the Poisson’s ratio. The perturbation of the stress-vector σs∆ (xs, xs∆ ) is defined

as

σs∆ (xs, xs∆ ) = Cs∆ (xs, xs∆ ) (B.9)



 
 ∆
u x0 

  ©­        ∆  ª
∂ ∂ ∂w ∂w
 1−ν2 0  ­­  ∂ x (z0 + z) ∂ x  
 E    
 
 
 
 ®
   ∂x
  ∂x 

 ®®
=   w∆ +
  
 ­ ®
 ­   
 0 G ­  0 ∂ −1    
0

 ®
∂x
  ­  
 
 
 
 ®
β 
   
 ∆

   ®

«   ¬
The finite element equations are derived using the principle of virtual work, which defines the

virtual displacement vector, δxs = {δu x0 , δw, δβ}T , and the virtual strain vector, s∆ (δxs, xs ). The

variational statement is expressed as: Find xs such that for all δxs
∫ ∫ h ∫
2
s∆ (δxs, xs )T σ(xs ) dz dΩ = δxs p · nz dΩ (B.10)
Ω − h2 Ω

where p = {p x , p y, pz }T and nz is the vertical unit vector normal to the midaxis of the beam. We

can discretize the our parameter using a suitable discretization scheme,

u x0 = Bu h (B.11)

w = Bwh (B.12)

β = Bβh (B.13)
115
and

∂u x0 ∂B
= u h = Bx u h (B.14)
∂x ∂x
∂w ∂B
= wh = Bx w h (B.15)
∂x ∂x
∂β ∂B
= βh = Bx βh (B.16)
∂x ∂x

where B is the set of shape functions. Equation B.10 can be represented in a matrix-vector form as

F1 (Xs ) − FE (Xs ) = 0 (B.17)

where the matrices and force vector are related to the variational statement through

∫ ∫ h
2
δXs F1 (Xs ) = s∆ (δxs, xs )T Cs (xs ) dz dΩ
Ω − h2
∫ ∫ h T
2

= Ds δxs + vk

(xs, δxs ) C (Ds xs + vk (xs )) dz dΩ
Ω − h2
∫ ∫ h ∫
2
δXs FE (Xs ) = s∆ (δxs, xs )T Cα∆T {1, 0}T dz dΩ + δxs p · nz dΩ
Ω − h2 Ω
∫ ∫ h T ∫
2

= Ds δxs + vk

(xs, δxs ) Cα∆T {1, 0} dz dΩ +
T
δxs p · nz dΩ
Ω − h2 Ω

∫ ∫ h
2
δXs F1 (Xs ) = (Ds δXs )T C (Ds xs ) dz dΩ
Ω − h2
∫ ∫ h
2
+ (Ds δXs )T Cvk (xs ) dz dΩ
Ω − h2
∫ ∫ h
2
+ vk

(Xs, δXs )T C (Ds xs ) dz dΩ
Ω − h2
∫ ∫ h
2
+ vk

(xs, δxs )T Cvk (xs ) dz dΩ (B.18)
Ω − h2

116
∫ ∫ h
2
δXs F1 (Xs ) = δxsT DTs CDs xs dz dΩ
Ω − h2
  2
 21 ∂w
∫ ∫ h

 

2 ∂x
 
+ δxsT DTs C
 

dz dΩ
Ω − h2
0

 


 

 
    T
∂w ∂δw
∫ ∫ h

 

2  ∂x ∂x 
 
+
 
C (Ds xs ) dz dΩ
Ω − h2
0

 


 

 
  2
1 ∂w 
∫ ∫ h

 
2 2 ∂x
 
+ vk (xs, δxs ) C (B.19)
∆ T

 

dz dΩ
Ω − h2
 0 

 

 
 

117
∫ ∫ h
2
DTs CDs Xs dz dΩ =
Ω − h2
 T   
 x 0 (z + z0 )Bx   1−ν2 0   Bx 0 (z + z0 )Bx 
∫ ∫ Bh   E  
2
     Xs dz dΩ
Ω − h2 
    
 0 Bx −B   0 G  0 Bx −B 
   
    
 

 BTx 0   
∫ ∫ h 
2 
 E B
  1−ν2 x 0 1−ν
E
2 (z + z0 )B x

=

 0 BT  
x  
 Xs dz dΩ
Ω − h2 
 
 0 G Bx −G B 
  
(z + z )BT −BT   
 0 x 
 
 


E
BTx 1−ν 2 Bx 0 BTx 1−ν E
2 (z + z0 )B x


∫ ∫ h  
2 
=

 0 B T G B
x x −B TG B
x
 Xs dz dΩ
Ω − h2 
 

 
 BT (z + z ) E B −BT GB BT (z + z )2 E B + BT GB
 x 0 1−ν 2 x x x 0 1−ν 2 x 
 
 
E z0 h
 BT E h B
 x 1−ν2 x 0 BTx 1−ν 2 Bx


∫  
= (B.20)
 

 0 BTx Gh Bx −BTx Gh B  Xs dΩ

Ω  
 BT E z0 h B −BT GhB BT ( h3 + z2 h) E B + BT GhB
 
 x 1−ν2 x x x 12 0 1−ν 2 x 
 

118
  2  T    2
1 ∂w  1 ∂w 
0 +   E 0
h
  h  
B (z z )B
 
0 x   1−ν 2
∫ ∫ ∫ ∫
2  2 ∂x  2 x  2 ∂x 
 
dz dΩ =
 
  

DTs C     dz dΩ
− h2 Ω − h2 
   

0  0 Bx −B   0 G   0 

 
    
 


 
 
      
 

 BTx 0    2
 2(1−ν2 ) ∂w
E
∫ ∫ h   

2  ∂x
 
=
 

 0 B T
x
 dz dΩ
Ω − h2 
 
 0
 

  

(z + z )B T −B T
  
 0 x 
 
 
E
 T

B
 
2(1−ν 2
 x

 ) 
 ∂w 2
∫ ∫ h 
 
 
2

=

 
0 ∂x
dz dΩ
Ω − h2 





 
 Bx (z + z0 ) 2(1−ν2 ) E

 T


 
 
 
Eh 
 
BTx 2(1−ν

2) 

 
 ∂w 2

∫  
  
 
=

 
0 ∂x
dΩ
Ω





 
 T E z0 h 
 Bx 2(1−ν2 ) 

 

 

 


 B T 
 x 
2  
∂w
∫   
Eh
 
= (B.21)

 

0 dΩ
Ω ∂x 2(1 − ν 2 )  





 


B T z

 x 0
 
 

119
    T
∂w ∂δw
∫ ∫ h

 

2  ∂x ∂x 

 

C (Ds xs ) dz dΩ
Ω − h2
0

 


 

 
 T  
h  0 B 0   E B 0 E
(z + z )B
∂w  0 x
∫ ∫  
2 x   1−ν2 x 1−ν 2
=    Xs dz dΩ
Ω − h2 ∂ x 
   
0 0 0   0 G Bx −G B
  

   
 
 0 0
   
∫ ∫ h 
2 
 E B
 ∂w  1−ν2 x
 0 1−ν
E
2 (z + z )B
0 x

=  BT 0
 x  ∂x 
  Xs dz dΩ
Ω − h2 

  0 G Bx −G B 
   
0 0
   
 
 
 

 0 0 0 

∫ ∫ h  
2 
=
    
 BT E ∂w B 0 BT E ∂w (z + z )B  Xs dz dΩ
x 1−ν 2 ∂x x x 1−ν 2 ∂x 0 x
Ω − h2 
 

 
0 0 0
 
 
 
 

 0 0 0 

∫  
=
     
 BT E h ∂w B 0 BT E z0 h ∂w B  Xs dΩ
 x 1−ν2 ∂ x x x 1−ν 2 ∂ x x
Ω  
 
0 0 0
 
 
 
 
 0 0 0 

E h ∂w 
∫   
= (B.22)

2 ∂x BTB 0 BT z B  Xs dΩ
x 0 x
Ω 1−ν
x
 x


 
0 0 0
 
 
 

120
 T   2
 0 B 0   E ∂w 
h h

∂w
∫ ∫ ∫ ∫  
2 2 x 2 ∂
 2(1−ν )
 
x
vk (xs, δxs )T Cvk (xs ) dz dΩ =

  
 

 dz dΩ
− h2 Ω − h2 ∂ x 
 

0 0 0   0
 
 

 

   
 
 0 0
    2
 2(1−ν2 ) ∂w
E
∫ ∫ h   
 ∂w 
  
2  ∂x
 
=


 BT 0 dz dΩ
Ω − h2 
x
 ∂x 
 
0
 

  
 

0 0
   
 
 
 
0

 


 

  3
∂w
∫   
Eh

 
= (B.23)
 

BTx  2(1 − ν 2 ) ∂ x dΩ
Ω

 

 
0

 

 
 

∫ ∫ h
2
δXs FE (Xs ) = δXs DTs Cα∆T {1, 0}T dz dΩ
Ω − h2
∫ ∫ h
2
+ vk

(xs, δxs )T Cα∆T {1, 0}T dz dΩ
Ω − h2

+ δXs p · nz dΩ (B.24)

121
 

 BTx 0     
 E 0 1
∫ ∫ h ∫ ∫ h   
 

2 2  1−ν 2
 
Ds Cα∆T {1, 0} dz dΩ = α∆T
T T
   
 

 0 B T
x
   dz dΩ
− h2 Ω − h2 
   
Ω   0 G
0

 

    
(z + z )BT −B T
    
 0 x 
 
 

 B T
x 0 
  
1
∫ ∫ h   
 

2 E
 
= α∆T
  
 

 0 T
Bx  1 − ν 2
 dz dΩ
Ω − h2 

0
 
 

   
(z + z0 )Bx −B 
T T
   
 

 

BTx 

 

 

∫  
Eh
 
= α∆T (B.25)

 

2 0 dΩ
Ω 1−ν 






 

T
 Bx z0 

 

 

 
 0 0
   
1
h ∫ ∫ h   
 ∂w
∫ ∫    
2 2  E

vk (xs, δxs ) Cα∆T {1, 0} dz dΩ = α∆T
∆ T T
 

 BT 0  dz dΩ
 ∂x 1 − ν
x 2
− h2 Ω − h2 
 

0

 

   
 0 0
   
 

 
0
 


 
 
∂w
∫    
Eh

 
= α∆T (B.26)
 

BT dΩ
Ω ∂x 1 − ν
2  x

 

 
0

 


  
 

 
0

 


 

∫ 

∫  

 
p · nz dΩ = (B.27)
 

BT p z dΩ
Ω Ω





 
0

 

 
 

122
The nonlinear system of equations in Equation B.17 are solved using Newton-Raphson (N-R)

iterations. The Jacobian for each N-R iteration is defined in the perturbed form of

∫ ∫ h
2
 
 ∆ (δxs, xs, xs∆ )σ(xs ) +  ∆ (δxs, xs )C ∆ (xs, xs∆ ) dz dΩ (B.28)
Ω − h2

where
T
∂δw ∂w ∆

 ∆
(δxs, xs, xs∆ ) = ,0 (B.29)
∂x ∂x

123
The previous equation can be re-written as δXsT J Xs∆ where J is the Jacobian.

    T
0

 
 
 


 
  


 
 
   

   
 δXs
  
    
T ∆
0 0 X σ
h h
 T
  
B B
   
s
∫ ∫ ∫ ∫
2 2 x  xx 
 x
 
 ∆ (δxs, xs, xs∆ )σ(xs ) dz dΩ =
 
  

dz dΩ
 
 
 
− h2 − h2
 
Ω Ω
0  τ 

 
 
 
 
 


   
  

   
  

 

0

 


 

 
 
0

 


 

∫ ∫ h 
 
  
2
 
= δXs BTx σx x 0 Bx 0 Xs∆ dz dΩ
T

 

h
Ω −2 





 
0

 


  
 
 
0

 


 

∫ ∫ h   2 !
∂u ∂w 1 ∂w
   
2 E
 
x0
= δXs BTx + (z0 + z) + − α∆T 0 Bx 0 Xs∆ dz dΩ
T

 

Ω −2h 
 
 1 − ν 2 ∂ x ∂ x 2 ∂ x

 

0

 
 

 
 
0

 


 

  2 !
∂u ∂w 1 ∂w
∫    
E h
 
x 0
= δXs BTx + (z0 ) + − α∆T 0 Bx 0 Xs∆ dΩ
T

 

Ω 
 
 1 − ν 2 ∂ x ∂ x 2 ∂ x

 

0

 

 
 
 
0 0 0
 
2
!
∂u x0 ∂w 1 ∂w
∫    
T Eh
= δXs + + α∆T (B.30)
 
(z 0 ) − 0 BT B 0 dΩ
Ω 1 − ν2 ∂ x ∂x 2 ∂x 
 x x 

 
0 0 0
 
 

124
∫ ∫ h
2
 ∆ (δxs, xs )C ∆ (xs, xs∆ ) dz dΩ
Ω − h2
∫ ∫ h T
2
  
= Ds δxs + vk

(xs, δxs ) C Ds xs∆ + vk

(xs, xs∆ ) dz dΩ
Ω − h2
∫ ∫ h
2
= δXsT DTs CDs Xs∆ dz dΩ
Ω − h2
∫ ∫ h
2
+ δXsT DTs Cvk

(xs, xs∆ ) dz dΩ
Ω − h2
∫ ∫ h
2
+ vk

(xs, δxs )CDs Xs∆ dz dΩ
Ω − h2
∫ ∫ h
2
+ vk

(xs, δxs )Cvk

(xs, xs∆ ) dz dΩ (B.31)
Ω − h2

∫ ∫ h
2
δXsT DTs CDs Xs∆ dz dΩ
Ω − h2
 
E z0 h
 BT E h B
 x 1−ν2 x 0 T
Bx 1−ν2 Bx 

∫  
= δXS (B.32)
T
  ∆

 0 BTx Gh Bx −BTx Gh B  X dΩ
 s
Ω  
3
 
+ z02 h) 1−ν
 T E z0 h h
 Bx 1−ν2 Bx −BT GhBx BTx ( 12 E
2 B x + BT GhB

 

125
∫ ∫ h
2
δXsT DTs Cvk

(xs, xs∆ ) dz dΩ
Ω − h2
 T    
 x 0 (z + z0 )Bx   1−ν2 0   ∂w  0 Bx 0
∫ ∫ Bh   E
2
= δXsT    
 ∂x 
  X ∆ dz dΩ
 s
Ω − h2 
  
 0 Bx −B   0 G 0 0 0 
  
     
 

 BTx 0   
∫ ∫ h  0 E
B 0
∂w 
 
2  1−ν 2 x
= δXsT

 0 B T   X ∆ dz dΩ
x  ∂x   s
Ω − h2 

 0 0 0
  
(z + z )BT −BT   
 0 x 
 
 
0 BTx 1−ν E
0
 2 Bx
∫ ∫ h 
∂w 
 
2
= δXsT
 ∆
0 0 0 Xs dz dΩ
Ω − h2 ∂ x  
 
0 BT (z + z ) E B 0
 x 0 1−ν 2 x
 
 
0 BT E h B 0
 x 1−ν 2 x 
∂w 
∫   
= δXsT
 ∆
∂ x  0 0 0  X dΩ
 s
Ω  
 
0 BT E z0 h B 0
 x 1−ν 2 x 
 
 
0 BT B 0 
 x x 
∂w
∫    
Eh 
= δXs (B.33)
T
  ∆
0 0 0 X dΩ
Ω ∂x 1 − ν 
2   s

 
0 BT z B 0
 x 0 x 
 

126
∫ ∫ h
2
vk

(xs, δxs )CDs Xs∆ dz dΩ
Ω − h2
 

 0 0 0 

∂w
∫    
= δXsT
  ∆
 BT E h B 0 E z0 h  X dΩ
∂x  x 1−ν2 x BTx 1−ν 2 B x s
Ω  
 

 0 0 0 

 
 
 0 0 0 

∂w
∫    
E h 
= δXsT (B.34)
 ∆
∂x
T
1 − ν 2  Bx Bx 0 BTx z0 Bx  Xs dΩ
Ω 
 
 0 0 0 

 

∫ ∫ h
2
vk

(xs, δxs )Cvk

(xs, xs∆ ) dz dΩ
Ω − h2
 T    
h  0 B 0   E 0  0 B 0 
∂w   ∂w 
∫ ∫  
2 x 2 x
= δXsT   1−ν
      ∆
 X dz dΩ
Ω − h2 ∂ x 
    ∂x   s
0 0 0   0 G 
   0 0 0 
 
     
 
 0 0
   
∫ ∫ h   2 0 E B 0 
 BT 0 ∂w  1−ν

2  2 x
= δXsT
   ∆
 X dz dΩ
h  x  ∂x   s
Ω −2  0


  0 0

 0 0
   
 
 
0 0 0

2 
∂w 
∫  
= δXs (B.35)
T
 ∆
0 B T Eh B 0  X dΩ
Ω ∂x
 x 1−ν 2 x  s
 
 
0 0 0
 
 
 

127
ProQuest Number: 28967283

INFORMATION TO ALL USERS


The quality and completeness of this reproduction is dependent on the quality
and completeness of the copy made available to ProQuest.

Distributed by ProQuest LLC ( 2022 ).


Copyright of the Dissertation is held by the Author unless otherwise noted.

This work may be used in accordance with the terms of the Creative Commons license
or other rights statement, as indicated in the copyright statement or in the metadata
associated with this work. Unless otherwise specified in the copyright statement
or the metadata, all rights are reserved by the copyright holder.

This work is protected against unauthorized copying under Title 17,


United States Code and other applicable copyright laws.

Microform Edition where available © ProQuest LLC. No reproduction or digitization


of the Microform Edition is authorized without permission of ProQuest LLC.

ProQuest LLC
789 East Eisenhower Parkway
P.O. Box 1346
Ann Arbor, MI 48106 - 1346 USA

You might also like