Professional Documents
Culture Documents
Modelling Properties of Specia
Modelling Properties of Specia
Modelling Properties of Specia
U sin g A r t if ic ia l I n t e l l ig e n c e
By
HASSAN H. EL-CHABIB
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Library and Bibliotheque et
Archives Canada Archives Canada
NOTICE: AVIS:
The author has granted a non L'auteur a accorde une licence non exclusive
exclusive license allowing Library permettant a la Bibliotheque et Archives
and Archives Canada to reproduce, Canada de reproduire, publier, archiver,
publish, archive, preserve, conserve, sauvegarder, conserver, transmettre au public
communicate to the public by par telecommunication ou par I'lnternet, preter,
telecommunication or on the Internet, distribuer et vendre des theses partout dans
loan, distribute and sell theses le monde, a des fins commerciales ou autres,
worldwide, for commercial or non sur support microforme, papier, electronique
commercial purposes, in microform, et/ou autres formats.
paper, electronic and/or any other
formats.
Canada
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
THE UNIVERSITY OF WESTERN ONTARIO
FACULTY OF GRADUATE STUDIES
CERTIFICATE OF EXAMINATION
The thesis by
Hassan H. El-Chabib
Entitled:
11
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
ABSTRACT
The engineering properties of special concretes depend on the effect of and interaction between
complex parameters such as the non-homogeneous nature and inherently different properties of
their components, and the dual and/or contradictory effects of several ingredients on the overall
performance. Current practice resorts to costly and time-consuming laboratory testing on
numerous trial mixtures to investigate such properties. Furthermore, current semi-empirical and
regression analysis tools used to predict these properties have fundamental limitations, are based
on many simplifying assumptions, and thus are not reliable. Meanwhile, artificial intelligence
(AI) methods have recently emerged as powerful techniques for modeling and predicting the
behaviour of materials. However, their applications in concrete technology are still at the
embryonic stage. Hence, the goal of this work is to lay the ground for more mature applications
of AI in concrete materials and structures. This is achieved in this study through selected
applications that have proven to be challenging using traditional modeling techniques.
The first application focussed on modeling the segregation phenomenon of flowable concrete,
which remains to be the least understood of its properties. Data on the ability of flowable
concrete mixtures to resist segregation of coarse aggregate particles are practically nonexistent
partly due to lack of simple and reliable test methods to measure such a property. In this study, a
simple, yet reliable test method for quantitatively evaluating the segregation resistance of
flowable concrete mixtures was developed. The test was used to generate a comprehensive
database on the segregation of flowable concrete mixtures. A total of 123 self-consolidating
concrete mixtures were prepared and tested for segregation resistance, slump flow, slump Tso(Cm),
and compressive strength. In addition, properties of 175 underwater flowable concrete mixtures
were carefully selected from a larger database collected from published research. The database
thus generated was used to train artificial neural networks (ANN) models to predict engineering
properties of flowable concrete. The ANN models thus developed showed excellent
performance not only in predicting the properties of flowable concrete, but also in predicting the
effect of basic mixture ingredients on such properties.
The second major application focussed on modeling the shear strength of slender concrete
beams. Shear strength results for 656 normal and high-strength concrete slender beams (with
and without stirrups) were collected from published research. The database was used to train
iii
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
ANN models to study the shear behaviour of normal and high-strength concrete (HSC) slender
beams and to investigate the shear capacity contributed by each of the basic shear transfer
mechanisms. Results showed that current shear design provisions overestimate the effect of
compressive strength of concrete on the shear capacity of HSC beams and that simply adding the
shear capacity of stirrups to that of concrete beams without shear reinforcement will
significantly underestimate the shear strength of such beams. The ANN models allowed a more
advanced understanding of the effect of basic shear design parameters on shear strength.
Moreover, the ability of current shear design guidelines to calculate the shear capacity of FRP-
reinforced concrete beams was investigated in this study. Analysis of shear strength
experimental results for 150 FRP-reinforced concrete beams along with ANN predictions
showed that current shear design guidelines for FRP-reinforced concrete members significantly
underestimate the shear capacity of beams without shear reinforcement and overestimate the
shear capacity of FRP stirrups. The genetic algorithms approach was used to develop and
optimize new shear design equations for FRP-reinforced beams. It was shown that the effect of
the axial rigidity of FRP longitudinal reinforcement and FRP stirrups on the shear strength can
be best represented by a cubic root and square root functions rather than linear or square root
relationships as proposed by current guidelines, respectively, which provided a more accurate
design equations compared to other existing methods.
The main contribution of this study is that it demonstrates that artificial intelligence techniques
can be mobilized to predict the engineering properties of special concretes with accuracy that
cannot be matched by existing semi-empirical and regression analysis-based techniques.
Furthermore, there are substantial research data on various aspects of concrete materials and
structures, but such data remains scattered and not fully exploited. This study demonstrates that
databases carefully generated from such existing data can be exploited to develop design
equations that outperform current design code provisions by far in terms of accuracy and
sensitivity to the effect of the basic design parameters involved. This work can initiate future
effort in developing commercial Al-based tools to assist engineers and contractors in the
formulation of special concretes and the prediction of their engineering properties.
Keywords: Flowable, high strength, fibre reinforced polymer, artificial neural networks, genetic
algorithms, modeling, behaviour, shear, mechanisms.
iv
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
CO-AUTHORSHIP
This thesis has been prepared in accordance with the regulation of manuscript format
Substantial parts o f this thesis were either published in or submitted for publication to
data analysis, modeling process, and writing the initial version of all publications listed
below were carried out by the candidate himself. The contribution of his research advisor
and any other co-author (if applicable) consisted of either providing advice, and/or
[1] El-Chabib H., Nehdi M., and Sonebi M., “Artificial Intelligence Model fo r
Flowable Concrete Mixtures used in Underwater Construction and Repair,” ACI
Materials Journal, V. 100, No. 2, 2003, pp. 165-173.
[2] El-Chabib H. and Nehdi M., “Neural Network Modeling o f Properties o f Cement-
Based Materials Demystified,” Journal of Advances in Cement Research, Elsevier,
V. 17, No. 3, 2005.
[3] El-Chabib H., Nehdi M., and Said A. “Predicting Shear Strength o f NSC and HSC
Slender Beams Using Artificial Intelligence,” Journal of Computers and Concrete,
Techno Press, V. 2, No. 1,2005.
[4] El-Chabib H., Nehdi M., and Said A. “Predicting Effect o f Stirrups on Shear
Strength o f Reinforced NSC and HSC Slender Beams Using Artificial
Intelligence,” submitted to the Canadian Journal of Civil Engineering.
[5] El-Chabib H. and Nehdi M., “Effect o f Mixture Design Parameters on Segregation
o f Self-Consolidating Concrete,” submitted to ACI Materials Journal.
[6] El-Chabib H., Nehdi M., and Said A. “Evaluation o f Shear Capacity o f FRP
Reinforced Concrete Beams Using Artificial Intelligence,” Submitted to the
Journal o f Smart Structures and Systems.
[7] El-Chabib H. and Nehdi, M., “Effect o f Mixture Ingredients on Segregation
Resistance o f Self-Consolidating Concrete,” SCC 2005, Joint Second North
American Conference on the Design and Use of Self-Consolidating Concrete and
The Fourth RILEM International Symposium on Self-Compacting Concrete,
Centre for Advanced Cement-Based Materials, Chicago, IL, October 2005.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
TO: M y -wife OLA,
M y son TA<
S£>'KJ and
M y Daughter (FA<
RJLK
vi
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
ACKNOWLEDGMENTS
The author would like to convey his sincere appreciation and gratitude to his advisor, Dr.
Moncef Nehdi for his guidance, advice and encouragement throughout the course of this
research. His mentorship, support and patience were of great essence and are gratefully
acknowledged.
Special thanks are due to my friends and colleagues Dr. Esam Hewayde and Dr. Aly Said
for the many constructive discussions we had during the course of this work. The author
would also like to thank all technicians, staff, and fellow graduate students at The
Finally, the author would like to express his genuine gratitude and appreciation to his
father, mother, brothers, sisters, wife, and beloved children for their continuous support
and encouragement.
vii
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
TABLE OF CONTENTS
CHAPTER I
I n t r o d u c t io n
CHAPTER II
A r t if ic ia l I n t e l l ig e n c e M o d e l in g and M ethodology
viii
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
2.3 Neural Networks A p p ro ach ........................................................................ 18
2.3.1 Neural Networks Applications in Modeling Cement-Based
Material.................................................................................................. 20
2.3.2 Feed-Forward Back-Propagation Neural N etworks.......................... 22
2.3.2.1 Architecture of multi-layer perceptron netw orks.................. 25
2.3.2.2 Training multi-layer perceptron networks ............................ 27
2.3.2.2.1 Selection and pre-processing o f training d a ta 29
23.2.2.2 Learning rate ......................................................... 30
2 3 .2 .2 3 Momentum................................................................ 31
2 3.2.2 A Activation functions ................................................. 31
2.3.2.2.5 Duration o f training................................................ 32
23.2.2.6 Training process .................................................... 34
2.3.23 Validation of trained MLP networks ................................... 39
2.3.3 Case Study: Effect of MLP network Architecture on its
Performance.......................................................................................... 39
2.3.3.1 Effect o f training duration....................................................... 42
2.3.3.2 Effect o f number of hidden units ........................................... 44
2.3.3.3 Effect o f number of hidden layers ......................................... 45
2.3.4 Optimizing the Performance o f MLP N etw orks................................ 47
2.4 Concluding R e m a rk s................................................................................... 49
2.5 R eferences...................................................................................................... 51
CHAPTER III
E ffect of M ix t u r e I n g r e d ie n t s on P r o p e r t ie s of Flo w able C oncrete
ix
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
3.3.2 Mixtures Proportions and Mixing Techniques .................................. 68
3.3.3 Testing Procedures............................................................................... 69
3.4 Effect of Basic M ixture Ingredients on SCC P ro p e rtie s....................... 75
3.4.1 Effect o f Cementitious Materials Content ......................................... 76
3.4.2 Effect o f Water to Cementitious Materials Ratio .............................. 80
3.4.3 Effect o f HRWR and VMA D osages.................................................... 84
3.4.4 Effect o f Coarse to Total Aggregate Ratio ........................................ 90
3.5 Concluding R e m a rk s................................................................................... 92
3.6 R eferences...................................................................................................... 94
CHAPTER IV
M odeling P roperties of F lowable C oncrete U sing A rtificial N eural
N etworks
4.1 Introduction .................................................................................................. 96
4.2 U nderw ater C o n c re te.................................................................................. 97
4.3 Application of Artificial Neural N etw orks................................................ 99
4.3.1 Selection and Pre-Processing o f D atabase......................................... 99
4.3.2 ANN-Based Models for Segregation Resistance of SCC Mixtures .. 102
4.3.2.1 Performance of ANN-STATIC and ANN-DYNAMIC 105
4.3.2.2 Sensitivity analysis of ANN-DYNAMIC .............................. 107
4.3.3 ANN Based Model for Properties of UWC M ixtures....................... 110
4.3.3.1 Performance of ANN models in predicting properties of
UWC mixtures.......................................................................... Ill
4.3.3.2 Sensitivity of ANN model to variations of UWC
ingredients................................................................................ 115
4.4 Concluding R e m a rk s ................................................................................... 117
4.5 R eferences...................................................................................................... 119
CHAPTER V
Application of ANN in P redicting Shear C apacity of NSC and HSC Slender
Beams w ithout Stirrups
5.1 Introduction .................................................................................................. 121
5.2 Shear Evaluation Techniques Included in this S tu d y ............................ 126
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
5.2.1 Statistical and Analytical Methods ..................................................... 126
5.2.2 Design Specifications........................................................................... 127
5.2.3 Artificial Neural Networks ................................................................. 129
5.2.3.1 Development of D atabase....................................................... 129
5.2.3.2 ANN Model ............................................................................. 130
5.3 Results and Discussion ................................................................................ 132
5.3.1 Effect o f Compressive Strength of Concrete ..................................... 140
5.3.2 Effect o f Tensile Steel R a tio ............................................................... 143
5.3.3 Effect o f Beam’s Effective D epth....................................................... 145
5.3.4 Effect o f Shear Span to Depth R atio ................................................... 147
5.4 Concluding R em arks................................................................................... 149
5.5 References...................................................................................................... 151
CHAPTER VI
E f f e c t s o f S t i r r u p s o n S h e a r D e s ig n P a r a m e t e r s of R e in f o r c e d
C oncrete S len der B eam s
xi
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER VII
E v a l u a t in g S h e a r C a p a c it y of F R P -R e i n f o r c e d C o n c r e t e B e a m s U s in g
A r t if ic ia l I n t e l l ig e n c e
CHAPTER VIII
S u m m a r y , C o n c l u sio n s and R e c o m m e n d a t io n s
xii
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
LIST OF TABLES
Table
5.1 Statistical data for shear design parameters and shear capacity of beams
used in database................................................................................................. 130
5.2 Performance o f shear calculation methods considered in this study 133
6.1 Statistical data for shear design parameters and shear capacity
of beams used in database................................................................................. 161
6.2 Performance o f shear design methods considered in this study
(beams with shear reinforcement).................................................................... 163
7.1 Statistical data for shear design parameters and shear capacity of
FRP-reinforced concrete beams used in database.......................................... 197
7.2 Parameters used in Genetic Algorithms setting................................................ 201
7.3 Performance o f shear calculation methods considered in this study
(FRP-reinforced concrete beams).................................................................... 203
xiii
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Table
B.2 Test results o f mixtures used in investigating properties of SCC mixtures... 244
D. 1 Properties and shear results of reinforced NSC and HSC slender beams
without web reinforcement (testing data)........................................................ 253
D.2 Properties of beams used to investigate the effect of compressive
strength o f concrete ( f’c ) on shear strength ( Vu/b wd ) .................................. 256
D.3 Properties o f beams used to investigate the effect of tensile steel ratio
(pi) on shear strength ( Vu/ b wd ) ...................................................................... 257
E. 1 Properties and shear results of reinforced NSC and HSC slender beams
(testing data)....................................................................................................... 261
E.2 Properties o f beams used to investigate the effect of stirrups capacity
(p /yv) on shear strength (N SC)......................................................................... 267
E.3 Properties o f beams used to investigate the effect of stirrups capacity
(p/yv) on shear strength (H SC)......................................................................... 268
E.4 Properties o f beams used to investigate the effect off c on shear strength.... 269
E.5 Properties o f beams used to investigate influence of stirrups on the effect
off c on shear strength....................................................................................... 220
E.6 Properties o f beams used to investigate the effect o f tensile steel ratio (pi)
on shear strength................................................................................................ 271
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
on shear strength of concrete beam s................................................................ 281
F.3 Properties o f beams used to investigate the effect of pfo
on shear strength o f concrete beams................................................................ 281
F.4 Properties o f beams used to investigate the effect of f c
on shear strength o f concrete beams................................................................ 282
xv
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
LIST OF FIGURES
Figure
2.1 Flow o f process in a general evolutionary genetic algorithm......................... 12
2.2 Selection using the roulette wheel method....................................................... 15
2.3 Schematic representation of a single-point crossover...................................... 17
2.4 Rosenblatt’s perceptron (artificial neuron)....................................................... 23
2.5 Example o f activation functions: hard limit and logsig
(sigmoidal functions)........................................................................................ 25
2.6 Typical architecture o f MLP back-propagation network with one
hidden layer........................................................................................................ 26
2.7 Flow of the training process of MLP networks................................................ 28
2.8 Over-trained vs. well-trained MLP networks................................................... 33
2.9 Performance o f MLP network in predicting the in-situ strength
of concrete.......................................................................................................... 41
2.10 Effect o f training duration (number of epochs) on the performance
of MLP network................................................................................................. 43
2.11 Effect o f number of hidden neurons on the performance of MLP network... 45
2.12 Effect o f number o f hidden layers on the performance of MLP
network (constant number of hidden units)..................................................... 46
xvi
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Figure
xvii
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Figure
xviii
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Figure
xix
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Figure
6.15 Influence o f stirrups on the effect of shear span to depth ratio a/d
on shear strength o f RC beam s......................................................................... 179
7.1 Typical stress-strain relationship for FRP and steel reinforcement................ 189
7.2 Architecture o f neural network model for predicting the shear capacity
of FRP-reinforced concrete beams................................................................... 198
7.3 Measured versus calculated shear strength using GA
(beams without stirrups).................................................................................... 204
7.4 Measured versus calculated shear strength using GA (beams with stirrups). 204
7.5 Performance o f ACI 440 in calculating the shear strength of
FRP-reinforced concrete beams without stirrups............................................ 207
7.6 Performance of CSA S806 in calculating the shear strength of
FRP-reinforced concrete beams without stirrups............................................ 208
7.7 Performance of JSCE shear equation in calculating the shear strength
of FRP-reinforced concrete beams without stirrups....................................... 208
7.8 Performance o f ISIS Canada in calculating the shear strength of
FRP- reinforced concrete beams without stirrups........................................... 209
7.9 Performance o f the proposed equation in calculating the shear strength
of FRP-reinforced concrete beams without stirrups....................................... 209
7.10 Performance o f ACI 440 in calculating the shear strength of
FRP-reinforced concrete beams with stirrups.................................................. 212
7.11 Performance of CSA S806 in calculating the shear strength of
FRP-reinforced concrete beams with stirrups.................................................. 212
7.12 Performance o f JSCE shear equations in calculating the shear strength
o f FRP-reinforced concrete beams with stirrups............................................ 213
7.13 Performance of ISIS Canada in calculating the shear strength of
FRP-reinforced concrete beams with stirrups.................................................. 213
7.14 Performance of the proposed shear equations in calculating the shear
strength o f FRP-reinforced concrete beams with stirrups............................... 214
7.15 Performance of ANN model in predicting the shear strength of
FRP-reinforced concrete beams with stirrups.................................................. 214
7.16 Effect o f FRP shear reinforcement on shear strength of RC beam s............... 215
7.17 Effect o f longitudinal reinforcement ratio on shear strength of RC
beams with FRP shear reinforcement.............................................................. 217
7.18 Effect o f concrete compressive strength on shear strength of RC
beams with FRP shear reinforcement.............................................................. 218
xx
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Figure
7.19 Effect o f shear span to depth ratio on shear strength of RC beams with
FRP shear reinforcement................................................................................... 219
7.20 Performance of the modified version of ACI 440 in calculating the
shear strength o f FRP-reinforced concrete beam s.......................................... 221
7.21 Performance o f the modified version of CSA S806 in calculating
the shear strength o f FRP-reinforced concrete beam s.................................... 222
7.22 Performance o f the modified version o f ANN model in calculating
the shear strength of FRP-reinforced concrete beam s.................................... 223
xxi
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
APPENDICES
xxii
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
NOTATIONS:
K
Connection strength between unit i in layer /-I and unit j in layer /,
rj
Output from unit j in layer /,
YY111
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
HRWR High range water reducer,
H SC High-strength concrete,
K Number of units in the output layer,
M LP Multi-layer perceptron,
Ned Axial force on a concrete member,
N SC Normal strength concrete,
ok Predicted output at unit (k) in the output layer
P Number o f patterns,
Pd Penetration depth,
PF Performance factor,
Pi Probability o f selecting an individual (i) in genetic algorithm,
PU Processing Unit,
SCC self-consolidating concrete,
sd Standard deviation o f a database,
SF Silica fume,
SI Segregation index,
SP Special concrete,
Sr Segregation ratio,
sus Stochastic universal sampling,
UW C Under water concrete,
Vc Shear capacity carried by concrete,
Veal Calculated shear capacity,
Vcf Shear capacity carried by concrete in FRP-reinforced concrete beams,
Vf Shear capacity carried by FRP shear reinforcement,
Vm Measured shear capacity,
VMA Viscosity modifying admixture,
v„ Nominal shear capacity,
xxiv
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Xi Input variable,
a Shear span,
bw Beam’s width,
cm Cementitious materials,
d Beam’s effective depth,
da Maximum size o f coarse aggregate,
e* Error at unit (k) in the output layer,
f c Compressive strength of concrete,
f ck Characteristic cylinder compressive strength,
fjbend Strength at the bend portion o f the FRP stirrups,
ffUV Ultimate tensile strength of FRP stirrups,
f yv Yield strength o f stirrups,
n Number o f units in the input layer,
s Spacing between stirrups,
tk Corresponding target at unit (&) in the output layer,
vcf Shear strength carried by concrete in a FRP reinforced concrete beam,
Vf Shear strength carried by FRP stirrups,
v, Fitness value o f individual (i),
v„ Shear strength o f concrete beam,
w/cm Water to cementitious materials ratio,
xmax Maximum value o f variable x in a database,
xmin Minimum value o f variable x in a database,
xt Scaled value o f variable x in a database,
xxv
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
pi Longitudinal reinforcement ratio,
pv Shear reinforcement ratio,
acp Axial stress on a concrete member,
X Variable to account for the effectiveness of stirrups,
xx vi
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 1 1
CHAPTER 1
INTRODUCTION
1.1- GENERAL
enhance their rheology, structural performance and durability. Special concretes (SC),
concrete (HSC), and concrete reinforced with fibre-reinforced polymer (FRP) materials
are a few examples o f such modified materials that emerged to meet the needs of such
mechanical properties and/or mode of failure, the properties and behaviour of special
concretes are likely to be different from those of ordinary concrete. The relatively recent
mechanical behaviour, and the limited laboratory and field data on their properties all
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 1 2
properties, which constitutes the leading cause of their limited acceptance by contractors
The purpose o f using special concretes includes simplifying the construction process of
durability o f such structures, and developing more sustainable and cost effective
structures over their life cycle. For example, self-consolidating concrete (SCC) was
Underwater concrete (UWC) was specially developed to facilitate the construction and
Despite the increasing demand for special types of concrete and the difference in their
behaviour from that of ordinary concrete, current standards and specifications are yet to
include special guidelines and rules for the design and/or testing of such materials. For
slightly modify existing design equations developed for conventional reinforced concrete
structures for use in the design of HSC and/or FRP-reinforced concrete members.
Moreover, fresh properties o f flowable concrete such as SCC and UWC are considerably
different from those o f normal concrete. Despite of the significant amount of research
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 1 3
conducted on this subject, most standards and specifications are yet to provide standard
The main objective o f this research is to enhance the understanding of and ability to
predict the behaviour and property of several types of special concrete using the artificial
networks (ANNs) and genetic algorithms (GAs) were used in this work. Focus was
models.
3- Develop an ANN-based model to predict properties of SCC and UWC as well as the
4- Evaluate the performance of existing shear design methods in calculating the shear
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 1 4
5- Investigate the ability o f existing shear design methods to accurately capture the
effect of basic shear design parameters on the shear capacity of normal and hig-
6- Investigate the feasibility of using existing shear design methods in the design of
7- Propose new modifications to enhance existing shear design equations for FRP-
This thesis has been prepared according to guidelines stipulated by the Faculty of
divided into 8 chapters, 6 o f which have been written independently and either published
optimization techniques for complex functions. The chapter also presents a detailed
description o f feed-forward multi layer perceptrons (MLP) neural networks and a state of
the art review on their application in modeling the behaviour of cement-based materials.
Details on how to build, train, and validate MLP networks are provided. Furthermore, a
parametric study to optimize the performance of MLP neural networks through a case
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 1 5
concrete. The experimental program including test methods used, materials properties,
and testing procedures is carefully described. The development of a simple, yet reliable,
test method to evaluate the ability of flowable concrete to resist segregation of large
coarse aggregate particles is presented in this chapter along with the performance of this
feasibility of using MLP networks to predict the properties of flowable concrete and to
capture the influence o f basic mixture ingredients on its behaviour. Results obtained
using MLP network models and those acquired experimentally are examined, compared,
and discussed.
Chapter 5 evaluates the capability of various empirical models and shear design
concrete slender beams without shear reinforcement. Shear design methods were
evaluated using experimental results compiled from 345 concrete beams tested in
different laboratories around the world. The chapter also describes the stages for the
development o f an ANN-based model to predict the shear strength of NSC and HSC
slender beams without shear reinforcement. The performance of the ANN model and its
ability to capture the effect o f basic shear design parameters on shear strength, along with
those provided by the various shear calculation methods are examined and discussed.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 1 6
In chapter 6, a database o f shear strength results for 645 concrete slender beams (with
and without stirrups) tested by several researchers for their shear capacity is developed
and discussed. The linear effect of stirrups on the shear strength of reinforced concrete
slender beams as assumed by current design practice is critically examined. The chapter
also describes the architecture, training and validation of a MLP neural network model to
simulate the shear behaviour o f reinforced concrete slender beams. A comparative study
between ANN model predictions of the effect of shear reinforcement on other shear
contribution mechanisms and those provided by various empirical models and design
equations proposed by various guidelines and standards for the design of FRP-reinforced
concrete beams is investigated. New shear design equations for FRP-reinforced concrete
beams, based on ANN predictions and GA optimizations, are proposed in this chapter.
recommended by current codes and standards for the design of concrete beams reinforced
with FRP.
Chapter 8 includes a summary o f the thesis, general and specific conclusions, along with
Finally, databases and pertinent information essential for the completeness of the content
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 1 7
This thesis takes the use of artificial intelligence in modeling engineering properties of
concrete beyond its embryonic stage to a more mature stage in which such powerful tool
concrete. Allowing, for the first time, the prediction of its engineering properties
strength.
• Proposing new more accurate shear equations for the design of concrete beams
and results.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 8
CHAPTER 2
2.1-INTRODUCTION
consequential events. However, in reality such events are mutually interacting and each
parameter cannot be studied without considering the entire context of its environment.
This study explores the feasibility of using an alternative approach, artificial intelligence,
Chamiak and McDermott (1985) is the study of mental faculties through the use of
people by imitating what the brain does through computing. In other words, Artificial
Intelligence is the science concerned with understanding intelligent behaviour and how it
can be artificially created (Smithers et al. 1990). The use of Artificial Intelligence (AI) in
*A version o f this chapter has been published in the Journal o f Advances in Cement Research, volume 17,
No. 3,2005.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 9
has received a significant attention in the literature during the last decades. Despite the
significant progress achieved in the application of AI, its technology is still in its infancy
and the ultimate goal of building an intelligent system capable of performing humanly
Most engineering problems can be solved using AI techniques and the technology has
been used successfully in several complex applications. The automotive and aerospace
industries have extensively used both robotic technology and expert systems in their
manufacturing processes. Expert systems have also been used successfully in computer-
aided design applications, electronic design, computer system design, and system
selection and installation. The applications of artificial intelligence for laboratory robots
in Bioengineering are also becoming a valuable tool for researchers. The potential for
using artificial intelligence in civil engineering and the construction industry is unlimited.
Most Al-based research has been dominated by studying design applications. Its use for
predictive tools and modeling techniques, especially for construction materials is fairly
new. The basic methodology o f using AI in design is to build artificial systems capable of
creating design tasks on their own or simply to provide intelligent support for people
doing design. The philosophy, which has dominated AI research since its inception can
be found in Newell and Simon’s (1976) and Smith (1985). This study will explore the use
of AI to develop predictive tools and modeling techniques for construction materials and
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 10
essential to capture the full range of knowledge required for intelligent behaviour, while
for a problem to find the most appropriate and final answer. Most Al-based techniques
modeling human performance, etc. offer reasonable answers for the above two concerns
and share similar approaches to problem solving. However, they tend to be brittle and
very sensitive to noise rather than degrading gracefully. Such techniques tend to either be
right or fail completely (Luger and Stubblefield, 1998). A different mechanism (artificial
neural networks) based on a structure parallel to that of neurons in the human brain tends
incomplete data exist. The following sections of this chapter outline the two AI
techniques used in this study namely, Genetic Algorithms and Artificial Neural
Networks.
The fast development of powerful computers gave a robust push in solving problems that
were first proposed by Holland (1975) as a global optimization technique for complex,
nonlinear, and noisy problems. It is a probabilistic global search method based on the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 11
mechanism o f natural selection and genetics of biological evolution. The power of using
solution.
Typically, GA starts the search with an initial coded set of random individuals
(genes) and represents a solution to the problem at hand. The fittest individuals in the
initial population are selected to mate and through genetic operators, a new generation of
their parents is produced. A new selection process is again performed and only the fittest
offspring are selected to mate so that another generation with a genetic structure superior
to that o f their parents is developed. After several generations, it is fair to assume that no
improvement will take place in the development of newer generations and the optimal
genetic content o f individuals that satisfy the objective and constraints imposed on the
system has been reached (optimal solution to the problem). A schematic representation of
The genetic algorithm approach has been successfully used in several engineering
applications. However, it was not until recently that GA has been used in concrete
research and structures optimization. Ramasamy and Rajasekaran (1996) investigated the
feasibility o f using expert systems, artificial neural networks and genetic algorithms, to
optimize the design o f industrial roofs. They concluded that results obtained using
genetic algorithms compared favourably with those of other expert system methods.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 12
deep beams. They reported a good agreement between model predictions and
experimental results. The genetic algorithms technique was also used in concrete research
to model the compressive strength of cement mortar (Akkurt et al. 2003) and in the
Start
Generate initial
population
Evaluate
objective
function
Generate new
population
Assess
Yes Stop
optimization
criteria
No
Selection
Crossover
Mutation
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 13
searched for a global optimum. It differs substantially from other traditional search and
optimization methods. The most significant aspects that differentiate GAs from other
• GAs deal with a representation of the parameter set, not the parameter set itself
The following sections describe the three major genetic operators that constitute the
2.2.1 S e l e c t io n
As stated earlier, the power of using GAs as optimization methods lies in their ability to
search a population o f potential solutions on the principal of survival of the fittest. Thus,
the selection process provides the driving force in a genetic algorithm. It is a simulation
of natural selection o f the best-fitted individuals who can adapt to changing environment
for reproduction in the forthcoming generations. A variety of selection methods are used
in the process o f GAs. A brief description of the most common ones is provided below,
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 14
The roulette wheel is the simplest selection method used in GAs. It is based on the fitness
value of the individual, meaning that healthier individuals have higher probability to be
selected for mating to produce more offspring for the next generation. The probability of
selecting an individual (pi) is simply calculated by dividing its fitness value (v;) by the
Each individual in the initial population is represented by a segment equal in size to its
generation size. The individual whose segment spans any of the random numbers is
selected. This process continues until the desired number of individuals in the new
generation is obtained. Note that an individual can be selected more than once. The
following is a simple example of how the roulette wheel selection method works. An
initial population o f 7 individuals along with their assigned fitness values is shown in
Table 2.1. The probability o f each individual being selected is calculated using equation
Table 2.1 - Selection probability of individuals using the roulette wheel selection method
Individual 1 2 3 4 5 6 7
Selection probability (pi) 0.2 0.1 0.26 0.13 0.16 0.09 0.06
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 15
Since individual 3 has the highest fitness value it will have the highest selection
probability and occupies the largest interval on the selection line (Fig. 2.2). Thus, it is
highly expected that this individual will be selected for mating. If a new generation of 5
individuals is needed, a set of 5 random numbers (less than one) is selected (0.5, 0.15,
0.9, 0.1, 0.36) and each number is mapped over the selection line (Fig. 2.2). The
individual whose segment spans the random number is selected. For this particular
example, the individuals selected for mating to generate a newer generation will be 1, 1,
+ ---- h I * I
0.0 0.2 0.3 0.56 0.69 0.85 0.94 1.0
1 5 I 6 I 7 I
Selection probability of individuals in initial population
The selection o f the fittest individuals from the initial population using the stochastic
universal sampling (SUS) is performed in a similar manner to that of the roulette wheel
method. However, equally spaced pointers are used instead of a single one. The pointers
are placed over a line representing the fitness values of all individuals in the population
and the individual whose segment maps any of the pointers is selected for mating. The
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 16
This method consists o f randomly choosing a group of individuals from a population and
the fittest individuals o f the group are selected for mating. The number of individuals in
the group is called tournament size. The process continues until the desired number of
individuals (parents) is selected. The selected parents then mate to produce a new
generation of offspring. This method employs both random and deterministic features
simultaneously.
2 .2 .2 C r o s s o v e r (R e c o m b in a t io n )
simulates the natural genetic process of creating a child (offspring). This is performed by
combining genes (strings) from two parents to hopefully create a better genetic structure
probability or crossover rate, which is defined as the ratio of the number of offspring
produced in each generation to the population size. The crossover rate should be carefully
chosen. A relatively high rate value permits the search to cover most of the solution space
and reduces the chances o f being trapped in local minima. Whereas, a too small or too
high rate value unnecessarily prolongs the time needed for the search process to converge
by either slowly exploring the solution space (too low rate value) or exploring fruitless
regions in the solution space (too high rate value). Many crossover or recombination
methods exist for both binary and real-value coding. Recombination methods for real
value coding such as discrete, intermediate, and linear combination are slightly more
complex than those o f binary coding. However, they are all based on the same principals.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 17
For binary-coding, another three methods of crossover exist: single-point crossover, two-
point crossover, and uniform crossover. Only the single point crossover is briefly
explained below while more detailed description on the other methods can be found in
offspring from two parents by simply merging genes (strings) on one side of a cut-off
single point from one parent with those on the other side of the cut-off point from the
other parent. Figure 2.3 shows a schematic representation of a single point crossover
method.
Cut-off point
h 1 1 1 0 0 1 1 0 1 0 1 0 1 1
\7
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 18
2 .2 .3 M u t a t io n
Mutation is considered as a background operator in GAs. Its role is to guarantee that the
probability o f searching any given string will never be zero and to recover good genetic
material that may have been lost through the action of selection and crossover (Goldberg,
1989). Mutation in natural evolution is a random process where one allele of a gene
replaces another to produce new genetic structures. In GAs, mutation is applied to modify
elements in the chromosomes with a given probability rate. This will introduce new genes
to the existing population and help guide the search. Again the mutation rate should be
carefully chosen so that useful genes in the chromosomes are not lost and/or the offspring
will not lose their resemblance to their parents. Wright (1991) and Janikow and
Michalewicz (1991) demonstrated how using a high mutation rate in real-coded GAs is
Artificial neural networks (ANNs) are powerful computational tools inspired by the
understanding and abstraction of the structure of biological neurons and the internal
operation of the human brain (Haykin, 1994). The concept and methodology of ANNs,
also known as connectionist models or parallel distributed processing, are not new.
Research in this field was first started in the early 1940’s after the introduction of
simplified neurons by McCulloch and Pitts (McCulloch and Pitts, 1943). However,
genuine applications o f ANNs effectively started in the early 1980’s with the work of
Hop field (1982) and the introduction of error back-propagation using the generalized
delta rule developed by Rumelhart et al. (1986). With new developments in learning
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 19
networks and their ability to learn from examples and to capture hidden behaviour
between patterns o f inputs and outputs have been well recognized. Nowadays, the interest
materials.
process, and model validation. The architecture of any neural network is based on the
same building block (neuron), in which simple computation is performed and might
contains several neurons or processing units. Depending on its current activation level,
each processing unit send signals to other processing units in the network and through
propagation, the network learns and adapts to new examples and stores the knowledge it
The nonlinear nature o f neural networks makes them suitable to solve virtually any
pattern recognition. For instance, a neural network has the ability to “learn” the mapping
between a set o f input data and its respective output, and to synthesize through training an
associative memory that may generate appropriate output when presented with an
unfamiliar set o f input. According to Rumelhart et al. (1986), the basic aspects that
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 20
connectivity among processing units, 3) state of activation for each processing unit, 4)
propagation rule, 5) activation function for each processing unit, and 6) rule of learning.
These aspects vary from one network to another leading to a wide variety of network
types. Some o f the well-known neural networks include the Hopfield network (Hopfield,
1982), Boltzmann machines (Ackley et al., 1985), the Kohonen network (Kohonen,
1982), and the Multi-layer back-propagation neural networks (Rumelhart et al., 1986).
The Multi-layer back-propagation networks are the most commonly used neural networks
concrete, illustrate some o f their important aspects, and also provide, through a case
2.3.1 N e u r a l N e t w o r k s A p p l ic a t io n s in M o d e l in g C e m e n t -B a s e d M a t e r ia l s
inherently different properties of various elements, and, sometimes on the dual and/or
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 21
often lack the ability to either account for the effect of all parameters involved, or to
account for the combined effects of variables considered. Over the last decade, artificial
neural networks (ANNs) have been widely investigated as a tool for modeling and
feed-forward back propagation (FFBP) networks (defined in detail later in this text), have
The attractiveness o f using ANNs in modeling materials behaviour arises because they
are trainable dynamic systems, capable of predicting the engineering properties of a given
material based on existing data, so that the researcher does not have to make assumptions
to fit the data to a certain model. Given adequate learning materials and proper training,
ANNs can be taught the embedded relationships between patterns of inputs and outputs,
and using its predictive capability to generalise to new domains in the neighbourhood of
the training data. Compared to traditional modeling techniques, the popularity of ANNs
has grown rapidly and applications in modeling the behaviour of cement-based materials
Ghaboussi et al. (1991) investigated the utilization of ANNs to model the stress-strain
behaviour o f plain concrete subjected to different loading conditions. Goh (1995), Sanad
and Saka (2001), Yeh (1998), Mukheijee and Biswas (1997), and Lee (2003) studied the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 22
feasibility o f using ANNs to predict the mechanical strength of concrete under different
conditions. ANNs have also been proposed as an effective tool to model concrete
durability. For instance, Glass et al. (1997) investigated the use of ANNs to model
chloride binding in concrete. Haj-Ali et al. (2001) and Buenfeld et al. (1995 and 1998)
proposed different neural network models to predict the durability of concrete subjected
to various degradation mechanisms. Recently, FFBP networks have also been used to
determine the mixture proportions of ordinary concrete (Oh et al. 1999), and for
predicting the rheological and hardened properties of special concretes such as self-
The interest in using back-propagation ANNs for modeling the behaviour of cement-
based materials and properties of special concrete is a relatively recent development and
the technique has been used by only a limited number of researchers. To date, there are
still no rules on how to choose a suitable architecture for a FFBP neural network and how
to optimize its performance. The rest of this chapter examines the important steps in
building a FFBP neural network-based model, the parameters that affect the behaviour of
the network during the training process, the acceptance/rejection of a trained neural
2.3.2 F e e d -F o r w a r d B a c k -P r o p a g a t io n N e u r a l N e t w o r k s
the most widely used neural networks in engineering applications, e.g. in modeling the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 23
problems, recognize logic functions, and subdivide the pattern space for classification
problems. MLP networks are parallel structures consisting of multiple layers, and each
layer may contain a large number of perceptrons or processing units. The perceptron,
similar to the artificial neuron introduced by McCulloch and Pitts (Fig. 2.4), was first
x.
(2 .2 )
calculates a weighed sum, and adds a threshold value dto form a net input U,
n
(2.3)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 24
where x, is an input parameter and w,- is the connection strength or weight between the
perceptron and each input. The result is then passed through a non-linear transfer
limit non-linear transfer function (Fig. 2.5), its capabilities were limited to the application
of simple logic functions and linear decision boundaries. However, grouping perceptrons
Rosenblatt’s perceptron constitutes the building block of a MLP network. However, its
understanding of all three stages necessary to build, train, and test such a network. First,
one has to decide on the appropriate architecture of the MLP network. Secondly, training
a MLP network is a complex operation and one has to balance the time consumed by
training against the anticipated performance. Finally, evaluating the performance of the
ANN and its ability to generalize its predictions to new data unfamiliar to the network is
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 25
Hard limit
1.0
0 = 1.0
0.8
0 = 0.2
0.4
0.2
0.0
■10 -8 -6 ■4 -2 0 2 4 6 8 10
U
Fig. 2.5 - Example o f activation functions: hard limit and logsig (sigmoidal functions).
and a number o f hidden layers. Each layer may contain several processing units that are
fully or partially connected to units in the subsequent layer with different strengths or
permitted. Fig. 2.6 shows a typical architecture of such networks. Although some
literature does not consider the input layer as an integral layer in the network architecture,
it is agreed that units in the input layer do not perform any computation and only serve as
a link between the input vector and units in the first hidden layer.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 26
The number o f units in the input and output layers depends on the input and output
parameters included in the training patterns, which are usually defined. However, there
are no commonly accepted rules to determine the optimum number of hidden layers or
the number o f hidden processing units (PUs) in each hidden layer used to optimize the
performance o f a particular MLP network in a given task; more research is still needed in
this area. It is understood, however, that the number of hidden layers and their respective
numbers of units depend on the number of patterns available for training along with the
Fig. 2.6 - Typical architecture o f MLP back-propagation network with one hidden layer.
hidden units. However, this issue is normally resolved by trial and error and is left to the
experience o f the operator. One of the rough rules of thumb for determining the number
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 27
where: PU is hidden processing unit in the hidden layers, P is the number of patterns
(input-output pairs) in the training database, n and K are the number of units in the input
and output layers respectively, and a is a constant with recommended value varying
Although equation (2.4) serves as a starting point for a process that to date is only
resolved by trial and error, the total number of processing units (PUs) in the hidden layers
explained later, the performance of a trained MLP network depends on the final set of
connection strengths (weights) between its processing units: decreasing the number of
hidden PUs to an excessively small number may decrease the number of connections and
their associated weights, thus reducing the ability of the network to implement non-linear
hidden PUs, however, to the point where the total number of connections and their
associated weights is much higher than the number of training patterns might slow down
the training process and reduce the ability of the network to generalize.
As stated earlier, ANNs are attractive predictive tools because they are data-driven
trainable systems that have the ability to learn from examples and to adapt to unknown
between a set o f inputs and outputs. However, this learning process is usually complex
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 28
and depends on several undefined parameters. The main objective in training (teaching) a
between its processing units for which the ANN can predict accurate values of outputs for
unsupervised manner. Unsupervised training means that the network is presented only
with input parameters and it must learn on its own the regularities and similarities among
training vectors, whereas supervised training provides the network with training patterns
that include both the input parameters and the corresponding outputs: the network is told
what to learn.
Training a MLP network involves two phases: i) feed-forward and ii) back-propagation.
In the feed-forward process, the data flow from the input units in a strictly forward
manner to predict network outputs and compare them to measured targets. In the back-
propagation process, the error between the predicted and measured outputs is propagated
backward from the output layer to modify the network connection strengths in order to
improve its performance. Fig. 2.7 schematically shows a typical flow pattern in the
"Weights^
Comparison
adjustmei
Measured
output
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 29
Computational activities in a MLP network essentially occur in the hidden and output
processing units, and the optimum set of connection weights after network convergence
(reaching a minimum global error) determines the network’s performance and its ability
to generalise. However, the training process is not only a function of the network
architecture and the learning method (supervised or unsupervised training) but it also
depends on the selection o f training data, as well as several other parameters that need to
be defined before training starts. These include the learning rate, the network momentum,
the transfer function, and the duration of training. These parameters will be discussed in
sequence below.
ANNs are data-driven modeling systems, and their success largely depends on the
learning materials provided for their training. Therefore, it is vital to generate a training
database that contains adequate information necessary to teach the ANN to capture the
underlying relationships between a set of inputs and outputs. Two important principals
i. The training data should be comprehensive, meaning that it must contain relevant and
complete information on the relationships between the inputs and the outputs of the
ii. The training data should be large enough and continuous within the practical domain
of the application being studied. Furthermore, it should be free of outliers (data that
are uncharacteristic o f the application domain) to improve the training of the network
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 30
Although MLP networks can accept data from a broad range of sources, they are only
recommended to scale the training patterns (input and output vectors) in order to speed up
the training process and improve network generalisation. The transfer function
recommended for MLP networks is normally a sigmoid function with an upper limit of 1
Therefore, scaling the training data (and especially the outputs) to fall within the range of
the transfer function is critical and is usually performed using any desirable scaling
v (X ~ X m i n )
0 < x, < 1 (2.5)
x-x
0 < x, < 1 (2.6)
where: x, is the scaled value of variable x, xmin, xmax, x , and Sd are the lower limit, the
upper limit, the average value, and the standard deviation of the training data,
respectively.
The learning rate (a scalar parameter) is the step size of weight and threshold changes
conducted by the learning algorithm in the back-propagation process for tracking down
the global minima in the error surface of the network. The error surface represents the
changes in network errors, calculated as the difference between the predicted outputs and
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 31
the provided targets during the training process. A large learning rate tends to accelerate
the learning process. However, it may cause the network to diverge from the global
minima in the error surface to favour local minima or the network may never converge to
a desirable solution. Conversely, a small learning rate may dramatically reduce the
learning speed, hence limiting the ability of the network to escape local minima in the
error surface, and forcing it to diverge to an undesirable region of the weight space
(searching for undesirable values of connections’ strength). The selection of the learning
speed up the learning process without leading to network’s oscillations. In most cases, a
high learning rate is usually used along with a momentum term and it is normally set to a
2.3.2.2.3 Momentum
The Momentum in a MLP network is another parameter in the learning algorithm that
controls changes in the weight space and ensures that the search movement for global
and it usually varies between 0 and 1. However, it is recommended (if the flexibility of
the software used permits) to either start with large values of learning rate and
momentum and then reduce them as the learning progresses or compliment a high
learning rate with a low momentum and vice versa (NeuralWare, 2001).
evaluate and transmit incoming signals to units in the subsequent layer. This function
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 32
must ensure the stability o f the network and allow non-linear transformations to be
implemented at each layer, otherwise the MLP network reduces to an equivalent single
layer network. A wide variety of activation functions are used in ANNs such as linear,
sigmoid function (Fig. 2.5) be employed for MLP networks because it is non-linear,
1 and 1 (tan- sigmoid) as U (the net input calculated by each processing unit), varies from
(2 .8)
where /? is a constant that determines the steepness of the transition zone, that is, the zone
desired minimum error between its predicted outputs and the desired targets provided in
the training patterns. However, the performance of a MLP network is best evaluated by
its capability to generalize its predictions to unfamiliar data (new data not used in the
training process). A minimum training error does not necessarily assure a minimum
order to converge to a desired minimum error, it may accommodate weight values that fit
the relationships o f all training patterns, including the imprecise ones. This problem is
commonly known as over-fitting (Fig. 2.8); it may lead to a precise prediction of the
training patterns on one hand but poor generalisation of new patterns on the other hand.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 33
Conversely, if the training is stopped prematurely, the network might not adequately
learn the relationships between inputs and outputs, leading to unsatisfactory performance
2.5
Well-trained
x
-II 1.5
>* Over-trained
0.5
The duration o f training a MLP network is data-dependent, and the user has to balance
the time consumed in the training process and the degree of accuracy needed for the
training a MLP network is to divide the available database into training and validation
sets. The training set is used to teach the network the embedded relationships between
input and output data, while the validation set is used to monitor the generalization
capability o f the network during training. Training is usually stopped after the network
error generated from the validation set concurs with that generated from the training set
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 34
and jointly converge to a desired minimum error, or when the validation error starts to
Unfortunately training data, especially that generated experimentally, is often limited and
there is usually need for additional training information to achieve better network
performance. Therefore, it might not be feasible to create a validation set of data at the
expense o f reducing the training set. In such a case, the duration of training can be
decided by either a parametric study or by trial and error. Some of the mechanisms for
stopping training are: i) limiting the number of training epochs (each epoch equal to a
predetermined number o f training patterns after which a network error is calculated), ii)
setting a desired minimum error so that when it is reached the training is stopped and the
network is tested and iii) monitoring the trend of error improvement so that training is
stopped when little or no improvement in the training error is reported over a given
number o f epochs.
Once suitable network architecture is selected and the training data has been screened and
normalized, training can start. The supervised training process involves presenting the
training data in a patterned format. Each pattern contains an input vector and its
corresponding outputs (targets). The network then tries to capture the effects that each
input exerts on the outputs by adjusting a randomly initialized weight space to minimize
the error between the network prediction and the specified targets. Training usually
progresses as follows:
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 35
i) Initialize the weight space [W~ ] (usually by small random values between -0.1 and
+ 0 . 1).
....” 1
K ■ • wL
w 21 + 22 ■
fc]= (2.9)
.
X
.w m
‘ + fl2 •
3
i
Where: [W‘t ] is the weight matrix (strength of connections) between units in layer / and
/-I (/ varies from 2 to AO, w'j. is the strength of the connection between unit j in layer I
and unit i in layer /-l, N is the number of layers, and m and n are the number of units in
ii) The network is presented with the first training pattern and each processing unit j in
uf W2i
JS
V , •
U2 W 2, * 4 • • wL x\ el
>— X - >+ * > (2 . 10)
1=1
iii) Using the assigned activation functions {/j2}, each unit j in the first hidden layer
activates an output signal to form an output vector {Yj } that serves as an input vector
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 36
y2 \A2(u?) V '
y 2i fiW l) xl
tN
b
{Yj} = < ■ = {*)} = (2.11)
II
II
A
V
>
•
s
S.. fivl). *1
iv) The same process continues through all hidden layers until each unit k in the output
u ; w ,2 W l. xr' V
n
u 2 w 2, w 22 W 2„ xr (2 . 12)
+
>+ - V
II
. X <
p , ) = 2 v r
7=1
■
*r.
s*
V*. W t2
n
Where: £/* is the net input of unit k in the output layer, k is the number of units in the
output layer, n is the number of units in the last hidden layer, Xj~l is the input value from
unit j in layer N-l, and wkj is the strength of the connection between unit k in the output
With the network predictions listed as outputs in the output layer, the feed-forward phase
is completed. However, these predictions (based on the initially assigned weight space)
can substantially differ from the measured targets that were provided in the training
pattern. Therefore, the search must continue for a new set of weights that minimize the
difference between the outputs and the corresponding targets. This is when the back-
i) From the output layer, the network calculates and stores an error vector {£/},
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
where: e* is the error at unit k in the output layer, o* and tk are the predicted output at unit
k in the output layer and the corresponding target of the first input vector, respectively.
ii) The network is presented with the second training pattern, the same procedure described
above is followed and the error vector {E2} is calculated and stored.
iii) The introduction of training patterns continues until all data available for training (or a
At the end of each epoch, the network calculates an average error (Esl) for all patterns as follows:
l 4 4
(2.14)
Est is the system error, p is a training pattern, P is the number of training patterns
assigned to one epoch, and opk and tpk are the predicted output and provided target of
v) If Est reaches a desired minimum value, training is stopped and the performance of the
network to adjust the weights and thresholds in a gradient search for the desired minimum
system error.
vi) The adjustment of the weight space takes place via an appropriate learning algorithm. The
generalised delta rules: a gradient descent algorithm, in which the weights are adjusted
iteratively by:
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 38
Where: w'(.(m) is the weight value at epoch m, w';(m + l) is the weight value at epoch
m+1, and Awlj i(m)is the weight adjustment generated by the system error at epoch m and
is equal:
Aw>)=z^ d ji
<2 1 6 >
where n is a positive constant called the learning rate and f — is the partial
dWji
derivative of the system error with respect to each weight value in the network. However,
d (E J d (E „ )J U ‘
(2.17)
dw^ dUlj dwL
The weight between unit j in layer I and units in layer /-I should be changed by an
27? \
amount proportional to the term ■ and the input value provided by unit i in layer /-l.
The implementation o f the chain rules and weight adjustment using a back-propagation
This iterative process continues until it finds a set of weights that minimizes the system
error to a desired value and/or leads to satisfactory performance of the network. If the
available training data are large enough to accommodate the creation of a validation set,
the network will calculate a validation error in a similar manner to calculating the training
error. However, this validation error plays no role in the adjustment of connection
weights and is only used to monitor the performance of the network during training.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 39
performance on the training data. In the latter case, the network is only trained to create
memory storage for all patterns used in training and to accurately match between each
successfully trained MLP network rather depends on its response to new input data; the
network has neither prior knowledge of such data nor of its associated outputs. Therefore,
to evaluate the performance o f a MLP network after successful training, it must be used
to predict the outcomes o f new input data (testing data) and its response compared to
corresponding targets. Only then can the ability of the network to generalise and associate
be evaluated. For this purpose, new data sets (testing sets) containing information only
about input parameters are presented to the network. Using the final set of weights
obtained after successful training, the network will be able to generate an output for each
input vector in the testing data sets (equations 8, 9 and 10). The generated outputs are
compared with known targets and the percentage error is calculated using an appropriate
evaluation method.
As stated earlier, the performance of MLP networks depends on several parameters such
as the network architecture, the length of training, the activation and learning functions,
the learning parameter, and the momentum. The effect of the activation and learning
parameter and momentum are well documented (Haykin 1994 and NeuralWare 2001)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 40
a suitable MLP architecture for a certain application and how long the training process
should continue are still unsolved problems, and this issue is normally resolved by trial
and error. The effect o f training duration, number of hidden units in one-hidden-layer-
investigated below.
A total o f 248 compressive strength data obtained on concrete cores drilled from various
concrete beams o f different compressive strengths and subjected to different drilling and
curing conditions were obtained from literature (Bartlett 1994). Such data are used herein
to study the ability o f MLP networks to predict the in-situ compressive strength of
concrete structures and to investigate the effect of network architecture on the training
and validation processes. The network in question has an input layer containing 9
cylinder strength was measured, curing condition (i.e. wet, air, or sealed), duration of
curing, location in the beam from which the core was obtained, direction of drilling
(parallel or perpendicular to concrete casting), age of concrete at which the core was
tested, diameter, and length to diameter ratio of concrete core. The network also has an
Analysis of the database identified 17 data sets as outliers; these were disregarded. A data
set was considered as an outlier if it does not contain all required information about
input-output parameters, and/or if the value of any of its input parameters is isolated from
the cluster o f the associated parameter in other data sets. Table A .l (Appendix A)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 41
includes sample data from the database. A total of 19 data points (randomly selected from
the database, shaded rows in Table A .l) were used to evaluate the network’s
performance, while the remainder of the data was used for training. Numerical input and
output data were scaled between 0 and 1. Each parameter was divided by the largest
value in its set while variables such as curing condition (moist, air, or sealed cured) were
assigned qualititative values. After several trials, a network architecture was adopted
consisting o f an input layer, an output layer and a single hidden layer of 5 units; it
provided best network performance (lowest prediction error on testing data). The learning
rate considered in this case was 0.1, the momentum factor was 0.5, and each unit was
assigned a logarithm sigmoid function (logsig) as a transfer function. Fig. 2.9 shows the
performance o f this MLP network in predicting the in-situ concrete strength of the
training and testing data sets with an average absolute error of 2% and 3%, respectively.
o Training data
B Testing data
— Equity line
o>0.8
= 0.6
■g 0.4
a. 0.2
AAE (training) = 2 %
AAE (testing) = 3 %
Fig. 2.9- Performance o f MLP network in predicting the in-situ strength o f concrete.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 42
Training a MLP network is achieved by teaching the network the embedded relationships
between a set o f inputs and their respective targets. In other word, the goal is to minimize
the differences between the outputs predicted by the network when it is presented with a
set of inputs and their respective specified targets. Perhaps due to the relatively recent
interest in using ANNs for modeling materials behaviour, new users might assume that
the smaller the training error (the error between the outputs predicted by the network and
the provided targets for the training data) the better the network will perform. However,
minimizing the training error does not necessarily assure better network performance and
one has to decide, through experience, on when to stop training and avoid over-fitting,
explained earlier in this text. If the database is large enough, a good method to decide on
when to stop the training process is to divide the database into two sets (training and
validation) and monitor the training and validation errors during the training process as
explained earlier.
In this study, the effect o f the training duration of a MLP network with one hidden layer
and 10 hidden units was investigated. Other network’s parameters and initial values of
weights (connection strengths) and biases were kept constant for each training process,
and the network was trained for different numbers of epochs. In this study, an epoch is
considered to be complete after all patterns available for training are presented to the
network. After each training process, network performance was evaluated by its ability to
predict the in-situ concrete strength of a set of testing data (not used in the training) and
the average absolute error (AAE) of training and testing data was calculated as per Eq.
(2 . 18).
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 43
i * i yme<w
i„ . pred
AAE=-'Z (2.18)
Where 7meai = measured value o f in-situ concrete strength from experimental data, Ypred =
predicted value o f in-situ concrete strength by the network, and n = number of data
points.
It is clear from Fig. 2.10 that the training error decreased when the training time (number
of epochs) increased. However, the testing error decreased until a threshold number of
epochs around 500 beyond which it increased rapidly, probably due to over-training. As
explained earlier, there is no clear rule to accurately determine the duration of training
that provides best network’s performance (minimum testing error); this is best resolved
12.0
10.0
o5N
■_ 7.37 7.65
O 8.0 6.84
u
ui
6.0
4.0
2.3< ■
2.0
0.0
200 400 800 1600 3000 5000
Number o f ep o ch s
Fig. 2.10 - Effect o f training duration (number of epochs) on the performance of MLP
network.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 44
MLP networks with one hidden layer may be capable of modeling a large number of
optimum number o f hidden units that can provide best performance of such one-hidden-
layer networks is not easy to define. It is believed that using a large number of hidden
units could enhance the training process but might not be beneficial to the overall
The study was based on the network’s architecture adopted earlier in predicting the in-situ
strength o f concrete structures, which consists of one hidden layer with 5 units.
Moreover, 6 additional single-layer MLP networks with the same network parameters but
having different numbers of hidden units were created from the above-adopted network.
All networks were trained for the same number of epochs (500), where each epoch
consisted o f all (212) training patterns. It should be mentioned that the number of epochs
(500) was chosen after each network was trained and evaluated separately for different
numbers o f epochs and it was observed that all networks performed best at around 500
epochs. Fig. 2.11 shows the effect of the number of hidden units on the performance of
the various networks. It can be observed that adding more hidden units consistently
training error with increasing number of hidden units. However the enhancement of
units (5 units for this particular application) but decreased thereafter. Such a trial and
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 45
error approach is recommended to define the number of hidden units that assures lowest
6.0
5.0
g
£ 4 0
L_
LU
3.0
2.0
1.0
0.0
3 4 5 7 10 15 20
Number of hidden neurons (one hidden layer)
Fig. 2.11 - Effect o f number of hidden neurons on the performance of MLP network.
To study the effect o f the number of hidden layers on the performance of MLP networks,
a set of network architectures were trained and validated using the compressive strength
where MLP-«-(/, means that the MLP network has n hidden layers with
including the total number of hidden units, the minimum desired error, and the number of
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 46
Table 2.2 shows that the number of connections between units for the various networks
depends on the scheme of hidden layers and the number of units they contain. For the
same number o f epochs, the duration of training increased with the number of
connections but did not depend on the number of hidden layers. Moreover for the same
total number o f hidden units, training epochs, and constant network parameters, Fig. 2.12
shows that the number o f hidden layers and the number of connections between units did
not have a clear trend, whether on training error or testing error. This exercise shows that
the single layer network had significantly lower testing error. Therefore, no general
recommendation with regards to the number of hidden layers can be made and a similar
trial and error procedure is needed to define the best network architecture.
10.0
8h .
UJ
*
S ' V '
Network architecture
Fig. 2.12 - Effect of number of hidden layers on the performance of MLP network
(constant number of hidden units).
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 47
Table 2.2 - Effect o f number of connections and hidden layers on training and testing
error of MLP Networks (after 500 epochs)
It is agreed upon that the selection of the learning material (training data) is the most
performance o f a MLP network and its ability to generalize its predictions depend to a
great extent on the quantity and quality of the database generated for training. The
quantity o f the learning material is important to continuously cover the practical range of
input data, while its quality helps the network to accurately learn the important factors
that affect the behaviour o f the phenomenon being modeled. In addition to the importance
• If the data available for training is large enough, creating a validation set (separate
from the training set) and monitoring the performance of the network during
training is the most recommended approach to determine when the MLP network
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 48
• It is suggested that using a relatively large learning rate along with a small
• It is proposed that the learning rate and momentum should be reduced as training
progresses, to speed up the learning at the early stage but avoid network
escalation (reducing the step size of weight changes so that the network will not
skip the global minimum error) when training is near its end.
However, the most adequate number of hidden layers depends on the complexity
of the process being modeled. The general rule of thumb is to use more hidden
relationship, and to use more hidden units (also up to an optimum number) in the
Optimum numbers o f hidden layers and hidden units can be defined by operator
• Using an epoch with fewer training patterns has proved to speed up the training
ANNs do not provide simple rules or equations for predicting material behaviour. The
product o f the ANN methods is trained networks which when presented with new input
data can rapidly predict the corresponding output data. However, the process by which an
ANN conducts its predictions is very complex as explained earlier and cannot usually be
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 49
There is growing interest in using ANNs for predicting the behaviour of concrete
materials and structures. However, the method still remains obscure for many researchers
applications, have been explained in detail. In particular, the mathematical basis, the
The ability o f MLP networks to predict the in-situ compressive strength of concrete
structures was used as a case study to investigate the effect of network parameters on its
overall performance. Using a database of 231 strength values obtained on concrete cores,
the MLP network was able to predict the in-situ compressive strength with an average
absolute error o f 3 %. It was shown that longer training and/or minimizing the training
error did not necessarily lead to a better network performance in the generalization of
predicting new data unfamiliar to the network, and it was recommended to optimize the
duration o f training using a validation set of data. It was also observed that the
performance o f MLP networks during training improved with higher numbers of hidden
units, whereas their generalization was best around an optimum number of hidden units.
For a constant number o f hidden units, there was no clear trend for the effect of the
number o f hidden layers and number of unit connections on the network’s performance.
Therefore, best network architecture could only be defined by trial and error.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 50
Unlike traditional modeling methods, no assumptions are needed in ANN model creation.
Thus, the use o f ANNs in concrete materials and structures modeling is expected to go
beyond its current infancy level. This chapter should clarify the method and make it more
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 51
2.5- REFERENCES:
Ackley D. H., Hinton G. E., and Sejnowski T. J., (1985), “A Learning Algorithm fo r
Boltzmann Machines,” Cognitive Science, V. 9, No. 1, pp. 147-169.
Akkurt S., Ozdemir S., Tayfur G., and Akyol B., (2003), “The Use o f GA-ANNs in
Modeling o f Compressive Strength o f Cement Mortar,” Cement and Concrete
Research, V. 33, No. 7, pp. 973-979.
Ashour A. F., Alvarez L. F., and Toropov V. V, (2003), “Empirical Modeling o f Shear
Strength o f RC Deep Beams by Genetic Programming,” Computers and Structures, V.
81, No. 5, pp. 331-338.
Buenfeld N. R., Hassanein N. M. and Jones A. J. (1998), "An Artificial Neural Network
fo r Predicting Carbonation Depth in Concrete Structures," in "Artificial Neural
Networks for Civil Engineers: Advanced Features and Applications", (Flood I. and
Kartam N. editors), ASCE, pp. 77-117.
El-Chabib H., Nehdi M., and Sonebi M., (2003), “Artificial Intelligence Model fo r
Flowable Concrete Mixtures Used in Underwater Construction and Repair, ” ACI
Materials Journal, V. 100, No. 2, pp. 165-173.
Ghaboussi J., Garrett J. H., and Wu X., (1991), ‘‘‘K nowledge-Based Modeling o f Material
Behavior with Neural Networks,” Journal of Engineering Mechanics, V. 117, No. 1,
pp. 132-153.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 52
Glass G. K., Hassanein N. M., and Buenfeld N. R., (1997), “Neural Network Modeling o f
Chloride Binding,” Magazine of Concrete Research, V. 49, No. 181, pp. 323 - 335.
Goh A. T. C., (1995), “Prediction o f Ultimate Shear Strength o f Deep Beams Using
Neural Networks, ” ACI Structural Journal, V. 92, No. 1, pp. 28-32.
Haj-Ali R. M., Kurtis K. E., and Sthapit A. R., (2001), “Neural Network Modeling o f
Concrete Expansion During Long-term Sulphate Exposure, ” ACI Materials Journal,
V. 98, No. l,p p . 36-43.
Hopfield J. J., (1982), “Neural Networks and Physical Systems with Emergent Collective
Computational Abilities,” Proceedings, National Academy of Science, USA, V. 79,
No. 8, pp. 2554-2558.
Hush D. R., and Home B. G., (1993), “Progress in Supervised Neural Networks,” IEEE
Signal Processing Magazine, V. 10, No. 1, pp. 8-39.
Janikow C. Z., and Michalewicz Z., (1991), "An Experimental Comparison o f Binary and
Floating Point Representations in Genetic Algorithms,” Proceedings ICGA, 4, pp. 31-
36.
Lee S-C., (2003), “Prediction o f Concrete Strength Using Artificial Neural Networks, ”
Engineering Structures V. 25, No. 7, pp 849-857.
Lim C-H, Yoon Y-S, (2004), “Genetic Algorithm in Mix Proportioning o f High-Strength
Concrete,” Cement and Concrete Research, V. 34, No. 3, pp. 409-420.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 53
Luger G., and Stubblefield W., (1998), “Artificial Intelligence: Structures and Strategies
fo r Complex Problem Solving,'" Addison-Wesley Longman, Inc., Reading
Massachusetts, USA, 824 pages.
McCulloch W. S., and Pitts W., (1943), “A Logical Calculus o f the Ideas Immanent in
Nervous Activity,” Bulletin of Mathematical Biophysics, V. 5, pp. 115-133.
Mukheijee A., and Biswas N. S., (1997), “Artificial Neural Networks in Prediction o f
Mechanical Behaviour o f Concrete at High Temperature, ” Nuclear Engineering and
Design V. 178, pp. 1-11.
Nehdi M., Djebbar Y., and Khan A., (2001), “Neural Network Model fo r Preformed-
Foam Cellular Concrete, ” ACI Materials Journal, V. 98, No. 5, pp. 402-409.
Newell A., Simon H. A., (1976), “Computer Science as Empirical Enquiry: Symbols and
Search,” Communications of the ACM, V. 19, No. 3, pp. 113-126.
Oh J. W., Kim J. T., and Lee G. W., (1999), “Application o f Neural Networks fo r
Proportioning o f Concrete Mixes, ” ACI Materials Journal, V. 96, No. 1, pp. 61-67.
Ramasamy J. V., and Rajasekaran S., (1996), “Artificial Neural Network and Genetic
Algorithm fo r the Design o f Industrial Roofs-A Comparison,” Computers and
Structures, V. 58, No. 4, pp. 747-755.
Rumelhart D. E., Hinton G. E., and William R. J., (1986), “Learning Internal
Representation by Error Propagation,” Parallel Distributed Processing, V. 1,
Foundation, MIT Press, Cambridge, Mass, pp. 318-362.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2 54
Sanad A., and Saka M. P., (2001), “Prediction o f Ultimate Shear Strength o f Reinforced-
Concrete Deep Beams Using Neural Networks, ” Journal of Structural Engineering, V.
127, No. 7, pp. 818-828.
Smithers T., Conkie A., Doheny J., Logan B., Millington K., and Tang M. X., (1990),
“Design as Intelligent Behavior: an A I in Design Research,” Journal of Artificial
Intelligence in Engineering, V. 5, No. 2, pp. 78-109.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 55
CHAPTER 3
3.1-INTRODUCTION
The main objective o f using flowable concrete is to enhance the structural performance of
concrete members and to alleviate the construction process of challenging and complex
applications. The required workability of such concrete depends on the complexity and
shape of the formwork, type of structural member (reinforced, highly reinforced, or non
reinforced), placement method, and type of construction. The focus of this chapter is on
investigating the behaviour and properties of a highly flowable special concrete called
can easily flow through congested reinforced concrete elements, compact itself, and
adequately fill voids and narrow formwork with minimal or no vibration and without
aggregates) and good stability (minimal segregation and bleeding), Khayat (1999).
’Parts of this chapter have been published in SCC 2005 conference, Joint Second North American
Conference on the Design and Use o f Self-Consolidating Concrete and the Fourth RILEM International
Symposium on Self-Compacting Concrete, Centre for Advanced Cement-Based Materials, Chicago, IL,
2005. A substantial part o f this chapter was also submitted to ACI Materials Journal for review.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 56
However, achieving these properties simultaneously is highly complex and not precisely
defined. The successful development of SCC requires manipulating some of the mixture
introducing some chemical admixtures such as high-range water reducing (HRWR) and
The development o f SCC represents one of the most significant recent advances in
concrete technology (Zhu et al., 2001). Since its inception in the late 1980’s, it has
become a very active research topic. The superior workability of such a concrete allows it
to flow through congested structural elements under gravity and adequately fill voids and
formwork without segregation and/or excessive bleeding, and with little or no vibration.
The ability o f SCC to flow under its own weight can be achieved by the addition of
chemical admixtures such as high-range water reducing admixtures (HRWR). Its ability to
to achieve without decreasing the viscosity o f the mortar matrix, which reduces the
ability o f SCC to resist the segregation o f large and dense coarse aggregate particles.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 57
(EFNARC, 2002). The amount of research literature on SCC has also been growing
rapidly. However, its technology is still evolving and additional research is needed to
further improve its behaviour. Several reliable and widely accepted test methods have
been developed worldwide to evaluate the properties of SCC, and significant progress has
been made to standardize such methods. Table 3.1 includes the most commonly used test
(2002).
Table 3.1 - Most common test methods used to measure the workability of SCC
Unlike other properties o f SCC mixtures, which can be assessed using numerous non
standard but fairly reliable and widely used test methods, there is still no consensus on
which technique can best characterize the potential of segregation in such mixtures. Thus,
setting remains the least understood property of SCC mixtures, and dedicated research is
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 58
still needed in this area. Previous research provided recommendations to either reduce
segregation in SCC by reducing the total volume of coarse aggregates (increasing the
Segregation is defined as the separation of coarse aggregate particles from the mortar
matrix during transporting, placement (dynamic state), and/or setting of fresh concrete
(static state). It is mainly due to the difference in materials densities and the relatively
low viscosity o f the mortar fraction considering the highly flowable nature of SCC
(Betancourt, 1988). A fresh SCC with poor segregation resistance can lead to a non-
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 59
Several test methods have been proposed to assess the risk o f segregation in SCC
mixtures, none o f which is widely accepted. Some of these tests are only based on visual
observations and offer no quantitative rating of segregation (e.g. the slump-flow test in
which the risk of segregation is visually inspected for the formation of a mortar hallow or
a coarse aggregate pile after the slump cone is removed). Other tests have been recently
developed including the settlement column test (Rooney and Bartos 2001), the
penetration apparatus (Bui et al. 2002), the V-funnel test (Ozawa et al. 1995), and the
GTM screen stability test (EFNARC 2002), which are claimed to capture the segregation
mechanism and numerically evaluate the ability of SCC mixtures to resist segregation.
However, several o f these tests usually require a large concrete sample, are time-
consuming and relatively difficult to perform, and more data on both their laboratory and
The primary objective o f this experimental research is to develop a simple yet reliable
test method to assess the risk of segregation in SCC mixtures based not only on the
difference in materials densities and viscosity of the mortar compound (static state), but
also to resemble conditions during concrete placement, such as the free fall and concrete
discharge from the chute o f a ready-mix concrete truck. The test is based on analyzing the
variation of the coarse aggregate content along the height of a fresh concrete sample for
different SCC mixture proportions and different placement conditions. The results of this
test were correlated to the average penetration depth of four penetration heads using a
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 60
mixtures, such as slump flow, slump flow T50cm, and compressive strength at various
In addition to measuring compressive strength at 1, 7, 28, and 91 days, five test methods
were used in this study to evaluate the workability of fresh SCC mixtures. These tests are
3.2 .1 S lu m p -F lo w a n d S lu m p -F lo w T 50Cm T e s t s
The slump-flow test was first proposed in Japan to evaluate the quality of underwater
concrete (JSCE, 1992). It is based on the conventional slump test method and consists of
using an ordinary slump cone as per ASTM C 143/C 143M “Standard Test Method for
Slump of Hydraulic Cement Concrete”. The average diameter of the concrete spread after
lifting the slump cone is considered as an indication of the ability of SCC to freely flow
under its self-weight in the horizontal direction with the absence of obstacles (Fig. 3.2).
The time for the SCC to reach a 50 cm spread (T50) is taken as an indication of the rate of
concrete deformability. In general, a slump flow value between 500 and 750 mm and a
value of T5o varying between 2 and 7 sec are considered acceptable for flowable concrete
design, whereas some associations require a minimum slump flow value of 600 mm for
concrete to be classified as SCC (EFNARC, 2002). Although the slump flow test has
been proposed as a method to assess the segregation resistance of SCC, it was not used
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 61
50 cm
Slump-flow
3 .2 .2 L -B o x T est
The L-Box test (Fig. 3.3) was first developed by Mitsui et al. (1994) for the design of
underwater concrete. It consists of an L-shaped box with a gate separating the vertical
and horizontal parts. The vertical part is filled with approximately 12 liters of concrete
and the gate is opened to allow the concrete to flow between three 12 mm- diameter steel
bars spaced at 50 mm center-to-center. The distance between steel bars varies according
to the type o f construction and the maximum size of aggregate used. The ratio of concrete
height in the horizontal section to that in the vertical one, Ayft/ is an indication of the
lt2/hi > 0.8 for a concrete mixture to meet the passing ability of SCC. The time for the
concrete to reach 20 cm {T2 0 ) and 40 cm (T40 ) from the inside face of the gate are other
parameters that can be measured using this test and are considered as an indication for the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 62
SCC filling ability. However, no suitable values for such parameters have been generally
agreed upon.
3 or 4 <b12
3 .2 .3 V -F u n n e l test
The V-funnel was also developed in Japan by Ozawa et al. (1995) and is used to
determine the filling ability of SCC. It simply consists of a V-shaped funnel (Fig. 3.4).
The funnel is filled with SCC and the flow-time (7>) needed for the concrete to discharge
from an opening at the bottom of the funnel is taken as an indication of the SCC filling
ability. A value o f Tv < 12 sec is considered appropriate for SCC design (EFNARC,
2002). The test has also been proposed in Japan to assess the segregation resistance of
SCC mixtures by measuring the difference between the initial flow-time {Tvo) and that
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 63
51,5
LD
3.2.4 GTM S c r e e n S t a b il it y T e s t
The GTM screen stability test (Fig. 3.5) was developed by a French contractor to
calculating a segregation ratio (Sr), which is the percentage of mortar passing through a 5
observations suggest that 5% < Sr < 15% is acceptable for SCC design. Although the test
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 64
»
K
3 .2 .5 P r o p o s e d S e g r e g a t io n A s s e s s m e n t M e t h o d
The segregation assessment method proposed herein is based on studying the profile of
the coarse aggregate content (particles larger than 9.5 mm) along the height of a SCC
sample. The apparatus simply consists of a PVC tube (300 mm in height and 150 mm in
diameter) and a modified version of the penetration apparatus proposed by Bui et al.,
(2002). The tube is divided into three (150 x 100 mm) equal parts using leak-free joints
that are hinged to a vertical steel rod for easy sliding (Fig. 3.6-a). The modified version of
the penetration apparatus consists of four penetration heads (instead o f one) mounted on a
steel frame. Each penetration head is about 25 g in mass and 20-mm in diameter with a
semi-spherical end (Fig. 3.6-b). First, the average depth o f the penetration heads is
measured by allowing the heads to penetrate under their self-weight into concrete just
after the cylinder is filled. The three parts of the cylinder are then separated after a rest
period o f about 30 minutes and concrete in each part is washed out over a 9.5 mm sieve.
Coarse aggregates with particle size larger than 9.5 mm in each part of the cylinder are
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 65
then retrieved and their mass is determined. The segregation index (SI) is taken as the
coefficient o f variation (COV) of the coarse aggregate content in all three parts and is
M .- M avg
xlOO (3.1)
/=! M avg
I 3
Where M avg = and M, = mass of coarse aggregate particles larger than 9.5 mm
3 i=i
in each part o f the cylinder. Results of SI are then correlated to the corresponding average
penetration depth (Pd) o f the penetration heads, which is a more rapid field-oriented test.
(a ) (b)
Fig. 3.6 - Proposed test set-up: (a) Three-compartment hinged cylindrical mould
and (b) modified penetration depth apparatus.
A total o f 123 flowable concrete mixtures were prepared and tested in this study. All
and to satisfy generally recognized requirements of SCC (flowing ability and self-
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 66
compactibility). The GTM screen stability test and the segregation method proposed
herein were used in conjunction with the slump-flow and T50, L-box test, V-funnel test,
and compressive strength test to evaluate the properties of all prepared mixtures.
3.3 .1 M a t e r ia l s
including ground granulated blast furnace slag (GGBFS), class C fly-ash (FA-C), and
silica fume (SF) was used in all concrete mixtures investigated in this study. Table 3.2
lists the chemical composition and physical characteristics of the cement and cement
replacement materials. A continuously graded siliceous crushed stone with minimum and
maximum particle size of 5 and 19 mm respectively was used as coarse aggregate, and
well-graded silica sand was used as fine aggregate. The relative specific gravity and
water absorption o f the coarse aggregate (CA) were 2.68 and 0.7 %, respectively, whereas
the fine aggregate (FA) had a relative specific gravity o f 2.65, water absorption of 2.3 %,
and a fineness modulus of 2.67. Figure 3.7 shows the particle size distribution of coarse
and fine aggregates. A polycarboxylate-based HRWR with specific gravity of 1.08 was
(VMA).
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 67
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 68
100
Fine aggregate
90 C oarse aggregate
80
o> 60
50
Q.
40
30
20
10
3.3.2 M ix t u r e P r o p o r t io n s and M ix in g T e c h n iq u e
Concrete mixtures were prepared in accordance with ASTM C 192 (2000) “Standard
Practice for Making and Curing Concrete test Specimens in the Laboratory” using a 150
litre open-pan mixer. The mixing sequence consisted of mixing coarse aggregates and
silica fume for 30 sec, adding sand, and then homogenizing both fine and coarse
aggregates for another 30 sec. Cement and other cementitious materials were
subsequently added with the mixing water, and finally, the HRWR and VMA, which were
diluted with part o f the mixing water, were added and mixing resumed for another 3 min.
During testing, the mixer was covered with a plastic sheet and between two subsequent
tests, mixing was resumed for 1 min. Mixtures were proportioned to satisfy common
requirements o f SCC and cover a wide range of SCC mixture designs. The total
cementitious materials content varied between 350 and 600 kg/m3 of concrete, and the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 69
water to cementitious materials ratio (w/cm) varied between 0.35 and 0.60. Proportions of
other mixture ingredients were also varied to investigate their effect on the segregation
potential o f SCC mixtures. The minimum, maximum average and standard deviation
(STDV) values o f all mixture ingredients used are listed in Table 3.3.
Table 3.3 - Range and average values of mixture ingredients and test results
3.3.3 T e s t in g P r o c e d u r e s
The objective o f this experimental study is to develop a simple and reliable test method to
evaluate the ability o f SCC mixtures to resist segregation and to investigate the effect of
basic mixture ingredients on segregation and other properties o f SCC. The focus is
mainly on examining the effect of the cementitious materials content, w/cm ratio, coarse
aggregate to total aggregate ratio (CA/TA), and dosages o f HRWR and VMA. To
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 70
investigate the effect o f each ingredient on the segregation of SCC, a set of mixtures was
prepared by changing the volume of the particular ingredient and the proportions of the
other ingredients accordingly. For example, in the case of studying the effect of cm
content, the w/cm ratio, and the dosages of HRWR and VMA as percentage of cm content
were kept constant, whereas the volume of total aggregates was decreased or increased to
accommodate the changes in cm and water volumes. However, the ratio of CA/FA was
maintained constant. On the other hand, in investigating the effect of the CA/TA ratio,
only the volumes o f CA and FA were changed, whereas the content and ratios of other
ingredients were kept constant. Proportions and test results of mixtures used to study the
effect of mixture ingredients on the properties of SCC are included in Tables B.l and B.2,
Appendix B.
The L-box and V-funnel tests were only used in this study to ensure that mixtures
conform to common requirements of SCC. The 1i2/hi values of the L-Box and the flow
time of the V-funnel, TVo varied from 0.78 to 0.97 and from 5 to 11 sec, respectively for
all mixtures tested. The slump flow, slump-flow (7}0), and GTM tests in addition to
compressive strength at 1, 7, 28, and 91 days were performed on all SCC mixtures. The
minimum, maximum, and average values of test results shown in Table B.2 (Appendix B)
of coarse aggregate particles from the mortar compound. To capture the true nature of
First, SI-STATIC in which segregation occurs under normal placement conditions and
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 71
during concrete setting, mainly due to settlement of large aggregate particles under
gravity because o f the difference in relative density of materials and the low viscosity of
static effects. Using the segregation method proposed in this study, two values of Pd were
determined using the same test procedure explained earlier, except for the way the
concrete sample was prepared. For SI-STATIC, the test cylinder was simply filled and left
undisturbed for 30 min before the aggregate content was analyzed, whereas for SI-
DYNAMIC the concrete sample was prepared using the V-funnel apparatus before it was
left undisturbed for 30 minutes. Using the V-funnel as a preconditioning method for
determining the dynamic segregation index of SCC mixtures was to account for dynamic
effects such as the discharge of SCC through a chute of a ready-mix concrete truck
(simulated by the 2:1 slope o f the V-funnel sides) and concrete subjected to free-fall from
a height during placement (simulated by placing the test cylinder at a distance d below
An important aspect in testing the resistance of fresh SCC to segregation is the quality
control in sampling and handling the tested SCC sample. Often during the sampling
process and before the SCC sample is left to rest undisturbed, it would have already
original distribution o f aggregate particles. The results thus obtained could be misleading.
To overcome this difficulty, all concrete samples were filled in a similar manner directly
from the mixer, and all coarse aggregates used in the original batch were retrieved,
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 12
washed, and surface dried for re-use in subsequent tests so that variation in coarse
aggregate properties and sample handling are no longer hidden variables. The saturated
surface dry mass o f the coarse aggregates did not change significantly for the 123 SCC
mixtures, and no coarse aggregate was added to that of the original batch during the
entire test program (Note: test specimens for compressive strength were made from a
A few other segregation test methods that emerged recently, such as the settlement
column test (Rooney and Bartos 2001) and the column segregation test (Assaad et al.
2004), are also based on measuring the profile of coarse aggregate content along the
height o f a concrete sample. Since these methods require relatively large size SCC
samples, it was decided in the preliminary stage of this research to investigate the effect
of the sample size on segregation results. Thus, three SCC mixtures: Mi (highly
the effect o f the specimen’s height on the vertical profile of coarse aggregate distribution.
Their mixture proportions are shown in Table 3.4-(a). The segregation resistance of the
above three SCC mixtures was investigated using three sample-tubes of different height
(300, 450, and 600 mm) but having similar 150 mm diameter cross-section. Three
specimens (S 3 0 0 , S450, and S 6 0 0 ) from each of the above three SCC mixtures were prepared
and tested. A constant free fall height was achieved by placing the bottom of each tube at
the same distance o f 750 mm from the bottom of the V-funnel. All specimens were cast
by pouring the concrete into the V-funnel on its inclined sides and were left undisturbed
for a period o f 30 min. The vertical profile of coarse aggregate content was determined
for each test tube and the segregation indexes were determined using equation (3.1). The
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 73
segregation results thus obtained are included in Table 3.4-(b). Each data point in Table
3.4-(b) is the average value obtained on three similar samples prepared from the same
batch.
Results show that variations o f coarse aggregate content along the height of all samples
prepared from the same mixture were generally similar. Thus, it was decided to use a
tube of 300 mm in height since it requires less material, is easier to handle, and makes the
test more practical and less time consuming. In addition, d (the distance between the top
of the tube and the bottom o f the V-funnel) was determined to be 300 mm. This value
was chosen because it proved to be the minimum free fall height that can allow capturing
the segregation tendency of various SCC mixtures, while it still permits a convenient
height to pour the concrete by the operator into the V-funnel. From visual inspection of
several SCC mixtures inside the mixer and their behaviour during slump-flow tests, and
by comparison o f segregation results obtained using the proposed test method and those
of two other existing test methods, namely the settlement column test and the GTM
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 74
screen stability test, it was concluded that a SCC mixture can be considered stable (no
Finally, the penetration depth, Pd and the static and dynamic SI values for all 123 SCC
mixtures prepared in this study were measured. A relationship between Pd and SI values
for both static and dynamic tests was established (Figs. 3.8 and 3.9) and a value of Pd at
examining Figs. 3.8 and 3.9, two threshold values of Pd can be determined, depending on
the placement conditions o f SCC on site. Thus, a SCC mixture can be considered to be
45
40
E 35
30
25
20
c 15
10 —■ Pd = 10 mm
0 5 10 15 20 25 30 35 40
SI (%)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 75
45
40 y = 0.0035X2 + 0.3821x
E 35 SI = 10% R2 = 0.8928
30
25
20
c 15
10
5 — — Pd = 5 mm
0 10 20 30 40 50 60 70
SI (%)
to attain due to dual and sometimes contradictory effects of some mixture ingredients.
Thus, good understanding of the effect of basic SCC mixture design parameters on
of SCC mixture design can be achieved. In this study, effects of the most influential
parameters, such as the cementitious materials content, w/cm ratio, coarse aggregate to
total aggregate ratio (CA/TA), and dosages of HRWR and VMA on the properties of SCC
are investigated.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 76
3 .4 .1 E ffect of C e m e n t i t io u s m a t e r ia l s Content
To investigate the effect o f the cementitious materials content on the properties of SCC
mixtures, a set o f eight SCC mixtures was prepared and tested (Mixtures CM1-CM8,
Table B.l). The w/cm ratio and dosages of HRWR and VMA (as percentage of the total cm
content) were kept constant for all mixtures, while the total content of cm varied between
350 and 550 kg/m3. The total aggregates content was adjusted to accommodate changes
in cm and water contents. However, a constant coarse/fine aggregate ratio was maintained
in all mixtures. Common practice of SCC mixture design often recommends increasing
the total content o f fine materials and decreasing the content of coarse aggregate to
reduce the inner-particle friction between large aggregate particles, thus increase the
flowability o f SCC mixtures and reducing the risk of blockage and/or segregation.
Results obtained from the slump-flow test (Fig. 3.10) confirm that increasing the cm
content increased the flowability of SCC mixtures. However, the rate of deformability of
such mixtures was slightly decreased, especially at high value o f cm content as shown in
Fig. 3.11.
The effect of cm content on the ability o f SCC mixtures to resist segregation and on their
conventional concrete, the compressive strength of SCC mixtures increased when the cm
content increased as shown in Fig. 3.13. However, Fig. 3.12 shows that increasing the cm
content (at constant w/cm = 0.45) slightly increased the segregation tendency of coarse
aggregates, which is unexpected. Such a behaviour was observed in results of both the
GTM and SI-DYNAMIC methods. For SCC mixtures subjected to normal placement
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3
650
600
550
Q.
E 500
450
400 4 -
350 375 400 425 450 475 500 525 550
cm (kg/m3)
o' 4
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 78
20.0
♦ SI-DYNAMIC □ GTM a SI-STATIC
18.0
16.0
Slump flow < 500 mm , Slump flow > 500 mm
14.0
12.0
10.0
<0
8.0
6.0
4.0
2.0
0.0
350 375 400 425 450 475 500 525 550
cm (kg/m3)
Fig. 3.12 -Effect o f cm content on segregation resistance of SCC mixtures (w/cm = 0.45).
60
49.3
50
41.8
44.3 43.8
£40 42.9
ra 39.3
2 20
OL 19.8
17.5
15.72
14.7
10
♦ 1-day • 28-days a 91 -days
350 370 390 410 430 450 470 490 510 530
cm (kg/m3)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 79
Another set o f five SCC mixtures (Table B .l) were prepared and tested to investigate the
lower w/cm. Mixtures in this set were designed to have a constant dosage of VMA = 0.01
%, a constant w/cm = 0.40, and a dosage of HRWR varying between 0.22 % and 1.13 %
to achieve a similar slump flow of 615 ±15 mm. It was observed that for constant dosage
of VMA and w/cm, the SCC mixture with the highest amount of cm required the least
dosage o f HRWR to achieve the required slump flow. The influence of cm content on the
segregation resistance o f SCC mixtures having relatively lower w/cm was evaluated using
the proposed test method along with the GTM screen stability test, and results are shown
in Fig. 3.14. Results o f SI-DYNAMIC show that increasing the cm content slightly
reduced segregation in SCC mixtures, while those obtained using the GTM screen
stability test indicate a negligible increase of segregation when cm content increased. The
GTM screen stability test is based on the percentage of mortar passing through a 5-mm
sieve of a concrete sample, thus SCC mixtures with higher mortar fraction are expected
to exhibit higher percentage o f mortar passing. For SCC mixtures subjected to normal
Fig. 3.14.
It is believed that this discrepancy in the effect of the cm content on segregation in Figs.
3.12 and 3.14 is due to the fact that increasing the cm content in mixtures of the first set
while maintaining a constant w/cm = 0.45 increased the amount of free water. The HRWR
deflocculates agglomerated cement particles and also releases entrapped water, hence
increasing the amount o f free water available for fluidification. This can offset the
beneficial effect o f increasing the volume of the mortar compound due to increasing the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 80
such mixtures. Conversely, the amount of free water in mixtures with w/cm = 0.40 is
relatively lower, thus reducing the ability of excess water to neutralize the beneficial
20.0 n
♦ SI-DYNAMIC aGTM a SI-STATIC
18.0
16.0
14.0
12.0
8.0
6.0
4.0
2.0
0.0
350 375 400 425 450 475 500 525 550 575 600
cm (kg/m3)
3 .4 .2 E ffect of W ater to C e m e n t it io u s M a t e r ia l s r a t io
To investigate the effect o f the w/cm ratio on properties of SCC mixtures, another set of
mixtures was prepared (Mixtures CM3 and W1-W6, Table B.l). Mixtures in this set had
constant dosages o f HRWR and VMA, a constant ratio of coarse/fine aggregates contents
and a w/cm ratio varying between 0.4 and 0.6. As expected, increasing the w/cm ratio
significantly increased the flowability o f SCC mixtures through reducing the viscosity of
the mortar compound, thus reducing the ability to maintain a uniform distribution of large
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 81
coarse aggregate particles. The effect of the w/cm ratio on the flowability and rate of
750
700
^ 650
X 600
550
500
450
400
35 40 45 50 55 60
w/cm (%)
3.5
2.5
0.5
35 40 45 50 55 60
w/cm (%)
Fig. 3.16 - Effect o f w/cm ratio on the rate of deformability of SCC mixtures.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 82
It can be found from Fig. 3.15 that increasing the w/cm ratio linearly increases the
flowability of SCC mixtures. Such relationship is not as obvious in Fig. 3.16 due to the
scattered data points. However, it can be observed that the effect of the w/cm ratio on the
The effect o f w/cm ratio on the segregation tendency of coarse aggregates in SCC
mixtures was also investigated using two segregation methods, namely the GTM screen
stability test and the proposed method. Test results are shown in Fig. 3.17, in which data
from both methods show a similar trend for the effect of the w/cm on SI. However, such
an effect was more pronounced in the case of SI-DYNAMIC for w/cm > 0.45. This is
likely due to the effect o f filling the concrete tubes using a free fall o f SCC from a V-
the concrete to more severe placement conditions. Figure 3.17 also shows that for the
particular dosages o f HRWR and VMA used, all test methods captured a significant
increase in segregation for w/cm > 0.45, whereas for w/cm < 0.45, the rate of increase in
The effect o f w/cm ratio on the compressive strength of ordinary concrete is well
documented. However, all mixtures prepared in this study contain a variety of mineral
and chemical admixtures and the combined effect of such admixtures on the compressive
strength of SCC mixtures is not clear. Changes in compressive strength at various ages
due to variation in w/cm ratio are shown in Fig. 3.18. The figure shows that, similar to
ordinary concrete, increasing the w/cm ratio decreased the compressive strength
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3
regardless o f age. However, the rate of such effect is more significant at early age and
50.0
♦ SI-DYNAMIC □ GTM A SI-STATIC
45.0
40.0
35.0
Slump flow > 550 mm ♦
30.0
Slum p flow < 550 mm
25.0
CO
20.0
15.0
10.0
5.0
0.0
35 40 45 50 55 60
w/cm (%)
70.0
♦ 1-day • 28-days a 91-days
60.0
nT
Q. 51.40
49.84 49.42
§ . 50.0
£4-» 42.90
o> 41.34
c 39.00
.37.95.
£ 40.0
»
0>>
w 30.0
(0
£
“ ■ 20.0 15.72
O 13.17
O 11.02
9.50
8.44
10.0
40 45 50 55 60
w/cm (%)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 84
To examine the effect of HRWR and VMA dosages on properties o f SCC mixtures to
resist segregation, two sets of mixtures were prepared. The first set was intended to
investigate the effect o f the HRWR dosage and it included 10 mixtures (mixtures CM3
and HRWR1-HRWR9, Table B.l), while the second set contained 12 mixtures (mixtures
HRWR8 and VMA1-VMA11, Table B .l) and was prepared to investigate the effect of
the VMA dosage. Mixtures in both sets had similar proportions of all ingredients other
than the HRWR and VMA. Those in the first set were developed based on the control
mixture CM3 by gradually increasing the dosage of HRWR from 0.2 to 0.6 %. Similarly,
mixtures in the second set were developed based on mixture HRWR8 by gradually
increasing the dosage o f VMA from 0.01 to 0.03 %. Proportions and test results of these
mixtures are also listed in Tables B.l and B.2. The above values of HRWR and VMA
were chosen from within the practical range and to maintain the slump flow values of all
The effect o f HRWR and VMA dosages on the flowability of SCC mixtures is shown in
Figs. 3.19 and 3.20, respectively. As anticipated, the figures show that these admixtures
exhibit a contradictory effect on the flowability of SCC. For the range of dosages shown,
both admixtures showed a linear relationship for such effect. However, the rate of their
influence is slightly different. Increasing the slump flow value from 550 mm to 750 mm
required an increase in the dosage of HRWR of about 3 times (Fig. 3.19) whereas, only an
increase o f about 2 times o f the dosage of VMA was required to decrease the slump flow
value from 750 mm to 550 mm (Fig. 3.20). The objective of using VMA in the
development o f SCC is not to modify the workability o f such concrete, rather to enhance
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 85
the viscosity o f the paste and reduce segregation. However, understanding the reverse
effect of VMA on one o f the most important properties of SCC is vital for better design
800
750
s' 700
S 650
600
550
500
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
HRWR (%)
800 -i
750
700
E 650
s
o 600
550
500
450
400
0.005 0.010 0.015 0.020 0.025
VEA (%)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 86
Similar behaviour was observed for the effect of HRWR and VMA on the ability of SCC
mixtures to deform. Such effect is shown in Figs 3.21 and 3.22 as a relationship between
dosages of HRWR and VMA and the time needed for the concrete to reach a spread of 50
cm (Tso), respectively. The figures show that the ability of SCC mixtures to deform is
more sensitive to VMA dosages than that of HRWR. Figure 3.21 shows that increasing the
dosage o f HRWR by about 3 times only decreased T50 from 3.5 to 1.0 sec whereas, an
increase o f about 2 times o f the VMA dosage increased T50 from 2 to 7 sec (Fig. 3.22).
4.5
4.0
3.5
3.0
« 2.5
S2.0
1.5
0.5
0.0 4 -
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
HRWR (%)
Fig. 3.21 - Effect o f HRWR dosage on the rate of deformability of SCC mixtures.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 87
8.0
7.0
6.0
5.0
3.0
2.0
1.0
0.0
0.005 0.010 0.015 0.020 0.025
VEA (%)
Fig. 3.22 - Effect o f VMA dosage on the rate of deformability of SCC mixtures.
The effect o f HRWR and VMA dosages on the ability of SCC mixtures to resist
segregation is shown in Figs. 3.23 and 3.24, respectively. It is shown in Fig. 3.23 that for
constant w/cm ratio and VMA content, the ability of SCC mixtures to resist segregation
linearly decreased with increasing HRWR dosage regardless of the test method used.
Similar to the effect o f the w/cm ratio, higher HRWR dosage tended to decrease the
stability o f SCC mixtures and this effect was more pronounced in the case of dynamic
segregation. A reverse effect is exhibited by increasing the VMA dosage. Figure 3.24
shows that for constant w/cm ratio and HRWR dosage, higher VMA dosage increased the
ability o f SCC mixtures to resist segregation as expected. However, such a trend is shown
to be nonlinear (except for the GTM test) with a threshold VMA dosage, beyond which
the effect o f VMA in decreasing segregation became significant for a constant dosage of
HRWR. Figures 3.23 and 3.24 also show that for SCC mixtures with low risk of
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 88
segregation, the difference between S I values obtained using different test methods
diminishes, showing that the test method proposed herein is reasonably sensitive to
distinguish between static and dynamic segregation based on the placement conditions of
SCC. In other words, measuring the SI o f an SCC mixture having a moderate viscosity
and low risk o f segregation using either the static or dynamic condition should yield
comparable values. It is important to note that the relationships shown in Figs. 3.23 and
3.24 reflect the effect of admixtures used in this study, and that other types of admixtures
60.0 -i
♦ SI-DYNAMIC □ GTM A SI-STATIC
50.0
40.0
'-'30.0
20.0
a , a- . A
10.0
For a constant w/cm ratio and cm content, changing the dosage of HRWR and VMA used
showed negligible effect on the compressive strength of SCC mixtures, especially at later
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3
50.0 i
♦ SI-DYNAMIC □ GTM A SI-STATIC
45.0
40.0
35.0
30.0
25.0
20.0
15.0
10.0
5.0
0.0
0.005 0.010 0.015 0.020 0.025
VMA (%)
60.0
A
JL— 53.65
~ 50.0 . . . A . . .
52.33
n 51.40 51.54 50.83
Q.
•— 45.70 45.5
44.46 44.74
£ 40.0 42^0
o>
c
£
w 30.0
o>
'</>
(0
2 20.0
a.
E
o 15.72
° 10.0
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 90
60.0
£ 40.0
o> 40.73
w 30.0
£ 20.0
Q.
15.72 15.75 16.07 15.36
14.02
10.0
3 .4 .4 E ffect of C o a r se to T o t a l A g g r e g a t e R a t io
General practice (with some exceptions) of SCC mixture design often recommends that
the ratio of coarse/total aggregate (CA/TA) content should be limited to 0.5. Such a limit
is imposed to reduce the inter-particle friction between coarse aggregate particles, thus
enhancing the ability o f SCC to flow. However, the effect o f the coarse aggregates
content on segregation of SCC needs to be quantified. For this purpose, a set of 7 SCC
mixtures with different CA/TA ratios were prepared and tested using the various
segregation methods used in this study. Mixtures in this set were based on the control
mixture (CM3) and generated by changing the coarse/total aggregate ratio from 0.4 to
0.6. The contents o f all other ingredients were kept constant and only the contents of
coarse and fine aggregates were adjusted. Mixtures proportions and test results of all
mixtures are included in Tables B.l and B.2, respectively. Changing the ratio of coarse
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 91
flowability, rate o f deformability and compressive strength of SCC mixtures. The effect
of such ratio on the ability o f SCC mixtures to resist segregation is shown in Fig. 3.27,
which also indicates a slight to negligible increase in 57 values obtained from all test
methods over the range o f aggregate ratio investigated. The figure also shows that the
risk of dynamic segregation in SCC mixtures decreased when increasing the CA/TA ratio
below a threshold value of about 0.45, and increased beyond that value. CA/TA ~ 0.45
mixture design (EFNARC, 2002). The GTM test exhibited a more uniform increase of SI
14.0 i
♦ SI-DYNAMIC □ GTM A SI-STATIC
0.0
0.40 0.45 0.50 0.55 0.60
C oarse aggregate /Total aggregate
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 92
A simple segregation test for SCC was developed in this study based on quantifying the
profile o f coarse aggregate content along the height of a cylindrical concrete sample. In
addition to segregation caused by the settlement of large aggregate particles under gravity
(static segregation), the test also considers other dynamic aspects that can cause SCC to
segregate, including concrete discharge from the chute of a ready-mix concrete truck and
concrete free-fall from a height (dynamic segregation). The proposed test is easy to
mixture based on its resistance to segregation. This study also investigated the effect of
basic mixture ingredients on the ability of SCC mixtures to resist segregation using
were based on establishing a correlation between penetration depth and the COV of
coarse aggregate distribution along the height of concrete samples (SI) under both
static and dynamic conditions. The ability of SCC mixtures to resist segregation was
increasing w/cm and/or dosage of HRWR. Such an effect is more pronounced for
w/cm > 0.45 and at high dosages of HRWR for a constant dosage of VMA. A reverse
relationship was observed for the effect of w/cm and HRWR on the flowability of
SCC mixtures.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 93
• As expected, VMA proved to be an effective tool for reducing the segregation of SCC
mixtures. For a certain w/cm and HRWR dosage combination, there seems to be a
Moreover, VMA decreases the ability of SCC mixtures to flow and significantly
• The coarse/total aggregate ratio showed a slight to negligible effect on all SCC
Increasing the cm content caused an increase in segregation for SCC mixtures with
high w/cm, while for SCC mixtures having relatively lower w/cm, increasing the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 94
3.6- REFERENCES:
Assaad J., Khayat K. H., and Daczko J., (2004), “Evaluation o f Static Stability o f Self-
Consolidating Concrete,” ACI Materials Journal, V. 101, No. 3, pp. 207-215.
ASTM C 143/C 143M, (2000), ‘‘'Standard Test Method fo r Slump o f Hydraulic Cement
Concrete,” American Society for Testing and Materials, Philadelphia, PA, V. 4.02,
pp. 88-90
ASTM C 192, (2000), “Standard Practice fo r Making and Curing Concrete Test
Specimens in the Laboratory,” American Society for Testing and Materials,
Philadelphia, PA, V. 4.02, pp. 117-123.
Bui V. K., Montgomery D., Hinczak I., and Turner K., (2002), “Rapid Testing Method
fo r Segregation Resistance o f Self-Compacting Concrete,” Cement and Concrete
Research, V. 32, No. 9, pp. 1489-1496.
Kosmatka S. H., Kerkhoff B., Panarese W. C., Macleod N. F., and Mcgrath R., (2002)
“Designe and Control o f Concrete Mixtures,” EB101, 7th edition, Cement Association
of Canada, Ottawa, Ontario, Canada, 368 p.
Mitsui K., Yonezawa T., Kinoshita M., Shimono T., (1994), “Application o f New
Superplasticizer to Ultra Strength Concrete,” Fourth CANMET/ACI International
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3 95
Ozawa K., Sakata N., and Okamura H., (1995), “Evaluation o f Self-Compactibility o f
Fresh Concrete Using the V-funnel Test,” Concrete Library o f JSCE, 25, pp. 59-75.
Rooney M. J., and Bartos P. J. M., (2001), “Development o f the Settlement Column
Segregation Test fo r Fresh Self-Compacting C o n c r e te Proceedings of the Second
International Symposium on Self-Compacting Concrete, Tokyo, Corns Engineering
Corporation, pp. 109-116.
Zhu W., Gibbs J. C., and Bartos P. J. M., (2001), “Uniformity o f in Situ Properties o f
Self-Compacting Concrete in Full-Scale Structural E le m e n ts Cement and Concrete
Composites, V. 23, No. 1, pp. 57-64.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 4 96
CHAPTER 4
4.1- INTRODUCTION
The highly flowable nature and/or placement conditions of flowable special concrete
mixture, such as self-consolidating (SCC) and underwater concrete (UWC) can hinder
essential for better mixture design and structural performance. The effects of individual
ingredients on the behaviour of flowable special concrete mixtures are well known and
defined. However, some o f these ingredients have dual and sometimes contradictory
effects on the mixture performance, hence requiring the trial of several batches before
international standards do not include specifications for SCC and only a few of them
intended for casting under water. For example, the standards of the Japan Society of Civil
Engineers (JSCE, 1992) have recommended that the water-cement ratio (w/c) of concrete
mixtures designed for casting underwater reinforced concrete structure be limited to less
* A version o f this chapter was published in the ACI Materials Journal, V. 100, No. 2, 2003.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 4 97
than 0.50 and 0.55 in seawater and fresh water, respectively. They also require that
specimens cast under water should develop compressive strengths greater than 80% of
those cast and cured above water. However, these standards fall short in defining the
performance.
Because o f the contradictory effects of some of the mixture parameters and the absence
properties, the effect o f mixture proportioning on the rheological behaviour and in-place
concrete properties are often described using regression analysis tools and statistical
models (Khayat et al. 1996). These methods are generally based on an assumed or
observed behaviour and do not have the capability to account for the effect of all
parameters involved in the development of such concrete. This chapter explores the
concrete mixtures.
The growing need to simplify the relatively complex and costly operation of: 1) repairing
marine piles and hydraulic canals, or 2) constructing new underwater structures such as
building a base for a cofferdam or bridge foundations has lead to a growing interest in
developing a special type o f high performance concrete for underwater repair and
construction.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 4 98
Concrete intended for underwater repair and placement should have the ability to flow
readily into place to fill complex and narrow formwork, consolidate itself with minimal
and water dilution to develop adequate in-place mechanical properties. The successful
development o f such concrete must ensure a good balance between deformability and
stability, hence requiring a better understanding of the effect o f the mixture variables on
durability. Previous research concluded that sound and adequate in-place properties of
perforated basket filled with 2.2 ± 0.2 kg of fresh UWC sample in a 1.7 m column of
standing water. The basket is dropped and slowly retrieved after it remains 15 sec at the
bottom. Washout is the mass of concrete diluted in water after three drops (mass loss) as
a percentage o f the initial mass of the sample (CRD C61 1989). While some researchers
VMA) can significantly enhance concrete resistance to washout and segregation (Kawai,
1987, and Ghio et al. 1994), others suggested that the composition of cementitious
materials, water-cementitious materials ratio (w/cm), and the dosages of A WA and high-
range water reducing admixtures (HRWRA) directly affect, with different magnitudes, the
stability of highly flowable underwater concrete (Khayat et al. 1997, and Sonebi and
Khayat 2001).
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 4 99
As stated in chapter 2, the basic methodology for developing an effective ANN model is
to train a neural network to learn the relationship between a set of inputs and
corresponding outputs. In other words, the network should be trained using reliable
experimental data, that are comprehensive enough to capture all essential aspects that
might influence the input/output relationship, and large enough to effectively cover the
practical range o f all input parameters. A well-trained neural network should have the
capability of accurately predicting an output when it is presented with a new set of inputs
from within the practical range of the training data. ANN does not assume a relationship
between a set o f input parameters (mixture ingredients in this case) and the modeled
experimental data provided to build and train the network model. Accordingly, it is
believed that ANNs will have the ability of discovering unknown and hidden behaviour
and therefore, they should offer a superior capability in modeling concrete properties. Six
ANN-based models were developed in this chapter to predict properties of both SCC and
below.
4 .3 .1 S e l e c t io n and P r e -P r o c e s s in g of D a tabase
The selection o f the database to train a neural network is of paramount importance, and
for a neural network to actually capture the relationship between the parameters of
flowable concrete mixture and its engineering properties, it must be trained on large and
comprehensive sets o f reliable experimental data that contain the influential factors on
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 4 100
flowable concrete properties. Basic parameters that affect the behaviour of flowable
concrete properties such as cement, water, silica fume, fly ash, slag, fine and coarse
aggregates contents, and dosages of high-range water reducer and viscosity modifying /
anti-washout admixtures (VMA/AWA) were used for the development of the ANN
models. The models were trained to predict properties of flowable concrete such as
The segregation database generated in this study (see chapter 3) was used in the
SCC mixtures. Sample data is included in Appendix B. On the other hand, an extensive
literature review has identified a great deal of experimental data on flowable underwater
concrete cast in air and underwater. However, only data having mixture ingredients with
similar physical and chemical properties were considered in the development of the
network models for UWC. Moreover, a wide spanning discontinuity of some of the
mixture variables can destabilize the training process and reduce the prediction capability
of the network. This has prompted to disregard isolated data and reduce the range of
variables. Figures 4.1 and 4.2 show an example of the screening process based on HRWR
dosages in the data sets used to model the washout property (mass loss of flowable
concrete cast underwater). The final database consists of 175 flowable underwater
concrete mixtures. Table 4.1 includes the range, mean value, and standard deviation of all
input and output variables, while Table C.l (Appendix C) includes proportions and test
results for sample mixtures in the final database. The distribution of the selected database
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 4 101
between training and testing along with their corresponding original source are listed in
0.90
0.80
0.70
0.60
♦
>_-------
0.50 ♦ ♦
♦ ♦♦
0.40 »» ___
♦ ♦♦♦
0.30
♦ ♦ ♦♦«►♦
0.20
♦♦ ■r «** ♦ ♦
0.10
0.00
20 40 60 80 100 120 140 160 180
Mixtures
Fig. 4.1 - Values of HRWR in original mixtures collected to model the washout property
(before data screening).
0.60
0.50
♦♦
a. .
5 0.40 4
O'
x ♦ A ♦
1 0.30
♦ ^ m a * *
« 1 ♦ A* ♦♦ ♦
A *
o 0.20 ♦ *-* ♦♦ ♦♦♦
z
0.10
♦ a
0.00
0 20 40 60 80 100 120 140 160 180
Mixtures
Fig. 4.2 - Values o f HRWR in the selected mixtures used to model the washout property
(after data screening).
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 4 102
Table 4.1 - Range, average, and standard deviation (STDV) of measured input and
output variables (data shown is normalized)
Variables Training data Testing data
Range Average STDV Range Average STDV
Cement1 11.83 -25.27 18.70 4.06 12.55-25.23 19.06 3.96
Water1 5.50-11.26 9.35 1.68 6.94-11.25 9.40 1.76
Fly Ash1 0 .0 0 -5 .0 0 1.22 1.95 0 .0 0 -5 .0 0 0.96 1.80
Slag1 0.00 - 12.57 1.12 3.55 0.00 - 12.50 0.84 3.24
Silica fume1 0 .0 0 -5 .0 0 0.97 1.13 0.00 - 2.40 1.08 0.98
FA1 26.70-39.57 30.45 1.96 29.46-35.21 30.53 1.59
CA1 25.24-50.22 37.90 4.63 29.94 - 44.89 37.83 4.81
HRWR1 0.00 - 0.60 0.32 0.22 0.02 - 0.50 0.29 0.2
AWA1 0.00 - 0.06 0.02 0.01 0.00 - 0.04 0.02 0.01
Slump-Flow2 0.20 - 0.67 0.46 0.09 0.28 - 0.58 0.43 0.09
Washout3 0 .0 1 -0 .1 6 0.07 0.05 0.01 -0.13 0.06 0.03
/ c (28-d)4 0 .1 3 -0 .5 3 0.52 0.08 0.38 - 0.62 0.51 0.08
1Normalized as percentage of total weight of mixture;2 mm/1000;3 % weight loss; 4 MPa /100.
4.3.2 A N N -B a s e d M o d e l s for S e g r e g a t io n R e s is t a n c e of S C C M ix t u r e s
Two MLP neural network models were developed to predict the potential of segregation
in highly flowable concrete mixtures. The first model, ANN-STATIC (includes an input
layer with 9 units, one hidden layer with 5 units, and an output layer with one unit) was
trained to predict the segregation resistance of SCC mixtures under static conditions. The
second model ANN-DYNAMIC (includes an input layer with 9 units, one hidden layer
with 9 units, and an output layer with one unit) was trained to predict the segregation
and/or free fall). Figures 4.3 and 4.4 show the architecture of network models developed
for ANN-STATIC and ANN-DYNAMIC, respectively. Both models were trained and
tested using the database (123 SCC mixtures) generated in this study (see chapter 3).
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 4 103
Sample data is also included in Appendix B. The database includes the various SCC
mixture ingredients along with measured segregation index and covers a wide range of
mixture proportions usually used in the development of SCC. The chemical and physical
properties o f cementitious materials and chemical admixtures used along with the
maximum, average, and standard deviation of mixture ingredients and segregation results
are shown in Table 3.3. Full forward connection and variable learning and momentum
parameters were adopted for both models. Logarithm sigmoid function was also used as a
transfer function for all units in the hidden and output layers.
The database was compiled in a patterned format. Each pattern consists of an input vector
containing mixture ingredients (cement, water, fly-ash, Slag, silica fume, coarse and fine
aggregates, HRWR, and VMA) and an output vector containing the corresponding SI (SI-
STATIC or SI-DYNAMIC). All variables in the database were scaled between 0 and 1.
The database was then randomly divided into a training set containing 99 patterns and a
testing set o f 24 patterns. The training process of the ANN models consisted of teaching
the networks the embedded relationships between SCC mixture ingredients and its
corresponding SI. The generalization of the network was investigated using the testing
patterns (SCC mixtures unfamiliar to the network) whereby only the input vector
(mixture ingredients) is provided, and the network is asked to predict the corresponding
SI.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 4
Cement
Water
FA-C
SL
SF SI-STATIC
CA
FA
HRWR
VMA
Cement
Water
FA-C
SI-DYNAMIC
HRWR
VMA
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 4 105
The ability o f an ANN model to estimate the risk of segregation in SCC mixtures
depends on its capacity to generalize its predictions beyond mixtures used in the training
process to new mixtures so far unfamiliar to the network (see chapter 2). Good
generalization however, greatly depends on the degree o f success of the training process.
Thus the performance o f the ANN on both training and testing patterns should be
evaluated. Because no clear trend of either over or underestimating the property being
modeled was observed, the accuracy of ANN predictions was evaluated using the average
1 v »n |\Ymeas - Y prea,
(4.1)
Y„meas
Where Ymeas represents the measured value of segregation index from experimental data,
Ypred is the predicted value o f segregation index by ANN, and n = number of data points.
Both network models (ANN-STATIC and ANN-DYNAMIC) were trained and tested on
data from 99 and 24 SCC mixtures, respectively. Their performance in predicting the SI
of training mixtures and that o f new mixtures is shown in Figs. 4.5 and 4.6, respectively.
The figures show measured versus predicted SI values and clearly demonstrate that most
training and testing data points are located along the equity line with an average error of 4
ingredients and associated S I values of the training mixtures and were able to generalize
their predictions to new mixtures proportioned from within the range of mixture
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 4
25
o Training data
A Testing data
20
ris
■o
AAE (training) = 6 %
AAE (testing =8 %
0 5 10 15 20 25
Measured SI (%)
70
o Training data
A Testing data
60
50
<0 40
T3
£ 30
°o
20
10
AAE (training) = 4 %
AAE (testing) = 7 %
0 10 20 30 40 50 60 70
Measured SI (%)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 4 107
While both ANN models (ANN-STATIC and ANN-DYNAMIC) were successfully able
whether the models have captured the embedded effects of basic mixture ingredients (e.g.
w/cm ratio, and dosages o f HRWR and VMA) on the ability o f SCC to resist segregation.
To evaluate the response o f ANN models in predicting the effect of the above
ingredients, three sets o f mixtures (one set for each ingredient) were generated from a
single mixture randomly selected from the database. Proportions of all ingredients were
kept fairly constant except for the ingredient of interest, which was assigned different
values from within the range used in building the experimental database (Table C.3,
Appendix C). The newly generated mixtures were presented to the networks to predict
and HRWR and VMA dosages along with measured SI-DYNAMIC values of several SCC
The network’s response in Fig. 4.7 shows that for constant dosages of HRWR and VMA,
the risk o f segregation in SCC mixtures increased significantly for w/cm > 0.45, in
agreement with experimental data. Such a behaviour is also similar to that obtained from
the experimental investigation presented in Fig. 3.17 (chapter 3). Figure 4.8 shows that at
constant w/cm ratio and VMA dosage, the tendency of SCC mixtures to segregate
increased with increasing dosage of HRWR, especially for dosages between 0.35 % and
0.7 %. Beyond a HRWR dosage of 0.7 %, the segregation tendency of SCC mixtures
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 4 108
shown in Fig. 3.23, chapter 3. Similarly, Fig. 4.9 shows that at constant w/cm ratio and
HRWR dosage, the risk o f SCC mixtures to segregate decreased with increasing dosage of
VMA, and that such an effect was more pronounced for moderate to high values of VMA
dosage (see also Fig. 3.24, chapter 3). This suggests that the network has adequately
captured the influence o f the w/cm ratio, the HRWR and VMA dosages on the ability of
50
□ ANN prediction
45 • Experimental points
40
0 30
§ 25
1 20
<0
15
10
0
30 35 40 45 50 55 60
w/cm (%)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 4 109
□ ANN prediction
• Experimental points
U 40
1 30
Q
J
</>
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
HRWR (%)
45
□ ANN prediction
• Experimental points
40
35
?30
25
10
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 4 110
Several MLP network architectures were trained to predict properties of UWC mixtures.
hidden layers offered better predictions than a large number of other variations. The
network architecture adopted herein to model properties of UWC is shown in Fig. 4.10. It
consists o f an input layer, an output layer and two hidden layers. The input layer contains
9 units representing the influential parameters of UWC mixtures (same parameters used
in the case o f SCC). The output layer contains one unit representing the property of UWC
at hand. The first hidden layer includes 10 processing units while the second hidden layer
has only five processing units. Again, full connection between processing units in
adjacent layers was adopted and a sigmoid function logsig was assigned as the transfer
(activation) function for all processing units in the hidden and output layers.
Cement
Water
HRWR
VMA
Input layer First hidden layer Second hidden layer Output layer
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 4 111
The network architecture shown in Fig. 4.10 was trained to predict the slump-flow,
washout, and 28-day compressive strength of concrete mixtures designed for underwater
construction and repair. The network was first trained to predict the slump-flow value of
UWC mixtures using a total of 133 training patterns selected from seven different sources
(Table C.2, Appendix C). Each training pattern contains an input vector (mixture
value. The training process was terminated by choosing a minimum desired mean
squared error (MSE) o f 10"4. After 2400 epochs, the desired error was reached and
performance o f the network on the training data was evaluated. Again, because there is
usually no clear trend in neural network predictions for either over-or underestimating the
property at hand, the reliability of its predictions was evaluated using the average
absolute error (AAE) given by equation (4.1). Figure 4.11 shows the normalized
measured slump-flow values o f the training mixtures versus the predicted ones by the
network. Clearly all data points are located along the equity line with an AAE = 1.9 %
and the model response in predicting the slump flow of UWC mixtures in the training
After a successful training, the ability of the network to generalize its prediction to new
data was evaluated using new 14 UWC mixtures not used in the training process (Table
C.4 and C.5 show the normalized values of mixture ingredients and the normalized
measured and predicted slump-flow of mixtures used in the testing process). The network
was presented with only the input vector of each mixture and was asked to predict the
associated slump-flow. The predicted slump-flow values of mixtures in the testing data
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 4 112
are also plotted against the normalized measured ones in Fig. 4.11. Again, it is shown that
all testing points are located within the cluster formed by the training data points and
0 .8 i
♦ Training data
0.7 A Testing data
0.6
a. 0.5
0.4
0.3
0.2
A similar strategy was adopted in training the same network architecture (Fig. 4.10) to
predict the washout values of UWC mixtures using a total of 139 training patterns
selected from nine different sources as shown in Table C.2, Appendix C. After
completion o f training, the performance of the network model in predicting the washout
values o f the training mixtures is shown in Fig. 4.12. The difference between the network
predictions and the measured washout values of the training mixtures was slightly higher
than that of the slump-flow with an AAE = 13.5 %. The slightly lower performance o f the
network model in this case is believed to be due to the numerically very small values of
the washout; even small prediction errors would lead to high AAE. It could also be due to
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 4 113
the higher variability in the experimental data itself, which is more delicate to obtain in
the lab compared to slump-flow values. It is important to note herein that different
training strategies were attempted to lower the training error. However, they all lead to
0.18
♦ Training data
0.16 A Testing data
0.14
I
.c
0.12
(0
I 0.10
1o 0.08
u
2> 0.06
Q.
0.04
The performance o f the network model in predicting the washout values of new UWC
mixtures not used in the training process is also shown in Fig. 4.12. In addition to the
washout values o f training mixtures, the figure also show the predicted versus measured
washout values o f 16 new mixtures (Table C.4 and C.5, Appendix C). Again good
agreement between measured and predicted washout values can be observed for all
Finally, the same network architecture was also trained on 106 training patterns from six
different sources and 55 training patterns from two different sources to predict the 28-day
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 4 114
compressive strength of UWC mixture cast in the air and underwater, respectively. The
network performance in predicting the 28-day f c (air) for the training and testing
mixtures are plotted in Fig. 4.13. The figure shows satisfactory performance for both
cases with an AAE o f 1.2 % and 4.4 %, respectively. Figure 4.14 shows the performance
of the ANN model in predicting the 28-day f c (water) of UWC mixtures (training and
testing). Again good agreement can be observed between measured and predicted
compressive strength with an AAE of 1.0 % for training data and 2.1 % for testing data.
0.8
♦ Training data
A Testing data
0.7
-4.° 0.6
co
CM
■o 0.5
= 0.4
0.3
AAE (training) = 1.2 %
AAE (testing) = 4.4 %
0.2
0.2 0.3 0.4 0.5 0.6 0.7 0.8
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 4 115
0.6
♦ Training data
□ Testing data
0.5
oo
eg
"5 0.3
0.2
0.1
AAE (training) = 1.0 %
AAE (testing) = 2.1 %
0
0 0.1 0.2 0.3 0.4 0.5 0.6
Normalized m easured 28-d f c
Because the neural network model thus developed showed satisfactory performance and
demonstrated its ability to predict the properties of UWC mixtures designed within the
practical training range, it would be worth investigating whether the model has captured
the sensitivity o f mixture properties to the most influential individual ingredients on the
washout property (AWA and HRWR). Therefore, a set of six mixtures was created from a
randomly selected single mixture from the database by only changing the dosage of the
(AWA) and maintaining all other ingredients unchanged. Normalized ingredients of the
base mixture include 21.6 % cement, 11.07 % water, 1.87 % silica fume, 30 % sand, 35.2
% coarse aggregate, 0.005 %VMA, and 0.241 % HRWR. The six new mixtures were
presented to the ANN model to predict the associated washout and the model response is
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 4 116
shown in Fig. 4.15. It is indicated that by increasing the dosage of anti washout
0.2
0.18
0.16
0.14
» 0.12
0.1
•2 0.08
CL. 0.06
0.04
0.02
The effect o f HRWR dosage on the washout resistance o f UWC mixtures was
investigated in a similar fashion and results are illustrated in Fig. 4.16. The figure
presents the predicted washout values of eight UWC mixtures created by only changing
the dosage o f HR WR and maintaining all other ingredients constant. It is shown that at
low dosage, the HRWR has limited effect on washout. A slight decrease of washout was
observed as the HRWR dosage increased from 0.05 to 1 % likely due to the improved
homogeneity o f the mixture. As the HRWR dosage increased beyond this range, the
The effect o f the HRWR dosage on washout is, however, much less significant than that
o fth eAWA.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 4 117
0.09
0.08
0.07
(0
g 0.05
■o
o 0.04
T3
£ 0.03
Q.
0.02
0.01
Based on the investigation conducted in this chapter, the following conclusions can be
warranted:
1- This study demonstrated the feasibility of using feed-forward MLP neural networks
to predict properties and behaviour of flowable special concrete mixtures with good
accuracy. The models thus developed were not only able to predict mixtures
behaviour but also illustrated their ability to recognize the influence of basic mixture
2- The trained ANN models can predict properties of flowable concrete mixtures such as
SCC and UWC with good accuracy, so that the operator needs to carry out laboratory
testing only on a limited number of mixtures. This allows saving time and associated
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 4 118
network model can predict such a behaviour instantly and accurately. Therefore,
methods.
4- The models developed herein had a demonstrated ability to recognize and evaluate
could be used to develop concrete mixtures with tailor-made properties without need
for a large number o f trial batches as in the case of current practice; and
5- The present models did not incorporate durability characteristics o f flowable concrete
mixtures and did not account for the mixing, handling, and curing methods, which
may be easily incorporated provided that adequate experimental data for training the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 4 119
4.5-REFERENCES
Ghio, V.A., Monteiro, P.J.M., and Gjorv, O.E., (1994), “Effects o f Polysac-Charide
Gums on Fresh Concrete Properties, ” ACI Materials Journal, V. 91, No. 6, pp. 602-
606.
Khayat K.H., and Assaad J., (2003), “Relationship between Washout Resistance and
Rheological Properties o f High-Performance Underwater Concrete,” ACI Materials
Journals, V. 100, No. 3, pp. 185-193.
Khayat K.H., and Hester H.T., (1991), “Evaluation o f Concrete Mixtures fo r Underwater
Piles Repairs,” ASTM Cement, Concrete, and Aggregates, V. 13 No. 1, pp. 32-41.
Khayat K.H., El Gattioui M., and Nmai C., (1997), “Effects o f Silica Fume and Fly Ash
Replacement on Stability and Strength o f Fluid Concrete Containing Anti-Washout
Admixtures,” Superplasticizers and other Chemical Admixtures in Concrete, 5th
International CANMET/ACI Conference, SP-173, (Malhotra V. M., editor), American
Concrete Institute, Farmington Hills, Mich., pp 695-718.
Khayat K.H., Sonebi M., Yahia A., and Skaggs C.B., (1996), “Statistical Models to
Predict Flowability, Washout Resistance, and Strength o f Underwater Concrete,”
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 4 120
Sonebi M., and Khayat K.H., (1999), “Effect o f Water Velocity on Performance o f Self-
Consolidating Underwater-Cast Concrete,” ACI Materials Journal, V. 96, No. 5, pp
519-528.
Sonebi M., and Khayat K.H., (2001), “Effect o f Free-Fall Height in Water on
Performance o f Flowable Concrete,” ACI Materials Journal, V. 98, No. 1, pp 72-78.
Sonebi, M., and Khayat K.H., (2001), “Effects o f Mixture Composition on Relative
Strength o f Highly Flowable Underwater Concrete,” ACI Materials Journal, V. 98,
No. 3, pp 233-239.
Yamaguchi M., Tsushida T., and Toyoizumi H., (1986), “Development o f High-Viscosity
Underwater Concrete fo r Marine Structures,” Marine Concrete, International
Conference on Concrete in Marine Environment, Concrete Society, pp. 235-245.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 5 121
CHAPTER 5
5.1-INTRODUCTION
High-strength concrete (HSC) emerged in response to the increasing demand for high-
applications such as offshore structures (Jackobsen 1989), tall buildings (Walther, 1987),
highway bridges, and hazardous waste storage facilities. Current advances in concrete
technology allowed using very low water/cement ratio mixtures along with ultrafine
particular, the transition zone between aggregates and cement paste. HSC exhibits a much
stronger bond between cement paste and aggregate particles and less internal micro
Despite the fact that concrete properties generally improve as its compressive strength
* A version o f this chapter has been published in the Journal of Computers and Concrete, V. 2, No. 1, 2005.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapters 122
for adequate use in current design procedures and empirical equations. For instance, due
to the enhancement o f the aggregate-cement paste transition zone in HSC, micro cracks
tend to propagate through aggregates rather than around them, and the fracture surface is
therefore smoother than that in NSC as shown in Fig. 5.1. Moreover, the stress-strain
relationship o f HSC under uniaxial compression shows an increase in the linear segment
of its ascending portion and that the failure of HSC is more brittle than that of NSC as
shown in Fig. 5.2. This behaviour might have serious implications especially in the shear
eo
so
g
I 40
20
I 0002
Concrete «raln. mm/mm
Fig. 5.2 - Stress-strain relationship of NSC and HSC (MacGregor and Bartlett, 2000).
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 5 123
Shear transfer mechanisms that constitute the shear capacity of RC beams are well known
and established. The 1998 ACI-ASCE committee 445 identified these mechanisms as
transferred by friction across shear cracks or aggregate interlock; 3) the dowel action of
the longitudinal reinforcement; 4); the shear resisted by residual tensile stresses across
cracks; 5); arch action; and 6) the shear carried by the shear reinforcement. Shear
contribution due to arch action mechanism is associated with the shear span to depth ratio
(a/d) and is o f significant value for a/d <2.5. The focus of this study is on slender beams
with a/d > 2.5, and therefore the arch action mechanism was not discussed. Fig. 5.3
shows a schematic o f shear parameters and internal forces across an inclined crack in RC
slender beams.
Current shear design procedures and standards estimate the nominal shear capacity of
reinforced concrete (RC) slender beams V„ by simply adding the contribution of all shear
(5.1)
Vs represents the shear capacity carried by the shear reinforcement, and Vc (concrete
contribution) is equal to the total shear resisted by all other shear transfer mechanisms
considered in slender beams. Vs is well understood and can be quantified. However, the
remains unresolved despite the numerous studies carried out on this subject since the start
of last century.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 5 124
ci
t>T
Several shear calculation techniques for RC beams without shear reinforcement have
appeared in the literature since the original method proposed by Ritter (1899) and the
statistical, and analytical methods. Semi-empirical and statistical methods are normally
based on observations from available experimental data, whereas analytical methods use
more rational approaches yet they often require extensive calculations. Most current shear
estimating the shear capacity of large, lightly reinforced beams) is still debatable
(Angelakos et al. 2001). Moreover, such methods have typically been developed for NSC
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 5 125
with compressive strength o f less than 50 MPa, and their validity to estimate the shear
concrete slender beams without shear reinforcement (Fig. 5.3) constitutes a significant
part of the total shear resisted by concrete. The aggregate interlock mechanism usually
contributes about 35-50 % o f the total shear capacity of reinforced concrete beams
without shear reinforcement (Taylor, 1970). Thus, the shear behaviour o f such beams can
leading to smoother fracture surfaces and hence reducing the contribution of the
in NSC (Duthinh and Carino 1996, and Walraven 1995). This imposes constraints on
using current shear design methods to calculate the shear strength o f HSC beams.
This chapter presents a study to investigate the feasibility o f using artificial neural
networks (ANNs) to predict the shear capacity of reinforced NSC and HSC slender
beams (a/d > 2.5) without shear reinforcement, and to compare such predictions to results
obtained from five different existing methods namely, Zsutty’s equation, Response 2000
(based on the modified compression field theory), the shear provisions of ACI building
code (ACI-318-03), the simplified method of the Canadian Standards Association (CSA.
A23.3-94), and the method provided by the final draft of Eurocode-02. A sensitivity
analysis to evaluate the ability of the various methods to accurately capture the effects of
basic shear design parameters on the shear capacity of NSC and HSC beams is also
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 5 126
Several shear calculation methods have been developed during the last century and they
all concur that the shear strength of reinforced concrete slender beams without shear
longitudinal steel, pi, the beam’s effective depth, d, and the shear span to depth ratio, a/d.
However, these methods vary considerably in evaluating the effects of the above
trusses, exact solutions, plastic analysis, shear friction, or statistical approaches. They
went through series o f refinements since their initial development and are described in
detail elsewhere (Duthinh and Carino 1996). Only the shear methods that have been used
5.2.1 S t a t is t ic a l and A n a l y t ic a l M e t h o d s
Zsutty (1968 and 1971) formulated semi-empirical equations for predicting the shear
estimating the shear strength o f NSC beams without stirrups (MacGregor and Bartlett,
2000). Thus, this statistical method has become widely used in the literature. The
equation proposed by Zsutty to estimate the shear capacity of reinforced concrete slender
r • \ 1/ 3
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapters 127
where, d is the effective beam’s depth (mm), bw is the beam’s width (mm), a is the shear
Recently, Bentz and Collins (2000) developed a computer program called Response 2000
(R2K), which is a versatile tool for sectional analysis of reinforced concrete members. It
is an extension o f the modified compression field theory MCFT (Vecchio, and Collins,
1986), which explicitly incorporates rigid slipping along crack surfaces into compatibility
relations. The MCFT accounts for the load deformation responses o f structural elements,
and determine the angle of shear cracks by considering the deformation compatibility of
the vertical shear reinforcement, the longitudinal reinforcement, and the diagonally
stressed concrete. It is also the foundation of the CSA A23.3 general method, and is
considered as one o f the most refined analytical methods for shear analysis of reinforced
concrete members.
5.2.2 D e s ig n S p e c if ic a t io n s
Three commonly used shear design methods: The A C I318 (2003), the simplified method
of CSA. A23.3 (1994), and the final draft of Eurocode-2, (2002) were employed in this
study. The ACI 318-03 and the simplified method of CSA A23.3 consider the shear
capacity of reinforced concrete beams without web reinforcement, Vc, as the shear value
at which diagonal cracking is initiated. ACI 318 calculates Vc using one of two equations.
The first one, ACI 11-3, (approximate method) directly relates the contribution of
concrete to the shear resistance of a beam to the concrete compressive strength / tensile
strength as follows:
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 5 128
while the second equation, ACI 11-5, takes into account the effect of longitudinal
'v d "
K = 0 . 1 5 8 ^ + 17.2^ — M * 0.3J f e b j i (5.4)
v M jj
where M and V are the moment and shear force, respectively at a section subjected to
260
V. = ^ y J 7 ^ b wd <0.\A^[f^bwd d> 300m m (5.5b)
1000 + d
The final draft o f Eurocode-2, EC-2 was adopted in April 2002 and includes several
changes from its predecessor in terms of the shear design procedure. The expression
provided by EC-2 to calculate the shear resistance of concrete members not requiring
Vr d ,c = 0.18* (I0 0 p ,fcky 3 + 0.15 O-cp K d > vrdMa = 0.035 k ^ f cky bwd (5.6)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapters 129
Ngd is the axial force at the cross-section (N), Ac is the area of the concrete cross-section
5 .2 .3 A r t if ic ia l N e u r a l N e t w o r k s
As mentioned in chapter II, the main objective in building a neural network-based model
optimum set of connection strengths (weights) between its processing units. Using the
final set o f weights, the trained ANN can predict accurate values of outputs for a given
set of inputs within the range of the training data. The performance of ANNs depends, to
a great extent, on the learning material provided for their training. Therefore, an adequate
database must be generated to train a network model to predict the shear capacity of
reinforced concrete beams. The training database should be large enough, accurate,
comprehensive, and must contain the necessary information to assist in capturing the
Experimental shear strength results for 523 reinforced concrete beams without shear
reinforcement that have been tested in different laboratories were collected from the
literature, 358 o f which were made of NSC and 165 were made of HSC. Only slender
beams with a shear span to depth ratio (a/d > 2.5) that exhibited shear failures during
testing were considered. All beams were simply supported and subjected to either three-
point or four-point loading acting symmetrically with respect to the centreline of the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapters 130
beam. The database thus generated was further screened to eliminate outliers and beams
having one or more o f their design parameters isolated from the range of values of the
same parameter for the rest of the beams were disregarded. The final database was made
of 387 beams (263 NSC and 124 HSC). Table 5.1 includes the maximum, minimum, and
average values o f all shear design parameters and shear capacity at failure for beams in
the final database. Details o f beams used to evaluate the ANN model and shear
Table 5.1 - Statistical data for shear design parameters and shear capacity of beams used
in database
Two feed-forward back-propagation MLP networks having the same input and output
variables were developed. The first network, (2 hidden layers) was trained on
experimental results o f 329 beams (226 NSC and 103 HSC) and tested on 58 different
beams (37 NSC and 21 HSC). The second network (one hidden layer) was trained and
tested on results obtained from 124 HSC beams only (103 beams for training and 21
beams for testing). The performance of both networks in predicting the shear capacity of
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 5 131
reinforced HSC beams without web reinforcement and their sensitivity analysis (ability to
capture the effect of basic shear design parameters on shear strength) were found to be
similar. Therefore, only the first model is considered herein for its robustness and ability
to study the effect o f basic shear design parameters on the shear strength of beams made
The ANN model considered (Fig. 5.4) consists of an input layer containing 5 variables
representing the basic shear design parameters (d, bw, a/d, pi, and f c), an output layer with
one unit representing the ultimate shear value ( Vu), and two hidden layers having 10 and
5 units, respectively. Full forward connection (between units of one layer and those of the
subsequent layer) was adopted and variable learning rate and momentum were used to
avoid lengthy training and ensure global convergence. The transfer function used for all
units in this model was a logarithm sigmoid function (see chapter 2) with outputs values
varying between 0 and 1. Therefore, prior to the training process, all variables in the
generated database were scaled between 0 and 1, using equation (2.5) to speed up the
vu
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapters 132
The acceptance or rejection of the proposed ANN model is based not only on its ability to
accurately predict the shear failure load of beams used for training, but more importantly
on its ability to generalize its predictions to new beams, not familiar to the network, but
having input variables from within the range of input variables of the training data.
Therefore, the performance o f the ANN model thus developed was evaluated based on its
ability to predict the shear strength of a new set of beams (testing data) not used in the
training process and randomly selected from the collected beams in the database.
The ANN model along with equation 11-5 of ACI 318, the simplified method of CSA
A23.3, Response 2000, Zsutty’s equation, and shear provisions of Eurocode-2 were
employed to calculate the shear strength of all beams selected for testing (Table D .l,
Appendix D). The performance of each method in calculating the shear capacity of NSC
and HSC beams was evaluated using the average absolute error (AAE) calculated using
equation (5.7) and the ratio o f measured to calculated shear capacity ( V„/Vcai).
IV -V I
AAE = ^-ix lO O (5.7)
Vm
where Vm and Vcai are the measured and calculated/predicted shear capacity, respectively.
The average, standard deviation (STDV), and coefficient of variation (COV) of measured
/ calculated shear strength ratio and average absolute error (AAE) for all shear calculation
methods investigated herein are listed in Table 5.2. It is shown that the ACI (11-5) and
the CSA simplified method provided the least accurate shear capacity values for both
reinforced NSC and HSC beams without shear reinforcement with AAE values varying
between 22 % and 34 %. These are followed by Zsutty’s method and the Eurocode-2 with
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 5 133
AAE values o f 26 % and 21 % (in the case of HSC), and 14 % and 13 % (in the case of
NSC), respectively. R2K provided relatively more accurate results with lower AAE
values o f 19 % and 13 % for HSC and NSC, respectively. The ANN model outperformed
all of the above methods with an AAE of 10 % in the case of HSC beams and 9 % in the
case of NSC beams. Table 5.2 also shows that the ANN model had the lowest COV
values in predicting shear strength. It is of particular interest that the AAE and COV of
the shear calculation methods considered herein in calculating the shear capacity of
reinforced HSC beams are significantly higher than the corresponding ones for NSC
beams, which indicates that the applicability of such methods for HSC is questionable.
ACI 11-5 34.0 1.24 0.47 38.05 24.7 1.30 0.31 23.64
CSA (simp) 28.0 1.12 0.42 37.11 21.8 1.22 0.30 24.42
EC-2 21.0 1.07 0.35 32.54 13.2 1.08 0.20 18.74
R2K 19.0 1.20 0.35 28.82 13.1 1.15 0.19 16.47
Zsutty 25.5 1.02 0.31 30.25 14.2 1.04 0.20 19.44
ANN 10.0 1.03 0.17 16.77 9.0 0.99 0.12 12.08
The calculated shear capacities of NSC and HSC beams by the various shear calculation
methods are also plotted against the experimentally measured ones in Figs. 5.5 to 5.16. It
can be observed that the data points predicted by the ANN are located either on or
slightly over/under the equity line for both NSC and HSC (Figs. 5.10 and 5.16), whereas
those calculated by the other methods are scattered over a relatively wider range. For
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapters 134
slender reinforced concrete beams without web reinforcement having shear capacity of
up to 200 kN, Figs. 5.5, 5.6 5.11, and 5.12 illustrate that the ACI (11-5) and the CSA
simplified method, tended to underestimate the shear capacity of such beams for both
NSC and HSC, whereas the R2K, EC-2, and Zsutty’s method provided comparatively
more accurate results as shown in Figs. 5.7 - 5.9 and 5.13 - 5.15. For beams having shear
capacity larger than 200 kN (beams with either large d and/or having a large amount of
longitudinal steel), all five methods (other than ANN) demonstrated poor estimation
ability with data points scattered over a wide range away from the equity line. In most
cases, especially for HSC beams, the ACI 11-5, the CSA simplified method, and Zsutty’s
equation in particular, tended to overestimate shear capacity as shown in Figs. 5.11, 5.12,
and 5.15, respectively which can have serious implications in designing high-shear
450
o Training data
400 a Testing data
■5 200
■S 150
Training data : AAE = 25%
Vm/ V c = 1.3
Fig. 5.5 - Performance o f ACI-11 5 in predicting the shear capacity of NSC beams.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 5 135
450
o Training data
400 a Testing data
300
250
A O
200
Calculated
150
Training data : A A E = 22 %
100
Vm/V c = 1.22
50 Testing data : AAE = 22 %
Vm/Vc = 1.23
0
0 50 100 150 200 250 300 350 400 450
M e a su re d s h e a r c a p a c ity (kN)
450
o Training data
400 a Testing data
(kN)
350
shear capacity
300
250
200
Calculated
150
CA
Training data : A A E = 13 %
100 V m/V c = 1.08
Testing data : A A E = 14 %
Vm/Vc =1.11
Fig. 5.7 - Performance o f EC-2 in predicting the shear capacity o f NSC beams.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 5 136
450
o Training data
400 a Testing data
350
p 250
■= 200
oo
150
Training data : A A E = 13 %
« 100 V m/V c = 1.15
50 Testing data : A A E = 15 %
Vm/Vc = 1.16
Fig. 5.8 - Performance o f R2K in predicting the shear capacity o f NSC beams.
450
o Training data
400 a Testing data
■S 150
Training data : A AE = 14 %
■=100 V „ /V c =1.04
Testing data : A A E = 14 %
Vm/Vc = 1.02
Fig. 5.9 - Performance o f Zsutty’s equation in predicting the shear capacity of NSC
beams.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 5 137
450
o Training data
400 a Testing data
350
o 300
<■> 250
-£ 200
Training data : A A E = 8 %
■S 100 V m/V e = 0.99
Testing data : A A E = 11 %
Vm/Vc = 0.97
Fig. 5.10 - Performance o f ANN model in predicting the shear capacity of NSC beams.
550
o Training data
500 a Testing data
Z 450
300
£ 250
200
50 Testing data : A A E = 35 %
Vm/Vc = 1.18
0 50 100 150 200 250 300 350 400 450 500 550
Measured sh ear capacity (kN)
Fig. 5.11 - Performance of ACI-11.5 in predicting the shear capacity of HSC beams.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 5 138
550
o Training data
500
a Testing data
Z 450
S. 350
300
250
"S 200
150 Training data : A A E = 28 %
100 V m/V c = 1.12
Testing data : A AE - 27 %
Vm/Vc =1.11
0 50 100 150 200 250 300 350 400 450 500 550
M e a su re d s h e a r c a p a c ity (kN)
Fig. 5.12 - Performance o f CSA A23.3 in predicting the shear capacity of HSC beams.
550
o Training data
500
a Testing data
? 450
§.350
300
250
■g 200
150 Training data : A A E = 22 %
100 QDJ V „ /V c = 1.07
Testing data : A A E = 21 %
Vm/Vc =1.06
0 50 100 150 200 250 300 350 400 450 500 550
M e a su re d s h e a r c a p a c ity (kN)
Fig. 5.13 - Performance of EC-2 in predicting the shear capacity of HSC beams.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 5 139
o Training data
a Testing data
o
re
a.
re
o
re
re
.c
</>
TJ
re
■*-<
re
3 Training data : A A E = 19 %
O
re V m/ V c = 1.22
O
Testing data : A A E = 17 %
Hw/Kc = 1.15
50 100 150 200 250 300 350 400 450 500 550
Measured sh ear capacity (kN)
Fig. 5.14 - Performance of R2K in predicting the shear capacity of HSC beams.
550
o Training data
A Testing data
£•400
” 300
S. 250 0-40 / 0
Testing data : A A E = 26 %
Vm/Vc = 0.96
50 100 150 200 250 300 350 400 450 500 550
Measured sh ear capacity (kN)
Fig. 5.15 - Performance of Zsutty’s equation in predicting the shear capacity of HSC
beams.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 5 140
550
o Training data
500
a Testing data
Z 450
£>400
o
§.350
m
“ 300
f 250
200
0 50 100 150 200 250 300 350 400 450 500 550
Measured sh ear capacity (kN)
Fig. 5.16 - Performance o f ANN model in predicting the shear capacity of HSC beams.
All shear design provisions and other developed shear models account for the effect of
reinforced concrete beams without web reinforcement. While these methods can be
adequate for NSC slender beams ( /J < 50 MPa), results discussed above and some
literature (Ahmad et al., 1986) indicate that current design codes could be unconservative
in calculating the shear capacity of HSC beams (/J > 50 MPa). As stated earlier, about
35-50 % o f the ultimate shear capacity of slender concrete beams without shear
Since the shear fracture surface in HSC members is usually smoother than that in NSC
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapters 141
members (cracks propagate across aggregates), the ultimate shear strength of HSC beams
strength o f slender reinforced concrete beams without web reinforcement, a set of ten
beams (Table D.2, Appendix D) was generated from the properties of a randomly
selected single beam (shaded row in Table D.2) from the database. All design parameters
were kept constant except f'c, which was varied between 36 and 80 MPa. Figure 5.17
illustrates the effect o f concrete compressive strength on the ultimate shear strength of
concrete beams without shear reinforcement as simulated by the various shear equations
used in this study. It is shown from the ANN analysis that the shear strength of concrete
beams without web reinforcement increased with the increase of the concrete
decrease with increasing f c beyond that value, thus capturing the effect of loss in shear
friction provided by aggregate interlock due to smoother fracture surfaces in HSC. This
those in the generated set as shown in Fig. 5.17, and is also supported by a previous
experimental investigation in which Walraven (1995) indicated that the shear friction in
HSC beams is up to 35 % lower than that in NSC beams. Thorenfeldt and Drangsholt
(1990) also stated that the shear strength of concrete decreases as its compressive strength
Similar behaviour to that o f the ANN results is exhibited in results obtained using
Response 2000 in which the reduction in shear carried by aggregate interlock due to
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapters 142
design parameter representing the maximum size of aggregate, which is reduced as the
method), and Zsutty’s equation tended to seriously overestimate the effect of concrete
compressive strength on the shear strength of reinforced concrete beams in the high f c
range. On the other hand, the limitation proposed by the ACI 11-5 on concrete
compressive strength ( / c < 70 MPa) seems to be reasonable since the decrease in shear
strength o f reinforced concrete beams without shear reinforcement started in the region of
f'c~ 70 MPa. For over a 100 % increase in f c, the ACI 11-5, the CSA (simplified
-a-A C I
-a-C S A
2.0
Zsutty
— R2K
-K-EC-2
-©-ANN
1.7 • Exp. points
O) Reference beam
5 1-5
1.4
1.2
bw = 203 mm d = 356 mm
p , =0.0173 a/d = 3.0
reinforcement.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 5 143
All methods other than the ANN model and Response 2000 failed to accurately capture
the impact o f high compressive strength on the shear strength of concrete beams without
web reinforcement, showing unconservative predictions for HSC beams. This is also
confirmed by findings o f Ahmad et al. (1986) who argued that existing shear design
methods (including ACI 11-5) overestimate the effect of concrete compressive strength
and might be unconservative in calculating the shear strength of HSC slender beams (a/d
5 .3 .2 E ffect of T e n s il e S t e e l R a t io
This section explains the effect of the tensile steel ratio, pi on the ultimate shear capacity
o f reinforced concrete slender beams without web reinforcement, two sets of beams were
generated using a single beam from the database (Table D.3, Appendix D). Beams in the
first set have a compressive strength of 36.2 MPa and share the same design parameters
except the tensile steel ratio, which was varied between 1.13 % and 3.50 %. Beams in the
second set have the same design parameters of those in the first set but have a concrete
compressive strength of 70 MPa. Figure 5.18 shows the effect of pi on the shear strength
of NSC beams (set # 1). It includes the calculated shear strength o f the generated beams
along with the experimental shear strength of three additional beams having similar
properties. Contrary to some o f the current shear design methods in which the effect of
the tensile steel ratio on shear strength was ignored (e.g. the CSA simplified method and
ACI 11-3), such an effect is evident in experimental results and in results of the ACI (11-
5), R2K, Zsutty’s equation, and the ANN model. Figure 5.18 also shows that the ANN
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 5 144
-B-ACI
-ir-C S A
2.0 - o - Zsutty
~«"-R2K
-* -E C -2
• Exp
ANN
o>
Reference beam
1.4
bw = 203 mm d = 356 mm
f ' = 36 MPa a/d = 3.0
0.9 4 -
0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0
T ensile steel ratio (%)
Figure 5.19 shows that the impact of the tensile steel ratio on the shear capacity of HSC
beams is more pronounced than in the case of NSC beams. It is shown in Figs. 5.18 and
5.19 that increasing the tensile steel ratio from 1.13 % to 3.0 % increased the shear
NSC and HSC, respectively. The rate of increase however, is noticeably lower for both
NSC and HSC beams having tensile reinforcement ratio, pi > 2.5 %. As stated earlier, the
CSA simplified method does not consider the effect of longitudinal tensile reinforcement
in the calculation o f shear strength, whereas ACI 11-5 includes a slight effect as shown in
Figs. 5.18 and 5.19. Zsutty’s equation and Response 2000 however, recognize such an
effect and capture the variation in shear strength of NSC and HSC beams for different
tensile steel ratios in a comparable fashion to that of the ANN model. On the other hand,
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 5 145
the Eurocode-2 reasonably accounts for the effect of pi < 0.02, but the recommended
upper limit o f pt can lead to more conservative predictions of shear strength. The
relatively higher values o f shear strength predicted by Zsutty’s equation in Fig. 5.19 are
due to the fact that such a method, which was developed in the late 1960’s, overestimates
the shear strength o f beams made with concrete having high compressive strength values
2.0
Reference beam
e-1.7
1.4
b w = 203 mm d = 356 mm
f' = 70 MPa a/d = 3.0
5.3.3 E ffect of B e a m ’ s E f f e c t iv e D e p t h
A similar approach to that used to investigate the effect of f c and pi was carried out to
study the influence o f the effective beam’s depth, d on the ultimate shear strength of
slender reinforced concrete beams without web reinforcement. All beams used in this
investigation share the same design parameters except the effective depth, which was
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 5 146
varied between 219 and 466 mm (Table D.4, Appendix D). Figure 5.20 shows the
influence o f the beam’s effective depth, d on the shear strength of NSC beams as
calculated by the various methods considered in this study. It is apparent from this figure
that the ANN prediction best correlated with experimental data, showing a significant
effect of the beam’s depth on the ultimate shear strength. Similar behaviour was also
reported in several previous experimental investigations (Kim and Park 1994) and
fracture-mechanics based analysis, especially those carried out by Bazant and Kim,
(1984) and Bazant and Kazemi (1991). Experimental results and results obtained by the
ANN model and the Eurocode-2 in Fig. 5.20 illustrate that for a constant shear span to
depth ratio, the ultimate shear strength decreases as the beam’s effective depth increases.
However, such an influence becomes less significant for beams with large depth. The
CSA (simplified method) and R2K were somewhat able to capture this behaviour for
beams with depth larger than 300 mm. However, these methods seemed to ignore such an
effect for d < 300 mm. Conversely, the ACI 11-5 and Zsutty’s equation seemed to
disregard the effect o f d on the ultimate shear strength of beams with constant shear span
The influence o f the beam’s effective depth, d on the ultimate shear strength of HSC
beams is illustrated in Fig. 5.21. It can be observed that for beams with relatively small
depths (d < 300 mm), there is no clear difference in the effect of beam’s effective depth
on the shear strength o f reinforced concrete beams without shear reinforcement with
increasing compressive strength. However, for beams with d > 300 mm, it is shown that
the effect o f the beam’s depth on the shear strength was less significant at low
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapters 147
compressive strengths. This is confirmed by findings of Fujita et al. (2002) who argued
that the effect o f beam’s size on the shear capacity differs depending on the compressive
strength of concrete and that the shear strength of NSC is proportional to the effective
depth to the power o f -1/4, while that of HSC is proportional to the effective depth to the
power o f -1/2.
2.1
-e -A C I
—A—CSA
b w = 203 mm fc = 36 MPa
2.0 ■ o Zsutty Pi =0.0173 aid =3.0
R2K
—#—EC-2
— ANN
— Exp.
o Reference beam
£ 1-5 ' ” 7 ......... ............................ .. .............
1.4
1.2
0.9
150.0 200.0 250.0 300.0 350.0 400.0 450.0 500.0
Beam's depth (mm)
5 .3 .4 E ffect of S h e a r S pan to D e p t h R a t io
The shear span to depth ratio (a/d) has been determined to have little effect on the shear
strength o f slender NSC beams with a/d > 2.5 and therefore, can be neglected (Rebeiz et
al. (2001). However, some shear design methods such as the ACI 11.5 and Zsutty’s
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 5 148
2.1
—o- f c = 80 MPa b w = 203 mm p, = 0.0173
"" A ' f c ■ 70 MPa a/d = 3.0
1.9 - o - f c = 36 MPa
1.7
O)
£1-5
® 1.3
0.9
100.0 150.0 200.0 250.0 300.0 350.0 400.0 450.0 500.0
B e a m 's d e p th (m m )
An investigation was carried out to evaluate the influence that a/d exerts on the shear
strength of HSC beams. The response of all shear calculation methods considered in this
study in capturing the effect o f a/d on the shear strength of reinforced HSC slender beams
not having shear reinforcement are plotted in Fig. 5.22. All methods show a satisfactory
agreement in evaluating the effect of shear span to depth ratio for a/d > 3.5. However, for
lower values o f shear span to depth ratios, 2.5 < a/d < 3.5, these methods differ in
evaluating the magnitude of such an effect, which is minimal for ACI 11-5, does not exist
for the CSA simplified method and the Eurocode-2, and is somewhat significant for ANN
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 5 149
2.0
o>
Reference beam
1.4
1.2
bw = 200 mm d = 359 mm
p, = 0.0224 f c = 87 MPa
0.9
2.0 2.2 2.4 2.6 2.8 3.0 3.2 3.4 3.6 3.8
S h e a r s p a n to d e p th ra tio (a /d )
alternative method for predicting the ultimate shear capacity of reinforced NSC and HSC
beams without web reinforcement, and to compare its results to those of several existing
shear design and calculation methods. Furthermore, a sensitivity analysis was carried out
to evaluate the effects o f basic shear design parameters on the shear strength of concrete
members as simulated by various methods including the ANN model. The following
The ANN approach outperformed all other methods considered herein and reasonably
predicted the shear capacity o f concrete beams regardless of their compressive strength.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapters 150
The ANN approach adequately captured the influence of compressive strength on shear
capacity o f reinforced concrete beams without web reinforcement. It showed that shear
strength tends to decrease when concrete compressive strength increases above 70 MPa.
Conversely, current shear design and calculation methods, except the R2K, tended to
overestimate the shear capacity of HSC beams, and their applicability for HSC should be
re-evaluated.
This investigation showed that the amount of longitudinal tensile steel influences the
ultimate shear strength o f concrete beams without shear reinforcement in general, and
that this influence is more pronounced for HSC beams. This observation is supported by
results of the ANN model, Response 2000 and Zsutty’s equation regardless of pi, and by
Euro-2 for pi < 2 %. Conversely, the A C I11-5 underestimates such an effect, while the
For beams with a relatively small effective depth (d < 300 mm), the effective depth
exerts a similar effect on the ultimate shear capacity for both NSC and HSC beams
without web reinforcement. However, for high compressive strength, such an effect
Finally, the ANN analysis showed that similar to the case of NSC beams, the shear span
to depth ratio, a/d slightly affects the shear strength of HSC beams, and that such an
effect diminishes at higher values of a/d ratio. This behaviour is also observed using R2K
and Zsutty’s equation, whereas other methods either ignored this effect or underestimated
it.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 5 151
5.5- REFERENCES:
Adebar P., Collins M. P., (1996), “Shear Strength o f Members without Transverse
Reinforcement,” Canadian Journal of Civil Engineering, V. 23, pp. 30-41.
Ahmad S. H., Khaloo A. R., and Poveda A., (1986), “Shear Capacity o f Reinforced High-
Strength Concrete Beams,” ACI Journal, V. 83, No. 2, pp. 297-305.
Angelakos D., Bentz E. C., and Collins M. P., (2001), “Effect o f Concrete Strength and
Minimum Stirrups on Shear Strength o f Large Members,” ACI Structural Journal, V.
98, No. 3, pp. 290-300.
Bazant Z. P., and Kazemi M. T., (1991), “Size Effect on Diagonal Shear Failure o f
Beams without Stirrups,” ACI Structural Journal, V. 88, No. 3, pp. 268-276.
Bazant Z. P., and Kim J-K., (1984), “Size Effect in Shear Failure o f Longitudinally
Reinforced Concrete Beams,” ACI Journal, Proceedings V. 81, No.5, pp. 456-468.
Chana P. S., (1981), “Some Aspects o f Modeling the Behaviour o f Reinforced Concrete
under Shear Loading,” Tech. Rep. No. 543, Cement and Concrete Association,
Wexham Springs, 21 p.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 5 152
Collins M. P., and Kuchma D. (1999), “How Safe are Our Large, Lightly Reinforced
Concrete Beams, Slabs and Footings?” ACI Structural Journal, V. 96, No. 4, pp. 482-
490.
Duthinh D., and Carino N. J., (1996), “Shear Design o f High-Strength Concrete Beams:
A Review o f the State-of-the-Art”, Research Report NISTIR 5870, National Institute of
Standards and Technology, Building and Fire Research Laboratory, Gaithersburg,
MD, 198 p.
Elzanaty A. H., Nilson A. H., and Slate F. O., (1986), “Shear Capacity o f Reinforced
Concrete Beams Using High-Strength Concrete,” ACI Journal, V. 83, No. 2, pp. 290-
296.
Fujita M., Sato R., Matsumoto K., and Takiki Y., (2002), “Size Effect on Shear Capacity
o f Reinforced Concrete Beams Using HSC without Shear Reinforcement,” Proceedings
of the 6th International Symposium on Utilisation of High Strength/High Performance
Concrete,” Edited by Koenig, Dehn, and Faust, Leipzig, Germany, pp. 235-245.
Islam M. S., Pam H. J., and Kwan A. K. H., (1998), “Shear Capacity o f High- Strength
Concrete Beams with Their Point o f Inflection within the Shear Span”, Proceedings of
the Institution o f Civil Engineers, Structures and Buildings, V. 128, No. 1, pp. 91-99.
Jackobsen B., (1989), “High Strength Concrete in Offshore Structures,” Design Aspects
of HSC, CEB Bulletin d ’Information, No. 193.
Kani M. W., Huggins M. W., and Wittkopp R. R., (1979), “Kani on Shear in Reinforced
Concrete,” Department o f Civil Engineering, University o f Toronto Press, Toronto,
Canada, 225 p.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 5 153
Kim J. K., and Park Y. D., (1994), “Shear Strength o f Reinforced High Strength Concrete
Beams without Web Reinforcement,” Magazine of Concrete Research, V. 46, No. 166,
pp. 7-16.
MacGregor J. C., and Bartlett F. M., (2000), “Reinforced Concrete: Mechanics and
Design,” First Canadian edition, Prentice-Hall, Scarborough, Ont., Canada, 1042 p.
Morsch E., (1909), “Concrete Steel Construction,” 3rd ed., (1st ed. In 1902), translated to
English by E. P. Goodrich of Der Eisenbetonau, Engineering News Publishing Co,
New York, 368 p.
Mphonde A. G., and Frantz G. C., (1984), “Shear Tests o f High and Low-Strength
Concrete Beams without Stirrups,” ACI Journal, V. 81. No. 4, pp. 350-357.
Papadakis G., (1996), “Shear Failure o f Reinforced Concrete Beams without Stirrups,”
Ph.D. Thesis, Department o f Civil Engineering, Aristotle University of Thessaloniki,
Greece, (in Greek).
Pellegrino C., Bemardini A., and Modena C., (2002), “Shear Failure o f HSC Beams with
Variable Shear Span-to-Depth Ratio,” Proceeding of the 6th International Symposium
on Utilisation o f High Strength/High Performance Concrete,” Edited by Koenig,
Dehn, and Faust, Leipzig, Germany, pp. 473-485.
Rebeiz K. S., Fente J., and Frabizzio M. A., (2001), “Effect o f Variables on Shear
Strength o f Concrete Beams,” Journal of Materials in Civil Engineering, ASCE, V. 13,
No. 6, pp. 467-470.
Ritter W., (1899), “Die Bauweise Hennebique, Schweizerische Bauzeitung,” V. 33, No. 7,
pp. 59-61.
Taylor H. P. J., (1970), “Investigation o f the Forces Carried Across Cracks in Reinforced
Concrete Beams in Shear by Interlock o f Aggregate,” Technical Report 42.447,
Cement and Concrete Association, London, UK, 22 p.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 5 154
Taylor H. P. J., (1972), “Shear Strength o f Large Beams,” Journal of the Structural
Division, ASCE, V. 98, No. ST11, pp. 2473-2489.
Thorenfeldt E., and Drangsholt G., (1990), “Shear Capacity o f Reinforced High Strength
Concrete Beams,” ACI 2nd International Symposium on HSC, ACI SP 121.8,
American Concrete Institute, Farmington Hills, Michigan, pp. 129-154.
Vecchio F. J., and Collins M. P., (1986), “The Modified Compression Field Theory fo r
Reinforced Concrete Elements Subjected to Shear,” ACI Structural Journal, V. 83, No.
2, pp. 219-231.
Walraven J. C., (1978), “The Influence o f Depth on the Shear Strength o f Light Weight
Concrete Beams without Shear Reinforcement,” Stevin Lab. Rep. No. 5-78-4, Delft
University o f Technology, the Netherlands.
Walther R., (1987), “Potentiality o f Using High Strength Concrete in Structures,” First
International Symposium on Utilization of High Strength Concrete, Stavanger,
Norway, pp. 365-378.
Zsutty T. C., (1968), “Beam Shear Strength Prediction by Analysis o f Existing Data,”
ACI Journal, V. 65, No. 11, pp. 943-951.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6 155
CHAPTER 6
6.1- INTRODUCTION
Current shear design practice assumes that the shear capacity o f reinforced concrete (RC)
slender beams varies linearly with the amount of shear reinforcement, and that the
addition o f stirrups does not alter the basic effects of shear design parameters and the
similar beam not containing shear reinforcement), to that of stirrups, Vs. Concrete
chapter 5 and Vs is simply determined based on the parallel truss model with 45° constant
Vt =<hL =pj (6 . 1)
s
*A version o f this chapter has been submitted for review to the Canadian Journal o f Civil Engineering.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6 156
Where Av is the cross-sectional area of stirru p s,^ is the yield strength o f stirrups, d is the
beam’s effective depth, s is the spacing between stirrups, and p v is the ratio of shear
reinforcement.
Previous studies conducted by (Mphonde 1989 and Russo and Puleri 1997) reported that
the addition of stirrups significantly enhances the resistance of concrete to shear and
usually leads to a higher Vc than that determined from the equilibrium consideration of
the 45° truss model. Moreover, Chana (1987) suggested that the shear failure mechanism
without shear reinforcement, and therefore Vc and Vs mutually influence each other.
Simply adding these parameters to determine the shear capacity of RC beams with
stirrups can induce significant error and has not been theoretically validated.
MPa has been employed in various challenging applications and its use is growing
rapidly. The mechanical properties of HSC are different from those of normal-strength
concrete (NSC) as explained in chapter 5. Thus, empirical design equations, which are
evaluated for HSC. Despite the numerous studies investigating the shear behaviour of RC
slender beams, the mechanism of shear failure is not yet fully understood. The effect of
fully embedded in the shear design equations, and the influence of stirrups on other shear
resisting mechanisms is simply not accounted for. Therefore, the ability of current shear
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6 157
design methods to accurately predict the shear resistance of RC slender beams and
especially those made o f HSC is still debatable (Russo and Puleri 1997).
Chapter 5 demonstrated that shear design equations and prediction models, which were
developed for NSC slender beams are not suitable for calculating the shear capacity of
beams made with HSC. Moreover, previous comparative studies (Russo and Puleri 1997,
and Chana 1987) concluded that current shear design techniques are either very
conservative in calculating the shear capacity of slender concrete beams with stirrups by
just adding Vc and Vs, or simply not applicable for RC beams made of HSC. In this
predict the shear strength of RC slender beams, Its results are compared to those obtained
from five different shear strength calculation methods namely, the provisions of the
American Concrete Institute (ACI 318 2003) and the Canadian Standards Association
(CSA A23.3 1994), Zsutty’s equation (Zsutty 1971), Bazant and Kim’s equation (Bazant
and Kim 1984), and Mphonde’s equation (Mphonde 1989). A parametric study was also
carried out to evaluate the effect of stirrups on other shear resisting mechanisms of
slender beams having shear reinforcement, and the ability of shear prediction techniques
to quantitatively account for the effects of basic shear design parameters on the shear
strength of reinforced NSC and HSC slender beams with or without shear reinforcement.
calculate the shear capacity Vc, of RC beams without shear reinforcement, and simply add
the contribution o f stirrups Vs, which is calculated based on the parallel truss model with
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6 158
45° constant inclination diagonal shear cracks, to obtain the shear capacity of RC beams
with stirrups. These codes vary considerably in evaluating the effect of basic shear design
parameters on Vc. Moreover, it has been argued that stresses in stirrups are consistently
lower than those predicted by the 45° truss model (Duthinh and Carino 1996) and that
other shear resistance mechanisms interact with that of stirrups in different ways, leading
to variable contribution o f the truss mechanism and to some enhancement of the beam
action contribution (Russo and Puleri 1997), which leads to underestimating the shear
strength o f RC slender beams. The various shear prediction methods used in this study
Similar to all other specifications and codes, the shear design provision of the ACI
building code (ACI 318-03) and the simplified method o f the Canadian Standards
Association (A23.3) assume that the shear capacity of RC slender beams with shear
reinforcement is simply equal to the superposition of the contribution of concrete (Vc) and
that of stirrups (Vs). Vs is calculated using equation (6.1) and Vc is assumed to be equal to
the shear capacity o f a similar beam that has no shear reinforcement, and is calculated
The ACI method (ACI 11-5) applies only for concrete with compressive strength f'c < 70
MPa and limits the shear contribution of stirrups Vs, to 0.66-Jf^bwd , while the CSA
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6 159
Two semi-empirical equations developed by Zsutty (1971) and Mphonde (1989), and one
equation (derived based on fracture mechanics) developed by Bazant and Kim (1984)
were used in this study to predict the shear strength of reinforced NSC and HSC slender
beams. Zsutty’s equation assumes that the shear capacity of RC slender beams with web
Mphonde and Bazant-Kim were developed to account for the effect of stirrups on other
/ , \ 1/3
V=K+V=2.2
L pi A A f y v d
(N) Zsutty (6.2)
M +
a
yv
V = 0.83 f z 3J p i + 1.67 ^ P vf> M (N) Bazant-Kim (6.4)
where ^ and %are functions to account for the size effect and effectiveness of stirrups,
1
# = ■
(6.4a)
1+ -
25 d„
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6 160
6 .2 .3 A r t if ic ia l N e u r a l N e t w o r k s A p p r o a c h
model. However, the learning material provided for training remains the most important
factor that affects the network’s performance and generalization. To capture the
embedded relationship between the most influential parameters of RC slender beams and
their corresponding shear capacity as well as the influence of stirrups on other shear
sufficiently large shear database that include data on beams with and without shear
reinforcement.
The database developed to investigate the feasibility of ANN to predict the influence of
vertical shear reinforcement on shear capacity of reinforced concrete slender beams and
on the effect o f other shear design parameters consists of shear strength results for 656
reinforced NSC and HSC beams with or without shear reinforcement. The database
includes shear results for the 387 beams (without stirrups) used in chapter 5 and another
269 beams (with stirrups) collected from literature. Only simply supported beams with a
shear span to depth ratio (a/d > 2.4) that exhibited shear failures were considered. About
60% of the database represent shear results obtained on NSC beams (36% of which are
with shear reinforcement and 64% are without shear reinforcement). The remaining 40%
of the database represent results obtained on HSC beams (53% of which are shear
reinforced and 47% are without shear reinforcement). The database was compiled in a
patterned format and normalized between 0 and 1 as explained in chapter 2. Each pattern
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6 161
consists o f an input vector containing the geometrical and mechanical properties of the
beams, and an output vector containing the corresponding shear capacity of the beam.
Table 6.1 provides the range and average values of all parameters used in the database.
Table 6.1 - Statistical data for shear design parameters and shear capacity of beams used
in database
6 .2 .3 .2 A N N M o d e l
Several ANN architectures were trained and tested to develop a feed-forward back-
propagation MLP network that can accurately predict the shear strength of RC slender
beams. The network’s architecture that was adopted consists of an input layer containing
6 variables representing the commonly known shear design parameters (d, bw, a/d, p f yv,
pi, and f c), an output layer with one unit representing the shear capacity (V„), and a
hidden layer o f 10 processing units (Fig. 6.1). The ANN was trained using 542 data
patterns and tested on the remaining 114 patterns. The testing patterns were randomly
selected from the original database and were not used in the training process. Among the
testing patterns 58% were NSC beams (32% without shear reinforcement and 26% with
shear reinforcement) and 42% were HSC beams (19% without shear reinforcement, and
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6 162
23% with shear reinforcement). Table E.l in Appendix E includes beams characteristic
and shear strength results of all beams used as testing data. Again, Variable learning rate
and momentum were used to avoid lengthy training and ensure global convergence of the
network.
The performance of the ANN model thus developed and trained was evaluated using the
testing database described earlier. The network was presented with the input vectors of
the testing patterns and asked to predict the corresponding shear capacities. The predicted
shear capacities were subsequently compared to the experimentally measured values and
the performance o f the network was evaluated based on the average absolute error (AAE)
and a performance factor calculated using equations (6.5) and (6.6), respectively.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6 163
V
PF = — ( 6 .6)
cal
where Vm and Vcai are the measured and calculated/predicted shear capacity, respectively.
The shear capacities o f all beams in the testing database, which was used to evaluate the
performance o f the ANN model, were also calculated using the five shear strength
calculation procedures discussed earlier. The average, standard deviation (STDV), and
coefficient o f variation (COV) for the performance factor (PF) and the average absolute
error (AAE) o f all shear design methods investigated are listed in Table 6.2 for both NSC
Table 6.2 - Performance o f shear design methods considered in this study (beams with
shear reinforcement)
It is shown that the ACI (11-5) and the simplified method of the CSA are fairly
stirrups with an AAE varying between 20% and 27% and an average PF reaching as high
as 1.36 in the case o f HSC beams. Zsutty’s equation provided reasonably accurate
predictions and tended to underestimate the measured shear capacity by about 10% in the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6 164
case of NSC beams. However, its accuracy is clearly lower for HSC beams with an AAE
of 17% and average PF o f 1.13. Table 6.2 also shows that the ANN model and Bazant-
Kim’s equation provided the most accurate shear capacity predictions in the case of NSC
beams with an AAE o f 12% and average PF of 1. However, the accuracy of Bazant-
Kim’s equation in predicting the shear capacity of HSC beams is lower than that for NSC
beams, while that o f ANN was similar for both NSC and HSC beams.
Figures 6.2 to 6.7 illustrate the calculated/predicted shear capacities of reinforced NSC
and HSC beams with shear reinforcement versus the corresponding experimentally
measured ones. It can be observed that the data points for the ANN (Fig. 6.7) and Bazant-
Kim’s method (Fig. 6.6) are the closest to the equity line, whereas those for the other
methods (Figs. 6.2 - 6.5) are scattered over a relatively wider range, especially for beams
having shear capacity values larger than 200 kN (beams with either large effective depth,
large amount o f longitudinal and/or transverse reinforcement). Figures 6.2 and 6.3 also
show that in most cases, the ACI (11-5) and the simplified CSA methods underestimate
the shear capacity o f RC beams with shear reinforcement, and that their accuracy in
predicting the shear capacity of HSC beams is slightly lower than that for NSC beams.
Despite its relatively accurate predictions in the case of NSC beams without shear
the case o f RC beams containing shear reinforcement (Fig. 6.4). The conservative
predictions provided by the CSA simplified method, ACI (11-5), and Zsutty’s equation
are partially due to the fact that all three methods assume that no interactions occur
between stirrups and other shear transfer mechanisms and simply calculate the shear
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6 165
that of concrete.
900
OHSC ANSC
800
700
o 600
o 500
£<0 400
I 300
HSC: AAE = 23%
S> 200 Vm/V c =136
NSC: AAE = 27 %
100 Vm/Vc = 1JO
Fig. 6.2 - Performance o f ACI method in calculating the shear capacity of RC beams
with stirrups (only testing points are shown).
900
OHSC A NSC
Fig. 6.3 - Performance o f the simplified method of CSA method in calculating the shear
capacity o f RC beams with stirrups (only testing points are shown).
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6 166
900
OHSC A NSC
800
700
500
400
300
1000
O HSC A NSC
900
(kN)
800
700
sh ea r capacity
600
500
400
Predicted
300
HSC: AAE =21 %
200 Vm/V c = 1.03
NSC: AAE =25%
100
Vm/Vc = 0.9
0 100 200 300 400 500 600 700 800 900 1000
M e a s u re d s h e a r c a p a c ity (kN)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6 167
900
OHSC A NSC
800
700
o 600
o 500
f 400
£ 300 o .
900
OHSC A NSC
800
O 600
CO
o 500
j= 400
a> 300
HSC.- AA E = 12 %
S 200
Vm/Vc = 1-00
100 NSC: A AE =12 %
Vm/Vc = 1.00
Fig. 6.7 - Performance o f ANN model in predicting the shear capacity of RC beams with
stirrups (only testing points are shown).
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6 168
Shear provisions o f design codes and standards generally assume that adding shear
reinforcement to a RC slender beam will only enhance its shear strength by the shear
capacity o f stirrups calculated using equation (6.1). Such a practice ignores the influence
of stirrups on the contribution of other shear design parameters and assumes that the
presence o f stirrups does not interfere with other shear resisting mechanisms, which
presumes a linear relationship between the amount of shear reinforcement and shear
strength. A sensitivity analysis was conducted in this study to investigate the effect of
stirrups on the shear strength of RC slender beams using the various shear prediction
procedures discussed earlier. A set of beams was generated from a NSC beam randomly
selected from the database, in which all beams share the same geometrical and
mechanical properties, except for the amount of shear reinforcement. Properties of the
original beam and all generated beams in this set along with their calculated shear
The shear strength o f the generated beams (calculated using the various shear design
methods) along with the experimentally measured shear strength of beams tested by
Placas and Regan (1971) are also plotted versus the capacity o f stirrups in Fig. 6.8. It can
be observed that the ANN response was the closest to the experimentally measured data
points and that the effect o f stirrups on the shear strength of beams was greater at low
shear reinforcement ratio than that at relatively higher ratio. This behaviour was also
Scordelis (1963), and Haddadine et al. (1971), in which they found that stirrups were
more efficient at low to moderate shear reinforcement ratio than at higher reinforcement
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6 169
ratio. Figure. 6.8 also show that shear strengths predicted by the ANN are 65% to 75%
higher than those calculated by the ACI. These values are similar to findings of Bresler
and Scordelis, and Haddadine et al. who reported that the shear strength of RC beams
having moderate shear reinforcement are 75% to 80% higher than corresponding values
5.5
E — ANN
5.0 -®— Bazant-Kim
M phonde
4.5
Q - Zsutty
O - C SA simplified
A - ACI
1 . 3.5 • E xperim ental points
0 ) 3 .0
c
4)
= 2.5
CO
i! 2.0
a>
</5 1.5
1.0
0.5
0.0
0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4
Shear reinforcem ent p vfyv (MPa)
Similar behaviour was observed in the case of HSC beams (Fig. 6.9). Again the ANN
response was the closest to the experimental shear strength of beams having similar
geometrical and mechanical properties tested by Kong and Rangan (1998). Figure 6.9
also show that the Bazant-Kim’s equation overestimates the shear strength of HSC beams
in agreement with data o f Table 6.2. This appears to be due to an overestimation of the
effect of the concrete compressive strength on the shear capacity at high strength values.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6 170
a -A N N
■®— Bazant-Kim
■A— M phonde
- £> Z sutty
O - C SA simplified
b 2.5
1.0
d = 294 mm bw =250 mm a/d = 3.3
0.5
Pz= 4.4 f c = 75 MPa
0.0
0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6
Shear reinforcement p vfyv (MPa)
6 .3 .2 Influence of S t ir r u p s on E ffect of C o n c r e t e C o m p r e s s iv e S t r e n g t h
the shear strength o f RC slender beams. This is particularly true for NSC slender beams.
However, chapter 5 and several experimental studies (i.e. Duthinh and Carino 1996, and
Thorenfeldt and Drangsholt 1990) demonstrated that the shear strength of RC beams
without shear reinforcement slightly decreases with increasing f c beyond 70 MPa due to
loss in the shear resisting mechanism of aggregate interlock. Furthermore, most current
shear design techniques either do not acknowledge such a variation in the effect of
concrete compressive strength on the shear capacity of beams or simply do not account
for the influence o f adding shear reinforcement on other shear transfer mechanisms.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6 171
the shear strength of RC slender beams, the success of the ANN model in capturing the
relationship between the compressive strength of concrete and the shear strength of RC
beams without stirrups ought to be evaluated. The shear strengths of a set of RC beams
generated from a single beam randomly selected from the database were calculated using
all shear prediction methods adopted in this study, including the ANN method. All beams
share the same geometrical and mechanical properties, except for f c which varies from
The shear strength o f the generated beams along with the experimentally measured shear
strength o f 3 beams having similar properties and tested by Mphonde (1989) is plotted
versus the compressive strength of concrete in Fig. 6.10. The figure shows that while all
methods considered herein (except that of Mphonde) captured the trend of the effect of
their accuracy), the predictions provided by the ANN model are the closest to the
and Ramirez (1989) indicate that for a constant low amount of shear reinforcement, the
overall reserve shear strength after diagonal cracking diminishes with an increase in
concrete compressive strength. Only the ANN model response captured such a behaviour,
while all other methods failed to show a decrease in the rate of growth of shear strength
with higher f 'c . (Note that shear provisions of ACI does not capture this effect, but simply
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6 172
5.0
—• —ANN
• Bzant-Kim
4.5 —A—Mphonde
- • G - • Zsutty
-••A
2.0
shear capacity, the shear strengths of three sets of beams calculated using Bazant-Kim’s
equation and the ANN model (the two methods that best captured the relationship
between f 'c and the shear strength of RC beams) were plotted versus f c in Fig. 6.11.
Beams in each set share the same properties except f c, and beams in different sets with
similar f c share the same properties except the amount of shear reinforcement, Pyfyv.
Each set also contains experimental shear strengths for three beams having similar
properties tested by Mphonde (1989). Characteristics o f the generated beams and those
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6 173
5.0
■•-ANN
2.0
Experimental results in Fig. 6.11 show that increasing p / yv by 100 % from 0.345 MPa
(set #1) to 0.69 MPa (set #2) leads to an average increase in shear strength of about 60 %,
while increasing p / yv by 50 % from 0.69 MPa (set #2) to 1.034 MPa (set #3) leads to an
observations from the ANN response, that the effect of stirrups is less significant at
higher value o f p / yv. Figure 6.11 also shows that the contribution of f c to shear strength
of NSC slender beams is not significantly affected by the amount of shear reinforcement,
p j yv. However, such contribution in the case of HSC slender beams decreases at higher
values of p / yv.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6 174
6 .3 .3 Influence of S t ir r u p s on E ffect of T e n s il e S t e e l R a t io
beams was carried out to determine the effect of the longitudinal reinforcement ratio on
the shear strength, and the influence of stirrups on such an effect. Figure 6.12 shows the
relationship between longitudinal reinforcement ratio and the calculated shear strength of
a set of beams (Table E.6, Appendix E) using all shear design methods considered herein.
The figure also includes experimental shear strength values for three similar beams tested
by Sarsam and Al-Musawi (1992). It can be observed that the ANN model provided the
most accurate shear strength predictions among all shear design methods. The figure also
show that only the equations proposed by Zsutty and Bazant-Kim captured the trend of
the effect o f longitudinal steel ratio on the shear strength of RC slender beams with shear
reinforcement, but their quantitative predictions were less accurate than those of the ANN
model.
Shear strength values predicted by Zsutty’s equation for RC slender beams with shear
reinforcement were considerably lower than both experimental values and predictions of
the Bazant-Kim’s equation and the ANN model. This is due to the fact that the equation
proposed by Zsutty does not account for the influence of stirrups on other shear resisting
mechanisms, and, like most current design codes simply superimposes the capacity of
stirrups to the shear capacity of a similar beam not containing shear reinforcement in
calculating the shear capacity of RC beams having stirrups. On the other hand, Fig. 6.12
shows that not adequately considering the effect of longitudinal steel ratio on shear
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6 175
resisting mechanisms (e.g. ACI and the CSA simplified method), leads to even more
conservative shear strength estimates for RC slender beams with shear reinforcement.
6.5
-B-ANN
6.0 -® - Bazant-Kim
-A -M phonde
5.5 - B - Zsutty
-e -C S A
5.0 - A - ACI
• Experimental points
4.5
4.0
o>
3.5
3.0
2.5
2.0
1.5
d = 233 mm b* = 180 mm a/d = 2.5
1.0 Pyfyy = 1.14 MPa f c = 73 MPa
0.5
1.0 1.5 2.0 2.5 3.0 3.5 4.0
T ensile steel ratio (%)
The ANN model and Bazant-Kim’s equation show a similar effect of the longitudinal
reinforcement ratio, pi on the shear strength of typically reinforced slender beams (pi <
2.5 %). However, for highly reinforced beams, such effect is noticeably larger for the
ANN response than that for the Bazant-Kim’s equation. Nevertheless, the ANN response
Kong and Rangan (1998) in which they stated that increasing the longitudinal
reinforcement ratio from 1.66 % to 2.79 % leads to a small increase in shear strength, yet
a sharp increase in the shear strength was reported when increasing the longitudinal steel
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6 176
The influence of stirrups on the effect of pi on the shear strength of RC slender beams is
shown in Fig. 6.13. The figure illustrates the shear strength of three sets of beams versus
pi for different values o f Pyfyv- All beams have the same properties of beams in Table E.6
except for pj~yv, which has been changed from 1.14 MPa (set # 3) to 0.74 (set # 2) and
0.00 MPa (set # 1 ). Figure 6.13 also show the measured shear strength of 6 similar beams
tested by Sarsam and Al-Musawi (1992). All tested beams contain shear reinforcement;
three have pJyV= 0.75 MPa and the other three have pj'yv =1.14 MPa. Although the ANN
model and Bazant-Kim’s equation offered a nearly identical response in the case of RC
slender beams without shear reinforcement, their response was different for beams with
shear reinforcement. The Bazant-Kim’s equation assumes that the effect of pi on the shear
enhancement to such effect for pi > 2.5 %. Such enhancement in the effect of pi due to the
results reported by Sarsam and Al-Musawi (1992) and Kong and Rangan (1998) as
the shear capacity o f RC beams with stirrups is believed to be due to the fact that stirrups
help confining the longitudinal steel bars in place, thus preventing shear cracks from
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6 111
6.5
-B-ANN
6.0 -© - Bazant-Kim
♦ Experimental points-1.14
5.5
B Experimental polnts-0.754
_ 5.0
m
I 45
J= 4.0
O)
m 3.5
Pvfyv = 0.00 MPa
3.0
® 2.5
2.0
1.5
d = 233 mm b* = 180 mm
1.0 a/d = 2.5 f , = 73 MPa
0.5
1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5
T ensile steel ratio (%)
Fig. 6.13 - Influence o f stirrups on the effect of tensile steel ratio, pi on shear
strength o f RC beams.
The effect o f shear span to depth ratio on the shear strength of RC slender beams and the
influence o f stirrups on such an effect are illustrated in Figs. 6.14 and 6.15, respectively.
Figure 6.14 shows the variation in shear strength of a RC beam with variable shear span
to depth ratio (a/d) as predicted by the various shear prediction methods considered in
this study. While some of these methods disregard the effect of a/d on the shear strength
of RC beams (Mphonde and the simplified CSA methods), others incorporate a slight
variation in shear strength for 2.5 < a/d < 3.5 that becomes negligible for a/d > 3.5. The
largest variation in shear strength versus a/d was captured by the ANN model and
Zsutty’s equation (Fig. 6.14), yet such a variation is still relatively minor. On the other
hand, Fig. 6.15 shows the impact of adding stirrups on the effect of a/d on the shear
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6 178
strength o f RC slender beams. The figure presents variation in shear strength with a/d for
different values o f p / yv (0.0, 0.5, and 1.3 MPa) as predicted by the ANN model, Bazant-
Kim and Zsutty’s equations. Responses of all three methods show similar relationship
between the shear strength o f RC slender beams and a/d, regardless of the amount of
shear reinforcement. Therefore, increasing the amount of stirrups will have no impact on
—B — ANN
—®— Bazant-Kim
—A— Mphonde
- a - Zsutty
- o - CSA
- A - ACI
• Experimental points
o>
5 3.5
2
Is 3.0
A - -A -A
(0 2.5
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6 179
5.0
-a-A N N d =351 i b,, = 200 mm
-® - Bazant-Kim Pi = 2.29 r c = 87 MPa
4.5
- a - Zsutty
4.0
3.5
p v f y v = 1 .2 9 M P a
O)
3.0
•© p v f y v = 0 .5 0 M P a
2.0 p v f y v = ( J .O O M P a
1.5
2.00 2.25 2.50 2.75 3.00 3.25 3.50 3.75 4.00 4.25 4.50
a/d
Fig. 6.15 - Influence o f stirrups on the effect of shear span to depth ratio a/d
on shear strength of RC beams.
This chapter investigated the applicability of several existing shear calculation methods
in predicting the shear capacity of reinforced concrete slender beams with shear
reinforcement. The feasibility of using artificial neural networks for predicting the shear
capacity o f reinforced NSC and HSC slender beams with or without shear reinforcement
was also investigated. Furthermore, a parametric study was carried out to evaluate the
effect o f basic shear design parameters on the shear strength of RC beams and the
The ANN approach outperformed all other shear capacity calculation methods
shear capacity o f RC beams. As expected, it showed that shear strength increases with
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6 180
methods, the ANN model predicted that such an increase diminishes at higher values
Experimental results along with results obtained using the ANN model and Bazant-
Kim’s equation showed that the shear strength of RC slender beams with shear
reinforcement is more than 50 % higher than values calculated using shear provisions
o f ACI and the simplified method of CSA. It was shown that calculating the shear
shear capacity o f RC slender beams with shear reinforcement. It showed that shear
diminish above that value. It also showed that increasing the capacity of stirrups does
not impact the effect o f f c on the shear strength of NSC beams. However, for HSC
beams, the effect of f c on shear strength tends to decrease with increasing capacity of
stirrups.
The ANN model also showed that the amount of longitudinal tensile reinforcement
general, and that this influence is more pronounced for higher values of longitudinal
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6 181
Kim and Zsutty’s equations. Conversely, the shear provisions of the ACI code
underestimate such an effect, while the CSA simplified method does not account for it.
Finally, the ANN analysis showed that the shear span to depth ratio, a/d slightly
affects the shear strength of RC slender beams with shear reinforcement, and that such
an effect diminishes at higher values of a/d. This behaviour is somewhat consistent for
all methods considered in this study that account for the effect of a/d on shear strength
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6 182
6.5- REFERENCES:
Adebar P., Collins M. P., (1996), “Shear Strength o f Members without Transverse
Reinforcement,” Canadian Journal of Civil Engineering, V. 23, pp. 30-41.
Ahmad S. H., Khaloo A. R., and Poveda A., (1986), “Shear Capacity o f Reinforced High-
Strength Concrete Beams,” ACI Journal, V. 83, No. 2, pp. 297-305.
Anderson N. S., and Ramirez J. A., (1989), “Detailing o f Stirrup Reinforcement,” ACI
Structural Journal, V. 86, No. 5, pp. 507-515.
Bazant Z. P., and Kazemi M. T., (1991), “Size Effect on Diagonal Shear Failure o f
Beams without Stirrups,” ACI Structural Journal, V. 88, No. 3, pp. 268-276.
Bazant Z. P., and Kim J. K., (1984), “ Size Effect in Shear Failure o f Longitudinally
Reinforced Beams,” ACI Structural Journal, V. 81, No. 5, pp. 456-468.
Bresler B., and Scordelis A. C., (1963), “Shear Strength o f Reinforced Concrete Beams,”
ACI Journal, Proceedings V. 60, No. 1, pp. 51-72.
Chana P. S., (1981), “Some Aspects o f Modeling the Behaviour o f Reinforced Concrete
under Shear Loading,” Tech. Rep. No. 543, Cement and Concrete Association,
Wexham Springs, 21 p.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6 183
Collins M. P., and Kuchma D. (1999), "How Safe are Our Large, Lightly Reinforced
Concrete Beams, Slabs and Footings?” ACI Structural Journal, V. 96, No. 4, pp. 482-
490.
Debaiky S. Y., and Elniema E. I., (1982), “Behaviour and Strength o f Reinforced
Hunched Beams Using High Strength Concrete,” ACI Journal Proceedings, V. 79, No.
3, pp. 184-194.
Duthinh D., and Carino N. J., (1996), “Shear Design o f High-Strength Concrete Beams:
A Review o f the State-of-the-Art”, Research Report NISTIR 5870, National Institute of
Standards and Technology, Building and Fire Research Laboratory, Gaithersburg,
MD, 198 p.
Elzanaty A. H., Nilson A. H., and Slate F. O. (1986), '’''Shear Capacity o f Reinforced
Concrete Beams Using High-Strength Concrete,” ACI Journal, V. 83, No. 2, pp. 290-
296.
Haddadine M., Hong S. T., and Mattock A. H., (1971), “Stirrup Effectiveness in
Reinforced Concrete Beams with Axial Force,” Proceedings, ASCE, V. 97, ST9, pp.
2277-2297.
Hsiung W., and Frantz G. C., (1985), “Transverse Stirrup Spacing on RJC Beams,”
Journal o f Structural Engineering, ASCE, V.l 11, No. 2, pp. 353-362.
Islam M. S., Pam H. J., and Kwan A. K. H., (1998), “Shear Capacity o f High- Strength
Concrete Beams with Their Point o f Inflection within the Shear Span”, Proceedings of
the Institution o f Civil Engineers, Structures and Buildings, V. 128, No. 1, pp. 91-99.
Johnson M. K., and Ramirez J. A., (1989), “Minimum Shear Reinforcement in Beams
with Higher Strength Concrete,” ACI Structural Journal, V. 86, No. 4, pp. 376-382.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6 184
Kani M. W., Huggins M. W., and Wittkopp R. R., (1979), “Kani on Shear in Reinforced
Concrete,” Department o f Civil Engineering, University o f Toronto Press, Toronto,
Canada, 225 p.
Kim J. K., and Park Y. D., (1994), “Shear Strength o f Reinforced High Strength Concrete
Beams without Web Reinforcement,” Magazine of Concrete Research, V. 46, No. 166,
pp. 7-16.
Mattock A. H., and Wang Z. H., (1984), “Shear Strength o f Reinforced Concrete
Members Subjected to High Axial Compressive Strength,” ACI Journal Proceedings,
V. 81, No. 3, pp. 287-298.
McGormley J. C., Creary D. B., and Ramirez J. A., (1996), “The Performance o f Epoxy-
Coated Shear Reinforcement,” ACI Structural Journal, V. 93, No. 5, pp. 531-537.
Mphonde A. G., and Frantz G. C. (1984), “Shear Tests o f High and Low-Strength
Concrete Beams without Stirrups,” ACI Journal, V. 81. No. 4, pp. 350-357.
Ozcebe G., Ersoy U., and Tankut T., (1999), “Evaluation o f Minimum Shear
Reinforcement Requirements fo r Higher Strength Concrete,” ACI Structural Journal,
V. 96, No. 3, pp. 361-368.
Papadakis G., (1996), “Shear Failure o f Reinforced Concrete Beams without Stirrups,”
Ph.D. Thesis, Department of Civil Engineering, Aristotle University of Thessaloniki,
Greece, (in Greek).
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6 185
Pellegrino C., Bemardini A., and Modena C., (2002), “Shear Failure o f HSC Beams with
Variable Shear Span-to-Depth Ratio," Proceeding of the 6th International Symposium
on Utilisation of High Strength/High Performance Concrete,” Edited by Koenig,
Dehn, and Faust, Leipzig, Germany, pp. 473-485.
Placas A., and Regan P. E., (1971), “Shear Failure o f Reinforced Concrete Beams f ACI
Journal Proceedings, V. 68, No. 10, pp. 763-773.
Roller J. J., and Russel H. G., (1990), “Shear Strength o f High Strength Concrete Beams
with Web Reinforcement," ACI Structural Journal, V. 87, No. 2, pp. 191-198.
Russo G., and Puleri G., (1997), “Stirrups Effectiveness in Reinforced Concrete Beams
Under Flexure and Shear," ACI Structural Journal, V. 94, No. 3, pp. 227-238.
Sarsam K. F., and Al-Musawi J. M. S., (1992), “Shear Design o f High-and Normal
Strength Concrete Beams with Web Reinforcement," ACI Structural Journal, V. 89,
No. 6, pp. 658-663.
Swamy R. N., and Andriopoulis A. D., (1974), “Contribution o f Aggregate Interlock and
Dowel Forces to the Shear Resistance o f Reinforced Beams with Web reinforcement,"
Shear in reinforced Concrete, SP-42, American Concrete Institute, Farmington Hills,
Michigan, pp. 129-166.
Tan K-H., Kong F-K., Teng S., and Guan L., (1995), “High-Strength Concrete Deep
Beams with Effective Span and Shear Span Variations," ACI Structural Journal, V. 92,
No. 4, pp. 395-404.
Taylor H. P. J., (1972), “Shear Strength o f Large Beams," Journal of the Structural
Division, ASCE, V. 98, No. ST11, pp. 2473-2489.
Thorenfeldt E., and Drangsholt G., (1990), “Shear Capacity o f Reinforced High Strength
Concrete Beams," ACI 2nd International Symposium on HSC, ACI SP 121.8,
American Concrete Institute, Farmington Hills, Michigan, pp. 129-154.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6 186
Walraven J. C., (1978), “The Influence o f Depth on the Shear Strength o f Light Weight
Concrete Beams without Shear Reinforcement,” Stevin Lab. Rep. No. 5-78-4, Delft
University o f Technology, the Netherlands.
Xie Y., Ahmad S. H., Yu T., Hino S., and Chung W. (1994), “Shear Ductility o f
Reinforced Concrete Beams o f Normal and High Strength Concrete,” ACI Structural
Journal, V. 91, No. 2, pp. 140-149.
Yoon Y.-S., Cook W. D., and Mitchell D. (1996), “Minimum Shear Reinforcement in
Normal, Medium and High-Strength Concrete Beams,” ACI Structural Journal, V. 93,
No. 5, pp. 576-584.
Zararis P. D., and Papadakis G., (1999), “Influence o f the Arrangement o f Reinforcement
on the Shear Strength o f RC Beams,” Proceedings of the 13th Hellenic Conference on
Concrete, Rethymnon, Greece, V. 1, pp. 110-119, (in Greek).
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 187
CHAPTER 7
7.1-INTRODUCTION
costly rehabilitation and significant reduction in their service life. To alleviate such a
epoxy-coated, and galvanized steel bars. However, this practice did not completely
eliminate the problem (Kessler and Powers 1988). Recently, fibre-reinforced polymer
(FRP) has been proposed as an alternative material for reinforcing concrete structures due
to its high resistance to corrosion and high strength to weight ratio (ACI 440R 1996). The
mechanical properties o f FRP and its mode of failure differ from those of steel. Thus, a
* A version o f this chapter has been submitted for review to the Journal of Smart Structures and Systems.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 188
polymer resin and produced in the form of bars, tendons, strands, grids, etc. The fibres
are generally made o f glass, carbon, or aramid and constitute the principal load-carrying
element in the composite matrix. The tensile strength of FRP could reach up to twice that
of steel (Yost et al. 2001), which allows FRP to replace longitudinal steel reinforcement
Previous experimental and analytical studies concluded that the fundamental principles of
flexural theory of concrete beams reinforced with steel bars are also valid for beams
reinforced with FRP (Shehata 1999). However, due to their brittle mode of failure, a
higher safety factor should be used for design purposes when FRP is considered for
explained in chapter 5. Such a contribution depends mostly on the axial rigidity (EjAj)
and tensile strength o f the reinforcing material. Since the axial rigidity of FRP is smaller
than that o f steel, the amount of shear resisted by concrete, Fc/ is expected to be lower and
needs to be evaluated.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 189
1750
1500
1250
CFRP
^<v 1000
M
M High strength steel
£ 750
■«-*
to
500
Mild steel
250
GFRP
Fig. 7.1 - Typical stress-strain relationship for FRP and steel reinforcement.
Current shear design provisions calculate the shear capacity o f concrete beams reinforced
with FRP using slightly modified versions of existing semi-empirical shear design
equations that were primarily derived from experimental databases generated on concrete
beams having steel reinforcement. They reasonably assumed that the shear resistance
a similar manner to the nominal shear capacity of concrete members reinforced with FRP.
However, FRP materials have different mechanical properties and mode of failure than
those of steel, and extending existing shear design equations for steel reinforced beams to
cover concrete beams reinforced with FRP is questionable. For example, FRP materials
have a lower axial stiffness than that of steel reinforcement. Thus, a larger tensile strain
reduction in the depth of the neutral axis, wider cracks, and a reduction in the overall
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 190
Furthermore, it was found that bending FRP bars to form stirrups and the orientation of
tensile forces initiated by shear cracks (at an angle with respect to stirrups) limit the
ultimate capacity o f FRP bars. Hence, their full tensile strength, parallel to the direction
the capacity o f FRP stirrups when added to concrete contribution to shear to calculate the
shear capacity o f FRP-reinforced concrete beams. For instance, Shehata (1999) reported a
strength parallel to the fibres, and observed that failure of FRP stirrups was more likely to
occur at bends.
Due to the lack o f detailed and reliable design standards and the fact that FRP materials
do not offer the desirable ductility provided by steel reinforcement (brittle sudden
failure), current guidelines tend to highly underestimate the shear capacity of FRP
resist shear could be both costly and likely to create reinforcement congestion problems
Several studies were conducted to compare the ability of various shear design guidelines
to calculate the contribution of concrete and/or FRP stirrups to the nominal shear capacity
of concrete beams reinforced with FRP (Yost et al. 2001 and Razaqpur et al. 2004). It
was concluded that current shear design guidelines for the design of FRP-reinforced
objective o f the study in this chapter is to investigate the feasibility o f using artificial
intelligence to predict the shear capacity of concrete beams reinforced with FRP,
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 191
compare such predictions to results obtained from different shear design guidelines
namely, the provisions of the American Concrete Institute (ACI 440.1R 2003), the
Canadian Standards Association (CSA S806 2002), and the Japan Society of Civil
Engineers (JSCE 1997). The chapter also evaluates the performance of shear design
Sensing for Innovative Structures (ISIS Canada 2001) and propose shear design
Zsutty for concrete beams reinforced with steel (Zsutty 1968 and 1971). A parametric
study was also carried out to evaluate the ability of the ANN model, the proposed
equations, and current shear design guidelines to quantitatively account for the effects of
basic shear design parameters on the shear strength of FRP-reinforced concrete beams.
As stated in chapters 5 and 6, shear provisions of current design codes and standards,
(5.1). They also calculate the shear capacity resisted by concrete, Vc using semi-empirical
or statistically derived equations and ignore the influence of stirrups on the contribution
o f other shear mechanisms. Chapter 6 also shows that simply adding Vc and Vs to
beams. Moreover, design guidelines recommend using existing guidelines for steel-
reinforced members with proposed modification factors to calculate the shear capacity of
FRP-reinforced concrete members. Due to the brittle failure of FRP materials, current
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
C h apter7 192
concrete members to achieve a more conservative design. The various shear design
guidelines and shear prediction equations used in this study are briefly discussed below.
Note that all strength reduction factors used in the following equations for design
7.2.1 ACI R e c o m m e n d a t i o n s
Recommendations provided by ACI Committee 440 assume that equation. 5.1 is also
valid to calculate the shear capacity of FRP-reinforced concrete beams as long as shear
cracks are adequately controlled. Thus, taking into consideration the axial rigidity of the
Vc f = — E fl. vc < vc (7 . 1 )
90 A /c
Vc is calculated according to shear provisions of the ACI building code (ACI 318R-03)
K = (o. 15 8 / / f + 17 . 2 p l
V
^T)
))
M * 0 3 j Z b w<] (ACI 11-5) (7.3)
where bw is the beam’s breadth (mm); f c is the compressive strength of concrete cylinder
(MPa); Pfi is the longitudinal reinforcement ratio; Vc is the shear capacity of the
reinforced concrete beam without shear reinforcement (N); Ep is the modulus of elasticity
of the longitudinal FRP reinforcement (MPa); /?, is a material reduction factor that
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 193
depends on f c, 0.65 < A < 0.85; and M and V are the moment and shear force, at a critical
section, respectively.
ACI 440-03 calculates the shear capacity provided by FRP stirrups using shear equations
of ACI 318 (equation 6.1). However, it limits the stress level in the FRP shear
reinforcement to control the width of shear cracks and avoid failure at the bent portions
of stirrups. According to ACI 440-03, the shear capacity of FRP shear reinforcement is
calculated as follows:
where fo is the strength at the bend portions of FRP stirrups, and and Efv are the ratio
7 .2 .2 CSA G u i d e l i n e s
(7.5)
For reasons similar to those explained above in ACI recommendations, the shear capacity
provided by concrete, Vcf and by FRP stirrups, I /in FRP-reinforced concrete beams with
• vr d 1
Vcf = 0.035 f c Pji Efi - j- b .d a/d> 1.0 (7.6)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 194
(7.6a)
For beams having an effective depth d > 300 mm and without shear reinforcement, or
with shear reinforcement less than the minimum specified, Fc/ is calculated using:
130
*V = 1000 \[ 7 ,b J < 0 -0 z J 7 ,b J (7.8)
+d.
= (79)
Yb
1100 p a E n
A = 1 /1 5 ^ S 1.5 (7.9,)
where fin = 1.0 for members without axial force, yb is a strength reduction factor, f cd =
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 195
The shear capacity carried by FRP shear reinforcement in non pre-stressed flexural
f ncd P f l E fl ^ f bbend
e fi =0.0001. (7.10a)
Pf<E .
0.05— +0.3 f ,
fu v
fb e n d (7.10b)
Y m fb
where is the design value to limit the maximum strain in the FRP shear
reinforcement, j d = d! 1.15, a s is the angle between the shear reinforcement and the
member axis, h is the member’s height, ymJb =1.3 (safety factor for the bend portions of
FRP bars), and r* and dj, are the bend’s radius and bar’s diameter, respectively.
7.2.4 ISIS C a n a d a M e t h o d
In its design manual for FRP-reinforced concrete structures, (ISIS Canada-01) adopted
the same principles used by CSA A23.3 (1994) for shear design of steel-reinforced
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 196
I f Ef.
V f = P p X f j u v K d ^ °-8b^ d J -J r :— (x is assumed = 0.4) (7.13)
7.2.5 A r t if ic ia l N e u r a l N e t w o r k s A p p r o a c h
Shear strength results for 150 FRP-reinforced concrete beams (50 of which are without
shear reinforcement) were collected from published literature. Only simply supported
beams that exhibited shear failure were considered. Beams without shear reinforcement
were reinforced with FRP bars as longitudinal reinforcement. All beams with shear
reinforcement were reinforced with FRP stirrups, 72 of which had FRP bars and the
others had steel bars for longitudinal reinforcement. The database was compiled in a
patterned format. Each pattern consists of an input vector containing the geometrical and
shear capacity o f the beam. Table 7.1 provides the range and average values of all
parameters used in the database while beams characteristics are included in Table F.l
(Appendix F).
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 197
Table 7.1 - Statistical data of shear design parameters and shear capacity for FRP-
reinforced concrete beams used in database
Without shear reinf. (50 beams) With shear reinf. (100 beams)
Parameters
Minimum Maximum Average Minimum Maximum Average
d (mm) 150.0 360.0 253.1 210.0 500.0 291.2
by, (mm) 150.0 1000.0 360.5 135.0 300.0 200.8
a/d 1.8 6.5 3.7 1.2 4.3 2.7
#Ei(GPa) 0.3 3.2 0.9 0.3 9.6 2.0
Pffbcnd (MPa) 0.0 0.0 0.0 0.3 13.4 3.2
f c (MPa) 22.7 49.0 38.5 22.5 84.2 39.5
Vn (kN) 31.5 190.0 87.8 56.2 375.5 164.1
The network architecture adopted in this study to predict the shear strength of FRP
reinforced concrete beams consists of an input layer, an output layer and one hidden
layer. The input layer contains six variables representing the common shear design
parameters o f reinforced concrete beams (d, bw, a/d, Pffbend, and f c), in addition to the
axial rigidity o f the longitudinal reinforcement, pi Ei. The output layer includes one unit
representing the shear capacity, V„ and the hidden layer consists of seven processing units
(Fig. 7.2). Full forward connection (between units of one layer and those of the
subsequent layer) was adopted and variable learning rate and momentum were used to
avoid lengthy training and ensure global convergence as explained in chapter 2. The
transfer function used for all processing units was a logarithm sigmoid function (see
chapter 2) with outputs varying between 0 and 1. Therefore, prior to the training process,
variables in the generated database were scaled between 0 and 1 to speed up the training
process and improve the network’s generalization. The output values were scaled by
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 198
simply dividing the shear capacity of each beam by the maximum shear capacity in the
database, and input parameters were scaled using the following equation:
x, = 0 .1 + 0 .8 (7.14)
(Vx max - X m ■
in)1
where xt is the scaled value of variable x, and xmin and xmax are the minimum and
The primary factor that influences the performance of an ANN model is the completeness
and accuracy o f the database. Therefore, due to the relatively small number of beams
without shear reinforcement, the ANN model thus developed was trained and validated
on data for beams containing shear reinforcement. The ANN model was trained on 80 %
of the data and tested on the remaining 20 %. The testing patterns were randomly selected
from the original database and were not used in the training process (shaded rows in
Fig. 7.2 - Architecture of neural network model for predicting the shear capacity
of FRP-reinforced concrete beams.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 199
7 .2 .6 G e n e t ic A l g o r it h m s A p p r o a c h
an equation for the shear design of FRP-reinforced concrete beams with or without shear
reinforcement. Thus, an original form of the equation defining the overall shear
behaviour and including all shear design parameters that influence the shear capacity of
concrete beams is required. Zsutty’s equation (eq. 5.2) has proven to be one of the
simplest and most accurate empirical shear equations for conventional reinforced
concrete beams (see chapter 5) and its empirical form was chosen herein for optimization
using GA.
Zsutty’s equations were developed for steel-reinforced concrete beams. Thus, their
empirical forms were modified to account for the axial rigidity of FRP longitudinal
reinforcement and the capacity of FRP stirrups. Shear parameters considered in the
original form o f the equation are: the beam’s effective depth (d) and width (bw); the shear
span to depth ratio (a/d); the longitudinal reinforcement ratio (pj)\ the concrete
compressive strength ( f c)\ the ratio of FRP shear reinforcement (/?/„); the ratio of E /E s\
and the ultimate capacity of FRP shear reinforcement (ffuv). The original form of the shear
(7.15)
(7.15a)
(7.15b)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 200
The primary objective of using GA to optimize the above shear equations is to search for
an optimum set o f coefficients from within the solution space, for which the difference
between the measured shear strength of FRP-reinforced concrete beams and those
calculated using the final form of the optimized equations is minimal. The database used
to train and validate the ANN model was also used for the GA approach (Table FI,
Appendix F). GA is case-dependent and the appropriate selection of its key parameters
such as, selection method and pressure, recombination name and rate, mutation rate and
performance.
It is important to note herein that an attempt was made to optimize equation 7.15 with
both terms (vc/ and v/) included. A good agreement between measured nominal shear
strength and those calculated using the final equation obtained from GA was achieved.
However, such an equation was very conservative in calculating the shear strength of
decided to develop two separate GA models. The first one is to construct an equation to
predict the shear strength o f FRP-reinforced concrete beams without shear reinforcement,
vcf and to use such an equation in the objective function of the second GA model to
construct an equation to calculate the shear strength carried by FRP stirrups, v/. The final
setting o f both models is presented in Table 7.2 and the final form of the optimized
equations is as follows:
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 201
>7=0.74 (7 . 17)
As mentioned earlier, concrete members reinforced with FRP materials usually exhibit a
brittle and sudden shear failure. Thus, a higher material reduction factor or relatively
conservative shear equations should be used for design purposes when FRP is considered
for reinforcement. In this study and for comparative reasons, material reduction factors in
all shear equations considered herein were set to one. Furthermore, the fundamental task
measured shear strength o f FRP-reinforced concrete beams and its calculated values. In
other words, GA will provide equations capable of estimating shear strength values of
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 202
responsibility of the modeler to simplify such equations and modify them appropriately to
The performance o f equations 7.16 and 7.17 in calculating the shear strength of FRP-
reinforced concrete beams with and without shear reinforcement is shown in Figs. 7.3
and 7.4, respectively. The figures show a reasonable agreement between the measured
and calculated shear strengths. However, for the case of FRP-reinforced concrete beams,
a slightly more conservative result is preferred and equations 7.16 and 7.17 were slightly
modified to obtain the final proposed form (eqs. 7.18 and 7.19) for shear design equations
predicting the shear strength o f FRP-reinforced concrete beams with and without shear
(7.19)
Furthermore, the performance of the ANN model thus developed and trained was
evaluated using the testing database described earlier. The network was presented with
the input vectors o f the testing patterns and asked to predict the corresponding shear
capacities. The predicted shear capacities along with those calculated using ACI 440.1R-
03, CSAS806-02, JSCE-97, ISIS Canada-01, and the proposed equations were
method was evaluated based on both the ratio of measured to predicted (or calculated)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 203
shear capacity ( VJVP), and the average absolute error (AAE) calculated using equation
(7.20):
1 \v - V I
AAE = ~ y l— —x 100 (7.20)
« Vm
The average, standard deviation (STDV), and coefficient of variation (COV) for VdVp,
and the average absolute error (AAE) of the ANN model and other shear calculation
Table 7.3 - Performance of shear calculation methods considered in this study (FRP-
reinforced concrete beams)
Without shear reinf. (50 beams) With shear reinf. (100 beams)
Method Vmeasured / ^predicted ^measured ! ^predicted
AAE(%) AAE ( % )
Average STDV COV Average STDV COV
ACI 440.1R-03 6 8 .3 5 4 .0 2 2 .11 5 2 .4 1 4 5 .5 7 1 .90 0 .91 4 7 .8 2
CSA S806-02 3 3 .1 3 1.68 0 .6 6 3 9 .3 7 1 7 .4 9 1.13 0 .2 7 2 4 .1 4
JSCE-97 3 3 .1 6 1 .69 0 .7 1 4 1 .8 4 5 0 .0 5 2 .2 2 0 .7 3 3 2 .6 7
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 204
2.5
o Original GA equation (AAE = 21 %, Average Vm/Vp = 1.13)
o>
2 1.5
T3
■= 0.5
Fig. 7.3 - Measured versus calculated shear strength using GA (beams without stirrups).
8.0
o Original GA equation (AAE = 18%, Average Vm/Vp = 1.04)
7.0 A Modified GA equation (AAE = 22 %, Average Vm/Vp = 1.30)
6.0
o>
£ 5.0
4.0
■a 3.0
3 2.0
Fig. 7.4 - Measured versus calculated shear strength using GA (beams with stirrups).
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 205
7.3.1 B eam s w it h o u t S h e a r R e in f o r c e m e n t
A comparative study was conducted in this chapter to evaluate the performance of the
proposed shear equation (modified Zsutty’s equation) and those provided by current
shear design guidelines and recommendations in calculating the shear strength of FRP-
reinforced concrete beams without shear reinforcement. The shear strength of such beams
calculated using current shear design provisions and those calculated using the proposed
equation are plotted against the experimentally measured values in Figs. 7.5 - 7.9. It is
shown that ACI shear design guidelines for FRP reinforced beams are highly
conservative even without the application of reduction factors (Fig. 7.5). The guidelines
assume (eq. 7.1) that the shear strength of FRP-reinforced concrete beams increases
linearly with the axial rigidity, pi Ei of the longitudinal reinforcement and decreases as the
the shear capacity o f highly reinforced concrete beams, especially beams reinforced with
carbon fibre-reinforced polymer CFRP (high Ej), whereas the latter assumption yields
Shear design equations proposed by CSA S806-02, JSCE-97, and ISIS Canada-01
generally provided conservative results again even before applying reduction factors
(Figs. 7.6 - 7.8) with an average VJVP of 1.69. However, the ISIS shear equations (eqs.
7.11 and 7.12) generated more scattered results than those of CSA and JSCE with a COV
and recommendations simply assume that existing shear equations, initially derived for
steel-reinforced concrete beams, are applicable to concrete beams reinforced with FRP
bars provided that they are modified to account for the axial rigidity of the longitudinal
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 206
reinforcement. However, these methods differ in accounting for the magnitude of the
effect of axial rigidity o f FRP bars on the total shear resistance, Fc/ o f concrete beams.
The CSA S806 and JSCE-97 provisions assume that such an effect is to the magnitude of
(E/Es)m , whereas the ISIS method considers the effect to be (E/ES)U2. Furthermore, the
ISIS method does not consider the contribution of other shear design parameters on Vcf,
such as the longitudinal steel ratio, pi and the shear span to depth ratio, a/d. This could
explain the very conservative values of shear strength calculated using shear equations of
ISIS in the case o f slightly reinforced slender concrete beams (pi < 0.5 % and a/d > 2.5).
The proposed equation (eq. 7.18), provided the most accurate results for the shear
strength o f FRP-reinforced concrete beams (Fig. 7.9). The average value of VJVP for the
proposed equation is 1.35 compared to 4.02 for ACI and 1.68 for CSA, JSCE, and ISIS.
The proposed equation also provided the lowest COV and AAE of 33 and 22 %
respectively, compared to 52 and 68 % for ACI, 39 and 33 % for CSA, 41 and 33 % for
JSCE, and 48 and 34 % for ISIS. The equation provided by CSA S806 is similar to that
developed by Zsutty (1968) and it accounts for the effect of most shear design parameters
on the shear strength o f FRP-reinforced concrete beams. However, it does not consider
Zsutty’s recommendation in calculating the shear strength of beams having a/d < 2.5 for
The shear strength o f FRP-reinforced concrete beams calculated using the shear
equations provided by CSA S806 and Zsutty’s recommendation for beams with a/d < 2.5
(modified CSA) are also plotted in Fig. 7.6. It is shown that implementing Zsutty’s
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 207
suggestion to multiply Fc/b y 2.5d/a yields more accurate yet conservative results for the
CSA method and reduces the average Vm/Vp from 1.68 to 1.52, and the AAE from 33 % to
shear equations proposed by ACI 440, the shear strength of FRP-reinforced concrete
beams were calculated using equation 7.3 multiplied by (E /E s)m and results are plotted
in Fig. 7.5. It is shown that the shear strength of FRP-reinforced concrete beams
calculated using the modified ACI 440 became reasonably accurate and comparable to
those calculated by CSA, JSCE, and the proposed equation. This suggests that the shear
2.5
O ACI 440: AAE 68 %, Average(Vm/Vp) = 4.0
▲Modefied ACI 440: AAE = 30 %, Average (Vm/Vp) = 1.6
ra
1.0
O 0.5 A o
0.0
0.0 0.5 1.0 1.5 2.0 2.5
M easured sh e ar strength (MPa)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 208
2.5
O CSA S806: AAE = 33 %, Average (Vm/Vp) = 1.68
A Modified CSA S806: AAE = 28 %, Average (Vm/Vp) = 1.52
£ 2.0
o>
1.5
1.0
0.5
0.0
0.0 0.5 1.0 1.5 2.0 2.5
M e a su re d s h e a r s tr e n g th (MPa)
2.5
AAE = 33 %
Average (Vm/Vp) = 1.7
&2.0
o>
£ 1.5
m 1.0
T5
■=0.5 -A A * A
0.0
0.0 0.5 1.0 1.5 2.0 2.5
M e a su re d s h e a r s tr e n g th (MPa)
Fig. 7.7 - Performance o f JSCE shear equation in calculating the shear strength
of FRP-reinforced concrete beams without stirrups.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 209
2.5
AAE = 34 %
Average (Vm/Vp) = 1.7
<o
^ 2.0
o>
£ 1.5
w
k_
(0
tt
w 1.0
■o M
10
I-
o
0.0
0.0 0.5 1.0 1.5 2.0 2.5
M e a su re d s h e a r s tr e n g th (MPa)
Fig. 7.8 - Performance o f ISIS Canada in calculating the shear strength of FRP-
reinforced concrete beams without stirrups.
2.5
AAE = 26 %
Average (Vm/Vp) = 1.4
& 2.0
1.5
1.0
■o
f t * A f
0.5 AA
0.0
0.0 0.5 1.0 1.5 2.0 2.5
M e a su re d s h e a r s tr e n g th (M Pa)
Fig. 7.9 - Performance o f the proposed equation in calculating the shear strength
o f FRP-reinforced concrete beams without stirrups.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 210
7 .3 .2 Beam s w it h Shear r e in f o r c m e n t
ANN model predictions of shear strength for concrete beams reinforced with FRP
stirrups, and those calculated using the proposed equation and design guidelines
discussed earlier are plotted against the experimentally measured values in Figs. 7.10 -
7.15. Figure 7.15 shows that the data points predicted by the ANN model are located
either on, or slightly over/under, the equity line for both training and testing data, and
reinforced concrete beams compared to all other methods considered in this study.
Despite the highly conservative nature of shear equations proposed by ACI 440 in
calculating vc/(F ig. 7.5), these equations, in some cases, offered unsafe predictions of the
nominal shear strength v„ = vCf + v/ (Fig. 7.10). This indicates that ACI 440 tends to
overestimate the shear capacity carried by FRP shear reinforcement. Analysis of the
database indicated that most o f the beams for which ACI 440 overestimated the nominal
shear strength had either a high shear reinforcement ratio, ppv or were reinforced with
stirrups having a high modulus of elasticity (£/). Similar observations are valid for shear
equations proposed by ISIS (Fig. 7.13), and to a lower extent for equations proposed by
Shear provisions o f ACI 440, CSA S806, and ISIS adopted the principles of equation
(5.1) in calculating the nominal shear capacity, V„. However, such provisions assume that
FRP stirrups reach their capacity at failure and that such capacity varies linearly with ppv
E/v (in the case of ACI) or with pfvffuv (for CSA and ISIS). Such assumptions lead to
overestimating the shear capacity of concrete beams reinforced with FRP materials of
high ultimate strength or high modulus of elasticity. Furthermore, CSA S806 (eq. 7.7)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
C h apter7 211
and ISIS (eq. 7.13) consider that the capacity of shear reinforcement depends on the
strength o f FRP bars at the bent portion, which is assumed to be 40 % of the ultimate
tensile strength parallel to the direction of FRP fibres. However, an experimental study
conducted by Duranovic et al. (1997) to investigate the shear strength and mode of
failure of concrete beams reinforced with FRP stirrups lead to a conclusion that stresses
in FRP stirrups at failure never exceeded 65 % of their bend’s capacity. Another study by
Zhao et al. (1995) also lead to the conclusion that the strain distribution in stirrups along
a diagonal shear crack is not constant and can be expressed as a cubic root function. This
could explain the relatively high nominal shear strength of FRP-reinforced concrete
beams calculated using shear provisions of CSA S806 and ISIS Canada despite their
relative conservative calculation of the shear capacity carried by concrete alone, Vcf.
Figure 7.14 shows that the equation proposed in this study (eq. 7.19) provided better
results than other shear design methods considered herein in calculating the nominal
shear strength o f concrete beams reinforced with FRP stirrups. The first term in the
equation, vc/ is similar to that proposed by CSA S806, yet the proposed second term
assumes that the capacity o f FRP stirrups is a function of the square root of Pfvffuv, as
opposed to a linear relationship adopted by CSA. Shear equations of CSA S806 offered a
lower AAE than that o f the proposed equation (Table 7.3). However, Fig. 7.11 shows that
CSA S806 overestimated the shear strength of more than 30 % of beams in the database,
and in some cases the calculated shear strength was 40 % higher than the measured value.
Conversely, the proposed equation offered safe predictions for more than 95 % of the
beams with the highest overestimated shear strength was about 10 % higher than the
measured value. Due to the brittle and sudden failure of FRP materials, a safer design is
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 212
required, thus the proposed equation generally offered a safer yet accurate design
8.0
AAE = 45 %
7.0 A verage Vm/Vp = 1.9
ais
S
6.0
O)
IS 44
3 2.0
O
IS
O 4
0.0
0.0 1.0 2.0 3.0 4.0 5.0 6.0 7.0 8.0
Measured shear strength (MPa)
7.0
AAE = 17 %
A verage Vm/Vp = 1.13
Is 6.0
5 5.0
o>
* 4.0
■e 3.0
0.0
0.0 1.0 2.0 3.0 4.0 5.0 6.0 7.0
Measured shear strength (MPa)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 213
6.0
AAE = 50 %
Average Vm/Vp = 2.22
S. 5.0
S
o> 4.0
0)
</>
<5 3.0
£o>
(A
■0o) 2.0
(0
3
O
o 10
0.0
0.0 1.0 2.0 3.0 4.0 5.0 6.0
Measured shear strength (MPa)
Fig. 7.12 - Performance of JSCE shear equations in calculating the shear strength
o f FRP-reinforced concrete beams with stirrups.
9.0
AAE = 30 %
ra 8.0
Average Vm/Vp = 1.06
(L
S 7.0
2 > 6.0
5.0
4.0
3.0
2.0
0.0
0.0 1.0 2.0 3.0 4.0 5.0 6.0 7.0 8.0 9.0
Measured shear strength (MPa)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 214
8.0
▲Training data: AAE = 19 %, Average Vm/Vp = 1.23
O Testing data: AAE = 20 %, Average Vm/Vp = 1.24
3.0
2.0 A J Z k t? A A -A-'
0.0
0.0 1.0 2.0 3.0 4.0 5.0 6.0 7.0 8.0
Measured sh ear strength (MPa)
7.0
A Trainning data: AAE = 4 %, Average Vm/Vp = 1.00
5.0
2.0
0.0
0.0 1.0 2.0 3.0 4.0 5.0 6.0 7.0
Measured sh ear strength (MPa)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
C hapter7 215
Current shear design provisions for FRP-reinforced concrete members generally assume
that adding shear reinforcement to a RC beam will only enhance its shear strength by the
shear capacity o f stirrups. Such a practice presumes a linear relationship between the
amount o f shear reinforcement and the nominal shear strength and ignores the influence
of stirrups on other shear resisting mechanisms. A sensitivity analysis was also conducted
in this chapter using the ANN model to investigate the effect o f FRP stirrups on the shear
strength o f RC beams. The shear strength of a set of beams having geometrical and
mechanical properties similar to those of a beam randomly selected from the database
were calculated for different amounts of shear reinforcement using the shear methods
discussed earlier, and results are plotted in Fig. 7.16. Properties of beams used in this set
along with those o f two beams tested by Vijay et al., (1996) are shown in Table F.2,
Appendix F.
4.0
-Ar-Proposed
• EXPpoints
3.5
£•2.5
1.5
1.0
0.5 d = 260 mm, bw=150 mm, a/d - 1.89,
rc = 45 Mpa, and p, Eg = 0.78 GPa
0.0
1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4
Stirrups capacity (MPa)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 216
The figure shows the effect o f stirrups capacity on the nominal shear strength of FRP-
reinforced concrete beams. It can be observed that the ANN response is the closest to the
According to ACI 440, and for the range of FRP shear reinforcement considered in Fig.
7.16, increasing the amount o f shear reinforcement by 80 % increased the shear strength
by more than 60 %. The provisions of CSA S806 showed a slight increase of about 2 %,
whilst ANN, JSCE, and the proposed equation all illustrated an increase of 11 to 14 %.
The large increase in shear strength due to increase in pfv as calculated by ACI 440 is due
to the distinguished linear relationship assumed for the effect of ppv on Vn by such
provisions.
A similar analysis to that used in investigating the effect of FRP stirrups was carried out
to determine the effect of the longitudinal reinforcement ratio on the shear strength of
FRP-reinforced concrete beams. Figure 7.17 shows the relationship between piEi and the
shear strength o f a set o f beams calculated using the ANN model and shear design
methods considered herein. The figure also includes the experimental shear strength of
two similar beams measured by Tottori and Wakui (1993). All beams share similar
characteristics except the longitudinal reinforcement ratio, pi (Table F.3, Appendix F). It
is apparent that all methods, including the ANN model consider similar influence for the
effect of pi on shear strength. However, ACI 440 assumes a linear relationship for such an
effect, as opposed to a non-linear effect for the other methods. The figure also shows that
the amount o f FRP flexural reinforcement exerted a moderate to slight effect on the shear
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 217
strength o f RC beams with FRP shear reinforcement. Moreover, the ANN model
response as well as all other methods considered herein except ACI 440 show that the
3.0
- s - c s a -A -P ro p o s e d
-♦ -A N N • E X P p o in ts
2.5
2.0
o>
£1-5
o 1.0
0.5
d = 325 mm, b w = 200 mm, aid = 3.23,
r e = 39.4 MPa, ptv = 0.23 %, and = 1236 MPa
0.0
0.0 0.5 1.0 1.5 2.0
p, E, (Gpa)
the shear strength o f RC beams. In conventionally reinforced concrete beams with shear
reinforcement, stirrups restrain shear cracks from widening and provide confinement to
concrete in the compression zone, leading to a significant increase in the shear capacity
of concrete. Due to the high tensile strain and low modulus of elasticity of FRP, such a
benefit needs to be examined and validated. To investigate the ability of shear design
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
C h apter7 218
guidelines considered in this study to quantitatively account for the effect off c on shear
randomly selected from the database and tested by Tottori and Wakui (1993) were
considered (Table F.4, Appendix F). The shear strength of all beams calculated using the
various methods and those predicted by the ANN model are plotted against f c in Fig.
7.18. It is shown that all shear design methods account for the effect off c, yet they vary
in the magnitude o f such an effect. While ISIS assumes that the effect of f c on shear
increases (Fig. 7.18). All other methods, including the ANN model and the proposed
7.0
-A -P ro p o s e d
-♦ -A N N • E X P p o in ts
6.0
5.0
4.0
ra
3.0
2.0
1.0
d = 285 mm, bw = 200 mm, a/d =2.11,
Ph =0.27%, ffuv =602 MPa, and p, E, = 8.38
0.0
20.0 25.0 30.0 35.0 40.0 45.0 50.0 55.0
Tc (MPa)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
C h apter7 219
Figure 7.19 shows the variation in shear strength of FRP-reinforced concrete beams with
variable shear span to depth ratio (a/d). The figure illustrates the effect of a/d as
estimated by the various shear provisions and calculation methods considered in this
chapter. While some of these methods disregard the effect of a/d on the shear strength of
RC beams (ACI 440 and JSCE provisions), others (CSA S806 and proposed method)
incorporate a slight to negligible effect regardless of the value of a/d. The ANN response
beams when a/d > 2.6, but showed a larger effect for smaller values of a/d. Such a
behaviour is valid for steel-reinforced concrete beams due to the arch action mechanism
of deep beams and it is reasonable to assume a similar effect for FRP-reinforced beams.
However due to the limited experimental data on deep beams reinforced with FRP,
4.0
- A - P ro p o se d
• E X P p o in ts
3.5
O)
£ 2.0
1.5
1.0
0.0 4 -
Fig. 7.19 - Effect o f shear span to depth ratio on shear strength o f RC beams
with FRP shear reinforcement.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 220
As stated earlier, most shear design provisions for FRP-reinforced concrete beams are
based on modified versions of existing shear equations initially derived for steel
reinforced concrete beams. These provisions vary greatly in the number of parameters
they consider and the intensity of the effect of such parameters on shear strength. Based
on the analysis conducted above and bearing in mind the brittle and sudden failure of
FRP materials, the following modifications to ACI 440 and CSA S806 equations are
and 7 .2 2 b ), respectively. The ability of these modified equations to safely estimate the
'Vd'
C = 0 . 1 5 8 ^ + 17.2^ b j < 0 .3 ^ M ( 7 .2 1 )
k M jj
V Pfv Jfuv
( 7 .2 1 a)
Vf = ^ ~ 2 b* d
sl/3
Vcf = 0 .0 3 5 fc Pfi E f Kd fox a/d > 2 . 5 ( 7 .2 2 )
Mf
0.1J 7 , b , d Z V 4 Z 0 . 2 / f cb„d ( 7 .2 2 a)
( 7 .2 2 b)
Figure 7 . 2 0 shows the shear strength o f RC beams calculated using current and suggested
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 221
equations of ACI 440) yield safer and less scattered results than current equations (eq. 7.1
and 7.4). Incorporating the suggested modifications into current equations of ACI 440
lowered the average Vm/Vp from 1.90 to 1.43, and the STDV, COV, and the AAE from
0.91, 47.82, and 45.57 to 0.37, 25.70, and 27.61, respectively. The performance of the
equations adopted by CSA S806 and the suggested modified version of such equations
(eq. 7.22 and 7.22b) is shown in Fig. 7.21. It is shown that current equations of CSA S806
yield more accurate results than the modified ones. However, these equations
overestimated the shear strength of a large number of concrete beams reinforced with
FRP stirrups, and in some cases the overestimation was more than 35 %. Bearing in mind
that the shear failure o f FRP-reinforced concrete beams is brittle and sudden, shear
equations for FRP-reinforced concrete beams should be accurate yet conservative. Such
requirements are satisfied by the modified version of CSA S806 equations as shown in
Fig. 7.21.
10.0
O ACI 440: AAE = 46 •/., Average Vm/Vp = 1.9
9.0 A Modified ACI 440: AAE = 28 %, Average Vm/Vp = 1.43
a.ra 8.0
S
oc> 7.0
00
£w
5.0
(A 4.0
U
« 30
J 2.0
M
° 1.0
0.0
0.0 2.0 4.0 6.0 8.0 10.0
Measured shear strength (MPa)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 222
7.0
O CSA S806: AAE = 17 %, Average Vm/Vp = 1.13
o>
■a 4.0
M 3.0
W,V
« 2.0
Furthermore, the ANN model thus developed is highly accurate but slightly over or
underestimates the shear strength. Thus, to use it for design purposes, a modified version
of ANN predictions is expressed in equation (7.23) and its predictions are shown in Fig.
7.22.
Vcf (ANNm) is the modified value of the predicted shear capacity Vcf (ANNP). Note that
these equations do not include strength reduction factors and if considered for design
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 223
7.0
A Training: AAE = 9 %, Average Vm/Vp = 1.07
~ 5.0
o>
c
A
£ 4.0
3.0
■o
2.0
1.0
0.0
0.0 1.0 2.0 3.0 4.0 5.0 6.0 7.0
Measured sh ear strength (MPa)
The study presented in this chapter investigated using artificial neural networks for
predicting the nominal shear strength of FRP-reinforced concrete beams with shear
reinforcement, and comparing its predictions with those o f several existing shear design
and calculation methods. The genetic algorithms approach was also used to develop shear
equations for the design o f FRP-reinforced concrete beams. Moreover, a parametric study
was carried out to evaluate the effect of basic shear design parameters on the shear
strength of FRP-RC beams and the accuracy of current shear design provisions in
capturing the effect o f such parameters on shear strength. The following conclusions can
be drawn:
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 224
• A successfully trained ANN model can be used as an effective tool for predicting the
shear strength o f FRP-RC beams and evaluating the effect of basic shear design
parameters on the shear behaviour of such beams. The ANN approach outperformed
all other shear design and calculation methods considered in this study.
• Shear provisions o f ACI 440 are extremely conservative in estimating the shear
strength o f FRP-RC beams without shear reinforcement. All other shear provisions
• The equation proposed by ACI 440 to calculate the capacity of FRP stirrups, Vf is
overestimates the capacity of FRP stirrups with high modulus of elasticity to a degree
where the nominal shear capacity of FRP concrete beams reinforced with such
Vcf-
capacity supplied by FRP shear reinforcement for low shear capacity beams.
especially for beams having high shear capacity. The CSA S806 on the other hand,
provided reasonably accurate estimations for the shear strength of concrete beams
having relatively low shear reinforcement ratio, but it overestimated the shear
capacity o f highly reinforced beams with FRP stirrups of high tensile strength. Due to
the brittle failure o f FRP materials, such equations need to be re-evaluated for the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 225
• The shear equations proposed in this study incorporated the axial rigidity of FRP
longitudinal bars as a cubic root function and assumed that the contribution of FRP
stirrups to shear strength is a square root function rather than a linear function as
proposed by CSA S806 and ACI-440. It provided the most accurate (but still
conservative) results in calculating both VCf and Vf amongst existing shear evaluation
methods considered.
• The effect of the axial rigidity of FRP longitudinal reinforcement on the shear
function rather than linear or square root relationships as proposed by ACI 440 and
ISIS-01, respectively. CSA S806, JSCE, and the proposed equation adopted such
relationship and incorporating it into ACI 440 yielded better estimations of shear
strength.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 226
7.7- REFERENCES:
ACI 440.1R, Technical Committee Document, (2003), “Guide fo r the Design and
Construction o f Concrete Reinforced with FRP Bars,” American Concrete Institute,
Farmington Hills, Michigan, 42 p.
ACI Committee 318, (ACI 318R-03), (2003), “Building Code Requirements for
Structural Concrete,” American Concrete Institute, Farmington Hills, Michigan, USA,
445 p.
Alkhrdaji, T., Wideman, M., Belarbi, A., and Nanni, A., (2001), “Shear Strength o f RC
Beams and Slabs,” Composites in Construction Journal, pp. 409-414.
Alsayed S.H., Al-Salloum Y. A., and Almusallam, T.H., (1996), “Evaluation of Shear
Stress in Concrete Beams, Reinforced by FRP Bars,” Proceedings of the srcond
International Conference on Advanced Composite Materials for Bridges and
Structures, Montreal, Quebec, pp. 173-179.
Duranavic, N., Pilakoutas, K., and Waldron, P., (1997), “Test on Concrete Beams
Reinforced with Glass Fiber Reinforced Plastic Bars,” Proceeding of the Third
International Symposium, Non-Metalic (FRP) Reinforcement for Concrete Structures,
Sapporo, Japan, pp. 479-486.
El-Sayed, A., El-Salakawy, E., and Benmokrane, B., (2005) “Shear Strength of One-Way
Concrete Slabs Reinforced with Fiber-Reinforced Polymer Composite Bars,” Journal
of Composites for Construction, ASCE, Vol. 9, No. 2, pp. 147-157.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 227
Kessler, R.J., and Powers, R.G., (1988), ‘‘Corrosion o f Epoxy Coated Rebars-Keys
Segmental Bridge-Monroe County,” Florida Dept, of Transportation, Materials Office,
Corrosion Res. Lab., Gainesville, FI., Rep. N 0 .8 8 - 8 A.
MacGregor, J.C., and Bartlett, F.M., (2000), “Reinforced Concrete: Mechanics and
Design,” First Canadian edition, Prentice-Hall, Scarborough, Ont., Canada, 1042 p.
Nagasaka, T., Fukuyama, H., and Tanigaki, M., (1993), “‘Shear Performance o f Concrete
Beams Reinforced with FRP Stirrups,” American Concrete Institute, SP-138, (Nanni,
A., and Dolan, C., editors), Farmington Hills, Michigan, USA, pp. 789-811.
Nakamura, H., and Higai, T., (1995), ‘Evaluation o f Shear Strength o f Concrete Beams
Reinforced with FRP,” Concrete Library International, Proceedings of Japan Society
of Civil Engineers, No. 26, pp. 111-123.
Razaqpur, A.G., Isgor, B.O., Greenaway, S., and Selley, A., (2004), “Concrete
Contribution to the Shear Resistance o f Fiber-Reinforced Polymer Reinforced
Concrete Members,” Journal of Composites for Construction, ASCE, Vol. 8 , No. 5, pp
452-460.
Tottori, S., and Wakui, H., (1993), “Shear Capacity o f RC and PC Beams using FRP
Reinforcement,” American Concrete Institute, ACI SP-138, Nanni, A., and Dolan,
C.W., editors, Farmington Hills, Michigan, USA, pp. 615-631.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7 228
Tureyen, A.K., and Frosh, R.J., (2002), ‘‘Shear Tests o f FRP-Reinforced Concerete
Beams without Stirrups,” ACI Structural Journal, Vol. 99, No. 4, pp. 427-444.
Vijay, P., Kumar, S., and Gangarao, H., (1996), “Shear and Ductility Behaviour o f
Concrete Beams Reinforced with GFRP Bars,” Proceedings of the Second
International Conference on Advanced Composite Materials for Bridges and
Structures, (ACMBS-II), (El-Badry, M.), editor, Montreal, Quebec, pp. 217-226.
Yost, J.R., Gross, S.P., and Dinehart, D.W., (2001), “Shear Strength o f Normal-Strength
Concrete Beams Reinforced with Deformed GFRP Bars,” Journal of Composites for
Construction, ASCE, Vol. 5, No. 4, pp 268-275.
Zhao, W., Maruyama, K., and Suzuki, H., (1995), “Shear Behavior o f Concrete Beams
Reinforced by FRP Rods as Longitudinal and Shear Reinforcement,” Proceedings of
the Second International RILEM Symposium (FRPRCS-2), Non-Metallic (FRP)
Reinforcement for Concrete Structures, (Taerwe, L., editor), E & FN Spon, London,
pp. 352-359.
Zsutty, T.C., (1968), “Beam Shear Strength Predictions by Analysis o f Existing Data,”
ACI Structural Journal, Vol. 65, No. 11, pp. 943-951.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 8 229
CHAPTER 8
8.1- SUMMARY
The main objective o f the research reported in this thesis is to enhance the understanding
concrete, high-strength concrete, and FRP-reinforced concrete and to create novel tools
for predicting their properties. Multi layer perceptrons (MLP) artificial neural networks
were primarily used to model properties of such concrete owing to their proven ability to
discover unknown and hidden behaviour. Moreover, the genetic algorithms approach
along with analysis o f neural network model results were used to optimize and develop
materials/structures is a relatively recent development and the technique has been used by
only a limited number o f researchers. Therefore, to make the technique clear and
accessible, and to improve the understanding of the way MLP networks operate and
building network’s architecture, the parameters that affect the behaviour of the network
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 8 230
recommendations for optimizing the model performance were also presented in that
chapter through a case study. A brief description of the most commonly used methods in
parameters on its engineering properties are the main reasons of the limited acceptance of
segregation tendency o f coarse aggregate particle during mixing and placement remains
the least understood property o f SCC mixtures. Thus, an extensive experimental study to
develop a simple, yet reliable, test method to quantitatively measure segregation in highly
flowable concrete mixtures was conducted during the course of this research and
Although the demand for special concretes is increasing rapidly, current standards and
design codes are yet to provide special specifications for their design. For example (see
from that o f conventional concrete. For instance, the shear behaviour of HSC members is
of particular concern. Nevertheless, current shear design provisions either permit the use
of existing shear equations developed for normal strength concrete members or merely
provide simple recommendations and guidelines when HSC is used. To fully understand
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 8 231
the difference in shear behaviour between NSC and HSC slender beams, Two MLP
neural networks were developed and trained to predict the shear capacity of reinforced
concrete beams with and without shear reinforcement. Furthermore, a parametric study
was also carried out in order to investigate the effects of basic shear design parameters on
shear strength o f normal and high-strength concrete beams and to evaluate the capability
Likewise, current shear design provisions slightly modified existing shear design
equations developed for concrete beams reinforced with steel for use in the design of
FRP-reinforced concrete beams. Most of these equations are either highly conservative,
as in the case o f beams without shear reinforcement, or not accurate when used to
estimate the shear strength o f concrete beams with shear reinforcement. Again, MLP
networks were used to model the shear behaviour of FRP-reinforced concrete beams in
this research and results are presented in chapter 7. Moreover, based on analysis of MLP
equations for the design of FRP-reinforced concrete beams were also proposed in chapter
7.
8.2- CONCLUSIONS
The current study has shown that the multi-layer perceptron neural networks and genetic
algorithms approaches can be used as effective tools to model the behaviour and
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 8 232
Provided that sufficient and comprehensive experimental data is available, MLP neural
concretes. They were not only able to predict such properties instantly with acceptable
accuracy, but also demonstrated an ability to recognize and quantitatively evaluate the
outperformed all methods investigated in this thesis and currently used in the design and
To date, there is still no theoretical means to determine the appropriate architecture and
training process o f a particular MLP neural network, and these issues are usually defined
that longer training and minimizing the training error did not necessarily lead to a better
network. It was also observed that increasing the number of hidden layers and/or the
number o f hidden units improves the performance of MLP networks during training, but
not necessarily their generalization. Therefore, the best network architecture could only
be defined by trial and error (starting with a small number of hidden units in one hidden
layer) and the best way to minimize the number of trials and optimize the network
generalization is to use a validation set of data along with the training set to monitor the
and setting remains the least understood property of SCC mixtures. A simple segregation
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 8 233
test for flowable concrete was developed in this study and presented in chapter 3. The test
is based on quantifying the profile of coarse aggregate content along the height of a
cylindrical concrete sample. The test is easy to conduct and offers a numerical value to
segregation. The numerical values recommended by this test were based on establishing a
correlation between a penetration depth of semi-spherical heads into the concrete sample
A parametric study was also conducted to better understand the effect of basic mixture
ingredients on the potential of segregation and other properties of SCC mixtures. It was
concluded that the effect o f aggregate content on most properties of SCC mixtures was
minimal for the range o f aggregates content used in this investigation. The total content
o f cementitious materials (cm) on the other hand, showed a mixed effect on segregation.
Increasing the cm content caused an increase in segregation for SCC mixtures with high
w/cm ratio, while for SCC mixtures having relatively lower w/cm ratio, increasing the
content of cm slightly enhanced the ability of SCC mixtures to resist segregation. The
the w/cm ratio or the dosage o f HRWR. Such an effect was more pronounced for w/cm
>0.45 and for high dosage o f HRWR. Viscosity-modifying admixtures (VMA) proved to
be an effective tool for reducing segregation of SCC mixtures. However, there seems to
Analysis using MLP neural networks in predicting the shear behaviour of reinforced
high-strength concrete slender beams proved that current shear design provisions
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 8 234
slender beams without shear reinforcement, especially for concrete having strength
higher than 70 MPa. Unlike several current shear design methods, ANN analysis also
showed that the amount o f longitudinal tensile steel influences the shear capacity of such
beams in general, and that this influence is more pronounced for HSC beams. A similar
conclusion was reached for the effect of the effective depth of beams for both NSC and
HSC beams without shear reinforcement. However, at high compressive strength, such an
In the case o f HSC beams with shear reinforcement, ANN analysis showed that the shear
existing shear calculation methods, the ANN predicted that such an increase diminishes
that the shear capacity o f slender beams with shear reinforcement is more than 50 %
higher than values calculated using some current shear provisions and that simply
conservative results.
provisions included in this study are either very conservative in predicting the shear
capacity of FRP-reinforced concrete beams without stirrups or not adequate for the
design o f FRP-reinforced beams with stirrups. Results provided in chapter 7 show that
incorporating the axial rigidity of the longitudinal reinforcement and the capacity of FRP
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 8 235
stirrups in existing shear design equations developed for steel-reinforced concrete beams
could be satisfactory for the design of FRP-reinforced beams. However, adequate values
for those parameters should be carefully assessed. The effect of the axial rigidity of FRP
be better represented by a cubic root function rather than linear or square root
contribution o f FRP stirrups to the overall shear capacity can be represented more
accurately by a square root function rather than a linear function as proposed by ACI and
CSA recommendations.
The shear equations proposed in chapter 7 for the design of FRP-reinforced concrete
beams incorporated the axial rigidity of FRP longitudinal bars as a cubic root function
and assumed that the contribution of FRP stirrups to shear strength is a square root
function. They provided the most accurate yet conservative results in calculating both the
shear capacity o f beams with and without shear reinforcement amongst existing shear
actually measure their engineering properties and/or effective tools to predict their
test methods for such concretes, and create predictive models for their properties.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 8 236
algorithms are data driven modeling systems and their performance depends to a
great extent on the quality of the data used in developing the model. Therefore,
technique, curing conditions and durability aspect of flowable concrete are also
accounted for.
virtually does not exist and more testing is needed in this area to develop
design methods in estimating the shear capacity o f FRP- reinforced concrete deep
beams.
based tools such as artificial intelligence in the area o f concrete materials and
support systems in this area that can be used commercially by the industry.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
A ppendix A 237
APPENDIX A
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Appendix A
Table A.1 - Properties and compressive strength o f core samples (sample data) used in parametric study to investigate the
performance o f MLP networks (Bartlett, F. M., 1994)
Appendix A
Table A.1 - Continue
Cylinder Age of Duration of Age at Core
Curing Direction of Length Diameter
Beam strength cylinder curing Location testing strength
type drilling (mm) (mm)
(MPa) (days) (days) (days) (MPa)
21 52.7 56 2 8 1 2 128 100 50 42.1
22 52.7 56 1 8 1 2 128 100 50 60.1
23 52.7 56 2 8 1 2 128 100 50 41.7
24 52.7 56 3 8 1 2 128 96 50 57.7
25 52.7 56 2 19 2 1 125 89 50 39.6
26 52.7 56 2 19 2 1 125 101 50 32.2
27 52.7 56 2 19 2 1 125 102 50 26.9
28 52.7 56 2 19 2 1 125 95 50 39.9
29 52.7 56 2 19 2 2 125 99 50 38.8
30 52.7 56 2 19 2 2 125 100 50 38.2
31 52.7 56 2 19 2 2 125 100 50 39.4
32 52.7 56 2 19 2 2 125 98 50 34.9
33 52.7 56 2 19 2 2 125 99 50 37.9
34 52.7 56 2 19 2 2 125 99 50 36.1
35 52.7 56 2 19 2 2 125 101 50 34.6
36 52.7 56 2 19 2 2 125 97 50 33.1
37 52.7 56 2 19 2 1 125 100 50 35.5
38 52.7 56 2 19 2 1 125 100 50 34
39 52.7 56 2 19 2 1 125 98 50 37.6
40 52.7 56 2 19 2 1 125 99 50 35.5
to
'O
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Appendix A
Table A.1 - Continue
Cylinder Age of Duration of Age at Core
Curing Direction of Length Diameter
Beam strength cylinder curing Location testing strength
type drilling (mm) (mm)
(MPa) (days) (days) (days) (MPa)
41 54.1 38 2 11 1 2 125 99 50 55.8
42 54.1 38 1 11 1 2 125 100 50 74.3
43 54.1 38 3 11 1 2 125 103 50 63.5
44 54.1 38 2 11 1 2 125 97 50 56
45 54.1 38 1 11 1 2 125 99 50 71.6
46 54.1 38 3 11 1 2 125 101 50 58.8
47 54.1 38 2 11 1 2 125 98 50 56.7
48 54.1 38 1 11 1 2 125 100 50 71
49 54.1 38 3 11 1 2 125 93 50 57.2
50 54.1 38 2 11 1 2 125 95 50 52.2
51 54.1 38 1 11 1 2 125 102 50 68.2
52 54.1 38 3 11 1 2 125 91 50 46.5
53 54.1 38 3 11 1 2 125 92 50 64.8
54 54.1 38 2 11 1 2 125 91 50 54.7
55 54.1 38 1 11 1 2 125 86 50 70.6
56 54.1 38 2 11 1 2 125 98 50 56.1
57 54.1 38 1 11 1 2 125 91 50 66.4
58 54.1 38 3 11 1 2 125 99 50 63.2
59 54.1 38 1 11 1 2 125 101 50 69.1
60 54.1 38 2 11 1 2 125 98 50 56.2
* Curing type: 1 = air cured, 2 = water cured, and 3 = sealed cured. Location: 1 = south end of beam, 2 = north end o f beam. Direction o f drilling: 1 = horizontal,
2 = vertical.
A ppen dix B 241
APPENDIXB
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
A ppen dix B 242
cm3 0.0107
Effect of cm
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
A ppen dix B 243
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
A ppendix B 244
Table B.2 - Test results o f mixtures used in investigating properties of SCC mixtures
1
cm 5 580 3.5 2.04 1 0 .2 1 10.14 - -- -
© cm 6 605 3.0 3.40 13.21 14.69 17.5 44.3 49.3
o
is cm 7 605 4.0 5.04 13.97 13.47 19.8 43.8 48.0
H
cm 8 615 4.0 6 .1 0 16.35 17.13
Effect of cm (w/cm
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
A ppendix B 245
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
A ppen dix C 246
APPENDIX C
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
A ppendix C 247
Table C .l - Proportions and test results of UWC mixtures used in the database (sample
mixtures)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
A ppen dix C 248
57 261 497 55 0 0 822 700 4.1 0.15 380 4.1 55.1 46.8
58 261 497 55 0 0 822 700 4.6 0.15 440 4.9 54.7 46.3
59 261 497 55 0 0 822 700 5.1 0.15 500 6.3 55.2 45.8
60 261 405 33 110 0 838 698 4 0.07 410 7.9 49.3 35.6
61 261 405 33 110 0 838 698 5 0.07 460 9.7 49.1 35.4
62 261 405 33 110 0 838 698 5.9 0.07 500 13.9 49.4 32.1
63 261 405 33 110 0 838 698 7.0 0.07 610 30.0 49.7 28.6
64 258 278 0 0 278 827 706 0.7 0.07 400 9.8 40.8 21.7
65 258 278 0 0 278 827 706 1.2 0.07 420 14.4 40.3 20.1
1 W = water, C = cement, SF - silica fume, FA = fly ash, SL = slag, CA = coarse aggregates, S-F = slump-
flow, f c air and water = f c for concrete cast in air and water, respectively.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
A ppen dix C 249
Table C.2 - Distribution o f database between training and testing data sets
Table C.3 - Measured and predicted values (%) of SI-DYNAMIC for different w/cm
ratio, and HR WR and VMA dosages
w/cm 35.3 37.6 40.0 42.5 45 47.5 50.0 51.8 52.5 54.1 57.5
55.0 56.5
(%)
AN N 1.6 2.4 3.5 5.2 7.7 11.9 16.4 22.8 - 32.4 - 45.8 -
Exp. - - 3.2 4.9 8.3 11.8 10.7 - 31.9 - 38.4 - 41.9
HRWR .23 .23 .29 .33 .35 .45 .53 .56 .59 .60 .65 .71 .82
.76
(% cm)
AN N - 9.4 13.0 - 17.8 29.5 38.5 - 45.7 - 52.8 58.7 61.6 62.6
Exp. 8.3 - - 14.3 - 28.6 - 36.4 - 49.3 - - - -
VMA .0107 .0118 .0141 .0160 .0165 .0176 .0188 0.0200 0.0212 .0224 .0235 .0247
(%cm)
AN N 40.44 39.6 38.14 34.22 28.43 22.58 17.74 14.19 11.94 10.31
Exp. 36.41 42.63 29.20 15.82
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
A ppen dix C 250
Table C.4 - Input variables for UWC mixtures used to test the predictions o f the A N N model
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
A ppendix C 251
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
A ppendix D 252
APPENDIX D
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Appendix D
Table D .l - Properties and shear results o f reinforced NSC and HSC slender beams without web reinforcement (testing data)
Beams property 1
^test I ^calculated
Reference Ttest
d K Pi fc (kN) ACI CSA
a/d EC-2 R2K Zsutty ANN
(mm) (mm) (%) (MPa) 11-5 Simplified
Mphonde and Frantz 298 152 3.60 3.36 37.66 81.83 1.60 1.47 1.31 1.30 1.16 1.07
(1984) 298 152 2.50 3.36 20.58 77.45 1.80 1.88 1.51 1.47 1.19 1.11
Elzanaty et al. 270 178 4.00 2.50 20.68 54.40 1.37 1.24 0.98 1.38 1.02 0.94
(1986) 270 178 4.00 2.50 40.00 64.93 1.22 1.07 0.94 1.05 0.97 0.92
465 200 3.00 1.35 28.9 104.25 1.21 1.17 1.11 1.14 1.01 1.14
139 60 3.00 1.35 34.0 12.10 1.45 1.24 1.13 1.09 1.23 1.14
359 229 3.50 4.32 26.7 126.80 1.50 1.56 1.30 1.19 1.02 1.09
359 229 3.50 4.32 25.9 111.20 1.33 1.39 1.15 1.07 0.90 0.97
Taylor (1972)
359 229 4.25 4.32 36.3 129.00 1.39 1.36 1.20 1.12 0.99 0.99
359 229 4.25 4.32 44.9 142.30 1.40 1.35 1.23 1.15 1.02 1.05
359 305 3.50 3.24 44.5 184.60 1.39 1.32 1.20 1.19 1.03 1.05
447 229 2.81 3.46 26.4 153.50 1.46 1.62 1.33 1.01 0.99 0.98
Ghannoum et al. (1998) 440 400 2.50 1.20 34.2 187.50 1.06 1.01 1.02 1.29 0.88 0.99
Podgomiak - Stanik 450 300 3.00 1.11 37.0 162.90 1.18 1.11 1.17 1.26 1.06 1.01
(1998) 925 300 2.92 0.76 43.0 237.20 0.79 0.97 1.01 1.08 0.81 1.05
895 300 3.01 2.09 32.0 257.00 0.94 1.23 0.90 1.09 0.72 0.82
Angelakos (2001)
925 300 2.92 1.01 38.0 180.00 0.63 0.78 0.73 0.85 0.58 0.76
274 153 3.47 0.51 27.6 25.35 0.71 0.58 0.75 1.06 0.80 0.80
Kani (1979) 272 152 2.50 0.78 18.0 43.66 1.46 1.24 1.31 1.31 1.25 1.17
272 154 2.49 0.76 27.0 43.30 1.18 0.99 1.13 1.14 1.08 1.10
NJ
Ui
U>
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Appendix D
Table D .l - Continue
254
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Appendix D
Table D .l - Continue
to
C /l
c/1
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Appendix D
Table D.2 - Properties o f beams used to investigate the effect o f compressive strength of concrete (f*c ) on shear strength ( Vu/b wd )
NJ
Os
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Appendix D
Table D.3 - Properties o f beams used to investigate the effect of tensile steel ratio {pi) on shear strength (Vu/b wd )
257
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Appendix D
Table D.4 - Properties o f beams used to investigate the effect of beam’s effective depth {d) on shear strength ( Vu/b wd )
to
00
A ppendix D
cn 00 CO VO cs ON 00 00
o VO r o vo Tt On r^
f On o O n O n 00 00 00 00 t***
N (N
CM VO vo n in in vo VO VO VO VO
1 in in in in in vo VO VO vo VO
o r-* r^ r-»
UJ
© in o <N 00 VO ro o
*4 o vo ro Ti ro On
vo m ro ro ro CN CN
a 1-H
. §*
6 n vo in in VO VO vo vo VO vo
"w 00 00 00 00 00 00 00 00 00 00
Table D.5 - Properties of beams used to investigate the effect of (aId) on shear strength ( Vu/b wd )
§ , f- r- r- r- r- r- r-
*
<75
U
■c
ixa VO
<N vo n 00 VO CN o VO CN 00
o <N o O n O n Os 00 00
I—i VO vo 3 VO VO vo VO VO vo vo
U
<
co r- Os O i i VO
r-H o 00 r-* vo Tf 00
VO 00 r- VO vo VO vo VO vo Tf
1""‘ y~^ 1"“* 1—♦
P
1 .6 4
%
CO
c3
0)
s
(M P a)
r-'
87
87
87
87
87
87
87
87
87
< 00
Tt ■^r ^r Tf ^ r Tt
CM CM CM CM (N CN CM CN CM (N
CM CM CM CM <N CM CM (N CM CN
■'t m N- CO VO N- o>
a /d
VO £ 00 O CM CO lO CO
CM CM CM CO CO CO CO CO CO CO
(m m )
200
200
200
200
200
200
200
200
200
200
*
(m m )
Ov
359
359
359
359
359
359
359
359
359
VO
CO
Beam
Q cm
o
o
cs
to
C3
W)
o s u in a q p a jK J a u a j)
CQ
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
A ppendix E 260
APPENDIXE
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Table E .l - Properties and shear results of reinforced NSC and HSC slender beams (testing data)
Appendix E
Beams property Vtest Vtest / Vcalculated
Reference d bw Pi fc P /w ACI CSA Bazant-
a/d (kN) Zsutty Mphonde ANN
(mm) (mm) (%) (MPa) (MPa) 11-5 Simplified Kim
Mphonde and Frantz 298.0 152.0 3.6 3.36 37.66 0.000 81.83 1.60 1.47 1.16 1.06 1.08 1.14
(1984) 298.0 152.0 2.5 3.36 20.58 0.000 77.45 1.80 1.88 1.19 1.14 0.89 1.08
Elzanaty et al. 270.0 178.0 4.0 2.50 20.68 0.000 54.40 1.37 1.24 1.02 0.75 1.01 1.10
(1986) 270.0 178.0 4.0 2.50 40.00 0.000 64.93 1.22 1.07 0.97 0.78 0.92 1.01
465.0 200.0 3.0 1.35 28.90 0.000 104.25 1.21 1.17 1.01 0.70 1.10 1.07
139.0 60.0 3.0 1.35 34.00 0.000 12.10 1.45 1.24 1.23 0.87 1.08 0.78
359.0 229.0 3.5 4.32 26.70 0.000 126.80 1.50 1.56 1.02 0.97 0.94 1.06
359.0 229.0 3 5 4.37 25.90 0.000 111.20 1.33 1.39 0.90 0.86 0.84 0.93
Taylor (1972)
359.0 229.0 4.2 4.32 36.30 0.000 129.00 1.39 1.36 0.99 0.93 0.96 1.10
359.0 229.0 4.2 4.32 44.90 0.000 142.30 1.40 1.35 1.02 0.97 0.97 1.09
359.1 305.0 3.5 3.24 44.50 0.000 184.60 1.39 1.32 1.03 0.95 0.98 1.54
447.3 229.0 2.8 3.46 26.40 0.000 153.50 1.46 1.62 0.99 0.95 0.89 1.04
Ghannoum et al. (1998) 440.0 400.0 2.5 1.20 34.20 0.000 187.50 1.06 1.01 0.88 0.64 0.90 1.01
Podgomiak - Stanik 450.0 300.0 3.0 1.11 37.00 0.000 162.90 1.18 1.11 1.06 0.71 1.17 1.37
(1998) 925.0 300.0 2.9 0.76 43.00 0.000 237.20 0.79 0.97 0.81 0.49 1.12 1.13
895.0 300.0 3.0 2.09 32.00 0.000 257.00 0.94 1.23 0.72 0.58 0.89 1.04
Angelakos (2001)
925.0 300.0 2.9 1.01 38.00 0.000 180.00 0.63 0.78 0.58 0.38 0.79 0.82
274.0 153.0 3.5 0.51 27.58 0.000 25.35 0.71 0.58 0.80 0.38 0.88 0.63
Kani et al. (1979) 272.0 152.0 2.5 0.78 18.00 0.000 43.66 1.46 1.24 1.25 0.72 1.24 1.07
272.0 154.0 2.5 0.76 27.00 0.000 43.30 1.18 0.99 1.08 0.65 1.07 0.92
KJ
Os
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Appendix E
Table E .l - Continue
262
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Appendix E
Table E .l - Continue
Appendix E
Table E .l - Continue
264
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Appendix E
Table E .l - Continue
Appendix E
Table E .l - Continue
266
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Appendix E
Table E.2 - Properties o f beams used to investigate the effect of stirrups capacity (P\fyv) on shear strength (NSC)
267
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Appendix E
Table E.3 - Properties o f beams used to investigate the effect of stirrups capacity (p jyv) on shear strength (HSC)
Bazant-
ID (mm) (mm) (%) (MPa) (MPa) (MPa) ANN A C I11-5 CSA simp. Mphonde Zsutty
Kim
Kong & Rangan 3.13 3.76 3.12
294.0 250.0 3.3 4.47 74.8 0.89 3.35 3.44 2.45 2.62
(1998)
1 294.0 250.0 3.3 4.47 74.8 0.00 2.09 1.55 1.73 1.70 2.62 2.22
2 294.0 250.0 3.3 4.47 74.8 0.13 2.29 1.68 1.86 1.91 2.78 2.35
3 294.0 250.0 3.3 4.47 74.8 0.25 2.49 1.81 1.98 2.11 2.94 2.48
5OS
V 4 294.0 250.0 3.3 4.47 74.8 0.38 2.69 1.93 2.11 2.31 3.10 2.60
JO
T3 5 294.0 250.0 3.3 4.47 74.8 0.51 2.88 2.06 2.24 2.51 3.27 2.73
■8 3.07 2.19 2.36 2.72 3.43 2.85
24> 6 294.0 250.0 3.3 4.47 74.8 0.63
a!U 7 294.0 250.0 3.3 4.47 74.8 0.76 3.25 2.31 2.49 2.92 3.59 2.98
O 8 294.0 250.0 3.3 4.47 74.8 1.01 3.59 2.57 2.74 3.32 3.91 3.23
9 294.0 250.0 3.3 4.47 74.8 1.14 3.73 2.69 2.87 3.52 4.07 3.36
10 294.0 250.0 3.30 4.47 74.8 1.39 3.88 2.94 3.12 3.93 4.39 3.61
294 250 3.30 4.47 74.8 0.597 2.96
Experimental beams 294 250 3.30 4.47 74.8 0.717 2.79
Kong Rangan (1998) 294 250 3.30 4.47 74.8 1.115 3.72
294 250 3.30 4.47 74.8 1.275 4.14
to
o\
00
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Appendix E
Table E.4 - Properties o f beams used to investigate the effect o f f c on shear strength
269
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Appendix E
Table E.5 - Properties o f beams used to investigate influence of stirrups on the effect of f c on shear strength
270
A ppen dix E
<N ON O ^ o os r*** ro
<N « ° o on © © ^ cn co co
ro (S N N ro ro ro ro ro ro ro
CN
Q\
ro
Table E.6 - Properties of beams used to investigate the effect of tensile steel ratio (pi) on shear strength
CN
co
ON
00
ON
00
cT
oo
CN
00
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
A ppen dix F 272
APPENDIX F
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Appendix F
T a b le F . l - Properties and shear results o f FRP-reinforced concrete beams
Razaqpur et al.
225.0 200.0 2.7 40.5 1.276 0.000 42.7 1.94 1.01 1.11 0.88 0.95 —
(2004)
225.0 200.0 1.8 40.5 0.725 0.000 96.2 7.76 2.42 3.01 1.97 1.66 —
225.0 200.0 3.6 40.5 0.725 0.000 47.0 3.88 1.48 1.47 0.96 1.39 —
225.0 200.0 4.5 40.5 0.725 0.000 38.5 3.19 1.31 1.20 0.79 1.23 —
225.0 229.0 4.1 36.3 0.528 0.000 38.1 3.59 1.23 1.20 1.37 1.19 --
225.0 178.0 4.1 36.3 0.685 0.000 31.5 2.90 1.23 1.17 1.45 1.16 -
Yost et al. 225.0 229.0 4.1 36.3 0.794 0.000 44.4 2.71 1.28 1.22 1.59 1.21 —
(2001) 225.0 279.0 4.1 36.3 0.866 0.000 45.3 2.06 1.04 0.99 1.33 0.98 --
225.0 254.0 4.1 36.3 0.979 0.000 45.1 1.97 1.09 1.04 1.46 1.03 —
225.0 229.0 4.1 36.3 1.064 0.000 42.2 1.87 1.10 1.05 1.51 1.04 -
279.0 178.0 2.7 24.1 0.920 0.000 106.8 4.67 3.04 3.50 4.90 2.87 —
Appendix F
Table F .l - Continue
21A
325.0 200.0 2.2 46.9 0.534 0.000 152.0 11.45 3.48 3.81 3.23 2.39 —
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Appendix F
Table F .l - Continue
Tottori and 325.0 200.0 3.2 46.9 0.534 0.000 61.8 4.76 1.42 1.55 1.31 1.29 —
Wakui (1993) 325.0 200.0 4.3 46.9 0.534 0.000 47.1 3.66 1.08 1.18 1.00 1.08 —
165.0 1000.0 6.1 40.0 0.445 0.000 140.0 5.20 1.34 1.37 0.89 1.59 —
165.0 1000.0 6.1 40.0 0.889 0.000 167.0 3.07 1.60 1.29 1.06 1.51 —
159.0 1000.0 6.3 40.0 1.345 0.000 190.0 2.37 1.67 1.33 1.25 1.57 —
El-Sayed et aL 162.0 1000.0 6.2 40.0 0.344 0.000 113.0 5.45 1.10 1.22 1.23 1.44 —
(2005) 159.0 1000.0 6.3 40.0 0.680 0.000 142.0 3.46 1.41 1.25 1.58 1.48 —
162.0 1000.0 6.2 40.0 0.684 0.000 163.0 3.87 1.59 1.40 1.78 1.65 —
159.0 1000.0 6.3 40.0 0.976 0.000 163.0 2.71 1.59 1.27 1.81 1.50 —
155.0 1000.0 6.5 40.0 1.052 0.000 168.0 2.65 1.66 1.31 1.92 1.56 —
264.5 150.0 1.9 44.8 0.772 2.306 126.8 1.38 0.95 2.59 1.02 1.26 0.98
Vijay et al. 264.5 150.0 1.9 44.8 0.772 1.538 115.0 1.76 1.14 2.55 1.25 1.25 1.02
(1996) 264.5 150.0 1.9 31.0 0.346 2.306 123.2 1.44 1.01 4.02 1.03 1.46 0.95
264.5 150.0 1.9 31.0 0.346 1.538 123.3 2.10 1.38 4.31 1.41 1.64 1.09
309.5 200.0 3.2 35.5 0.493 0.475 68.5 2.02 1.02 1.57 1.13 0.92 1.10
Alsayed et al. 0.76 0.80
309.5 200.0 3.2 39.5 0.493 0.475 57.9 1.72 0.84 1.27 0.93
(1996)
309.0 200.0 2.4 35.7 0.559 0.904 108.9 1.96 1.09 2.19 1.21 1.13 1.00
250.0 150.0 3.0 34.3 3.182 1.848 113.0 1.46 1.00 1.89 1.12 1.30 1.04
Zhao et al. 125.9 1.09 1.00 1.77 1.11 1.39 0.99
250.0 150.0 3.0 34.3 3.182 2.184
(1995)
250.0 150.0 3.0 34.3 2.384 1.848 116.2 1.86 1.07 2.18 1.15 1.40 1.08
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Appendix F
Table F .l - Continue
Appendix F
Table F .l - Continue
so © © © © \o » n T f o > - i ' 5 f O m o o \ o t ^ - v o t ~ o \
< N © O s © © ^ 0 0 © r - i © . - * © © © o q © © © 0 0 ©
. - < © © © © i-<. © © • - < © ■ ■c*i - i i - i i —
i© i—
ii—
i-*©©
^ a
55 o in© it n N © © ^ © ■ s t © © o i s N ‘ < N < N N - © © r - -
oa. -a
a
© --< d s c o © cn •— — n ( S N O \ » o o « i » H o o
O a -H -1 —1 © ©
£ a
CW S J
© <N so so ds © <n v i H i » © ^ N O \ i t h a h a ©
HH O
CO o
© i—i i—1 i—1 d s © <N^f©om©ooN^J' r-- © t~~ ©
1 t-l CN
J
X H P fO H io^'O 'noo^H in
<J © r^oo©iocsv©ooo© i n or*.v rc n^ o© o cc os o o co ns vo oo ©
£S 2 h (S N M —I dS <N <N CN <N
< s oo
©
©
CN
t'-
'
so
ch ' T>-it Od Sf NO nT <N
■— j - ^© H r—
fr
t f i - i f©N-CN
— i © ti-
©
<-•
oo
so ©
oo
•**
oo
CO
U00
1
co
H- S in os c s Tf r ~ © © ' O S f r H S f O ' t / l M
VO © ©r"; ds
n;
©
© © •-< ds © < N © © - ^ - r ^ O s < N O S ©
< N < N d s < N < N d v < N d s c i s c N c N C N CN © <N dS CN ds
^r
© © N; ^1; dS © N; itf; © N- ©_ is w n o
©
ds
© co oo CN
m ~© so cn
© © o© X-
N Tf
© © c»s
Tf 00
oo o\ h- oo © oo r- oo *> ©
oo
S sm N © t tt s Mh Nw Nm t n So h\ ©H Nt ©n H
© M o o o o ( C i O M o o
H O r t o o n © ^ ^
©©■*t©^H©©©<N©C'»©©cnOS©0 0 0 © 0 0
o d d d ^ d d o d o i - i d o d d r i H ^ N d
o o o o c n c i s c i s r c i c o r o r i s i —i ^ - i o o o o © T t r ^ r ' t ^ - r ^ r ^
«uO O©l ©
Ol © © © ' © . © © © ©
© © O O © © © © © ©
© . » W ^
O
O
,-1
l N
-H
M N N N
09
.Si
t,4> t-; ■a; -ct ^ ^ "^r 00 N" © © fH (S © (S © © CN CN ©
O. ^
o© n© ©tn m
© Os c ~ - © © r ~ ^ j - ^ - i —
n© ci © c i n w n m n si cn t f' 0N0 if sf N
OO 0 0 0 0
2
a
09
e C N C N C N C N C N C N C N C N C N © © © © © © © © © ©
4a> 1 ( l i m r n m M ' m m c n m N N N N N CN CIS d S CIS CIS
n
© © © © © © © © O O © © © © © © © © © ©
*s 0 00 0© 00 00 0© 0 0 0 0 ©0 ©© ©© ©0 ©0 ©
0 0© 00 ©
0 00 ©
0 00
s ©
N N N M N IN N N N h h h h c i i h N N N P I C I
Table F.l - Continue
© © © © © © p p p p © p p p p p © o p p
© © © © © . © ' © © © O O O O O O O O O O O
< N < N C N < N C N <N C N C N C N © © © © © S O © © © © ©
ds d S CIS CIS CIS CCS CCS CIS CIS <N CN < N < N © C N < N < N < N < N C N
U
W SjCS
a
2 •c
© Ia a
©
.a
H
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Appendix F
Table F .l - Continue
Appendix F
Table F .l - Continue
N)
oOO
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Appendix F
Table F.2 - Properties o f beams used to investigate the effect o f FRP stirrups on shear strength o f concrete beams
Table F.3 - Properties o f beams used to investigate the effect of ptEi on shear strength o f concrete beams
Appendix F
Table F.4 - Properties o f beams used to investigate the effect of f c on shear strength of concrete beams
K)
00
N)
283
VITA
Northeastern University
Boston, Massachusetts, USA
1990 - 1992, M.Sc. (Civil Engineering/Structure)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
284
Structural Engineer
Tripoli, Lebanon
1993 -1 9 9 9
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.