Index

EXPERIMENTAL AND NUMERICAL INVESTIGATION OF BALLISTIC
IMPACT BEHAVIOUR OF HIGH STRENGTH ALUMINIUM PLATES
A THESIS SUBMITTED TO
THE GRADUATE SCHOOL OF NATURAL AND APPLIED SCIENCES
OF
MIDDLE EAST TECHNICAL UNIVERSITY
BY
GÜRALP BAŞARAN
IN PARTIAL FULFILLMENT OF THE REQUIREMENTS

FOR
THE DEGREE OF MASTER OF SCIENCE
IN
AEROSPACE ENGINEERING
DECEMBER 2019
Approval of the thesis:

submitted by GÜRALP BAŞARAN in partial fulfillment of the requirements for the

degree of Master of Science in Aerospace Engineering Department, Middle East
Technical University by,
Prof. Dr. Halil Kalıpçılar

Dean, Graduate School of Natural and Applied Sciences
Prof. Dr. İsmail Hakkı Tuncer

Head of Department, Aerospace Engineering
Assoc. Prof. Dr. Ercan Gürses

Supervisor, Aerospace Engineering, METU
Examining Committee Members:
Prof. Dr. Altan Kayran

Aerospace Engineering, METU
Assoc. Prof. Dr. Ercan Gürses

Assoc. Prof. Dr. Demirkan Çöker

Assoc. Prof. Dr. Melin Şahin

Assist. Prof. Dr. Recep M. Görgülüarslan

Mechanical Engineering, TOBB
Date:
I hereby declare that all information in this document has been obtained and
presented in accordance with academic rules and ethical conduct. I also declare
that, as required by these rules and conduct, I have fully cited and referenced all
material and results that are not original to this work.
Name, Surname: GÜRALP BAŞARAN
Signature :
iv
ABSTRACT

BAŞARAN, GÜRALP
M.S., Department of Aerospace Engineering
Supervisor: Assoc. Prof. Dr. Ercan Gürses
December 2019, 90 pages
A hybrid solution method was used to examine the ballistic collision situation to be
used in armored vehicle design. This hybrid solution method includes an Artificial
Neural Network (ANN) and a Finite Element (FE) solver. MATLAB was used for
ANN model, and LSDYNA® was used as FE solver. For this purpose, first ballis-
tic tests were performed, and projectile residual velocities and depth of penetrations
were measured. The FE model was confirmed by ballistic tests. After the FE model
validation, FE analyses were performed for different armor thicknesses, and the re-
sults were transferred in the ANN model. The ANN model and FE method results
were compared for different armor thicknesses, and the ANN model was validated.
The validated ANN model was transferred to MATLAB® SIMULINK® and a tool
that is capable of predicting the results of ballistic collision in a short time.
Keywords: ballistics, armor, bullet, Finite Element Analysis, neural network
v
ÖZ
YÜKSEK MUKAVEMETLİ ALÜMİNYUM LEVHALARIN BALİSTİK

ÇARPMA DAVRANIŞLARININ DENEYSEL VE SAYISAL OLARAK
İNCELENMESİ
BAŞARAN, GÜRALP
Yüksek Lisans, Havacılık ve Uzay Mühendisliği Bölümü
Tez Yöneticisi: Doç. Dr. Ercan Gürses
Aralık 2019 , 90 sayfa
Zırhlı araç tasarımında kullanılmak üzere, balistik çarpışma durumunu incelemek için
hibrit bir çözüm metodu kullanılmıştır. Bu hibrit çözüm yöntemi, Yapısal Sinir Ağı
(YSA) ve Sonlu Elemanlar (SE) çözücüsünü içermektedir. YSA modeli için MAT-
LAB, SE çözücüsü olarak da LSDYNA® kullanılmıştır. Bu amaçla, ilk olarak ba-
listik testler gerçekleştirilmiş, mermi çıkış hızları ve giriş derinlikleri ölçülmüştür.
Gerçekleştirilen balistik testler ile SE modeli doğrulanmıştır. SE modeli doğrulama-
sından sonra, farklı zırh kalınlıkları için SE analizleri yapılmış ve sonuçlar YSA mo-
delinde kullanılmak üzere kaydedilmiştir. Farklı zırh kalınlıkları için YSA modeli
ve SE yöntemi sonuçları karşılaştırılmış ve YSA modelinin doğrulaması yapılmıştır.
Doğrulanan YSA modeli MATLAB® SIMULINK®’e aktarılmış ve balistik çarpışma
sonuçlarını kısa sürede tahmin edebilen bir gereç yapılmıştır.
Anahtar Kelimeler: balistik, zırh, mermi, Sonlu Eleman Analizleri, Yapısal Sinir Ağ
vi
Dedicated to my family
vii
ACKNOWLEDGMENTS
First and foremost, I would like to express my gratitude to my supervisor Assoc.

Prof. Dr. Ercan Gürses for his constant support, guidance and patience throughout
the preparation of this study.
Besides my supervisor, I would like to thank the rest of my thesis committee: Prof.
Dr. Altan Kayran, Assoc. Prof. Dr. Demirkan Çöker, Assoc. Prof. Dr. Melin Şahin
and Dr. Recep Görgülüaslan, for their insightful comments and contributions during
thesis defense.
I would like to give my special thanks to Erkal Özbayramoğlu, Onur Bütün, Eren
Öney and İlker Kurtoğlu for their support. They have always been there for me when-
ever I needed help.
Lastly, I also would like to thank my wife Esra Başaran and my parents Perihan
Başaran and Gazi Başaran for their caring and love throughout all the time. Their
encouragement and support brought me to these days.
viii
TABLE OF CONTENTS
ABSTRACT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v
ÖZ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi
ACKNOWLEDGMENTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii
TABLE OF CONTENTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii
LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xv
CHAPTERS
1 INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Ballistic Impact . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Types of Ballistic Collision and Projectiles and Target . . . . . . . . 2
1.2.1 Categorization of Ballistic Collision Types . . . . . . . . . . . 2
1.2.2 Projectile Types . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.3 Target Types . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 The ballistic response of materials . . . . . . . . . . . . . . . . . . . 4
1.4 Penetration Mechanics . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.5 Limit Velocities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.5.1 Elastic Limit Velocity . . . . . . . . . . . . . . . . . . . . . . 7
1.5.2 Plastic Limit Velocity . . . . . . . . . . . . . . . . . . . . . . 8
ix
1.5.3 Hydrodynamic Limit Velocity . . . . . . . . . . . . . . . . . 9
1.6 Ballistic Limit and Aerial Density . . . . . . . . . . . . . . . . . . . 10
1.7 Literature Survey on ballistic FEA combined ANN studies . . . . . . 11
1.8 Aim of this study . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2 NUMERICAL MODELLING . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.1 Numerical modeling approaches . . . . . . . . . . . . . . . . . . . . 16
2.1.1 Lagrangian approach . . . . . . . . . . . . . . . . . . . . . . 16
2.1.2 Eulerian approach . . . . . . . . . . . . . . . . . . . . . . . . 18
2.1.3 Arbitrary Lagrange Euler coupling . . . . . . . . . . . . . . . 19
2.1.4 Smooth Particle Hydrodynamics . . . . . . . . . . . . . . . . 20
2.2 Integration methods . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2.1 Implicit integration . . . . . . . . . . . . . . . . . . . . . . . 23
2.2.2 Explicit integration . . . . . . . . . . . . . . . . . . . . . . . 24
2.3 CONCEPTS IN NUMERICAL SIMULATION . . . . . . . . . . . . 25
2.3.1 Hourglass Deformation and Hourglass Damping . . . . . . . . 26
2.3.1.1 Hourglass Deformation . . . . . . . . . . . . . . . . . 26
2.3.1.2 Hourglass Damping . . . . . . . . . . . . . . . . . . . 27
2.3.2 Contact-Impact Methods . . . . . . . . . . . . . . . . . . . . 28
2.3.2.1 Kinematic Constraint Method . . . . . . . . . . . . . . 29
2.3.2.2 Penalty Method . . . . . . . . . . . . . . . . . . . . . 30
3 MATERIAL MODELLING AND FAILURE MODELS . . . . . . . . . . . 31
3.1 Gruneisen parameter . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.1.1 Shock Rankine-Hugoniot conditions and relation . . . . . . . 33
x
3.1.2 Mie-Gruneisen equation of state . . . . . . . . . . . . . . . . 34
3.2 Plasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.2.1 Johnson-Cook Model . . . . . . . . . . . . . . . . . . . . . . 36
3.3 Material Failure Modelling . . . . . . . . . . . . . . . . . . . . . . . 37
3.3.1 Failure Models . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.3.2 Johnson-Cook Failure Model . . . . . . . . . . . . . . . . . . 40
3.3.3 Modified Johnson-Cook Model and Cockroft-Lathan Failure

Criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4 BALLISTIC IMPACT EXPERIMENTS OF HIGH STRENGTH ALUMINUM

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.1 Configuration and Experimental Setup and Results . . . . . . . . . . 43
4.2 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5 FINITE ELEMENT ANALYSIS . . . . . . . . . . . . . . . . . . . . . . . 53
5.1 Computational Domain and Numerical Setup . . . . . . . . . . . . . 53
5.1.1 Mesh sensitivity . . . . . . . . . . . . . . . . . . . . . . . . . 54
5.2 Hourglass Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5.3 Element Type Solutions . . . . . . . . . . . . . . . . . . . . . . . . 59
5.4 Analysis results and experiment comparison . . . . . . . . . . . . . . 61
5.4.1 Results of 25.4mm aluminum plate . . . . . . . . . . . . . . . 61
5.4.2 Results of 31mm aluminum plate . . . . . . . . . . . . . . . . 63
5.4.3 Results of 38mm aluminum plate . . . . . . . . . . . . . . . . 65
6 ARTIFICIAL NEURAL NETWORK . . . . . . . . . . . . . . . . . . . . . 69
6.1 Theory of Artificial Neural Network . . . . . . . . . . . . . . . . . . 70
6.2 ANN Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
xi
6.3 FEA and ANN results comparison . . . . . . . . . . . . . . . . . . . 77
7 SUMMARY AND CONCLUSION . . . . . . . . . . . . . . . . . . . . . . 81
7.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
7.2 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
APPENDICES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
xii
LIST OF TABLES
TABLES
Table 1.1 Classification of ballistic impact [1] . . . . . . . . . . . . . . . . . 2
Table 1.2 Classification of projectile shape [1] . . . . . . . . . . . . . . . . . 3
Table 1.3 Classification of target [1] . . . . . . . . . . . . . . . . . . . . . . . 4
Table 1.4 Response of materials under the ballistic impact [2] . . . . . . . . . 6
Table 4.1 Velocity levels for protection levels in STANAG4569 [3] . . . . . . 44
Table 4.2 Calibration results . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
Table 4.3 Ballistic test plan . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
Table 4.4 The results of ballistic experiments for 25.4mm aluminum plate at
960m/s velocity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
Table 4.5 The results of ballistic experiments for 25.4mm aluminum plate at
800m/s velocity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
Table 4.6 The results of ballistic experiments for 31mm aluminum plate at
960m/s velocity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
800m/s velocity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
960m/s velocity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
xiii
800m/s velocity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
Table 5.1 Residual Velocity Values for different mesh sizes . . . . . . . . . . 55
Table 5.2 Residual velocity values for different hourglass formulations . . . . 58
Table 5.3 Residual velocity values for each element formulations . . . . . . . 61
Table 5.4 Residual velocity comparison between FEA and experiment for
25.4mm plate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
31mm plate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
38mm plate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
Table 6.1 FE analysis results for various strike velocities . . . . . . . . . . . . 73
Table 6.2 MSE and RMSE Values of Training, Validation and Test data . . . . 76
Table 6.3 ANN and FEA comparison of residual velocity values . . . . . . . 79
Table 6.4 ANN and Ballistic experiment comparison of residual velocity val-
ues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
xiv
LIST OF FIGURES
FIGURES
Figure 1.1 Deformed shape of target after strike and perforation process [4]. 4
Figure 1.2 Geometry and dimensions (in mm) of 20mm ogive-nose hard-
ened steel projectile(left) and 7.62mm APM2 bullet (right) [4]. . . . . . 5
Figure 1.3 Permanent deformations of thin target element showing bulging

and dishing [1]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
Figure 1.4 Definitions of perforation and partial penetration for defining the
ballistic limit [5]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
Figure 2.1 Numerical inputs and formulations governing computational bal-

listics [1]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
Figure 2.2 Deformed lagrangian elements [6]. . . . . . . . . . . . . . . . . 17
Figure 2.3 Lagrangian mesh and modeling [1]. . . . . . . . . . . . . . . . . 17
Figure 2.4 Eulerian mesh [1]. . . . . . . . . . . . . . . . . . . . . . . . . . 18
Figure 2.5 Eulerian mesh [2]. . . . . . . . . . . . . . . . . . . . . . . . . . 18
Figure 2.6 Difference between Lagrangian, Eulerian and ALE [6]. . . . . . 19
Figure 2.7 Time integration loop of ALE approach [6]. . . . . . . . . . . . 20
Figure 2.8 SPH (left) and Lagrange (right) representation of Fragment Sim-
ulating Projectile . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
Figure 2.9 The linear assumption for acceleration . . . . . . . . . . . . . . 23
xv
Figure 2.10 (a) An undeformed and deformed (b) one-point integration ele-
ment, (c) an undeformed and deformed (d) full integration element . . . 26
Figure 2.11 Kinematic constraint method [7] . . . . . . . . . . . . . . . . . 29
Figure 3.1 Schematic view of a shock front (line C) propagating through a

compressible material [8]. . . . . . . . . . . . . . . . . . . . . . . . . . 33
Figure 3.2 Failure types from left to right; brittle, ductile from brittle and
ductile, respectively [9]. . . . . . . . . . . . . . . . . . . . . . . . . . . 38
Figure 4.1 The dimension of 12.7mm and 20mm fragment simulating pro-
jectile [3]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
Figure 4.2 An overview of the ballistic laboratory . . . . . . . . . . . . . . 44
Figure 4.3 Impact room experimental setup (left) and high-speed camera
casing (right). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
Figure 4.4 Demonstration of camera position to the target plate. . . . . . . . 46
Figure 4.5 Target plate fixture. . . . . . . . . . . . . . . . . . . . . . . . . 46
Figure 4.6 High-speed camera images from calibration test. . . . . . . . . . 47
Figure 4.7 The casing to prevent flare effects (above), high-speed camera
image of the projectile with casing setup (below). . . . . . . . . . . . . 48
Figure 4.8 Dimensions of target plate and thickness configurations. . . . . . 49
Figure 4.9 Deformed shape of 25.4mm (left), 31mm (middle), and 38mm
(right) plate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
Figure 4.10 800m/s velocity impact to 38mm aluminum plate . . . . . . . . 52
Figure 5.1 The model for impact simulations . . . . . . . . . . . . . . . . . 53
Figure 5.2 The coarsest mesh (left) and the finest mesh (right) used in mesh
sensitivity analyses. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
xvi
Figure 5.3 Mesh sensitivity results . . . . . . . . . . . . . . . . . . . . . . 56
Figure 5.4 Crater shapes of the coarsest mesh and the finest mesh . . . . . . 56
Figure 5.5 Perforating state and final state of the model for each hourglass
formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
Figure 5.6 Hourglass energy comparison . . . . . . . . . . . . . . . . . . . 59
Figure 5.7 Reduced integration and fully integration elements . . . . . . . . 60
Figure 5.8 Deformed shapes of each element formulation . . . . . . . . . . 60
Figure 5.9 Perforation of FSP-25.4mm at 954m/s (above) and 804m/s (be-

low) velocity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
Figure 5.10 Crater shape comparison of 25.4mm plate impacted with 954m/s
velocity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
Figure 5.11 Perforation of FSP-31mm at 968m/s (above) and 824m/s (be-

low) velocity cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
Figure 5.12 Crater shape comparison of 31mm plate . . . . . . . . . . . . . 65
Figure 5.13 Perforation of FSP-38mm at 978m/s (above) and 800m/s (be-

low) velocity case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
Figure 5.14 38mm plate 800m/s test (left) and analysis (right) results . . . . 67
Figure 5.15 38mm plate 800m/s test (left) and analysis (right) results com-
parison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
Figure 6.1 The multilayer structure of ANN [10] . . . . . . . . . . . . . . . 70
Figure 6.2 One neuron calculation of ANN . . . . . . . . . . . . . . . . . . 71
Figure 6.3 Representation of the ANN setup . . . . . . . . . . . . . . . . . 74
Figure 6.4 Performance of Levenberg- Marquardt algorithm performance. . 74
Figure 6.5 123 epoch results (left) and 20 hidden layers (right) results . . . 75
xvii
Figure 6.6 Regression analysis results for training, test, validation and over-
all data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
Figure 6.7 ANN Simulink setup for ballistic analysis . . . . . . . . . . . . 78
Figure 6.8 Response surface representation of ANN results . . . . . . . . . 80
Figure A.1 A detailed display of the 31mm thickness 800m/s impact veloc-
ity perforation - (0µs - 20µs) . . . . . . . . . . . . . . . . . . . . . . . 87
ity perforation - (30µs - 70µs) . . . . . . . . . . . . . . . . . . . . . . 88
ity perforation - (80µs - 100µs) . . . . . . . . . . . . . . . . . . . . . 88
Figure A.4 Experiment, FEA and ANN results for 25.4mm thickness . . . . 89
Figure A.5 Experiment, FEA and ANN results for 31mm thickness . . . . . 89
Figure A.6 Experiment, FEA and ANN results for 38mm thickness . . . . . 90
xviii
CHAPTER 1
INTRODUCTION
1.1 Ballistic Impact
Ballistics is a branch of science that examines the situation of ammunition from the
moment of the exit from the barrel until the collision with the target. Much effort has
been spent in recent years to understand and mathematically express the physics at
the moment of collision. The examination of ballistic impact with a computer-aided
approach has been developed over the years. New demands of ground forces are
faster and more agile transportation from one point to another with safe occupants.
Impact studies are mainly focusing on two main subjects, which are vulnerability
and survivability requirements. Vulnerability and survivability criteria also contain
equipment protection, which can be a risk for the personal and critical for the vehicle.
Plastic deformation, fracture mechanics, wave propagation, contact mechanics, and
thermodynamics are all involved with the highly non-linear and complex ballistic
impact phenomena. During a ballistic collision, large strain and strain rate hardening
mechanisms work together with thermal softening mechanisms. Temperature levels
can go beyond the melting point, and strain rate levels are above 1000 s−1 . Generally,
processes end in 30µs and projectile initial velocities are above 600 m/s. The first
chapter is mainly about the classification of ballistic collisions, projectile and target
types, and reaction characteristics of a target under high-velocity impact at changing
strain rates. In the second section, the technical terms used in ballistic impact studies
will be explained. Definitions of the ballistic limit, the fragment simulating projectile,
the aerial density are given, and the physics of impact is shortly explained.
1
1.2 Types of Ballistic Collision and Projectiles and Target
Mainly, the ballistic collision is classified using parameters such as the angle of inci-
dence of impact, material properties, impact velocities, etc. The main characteristics
of penetration and perforation mechanisms are defined with the speed of a bullet.
Therefore, most commonly, the initial velocity of a projectile is used for the classifi-
cation.
1.2.1 Categorization of Ballistic Collision Types
There are two main groups concerning projectile impact velocity for the categoriza-
tion of ballistic impact, which is used in military applications and engineering. In
Table 1.1 the classification of ballistic impact is given [1].
Table 1.1: Classification of ballistic impact [1]

Ballistic impact Classification Velocity Range Vs (m/s) Example
Sub ordnance 25-500 Drop tests
Nominal ordnance or ordnance 500-1300 Propellant fired conventional gun
Military application
Ultra-ordnance 1300-3000 Warhead Fragments
Hypervelocity >3000 Shaped Charge
Low/moderate velocity impact <50 Automobile Crashes
General Application High-velocity impact 50-1500 Projectile impacts, bird strike in aircraft engines, etc.
Hypervelocity impact >1500 Debris impact on satellites
Subgroups of ballistic collision categorization are mainly based on the strike ve-
locity of the projectile. In military applications, there are four main strike velocity
ranges, so-called sub-ordnance, nominal ordnance, ultra-ordnance, and hyperveloc-
ity. Their corresponding velocity ranges are given in Table 1.1. In order to perform
sub-ordnance ballistic tests, pneumatic guns can be used. For a nominal ordnance test,
commonly conventional guns are used. Finally, for ultra-ordnance and hyperveloc-
ity strike velocity ranges, special purpose guns and light-gas guns can be employed,
respectively.
In general applications, crashworthiness studies of vehicles are performed in veloc-
2
ity ranges less than 50 m/s, which refers to the low/moderate velocity impact range.
The high-velocity impact is encountered, for example, in projectile impacts and bird
strikes. Finally, space shuttles and satellites have requirements against meteor im-
pacts, which are examples of a hypervelocity impact.
1.2.2 Projectile Types
Projectiles may have different forms, depending on their main body structure and tip
shapes. Table 1.2 summarizes the common projectile types.
Table 1.2: Classification of projectile shape [1]
Projectile Shape
Sphere
Solid rod or bar
Hollow Shell
Main Body
Irregular Solid
Solid rod with two or three
different core materials
Conical
Ogival
Nose or frontal portion
Hemi-spherical
Flat
1.2.3 Target Types
In 1978, Backman and Goldsmith divided target types into four groups, which are
given in Table 1.3. Their classification of target types is based on the ratio of the
thickness (ht ) to the diameter (D) of the projectile.
3
Table 1.3: Classification of target [1]
Target Classification Ratio of ht /D
Thin <1
Intermediate 1-8
Thick 8-12
Semi-infinite > 12
1.3 The ballistic response of materials
When the impact velocity of a projectile is between 500m/s and 2000m/s global de-
formation characteristics of target material becomes less important. Relatively large
and localized deformations are observed in the neighborhood of impact area (about
3-4 times of bullet diameter) while the rest of the target remains almost undeformed.
The deformed shape of a target after impact is shown in Figure 1.1 for a better under-
standing.
Figure 1.1: Deformed shape of target after strike and perforation process [4].
4
Figure 1.2: Geometry and dimensions (in mm) of 20mm ogive-nose hardened steel
projectile(left) and 7.62mm APM2 bullet (right) [4].
Børvik et al. [4] studied the perforation process of aluminum 5083-H116 armor plate.
Figure 1.2 shows the geometry and dimensions (in mm) of 20mm ogive-nose hard-
ened steel projectile (left), and 7.62mm APM2 bullet (right) [4]. Plates of various
thicknesses were used in this study for a range of 480-950m/s striking velocities. A
comparison was made among ogival, nose-shaped, and 7.62mm APM2 projectiles in
terms of perforation processes. It was shown that the effect of lead brass and jacket are
negligible for residual velocity values compared to other nose-shaped bullets. Resid-
ual velocity is velocity value after the finishing of perforation.
Table 1.4 shows the upper and lower boundaries of strike velocity values concerning
their strain rate levels. These velocity values can be taken as reference points for the
classification.
The impact response of a bullet is directly related to stress waves which are generated
during the impact. These stress waves are functions of parameters such as veloc-
ity, geometry, strain rate, material, localized plastic flow, etc. The response time of
collided parts is in the range of microseconds for all ballistic impacts.
5
Table 1.4: Response of materials under the ballistic impact [2]
Strain rate s−1 Impact velocity(m/s) Effect or material response
0 - 100 < 50 Primarily elastic and some local plasticity

0 2
10 - 10 50-500 Primarily plastic
Viscous – material strength
102 - 104 500-1000
still significant
Fluid like behavior in materials; pressures approach or
104 - 105 1000-3000
exceed material strength; density a dominant parameter
Hydrodynamic behavior - material
105 - 107 3000-12000
compressibility not ignorable
107 - above >12000 Explosive impact-colliding solids get vaporized
1.4 Penetration Mechanics
The definition of penetration is the entrance of the missile or penetrator into a tar-
get after the collision, without exiting the target body. Bouncing back of a bullet or
burying into a target case is also included in the ballistic impact. During a collision,
the energy of the projectile is transferred to strain energy, energy loss due to friction
(i.e., heat, sound, light) of collided parts. The determination of the kinetic energy of
a projectile is essential for penetration analysis. There is not an easy way to measure
the amounts of these energies. To simplify the analyses, some assumptions/approxi-
mations about the energies of collided parts must be made.
Penetration process can be grouped into four stages [1]:
1. Transient
2. Primary penetration
3. Secondary Penetration
4. Recovery
The first stage is the initial contact of the impactor with the target interface. The
maximum value of pressure is attained in the transient stage of impact. In the primary
6
penetration stage, the impactor starts to use its driving force into a target, and material
starts to behave fluid-like material locally at the interface. The secondary penetration
phase starts with the end of the deformation of the bullet.The deformation of the target
by the projectile’s kinetic energy is not considered in this phase. Instead, a shockwave
which is distributed by the impactor deforms the target. The last stage of penetration
is the recovery phase in which the shrinkage of the crater takes place.
The conservation of energy law holds during a collision. The impact energy of the
bullet transforms into the internal energy and the kinetic energy of the target together
with some forms of energy losses.
Etrans = Eplate + KEplate + IEeroded + KEeroded (1.1)
In the above equation where Etrans denotes the total energy transmitted, subscript
‘plate’ refers to the energy of the plate; and the superscript ‘eroded’ denotes the en-
ergy in the eroded mass [1]. Eplate represents the strain energy of the plate, IEeroded is
the internal energy of the eroded parts and KEplate is the kinetic energy of the plate.
1.5 Limit Velocities
The impact velocity and the thickness of the target determine the failure mode of
collided parts. Limit velocities are grouped into three, which are the elastic limit
velocity, the plastic limit velocity, and the hydrodynamic limit velocity.
1.5.1 Elastic Limit Velocity
The elastic limit velocity, VEL , is the maximum velocity, which creates only elastic
deformation both in the projectile and the target. This situation can only happen at
very low impact speeds. In case of a flat nose-shaped bullet, or a planar contact
surface at the interface of impact, compressive stress equal to the yield stress σyc is
7
generated, and the elastic limit velocity is computed as given in Eq. (1.2)
σyc [ρt c1t + ρp c0p ] σyc [Z1t + Z0p ]

VEL = = (1.2)
(ρt c1t ρp c0p ) Z1t Z0p
where,
s s
λ + 2G 0.5E E
c1 = ,G = , c0 = (1.3)
ρ 1+ν ρ
In Eqs. (1.2) and (1.3), c1 and c0 represent the wave velocities Z1 and Z0 are impedances
of in longitudinal and transverse directions and λ is the Lame’s constant. E is Young’s
modulus, G is the shear modulus, ν is the Poisson’s ratio, ρ is the mass density, and
subscripts p and t denote the projectile and the target, respectively [11].
1.5.2 Plastic Limit Velocity
Plastic limit velocity, VPL (see Eq. (1.4)), is the maximum velocity, which leads to a
perdurable deformation on both target and impactor. For a steel with a compressive
strength of 152MPa and and tensile strength of 250MPa, the VEL is 3.75m/s and the
VPL is 90m/s [1]. At the plastic limit velocity, plastic stresses are less than hydrody-
namic stresses.
r
σy
VP L = (1.4)
ρ
Figure 1.3 shows an impact case without projectile drilling into a target. In this case,
two types of failure modes are activated because of plastic deformations. The first
mode occurs at the contact interface between the bullet and the target, which is called
bulging. In this mode, the contact surface of the target takes the form of a projectile
nose shape. The second mode is activated by bending, referred to as dishing, which
might widen far from the impact interface [5].
8
Figure 1.3: Permanent deformations of thin target element showing bulging and dish-
ing [1].
1.5.3 Hydrodynamic Limit Velocity
The hydrodynamic limit velocity, VHL defines this maximum strike velocity just be-
fore shock waves occur both in the projectile and the target. The whole deformation
process starts with plastic limit velocity and ends with hydrodynamic limit velocity.
This limit velocity can be related to the bulk modulus, K, of the material as given in
Eq. (1.5)
s
K
VHL = (1.5)
ρ
It is considered that, above the hydrodynamic speed limit, the compressibility of the
solid decreases, and therefore, shock waves occur. Also, at sufficiently high, impact
velocities (e.g.,> 3VHL ) [11], phase changes, vaporization, or even impact explosions
may be expected [12], [13].
9
1.6 Ballistic Limit and Aerial Density
Ballistic limit evaluation is a fundamental concept in order to determine the perfor-

mance of an armor. A good quality armor design has a high ballistic limit. The
ballistic limit is defined as an average of the fully perforating velocity and the partial
penetrating velocity.
Figure 1.4: Definitions of perforation and partial penetration for defining the ballistic
limit [5].
There are several measurement criteria for ballistic limit determination. Most com-
monly used criteria are:
1. Army ballistic limit
2. Protection ballistic limit
3. Navy ballistic limit
Above listed criteria are illustrated in Figure 1.4 [5]. The main objective of the bal-
listic limit tests is to determine the velocity of the projectile, which fails to perforate
in the target. This velocity can be obtained either using physical principles (i.e., con-
servation law and material relations) or probabilistic methods which are based on an
immense database of striking velocities. Details on deterministic and probabilistic
methods for ballistic determination limit can be found in [5].
10
1.7 Literature Survey on ballistic FEA combined ANN studies
One of the most critical subjects in the defense industry is the ballistic protection.
Numerical methods have been widely used for the determination of the ballistic limit.
Furthermore, analytical and empirical solutions are widely used for ballistic prob-
lems. Analytical models and empirical equations are generated from numerous ex-
periments and are employed as a solution method for ballistic collision problems [14].
Since empirical formulations are case-specific, they are generally not valid for differ-
ent cases. They must be derived for each problem, and this derivation is complicated
and costs too much time.
Analytical models can be used on various problems directly. They are convenient
for ballistic problems. However, their derivations involve simplifications, which then
leads to differences in outcomes [2].
In a conclusion, one can see that determination of the impact characteristics is a very
complicated task, and generally, it cannot be determined using the analytical and
empirical. At this point, FEA is an alternative tool to model the ballistic impact. With
the help of the developments in FEA and computational resources, these complex
impact phenomena could be predicted accurately in less time.
Although the FEA is a powerful tool, it might sometimes require high computational
power, depending on the problem. The number of numerical simulations and ballis-
tic experiments can be reduced by creating an Artificial Neural Network (ANN). The
neural network uses statistical non-linear regression-based calculation, and it is one of
the most robust techniques in computer science. The method has been used in many
fields other than engineering to simulate complex systems. The power of an ANN
solution comes from the learning capability of the computer. The neural network sys-
tem inspired by the human brain system which can imitate the way that humans learn,
and after the end of the training, similar problems can be solved without the need for
any preparation of the system set up for a similar system. In this section, a litera-
ture review of ballistic impact simulations and artificial neural network researches is
given. An investigation was conducted by Arndt et al. [15] to combine the FE method
with a neural network for a groundwater engineering problem. Numerous numbers
11
of experiments and FE analyses were performed, and results were used for the ANN
training.
Chan et al. [16] are studied a similar integration of the FEM and the ANN for metal-
formed product design. Examining the essential design parameters for metal forming
applications, evaluation of design performance is done with FE simulations. The
metal forming application contains a large number of parameters which are tooling
design, material properties selection, metal formed part design, etc. Therefore, doing
the parameter sensitivity analysis for each design parameter by FE simulations is a
very time-consuming task to find the optimal design. In this study, the design param-
eters which are geometric parameters of the punch, are used as input parameters. The
effective stress on punch and the maximum punch load to make forming applications
are taken as an output parameters for FE simulations. A sufficient amount of analyses
were performed and results were used to create the ANN setup. The optimum design
prediction time is reduced by about 60% with the help of ANN training.
Additionally, Shabani et al. [17] used ANN to find mechanical properties of Al-
Si(356) material which are the yield strength, the ultimate tensile strength, and elon-
gation percentage for the solidification process. Experiments and FE simulations
were performed for the solidification of A356 alloy. Simulations were performed to
create input data for the cooling rate and the temperature gradient. The ANN tool was
used to predict the material yield strength, the elongation percentage, and the ultimate
tensile strength. Also, Haghdadi et al. [18] worked on the A356 aluminum alloy to
predict the material behavior under various thermomechanical conditions by using
the ANN and the FE method. Results showed that the ANN tool is such a robust tool
to predict the flow behavior of cast A356 material.
Furthermore, Hambli [19] developed a procedure for the simulation of trabecular

bone adaptation process with a FE model and ANN model. The FE model was pre-
pared at the macroscopic level, and neural network training was used for mesoscale
predictions. Applied stress and boundary conditions were taken as input parame-
ter for the neural network setup and the averaged bone properties are taken as output.
Results of the FE-ANN method found beneficial because remodeling of bones is time-
consuming. Gudur et al. [20] also studied the ANN in combination with the FEM. In
12
this study, the computational time of cold flat rolling simulation was decreased with
the help of the ANN. Input data of the problem was taken from both experiments
and FE solutions. Artificial neural network-assisted FEM provides a good agreement
with reduced computational time and therefore is appropriate for on-line control or
optimization. Haj-Ali et al. [21] used ANN to define the constitutive behavior of ma-
terials from nanoindentation tests. The load indentation tests were only conducted for
the monotonic loading scenario to build the ANN model. Nanoindentation tests were
performed on a silicon (Si) substrate with and without a nanocrystalline copper (Cu)
film. Inverse neural network (back-calculate) run was trained for material parame-
ters for different copper materials. Comparison with literature showed close results
with the FE-ANN approach. In light of the studies mentioned above, the FE-ANN
approaches can significantly reduce the simulation times.
1.8 Aim of this study
The determination of the ballistic performance of a vehicle armor is based on several

parameters such as the armor thickness and the strike velocity. The FEM is a powerful
tool to determine ballistic characteristics; however, the FE simulations of all armor
and strikes velocity combinations would be computationally costly. To avoid this, a
promising approach is to combine the FEM and ANN to predict the ballistic limit of
different armor configurations. Creating a well trained ANN with a sufficient amount
of numerical analyses can save computational power and time drastically. This work
aims to perceive whether the outcome from a ballistic impact can be predicted with
the help of material testing, FE simulations, and Artificial Neural Network (ANN).
Experimental and numerical studies were conducted to determine the impact response
of high strength aluminum armor at level four protection level of STANAG 4569 [3]
standard. A high-speed camera was used for calculations of the residual projectile
velocities and projectile output images. An artificial neural network was trained using
the tests and numerical results to create a simple ballistic limit tool.
13
14
CHAPTER 2
NUMERICAL MODELLING
High-velocity impact problems are highly non-linear problems, which include non-
linear material behavior, contact nonlinearity, large deformation of collided parts and
nonlinear boundary conditions. Creating a numerical model for a high-velocity im-
pact case requires several parameters to simulate the real impact process in a computer
environment. These parameters are mostly related to the material model, the equation
of state and the failure model employed. In Figure 2.1 an overview of key ingredients
that govern computational ballistics are given. The details of numerical models and
numerical formulation are given in Chapter 3.
Figure 2.1: Numerical inputs and formulations governing computational ballistics [1].
The FEA starts with Computer-Aided Design (CAD) software. Firstly, projectile and
target solid geometry models are generated using relevant design tools. Secondly, a
mesh is generated for the solution domain. Thirdly, a suitable material model (details
15
are given in Chapter 3) is employed and boundary conditions are defined. Then, the
solution process can be initialized. During the solution stage, the conservation equa-
tions of mass, momentum, energy and entropy are solved. All the material model,
damage model and equation of states are integrated into the formulations and these
equations are solved for each time-step for each element in the domain. Finally,
when the case is solved for a sufficient time, using any post-processing tool, the re-
sults (such as strain, stress, velocity, etc.) can be evaluated and visualized. In this
study, LSDYNAr is used for the solution phase and LS-Prepost is utilized as a post-
processing tool.
2.1 Numerical modeling approaches
Various approaches are used for computational ballistic analyses. The most com-
monly used approaches are the Lagrangian approach, the Eulerian approach, the Ar-
bitrary Lagrange Euler (ALE) approach and the Smooth Particle Hydrodynamics. In
this section, these approaches are briefly explained.
2.1.1 Lagrangian approach
When the Lagrangian solver is used, grid points are fixed on the body, which is ana-
lyzed. The Lagrangian solver, therefore, calculates the motion of elements of constant
mass [19]. As the body deforms, the grid points move together with the material and
the finite elements distort. Figure 2.2 and 2.3 shows the initial mesh and the deformed
Lagrangian mesh.
For the problems with small deformations, the Lagrangian approach gives very accu-
rate results because all the material points are tracked. Therefore, mass conservation
is precisely satisfied. Compared to other approaches (Eulerian and ALE), the La-
grangian approach requires fewer computational calculations per cycle. Therefore,
it is computationally cheaper. However, when extensive deformations occur in FE
analysis, elements can be highly distorted during the deformation process. Highly
16
distorted elements reduce the accuracy on one side and require smaller time incre-
ments that result in higher computational cost on the other side. In order to prevent
this problem, the material failure definition and the material model, which are em-
ployed in FE analysis, must be chosen carefully. With the help of accurate material
failure modeling, unrealistic solutions and small times steps can be prevented within
the Lagrangian approach.
Figure 2.2: Deformed lagrangian elements [6].
Figure 2.3: Lagrangian mesh and modeling [1].
17
2.1.2 Eulerian approach
The Eulerian approach is used mostly for fluid-like material behavior modeling under
very large deformations. For example, the bullet and armor materials undergo such
deformations that the material can reach the melting point, which makes the Eule-
rian formulation very suitable. The essential advantage of this approach is that the
arbitrary material motion is permitted.
Figure 2.4: Eulerian mesh [1].
Figure 2.5: Eulerian mesh [2].
In Figure 2.4 and Figure 2.5 an Eulerian domain is illustrated. In Figure 2.5 grid
points remain stationary in space during the simulation. Material flows from one ele-
ment to the other throughout the mesh, where the mass is conserved. The distortion of
the element is prevented by transferring material through the mesh domain. Therefore
the Eulerian approach is suitable for problems which involve very large deformations.
As given in Figure 2.5, the projectile and target flow do not distort the Eulerian mesh
18
while the deformation of material can still be observed.
The Eulerian approach is computationally expensive compared to the Lagrangian ap-

proach. The computational cost difference between these two approaches comes from
the material transfer from one element to another and recording the element data for
each material that exists within the Eulerian approach. Every element can contain
one to many different materials, and the location of each material needs to be traced.
Additionally, tracking material transfer interfaces, plasticity, and failure modeling are
difficult tasks in the Eulerian approach.
2.1.3 Arbitrary Lagrange Euler coupling
The Arbitrary Lagrange Euler (ALE) approach is a hybrid method which combines
the Lagrangian formulation with the Eulerian approach. Mine blast simulations,
sloshing analysis and other fluid-structure interaction problems are commonly mod-
eled with the ALE approach. The ALE method can be explained by the help of Figure
2.6.
Figure 2.6: Difference between Lagrangian, Eulerian and ALE [6].
In Figure 2.6 overlapping meshes are given. One is a background mesh which can
move arbitrarily in space (i.e., a Lagrangian Mesh), and the other is attached to the
material which “flows” through the former moving mesh. This can be visualized in
two steps. First, the material is deformed in a Lagrangian step just like the Lagrangian
19
formulation. Then, the element state variables in the “Lagrangian elements” (red) are
remapped or advected or distributed back onto the moving (background) reference
ALE mesh (green). The main difference between the pure Eulerian approach versus
the ALE method is different amounts of material being advected through the meshes
due to the reference mesh positions [6].
Figure 2.7: Time integration loop of ALE approach [6].
In Figure 2.7 time integration loop of the ALE method is shown. Firstly, Lagrangian
time derivatives of displacement are calculated, and the acceleration and velocity his-
tory variables are updated. Afterward, the relative motion between the mesh and the
material is computed and the acceleration and velocity history variables are updated
once again. Compared with the Lagrangian approach, all of these calculations are
done a minimum twice as much. After the Lagrangian step, the ALE element calcu-
lation loop is started.
2.1.4 Smooth Particle Hydrodynamics
The Smooth Particle Hydrodynamics (SPH) represents a mesh-free solution which

can also be used for cases that involve large deformations. In the SPH numerical
discretization of the continuum is achieved with a set of nodes without the node to
node connecting mesh. Comparison of a Lagrange mesh and SPH model is shown in
Figure 2.8. Within the SPH formulation, the particles are free movable points with
20
a fixed mass, and they all have coherence only through an interpolation function.
The kernel estimate allows us to describe the conservation of mass, momentum, and
energy in terms of interpolation sums [22]. The disadvantage of the SPH approach
is its limited accuracy in modeling high energy events. Moreover, the calculation of
contact forces between SPH particles is problematic. Due to the high oscillations of
interface forces between SPH particles low-level accuracy results can be obtained.
Figure 2.8: SPH (left) and Lagrange (right) representation of Fragment Simulating
Projectile
2.2 Integration methods
There are two types of integration methods used in numerical simulations:
1. Implicit time integration
2. Explicit time integration
The implicit time integration is widely used in simulations that involve low strain rates
(less than (102 s−1 ) and low-velocity values. However, the implicit time integration
provides unconditional numerical stability for the solution. Impact problems must be
examined in detail for the selection of the type of solver. The equation of motion (2.1)
21
is used for both integration methods to advance in time.
Mü + Cu̇ + Ku = F(t) (2.1)
On the left side of (2.1), M represents the mass matrix, C is the damping coefficient
matrix, and K is the stiffness matrix. Nodal displacement vectors are given with u.
Therefore, u̇ and ü are the nodal velocity and nodal acceleration vectors. On the
right side of the (2.1), F(t) is the external nodal forces. At any point in time, (2.1)
is solved for nodal accelerations from nodal displacements, velocities and external
forces respectively.
Explicit time marching method is frequently used for any impact problems. Imple-
mentation of an explicit finite element code is quite easy and requires low memory
storage. However, explicit integration is conditionally stable, which is its major dis-
advantage. To explain it further, there is a restriction for time step value, which cannot
be exceeded during the solution process. The critical time step (∆tcr ) value is con-
trolled by the size of the smallest element in the FE model, and the largest natural
frequency of model. The critical time step is found as shown in (2.2).
r
`s ρ 2
∆t = = `s ≤ ∆tcr = (2.2)
c E wmax
In (2.2), `s represents the minimum edge length of the element, ρ represents the den-
sity, wmax is the highest eigenfrequency of the system and E represents Young’s mod-
ulus of the material. Based on the principle that time step is proportional to the square
root of mass density (sound speed inversely proportional to the square root of mass),
an increase in mass density (or adding element mass) will increase element time step
and thereby reduce the computational time. Robustness of the explicit time integra-
tion is hinged upon to the value of the critical time-step of the FE model [23].
22
2.2.1 Implicit integration
Within the implicit integration method, equations of motion are evaluated at t + ∆t

time. However, the nodal accelerations and displacements at t + ∆t are unknown.
Expanding the Taylor series for displacement equation at third order:
∆t2 üt ...

ut+∆t = ut + ∆tu̇t + + βt ∆t3 u t+∆t (2.3)
2
∆t2 ...
u̇t+∆t = u̇t + ∆tüt + γT u t+∆t (2.4)
2
where βt and γT are constants for Taylor series expansion and ∆t is the time step size.
To be able to determine the third order derivative of displacement at time t + ∆t, a
linear change of acceleration assumption at a given time step is made:
Figure 2.9: The linear assumption for acceleration
... ü − üt
u t+∆t = t+∆t (2.5)
∆t
Using (2.5) in both Eqs. (2.3) and (2.4) will give the Newmark’s solution for the
23
ut+∆t and u̇t+∆t at time t + ∆t:
∆t2
ut+∆t = ut + ∆tu̇t + (1 − 2βT ) üt + βT ∆t2 üt+∆t = u∗t + βT ∆t2 üt+∆t (2.6)
2
u̇t+∆t = u̇t + (1 − γT ) ∆tüt + γT ∆tüt+∆t = u̇∗t + γT ∆tüt+∆t (2.7)
where u∗t and u̇∗t used for a compact representation. Setting the βT and γT constants
zero, changes the Eqs. (2.6) and (2.7) into an explicit form. Evaluation of the dis-
placement and the velocity dependent only on one unknown at time t Eqs. (2.6) and
(2.7) unconditionally stable. Hughes and Taylor [24] have proven that this integra-
tion method is unconditionally stable if 2βT ≥ γT ≥ 1/2 and conditionally stable if
γT ≥ 1/2 and βT ≤ (1/2)γT . Usually, these constants are chosen as βT = 1/4 and
γT = 1/2 , which is known as the average acceleration method [24]. Using these
constants values in Eqs.(2.6) and (2.7) and insertion of (2.6) and (2.7) into (2.1) give:

1 1
M + C∆t + K∆t üt+∆t = Füext
2
− Cu∗t − Ku∗t (2.8)
2 4 t+∆t
Rewriting (2.8) gives:
M∗ üt+∆t = Füext
t+∆t
− F∗ = Fresidual
t+∆t (2.9)
where F∗ = Cu∗t + Ku∗t and Fresidual

t+∆t is the subtraction of the forces seen in (2.9).
The nodal acceleration values are found by solving the equation set with the Newton-
Raphson iteration method. After the calculation of nodal accelerations, nodal veloci-
ties and displacements are found from (2.6). This solution process is repeatedly done
for each node for every time step until the solution terminates [24].
2.2.2 Explicit integration
The explicit integration solution uses the central difference method for most applica-
tions. Nodal accelerations at t + ∆t are predicted at t time from (2.1). Nodal acceler-
24
ation is assumed to be constant during the time step, and its value is calculated from
the initial state of equilibrium. Therefore, the time step must be sufficiently small,
considering the linear change of acceleration (see implicit methods). Calculation of
the acceleration is given in (2.10):
üt+∆t = M−1 Fresidual

t (2.10)
Using the central difference method displacements and velocities are calculated as
follows:
u̇t+∆t = u̇t + ∆tü (2.11)
ut+∆t = ut + ∆tu̇t+∆t (2.12)
The explicit integration method is only stable if the time step is smaller than the so-
called critical time step (conditionally stable). The critical time step is related to the
highest eigenfrequency of the system and the mesh size of the model. Generally, the
time-steps for explicit codes are 100-1000 times smaller than implicit codes. How-
ever, since the values at the next time step are computed directly, the cost of each time
increment is relatively low. Explicit methods have a greater advantage over implicit
methods if the time step of the implicit solution has to be small, and if the model size
is large [22].
2.3 CONCEPTS IN NUMERICAL SIMULATION
In this chapter, simulation concepts such as hourglass deformation and contact defi-
nition methods are discussed.
25
2.3.1 Hourglass Deformation and Hourglass Damping
Hourglass deformation is a mesh-based problem which generally occurs in impact

simulations employing a Lagrangian approach. Even though the hourglass deforma-
tion is not the primary problem, it has to be handled in order to reduce the unrealistic
behavior of collided parts.
2.3.1.1 Hourglass Deformation
The source of the hourglass problem emerges from the reduced integration (one inte-
gration point). One integration point element may produce zero energy deformation
modes (hourglass mode) in Lagrangian elements.
Figure 2.10: (a) An undeformed and deformed (b) one-point integration element, (c)
an undeformed and deformed (d) full integration element
The hourglass mode is a deformation mode that produces zero strain at all integration
points. Figure 2.10 shows a comparison between one point integration element and a
fully integrated element and their deformation characteristics. Blue points represent
the integration points, and 1,2,3,4 are the node numbers of the element. f1, f2, f3,
and f4 are hourglass penalty forces. In Figure 2.10 (a) and (b), one point integration
26
element and its deformation are shown schematically. As it is shown in Figure 2.10
(a) during the deformation of a one point integration element, no strain is generated
at the integration point. If the integration point (Gauss point) senses no strain under a
certain deformation mode, the resulting element stiffness matrix will have no resistant
to that deformation mode. Therefore, the element shows a very soft response to this
deformation mode. On the other hand, the strains at integration points, are not zero
for a fully integrated element because the change of length can be taken into account
over integration points as seen in Figure 2.10 (c)-(d).
2.3.1.2 Hourglass Damping
A penalty stiffness is predefined by the user in order to prevent the hourglass modes.
Consider the one integration point element in Figure 2.10 again. One needs to in-
troduce a nodal force field that opposes the hourglass component of nodal velocities.
These forces must thus be opposed to the hourglass base vector. Viscous hourglass
forces are defined as
   
f1 1
   
 f 
 2 
 −1 
 = hm au̇hx  (2.13)
 
 
 f3   1 
   
f4 −1
cρ 2
a= (tA) 3 (2.14)
4
where hm is a dimensionless user-defined penalty factor (default value is 0.1), u̇hx is
the nodal velocities and a has the dimension of viscosity (Ns/m) and given as (2.14).
In (2.14) A represents the area of the element and t is time. Using (2.13) a viscous
hourglass damping force field is generated.
27
Alternatively, stiffness hourglass forces can be used which are defined as follows:
   
f1 1
   
 f  −1
Z
 2  Et  
 = hm u̇hx   dt (2.15)
 
8

 f3   1 
   
f4 −1
In (2.15) stiffness based hourglass damping formulation is shown. E represents Young’s

modulus of the element and t represents the time. The viscous hour glassing only
stops the hourglass mode from being further developed. On other the hand, the stiff-
ness hourglass damping forces push the element back towards its undeformed config-
uration. The stiffness based hourglass considerably stiffens the behavior of the solid
elements. Since the hourglass forces cannot be expected to compensate exactly the
missing element stiffness, an accurate result requires low hourglass energy compared
to deformation energy. Ideally, the hourglass energy should be less than ten percent
of the internal energy.
2.3.2 Contact-Impact Methods
Modeling the interaction between the impactor and the target is another essential point
of numerical modeling. Using implicit codes, nonlinear contact modeling is a very
challenging task. On the other hand, explicit codes are so powerful and useful for
nonlinear contact problems.
Considering explicit codes, there are three distinct methods for defining contacts:
1. the kinematic constraint method
2. the penalty method
3. the distributed parameter method
On the contact interface, one side of the contact is named as the slave surface, and
another side is called the master surface. The selection of the master surface and the
28
slave surface is arbitrary in case of modeling symmetric contact. In other cases, there
are several rules of thumbs to determine master and slave surfaces (for details see
2.3.2.1). Since the most commonly used contact definitions are the kinematic con-
straint method and the penalty method, below, only these methods will be discussed.
2.3.2.1 Kinematic Constraint Method
The kinematic constraint method is also known as the nodal constraint method. In
this method, the degree of freedom constraints is imposed on the global equations by
a transformation of the nodal displacement components of the slave nodes along with
the contact interface [7]. The transformation eliminates normal degrees of freedom
of slave nodes. Additionally, normal forces are distributed from slave nodes to master
nodes during this transformation process.
In Figure 2.11, two surfaces, one master and one slave (x nodes belong to the master
surface) with different sizes of meshes, are illustrated. The slave surface has a coarser
mesh compared to the master surface. As a result of this, nodes of the master surface
can easily penetrate the slave surface without any resistance. This penetration can lead
hourglassing to elements in contact and reduce the accuracy of the model. Therefore,
the slave contact surface should have a finer mesh to the master side.
Figure 2.11: Kinematic constraint method [7]
29
2.3.2.2 Penalty Method
This method is based on inserting an elastic compression-only spring in the normal

direction between the penetrating element nodes and the contact surfaces. The spring
force is calculated from the thickness and bulk modulus of the element for each slave
and master segment. Three types of penalty algorithms are used, which are:
1. The standard penalty formulation is employed when the parts in contact have
similar stiffness values.
2. The soft constraint penalty formulation is widely used whenever there are
significant differences of stiffness, density, mechanical properties between parts
in contact.
3. The segment-based penalty formulation uses a slave-master segment ap-

proach different from the classical slave node-master segment approach.
In this work, standard penalty formulation is employed for contact interactions.
30
CHAPTER 3
MATERIAL MODELLING AND FAILURE MODELS
Selection of the right material model and determination of the real material constants
from experiments are the most challenging parts of FE simulations. LSDYNA® soft-
ware includes two hundred material models. For ballistic problems, suitable material
models are explained in this chapter. Additionally, the most commonly used theoret-
ical failure criteria and failure models are explained
3.1 Gruneisen parameter
Mie-Gruneisen equation of state comes from statistical mechanics which considers

energies of atoms. The equation state (EOS) is a function of pressure p, density ρ, and
internal energy e of the material, and it defines the volumetric response (or hydrostatic
response) of a material. The Gruneisen parameter can be regarded as the change in
pressure through an increase of specific internal energy at constant volume [25].
The Gruneisen parameter is commonly represented with Γ symbol, and the definition
is:

1 δp
Γ= (3.1)
ρ δe v
δp

To evaluate the Gruneisen parameter in measurable properties, δe v
can rewritten
31
as:
δp

δp δT v
= δe
(3.2)
δe v δT v
δe δp

where δT v
= CV is the specific heat per unit mass. The term δT v
can be derived
using the Maxwell’s relation [26]

δp δp δV
= (3.3)
δT v δV T δT p
δp δV

where δV T
= KT is the isothermal bulk modulus and δT p
= β is the coefficient
of the volumetric expansion. In the case of an isotropic solid element, volumetric
expansion can be used as proportional to linear thermal expansion, α as β = 3α.
Insertion of the parameters CV , α, KT in Eq.(3.2) gives:

δp 3αKT
= (3.4)
δe v CV
Then using (3.4) and (3.3) a physically measurable Gruneisen parameter is obtained
as
3αKT
Γ= (3.5)
ρCV
The Gruneisen parameter is taken as Γ = 2.0 for most ambient conditions. Further-
more, Γ is assumed to be temperature independent and, Γρ is taken as a constant for
a given solid material for a wide pressure range [27].
32
3.1.1 Shock Rankine-Hugoniot conditions and relation
Figure 3.1: Schematic view of a shock front (line C) propagating through a compress-
ible material [8].
The Rankine-Hugoniot equations involve the conservation of mass, momentum, and

energy equations which are shown in Eqs. (3.6), (3.7) and (3.8).
ρ0 us = ρ (us − up ) (3.6)
Over time increment dt change of momentum can be written as ρ0 us dtup , and pres-
sure impulse is p − p0 dt. Therefore, the conservation of momentum can be written
as
p − p0 = ρ0 us up (3.7)
In (3.7) ρ0 us , is known as the shock impedance.
After time increment dt the kinetic energy of the plate is 21 ρ0 us dtu2p , the change in
the internal energy is ρ0 us dt(e − e0 ), and the work done by the pressure is pup dt.
Therefore, the conservation of energy can be written as
1
ρ0 us dtu2p + ρ0 us dt(e − e0 ) = pup dt (3.8)
2
A detailed version of the derivation of these equations is given in Zukas et al. [3].
33
Excluding us and u from (3.8) gives Rankine-Hugoniot relation as:
1
(e − e0 ) = (p + p0 ) (v0 + v) (3.9)
2
1 1
where ρ0
= ν0 and ρ1
= ν1 represent the uncompressed and the compressed specific
volumes, respectively. Equations (3.6) - (3.9) are generally referred to as jump con-
ditions. These conditions should always be satisfied at the shock front for all impact
applications.
3.1.2 Mie-Gruneisen equation of state
The Mie-Gruneisen equation of state (EOS) comes from statistical mechanics and can
be expressed with the Gruneisen parameter given in (3.5):
δp = Γρδe (3.10)
where Γρ is taken as a material constant as already mentioned before in Section 3.1
The difference in the internal energy is proportional to the difference in pressure at

0K. One can write (3.10) as:
(p − pref ) = Γρ(e − eref ) (3.11)
This reference state can easily be rewritten in other forms. The most common known
form of the Mie-Gruneisen is given in [28], [29], [30]. The full description of the
Mie-Gruneisen EOS comes from Hugoniot pressure and energy equations which can
be found in [7] in detail.
3.2 Plasticity
High-velocity impact problems generally include large deformation of both the im-
pactor and the target. Therefore, the impact process is dominated by plasticity. For
the low-velocity impact cases, target and impactor are first deformed elastically, and
34
then they return to their original state. If the distortion is too significant, the projectile
and the target will exceed their elastic limits and deform plastically. Whether plas-
tic deformations occur or not is generally controlled by the von Mises yield criterion
which is given in (3.12)
2
F = σeq − σy2 (3.12)
In (3.12) F is the yield function of the material and σy is the yield stress. The von
Mises stress, σeq , is given by:
1
q
σeq = (σ1 − σ2 )2 + (σ2 − σ3 )2 + (σ3 − σ1 )2 (3.13)
2
where σ1 , σ2 , σ3 represent the principal stresses. The restrictions proposed by Kuhn-
Tucker rule, which are shown in Eqs. (3.14) and (3.15), indicate whether the stress is
inside or on the yield surface.
(F < 0) ∨ (F = 0 ∧ Ḟ < 0), Elastic (3.14)
(F = 0 ∧ Ḟ = 0), Elastic − Plastic (3.15)

Generally, for metals with elastic-perfectly plastic behavior, yielding characteristics
are determined by using the Von-Mises criterion. The von-Mises criterion is suitable
for high strength materials, such as armor steels and the high strength aluminum
alloys [27]. The collision causes large deformations in both the projectile and the
target. Additionally, for materials which go through the large deformation in high-
velocity impact events, the progress is an adiabatic process. During the deformation
process for impact scenarios, temperatures can reach 200-500◦ C, which is around the
melting point of few metals, as given in [1]. The temperature directly affects the
thermal softening/hardening behavior of the materials. Therefore, the effect of the
strain rate and the thermal effects must be included for high-velocity impact problems.
There are many material models for the high-velocity impact behavior of metals
such as the kinematic hardening model, the thermo elastic-plastic material model,
the power-law isotropic material model, Johnson-Cook material model, the Zerilli-
Armstrong material model, and the Steinberg-Guinan model. In this study, the Johnson-
Cook material model is employed both for the projectile and the target. Therefore,
only the Johnson-Cook material model will be explained in the next sections.
35
3.2.1 Johnson-Cook Model
The Johnson-Cook model was introduced in 1983. It is an empirical material model,

which is prepared by a sufficient amount of test data at different strain rates and
temperatures. This is the appropriate model for problems, including high strain rate
range and extensive temperature changes. For each phenomenon (strain hardening,
strain rate hardening, and thermal softening), an independent term is created. By
multiplying these terms, the flow stress is defined as a function of the effective plastic
strain, strain rate, and temperature [1]. The constitutive relation of the Johson-Cook
model is shown in (3.16).
σy = A + Bεnp [1 + C ln (ε̇∗ )] [1 − (T∗ )m ]

(3.16)
with
T − Tref ε̇p
T∗ = , and, ε̇∗ = (3.17)
Tm − Tref ε̇0
where A is the yield stress, B is the strain hardening coefficient, n is the strain harden-
ing exponent, εp is the effective plastic strain, (ε̇p ) is the effective plastic strain rate,
ε̇0 is the reference strain rate used in quasi-static tensile tests, ε̇∗ is effective plastic
strain rate, C is the strain rate coefficient, Tm is the melting temperature and Tref is
the reference temperature.
The parameters in J-C model (n, C, and m) are determined from an empirical curve fit
of dynamic yield stress. Firstly, the second and third brackets of (3.16) are eliminated
to determine n. In order to eliminate second bracket, ε̇p is chosen as the strain rate of
interest (for example maximum strain rate in test data), and ε̇0 is set to the same value
of ε̇p so that ε̇∗ = 1s−1 . Hence, the second term of the J-C equation is now equal
to one. To eliminate the third bracket, Tref is set to test temperature which yields to
T∗ = 1. After eliminating the second and the third terms, n can be calculated from
the slope of a plot between ln(σy − A) vs ln(Bεp ). Secondly, to determine C, only
the third bracket of (3.16) is eliminated by setting again Tref to test temperature. This
reduces (3.16) to (3.18):
σy = A + Bεnp [1 + C ln (ε̇∗ )]

(3.18)
36
From dynamic yield stress for every fixed plastic strain at various strain rates, C is the
σy ∗
slope of ln([ (A+Bε n ) ] − 1) vs ln(ε̇ ). Thirdly, to determine m, only the second term is
p
eliminated. To do this, as already mentioned above, ε̇∗ is arranged to 1s−1 . After this
elimination (3.16) reduces to (3.19):
σy = A + Bεnp [1 − (T∗ )m ]

(3.19)
σy ∗
In (3.19) m is the slope of the ln(1 − [ (A+Bε n ) ]) vs ln(T ) plot. Finally, the material
p
constants C and m are determined using the least-square method [31].
3.3 Material Failure Modelling
Evaluation of a failure in ballistic problems is a very complicated task. The impact

process contains so many complexities like large deformation, high strain rates, high
pressures, and high temperatures. All of these phenomena should be taken into ac-
count for true material failure. There are two ways to model failure in the numerical
environment. The first method is called as defining a failure criteria. This option
assigns a failure value for specific parameters like maximum effective plastic strain
and maximum pressure attained in elements. If the element exceeds the value of the
failure limit, it is eroded from the FE model. Therefore, the real failure of the material
cannot be modeled with that option because the strain rate effects and the continuum
damage mechanisms are eliminated. The second option is named as failure model,
which accounts for the degradation of the stiffness of the structure after the failure
occurs. Therefore, it provides a more realistic behavior of the structure.
Both of above mentioned methods are used in ballistic simulations, but especially
for high-velocity impact cases, the use of material failure models is more common.
Defining failure criteria could cause unrealistic element erosion for significant defor-
mation/high strain rate problems.
37
3.3.1 Failure Models
The material model determines the material’s strength according to its strain harden-
ing and softening response. The failure and the damage models predict the damage in
the material during a collision. The damage of material is estimated by the loss of co-
hesion in its interior, leading to either its complete disintegration or to some internal
damage which is manifested by the appearance of new free surfaces inside the mate-
rial specimen. Several models are available to predict the material fracture and they
are developed based on three broad approaches such as physical, micro-statistical and
phenomenological [27]. As damage progresses, discontinuities in the form of micro-
cracks may develop and stress concentration points arise from these discontinuities.
Two types of fracture, namely brittle and ductile, occur in the materials.
Figure 3.2: Failure types from left to right; brittle, ductile from brittle and ductile,
respectively [9].
Brittle fracture is the fracture between atomic bonds under relatively small deforma-
tion. When the local strain energy exceeds the energy of the atomic bonds, fracture
occurs. In brittle fracture, the crack grows abruptly with a small amount of plastic
strain energy. This type of fracture generally occurs on high strength metals with
poor ductility and toughness like ceramics, ice, cold metals, etc..
A ductile fracture occurs when materials go through large deformation before frac-
38
ture. The crack grows very slowly and moves with a high amount of plastic defor-
mation. Therefore, failure modeling depends on the material type and characteristics.
The schematic stress-strain curves for brittle and ductile materials are given in Figure
3.2.
The failure definition of a material can be done according to maximum stress or strain
or temperature or strain rate, etc. Many of the material models may depend on stress,
strain, or a combination of stress and strain. In this work, employed failure mod-
els depend on these variables, and if the variable reaches its critical value, then the
material fails.
The response of many materials can be grouped into two: the volumetric response
(equation of state) and the deviatoric response (strength model). The plastic defor-
mation of metallic materials is generally independent of the hydrostatic stress com-
ponents. This component is only related to the volumetric change of the material,
whereas the deviatoric stress components determine shape changes. On the other
hand, when a metallic material deforms plastically, the volume of the material does
not change, but its shape. The decomposition of stress state into volumetric and devi-
atoric parts reads
     
σ11 σ12 σ13 σm 0 0 S11 S12 S13
     
 σ21 σ22 σ23  =  0 σm 0  +  S21 S22 S23  (3.20)
     
σ31 σ32 σ33 0 0 σm S31 S32 S33
where
σ11 + σ22 + σ33
σm = (3.21)
3
is the mean stress and
   
S11 S12 S13 1/3(2σ11 − σ22 − σ33 ) σ12 σ13
(3.22)
   
 S21 S22 S23  =  σ 21 1/3(2σ11 − σ 22 − σ 33 ) σ23 
   
S31 S32 S33 σ31 σ32 1/3(2σ11 − σ22 − σ33 )
is the deviatoric stress.
39
3.3.2 Johnson-Cook Failure Model
The Johnson-Cook failure model is considered as a fracture criterion and is appro-

priate for the explicit solvers. This failure criterion depends on the maximum strain
of an element, which is the function of strain rate and temperature. Additionally the
stress triaxiality
σm
σ∗ = (3.23)
σeq
is taken into account in this model. The stress triaxiality is an important parameter
since the fail tendency of ductile materials is strongly depends on the pressure exerted
on them [27]. This failure model contains a damage parameter marked with D and its
defined as follows
εp
D= (3.24)
εf
where
Z t
εp = ε̇p dt (3.25)
0
r
2
ε̇p = Dp : Dp (3.26)
3
ε̇f = [D1 + D2 exp(D3 σ ∗ )] [1 + D4 (ln ε̇∗ )] [1 + D5 T∗ ] (3.27)
In (3.27) εf represents the equivalent plastic strain at fracture, εp is the effective plastic
strain, ε̇p is the effective plastic strain rate, ε̇∗p is the ratio of effective plastic strain
rate to reference strain rate as already given in (3.17). Dp is the plastic rate tensor,
and D1 ,. . . D5 are empiric parameters. When the damage parameter D reaches 1, the
element fails. The damage parameter can also be defined to develop in continuously
with the stress to describe the weakening of the material [22].
40
3.3.3 Modified Johnson-Cook Model and Cockroft-Lathan Failure Criterion
In LSDYNAr Modified Johnson-Cook model uses the Cockroft-Latham failure model.

This model calculates the flow stress by:

ε̇p
Bεneq (1 − T∗m )

σeq = A + 1+ (3.28)
ε̇0
where A, B, C, n and, m are material parameters as mentioned in subsection 3.2.1. εeq
ε̇p
is the equivalent plastic strain and ε̇0
is the normalized equivalent plastic strain-rate.
The Cockcroft-Latham failure criterion is based on both stresses and strains, so the
plastic work becomes an essential factor. It is defined as plastic work per unit volume
as:
Z εeq
1
D= (σ1 , 0)dεpeq (3.29)
Wcr 0
σ1 represents the maximum principal stress and Wcr is the Cockroft-Latham param-
eter, which denotes the total plastic work. Employing the Modified Johnson-Cook
model, failure of the element starts when D=1. Comparing this model with the
Johnson-Cook model, the most important differences are the strain rate dependence
term and failure criterion [7]. Johnson-Cook (J-C) model needs five input parame-
ters, which are D1 , D2 , D3 , D4 , andD5 for damage modeling, whereas the Cockroft-
Latham (C-L) failure model needs only one parameter. The C-L failure model can
be generated by doing a simple uniaxial tensile test. The area under the stress-strain
graph gives the limiting Wcr (plastic work equals to strain energy), and other The
Johnson-Cook parameters can be found by the same procedure explained in subsec-
tion 3.2.1. Compared to the procedure for the determination of J-C failure parameters,
C-L failure model needs less effort to obtain material model and failure. To be able
to identify material parameters more easily, the C-L failure model is selected in this
study.
41
42
CHAPTER 4
BALLISTIC IMPACT EXPERIMENTS OF HIGH STRENGTH ALUMINUM
This chapter presents the setup of the ballistic experiments conducted and the corre-
sponding results.
4.1 Configuration and Experimental Setup and Results
Ballistic tests were performed in the ballistic laboratory of FNSS with a 20mm frag-
ment simulating projectile (FSP) for strike velocites in the range 800-960m/s. Techni-
cal drawings and dimensions of 20mm projectile are given in Figure 4.1. The reason
behind using the 20mm FSP is to simulate the artillery threat, which corresponds to
Protection Level 4 and 5 in STANAG4569 [3] standards. Dimensions and weight
of 20mm FSP have to assure the values given in the drawing and table provided in
Figure 4.1. The 20 mm FSP is mandatory for Protection Levels 4 and 5 component
acceptance tests [32].
In Table 4.1, the protection level for bullet velocity and type is shown. Level-3 protec-
tion level covers a maximum of 770m/s strike velocity with 20mm FSP. Beyond this
velocity, protection levels are Level 4 and Level 5 for 20mm FSP. In order to evaluate
the protection Level 4 and Level 5, the projectile velocity is set between 800-960m/s
in the experiments.
An overview of ballistic laboratory, gun barrel, and impact velocity measurement

screens are shown in Figure 4.2.
43
Figure 4.1: The dimension of 12.7mm and 20mm fragment simulating projectile [3].
Table 4.1: Velocity levels for protection levels in STANAG4569 [3]
STANAG 4569 Kinetic Energy Bullet Artillery Threat (FSP

Protection Level Threat Type 20mm)
7.62x51 NATO Ball (Ball M80) 20 MM FSP Velocity 520 m/s
1 Riffle
Distance 30m Velocity 833 m/s tolerance +/- 20m/s
7.62 x 39 API BZ 20 MM FSP Velocity 630 m/s
2 Infantry Rifle
Velocity 695 m/s tolerance +/- 20m/s
Sniper Rifle 7.62 x 51 AP (WC core) 20 MM FSP Velocity 770 m/s
3/3+
Medium Machine Gun 7.62 x 54R B32 API (Dragunov) tolerance +/- 20m/s
14.5x114AP / B32 20 MM FSP Velocity Velocity 960 m/s
4 Heavy Machine Gun
Velocity 911 m/s tolerance +/- 20m/s
25mm APDS-TM-791 or 20 MM FSP Velocity 960 m/s
5
Automatic Cannon TLB 073 Velocity 1258 m/s 20 MM FSP Velocity Velocity 960 m/s
Figure 4.2: An overview of the ballistic laboratory
44
The projectile passes between two parallel laser windows for the impact velocity mea-
surement. Only the impact velocity of the projectile is measured with the laser system.
A high-speed camera is used to measure the exit velocity of a bullet from the target.
This fast cam system setup is located in the impact room.
In Figure 4.3 left, demonstration of the impact room is given. In order to cover the
high-speed camera from fragment effects of the target plate, an aluminum casing is
designed. The aluminum casing is shown in Figure 4.3 (right).
Figure 4.3: Impact room experimental setup (left) and high-speed camera casing
(right).
The camera is placed on the side of the target plate such that the camera axis is
perpendicular to the trajectory of the projectile. In Figure 4.4 the camera position is
shown. This placement is done for the measurement of the residual velocity values
with high accuracy. Fragmentation and spall effects are observed with a 30.000 fps
high-speed camera with a resolution of 256x176 pixels. Additionally, for target plate
positioning, a fixture is designed which is given in Figure 4.5. This fixture holds the
target plate from four corners with M16 bolts.
45
Figure 4.4: Demonstration of camera position to the target plate.
Figure 4.5: Target plate fixture.
Before starting the experiments, a calibration between the high-speed camera mea-
surements and the laser measurement system is needed.
46
Figure 4.6: High-speed camera images from calibration test.
In Figure 4.6 high-speed camera images are shown from calibration tests. The projec-
tile marching is shown frame by frame for every millisecond until the impact moment.
Three shots were performed for calibration studies, and the results are shown in Table
4.2.
Table 4.2: Calibration results
Laser High Speed

Test Number
Measurements (m/s) Camera (m/s)
1 951 903
2 957 909
3 960 912
A strike velocity of 960m/s was used for calibration tests. It is seen that the high-speed
camera results are lower than the laser results for about five percent. This information
will be later used in the computation of residual velocities.
In measuring the exit velocity of the bullet, difficulties were faced with due to the
flare effect. The flare affects the accurate exit velocity measurement because the
bullet cannot be captured in the recordings. This situation can be seen in Figure 4.6.
To avoid flare effects, a casing was designed, as shown in Figure 4.7. Using this new
casing, the tests were performed again. A high-speed camera image of the projectile
with a casing setup is also given in Figure 4.7 (below).
47
Figure 4.7: The casing to prevent flare effects (above), high-speed camera image of
the projectile with casing setup (below).
The target plate is a 300x500mm single monolithic high strength aluminum plate.
Three different thickness configurations were tested, which are shown in Figure 4.8.
This study only covers monolithic types of targets with variable thickness values, so
double-layered or triple-layered armor configurations were not tested.
48
Figure 4.8: Dimensions of target plate and thickness configurations.
The test plan is demonstrated in Table 4.3. These experimental results will be used to
validate the numerical model and creating an artificial neural network.
Table 4.3: Ballistic test plan
Number of shots 25.4mm 31mm 38mm

800 m/s 6 2 3
960 m/s 3 4 3
4.2 Experimental Results
Twenty-one shots were performed for ballistic experiments. It was aimed to set the
maximum velocity at 960m/s and the minimum velocity at 800m/s. For all shots,
residual velocity values were measured. The tests results for different aluminium
thicknesses are given in Tables 4.4 to 4.9. In the tables, the results are given for the
impact velocities and the residual velocities. The thicknesses of aluminum armors
used in tests are 25.4mm, 31mm and 38mm. The zero residual velocity value is
shown in Table 4.9 means, the projectile cannot exits from the target plate.
49
Table 4.4: The results of ballistic experiments for 25.4mm aluminum plate at 960m/s
velocity
Test Impact Residual

Number Velocity [m/s] Velocity [m/s]
1 950 536
2 956 543
3 957 547
Test Average 954 542
Table 4.5: The results of ballistic experiments for 25.4mm aluminum plate at 800m/s
velocity

4 785 385
5 800 410
6 794 394
7 799 411
8 814 418
9 834 429
Table 4.6: The results of ballistic experiments for 31mm aluminum plate at 960m/s
velocity

10 973 417
11 963 406
12 977 431
13 957 404
50
velocity

14 829 310
15 819 306
velocity

16 969 237
17 982 247
18 984 270
velocity

19 807 0
20 804 0
21 787 0
Test Average 800 0
The plates were perforated in all cases except for the 38mm thickness plate impacted
with 800m/s velocity. The perforated aluminum armor plates (25.4mm, 31mm and
38mm) after impact with 960m/s velocity are shown in Figure 4.9.
51
Figure 4.9: Deformed shape of 25.4mm (left), 31mm (middle), and 38mm (right)
plate
The last set of experiments were performed for 38mm thick aluminum armor at
800m/s strike velocity. In the last experiment set, residual velocities were very low or
cannot be read because the projectile could not exit from the target. As an example in
Figure 4.10 test number 19 is shown.
Figure 4.10: 800m/s velocity impact to 38mm aluminum plate
For all experiment sets, average residual velocity values will be compared with the
FEA results in Chapter 5.
52
CHAPTER 5
FINITE ELEMENT ANALYSIS
5.1 Computational Domain and Numerical Setup
Numerical simulations were performed with the explicit solver of LS-DYNAr . Anal-
yses were done for an impact velocity range of 800-960m/s. The model is presented
in Figure 5.1.
Figure 5.1: The model for impact simulations
The target plate is modeled as 300x500mm high strength aluminum. The thickness
of the plate is 25.4mm. The target plate is clamped at the four corners. The MPP
LS-DYNA R10.1 solver is selected as a solver version. Owing to the symmetry, only
a quarter of the model is simulated. The residual velocity of the projectile is measured
with part rigid body velocity option in the preprocessor of LSDYNAr .
The simplified version of the Johnson-Cook material model was used for FSP with
4340 steel material and, the modified Johnson-Cook material model with the Cockroft-
53
Latham failure model was used for high strength aluminum target material. In order
to determine the FSP material model constants of the Johnson-Cook model, it is nec-
essary to perform tests for different speed ranges. Quasi-static tests were performed
on a standard strength testing machine for the determination of A, B, n constants.
Furthermore, dynamic tests were performed to determine the value of strain rate sen-
sitivity. In this work, the Split Hopkinson Bar test was performed for the strain rate
parameters characterization of the FSP material. Thermal effects are neglected for
the simplified version of the Johnson-Cook material model. The failure model was
not used for the FSP.
Additionally, simple uniaxial tests were performed for the high strength alumınum
armor material parameter characterization. The test procedure was used for the target
material is the same procedure which is explained in subsection 3.3.3.
The viscous form of hourglass damping (IHQ=6) is used in all analyses, and the effect
of the type of hourglass formulation is examined. Also, the viscous form and the
stiffness form of hourglass formulation results are compared with residual velocity
values. Additionally, hourglass energy values are examined and compared with the
internal energy values to select accurate hourglass formulation for further analysis.
5.1.1 Mesh sensitivity

A mesh sensitivity study is performed for four different mesh sizes. The element
size is changed through the thickness direction of the target, and residual velocity
comparison is made for four different models. The mesh of the FSP is kept the same.
54
Table 5.1: Residual Velocity Values for different mesh sizes
Model Number of elements Through Residual Velocity [m/s]

Number Thickness Direction
1 15 510
2 30 570
3 42 576
4 60 577
Figure 5.2: The coarsest mesh (left) and the finest mesh (right) used in mesh sensitiv-
ity analyses.
In Figure 5.2, the finest mesh, and the coarsest meshes are shown. Note that in Figure
5.2 not the entire target plate, but only the region close to impact is visualized. The FE
mesh of the impact region of the target plate consists of 0.2x0.3mm reduced integra-
tion hexahedron elements for the finest mesh model, whereas in the coarsest model,
the mesh size is 1.6x1.6mm. Elements away from the impact region have larger sizes.
In Table 5.1 and Figure 5.3, residual velocity values for different mesh sizes are
shown. Model number one behaves stiffer compared to other three mesh models.
Except model number one, all three models give very similar results.
55
Figure 5.3: Mesh sensitivity results
Although the model two and three give similar residual velocity values, model three
gives better crater shapes when compared to model two. Results of the model three
and model four are very similar in terms of the residual velocity and the crater shapes.
The size of the elements in the thickness direction is vital in the event of tracing the
material’s behavior through the penetration process.
Additionally, all models are examined by their crater shape results. The crater shape
is another critical parameter in ballistic impact simulations. A comparison is made
with the quarter model, and results of the coarsest and the finest models are shown in
Figure 5.4. Larger elements may create an artificial eroding effect in the simulation
and, therefore, lead to larger crater size. According to residual velocity values and
crater shape results, model three is founded as the most suitable numerical model for
further simulations.
Figure 5.4: Crater shapes of the coarsest mesh and the finest mesh
56
5.2 Hourglass Solutions
The hourglass and element type selection could affect the accuracy of simulations.
In this section, stiffness based hourglass formulations are compared in terms of their
hourglass energy values. A shortcoming of the standard hourglass control (IHQ=1)
Type 1 is that the hourglass resisting forces are not orthogonal to the linear veloc-
ity field when elements are not in the shape of parallelpipeds. As a consequence,
such elements can generate a hourglass energy with a constant strain field or a rigid
body rotation. Flanagan and Belytschko [33] developed an hourglass control that
is orthogonal to all modes except for the zero energy hourglass modes [7]. Type 4
(IHQ=4) and Type 5 (IHQ=5) hourglass algorithms originate from the Flanagan and
Belytschko hourglass control. Type 5 hourglass algorithm is similar to Type 4, ex-
cept that the shape function derivatives are evaluated at the centroid of the element
rather than at the origin of the referential coordinate system. This method produces
the exact element volume [7]. Type 6 hourglass control improves Type 5 by scaling
the stiffness such that the hourglass forces match those generated by a fully integrated
element control [7].
This study aims to find the maximum hourglass energy among all types of hourglass
formulations. The lowest energy value means the best formulation, which prevents
non-physical hourglass problems.
The standard hourglass damping formulation with a coefficient of 0.1 is compared

with three forms of stiffness based hourglass algorithms that exist in LSDYNAr .
A comparison is made for the residual velocity values and the hourglass damping
energy values. In Table 5.2 residual velocity comparison is made among all hourglass
formulations.
57
Table 5.2: Residual velocity values for different hourglass formulations
Hourglass Residual
Control Velocity [m/s]
IHQ=1 576
IHQ=4 577
IHQ=5 578
IHQ=6 581
Deformed shapes of each hourglass formulation are shown in Figure 5.5. The damp-
ing coefficient is set to a constant value of 0.1 for each formulation.
Figure 5.5: Perforating state and final state of the model for each hourglass formula-
tion
Also, the amount of hourglass damping energy is compared for each formulation. In
58
Figure 5.6 comparison of the hourglass, energy is shown. It is seen that IHQ=1 adds
significantly more energy to the system compared to the other damping formulations.
Hourglass energies for IHQ=4 and IHQ=5 are very similar. Type six (IHQ=6) shows
the best results among all hourglass formulations.
Figure 5.6: Hourglass energy comparison
5.3 Element Type Solutions
In this section, the comparison of residual velocity values is presented for fully inte-
grated elements and one-point integration elements. This element formulation study
aims to find the robust element formulation for this impact simulations. LSDYNAr
contains four different element formulations (ELFORM=1, -1, -2, and 2), which are
used for explicit dynamic impact analysis. ELFORM 1 represents the one-point (re-
duced) integration element, and ELFORM -1, ELFORM -2, and ELFORM 2 repre-
sent the full integration elements. The full integration formulations ELFORM -1 and
ELFORM -2 may offer an improved behavior over the formulation in the ELFORM
2 by accounting for poor element aspect ratios in a manner to reduce the transverse
shear locking effects in formulation ELFORM 2. ELFORM-1 is a more computa-
tionally efficient implementation of ELFORM -2. However, ELFORM-1’s resistance
to a particular deformation mode, similar to an hourglass mode, is weakened [25].
59
In Figure 5.7, a comparison of reduced integration and fully integration solid ele-
ment is shown. Black points at the corner of an element represent the nodes, and
red crosses represent the integration points. ELFORM 1 has one integration point
whereas others have eight integration points. This difference directly affects the com-
putational cost because fully integrated elements need significantly more computa-
tion time. Detailed information about fully integration element types is given in [7].
Very close crater shapes are observed in analysis results among all element types.
Figure 5.7: Reduced integration and fully integration elements
In Figure 5.8, deformed plates after the impact are shown.
Figure 5.8: Deformed shapes of each element formulation
Additionally, elements formulation effects were analyzed for residual velocity val-
60
ues of the projectile. In Table 5.3 residual velocity results of each element for-
mulation are shown. Close results are seen between all element formulation. The
maximum difference in residual velocity is in the order of 1.8 percent. As a re-
sult, similarities are seen in both residual velocity and crater shape results. There-
fore ELFORM 1 is selected for further applications to reduce computational time.
Table 5.3: Residual velocity values for each element formulations
ELFORM Residual Velocity[m/s]

1 576
-1 574
2 574
-2 577
5.4 Analysis results and experiment comparison
In this section, for each set of experiments, average of the residual velocity values
is compared with the FEA results. The impact velocity of the FE model is set to
the average impact velocity in the related experiment. Detailed perforation results is
shown in Appendix A.
5.4.1 Results of 25.4mm aluminum plate
In Figure 5.9, perforation of the FSP to 25.4mm thickness armor, is shown for 954m/s
and 804m/s strike velocity cases
61
Figure 5.9: Perforation of FSP-25.4mm at 954m/s (above) and 804m/s (below) ve-
locity
In Table 5.4 the residual velocity is presented for 25.4mm thickness armor. In Table
5.4, the residual test velocity represents the average values of all the experiments.
Table 5.4: Residual velocity comparison between FEA and experiment for 25.4mm
plate
Analysis FEA Impact Average Test Impact FEA Residual Average Test Residual
Number Velocity [m/s] Velocity [m/s] Velocity[m/s] Velocity[m/s]
1 954 954 560 536
2 804 804 438 407
A good match between experiment and FEA results is obtained for both 954m/s and
804m/s impact cases. There is a 4% percentage difference for the 954m/s strike veloc-
62
ity case. The difference between exit velocity of test and FEA result is 7% percentage
for the 800m/s strike velocity.
Additionally, crater shapes of both experiments and FEA results are compared. In
Figure 5.10, crater shapes of the FE analysis, and the experiment are shown for the
25.4mm plate impacted with 804m/s velocity. Note that in Figure 5.10 FE model
is solved as a quarter symmetric model, but the results are displayed in full form.
Furthermore, crater diameters are measured and compared. From the FEA results,
the crater diameter is measured as 33.6mm whereas the experiment is measured as
32.4mm. Similar results obtained for the crater shape and the crater diameter.
Figure 5.10: Crater shape comparison of 25.4mm plate impacted with 954m/s veloc-
ity.
5.4.2 Results of 31mm aluminum plate
In Figure 5.11, perforation of FSP to 31mm thickness aluminum plate is shown for
968m/s, and 824m/s strike velocities.
63
Figure 5.11: Perforation of FSP-31mm at 968m/s (above) and 824m/s (below) veloc-
ity cases
In Table 5.5, the residual velocities is presented. A good agreement is found in terms
of residual velocity values between the test and the FEA results. In the case of 968m/s
impact case, 9% percentage difference is seen between the test and the FEA. Simi-
larly, in the 824m/s impact case, 8% variation is observed between the experiment
and the FEA results.
64
Table 5.5: Residual velocity comparison between FEA and experiment for 31mm
plate
3 968 968 452 414
4 824 824 335 308
Furthermore, the crater shape comparison is made for the 31mm plate impact sce-
nario. In Figure 5.12, the crater shape for a 31mm plate at 960m/s velocity is shown
for both the experiments and the FEA. The measured crater diameter is 34.2mm for
the FEA and 33.8mm for the experiment. VIn conclusion, close results are found both
for the crater shape and the diameter values between the FEA and the test results.
Figure 5.12: Crater shape comparison of 31mm plate
5.4.3 Results of 38mm aluminum plate
In Figure 5.13 perforation of FSP to 38mm thickness aluminum plate is shown for,
978m/s, and 800m/s strike velocities.
65
Figure 5.13: Perforation of FSP-38mm at 978m/s (above) and 800m/s (below) veloc-
ity case
In Table 5.6 residual velocities is presented. For the first impact case (978m/s), the
difference between the test and FEA is found as 6% for the residual velocities. In the
second impact scenario (800m/s), FSP cannot exit from the target plate in both the
experiments and the FEA.
66
Table 5.6: Residual velocity comparison between FEA and experiment for 38mm
plate
5 978 978 266 251
6 800 800 0 0
The depth of penetration values is measured for the 800m/s case. For the experiment,
the depth of penetration value is found as 38mm, whereas, in FEA gives 38.3mm. In
Figure 5.14 and 5.15 different views of the deformed aluminum plate are shown for
the test and the FEA.
Figure 5.14: 38mm plate 800m/s test (left) and analysis (right) results
67
Figure 5.15: 38mm plate 800m/s test (left) and analysis (right) results comparison
Additionally, backplane deflection values are measured, and 0.3mm variation was
seen between test and numeric model results. The backplane deflection measurement
is shown in Figure 5.15. In the test case, backplane deflection is measured as 14.7mm,
whereas numeric model gives 15mm.
Looking at both residual velocity and depth of penetration results, good correlation is
seen between FEA and experiment. The FEA model is validated with this comparison
study for the 38mm plate case.
68
CHAPTER 6
ARTIFICIAL NEURAL NETWORK
Artificial neural networks (ANN) connected to a set of data-processing techniques

which are used to find solution patterns or models from a sufficient amount of infor-
mation. Additionally, other benefits can be shown by creating a hybrid solution with
the FE method, which can give data to create ANN. Machine learning algorithms
can be used to solve complex problems in a very short time. Although the ANN is
widely used for many engineering problems, there are very few studies which contain
ballistic impact problems. Fernandez-Fdz and Zarea [34] developed a tool to predict
ceramic-metal armor ballistic performance by using a combined FE-ANN method.
They performed 200 impact scenario FE analyses to create ANN and 7.5% percent-
age difference was found for residual velocity values between analyses and ANN
results.
Similarly, Garci-Crespo [35] studied the performance of steel armors and created
different analysis cases for projectile radius, length of the projectile and thickness of
the target. The residual mass and the residual velocity values were taken as outputs
for each analysis and all these results were used to train an ANN. As a result, a
good correlation was seen between analyses and ANN results. Gonzales-Carrasco et
al. [36] worked on various neural network types, and compared training algorithms,
error cost functions and data selection methods for ballistic problems. Their study
showed that ANN is a suitable tool for ballistic limit problems. In this study, 18
ballistic experiments and 60 FE analyses were performed to build an ANN. First,
experimental results were used to validate the numerical model results than all 60
ballistic analysis results were used as input for the ANN.
69
6.1 Theory of Artificial Neural Network
Neural networks are constituted from simple components which work in parallel.
These components have so many similarities with a human learning system. They
contain three parts of layers, which are represented in Figure 6.1.
Figure 6.1: The multilayer structure of ANN [10]
Input layers, hidden layers, and output layers are linked through named nodes, or
neurons and each hidden layer uses the output of the previous layer as input.
The function of the input layer is to keep all the parameters and data as input and
transfer these data to the hidden layer. Then, the hidden layer takes the data from
the input layer and processes it to transfer all the data to the output layer. Lastly, the
output layer shows the results of hidden layer calculations. The computational power
of ANN is determined by the number of neurons or hidden layers used in the setup.
The optimum number of neurons or hidden layers can only be found by a trial and
error method [37].
In Figure 6.2, the calculation steps of a straightforward neuron are shown. Inputs are
represented by x1 , x2 and x3 and weights are represented by w1 , w2 and w3 . Each
input has its weight parameter during operations, and weight parameters are the only
parameters that change during the learning process. Weight parameters provide a con-
nection between two neurons during the process. At the beginning of the calculation,
each input parameter is multiplied by the weight parameter and summed with a bias
70
value, then this value, and then is sent to the activation function in order to prepare the
neuron to output. The activation function usually used to convert the total calculated
value to a number between 0 and 1. The sigmoid function is generally used for the
activation function phase [10].
Figure 6.2: One neuron calculation of ANN
Another significant point is in the ANN application is the backpropagation algorithm.

The weight parameter variation is processed with backpropagation. The backpropa-
gation method controls how the output changes according to weight change and cal-
culates an error value for each weight value. To minimize the error value, fine-tuning
operations are performed on the weight value. A detailed explanation can be found
in [10].
In this study, the Levenberg- Marquardt backpropagation algorithm is used. Training

of ANN stops when the mean square error of data samples increases. The mean
squared error (MSE) is calculated by the difference between output and target values.
Therefore, the MSE values close to zero indicate less error for output.
In (6.1), MSE calculation of the ANN for this work is shown. N represents the number
of samples, k represents the number of the input set,VFEA represents the residual
velocity obtained from FEA analysis (i.e., the actual value), VANN represents the
residual velocity predicted the ANN model. The square root of (6.1) is defined as
the Root Mean Square Error (RMSE) which represents the difference between FEA
residual velocity results and ANN residual velocity results. The definition of RMSE
is given in (6.2).
71
N
1 X k k
2
M SE = VF EA − VAN N (6.1)
N k=1
v
u
u1 X N
2
RM SE = t VFkEA − VAN
k
N (6.2)
N k=1
Additionally, a correlation between output and target is measured by performing the

regression analysis. If R-value is close to 1, it means that there is a good correlation
between the target and the outputs, whereas zero R-value indicates that a random
correlation.
The most crucial feature of the Levenberg-Marquardt algorithm is that it converges

much faster than other algorithms. The speed of convergence is much higher than
other backpropagation algorithms. Additionally, the Levenberg-Marquardt algorithm
provides lower MSE values, so the accuracy of this algorithm is higher than other
algorithms [10]. The Levenberg-Marquardt algorithm is selected in this study because
of the benefits above.
6.2 ANN Training
The preparation of the ANN is initialized with the 18 FE analyses at various impact
velocities. In this study, two input parameters are used in FE analyses, which are the
target thickness and the impact velocity. Residual velocity values are measured for
all FE analyses. Therefore, the residual velocity values are chosen as output for the
ANN training. The set of the performed analyses and results are shown in Table 6.1.
72
Table 6.1: FE analysis results for various strike velocities
Analysis Plate Impact Residual

Number Thickness [mm] Velocity [m/s] Velocity[m/s]
1 25.4 800 436
2 25.4 850 484
3 25.4 875 503
4 25.4 900 525
5 25.4 925 546
6 25.4 960 576
7 31 800 308
8 31 850 358
9 31 875 383
10 31 900 406
11 31 925 429
12 31 960 460
13 38 800 0
14 38 850 92
15 38 875 160
16 38 900 193
17 38 925 229
18 38 960 266
All the inputs and outputs presented in Table 6.1 are embedded in the ANN, and
then the training process of ANN is started. Data shown in 6.1 were selected at
intervals of 25m/s and 50m/s from 800m/s to 960m/s range. The neural network
tool of MATLAB is used for the ANN training. Then, SIMULINK is used for the
postprocessing phase. Default parameters for hidden layers and hidden neurons are
used for the ANN training. MSE values and regression analysis results are to be
examined to check the accuracy of the ANN. The ANN setup is shown in Figure 6.3
73
Figure 6.3: Representation of the ANN setup
All the data samples are divided into three groups in MATLAB for the ANN setup.
The first set of data is called the training, which is used for the weight and the bias
calculation. The other two sets are the validation and the test. The validation and
the test data sets are used for the performance checking of the ANN. Default values
are used for data participation, so 70% of data is sent to the training, 15% is sent to
validation and 15% is sent to the test part. All the training, validation, and test dataset
samples shown in Table 6.1 of ANN setup assigned randomly.
In Figure 6.4, the Levenberg- Marquardt algorithm performance graphic is shown.

Here epoch number represents the number of sweeps in one forward direction and
backward through all training data. According to Figure 6.4, the lowest validation
error is taken at the 6th epoch of the training. All of the training, validation, and test
MSE values are shown in Table 6.2.
Figure 6.4: Performance of Levenberg- Marquardt algorithm performance.
74
The lowest validation error is taken at the 6th epoch. The ANN run stops at the 12th
epoch to prevent overfitting problems. One of the problems that arises during neural
network preparation is called overfitting. Generally, the MSE error decreases as the
number of epochs increases; however, MSE can start to increase on the validation
data set as the neural network starts to overfit the training data. The early stopping
of ANN can prevent overfitting problems [10]. Additionally, the number of neurons
which are used in ANN can lead to overfitting results. In Figure 6.5 some of the trial-
error studies for ANN performance are shown for 123 epochs and 20 hidden layers of
the ANN setup separately.
Figure 6.5: 123 epoch results (left) and 20 hidden layers (right) results
In Figure 6.5, the MSE error starts to increase after 117 epoch while the training error
decreases. Therefore, it can be seen that the ANN starts to overfit after 117th of the
epoch. Also, the MSE values are found for both training and validation data sets. The
minimum MSE value is found around 592.26 which is much higher than the MSE
value shown in Figure 6.4. Similar overfitting results are seen for the twenty hidden
layers ANN setup. The training error starts to decrease after the second epoch while
the validation error starts to increase. After the trial-error study for the number of
epoch and hidden layers used in ANN, it is concluded that the ANN performance
with five hidden layers and twelve epochs can be used for further studies. The MSE
values are shown in Table 6.2 for five hidden layers and twelve epoch ANN setup.
75
Table 6.2: MSE and RMSE Values of Training, Validation and Test data
Samples MSE RMSE

Training 12 50.43 7.1
Validation 3 2.71 1.64
Testing 3 16.5 4.06
The MSE value of the training is 50.43, whereas the validation MSE is 2.71 and the
test MSE is 16.5. Additionally, the root mean square (RMSE) values are shown in
Table 6.2. As mentioned in Section 6.1, the RMSE values represent the difference
between the residual velocity values of the FEA (target) and the ANN (predicted).
The average difference between the target and the predicted velocity value is 7.1m/s.
Furthermore, the validation and the test MSE and RMSE values are examined. The
average difference between the FEA and the ANN results are found as 1.64m/s for
the validation and 4.06m/s for the test phase. The maximum difference between the
FEA and the ANN residual velocity value is 7.1m/s which is an acceptable difference
for ballistic impact problems. According to differences between the target and the
residual output, it can be concluded that the ANN gives quite a good correlation
between the output and the target.
Additionally, a regression analysis is performed to check ANN performance. In Fig-

ure 6.6, the target axis represents the FEA results, and the output axis represents the
ANN results. According to the graphics, the FEA and the ANN give quite similar
results for the training, test, validation and overall results. In all of the graphics, the
R-value is so close to 1 which means a very good correlation is provided between the
target and the outputs.
76
Figure 6.6: Regression analysis results for training, test, validation and overall data.
6.3 FEA and ANN results comparison
Figure 6.7 shows the SIMULINK setup for the ANN. The ANN is performed for
different velocities and thicknesses. The velocity range is set between 800m/s and
960m/s, and the target thickness is set according to FE analyses between 25mm and
38mm. The graphical representation of the residual velocity difference between the
experiments, FEA and ANN is shown in Appendix A.
77
Figure 6.7: ANN Simulink setup for ballistic analysis
After entering the plate thickness and impact velocity parameters the ANN is run,
and the residual velocity is obtained. ANN is performed for various thicknesses and
impact velocity values. The FE analyses are performed for the cases which are shown
in Table 6.3. Also, the ANN is performed for each thickness and impact velocity
values shown in Table 6.3. According to Table 6.3, consistent residual velocity results
are generated with the ANN. In Table 6.3, difference column represents the difference
of ANN results from FEA results. The ANN analyses run about 2-3 seconds for each
case, whereas the FEA analyses continue about 8 hours in an eight-core 64GB RAM
computer.
78
Table 6.3: ANN and FEA comparison of residual velocity values
27mm Thickness 32mm Thickness

Residual Residual
Velocity [m/s] Velocity [m/s]
Impact Velocity [m/s] ANN FEA Difference % ANN FEA Difference %
800 415 401 3.49 264 247 6.88
820 433 422 2.61 290 274 5.84
850 457 449 1.78 316 298 6.04
860 466 457 1.97 338 318 6.29
875 478 472 1.27 355 332 6.93
880 481 472 1.91 360 334 7.78
900 496 491 1.02 382 354 7.91
910 503 500 0.60 392 393 -0.25
925 512 512 0.00 407 376 8.24
930 516 516 0.00 411 379 8.44
960 534 541 -1.29 438 411 6.57
In order to show that the results are consistent with the FEA, FE analyses are per-
formed for 27mm thickness and 32mm thickness plates for each velocity, see Table
6.3. These analyses are performed for a comparison between the FEA and the ANN
results. According to Table 6.3, a good correlation is found between the ANN and the
FEA results. The maximum residual velocity value difference is found around 8.4%
for 32mm thickness and 930m/s impact velocity case.
Additionally, the ANN residual velocity results are compared with the experiment
results. In Table 6.4, the difference between ANN and experiment is represented.
The maximum difference between these two residual velocity result is found around
8.3%. Very close results are also seen between ballistic experiments and ANN results.
79
Table 6.4: ANN and Ballistic experiment comparison of residual velocity values
Thickness Average Test Impact Average Test ANN Residual Difference

[mm] Velocity [m/s] Residual Velocity [m/s] Velocity [m/s] [%]
954 542 569 4.7
25
804 407 444 8.3
968 414 451 8.2
31
824 308 332 7.2
978 251 267 5.9
38
800 0 0 0
In Figure 6.8 relationship between thickness of the target, impact velocity of projectile
and residual velocity values is shown in a response surface graphic.
Figure 6.8: Response surface representation of ANN results
80
CHAPTER 7
SUMMARY AND CONCLUSION
7.1 Conclusion
The study aims to reduce the number of expensive tests and analyses needed, prepro-
cessing and postprocessing times of FE in ballistic studies with the help of a combined
FEA-ANN approach. To this end, a tool is proposed by combining FEA and ANN
methods to make a quick prediction in aluminum armor design.
First, a numerical model verification study was conducted, and mesh sensitivity, hour-
glass formulations, and element formulations were examined. The effects of these
parameters on the residual velocity values were examined.
At the end of the verification of the numerical model, parameters that have significant
effect the on the FE model were determined. After that phase, ballistic experiments
were performed to validate the numerical model.
In ballistic experiments, the impact velocity, the residual velocity and the depth of
penetration were measured. Additionally, crater shapes and crater diameters were
examined in order to make comparisons with the FEA model.
After the ballistic test phase, a comparison of FEA results and ballistic test results
was performed. At first stage of comparison, the residual velocity values were con-
trolled. The maximum difference in residual velocity values was found around 9%
percent. A good correlation was seen between the residual velocity values for each
target thickness and impact velocity.
Furthermore, the crater shape, the crater diameter, the depth of penetration and the
81
backplane deflection values were compared.
Then, additional analyses were performed to expand the ANN database. All the ad-
ditional FEA results were added into the ANN database and then ANN runs were
performed for different thickness and impact velocity values. ANN analyses showed
reasonable results which were found to be consistent with FE analyses. For 27mm
plate thickness case, the residual velocity difference was founded as maximum 3.49%
whereas for 32mm plate thickness, the residual velocity difference was founded maxi-
mum 8.49%. Therefore, the ANN ballistic tool can be used for the preliminary design
of an aluminum armor for the determination of the armor thickness.
7.2 Future work
This study was performed for a single layer armor plate made of high strength alu-
minum material within a defined thickness range. This ANN model will be extended
for double-layered and tree-layered armor combinations in the future. This was the
main reason for choosing the ANN for this study. In the concept design stage, double-
layered or three-layered armor designs were also evaluated for armor plates. Double-
layer and three-layer ballistic armors were also created using a steel-aluminum com-
bination. The procedure followed in this study could be applied to double-layered
and three-layer aluminum armors to expand the ANN database. Input parameters of
the ANN could contain double-layer and three-layer armor designs. Also, the weight
of the armor could be added as output of the ANN model to determine the armor per-
formance together with the weight. In Steel-Al combined armors, the weight of total
armor is a very significant point for the armor design. Design of experiment approach
will be used to optimize back-propagation of the ANN parameters, so, the accuracy
of the ANN residual velocity results can be increased. Additionally, new graphical
user interface can be organized to make the tool more user-friendly.
82
REFERENCES
[1] C. L. Rao, V. Narayanamurthy, and K. R. Y. Simha, Applied Impact Mechanics.

John Wiley & Sons, Ltd, jun 2016.
[2] J. A. Zukas, High Velocity Impact Dynamics-Wiley-Interscience 1990.
[3] NATO and PfP, “Procedures Evaluating Protection Level Logistic Light Ar-
moured Vehicles,” Allied Engineering Publication, vol. 1, no. FEBRUARY,
2005.
[4] T. Børvik, M. J. Forrestal, and T. L. Warren, “Perforation of 5083-H116 Alu-

minum Armor Plates with Ogive-Nose Rods and 7.62 mm APM2 Bullets,” Ex-
perimental Mechanics, vol. 50, no. 7, pp. 969–978, 2010.
[5] M. E. Backman and W. Goldsmith, “The mechanics of penetration of projectiles

into targets,” International Journal of Engineering Science, vol. 16, no. 1, pp. 1–
99, 1978.
[6] A. L. Eulerian, “LSDYNA ALE Arbitrary Lagrangian Eulerian Traning Notes,”
[7] Livermore Software Technology Corporation (LSTC), Theory Manual

02/20/19, vol. 19.
[8] A. Mehmood, “Governing equations,” in Mathematical Engineering,

no. 9783319554310, pp. 13–21, Springer Verlag, 2017.
[9] L. V. Zhigilei, “University of Virginia, Department of Materials Science and

Engineering, Introduction to the Science and Engineering of Materials Lecture
Notes,” Materials Science, pp. 1–34.
[10] M. Beale and H. Demuth, “Neural network toolbox,” Neural Network Toolbox,
User’s Guide, 2004.
[11] H. Hopkins and H. Kolsky, “In: Proceeding of the 4th Hypervelocity Impact
Symposium. APGC-TR-60-39,” vol. 1, 1960.
83
[12] W. Goldsmith, High-Speed Physics. Springer-Verlag, Berlin,. 1967.
[13] A. E. Olshaker and R. L. Bjork, “Proceedings of 5th Symposium on Hyperve-

locity Impact,” Proceedings of 5th Symposium on Hypervelocity Impact, vol. 1,
no. 1, pp. 225–239, 1962.
[14] G. Ben-Dor, A. Dubinsky, and T. Elperin, “Ballistic Impact: Recent Advances

in Analytical Modeling of Plate Penetration Dynamics–A Review,” Applied Me-
chanics Reviews, vol. 58, no. 6, p. 355, 2005.
[15] O. Arndt, T. Barth, B. Freisleben, and M. Grauer, “Approximating a finite el-

ement model by neural network prediction for facility optimization in ground-
water engineering,” European Journal of Operational Research, vol. 166, no. 3,
pp. 769–781, 2005.
[16] W. L. Chan, M. W. Fu, and J. Lu, “An integrated FEM and ANN methodology
for metal-formed product design,” Engineering Applications of Artificial Intel-
ligence, vol. 21, no. 8, pp. 1170–1181, 2008.
[17] M. Ostad and A. Mazahery, “The ANN application in FEM modeling of me-
chanical properties of Al – Si alloy,” Applied Mathematical Modelling, vol. 35,
no. 12, pp. 5707–5713, 2011.
[18] N. Haghdadi, A. Zarei-Hanzaki, A. R. Khalesian, and H. R. Abedi, “Artificial

neural network modeling to predict the hot deformation behavior of an A356
aluminum alloy,” Materials and Design, vol. 49, pp. 386–391, 2013.
[19] R. Hambli, “Numerical procedure for multiscale bone adaptation prediction

based on neural networks and finite element simulation,” Finite Elements in
Analysis and Design, vol. 47, no. 7, pp. 835–842, 2011.
[20] P. P. Gudur and U. S. Dixit, “A neural network-assisted finite element analysis

of cold flat rolling,” Engineering Applications of Artificial Intelligence, vol. 21,
no. 1, pp. 43–52, 2008.
[21] R. Haj-Ali, H. K. Kim, S. W. Koh, A. Saxena, and R. Tummala, “Nonlinear

constitutive models from nanoindentation tests using artificial neural networks,”
International Journal of Plasticity, vol. 24, no. 3, pp. 371–396, 2008.
84
[22] B. Adams, Simulation of ballistic impacts on armored civil vehicles,Department
of Mechanical Engineering, Eindhoven University of Technology, Netherlands.
PhD thesis, 2003.
[23] M. Dokainish and K. Subbaraj, “A survey of direct time-integration methods in

computational structural dynamics—I. Explicit methods,” Computers & Struc-
tures, vol. 32, pp. 1371–1386, jan 1989.
[24] T. J. Hughes and R. L. Taylor, “Unconditionally stable algorithms for quasi-

static elasto/visco-plastic finite element analysis,” Computers and Structures,
vol. 8, no. 2, pp. 169–173, 1978.
[25] O. Anderson, “The Gruneisen ratio for the last 30 years, 143, 279-294, 2000,”
Geophys. J. Int.,, pp. 279–294, 2000.
[26] A. Shavit and C. Guthinger, Thermodynamics, from concepts to applications,

vol. 0132882671. Pearson Education limited, 1995.
[27] Z. Rosenberg and E. Dekel, Terminal ballistics, vol. 9783642253058. Springer-

Verlag Berlin Heidelberg, april 2012.
[28] M. A. Meyers, Dynamic behavior of materials, vol. 047158262X. John Wiley

and Sons, Inc. New York, january 1994.
[29] F. E. Prieto and C. Renero, “The equation of state of solids,” Journal of Physics
and Chemistry of Solids, vol. 37, pp. 151–160, jan 1976.
[30] ABAQUS, “Version 6.10-1 Analysis User’s Manual,” 2017.
[31] C. Kiran Sagar, A. Chilukuri, and A. Priyadrashini, “Determination of johnson

cook material model constants for 93% WHA and optimization using genetic al-
gorithm,” Materials Today: Proceedings, vol. 5, no. 9, pp. 18911–18919, 2018.
[32] Dytran, “Theory Manual 2008 r1,” Analysis, 2008.
[33] D. P. Flanagan and T. Belytschko, “A uniform strain hexahedron and quadri-

lateral with orthogonal hourglass control,” International Journal for Numerical
Methods in Engineering, vol. 17, no. 5, pp. 679–706, 1981.
85
[34] D. Fernández-Fdz and R. Zaera, “A new tool based on artificial neural networks
for the design of lightweight ceramic-metal armour against high-velocity impact
of solids,” International Journal of Solids and Structures, vol. 45, no. 25-26,
pp. 6369–6383, 2008.
[35] A. García-Crespo, B. Ruiz-Mezcua, D. Fernández-Fdz, and R. Zaera, “Predic-

tion of the response under impact of steel armours using a multilayer percep-
tron,” Neural Computing and Applications, vol. 16, no. 2, pp. 147–154, 2007.
[36] I. Gonzalez-Carrasco, A. Garcia-Crespo, B. Ruiz-Mezcua, and J. L. Lopez-

Cuadrado, “Dealing with limited data in ballistic impact scenarios: An empir-
ical comparison of different neural network approaches,” Applied Intelligence,
vol. 35, no. 1, pp. 89–109, 2011.
[37] H. M. Gomes and A. M. Awruch, “Comparison of response surface and neural

network with other methods for structural reliability analysis,” Structural Safety,
vol. 26, pp. 49–67, jan 2004.
86
APPENDICES
Figure A.1: A detailed display of the 31mm thickness 800m/s impact velocity perfo-
ration - (0µs - 20µs)
87
88
Figure A.4: Experiment, FEA and ANN results for 25.4mm thickness
Figure A.5: Experiment, FEA and ANN results for 31mm thickness
89
Figure A.6: Experiment, FEA and ANN results for 38mm thickness
90

Index

Uploaded by

Copyright:

Available Formats

Index

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Index

Uploaded by

Copyright:

Available Formats

EXPERIMENTAL AND NUMERICAL INVESTIGATION OF BALLISTIC

IMPACT BEHAVIOUR OF HIGH STRENGTH ALUMINIUM PLATES

IN PARTIAL FULFILLMENT OF THE REQUIREMENTS

EXPERIMENTAL AND NUMERICAL INVESTIGATION OF BALLISTIC

submitted by GÜRALP BAŞARAN in partial fulfillment of the requirements for the

Prof. Dr. Halil Kalıpçılar

Prof. Dr. İsmail Hakkı Tuncer

Assoc. Prof. Dr. Ercan Gürses

Examining Committee Members:

Prof. Dr. Altan Kayran

Assoc. Prof. Dr. Ercan Gürses

Assoc. Prof. Dr. Demirkan Çöker

Assoc. Prof. Dr. Melin Şahin

Assist. Prof. Dr. Recep M. Görgülüarslan

Name, Surname: GÜRALP BAŞARAN

EXPERIMENTAL AND NUMERICAL INVESTIGATION OF BALLISTIC

December 2019, 90 pages

Keywords: ballistics, armor, bullet, Finite Element Analysis, neural network

YÜKSEK MUKAVEMETLİ ALÜMİNYUM LEVHALARIN BALİSTİK

Aralık 2019 , 90 sayfa

First and foremost, I would like to express my gratitude to my supervisor Assoc.

LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii

1.1 Ballistic Impact . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 Types of Ballistic Collision and Projectiles and Target . . . . . . . . 2

1.2.1 Categorization of Ballistic Collision Types . . . . . . . . . . . 2

1.2.2 Projectile Types . . . . . . . . . . . . . . . . . . . . . . . . . 3

1.2.3 Target Types . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

1.3 The ballistic response of materials . . . . . . . . . . . . . . . . . . . 4

1.4 Penetration Mechanics . . . . . . . . . . . . . . . . . . . . . . . . . 6

1.5 Limit Velocities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

1.5.1 Elastic Limit Velocity . . . . . . . . . . . . . . . . . . . . . . 7

1.5.2 Plastic Limit Velocity . . . . . . . . . . . . . . . . . . . . . . 8

1.6 Ballistic Limit and Aerial Density . . . . . . . . . . . . . . . . . . . 10

1.7 Literature Survey on ballistic FEA combined ANN studies . . . . . . 11

1.8 Aim of this study . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

2.1 Numerical modeling approaches . . . . . . . . . . . . . . . . . . . . 16

2.1.1 Lagrangian approach . . . . . . . . . . . . . . . . . . . . . . 16

2.1.2 Eulerian approach . . . . . . . . . . . . . . . . . . . . . . . . 18

2.1.3 Arbitrary Lagrange Euler coupling . . . . . . . . . . . . . . . 19

2.1.4 Smooth Particle Hydrodynamics . . . . . . . . . . . . . . . . 20

2.2 Integration methods . . . . . . . . . . . . . . . . . . . . . . . . . . 21

2.2.1 Implicit integration . . . . . . . . . . . . . . . . . . . . . . . 23

2.2.2 Explicit integration . . . . . . . . . . . . . . . . . . . . . . . 24

2.3 CONCEPTS IN NUMERICAL SIMULATION . . . . . . . . . . . . 25

2.3.1 Hourglass Deformation and Hourglass Damping . . . . . . . . 26

2.3.1.1 Hourglass Deformation . . . . . . . . . . . . . . . . . 26

2.3.1.2 Hourglass Damping . . . . . . . . . . . . . . . . . . . 27

2.3.2 Contact-Impact Methods . . . . . . . . . . . . . . . . . . . . 28

2.3.2.1 Kinematic Constraint Method . . . . . . . . . . . . . . 29

2.3.2.2 Penalty Method . . . . . . . . . . . . . . . . . . . . . 30

3 MATERIAL MODELLING AND FAILURE MODELS . . . . . . . . . . . 31

3.1 Gruneisen parameter . . . . . . . . . . . . . . . . . . . . . . . . . . 31

3.1.1 Shock Rankine-Hugoniot conditions and relation . . . . . . . 33

3.2.1 Johnson-Cook Model . . . . . . . . . . . . . . . . . . . . . . 36

3.3 Material Failure Modelling . . . . . . . . . . . . . . . . . . . . . . . 37

3.3.1 Failure Models . . . . . . . . . . . . . . . . . . . . . . . . . 38

3.3.2 Johnson-Cook Failure Model . . . . . . . . . . . . . . . . . . 40

3.3.3 Modified Johnson-Cook Model and Cockroft-Lathan Failure

4 BALLISTIC IMPACT EXPERIMENTS OF HIGH STRENGTH ALUMINUM