Birhane Gemeda, Shear Capacity Prediction

ADDIS ABABA UNIVERSITY
ADDIS ABABA INSTITUTE OF TECHNOLOGY

SCHOOL OF CIVIL AND ENVIRONMENTAL
ENGINEERING
APPLICATION OF MACHINE LEARNING

METHODES FOR SHEAR CAPACITY OF
RC BEAMS
A Thesis in Structural Engineering
By Birhane Gemeda Wakjira

Addis Ababa
December 2021
A Thesis
Submitted in Partial Fulfillment of the Requirements for the Degree of Master of Science
i
ii
UNDERTAKING
I certify that research work titled “Application of Machine learning methods for shear
capacity of RC beams” is my own work. The work has not been presented elsewhere for
assessment. Where material has been used from other sources it has been properly
acknowledged / referred.
Signature of Student
Name of Student: Birhane Gemeda Wakjira
iii
ABSTRACT
Accurate determination of the capacity of reinforced concrete (RC) beams in shear remains
a demanding problem due to its complex failure mechanism and the nonlinear relationship
between different factors influencing the shear capacity. This research employs different
types of single and ensemble machine learning (ML) based techniques; namely, decision
tree, support vector machine, extremely randomized trees, gradient boosting, random
forest, and extreme gradient boosting (xgBoost) to correctly predict the shear capacity of
reinforced concrete beams. To this end, a dataset of experimental test results of RC beam
with and without stirrups comprised of various beam geometry, concrete strength,
reinforcing steel strength, longitudinal and shear reinforcement ratios, and shear span-to-
effective depth ratio is used to develop the models.
The proposed models were calibrated for different values of hyperparameters to achieve
optimized ML models. The results of the analysis evidenced that the xgBoost model can
be effectively utilized to predict the shear capacity of RC beams. The comparison of the
predictions of the proposed and existing models evidenced that the efficiency of the
proposed model is superior to the existing models and guidelines in terms of accuracy,
safety, and economic aspects with significantly lowest bias and variability.
A solid correlation exists between the shear capacities predicted using the proposed model
and the corresponding experimental values as evidenced by the value of 𝑅 2 (𝑅 2 = 0.99)
for RC beams without stirrup and (𝑅 2 = 0.995) for RC beams with stirrup.
The proposed xgBoost model is deployed into a user-friendly web-based application to

facilitate a quick and accurate prediction of capacity of RC beams in shear. The web-based
application can be used by both practitioners and researchers to accurately predict the shear
capacity of beams.
Keywords
Shear Capacity Prediction; RC Beams; Machine Learning
iv
ACKNOWLEDGMENTS
First and foremost, praises and thanks to the Almighty God, for His blessings throughout
my research work to complete the research successfully.
I would like to express my deep sincere gratitude to my thesis advisor Dr. Bedilu Habte
for his guidance and support
I am extremely grateful to my brother Dr. Tadesse Gemeda for his support and guidance.
He has taught me the methodology to carry out the research and to present the research
works as clearly as possible.
I am very much thankful to my mother for her love, prayers, caring and sacrifices for
educating me. Also I express my thanks to my husband for his support to complete this
research work.
v
TABLE OF CONTENTS
ABSTRACT.................................................................................................................... IV
ACKNOWLEDGMENTS ............................................................................................... V
TABLE OF CONTENTS .............................................................................................. VI
LIST OF TABLES ......................................................................................................... IX
LIST OF FIGURES ......................................................................................................... X
CHAPTER 1 INTRODUCTION ............................................................................... 1
1.1 Background .......................................................................................................... 1

1.2 Research objectives .............................................................................................. 5
1.2.1 General objective .......................................................................................... 6
1.2.2 Specific objective.......................................................................................... 6
1.3 Scope .................................................................................................................... 7

1.4 Research Significance .......................................................................................... 7
1.5 Layout of work ..................................................................................................... 8
1.6 Methodology ........................................................................................................ 9
CHAPTER 2 LITERATURE REVIEW ................................................................. 10
2.1 Shear strength of RC beam ................................................................................ 10

2.2 Parameters influencing shear strength of RC beam ........................................... 12
2.2.1 Concrete compressive strength ................................................................... 12
2.2.2 Transverse shear reinforcement .................................................................. 13
2.2.3 Depth........................................................................................................... 13
2.2.4 Longitudinal reinforcement ratio ................................................................ 13
2.2.5 Shear span-to-depth ratio ............................................................................ 14
2.3 Existing Shear Design Model and Guideline ..................................................... 14

2.3.1 ACI 318 [47] ............................................................................................... 17
2.3.2 Eurocode (EC2) [4]..................................................................................... 18
2.3.3 JSCE (2007) [57] ........................................................................................ 19
2.3.4 CSA A23.3-94 (Simplified Method) [58] ................................................... 20
2.3.5 Modified compression field theory............................................................. 21
vi
CHAPTER 3 EXPERIMENTAL DATABASE...................................................... 24
3.1 Experimental database for RC beams without stirrups ...................................... 24

3.1.1 Introduction................................................................................................. 24
3.1.2 Distribution of input parameters ................................................................. 28
3.1.3 Effect of input parameters on the shear capacity ........................................ 32
3.2 Experimental database for RC beams with stirrups ........................................... 36

3.2.1 Introduction................................................................................................. 36
3.2.2 Distribution of input parameters ................................................................. 37
3.2.3 Effect of input parameters .......................................................................... 41
CHAPTER 4 MACHINE LEARNING-BASED MODELS .................................. 45
4.1 Introduction to Machine Learning ..................................................................... 45

4.2 Normalization of dataset .................................................................................... 48
4.3 Performance measurement ................................................................................. 48
4.4 Hyperparameter optimization ............................................................................ 49
4.5 Single ML models .............................................................................................. 50
4.5.1 Support vector machine .............................................................................. 50
4.5.2 Decision tree ............................................................................................... 51
4.6 Ensemble Models ............................................................................................... 53

4.6.1 Random forest ............................................................................................. 53
4.6.2 Extremely Randomized Trees ..................................................................... 55
4.6.3 Gradient boosting........................................................................................ 55
4.6.4 Extreme gradient boosting .......................................................................... 56
CHAPTER 5 RESULTS AND DISCUSSION........................................................ 58
5.1 RC beams without stirrup .................................................................................. 58

5.1.1 Comparison with existing design guide line for RC beams without stirrup
67
5.2 RC beam with stirrup ......................................................................................... 70

5.2.1 Comparison with existing design guide line for RC beams with stirrup .... 78
5.3 Deployment of the ML model to a web-based application ................................ 83
vii
CHAPTER 6 CONCLUSIONS AND RECOMMENDATIONS .......................... 85
6.1 Conclusions ........................................................................................................ 85

REFERENCES ............................................................................................................... 87
viii
LIST OF TABLES
Table 2.1: Shear capacity of RC beams according to different models ............................ 15

Table 3.1: Geometry and material characteristics of the beams included in the database 31
Table 3.2: Geometry and material characteristics of the beams included in the database 38
Table 5.1: Optimized hyperparameters for each model for RC beams without stirrup .... 58
Table 5.2 : Performance indices for RC beams without stirrup ....................................... 59
Table 5.3: Descriptive statics for 𝑽𝒑𝒓𝒆𝒅/𝑽𝒆𝒙𝒑 for RC beams without stirrup.............. 67
Table 5.4: Performance indices for xgBoost, ACI-318 and EC-2 .................................... 69
Table 5.5: Optimized hyperparameters of each models for RC beams with stirrup......... 72
Table 5.6: Performance indices for RC beams with stirrup.............................................. 74
Table 5.7: Descriptive statics for 𝑽𝒑𝒓𝒆𝒅/𝑽𝒆𝒙𝒑 for RC beams with stirrup ................... 79
Table 5.8 : Performance indices for xgBoost, ACI-318 and EC-2 ................................... 81
ix
LIST OF FIGURES
Figure 2.1: Internal forces along diagonal shear crack [43]. ............................................ 10
Figure 2.2: Components of shear resistance at cracked concreate section for RC beams
without stirrup [46]. .......................................................................................................... 12
Figure 2.3: Concrete and stirrups shear resisting mechanisms (a) and graphical illustration
of their contribution (b) [59]. ............................................................................................ 18
Figure 2.4: MCFT equations [46] ..................................................................................... 22
Figure 3.1: Geometry of RC beams in the database. ....................................................... 26
Figure 3.2: Schematic of the loading types ...................................................................... 26
Figure 3.3: Distribution of RC beams in terms of the loading type. ................................ 27
Figure 3.4: Distribution of the input variables in the database. ........................................ 31
Figure 3.5: Relationship between the input parameters and the shear capacity. .............. 35
Figure 3.6: Beam type (slender beams/deep beams) ........................................................ 37
Figure 3.7: Distribution of the input variables in the database. ........................................ 41
Figure 3.8: Relationship between the input parameters and the shear capacity of RC beams
with stirrups ...................................................................................................................... 44
Figure 4.1: 10-fold cross-validation. ................................................................................ 49
Figure 4.2: model development ........................................................................................ 50
Figure 4.3: Decision tree flow chart. ................................................................................ 52
Figure 4.4: Conceptual schematic of random forest. ........................................................ 54
Figure 4.5: Conceptual schematic of gradient boosting. .................................................. 56
Figure 5.1: Comparisons of shear capacity predictions provided by the single ML models
to the experimental shear capacity for RC beams without stirrup. ................................... 61
Figure 5.2: Comparisons of shear capacity predictions provided by the ensemble ML
models to the experimental shear capacity for RC beams without stirrup. ...................... 63
Figure 5.3: Residual plots for RC beams without stirrup. ................................................ 66
Figure 5.4: Predictions of ACI and proposed model for RC beams without stirrup. ....... 68
Figure 5.5: Histogram of predicted to experimental shear capacity ratio for RC beams
without stirrup. .................................................................................................................. 69
Figure 5.6 : Experimental versus predicted shear capacities for RC beams without stirrup
based on the proposed xgBoost model and existing code equations. ............................... 70
Figure 5.7: Properties of dataset for RC beams with stirrups. .......................................... 73
x
Figure 5.8: Comparisons of shear capacity predictions provided by single ML models to
the experimental shear capacity for RC beam with stirrup............................................... 75
Figure 5.9: Comparisons of shear capacity predictions provided by ensemble ML models
to the experimental shear capacity for RC beam with stirrup. ......................................... 77
Figure 5.10: Predictions of ACI, Eurocode and proposed model for RC beams with stirrup.
.......................................................................................................................................... 80
Figure 5.11: Histogram of predicted to experimental shear capacity for RC beams with
stirrup ................................................................................................................................ 81
Figure 5.12 : Experimental versus predicted shear capacities for RC beams with stirrup
based on the proposed xgBoost model and existing code equations. ............................... 82
xi
Application of Machine learning methods for shear capacity of RC beams
CHAPTER 1 INTRODUCTION
1.1 Background
Shear failure is sudden and often catastrophic as it occurs without warning. On the
contrary, flexural failure involves the yielding of reinforcement bars which is ductile as it
provides enough warning. Thus, reinforced concrete (RC) beams should be designed with
adequate shear capacity in order to avoid such catastrophic and sudden shear failure and
assure ductile flexural failure. It is crucial to understand the shear behavior of RC beams
for a safe, economical, accurate design of the beams. Despite the various research efforts,
fully assessing the shear behavior of RC beams remains a challenging task when compared
to assessing its flexural behavior that can be predicted with an acceptable accuracy [1,2].
This challenge lies in a complicated failure mechanism associated with shear-critical RC
beams. Most design codes and guidelines including ACI 318 [3] and Eurocode 2 [4] for
the shear strength of RC elements are empirically derived using a limited number of
important factors. Hence, most of the available models results in unsafe and/or
uneconomical design. As a result, a more rigorous prediction model is required for better
understanding and safe design of reinforced concrete beams.
Machine learning (ML) has recently gained immense attention owing to its ability to
effectively determine the relationship between the input features and the response
(dependent) variable (s) in a complex system. It is considered as a powerful technique to
solve different civil engineering problems [5–15]. This is attributed to their ability to
estimate the relationship between the factors and the response parameter (s) [16]. In
contrast to the case of most empirical models, ML techniques does not require the prior
MSc Thesis Page 1

assumption or knowledge of the underlying mathematical and physical models [16]. Some
of the applications of ML techniques reported in the literature include the prediction of
mechanical properties of concrete [6,8–10], load capacity and failure modes of RC
columns and walls [11,17,18], and seismic damage assessment of RC buildings and
bridges [19–21]. Artificial neural network (ANN) is among the most popular ML
techniques [9]. It has been successfully employed for different structural engineering
studies; some of which include the prediction of mechanical properties of concrete [6,8–
10], damage assessment of bridges [22–24] and buildings [25], ultimate deformation
capacity of RC columns [11], shear strength of circular RC columns [26], and compressive
strength [13–15] and the stress-strain relationship [12] of concrete confined with FRP. A
review of the application of different artificial intelligent techniques in the field of civil
engineering is conducted by [5]. Support vector machine (SVM) is another powerful ML
technique that has been applied to structural engineering problems [27–29]. Solhmirzaei
et al. [29] used three different ML techniques, namely, SVM, k-nearest neighbor (kN), and
ANN to classify the failure modes of Ultra High Performance Concrete (UHPC) beams.
The results showed that ANN outperformed the other two methods in predicting the failure
modes of UHPC beams with an overall accuracy of 89%. In this study, several ML
techniques have been explored to propose the best predictive model for shear capacity of
beams. Ensemble models are another type of emerging ML technique with high
performance compared to other techniques. These algorithms combine two or more
learners known as base learners (e.g. kernel regression, support vector machine, and
decision tree) to enhance the efficiency, robustness, and stability of the base learners [30].
The basic idea behind the ensemble model is to combine multiple base learners in
computing the final response rather than relying on an individual model. Ensemble
MSc Thesis Page 2
learners can be formed in sequential or parallel styles with the aim to exploit the
dependence and independence between the base models, respectively. Bagging (short for
Bootstrap Aggregating) [31] is one of the ensemble models in which multiple base learners
are independently trained in parallel using different bootstrap sample. Each bootstrap
training dataset contains an average of 63.2% of the original training set. The final
prediction is then taken as the mean of the predictions from the base learners [31]. As a
result a better prediction with reduced variance is obtained [31]. Random forest (RF) is a
popular example of bagging ensemble learners. In RF homogeneous base learners,
particularly, decision trees are used. An ensemble model can also be formed from multiple
base learners of different types or heterogeneous base learners [30]. Stacking ensemble is
one of the heterogeneous ensemble models in which different types of base learners are
combined via a meta-model with the objective of enhancing the accuracy of the individual
model [32,33].
In the previous study, different researchers have attempted to examine the efficacy of ML
models to estimate the capacity of RC beams in shear [34–39]. Sanad and Saka [34] studied
the application of ANN in predicting the capacity of deep RC beams in shear using the test
results of only 111 RC beams considering 10 input variables. The results of the ANN-
based model were compared against that of the ACI equation. The authors concluded that
the shear capacity predictions from the ACI code equations are very conservative and thus
uneconomical. The proposed ANN model outperformed the ACI equation in predicting
the capacity of the beams in shear. However, it is worth mentioning here that a small
number of datasets used to train ML algorithms will lead to overfitting and a false sense
of high predictive performance. As per Friedman et al. [40], a sufficient training and testing
MSc Thesis Page 3

dataset should be at least 10 times the number of independent parameters. Chou et al. used
support vector regression (SVR) for the prediction of the shear strength of RC deep beams
based on a larger experimental data of 214 RC deep beams.
Cladera and Mari [35] investigated the application of ANN to predict the shear capacity
of RC beams without stirrups based on the dataset of 177 experiments on RC beams. Five
parameters; namely, effective beam depth, longitudinal reinforcement ratio, width of the
beam, concrete compressive strength, and aspect ratio were considered as input
parameters. The authors concluded that ANN can be applied to predict the shear capacity
of RC beams without stirrups. In another study, the same authors applied ANN to predict
the shear capacity of RC beams reinforced with stirrups based on the test results of 123
RC beams with stirrups.
Chou et al. [38] used different ML algorithms to estimate the shear strength of RC beams
using the experimental dataset of RC beams compiled by Collins et al. [41] and Zhang et
al. [42]. For RC beams reinforced without stirrups, the dataset comprises 1849 RC beam
test results. However, the database used for RC beams with stirrups comprised only 194
test results of RC beams. In this study, different single models including multi-layer
perceptron ANN, decision tree, support vector regression, linear regression, ensemble
methods (voting, bagging, stacking), and hybrid ML models are used.
Despite the investigation of the application of ML techniques to predict the capacity of RC
beams in shear, all the above studies generally failed to propose a user-friendly and
practical shear design approach based on the results of the trained ML models. As it is well
understood, ML algorithms are generally black-boxes and as such cannot be reproduced
MSc Thesis Page 4

by others. Thus, practical implementation and simplification of the ML-based models are
essential. Moreover, most of the models are developed using a limited number of
experimental datasets, which limits the generalization of the developed models.
In lieu of the above-mentioned limitations, the current study investigated the application
of several ML models (both single and ensemble models) to propose the best predictive
model based on a large database of slender and deep RC beams with/without stirrups.
Moreover, the best model among the developed models is deployed into a user-friendly
web-based application, which can be used by practitioners and researchers in the field of
civil engineering to accurately determine the capacity of RC beams in shear. The
developed web-based application facilitates a rapid and accurate prediction of the capacity
of RC beams in shear.
1.2 Research objectives
This is a pioneer research work leveraging the power of different machine learning models
to produce accurate and robust model for shear design of reinforced concrete (RC) beams.
Machine learning models generally exhibit higher prediction capacity compared to
statistical and empirical models; however, the accuracy of machine learning models can
drastically vary from one type of problem to other. One of the reasons for the variation in
the accuracy of the ML models with the problem type is attributed to the complex
relationship of input and output. In this context, among different ML models, ensemble
learners combine different base learners to produce a powerful model. Thus, this research
is aimed to produce a robust ensemble machine learning model to accurately estimate the
capacity of RC beams in shear considering key design factors that influence their shear
MSc Thesis Page 5

capacity. Moreover, a web-based application is developed using the best ML model. The
developed web-based application can be easily utilized by practitioners or researchers in
the field of Civil Engineering to accurately determine the capacity of RC beams in shear.
1.2.1 General objective

The main objectives of the current research are listed below.
 Develop a novel ML-based model for estimating the shear capacity of RC beams
with/without stirrups.
 Compare the prediction accuracy of the proposed ensemble model with other
machine learning models.
 Comparative investigation of the proposed model and existing shear design
models and code equations.
1.2.2 Specific objective

 Collect a large database of RC beams experiments and preprocess the database.
 Identify the main factors that influence the shear capacity of RC beams based on a
thorough literature review.
 Develop different ML models for predicting the capacity of RC beams in shear
and select the best predictive model using different statistical performance indices.
 Compare the predictive capability of the proposed model with that of the existing
models.
 Develop a simple and efficient web-based application, which can be used by
practitioners and researchers in the field of civil engineering without the need of
the knowledge of machine learning algorithms.
MSc Thesis Page 6

1.3 Scope
The scope of this thesis is to propose novel machine learning-based shear design models
for RC beams with/without stirrups. Various ML-based are developed in order to propose
the best shear capacity predictive models. Finally, the predictive performance of the
proposed model is compared with that of the existing models.
1.4 Research Significance
There are different approaches and guidelines to determine the shear capacity of RC
beams. Numerical, analytical, and mechanical models are widely-used models. All these
models are based on assumptions hence fail short of producing accurate predictions.
Besides, there are large discrepancies and uncertainty in the existing shear design models.
One of such discrepancies is related to the selection of important factors in determining
the shear capacity of RC beams. Furthermore, the existing empirical models are generally
based on a simple truss analogy without considering the interaction between different
factors and the variation of the shear crack angle. In contrast, a machine learning model
learns the relationship between the input features and the response variable from data
without prior assumption or knowledge of the underlying mathematical and physical
models. In nature, machine learning models work better on large data to solve
classification and regression problems. Therefore, this study is aimed to propose an
accurate and reliable shear design model using different ML techniques considering all
important parameters. To this end, a large database of experiments on RC beams was
collected. The collected data comprised various beam geometries, material strengths for
both concrete and steel, longitudinal and transverse reinforcement ratios, and shear span-
MSc Thesis Page 7

to-effective depth ratios. The performance of the ML model is greatly influenced by the
value of its hyperparameters, which are parameters that control the learning process of a
given ML model. The values of the hyperparameters for each ML model are tuned to
obtain the best model. Finally, the prediction performance of the proposed ML models is
compared with that of the existing models and code equations. A comparative
investigation of the proposed and existing models revealed the superior prediction
capability of the proposed model in terms of accuracy, safety, and economical aspects.
Hence, this research is significant in assessing the best model in determining the shear
capacity of RC beam and propose a robust predictive model.
1.5 Layout of work
Chapter 1 presents the background information of the study, objective, significance, scope,
and methodology of the study. Chapter 2 presents a literature review on the shear strength
of RC beam focused on parameters influencing the shear strength of RC beam and existing
shear design models and code equations. Chapter 3 presents the experimental database
used in this study. Different machine learning techniques are introduced in Chapter 4.
Moreover, Chapter 4 discusses different single and ensemble ML models. Six ML models
(two single models and four ensemble models) are analyzed in this chapter. Chapter 5
describes the results from the ML model and comparative study with existing models and
code equations. The performance of each model is evaluated using different performance
indices. Chapter 6 discuss the conclusion from this study and recommendations for future
research.
MSc Thesis Page 8

1.6 Methodology
The following procedures will be followed in this research work.
 Construction of database: an extensive database of experimental results of deep
and slender RC beams with/without stirrups critical in shear will be compiled form
the literature.
 Identification of input parameters: this is a critical stage in the development of
ML models. The input variables that affect the capacity of RC beams will be
identified based on a thorough literature survey.
 Data processing: the third step in the development of ML models is the
preprocessing of the collected database, which includes but is not limited to the
identification of outliers and normalization of the database.
 Hyperparameter tuning: the performance of ML models depends on the values
of hyperparameters, which are parameters that control the learning process of a
given ML model. The hyperparameters for each model will be tuned carefully to
achieve superior prediction accuracy. In order to select the best predictive model,
different performance indices will be used.
 Model training, validation, and testing: the database will be randomly split into test
and train datasets. The training database will be used for the model training and
development, while the testing dataset will be used to finally appraise the
performance of the ML models.
 Comparative study: the efficacy of the proposed models will be finally compared
with that of the existing models and code equations including ACI 318 and
Eurocode.
MSc Thesis Page 9

CHAPTER 2 LITERATURE REVIEW
2.1 Shear strength of RC beam
The formation of diagonal tension crack occurring at an angle with respect to the beam
axis causes the shear failures. Shear failure is sudden and occurs without prior warning,
thus it is a brittle failure, unlike flexural failure. Hence, it is vital to make sure that the
shear supply is greater than the shear demand in order to avoid such undesirable shear
failure. The primary goal of shear design is to prevent such failure and ensure flexural
failure. However, shear failure involves a complex failure mechanism, and as such difficult
to develop simple models like that used to characterize flexural behavior, where the
assumption of plane sections remain plane is valid. Several studies have been conducted
to propose shear strength models based on the principles of mechanics. Shear behavior is
the sum of various mechanisms, thus difficult to establish a mechanics-based shear design
model. The free-body diagram in Figure 2.1 shows the internal forces in RC beam after
the formation of diagonal shear crack.
Figure 2.1: Internal forces along diagonal shear crack [43].
MSc Thesis Page 10

In the above figure, C denotes uncracked compressive force in uncracked concrete, 𝑉𝑐𝑧 is
the shear in the compression zone, 𝑉𝑠 is the shear capacity provided by transverse
reinforcement, 𝑉𝑖 is the shear capacity provided by the aggregate interlock, 𝑇 is the tension
force in longitudinal reinforcement bars, and 𝑉𝑑 is the dowel action of longitudinal
reinforcement bars.
The shear force in RC beams is resisted by different mechanisms including transverse
reinforcement (stirrups), aggregate interlock, uncracked concrete, and dowel action of the
flexural bars [44,45]. For instance, Figure 2.2 illustrates the relative contributions of
different mechanisms to the capacity of RC beams without stirrups at a cracked section.
As can be seen in this figure, concrete aggregate interlock provides the highest shear
capacity compared to dowel action of flexural reinforcement and uncracked concrete
contribution in the compression zone. Various factors influence the shear capacity of RC
beams. The following section briefly discusses the influence of different key variables on
the capacity of RC beams in shear.
MSc Thesis Page 11

Figure 2.2: Components of shear resistance at cracked concreate section for RC beams
without stirrup [46].
2.2 Parameters influencing shear strength of RC beam
The shear capacity of RC beams is influenced by different factors: concrete compressive
strength, yield strength and reinforcement ratio of stirrups, cross-sectional depth of the
beam, 𝑎/𝑑 ratio, and longitudinal reinforcement ratio, as will be discussed below.
2.2.1 Concrete compressive strength

Previous studies reported an improvement in the shear strength of RC beams with
increasing concrete compressive strength (𝑓𝑐 ). However, there is inconsistency in the
formulations for the shear strength provided by concrete. For instance, according to ACI
318-14 [47], the concrete contribution to shear capacity has been assumed to be 2√𝑓𝑐, 𝑏𝑑.
However, in Eurocode 2 [4], the shear strength is correlated with the cubic root of 𝑓𝑐, , 𝑉 ∝
3 ,
√𝑓𝑐 . Shear forces are transmitted through concrete cracks by virtue of roughness of the
concrete interfaces.
MSc Thesis Page 12

2.2.2 Transverse shear reinforcement
As it is well understood, the provision of transverse reinforcement (stirrups) increases the
shear strength of RC beams which in turn increases their ductility. In addition to providing
additional shear strength, stirrups are used to hold the main reinforcements and link
together the flexural compression and tension sides of a member and ensure that the two
sides act as a unit. Breakdown of those linkages may cause shear failure. Besides, shear
reinforcement limits the development of inclined cracks and prevents the cracks from
widening. The strain in shear reinforcement prior to the onset of crack is equal to that in
concrete. However, with the formation of the inclined cracks shear reinforcement
contributes substantially to shear resistance as diagonal tension cracks form at very small
strains. The contribution of stirrups to the capacity of RC beams in shear depends on the
diameter and spacing of stirrups.
2.2.3 Depth
Increasing the effective depth of RC beams intuitively increases the shear capacity of the
beams. In most of the models and code equations, the shear capacity of RC beams is
directly proportional to the effective depth of the beam. The average shear strength at the
failure of beams is reduced by an increase in the beam depth.
2.2.4 Longitudinal reinforcement ratio
The longitudinal reinforcement provides shear resistance through the dowel action. In
addition to providing an enhanced dowel action, the presence of increased longitudinal
reinforcement controls the propagation of flexural cracks and leads to an increase in the
neutral axis depth, which in turn increases the depth of uncracked concrete in compression.
The increase in the depth of uncracked concrete in compression, in turn, enhances the shear
MSc Thesis Page 13

capacity of the beam. Thus, an increase in the longitudinal reinforcement ratio enhances
the shear strength of the beam owing to an increase in dowel action [48–50]. However,
there is a discrepancy in international codes and standards for the shear design of RC
beams in accounting for the shear strength provided by longitudinal reinforcement. For
instance, in Eurocode2 [4], the shear strength of RC beams has a direct relation with the
cubic root of the longitudinal reinforcement ratio (𝜌𝑠𝑥 ). However, this effect has not been
considered in ACI 318 [3].
2.2.5 Shear span-to-depth ratio
The shear span-to-depth (𝑎/𝑑) ratio is another important factor that influences the shear
capacity of RC beam. However, it has a distinct effect on different shear resisting
mechanisms. The concrete contribution to the shear capacity decreases as 𝑎/𝑑 ratio
increases [51]. However, the shear strength provided by transverse reinforcement increases
with an increase in 𝑎/𝑑 ratio [51]. In addition, shear resistance provided by first diagonal
crack is significantly larger than shear resistance provided by concrete when the shear
span-to-depth ratio is large but much smaller than shear resistance provided by concrete
when 𝑎/𝑑 ratio is small. Generally, an increase in 𝑎/𝑑 ratio cause reduction of ultimate
shear capacity of RC beams [51].
2.3 Existing Shear Design Model and Guideline
There exist different standards and models for shear capacity determination of RC beams.
Collins et al. [52] provided a review on the shear models proposed in the early 20th century
(1948 to 2008). In a traditional truss analogy of Ritter and Morsch, the shear stress was
assumed to be resisted by stirrups and diagonal compressive concrete struts inclined at 45°
MSc Thesis Page 14

to the beam height [53]. According to this model, the ultimate shear capacity of the beam
corresponds to the yielding of the stirrups, while neglecting the tensile stress in a cracked
concrete and thus gives a conservative prediction [54]. Moreover, the model ignores the
shear resisted by uncracked concrete section, aggregate interlock along the cracks, and
dowel action of the longitudinal reinforcement [53]. The truss model has provided a basis
for different codes of practice including ACI 318-14 [55] and Eurocode-2 (2004) [4]
standards. ACI 318-14 standard assumes a constant strut angle of 45° while considering
the contribution of the concrete in the tension. On the other hand, Eurocode-2 (2004) is
based on a variable angle truss model but ignores concrete contribution to the shear
strength for beams reinforced with ITSR. The most widely used models and code equations
are summarized in Table 2.1 and discussed in the following subsections.
Table 2.1: Shear capacity of RC beams according to different models

Reference Formulation
Eurocode 2 [4]  Members without internal transverse reinforcement
𝑉𝑛 = 𝑉𝑅𝑑,𝑐 = 𝐶𝑅𝑑,𝑐 𝑘(100𝜌𝑠𝑥 𝑓𝑐 ′ )1/3 𝑏𝑤 𝑑 (𝑓𝑐 ′ is in MPa)
200
𝑘 =1+√ ≤ 2.0 (𝑑 in mm)
𝑑
𝜌𝑠𝑥 ≤ 0.02
0.18
𝐶𝑅𝑑,𝑐 = , 𝛾𝑐 = 1.5
𝛾𝑐
 Members with internal transverse reinforcement
𝑉𝑛 = 𝑉𝑅𝑑,𝑠 = 𝜌𝑠𝑣 𝑓𝑠𝑣𝑑 𝑏𝑤 𝑍 cot 𝜃 ≤ 𝛼𝑐𝑤 𝑏𝑤 𝑍𝜈1 𝑓𝑐𝑑 /(cot 𝜃 + tan 𝜃)

𝑓𝑠𝑣𝑑 = 0.8𝑓𝑠𝑣
1 ≤ cot 𝜃 ≤ 2.5
𝑍 = 0.9𝑑
𝛼𝑐𝑤 = 1, for non-prestressed beam
MSc Thesis Page 15

0.6 for 𝑓𝑐 ′ ≤ 60 MPa

𝜈1 = { 𝑓𝑐 ′
0.9 − > 0.5 for 𝑓𝑐 ′ ≥ 60 MPa
200
AASHTO-
𝑉𝑛 = 𝛽𝜆√𝑓𝑐 ′ + 𝜌𝑠𝑣 𝑓𝑠𝑣 𝑏𝑤 𝑑 cot 𝜃 (𝑓𝑐 ′ is in MPa)
LRFD [56]
where,
𝛽
0.4
, for section with minimum amount of ISR
1 + 750𝜀𝑥
=
0.4 1300
, esle
{1 + 750𝜀𝑥 1000 + 𝑆𝑥𝑒
𝜃 = 29 + 3500𝜀𝑠𝑥
ACI 318 [47] 𝑉𝑛 = 𝑉𝑐 + 𝑉𝑠
√𝑓𝑐
𝑉𝑐 = ( ) 𝑏𝑤 𝑑, 𝑓𝑐 < 70 MPa
6
𝐴𝑠𝑤
𝑉𝑠 = 𝑑𝑓𝑦𝑡
𝑆
JSCE 2007 [57] 1⁄

3
𝑉𝑐 = 0.9𝛽𝑑 𝛽𝑝 𝛽0 𝑓𝑐 𝑏𝑤 𝑑
where,
1
100 ⁄4
𝛽𝑑 = ( ) ≤ 1.5
𝑑
1
𝛽𝑝 = (100𝜌) ⁄3 ≤ 1.5
𝛽0 = 0.75 + [1.4/(𝑎⁄𝑑 )]
𝐴𝑠𝑤
𝑉𝑠 = 𝑓 (𝑗 cos 𝜃)
𝑆 𝑦𝑤 𝑡
CSA A23.3-94 𝑉𝑛 = 𝑉𝑐 + 𝑉𝑠
(simplified
method) [58] 0.06√𝑓𝑐′ 𝑏𝑤 𝑠
𝑉𝑐 = 0.2√𝑓𝑐′ 𝑏𝑤 𝑑 𝑤ℎ𝑒𝑛, 𝐴𝑠𝑤 ≥ 𝑜𝑟 𝑑 ≤ 300 𝑚𝑚
𝑓𝑦
260
𝑜𝑟 𝑉𝑐 = √𝑓 ′ 𝑏 𝑑
1000 + 𝑑 𝑐 𝑤
0.06√𝑓𝑐′ 𝑏𝑤 𝑠
𝑤ℎ𝑒𝑛, 𝐴𝑠𝑤 ≤ 𝑜𝑟 𝑑 ≥ 300 𝑚𝑚
𝑓𝑦
𝐴𝑠𝑤
𝑉𝑠 = 𝑑𝑓𝑦𝑡 where, 𝑉𝑠 ≤ 0.8√𝑓𝑐′ 𝑏𝑤 𝑑
𝑆
MSc Thesis Page 16

2.3.1 ACI 318 [47]

According to ACI 318 [47], the shear capacity of RC beam is evaluated as a simple
superposition of the capacity provided by concrete (𝑉𝑐 ) and stirrups (𝑉𝑠 ). Figure 2.3 depicts
the shear contributions provided by stirrups and concrete. Thus, in ACI 318, the nominal
shear capacity (𝑉𝑛 ) is determined as:
𝑉𝑛 = 𝑉𝑐 + 𝑉𝑠 (1)
√𝑓𝑐
𝑉𝑐 = ( ) 𝑏𝑤 𝑑, 𝑓𝑐 < 70 MPa (2)
6
𝐴𝑠𝑤
𝑉𝑠 = 𝑑𝑓𝑦𝑡 (3)
𝑆
where,
𝑉𝑐 = Shear capacity provided by concrete
𝑉𝑠 =Shear capacity provided by transverse reinforcement
𝑓𝑦𝑡 = Yield strength of stirrups
𝑏𝑤 = Width of the web
𝑆 = Spacing of the stirrups
𝑑 = Effective depth of the beam section
MSc Thesis Page 17

(a)
(b)
Figure 2.3: Concrete and stirrups shear resisting mechanisms (a) and graphical illustration
of their contribution (b) [59].
2.3.2 Eurocode (EC2) [4]
According to Eurocode (EC2) [4], different models are used to determine the capacity of
RC beams with and without internal transverse reinforcement, as discussed below.
2.3.2.1 Members without internal transverse reinforcement

For beams without stirrups, the shear capacity is given by [4]:
1⁄
𝑉𝑛 = 𝑉𝑅𝑑,𝑐 = 𝐶𝑅𝑑,𝑐𝑘 (100𝜌𝑠𝑥 𝑓𝑐, ) 3 𝑏𝑤 𝑑 (4)
200
𝑘=1+√ ≤ 2.0
𝑑
(5)
𝜌𝑠𝑥 ≤ 0.02
MSc Thesis Page 18

0.18
𝐶 𝑅𝑑,𝑐 = , 𝛾𝑐 = 1.5
𝛾𝑐 (6)
where 𝜌𝑠𝑥 is the longitudinal reinforcement ratio.
2.3.2.2 Members with internal transverse reinforcement
For beams internally reinforced with stirrups, the shear capacity is given by [4]:
𝑉𝑛 = 𝑉𝑅𝑑,𝑐 = 𝜌𝑠𝑡 𝑓𝑠𝑡𝑑 𝑏𝑤 𝑍𝑐𝑜𝑠𝜃 ≤ 𝛼𝑐𝑤 𝑏𝑤 𝑍𝑣1 𝑓𝑐𝑑 /(𝑐𝑜𝑠𝜃 + 𝑡𝑎𝑛𝜃) (7)
𝑓𝑠𝑡𝑑 = 0.8𝑓𝑦𝑡 (8)
𝑍 = 0.9𝑑 (9)
1 ≤ 𝑐𝑜𝑠𝜃 ≤ 2.5 (10)
𝑣1 = 0.6 for 𝑓𝑐, ≤ 60𝑀𝑃𝑎

(11)
𝑓,
𝑐
𝑣1 = 0.9 − 200 > 0.5 for 𝑓𝑐, ≥ 60𝑀𝑃𝑎
where,
𝜃 = shear crack angle relative to the longitudinal axis of the beam, and
𝛼𝑐𝑤 = 1, for non-prestressed beam.
2.3.3 JSCE (2007) [57]
The Japanese JSCE code (2007) [57] considers the effect of effective depth of member,
the reinforcement ratio of longitudinal bars, and the concrete compressive strength in its
model to estimate the concrete contribution to shear strength of RC members, as follows:
𝑉𝑐 = 𝛽𝑑 𝛽𝑝 𝛽𝑛 𝑓𝑣𝑐𝑑 𝑏𝑤 𝑑/𝛾𝑏 (12)
MSc Thesis Page 19

3 ′
𝑓𝑣𝑐𝑑 = 0.20 √𝑓𝑐𝑑 ≤ 0.72 (MPa) (13)
1⁄
100 4
(14)
𝛽𝑑 = ( ) ≤ 1.5, (𝑑 𝑖𝑛 mm)
𝑑
1⁄
𝛽𝑝 = (100𝜌𝑠𝑡 ) 3 (15)
2𝑀0
0 ≤ 𝛽𝑛 = 1 + ≤ 2.0, 𝑓𝑜𝑟 𝑁𝑑′ ≥ 0 (16)
𝑀𝑢𝑑
where,
𝛽𝑑 = Depth factor,
𝑁𝑑′ = Design axial force,
𝛽𝑑 = Factor accounting for the effect of 𝜌𝑠𝑥 ,
𝛽𝑛 = Factor accounting for the effect of 𝑎/𝑑 ratio,
𝑀0 = Flexural moment required to balance the stress due to axial load at the extreme
tension fiber,
𝑀𝑢𝑑 = Pure flexural capacity neglecting the effect of axial force, and
𝛾𝑏 = Member factors, generally taken as 1.30.
For members without axial load the value of 𝛽𝑛 is unity.
2.3.4 CSA A23.3-94 (Simplified Method) [58]
The simplified method of Canadian Standard [58] is based on the 45-degree truss model.
The shear capacity of RC beam is evaluated as a simple superposition of the capacity
provided by stirrups and concrete, as follows:
MSc Thesis Page 20

0.06√𝑓𝑐′ 𝑏𝑤 𝑠
𝑉𝑐 = 0.2√𝑓𝑐′ 𝑏𝑤 𝑑 𝑤ℎ𝑒𝑛 𝐴𝑠𝑤 ≥ 𝑜𝑟 𝑑 ≤ 300 𝑚𝑚
𝑓𝑦
(17)
260 0.06√𝑓𝑐′ 𝑏𝑤 𝑠
𝑉𝑐 = √𝑓 ′ 𝑏 𝑑 𝐴𝑠𝑤 ≤ 𝑜𝑟 𝑑 ≥ 300 𝑚𝑚
1000 + 𝑑 𝑐 𝑤 𝑓𝑦
𝐴𝑠𝑤
𝑉𝑠 = 𝑑𝑓𝑦𝑡 ≤ 0.8√𝑓𝑐′ 𝑏𝑤 𝑑 (18)
𝑆
2.3.5 Modified compression field theory
The modified compression field theory (MCFT) [60] used the principles of equilibrium,
compatibility, and constitutive relationship of the cracked concrete section to develop a
shear model. Unlike the traditional truss-based models (e.g. ACI 318 [47]), the MCFT
accounts for the contribution of the concrete tensile stresses of a cracked section.
Moreover, a variable angle of diagonal struts was considered compared to the fixed value
of 45° strut angle. The principles of compatibility, equilibrium and constitutive
relationship of cracked concrete section of MCFT are shown in the figure below.
MSc Thesis Page 21

Figure 2.4: MCFT equations [46]
The MCFT entails solving a large number of equations iteratively [61,62]. To ease this
procedure, Collins et al. [61] proposed a general method of shear design based on the
MCFT where the shear capacity is determined as a function of tensile stress factor (𝛽) and
inclination of the principal compressive stress (𝜃). The authors provided a table for
approximating the values of 𝛽 and 𝜃, for members with as well as without stirrups, for a
limited range of longitudinal strain values (𝜀𝑠𝑥 ≤ 0.2%). On the other hand, a simplified
compression field theory (SCFT) was developed by Bentz et al. [62] in which the values
of 𝛽 and 𝜃 are determined as a function of the longitudinal tensile strain (ɛ𝑥 ) and crack
spacing (𝑆𝑥𝑒 ). In the SCFT, simple expressions for 𝛽 and 𝜃 are used, reducing the required
number of parameters and iterations, while predicting the shear capacity of the concrete
beam almost as accurately as the MCFT.
MSc Thesis Page 22

0.4 1300
𝛽= (19)
1 + 1500𝜀𝑥 100 + 𝑠𝑥𝑒
1
𝑆𝑥𝑒 =
sin 𝜃 cos 𝜃 (20)
𝑠𝑚𝑥 + 𝑠𝑚𝑦
For beams not reinforced with stirrup the crack spacing is given by as follows as per.
35𝑆𝑥
𝑆𝑥𝑒 = ≥ 0.85𝑆𝑥 (21)
𝑎𝑔 + 16
𝑣 𝑐𝑜𝑡 𝜃 − 𝛽 √𝑓𝑐, 𝑡𝑎𝑛 𝜃 (22)

𝜀𝑥 =
𝐸𝑠 𝜌𝑠𝑥
𝑆𝑥𝑒
𝜃 = (29 + 7000𝜀𝑥 ) × (0.88 + ) ≤ 75° (23)
2500
𝑣 = 𝛽√𝑓𝑐, + 𝜌𝑠𝑦 𝑓𝑠𝑦 cot 𝜃 (24)
𝑓𝑠𝑥 = 𝜀𝑥 𝐸𝑠 ≤ 𝑓𝑦𝑥
Thus, the shear capacity of RC beam is determined by, 𝑉 = 𝑣𝑏𝑤 𝑑
where,
𝑠𝑚𝑥 = Average spacing perpendicular to the x-reinforcement (stirrups spacing)
𝑠𝑚𝑦 = Average crack spacing perpendicular to y-reinforcement (effective depth of the

beam)
MSc Thesis Page 23

CHAPTER 3 EXPERIMENTAL DATABASE
This chapter presents the experimental database of shear critical RC beams (with and
without stirrups). Data curation, which involves collection of relevant experimental
results, data preprocessing, and identification of input and output variables is a critical step
in the development of a machine learning model. Thus, in the following sections, the
details on the experimental database of RC beams with or without stirrups are provided.
Moreover, the distribution of the database for both RC beams with and without stirrups is
provided. The development of an accurate shear model requires the incorporation of all
parameters affecting the shear capacity of RC beams. Thus, the input features (key design
parameters) are identified and the effects of these factors on the response variable (shear
capacity in this case) are discussed based on the collected database for both RC beams
with/without stirrups in Section 3.1 and 3.2.
3.1 Experimental database for RC beams without stirrups
3.1.1 Introduction
A large superset database of RC beams without stirrups, which is collected by Collins et
al. [41] is used in this study. The authors collected the test results of RC beams tested over
60 years of research between 1948 and 2006. A total of 1849 RC beams are contained in
the database. The experimental database covers a wide range of beam geometries, material
strengths (concrete and steel reinforcement), diameter and number of longitudinal
reinforcement steel bars, and shear 𝑎/𝑑 ratios. The following general criteria were used to
collect the database:
MSc Thesis Page 24

 Material strength: RC beams with no limit on the material strength, i.e., no limit
on the 𝑓𝑐′ and yield strength of longitudinal reinforcement bars. Details on the
material strength will be provided in the subsequent sections.
 Beam shape: both rectangular and T-section beams are included in the database.
 Reinforcement: beams without stirrups.
 Concrete: normal strength concrete without additional fibers.
 Bar type: deformed longitudinal steel bars.
 Load type: point load or uniform load (details will be provided in the following
subsection).
 Beam geometries: no limits on the geometric size of the beams.
As mentioned above, both RC rectangular and T-section beams are contained in the
database. As shown in Figure 3.1, among the total of 1849 RC beams included in the
database, 1599 beams were rectangular beams representing 86% of the beams, while 250
RC beams had T-section, which denotes 14% of the beams.
MSc Thesis Page 25

Figure 3.1: Geometry of RC beams in the database.
The beams were tested under three types of loading conditions: (a) point load (Type P),
(b) uniform load with one concentrated load (Type U1), and (c) uniform load with two-
point concentrated loads (Type U2). Figure 3.2 shows the details of the three loading
types.
Figure 3.2: Schematic of the loading types

Figure 3.3 shows the distribution of the beams loading type. Among the total 1849 RC
beams, 1701 beams were tested under three-point load (Type P), which represents 92% of
the beams. Thus, the majority of the beams were tested under Type P loading condition.
MSc Thesis Page 26

For the uniform loads, 11% were tested under Type U1, while the remaining 3% were
tested under Type U2 loading condition.
Figure 3.3: Distribution of RC beams in terms of the loading type.
As discussed earlier, the development of efficient machine learning models requires proper
identification of the input features or key design parameters that influence the response
variables; namely, the shear capacity of the beams in this case. In this study, a total of eight
important factors are identified based on a thorough literature review. These factors
include the width of the web (𝑏𝑤 ), the width of the flange (b), cross-sectional depth (ℎ),
effective depth of the section (𝑑), shear span-to-effective depth ratio (𝑎/𝑑), concrete
compressive strength (𝑓𝑐′ ), longitudinal reinforcement ratio (𝜌𝑠 ), and yield strength of steel
bars (𝑓𝑦 ). The details on the distribution of these factors and their effect on the shear
capacity are provided in the following sections.
MSc Thesis Page 27

3.1.2 Distribution of input parameters
A wide range of input parameters is comprised in the database. Figure 3.4 shows the
distribution of each factor. Moreover, the distribution of the input features in terms of the
average (mean) value, standard deviation (STD), minimum, first quartile (Q1), second
quartile (Q2), third quartile (Q3), and maximum values are listed in Table 3.1. As can be
observed in Figure 3.4 and Table 3.1, an extensive range of the factors is included in the
database. For instance, the values of 𝑓𝑐′ ranged between 6.1 MPa to 127.5 MPa with a mean
strength of 34.86 MPa and a standard deviation of 18.34 MPa. As shown in Figure 3.4
and Table 3.1, 75% of the beams were constructed with concrete having a compressive
strength of 24.1 MPa or higher strength. The compressive strength of concrete in most of
the beams ranged between 24.1 MPa and 50 MPa, as shown in Figure 3.4. With regard to
the beam section, the width of the beam web and flange ranged between 21 mm to 3000
mm, the height of the beam section was in the range of 51 mm–3140 mm, and the effective
depth of the beam ranged from 41 mm to 3000 mm, as listed in Table 3.1. Moreover, the
𝑎/𝑑 ratio is in the range of 0.25 to 15.06. Thus, both slender and deep beams are included
in the database. As per the ACI- ASCE Committee 445 [53], RC beams with 𝑎/𝑑 ratio of
MSc Thesis Page 28

less than 2.50 are characterized as deep beams, while slender beams have 𝑎/𝑑 ratio of
greater than 2.50.
MSc Thesis Page 29

MSc Thesis Page 30

Figure 3.4: Distribution of the input variables in the database.
Table 3.1: Geometry and material characteristics of the beams included in the database
Parameters Mean STD Minimum Q1 Q2 Q3 Maximum
𝑏𝑤 (𝑚𝑚) 213.01 212.94 21 150 153 200 3000
𝑏 (𝑚𝑚) 256.76 230.21 21 152 157 300 3000
ℎ (𝑚𝑚) 364.35 254.05 51 250 305 381 3140
𝑑 (𝑚𝑚) 320.25 237.77 41 210 270 318 3000
MSc Thesis Page 31

𝑎/𝑑 3.20 1.80 0.25 2 3 3.94 15.06
𝑓𝑐 (𝑀𝑃𝑎) 34.86 18.34 6.1 24.1 29.7 37.4 127.5
𝜌𝑠 (%) 2.24 1.52 0.1 1.24 1.87 2.73 9.5
𝑓𝑦 (𝑀𝑃𝑎) 462.37 172.14 267 379 420 483 1779
STD: standard deviation
As discussed earlier, the flexural reinforcement bars had a deformed shape. The yield
strength of flexural reinforcement bars ranged between 172.14 MPa to 1779 MPa, as listed
in Table 3.1; however, most of the beams were reinforcement with flexural bars with yield
strength in the range of 379 MPa–550 MPa, as shown in Figure 3.4. Finally, the
longitudinal reinforcement ratio ranged between 0.1% and 9.5% with a mean value of
2.24% and standard deviation of 1.52%. It is worth mentioning here that the concrete
compressive strength and yield strength of longitudinal reinforcement bars are limited to
100 MPa and 1000 MPa, respectively, in the final database used in the development of the
machine learning-based models, as will be discussed further in Chapter 5. The effect of
the input parameters on the shear capacity of RC beams is briefly discussed in the
following section.
3.1.3 Effect of input parameters on the shear capacity
In Section 2.2, the effect of different key design factors on the shear capacity is discussed
based on a thorough literature review. In this section, the collected database is used to
investigate the influence of the factors on the shear capacity of RC beams without stirrups.
Figure 3.5 depicts the variation of the shear capacity (𝑉𝑢 ) with the change in each factor
based on the test results of RC beams included in the database.
MSc Thesis Page 32

MSc Thesis Page 33

MSc Thesis Page 34

Figure 3.5: Relationship between the input parameters and the shear capacity.
As can be observed in Figure 3.5, each parameter significantly influences the shear
capacity of RC beams. Moreover, there exists a nonlinear relationship between the factors
and the shear capacity of RC beams. Thus, a simple linear model cannot be used to
determine the shear capacity of the beams. Generally, the shear capacity of the beam
increased with an increase in the width of the web, width of flange, height of the beam
section, and effective depth of the beam, as shown in Figure 3.5. However, an increase in
MSc Thesis Page 35

the shear span-to-effective depth showed a negative effect on the shear capacity, thus, the
shear capacity decreased with an increase in 𝑎/𝑑 ratio. The concrete compressive strength,
yield strength of longitudinal bars, and reinforcement ratio of longitudinal bars showed a
complex trend, which shows the complex nonlinear relationship between the factors and
the shear capacity of the beams and interaction between the factors.
3.2 Experimental database for RC beams with stirrups
3.2.1 Introduction
A database of RC beams with stirrups, which is collected by Zhang et al. [42], Jung et al.
[63], and Cladera et al. [64] is used in this study. The authors collected the experimental
results of a total of 348 RC beams. The experimental database covers a wide range of beam
geometries, concrete and steel reinforcement strengths, diameter and number of
longitudinal and transversal reinforcement steel bars, and shear span-to-effective depth
ratios. Both rectangular and T-section beams are included in the database. And also both
slender beam and deep beam are included in the database. As shown in Figure 3.6, among
the total of 348 RC beams included in the database, 116 beams were slender beams
representing 33.33% of the beams, while 232 RC beams had deep beam, which denotes
66.67% of the beams.
MSc Thesis Page 36

Figure 3.6: Beam type (slender beams/deep beams)
As discussed in section 4.1, the development of efficient machine learning models requires
proper identification of the input features or key design parameters that influence the
response variables; namely, the shear capacity of the beams in this case. In this study, a
total of seven important factors (input features) are identified based on a thorough literature
review. These factors include the width of the flange (b), effective depth of the section (𝑑),
shear span-to- depth ratio (𝑎/𝑑), compressive strength of concrete (𝑓𝑐′ ), reinforcement ratio
of longitudinal bars (𝜌𝑠𝑦 ), transversal reinforcement ratio (𝜌𝑠𝑦 ), and yield strength of steel
bars (𝑓𝑦 ). The details on the distribution of these factors and their influence on the shear
capacity are provided in the following sections.
3.2.2 Distribution of input parameters
Figure 3.7 illustrates the distribution of each input parameters. Moreover, the distribution
of the input features are listed in Table 3.2. As can be observed in Figure 3.7 and Table
3.2, an extensive range of the factors is included in the database. The values of 𝑓𝑐′ ranged
between 12.8 MPa to 125 MPa with an average strength of 41.21 MPa and a standard
deviation of 20.67 MPa, the yield strength of flexural reinforcement bars ranged between
MSc Thesis Page 37

0 MPa to 844 MPa, the longitudinal reinforcement ratio ranged between 0.76 MPa and 5.8
MPa with a mean value of 2.86 MPa, and standard deviation of 0.97, and the transversal
reinforcement ratio range between 0 MPA and 1.90 MPa, with a mean value of 0.31 MPa,
and standard deviation of 0.23 MPa, With regard to the beam section, the width of the
beam ranged between 76 mm to 457 mm, and the effective depth of the beam ranged from
95 mm to 1200 mm, and the shear span-to-effective depth ratio is in the range of 1 to 5.98,
both slender and deep beams are included in the database, as listed in Table 3.2.
Table 3.2: Geometry and material characteristics of the beams included in the database
Parameters Mean STD Minimum Q1 Q2 Q3 Maximum
𝑏 (𝑚𝑚) 221.39 80.70 76.00 152.00 203.00 290.00 457.00
𝑑 (𝑚𝑚) 365.07 151.38 95.00 292.00 345.00 425.00 1200.00
𝑎/𝑑 2.98 0.82 1.00 2.50 3.00 3.60 5.98
𝑓𝑐 (𝑀𝑃𝑎) 41.21 20.67 12.80 26.20 33.30 50.00 125.00
𝜌𝑠𝑥 (%) 2.86 0.97 0.76 2.20 2.80 3.40 5.80
𝜌𝑠𝑦 (%) 0.31 0.23 0.00 0.14 0.24 0.50 1.90
𝑓𝑦 (𝑀𝑃𝑎) 423.74 146.08 0.00 331.00 379.00 531.00 844.00
STD: standard deviation; Q1: first quartile; Q2: second quartile; Q3: third quartile
In the final database used in the development of the machine learning-based models, as
will be discussed further in Chapter 6. The effect of the input parameters on the shear
capacity of RC beams is briefly discussed in the following section
MSc Thesis Page 38

MSc Thesis Page 39

MSc Thesis Page 40

Figure 3.7: Distribution of the input variables in the database.
3.2.3 Effect of input parameters
In this section, the collected database is used to investigate the influence of the factors on
the shear capacity of RC beams with stirrups. Figure 3.7 shows the variation of the shear
capacity (𝑉𝑢 ) with the change in each factor based on the experimental results of RC beams
included in the database.
MSc Thesis Page 41

MSc Thesis Page 42

MSc Thesis Page 43

Figure 3.8: Relationship between the input parameters and the shear capacity of RC beams
with stirrups
As observed from Figure 3.8, all parameter significantly influences the shear capacity of
RC beams and there is the complex nonlinear relationship between the factors and the
shear capacity of the beams and interaction between the factors.
MSc Thesis Page 44

CHAPTER 4 MACHINE LEARNING-BASED MODELS
4.1 Introduction to Machine Learning
Machine learning has recently gained immense attention owing to its ability to effectively
determine the relationship between the input features and the response variable (s) in a
complex system. In contrast to the case of most empirical models, ML techniques does not
require the prior assumption or knowledge of the underlying mathematical and physical
models [16]. Artificial neural network (ANN) is one of the most widely utilized ML
techniques [9]. It has been successfully employed for different structural engineering
studies; some of which include the prediction of mechanical properties of concrete [6,8–
10], damage assessment of bridges [22–24] and buildings [25], ultimate deformation
capacity of RC columns [11], shear strength of circular RC columns [26], and compressive
strength [13–15] and the stress-strain relationship [12] of concrete confined with FRP.
ANN has also proved to be able to predict the shear capacity of pristine [35,39,65] and
FRP strengthened [66,67] RC beams. A review of the application of different artificial
intelligent techniques in structural engineering is conducted by Salehi and Burgueño [5].
Support vector machine (SVM) is another powerful ML technique that has been applied
to structural engineering problems [27–29]. Solhmirzaei et al. [29] used three different ML
techniques, namely, SVM, k-nearest neighbor (kN), and ANN to predict the failure modes
of Ultra High Performance Concrete (UHPC) beams. The results showed that ANN
outperformed the other two methods in predicting the failure modes of UHPC beams with
an overall accuracy of 89%.
MSc Thesis Page 45

Ensemble models are another type of emerging ML technique in the field of structural
engineering with high performance. These algorithms combine two or more learners
known as base learners (e.g. kernel regression, support vector machine, and decision tree)
in order to improve the prediction performance, robustness, and stability of the base
learners [30]. Ensemble learners can be formed in sequential or parallel styles with the aim
to exploit the dependence and independence between the base models, respectively.
Bagging (short for Bootstrap Aggregating) [31] is one of the ensemble models in which
multiple base learners are independently trained in parallel using different bootstrap
sample. Each bootstrap training dataset contains an average of 63.2% of the original
training set. The final prediction is then taken as the mean of the predictions from the base
learners [31]. As a result a better prediction with reduced variance is obtained [31].
Random forest (RF) is a popular example of bagging ensemble learners. In RF
homogeneous base learners, particularly, decision trees are used. An ensemble model can
also be formed from multiple base learners of different types or heterogeneous base
learners [30]. Stacking ensemble is one of the heterogeneous ensemble models in which
different types of base learners are combined via a meta-model with the objective of
enhancing the accuracy of the individual model [32,33].
In the previous study, different researchers have attempted to examine the efficacy of ML
models to estimate the capacity of RC beams in shear [34–39]. However, generally, a
limited number of experimental databases were used in the previous study. Moreover,
despite the investigation of the application of ML techniques to predict the capacity of RC
beams in shear, all previous studies generally failed to propose a user-friendly and practical
shear design approach based on the results of the trained ML models. As it is well
MSc Thesis Page 46

understood, ML algorithms are generally black-boxes and as such cannot be reproduced
by others. Thus, practical implementation and simplification of the ML-based models are
essential. Moreover, most of the models are developed using a limited number of
experimental datasets, which limits the generalization of the developed models.
In lieu of the above-mentioned limitations, the current study investigated the application
of several ML models (both single and ensemble models) to propose the best predictive
model based on a large database of slender and deep RC beams with/without stirrups. Four
types of homogeneous ensemble models that are based on decision trees; namely, random
forest, gradient boosting, extremely randomized trees, and extreme gradient boosting
(xgBoost) are considered. The xgBoost model, which is a comparatively new algorithm,
uses regularization parameters in order to overcome the overfitting problem [68]. It has
recently become popular across multiple disciplines [17,69–76].
Thus, this study adopted different advanced ensemble ML techniques, arguably for the
first time, to predict the shear capacity of RC beams. Furthermore, the prediction capability
of the ensemble models is compared with that of the base learners including support vector
regression and decision tree. The proposed model has been compared against the existing
models and code equations. The results of the analysis evidenced that the proposed model
is superior to other existing models and guidelines in predicting the RC beam shear
capacity. Finally, the best model among the developed models is deployed into a user-
friendly web-based application, which can be used by practitioners and researchers in the
field of civil engineering to accurately determine the capacity of RC beams in shear. The
developed web-based application facilitates a rapid and accurate prediction of the capacity
of RC beams in shear.
MSc Thesis Page 47

4.2 Normalization of dataset
To overcome the problems related to low rate of learning at the extreme values, the dataset
is normalized within a range of 0 and 1 as follow:
𝑥𝑜𝑟𝑖𝑔 − 𝑥𝑚𝑖𝑛
𝑥𝑛 = (25)
𝑥𝑚𝑎𝑥 − 𝑥𝑚𝑖𝑛
where 𝑥𝑜𝑟𝑖𝑔 is the original value of the variable, while 𝑥𝑛 is its corresponding normalized
value and 𝑥𝑚𝑖𝑛 and 𝑥𝑚𝑎𝑥 are the minimum and maximum values of the variable,
respectively.
4.3 Performance measurement
Different statistical indices are commonly used to evaluate the performance of ML-based
models. The following four indices are used in this study; namely, root mean squared error
(RMSE), mean absolute percentage error (MAPE), mean absolute error (MAE), and
coefficient of determination (𝑅 2 ). The performance indices are given as follows:
1
𝑀𝐴𝐸 = 𝑛 ∑𝑛𝑖=1|𝑦𝑖 − 𝑦̂𝑖 | (26a)
100 𝑛 𝑦𝑖 − 𝑦̂𝑖 (26b)

𝑀𝐴𝑃𝐸 = ∑ | |
𝑛 𝑖=1 𝑦𝑖
1 (26c)
𝑅𝑀𝑆𝐸 = √𝑛 ∑𝑛𝑖=1(𝑦𝑖 − 𝑦̂𝑖 )2
2
∑𝑛𝑖=1(𝑦𝑖 − 𝑦̂𝑖 )2 (26d)
𝑅 =1− 𝑛
∑𝑖=1(𝑦𝑖 − 𝑦̅)2
where 𝑦 is the target response, 𝑦̂ is the predicted response, and 𝑦̅ is the mean of the
responses.
MSc Thesis Page 48

4.4 Hyperparameter optimization
The performance of a given machine learning model depends on the values of its
hyperparameters, which are parameters that control the learning process of ML models. In
this study, the optimal hyperparameters are determined using grid search and K-fold cross
validation. Following the normalization of the dataset and identification of the input and
response vectors, the dataset is randomly divided into the training set and test set including
80% and 20% of the complete dataset, respectively, as shown in Figure 4.1.
Figure 4.1: 10-fold cross-validation.
The ML model is trained on the training set, while the test dataset is used to finally appraise
its performance. A standard K-fold cross-validation technique is adopted to evaluate the
model, where K denotes the number of partitions. Here, the data is randomly split into K
parts of equal sizes, and then the model is fitted on the 𝐾 − 1 parts, while the remaining
one part is used to validate the model, as shown in Figure 4.1 (for 𝐾 = 10). Therefore,
MSc Thesis Page 49

each fold is used as a validation set and the cross-validation is repeated K times. The
performance of the model is then determined as the average of the results from the K data
folds, as illustrated in Figure 4.1. This study adopted a ten-fold (𝐾 = 10) cross-validation,
as shown in Figure 4.1. The training of each model is performed using scikit learn [77],
which is a powerful ML packing in python programming language. Figure 4.2 summarizes
procedures followed in the development of the ML models.
Figure 4.2: model development
4.5 Single ML models
4.5.1 Support vector machine

Support vector machine is one of the supervised ML techniques with associated algorithms
primarily applied for classification problems using structural risk minimization principle
MSc Thesis Page 50

[78]. The SVM-based classifications are based on the optimal separation of classes. The
algorithm in the SVM finds a hyperplane or decision surface that distinctly classifies the
data points. It can also be used to efficiently perform both linear and non-linear regression
by indirectly mapping the original input vectors into a very high-dimensional feature space
in which they become separable, using kernel functions [78]. Different kernel functions
can be used for SVM with the radial basis function (RBF) kernel being the most popular
type. The RBF kernel, 𝐾𝑅𝐵𝐹 (𝑋1 , 𝑋1 ), on two vectors 𝑋1 and 𝑋2 is defined as follows:
𝐾𝑅𝐵𝐹 (𝑋1 , 𝑋1 ) = 𝑒𝑥𝑝(−𝛾‖𝑋1 − 𝑋2 ‖2 ) (27)
where 𝛾 = 0.5⁄𝜎 2 is a parameter that defines the spread of the kernel.
The other two types of nonlinear kernels used in SVM include sigmoid and polynomial
kernels defined as follow:
𝐾𝑠𝑖𝑔𝑚𝑜𝑖𝑑 (𝑋1 , 𝑋2 ) = tanh(𝛾〈𝑋1 , 𝑋2 〉 + 𝑟) (28)
𝐾𝑝𝑜𝑙𝑦𝑛𝑜𝑚𝑖𝑎𝑙 (𝑋1 , 𝑋2 ) = (𝛾〈𝑋1 , 𝑋2 〉 + 𝑟)𝑑 (29)
The above three nonlinear kernel functions; namely, polynomial, RBF, sigmoid kernels
are considered in this study. The other two hyperparameters that greatly affect the SVR
predictive capacity are the regularization parameter C and 𝜀-insensitive zone [79]. These
hyperparameters are also optimized in this study.
4.5.2 Decision tree

Decision tree also referred to as classification and regression tree (CART) is a supervised
ML algorithm that is similar to a flowchart-like structure. It is a tree-structured model with
three types of nodes; the root node, interior node, leaf node (terminal node) as showed in
Figure 4.3. The CART method splits the feature space into multiple smaller disjoint
MSc Thesis Page 51

regions with similar response values using a set of rules to predict a class label (in
classification) and value (in regression) of the response variable. Each internal node in
CART specifies a test on an attribute of the data, while each branch represents the test
output. The root node, which is the topmost node in CART denotes the most relevant
feature, while the leaf node or terminal node provides the predicted response variable.
Figure 4.3: Decision tree flow chart.
The performance of CART model can be optimized by tuning its hyperparameters
including the maximum depth of the tree, minimum samples number needed to be at a
leaf/terminal node, and minimum number of samples required to split an internal node.
Decision tree algorithm is easy to visualize and interpret; however, it has less
generalization ability, high bias, and high variance. Ensemble models can be used to
overcome the problems associated with a single decision tree as discussed below.
MSc Thesis Page 52

4.6 Ensemble Models
Ensemble learners are supervised machine learning paradigm that combine multiple single
learners (a.k.a. weak learners or base learners) into one predictive model to decrease
variance error, bias, and produce a strong model with enhanced generalization capability
and superior performance [80]. The basic idea behind the ensemble model is to combine
multiple base learners in computing the final response rather than relying on an individual
model. The ensemble models can be formed in either parallel or sequential manner.
Moreover, ensemble models can be formed from homogeneous base learners (same type
of base learner algorithms) or heterogeneous base learners (different types of base learner
algorithms).
The most popular type of meta-algorithms that combines base learners are bootstrap
aggregation (bagging) [31] and boosting [81] ensembles. In bagging ensemble (e.g.,
random forest), multiple base learners are independently trained in parallel mode on a
different bootstrap sample, while in boosting ensemble (e.g., gradient boosting) the base
models are trained sequentially.
4.6.1 Random forest
Random forest (RF) is a forest of randomly created CART models. It is the most used type
of tree-based ensemble. It randomly combines two or more decision trees to make a
decision [82]. However, it uses a bootstrap sample to train a series of decision trees and
make the final prediction unlike a single decision tree that is construct using the complete
training dataset [83]. It has emerged as a versatile and highly accurate methodology to
solve both classification as well as regression problems and with a capability to handle
MSc Thesis Page 53

large features with a small sample size and reduce the variance of prediction while keeping
low bias [82,84,85]. In this study, RF is combined with 10-fold cross-validation which
further acts to prevent overfitting.
Figure 4.4 shows the training process involved in random forest regression. Given training
dataset {(𝑥1 , 𝑦1 ), (𝑥2 , 𝑦2 ), … , (𝑥𝑛 , 𝑦𝑛 )} with 𝑛 observations, where 𝑥𝑖 𝜖 𝑋 =
〈𝑥𝑖,1 , 𝑥𝑖,2 , … , 𝑥𝑖,𝑗 〉 ⊆ 𝑅 𝑗 are 𝑗 independent variables and 𝑦𝑖 𝜖 𝑌 ⊆ 𝑅 is the response
parameter. The RF creates 𝑁 numbers of bootstrap sample (𝑆) and trains 𝑁 number of
decision trees each bootstrap sample, as can be observed in Figure 4.4. The final prediction
is then obtained as the average of the predictions from each decision tree, Figure 4.4.
Figure 4.4: Conceptual schematic of random forest.
MSc Thesis Page 54

4.6.2 Extremely Randomized Trees
An additional type of ensemble learner referred to as extremely randomized trees (ERT)
was proposed by Geurts et al. [86]. It can be applied to solve classification as well as
regression based problems. Extremely randomized trees differs from random forest in two
major features. Firstly, each decision tree in ERT is trained using the entire dataset in
contrast to random forest algorithm that uses a bootstrap sample. Secondly, ERT adds
randomization in selecting the split points of each node [86]. Similar type of
hyperparameters as that of random forest are used in extremely randomized trees.
However, the hyperparameters are independently tuned for both random forest and
extremely randomized trees.
4.6.3 Gradient boosting
Gradient boosting machine is a powerful boosting algorithm, which combines a sequence
of weak learners to generate an additive model whose performance is significantly
enhanced compared to the base learners [87]. Similar to other boosting algorithms,
gradient boosting regression (GBR) trains multiple learners sequentially. Given a training
examples 𝑋 = {(𝑥1 , 𝑦1 ), (𝑥2 , 𝑦2 ), … , (𝑥𝑛 , 𝑦𝑛 )}, the GBR fit the model 𝐹𝑚 (𝑥) using 𝑀 trees
[88]:
𝐹𝑚 (𝑥) = 𝐹𝑚−1 (𝑥) + ℎ𝑚 (𝑥) (30)

𝑛
ℎ𝑚 = arg min [∑ 𝐿(𝑦𝑖 , 𝐹𝑚−1 (𝑥𝑖 ) + ℎ(𝑥𝑖 ) )] (31)

ℎ
𝑖=1
where ℎ𝑚 is the 𝑚𝑡ℎ tree and 𝐹 represents set of all possible trees.
The objective of a newly added decision tree at each iteration is to reduce the loss given
by:
MSc Thesis Page 55

𝜕[𝑦, 𝐹𝑚−1 (𝑥)]

𝜗𝑚 = − (32)
𝜕𝐹𝑚−1 (𝑥)
Where 𝜗𝑚 is the loss function.
Figure 4.5: Conceptual schematic of gradient boosting.
4.6.4 Extreme gradient boosting
In 2016, Chen and Guestrin [68] proposed an enhanced form of gradient boosting
algorithm know as extreme gradient boosting (xgBoost). The main difference of the
xgBoost from the gradient boosting is in its objective function. The xgBoost algorithm
adds a regularization parameter to reduce complexity of the model and produce a better
generalization ability [68]. Its objective function (𝐽) is given as follows in which the
regularization term (𝛺) controls the complexity of the model [68]:
𝐽 = ∑ 𝐿(𝑦̂,𝑖 𝑦𝑖 ) + ∑ 𝛺(𝑓𝑚 )
𝑖 𝑚
(33)
The regularization parameter is determined as follows [68]:
MSc Thesis Page 56

1
𝛺(𝑓𝑚 ) = 𝛾𝑀 + 𝜆‖𝜔‖2 (34)
2
where 𝑀 is the number of decision trees, 𝛾 and 𝜆 are the regularization coefficients, and
𝜔 is the internal split tree weight.
MSc Thesis Page 57

CHAPTER 5 RESULTS AND DISCUSSION
5.1 RC beams without stirrup
The first step in the development of machine learning model is data collection and
preprocessing including normalization of dataset. As discussed in preview chapter, the
dataset is randomly partitioned into the train and test sets, which comprises 80% of the
dataset for training and 20% of the dataset for testing. The training set is used to train and
develop the models, while the test set is used to finally appraise the models based on
different performance indices. During the model development, the hyperparameters of
each model are tuned to find the most powerful and best model. The results of
hyperparameter tuning are summarized in Table 5.1.
Table 5.1: Optimized hyperparameters for each model for RC beams without stirrup
Models Optimum parameters
SVR Kernel type = radial basis function, C = 600, ε = 0.01
CART Maximum depth = 7, maximum features = 5, minimum sample leaf
= 1, minimum sample split = 3
RFR Number of estimators = 17, depth_max = 10, features_max = 5,
minimum sample leaf = 1, minimum sample split = 2
ERT Number of estimators = 33, maximum depth = 10, maximum features
= 8, minimum sample leaf = 1, minimum sample split = 2
GTBR Number of estimators = 62, learning rate = 0.15, depth_max = 7,
subsample = 0.3, features_max = 5, minimum sample split = 2
MSc Thesis Page 58

xgBoost Number of estimators = 180, learning rate = 0.15, maximum depth =
7, subsample = 0.3
The models are train in train dataset and tested in test dataset, as described in previous
chapter six models (two single models and three ensemble models) namely; support vector
machine, decision tree, random forest, extremely randomized tree, gradient boosting and
extremely gradient boosting are included in the study. The result from each model for both
train and test dataset in terms of performance indices; MAE, MAPE, RMSE, R2 for RC
beam without stirrup are discussed in Table 5.2.
Table 5.2 : Performance indices for RC beams without stirrup

Training dataset Test dataset
Model
MAE MAPE RMSE R2 MAE MAPE RMSE R2
SVR 30.081 0.359 56.378 0.872 29.100 0.382 56.274 0.836
CART 27.855 0.308 45.552 0.917 33.060 0.348 57.416 0.830
RFR 14.713 0.130 27.902 0.969 20.813 0.170 42.548 0.906
ETR 16.296 0.202 25.873 0.973 21.787 0.234 41.848 0.910
GBR 12.197 0.142 18.698 0.986 17.305 0.186 31.603 0.948
xgBoost 7.434 0.100 11.230 0.995 15.317 0.167 26.521 0.964
All ensemble models model showed higher prediction capability on the test dataset
compared to the base model (CART and SVM), as listed in Table 5.2. The evaluation of
the performance of the selected model showed a very small difference between the
experimental shear capacity and the predictions of the proposed xgBoost model in the
testing phases as indicated by the RMSE of 26.522 kN for the test dataset. For the GBR
MSc Thesis Page 59

model, this value was 31.603 kN. The RMSE for the other models on the test dataset
ranged between 41.848 kN to 57.416 kN. A similar trend was observed for other
performance indices, as listed in Table 5.2. This observation illustrates the excellent
prediction capacity of the proposed xgBoost model among all models, as listed in Table
5.2. Compared to all models, the CART model revealed the smallest performance with the
maximum root mean square error and least 𝑅 2 on both train and test dataset, as presented
in Table 5.2.
MSc Thesis Page 60

Figure 5.1: Comparisons of shear capacity predictions provided by the single ML models to
the experimental shear capacity for RC beams without stirrup.
MSc Thesis Page 61

MSc Thesis Page 62

Figure 5.2: Comparisons of shear capacity predictions provided by the ensemble ML

models to the experimental shear capacity for RC beams without stirrup.
MSc Thesis Page 63

Figures 5.1a–5.1b and Figures 5.2a–5.2d compare the shear capacity predicted by the
ML models against the corresponding experimental values. As can be seen in these figures,
the predictions provided by most of the models are in excellent agreement with the
experimental values. It can also be observed from the same figures that the xgBoost model
provided the best prediction for the load capacity of the beams, with the highest correlation
between the experimental and predicted shear capacity for both the training and test
datasets. A great correlation exists among the experimental and predicted shear capacities
based on the xgBoost model as evidenced by the value of 𝑅 2 (𝑅 2 = 0.99), as shown in
Fig. 5.2d and Table 5.2.
Figures 5.3a and 5.3f further compare the prediction capability of the models in terms of
the residuals of the predicted shear capacity of the beams, which is the difference between
the predicted shear capacity and the corresponding experimental value on the normalized
training and test datasets. The figures also provide the correlation coefficient for both the
training and test datasets. The residuals for all models are distributed around zero, as can
be seen in Figs. 5.3a–5.3f. In addition, the proposed models generally resulted in a good
correlation between the predicted and experimental load capaciies as can be evidenced
from the values of coefficient of determination, as can be seen in Figs. 5.3a–5.3f.
Compared to all models, xgBoost model showed the strongest correlation between the
experimental shear capacity and its predicted value, as can be observed in Fig. 5.3f.
MSc Thesis Page 64

(a)
(b)
(c)
MSc Thesis Page 65

(d)
(e)
(f)
Figure 5.3: Residual plots for RC beams without stirrup.
MSc Thesis Page 66

5.1.1 Comparison with existing design guide line for RC beams without stirrup
The accuracy of the proposed ML model; particularly, xgBoost model is compared with
that of the ACI 318 equation [55] and Eurocode 2. According to ACI 318 [55] and
Eurocode 2, the shear capacity (𝑉𝑛 ) of RC beam is evaluated as a simple superposition of
the capacity provided by concrete (𝑉𝑐 ) and transverse reinforcement (𝑉𝑠 ).
𝑉𝑛 = 𝑉𝑐 + 𝑉𝑠 (35)
Thus, for RC beams without stirrups, the nominal shear capacity (𝑉𝑛 ) is given by the shear
capacity provided by concrete (𝑉𝑐 ). It is determined by equation (2) and equation (4) for
AC1 318 and EC-2 respectively.
Table 5.3: Descriptive statics for 𝑽𝒑𝒓𝒆𝒅 /𝑽𝒆𝒙𝒑 for RC beams without stirrup
Model mean std min 25% 50% 75% max
ACI-
318 1.028372 0.231053 0.321645 0.938936 1.000643 1.07397 4.32554
EC-2 0.630559 0.344747 0.061932 0.41232 0.603127 0.773621 2.679892
xgBoost 0.526909 0.220215 0.055173 0.364288 0.561778 0.676812 1.331731
std: standard deviation, min: minimum, max: maximum
Figure 5.4 compares the predictive performance of the proposed xgBoost model with that
of the ACI 318 equation [55] and Eurocode 2. As can be observed from this figure, the
ACI and Eurocode 2 equation is inaccurate in predicting the shear capacity of the beams,
as opposed to the proposed model that showed high accuracy in predicting the shear
capacity of the beam. It can also be observed in Fig. 5.4 that the ACI 318 and Eurocode 2
provision significantly underestimates the shear capacity for most of the beams.
MSc Thesis Page 67

Figure 5.4: Predictions of ACI and proposed model for RC beams without stirrup.
Furthermore, the histogram in Figures. 5.5a –5.5c show the distribution of the predicted
to experimental shear capacity ratio using the proposed xgBoost model, ACI 318 code
equation and Eurocode 2. The proposed xgBoost model resulted in the most accurate and
stable predictions, as can be seen in Figures. 5.5a – 5.5c and table 5.3. The average (μ)
of the predicted to experimental shear capacity ratio was 1.03 for the proposed xgBoost
model compared with an average of 0.63 and 0.53 for ACI 318 code equation and
Eurocode 2 equation respectively. Furthermore, the standard deviation (σ) of 0.23 for
proposed model and 0.36 and 0.22 for the ACI 318 code equation and EC-2 respectively.
MSc Thesis Page 68

Figure 5.5: Histogram of predicted to experimental shear capacity ratio for RC beams
without stirrup.
Table 5.4: Performance indices for xgBoost, ACI-318 and EC-2

Model MAE MAPE (%) MSE RMSE R2
ACI 69.50109 44.24366 19111.12 138.243 0.149831
EC 76.73064 47.66816 19451.19 139.4675 0.183489
xgBoost 9.003983 11.2639 241.5076 15.54051 0.989865
Moreover, as showed in figure 5.13 and table 5.4 the root mean square error between
experimental and predicted shear capacity is 15.54 for xgBoost model when it is 138.23
MSc Thesis Page 69

for ACI-318, and 139.47 for EC-2. The coefficient of determination (R2) between
experimental and predicted shear capacity is 99% for proposed xgBooost model, when it
is 15% for ACI 138, and 18% Eurocode 2 equation. Compared to ACI-318 and EC-2 the
proposed xgBoost model showed high performance with the lowest RMSE and highest
correlation. Generally, it is concluded that the developed xgBoost model can reasonably
estimate capacity of RC beams in shear.
Figure 5.6 : Experimental versus predicted shear capacities for RC beams without stirrup
based on the proposed xgBoost model and existing code equations.
5.2 RC beam with stirrup
In chapter four the dataset is prepared, the total of 348 RC beam with stirrup are included
in the study. This section discusses the result from each models and comparative study
MSc Thesis Page 70
between proposed model, ACI 318 and EC-2. Similar to previous section, for RC beam
with stirrup the hyperparameters for single and ensemble ML models are optimized by
using 10-fold cross validation and presented in Table 5.5.
MSc Thesis Page 71

Table 5.5: Optimized hyperparameters of each models for RC beams with stirrup
Models Optimized hyperparameters
SVR Kernel type = radial basis function (rbf), ε = 0.1, C = 195
CART Maximum depth = 8, maximum features = 5, minimum sample leaf
= 1, minimum sample split = 2
RFR Number of estimators = 34, depth_max = 15, features_max = 5,
minimum sample split = 2
ERT Number of estimators = 29, depth_max = 10, maximum features = 7
GTBR Number of estimators = 55, depth_max = 8, rate of learn = 0.3, subs

ample = 0.3, maximum features = 6
xgBoost Number of estimators = 165, maximum depth = 9, rate of learn =

0.12, subsample = 0.4
Figure 5.7 shows a 7  7 matrix in which the diagonal of the matrix shows the histogram
for the distribution of each variable, whereas the lower and upper triangular matrices show
the scatter plot and Pearson correlation coefficient (𝑟) between the input variables,
respectively. As can be observed in this figure, there is complex relationship between the
parameters that influence shear capacity of RC beam.
MSc Thesis Page 72

Figure 5.7: Properties of dataset for RC beams with stirrups.
The total of 348 RC beams with stirrup are included in dataset, the dataset is randomly
split into 80% and 20% of test and train dataset respectively. The performance indices
regarding to the test and train dataset for each model (SVM, CART, RFR, ERT, GBR,
xgBoost) are evaluated and listed in Table 5.6. Compering the base models and ensemble
models, the ensemble models showed the highest performance to predict the shear capacity
of RC beam with RMSE of between 42.873 kN and 35.743 kN for test dataset, when it is
MSc Thesis Page 73

range between 112.416 kN and 55.851 kN for single or base models. From all models the
selected model showed a very small difference between the experimental shear capacity
and the predictions of the proposed xgBoost model in the testing phases as indicated by
the RMSE and 𝑅 2 of 35.743 kN and 0.995 respectively. A similar trend was observed for
other performance indices, as listed in Table 5.6. This observation illustrates the superior
prediction capability of the proposed xgBoost model among all models, as listed in Table
5.6. Compared to all models, the SVR model showed the lowest performance with the
highest root mean square error and least 𝑅 2 on both the test and train dataset, as presented
in Table 5.6.
Table 5.6: Performance indices for RC beams with stirrup

Training dataset Test dataset
Model
MAE MAPE RMSE R2 MAE MAPE RMSE R2
SVR 93.556 51.295 122.362 0.798 81.882 70.038 112.416 0.868
CART 19.083 7.628 32.958 0.985 37.435 15.471 55.851 0.967
RFR 17.800 6.252 32.836 0.985 29.044 12.509 42.873 0.981
ETR 10.897 4.167 19.794 0.995 26.710 11.447 38.341 0.985
GBR 12.219 6.055 19.825 0.995 25.058 17.046 32.698 0.989
xgBoost 6.425 2.513 11.823 0.998 27.036 14.009 35.743 0.987
MSc Thesis Page 74

Figure 5.8: Comparisons of shear capacity predictions provided by single ML models to the
experimental shear capacity for RC beam with stirrup.
MSc Thesis Page 75

MSc Thesis Page 76

Figure 5.9: Comparisons of shear capacity predictions provided by ensemble ML models

to the experimental shear capacity for RC beam with stirrup.
MSc Thesis Page 77

For RC beams with stirrup Figures 5.9a–5.9b compare the shear capacity predicted by
single ML models and the corresponding experimental values. From single model CART
models showed highest shear capacity predictions with R2 of 0.981. However Figures
5.10a–5.10d compare the shear capacity predicted by ensemble ML models and
corresponding experimental values. As can be observed in these figures, the predictions
provided by ensemble ML models are in excellent agreement with the experimental values.
It can also be observed from the same figures that the xgBoost model provided the best
prediction for the shear capacity of the beams, with the highest correlation between the
experimental and predicted shear capacity for both the test and train datasets. A solid
correlation exists between the experimental and predicted shear capacities based on the
xgBoost model as evidenced by the value of 𝑅 2 (𝑅 2 = 0.995), as shown in Figure. 5.10d
and Table 5.6.
5.2.1 Comparison with existing design guide line for RC beams with stirrup
The performance of the proposed xgBoost model ML model is compared with that of the
ACI 318 equation [55] Eurocode equation. According to ACI 318 [55] and Eurocode 2
design guideline equation, the nominal shear capacity (𝑉𝑛 ) of RC beam with stirrups is
evaluated as a simple superposition of the capacity provided by concrete (𝑉𝑐 ) and
transverse reinforcement (𝑉𝑠 ). It is evaluated by equation (1) and equation (7) respectively.
MSc Thesis Page 78

Table 5.7: Descriptive statics for 𝑽𝒑𝒓𝒆𝒅 /𝑽𝒆𝒙𝒑 for RC beams with stirrup
Model mean std min 25% 50% 75% max
ACI-
318 0.743987 0.446266 0.260748 0.583917 0.701994 0.807901 5.05479
EC-2 0.274128 0.258981 0.043825 0.173477 0.224379 0.305725 2.898384
xgBoost 1.009169 0.101862 0.743393 0.979759 1.000558 1.021289 2.065771
std: standard deviation, min: minimum, max: maximum
Figure 5.11 compares the predictive performance of the proposed xgBoost model with
that of the ACI 318 equation [55] and Eurocode 2. As can be observed from this figure,
the ACI equation and EC-2 equation are inaccurate in predicting the shear capacity of the
RC beam with stirrup, as opposed to the proposed model that showed high accuracy in
predicting the shear capacity of the beam. It can also be observed in Fig. 5.11 that the ACI
318 and Eurocode 2 provision significantly underestimates the shear capacity for most of
the beams.
MSc Thesis Page 79

Figure 5.10: Predictions of ACI, Eurocode and proposed model for RC beams with stirrup.
Moreover, the histogram in Figures. 5.12a –5.12c show the distribution of the predicted
to experimental shear capacity ratio using the proposed xgBoost model, ACI 318 code
equation and Eurocode 2. The proposed xgBoost model resulted in the most accurate and
stable predictions, as can be seen in Figures. 5.12a – 5.12c and table 5.7. The average (μ)
of the predicted to experimental shear capacity ratio was 1.01 for the proposed xgBoost
model compared with an average of 0.74 and 0.27 for ACI 318 code equation and
Eurocode 2 equation respectively. Furthermore, the standard deviation (σ) of 0.102 for
proposed model and 0.45 and 0.26 for the ACI 318 and EC-2 code equation respectively.
MSc Thesis Page 80

Figure 5.11: Histogram of predicted to experimental shear capacity for RC beams with
stirrup
Table 5.8 : Performance indices for xgBoost, ACI-318 and EC-2

Model MAE MAPE (%) MSE RMSE
ACI-318 111.6445 35.07028 44480.63 210.9043
EC-2 229.2035 75.60847 104698.1 323.5708
xgBoost 10.5715 4.825852 368.6531 19.20034
Furthermore, as showed in figure 5.13 and table 5.8 the root mean square error between
experimental and predicted shear capacity is 19.20 for xgBoost model when it is 210.90
MSc Thesis Page 81

for ACI-318, and 323.57 for EC-2. The value of R2 between experimental and predicted
shear capacity is 99.5% for proposed xgBooost model. Compared to ACI-318 and EC-2
the proposed xgBoost model showed high performance with the lowest RMSE and highest
R2. Generally, it can be concluded that the proposed xgBoost model can accurately predict
capacity of RC beams in shear.
Figure 5.12 : Experimental versus predicted shear capacities for RC beams with stirrup
based on the proposed xgBoost model and existing code equations.
MSc Thesis Page 82

5.3 Deployment of the ML model to a web-based application
The developed xgBoost model is deployed to a user-friendly web-based application. The

deployed model can be accessed at: https://shearcapaity.herokuapp.com/. Alternatively, it
can be accessed via the following QR code.
The deployed model is under development. It can be used for the reliable and accurate
shear design of RC beams. Particularly, it is of great interest to practitioners and designers
as it is user-friendly and superior in terms of its prediction capability compared to other
available models and code equations. The developed web-based application does not
require knowledge of the machine learning algorithms. This makes it very attractive to the
practitioners and researchers in the field of civil engineering. It can also be used for the
teaching purpose as a quick and accurate determination of the capacity of RC beams in
shear. The screenshot of the deployed model is shown below.
MSc Thesis Page 83

As can be seen in the screen shoot above, the input parameters are listed on the left hand
side. The user should define the input variables by scrolling under each input parameter.
Once the definition for the input variables are complete, the summary of the values of the
input parameters can be seen on the right hand side under “user defined variables”. The
capacity of RC beams in shear based on the defined variables can then be read on the same
screen. The developed application can be used in any device including mobile phones,
tablets, and computers. Hence, it is user-friendly to be used by any interested individual
without prior knowledge in the machine learning algorithms and also the theory behind
the shear mechanisms.
MSc Thesis Page 84

CHAPTER 6 CONCLUSIONS AND RECOMMENDATIONS
6.1 Conclusions
Despite several experimental studies aimed to understand the structural response of
reinforced concrete (RC) beams, accurate prediction of their shear capacity remains a
challenge. To this end, data-driven ensemble ML-based models to predict the shear
capacity of RC beams are presented in this paper. This study has shown that the ML model
can be used as effective tools to predict the shear capacity of RC beam. The proposed
models account for several input parameters that characterize the beam geometry, concrete
strength, and flexural reinforcement and transversal reinforcement. The prediction
capability of the ML-based models is compared with that of the ACI 318 equation and
Eurocode 2.
 The developed ML-based models are shown to be effective in predicting the shear
capacity of RC beams. Generally, all models showed good prediction capability
for RC beams with/withour stirrups.
 The ensemble models provided higher accuracy compared to the single models.
 Among ML models, xgBoost showed the highest prediction capability with the
lowest RMSE and highest correlation for the test dataset. The experimental shear
capacity and predicted values based on the xgBoost model showed the strongest
correlation with a coefficient of determination (𝑅 2 ) of 0.99 for RC beam without
stirrup and 0.995 for RC beam with stirrups.
 The comparisons of the proposed models with the existing code equation
confirmed the superiority of the xgBoost model. The ACI 318 and Eurocode 2
MSc Thesis Page 85

provision significantly underestimates the shear capacity of RC beams with and
without stirrups.
 The findings of this study showed the successful implementation of machine
learning techniques to predict the shear capacity of shear-critical RC beams.
 With an aim to develop an accurate and user-friendly shear design model, the
developed xgBoost model is deployed to a user-friendly web-based application:
https://shearcapaity.herokuapp.com/.
Future research is recommended to incorporate reliability study and other types of
reinforcement including fiber reinforced polymer (FRP) reinforcement. Moreover, it is
recommended to investigate the flexural response of RC beams using data-driven ML-
based models.
MSc Thesis Page 86

REFERENCES
[1] Baghi H, Barros JAO. New Approach to Predict Shear Capacity of Reinforced
Concrete Beams Strengthened with Near-Surface-Mounted Technique. ACI Struct
J 2016;114. https://doi.org/10.14359/51689433.
[2] Minelli F, Vecchio FJ. Compression Field Modeling of Fiber-Reinforced Concrete
Members Under Shear Loading. ACI Struct J 2006;103:244–52.
[3] ACI Committee 318. Building code requirements for structural concrete (ACI 318-
14). American Concrete Institute, Farmington Hills, MI, USA.; 2014.
[4] BS EN 1992-1-1. Eurocode 2: Design of concrete structures - Part 1-1 : General
rules and rules for buildings. vol. 1. Brussels, Belgium: European Committee for
Standardization; 2004.
[5] Salehi H, Burgueño R. Emerging artificial intelligence methods in structural
engineering. Eng Struct 2018;171:170–89.
https://doi.org/10.1016/j.engstruct.2018.05.084.
[6] Demir F. Prediction of elastic modulus of normal and high strength concrete by
artificial neural networks. Constr Build Mater 2008;22:1428–35.
https://doi.org/10.1016/j.conbuildmat.2007.04.004.
[7] Perera R, Ruiz A, Manzano C. An evolutionary multiobjective framework for
structural damage localization and quantification. Eng Struct 2007;29:2540–50.
[8] Lee S. Prediction of concrete strength using artificial neural networks. Eng Struct
2003;25:849–57. https://doi.org/10.1016/S0141-0296(03)00004-X.
[9] Chaabene W Ben, Flah M, Nehdi M. Machine learning prediction of mechanical
MSc Thesis Page 87

properties of concrete: Critical review. Constr Build Mater 2020;260:119889.
https://doi.org/10.1016/j.conbuildmat.2020.119889.
[10] Trocoli A, Dantas A, Leite MB, Nagahama KDJ. Prediction of compressive strength
of concrete containing construction and demolition waste using artificial neural
networks. Constr Build Mater 2013;38:717–22.
[11] Inel M. Modeling ultimate deformation capacity of RC columns using artificial
neural networks. Eng Struct 2007;29:329–35.
[12] Jiang K, Han Q, Bai Y, Du X. Data-driven ultimate conditions prediction and stress-
strain model for FRP-confined concrete. Compos Struct 2020;242:112094.
https://doi.org/10.1016/j.compstruct.2020.112094.
[13] Naderpour H, Kheyroddin A, Amiri GG. Prediction of FRP-confined compressive
strength of concrete using artificial neural networks. Compos Struct 2010;92:2817–
29. https://doi.org/10.1016/j.compstruct.2010.04.008.
[14] Jalal M, Ramezanianpour AA. Strength enhancement modeling of concrete
cylinders confined with CFRP composites using artificial neural networks. Compos
Part B Eng 2012;43:2990–3000.
https://doi.org/10.1016/j.compositesb.2012.05.044.
[15] Elsanadedy HM, Al-Salloum YA, Abbas H, Alsayed SH. Prediction of strength
parameters of FRP-confined concrete. Compos Part B Eng 2012;43:228–39.
https://doi.org/10.1016/j.compositesb.2011.08.043.
[16] Flood I. Towards the next generation of artificial neural networks for civil
engineering. Adv Eng Informatics 2008;22:4–14.

MSc Thesis Page 88
https://doi.org/10.1016/j.aei.2007.07.001.
[17] Mangalathu S, Jang H, Hwang S-H, Jeon J-S. Data-driven machine-learning-based
seismic failure mode identification of reinforced concrete shear walls. Eng Struct
2020;208:110331. https://doi.org/10.1016/j.engstruct.2020.110331.
[18] Keshtegar B, Nehdi ML, Trung N-T, Kolahchi R. Predicting load capacity of shear
walls using SVR-RSM model. Appl Soft Comput 2021:107739.
https://doi.org/10.1016/j.asoc.2021.107739.
[19] Hwang SH, Mangalathu S, Shin J, Jeon JS. Machine learning-based approaches for
seismic demand and collapse of ductile reinforced concrete building frames. J Build
Eng 2021;34. https://doi.org/10.1016/j.jobe.2020.101905.
[20] Mangalathu S, Hwang SH, Choi E, Jeon JS. Rapid seismic damage evaluation of
bridge portfolios using machine learning techniques. Eng Struct 2019;201.
https://doi.org/10.1016/j.engstruct.2019.109785.
[21] Mangalathu S, Sun H, Nweke CC, Yi Z, Burton H V. Classifying earthquake
damage to buildings using machine learning. Earthq Spectra 2020;36:183–208.
https://doi.org/10.1177/8755293019878137.
[22] Yan W, Deng L, Zhang F, Li T, Li S. Probabilistic machine learning approach to
bridge fatigue failure analysis due to vehicular overloading. Eng Struct
2019;193:91–9. https://doi.org/10.1016/j.engstruct.2019.05.028.
[23] Weinstein JC, Sanayei M, Asce M, Brenner BR, Asce F. Bridge Damage
Identification Using Artificial Neural Networks. J Bridg Eng 2018;23:04018084.
https://doi.org/10.1061/(ASCE)BE.1943-5592.0001302.
[24] Fathalla E, Tanaka Y, Maekawa K. Remaining fatigue life assessment of in-service
road bridge decks based upon arti fi cial neural networks. Eng Struct 2018;171:602–
MSc Thesis Page 89
16. https://doi.org/10.1016/j.engstruct.2018.05.122.
[25] Morfidis K, Kostinakis K. Approaches to the rapid seismic damage prediction of r/c
buildings using artificial neural networks. Eng Struct 2018;165:120–41.
[26] Caglar N. Neural network based approach for determining the shear strength of
circular reinforced concrete columns. Constr Build Mater 2009;23:3225–32.
[27] Jalal M, Arabali P, Grasley Z, Bullard JW, Jalal H. Behavior assessment, regression
analysis and support vector machine (SVM) modeling of waste tire rubberized
concrete. J Clean Prod 2020;273:122960.
https://doi.org/10.1016/j.jclepro.2020.122960.
[28] Pal M, Deswal S. Support vector regression based shear strength modelling of deep
beams. Comput Struct 2011;89:1430–9.
https://doi.org/10.1016/j.compstruc.2011.03.005.
[29] Solhmirzaei R, Salehi H, Kodur V, Naser MZ. Machine learning framework for
predicting failure mode and shear capacity of ultra high performance concrete
beams. Eng Struct 2020;224:111221.
[30] Tuv E. Ensemble learning. Stud Fuzziness Soft Comput 2006;207:187–204.
https://doi.org/10.1007/978-3-540-35488-8_8.
[31] Breiman L. Bagging predictors. Mach Learn 1996;24:123–40.
https://doi.org/10.3390/risks8030083.
[32] Wolpert D. Stacked Generalization. Neural Networks 1992;5:241–59.
[33] Zhou H, Huang G Bin, Lin Z, Wang H, Soh YC. Stacked extreme learning
MSc Thesis Page 90
machines. IEEE Trans Cybern 2015;45:2013–25.
https://doi.org/10.1109/TCYB.2014.2363492.
[34] Sanad A, Saka MP. Prediction of ultimate shear strenggth of reinforced concrete
deep beams using Neural Networks. J Struct Eng 2001;127:818–28.
[35] Cladera A, Mari AR. Shear design procedure for reinforced normal and high-
strength concrete beams using artificial neural networks. Part I: beams without
stirrups. Eng Struct 2004;26:917–26.
[36] Chou J, Ngo N, Pham A. Shear Strength Prediction in Reinforced Concrete Deep
Beams Using Nature-Inspired Metaheuristic Support Vector Regression 2012;30:1–
9. https://doi.org/10.1061/(ASCE)CP.1943-5487.0000466.
[37] Naderpour H, Mirrashid M. Shear strength prediction of RC beams using adaptive
neuro-fuzzy inference system. Sci Iran 2020;27:657–70.
https://doi.org/10.24200/sci.2018.50308.1624.
[38] Chou JS, Pham TPT, Nguyen TK, Pham AD, Ngo NT. Shear strength prediction of
reinforced concrete beams by baseline, ensemble, and hybrid machine learning
models. Soft Comput 2020;24:3393–411. https://doi.org/10.1007/s00500-019-
04103-2.
[39] Cladera A, Mari AR. Shear design procedure for reinforced normal and high-
strength concrete beams using artificial neural networks. Part II: beams with
stirrups. Eng Struct 2004;26:927–36.
[40] Friedman JH, Hastie T, Tibshirani R. The Elements of Statistical Learning. 2001.
https://doi.org/10.1111/j.1467-985x.2004.298_11.x.
MSc Thesis Page 91
[41] Collins MP, Bentz EC, Sherwood EG. Where is shear reinforcement required?
Review of research results and design procedures. ACI Struct J 2008;105:590–600.
https://doi.org/10.14359/19942.
[42] Zhang T, Visintin P, Oehlers DJ. Shear strength of RC beams with steel stirrups. J
Struct Eng 2016;142:04015135. https://doi.org/10.1061/(ASCE)ST.1943-
541X.0001404.
[43] Pathuri JR, Diaz M, Ph D, Montoya A, Ph D. Effect of longitudinal reinforcement
strength on shear strength of reinforced concrete beams and slabs by Presented to
the Graduate Faculty of The University of Texas at San Antonio in Partial
Fulfillment of the Requirements for the Degree of . MASTER OF. 2017.
[44] Fernández Ruiz M, Muttoni A, Sagaseta J. Shear strength of concrete members
without transverse reinforcement: A mechanical approach to consistently account
for size and strain effects. Eng Struct 2015;99:360–72.
[45] El-Ariss B. Behavior of beams with dowel action. Eng Struct 2007;29:899–903.
[46] Ormberg G. Evaluating Shear Capacity of Concrete Members with Deficient Shear
Reinforcement. 2010.
[47] ACI Committee 318. Building Code Requirements for Structural Concrete and
commentary (ACI 318–14). Am Concr Institute, Farmingt Hills, MI 2011:524.
[48] Tompos EJ, Frosch RJ. Influence of beam size, longitudinal reinforcement, and
stirrup effectiveness on concrete shear strength. ACI Struct J 2002;99:559–67.
[49] Bentz EC. Empirical modeling of reinforced concrete shear strength size effect for
members without stirrups. ACI Struct J 2005;102:232–41.

MSc Thesis Page 92
https://doi.org/10.14359/14274.
[50] Zararis PD, Zararis IP. Shear Strength of Reinforced Concrete Beams under
Uniformly Distributed Loads. ACI Struct J 2008;105:711–9.
[51] Hu B, Wu YF. Effect of shear span-to-depth ratio on shear strength components of
RC beams. Eng Struct 2018;168:770–83.
[52] Collins MP, Bentz EC, Sherwood EG. Where is shear reinforcement require?
Review of research results and design procedures. ACI Struct J 2009.
[53] Joint ACI-ASCE Committee 445. Recent approaches to shear design of structural
concrete. J Struct Eng 1998;124:1374–417. https://doi.org/10.1061/(ASCE)0733-
9445(1998)124:12(1375).
[54] Baghi H, Barros JAO. Design approach to determine shear capacity of reinforced
concrete beams shear strengthened with NSM systems. J Struct Eng
2017;143:4017061. https://doi.org/10.1061/(ASCE)ST.1943-541X.0001793.
[55] ACI Committee 318. Building Code Requirements for Structural Concrete and
commentary (ACI 318-14). American Concrete Institute, Farmington Hills, MI,
USA.; 2014.
[56] AASHTO. LRFD Bridge Design Specifications. 8th ed. Washington, DC:
American Association of State Highway Transportation Officials; 2017.
[57] Japan Society of Civil Engineers. Standard specifications for concrete structures
2007 “Design.” 2007.
[58] CSA Committee A23.3. Design of Concrete Structures (CSA A23.3-14). Can Stand
Assoc Mississauga, ON, Canada, 2004, 214 Pp n.d.
[59] Telkamp GJ. Shear Behavior of Reinforced Concrete Beams and Prestressed
MSc Thesis Page 93
Concrete Beams. Itinerario 1981;5:68–9.
[60] Vecchio FJ, Collins MP. The modified compression-field theory for reinforced
concrete elements subjected to shear. ACI J Proc 1986;83.
https://doi.org/10.14359/10416.
[61] Collins MP, Mitchell D, Adebar P, Vecchio FJ. A general shear design method. ACI
Struct J 1996;93:36–45. https://doi.org/10.14359/9838.
[62] Bentz EC, Vecchio FJ, Collins MP. Simplified Compression Field Theory for
Calculating Shear Strength of Reinforced Concrete Elements. ACI Struct J
2006;103:614–24. https://doi.org/10.14359/16438.
[63] Lee J-Y, Choi I-J, Kim S-W. Shear behavior of reinforced concrete beams with
high-strength stirrups. ACI Struct J 2011;108. https://doi.org/10.14359/51683219.
[64] Cladera A, Marí AR. Shear strength in the new Eurocode 2. A step forward? Struct
Concr 2007;8:57–66. https://doi.org/10.1680/stco.2007.8.2.57.
[65] Mansour MY, Dicleli M, Lee JY, Zhang J. Predicting the shear strength of
reinforced concrete beams using artificial neural networks. Eng Struct
2004;26:781–99. https://doi.org/10.1016/j.engstruct.2004.01.011.
[66] Abuodeh OR, Abdalla JA, Hawileh RA. Prediction of shear strength and behavior
of RC beams strengthened with externally bonded FRP sheets using machine
learning techniques. Compos Struct 2020;234:111698.
https://doi.org/10.1016/j.compstruct.2019.111698.
[67] Lee S, Lee C. Prediction of shear strength of FRP-reinforced concrete flexural
members without stirrups using artificial neural networks. Eng Struct 2014;61:99–
112. https://doi.org/10.1016/j.engstruct.2014.01.001.
[68] Chen T, Guestrin C. Xgboost: A scalable tree boosting system. In 22nd SIGKDD
MSc Thesis Page 94
Conference on Knowledge Discovery and Data Mining 2016.
[69] Rahman J, Ahmed KS, Khan NI, Islam K, Mangalathu S. Data-driven shear strength
prediction of steel fiber reinforced concrete beams using machine learning
approach. Eng Struct 2021;233. https://doi.org/10.1016/j.engstruct.2020.111743.
[70] Feng DC, Wang WJ, Mangalathu S, Hu G, Wu T. Implementing ensemble learning
methods to predict the shear strength of RC deep beams with/without web
reinforcements. Eng Struct 2021;235.
[71] Nguyen-Sy T, Wakim J, To QD, Vu MN, Nguyen TD, Nguyen TT. Predicting the
compressive strength of concrete from its compositions and age using the extreme
gradient boosting method. Constr Build Mater 2020;260.
https://doi.org/10.1016/j.conbuildmat.2020.119757.
[72] Nguyen HD, Truong GT, Shin M. Development of extreme gradient boosting model
for prediction of punching shear resistance of r/c interior slabs. Eng Struct
2021;235. https://doi.org/10.1016/j.engstruct.2021.112067.
[73] Mangalathu S, Shin H, Choi E, Jeon JS. Explainable machine learning models for
punching shear strength estimation of flat slabs without transverse reinforcement. J
Build Eng 2021;39. https://doi.org/10.1016/j.jobe.2021.102300.
[74] Machine learning-based prediction of CFST columns using gradient tree boosting
algorithm n.d.
[75] Marani A, Nehdi ML. Machine learning prediction of compressive strength for
phase change materials integrated cementitious composites. Constr Build Mater
2020;265. https://doi.org/10.1016/j.conbuildmat.2020.120286.
[76] Wong LS, Marani A, Nehdi ML. Gradient Boosting Coupled with Oversampling
MSc Thesis Page 95
Model for Prediction of Concrete Pipe-Joint Infiltration Using Designwise Data Set.
J Pipeline Syst Eng Pract 2021;12. https://doi.org/10.1061/(asce)ps.1949-
1204.0000557.
[77] Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, et al. Scikit-
learn: Machine Learning in Python. J Mach Learn Res 2011;12:2825–30.
[78] Cortes C, Vapnik V. Support-Vector Networks. Mach Learn 1995;20:273–297.
https://doi.org/10.1109/64.163674.
[79] Cherkassky V, Ma Y. Practical selection of SVM parameters and noise estimation
for SVM regression. Neural Networks 2004;17:113–26.
https://doi.org/10.1016/S0893-6080(03)00169-2.
[80] Sutton CD. Classification and Regression Trees, Bagging, and Boosting. Handb Stat
2005;24:303–29. https://doi.org/10.1016/S0169-7161(04)24011-1.
[81] Freund Y, Schapire RE. A Decision-Theoretic Generalization of On-Line Learning
and an Application to Boosting. J Comput Syst Sci 1997;55:119–39.
https://doi.org/10.1006/jcss.1997.1504.
[82] Fawagreh K, Gaber MM, Elyan E. Random forests: From early developments to
recent advancements. Syst Sci Control Eng 2014;2:602–9.
https://doi.org/10.1080/21642583.2014.956265.
[83] Svetnik V, Liaw A, Tong C, Christopher Culberson J, Sheridan RP, Feuston BP.
Random Forest: A Classification and Regression Tool for Compound Classification
and QSAR Modeling. J Chem Inf Comput Sci 2003;43:1947–58.
https://doi.org/10.1021/ci034160g.
[84] Breiman L. Random Forests. Mach Learn 2001;45:5–32.
[85] Segal M, Xiao Y. Multivariate random forests. WIREs DATA Min Knowl Discov
MSc Thesis Page 96
2011;1:80–7. https://doi.org/10.1002/widm.12.
[86] Geurts P, Ernst D, Wehenkel L. Extremely randomized trees. Mach Learn
2006;63:3–42. https://doi.org/10.1007/s10994-006-6226-1.
[87] Mining D. The Elements of Statistical learning - Springer Series in Statistics. Math
Intell 2009;27:83–85.
[88] Friedman JH. Greedy Function Approximation: A Gradient Boosting Machine. Ann
Stat 2001:1189–232. https://doi.org/10.1214/aos/1013203451.
MSc Thesis Page 97

Birhane Gemeda, Shear Capacity Prediction

Uploaded by

Copyright:

Available Formats

Birhane Gemeda, Shear Capacity Prediction

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Birhane Gemeda, Shear Capacity Prediction

Uploaded by

Copyright:

Available Formats

ADDIS ABABA UNIVERSITY

ADDIS ABABA INSTITUTE OF TECHNOLOGY

APPLICATION OF MACHINE LEARNING

By Birhane Gemeda Wakjira

Name of Student: Birhane Gemeda Wakjira

The proposed xgBoost model is deployed into a user-friendly web-based application to

TABLE OF CONTENTS .............................................................................................. VI

LIST OF TABLES ......................................................................................................... IX

LIST OF FIGURES ......................................................................................................... X

CHAPTER 1 INTRODUCTION ............................................................................... 1

1.1 Background .......................................................................................................... 1

1.2.2 Specific objective.......................................................................................... 6

1.3 Scope .................................................................................................................... 7

2.1 Shear strength of RC beam ................................................................................ 10

2.2.2 Transverse shear reinforcement .................................................................. 13

2.2.4 Longitudinal reinforcement ratio ................................................................ 13

2.2.5 Shear span-to-depth ratio ............................................................................ 14

2.3 Existing Shear Design Model and Guideline ..................................................... 14

2.3.2 Eurocode (EC2) [4]..................................................................................... 18

2.3.3 JSCE (2007) [57] ........................................................................................ 19

2.3.4 CSA A23.3-94 (Simplified Method) [58] ................................................... 20

2.3.5 Modified compression field theory............................................................. 21

3.1 Experimental database for RC beams without stirrups ...................................... 24

3.1.2 Distribution of input parameters ................................................................. 28

3.1.3 Effect of input parameters on the shear capacity ........................................ 32

3.2 Experimental database for RC beams with stirrups ........................................... 36

3.2.2 Distribution of input parameters ................................................................. 37

3.2.3 Effect of input parameters .......................................................................... 41

CHAPTER 4 MACHINE LEARNING-BASED MODELS .................................. 45

4.1 Introduction to Machine Learning ..................................................................... 45

4.5.2 Decision tree ............................................................................................... 51

4.6 Ensemble Models ............................................................................................... 53

4.6.2 Extremely Randomized Trees ..................................................................... 55

4.6.3 Gradient boosting........................................................................................ 55

4.6.4 Extreme gradient boosting .......................................................................... 56

CHAPTER 5 RESULTS AND DISCUSSION........................................................ 58

5.1 RC beams without stirrup .................................................................................. 58

5.2 RC beam with stirrup ......................................................................................... 70

5.3 Deployment of the ML model to a web-based application ................................ 83

6.1 Conclusions ........................................................................................................ 85

Table 2.1: Shear capacity of RC beams according to different models ............................ 15

This challenge lies in a complicated failure mechanism associated with shear-critical RC

understanding and safe design of reinforced concrete beams.

(dependent) variable (s) in a complex system. It is considered as a powerful technique to

MSc Thesis Page 1

of the applications of ML techniques reported in the literature include the prediction of

mechanical properties of concrete [6,8–10], load capacity and failure modes of RC

engineering is conducted by [5]. Support vector machine (SVM) is another powerful ML

performance compared to other techniques. These algorithms combine two or more

popular example of bagging ensemble learners. In RF homogeneous base learners,

MSc Thesis Page 3

based on a larger experimental data of 214 RC deep beams.

RC beams with stirrups.

methods (voting, bagging, stacking), and hybrid ML models are used.

Despite the investigation of the application of ML techniques to predict the capacity of RC

understood, ML algorithms are generally black-boxes and as such cannot be reproduced

MSc Thesis Page 4

experimental datasets, which limits the generalization of the developed models.