Engineering Structures 26 (2004) 681–691

www.elsevier.com/locate/engstruct

Simulating size effect on shear strength of RC beams

without stirrups using neural networks
Andres Winston C. Oreta

Department of Civil Engineering, De La Salle University, 2401 Taft Avenue, Manila, Philippines

Received 9 October 2003; received in revised form 12 January 2004; accepted 13 January 2004

Abstract

An artificial neural network (ANN) model was developed using past experimental data on shear failure of slender RC beams
without web reinforcements. The neural network model has five input nodes representing the concrete compressive strength (fc0 ),
beam width (b), effective depth (d), shear span to depth ratio (a/d), longitudinal steel ratio (q), five hidden layer nodes and one
output node representing the ultimate shear strength (vu ¼ Vu =bd). The model gives reasonable predictions of the ultimate shear
stress and can simulate the size effect on ultimate shear stress at diagonal tension failure. The ANN model performs well when
compared with existing empirical, theoretical and design code equations. Through the parametric studies using the ANN model,
the effects of various parameters such as fc0 , d, q and a/d on the shear capacity of RC beams without web reinforcement was
shown. This shows the versatility of ANNs in constructing relationships among multiple variables of complex physical processes
using actual experimental data for training.
# 2004 Elsevier Ltd. All rights reserved.

Keywords: RC beam; Diagonal shear; Size effect; Neural network; Modeling

1. Introduction Current design methods for shear in RC members

are based on almost entirely on the results of tests.
The prediction of the ultimate shear strength of rein- Majority of shear tests where the design equations were
forced concrete (RC) beams is critical especially when derived have been carried out on beams of depth less
this value is used in design. An unconservative value of than 300 mm [2]. However, more recent experiments on
shear capacity may lead to failure, a classic example is beams of larger depths have shown that as the size of
the partial collapse of the Wilkins Air Force Depot the beam increases, the intensity of shear stress decrea-
warehouse in Ohio, USA in August 1955 [1]. Investi- ses especially in lightly reinforced beams. This phenom-
gation showed that the beams, which failed in diagonal enon that is referred to as ‘‘size effect’’ simply means
tension, have depths of 914 mm (36 in.), at the failure that the shear strength is not constant for a given com-
location, did not contain stirrups, and had only 0.45% pressive strength of concrete but varies with the size of
longitudinal bars. Although the American Concrete the beam. This observation contradicts the assumption
Institute (ACI) design permitted an allowable shear given by some design codes like the simple equation for
stress of 0.62 MPa (90 psi) the beams failed at a lower shear of the ACI which assumes a constant value of
shear strength for a given compressive strength of con-
stress of 0.50 MPa (75 psi). A PCA test, on the other
crete [3]. This phenomenon of size effect gives one
hand, showed that a 12-in. beam could resist a shear
explanation for the collapse of some RC structures like
stress up to 1.0 MPa (150 psi) prior to failure. Why did
the Wilkins Air Force Depot warehouse. Design for
the beams fail in shear? shear using our present codes for some types of beams
may lead to unconservative results due to size effect.


Tel.: +63-2-52-44-6-11xloc. 226; fax: +63-2-52-40-5-63. Researches have been conducted to understand the
E-mail address: [email protected] (A.W.C. Oreta). shear problem, e.g. Refs. [4–10] and more recent
0141-0296/$ - see front matter # 2004 Elsevier Ltd. All rights reserved.
doi:10.1016/j.engstruct.2004.01.009
682 A.W.C. Oreta / Engineering Structures 26 (2004) 681–691

studies considered the shear effect, e.g. Refs. [1,2,11,12]. when the data available are ‘‘noisy’’ or incomplete.
Experiments have been conducted to understand shear One advantage of neural network modeling is that
failure of RC beams with and without stirrups for a there is no need to know a priori the functional
wide range of concrete strengths, steel ratios, effective relationship among the various variables involved,
depths, span lengths, subjected to both concentrated unlike in regression analysis. The ANNs automatically
and distributed loads. A database of some of these construct the relationships for a given network archi-
experimental results have been compiled and analyzed tecture as experimental data are processed through a
by researchers to verify a new theory or to develop a learning algorithm. This approach is ‘‘data driven’’,
new model on shear failure. Empirical equations of meaning that the network adopts to the training data
various forms, e.g. Refs. [1,12] were developed by presented to capture the relationship among input and
regression analysis of an assumed form of a function output parameters. For this reason, ANNs should be
relating different parameters which may affect the shear interesting to engineers and scientists as a tool to sup-
capacity of RC beams using a set of experimental data. port their task related to the modeling and prediction
Analytical equations based on a hypothesis of the of behavior of engineering and natural systems.
mechanism of shear failure, e.g. Zararis and Papadakis With the availability of more experimental data on
[11], have also been derived to estimate the critical diagonal shear of RC beams with a wider range of con-
shear intensity in terms of known parameters. Despite crete strengths, steel ratios, shear span to depth ratios,
the numerous studies over the last 50 years, the prob- and geometrical sizes, especially test data of large
lem of shear failure and the prediction of the shear beams, this study reanalyzes the new data to develop a
capacity of RC beams still remains an active research neural network model which will help us understand
area. Kani [5] states that the main obstacle to the shear more the diagonal shear problem and the different
problem is the large number of parameters involved, parameters which affect the shear capacity including
some of which may not be known. Krefeld and size effect. The present study aims to contribute to the
Thurston [7] explains that the major difficulties in continuing research about shear strength prediction of
developing a theoretical expression for the shearing RC beams using new computing technologies. An
strength of RC beams are due primarily to the inde- ANN model was developed to predict the ultimate
terminacy of the internal force system of a cracked shear strength of RC beams without stirrups using
reinforced member, the nonhomogeneity of concrete, available data from past experiments. The effects of
and the nonlinearity of its stress–strain diagram. various factors such as concrete strength, beam dimen-
Most of the empirical and analytical equations for sions, slenderness, longitudinal steel ratio and other
shear were developed using regression analysis of factors on the shear strength were considered. The per-
experimental data. To develop such models, the form formance of the ANN model was also compared with
of the empirical equation must be assumed and then existing empirical, theoretical and design code equa-
the unknown parameters in the equations are then tions. Hopefully, this study will contribute to a better
determined. The main problem with this approach is understanding of the influence of various beam para-
that it is difficult to determine the form and the num- meters on size effect of shear failure and eventually
ber of coefficients of the equation, which will best structural design provisions on shear of RC beams will
describe the physical process. With this approach, be updated leading to safer design of structures.
different expressions have been derived—ranging from
simple to complex, linear to nonlinear—depending on
the assumptions and details of the experimental data 2. Shear in slender RC Beams without web
used by the researchers. Because of the restriction of reinforcements
the assumed form of the equation, the model may not
be able to capture the interrelationship of the various 2.1. Failure mode
parameters considered in the model. The empirical
models of Zsutty [9] and Mphonde and Frantz [8], for Beams without web reinforcements or stirrups will
example, cannot capture the size effect, while the other fail when inclined cracking occurs. This type of failure
empirical equations by Collins and Kuchma [1] or referred to as diagonal tension failure (Fig. 1) pre-
Niwa et al. [12] usually perform well when tested using cipitates if the strength of the beam in diagonal tension
their own data , but perform poorly when applied to is lower than its strength in flexure. This behavior is
new data or data not used in the regression analysis. common in slender beams—beams with a/d between
Recently, researchers have found the potential of 2.5 and 5.5 or Lc =d between 11 and 16 [13]. The shear
artificial neural networks (ANNs) in the modeling of span, a, is the distance between the point of application
various engineering and natural systems. ANNs have of the concentrated load and the face of the support,
been found very powerful in modeling systems gov- Lc is the clear beam span if the loading is distributed,
erned by multiple variable interrelationships, especially and d is the effective beam depth. Diagonal shear fail-
 A.W.C. Oreta / Engineering Structures 26 (2004) 681–691 683

cements. Summarized in Table 1 are the ACI design

equations, some empirical and theoretical equations for
shear which consider the size effect. In this equations,
the value of fc0 is in MPa.
The ACI code defines the critical shear strength, Vc
or vc ð¼ Vc =bdÞ of a concrete beam as the stress at the
occurrence of the first inclined crack. The principal
diagonal tensile stress controlling the crack is the result
of the shearing stress due to the external factored shear
Vu and the horizontal flexural stress due to external
Fig. 1. Diagonal tension failure. bending moment Mu. The ACI code provides a simpli-
fied and detailed empirical model based on results of
extensive tests of failure of a large number of beams
ure starts with the development of a few fine vertical without web reinforcement in terms of Vu and Mu. The
flexural cracks at midspan, followed by the destruction quantity Vu d=Mu shall not be greater than 1.0 in com-
of the bond between the reinforcing steel and sur- puting Vc. The shear span to depth ratio, a/d, is often
rounding concrete at the support. Thereafter, without replaced by the quantity Mu =Vu d. The ACI equations
ample warning of impending failure, two or three diag- were formulated based on experimental tests—majority
onal cracks develop at about 1.5–2d distance from the of beams tested then have smaller depths. Hence, the
face of the support. As they stabilize, one of the diag- sensitivity of failure shear stress to size and reinforce-
onal cracks widens into a principal diagonal tension ment ratio was not recognized [1].
crack and extends to the top compression fibers of the Both empirical and theoretical equations were pro-
beam. This type of brittle failure mode occurs at a rela- posed to incorporate the size effect in the shear strength
tively small deflection. The shear capacity of the beam of RC beams without web reinforcements. Collins and
is considered to be equal to the inclined cracking shear. Kuchma [1] recommends modifications in the ACI
shear provisions by incorporating a crack spacing para-
2.2. Design for shear meter se which contributes to the size effect. The para-
meter se accounts for the influence of the crack spacing
RC beams are designed against shear failure by pro- sx and the maximum aggregate size, a. The parameter,
viding web reinforcement. The web reinforcement sx, is taken as 0.9d for members that have only con-
usually takes the form of vertical stirrups that enclose centrated reinforcement near the flexural tension face,
the longitudinal bars along the faces of the beam. Bent- or as the maximum distance between the layers of
v
up bars (usually at 45 angles) are another type of web longitudinal reinforcement if the member contains
reinforcement. The design for shear based on the ACI intermediate layers of crack control reinforcement.
code starts with the computation of the shear strength, A re-evaluation of the shear strength equation used
Vc, of the concrete beam cross-section and the in Japan was carried out by Niwa et al. [12] consider-
maximum shear, Vu, that the beam will resist. The ACI ing the result of large-sized beam tests and a new equa-
code assumes that Vc is equal to the shear strength of a tion was derived where the nominal shear strength at
beam without stirrups and this is taken equal to the failure, vu ð¼ Vu =bdÞ is inversely proportional to the
inclined cracking load. The required shear strength to fourth root of the effective depth.
be provided by the web reinforcement is then com-
puted as Vs ¼ ðVu  / Vc Þ=/. The required area of
web reinforcement or spacing of the stirrups can then
Table 1
be determined based on Vs. It must be observed that Shear strength equations for RC beams without web reinforcement
the value of Vc is critical in the design of shear reinfor- pffiffiffiffi
ACI vc ¼ 16 fc0
cing because if you overestimate the concrete shear
simplified
strength, then this will result to a theoretical area of equation [3]
 
stirrups which will be less than required or a spacing ACI detailed v ¼ 1 pffiffiffiffi pffiffiffiffi
fc0 þ 120q VMuud  vcmax ¼ 0:3 fc0
c 7
of stirrups which may be larger than the required equation [3]
theoretical spacing. Niwa et al. vu ¼ 0:20100q f 0 1=3 ðd 1=4 Þ½0:75 þ 1:4=ða=dÞ
c
[12]
 pffiffiffiffi
Collins and v ¼ 245 35sx
fc0 where se ¼ ðaþ16Þ
c 1275þse
2.3. Shear strength equations for RC beams without web Kuchma [1]
 
reinforcements Zararis and vu ¼ 1:2  0:2 da d dc fct
Papadakis    2=3
where 1:2  0:2 da d  0:65; fct ¼ 0:30 fc0
Various equations have been proposed to predict the [11]  c 2
and d þ600 fq0 dc  600 fq0 ¼ 0
shear strength of concrete beams without web reinfor- c c
684 A.W.C. Oreta / Engineering Structures 26 (2004) 681–691

A theory was presented by Zararis and Papadakis

[11] under which the diagonal shear failure of RC ordi-
nary (slender) beams without web reinforcement under
two-point loading (or one-point load at the middle of
the beam) is due to splitting of concrete occurring in a
certain region of the shear span. The theory results to
an expression where the shear stress, vu, at failure
depends on the ratio (c/d) of the neutral axis depth to
the effective depth of the beam times the splitting ten-
sile strength of concrete (fct). The equation has a cor-
rection factor to account for the size effect in slender
beams, according to which the size effect depends on
the size of the depth, d, and the ratio (a/d).

3. Neural network modeling

3.1. Applications of neural networks Fig. 2. A three-layer feedforward neural network.

ANNs are useful computing tools for which rules are

that will be received by the neighboring units. The fol-
either unknown or difficult to discover. No priori func-
lowing equations describe the mode of operation of a
tion is required before an ANN model can be
three-layer feedforward network:
developed. ANNs adapt solutions and are capable of !
capturing the interrelationships among multiple XM

variables by simply presenting them with data. Because hn ¼ f ðwx;n ix Þ þ bn ð1Þ

x¼1
of this capability of ANNs, applications to civil engin- !
eering, in general, and concrete structures, in parti- X
N

cular, have increased. ANN modeling have been oy ¼ f ðvn;y hn Þ þ by ð2Þ

n¼1
applied in the prediction of capacity of pin-ended RC
columns [14], modeling the confined strength and strain where ix, scaled input value transmitted from the xth
of circular columns [15], cost optimization of beam input neuron; hn, activity level generated at the nth hid-
design [16], deflection of externally reinforced RC den neuron; oy, activity level generated at the yth out-
beams [17], prediction of ultimate shear strength of put neuron; wx,n and vn,y, weights on the connections to
simply supported deep beams [18], and ordinary and the hidden and output layers of neurons, respectively;
high-performance concrete strength prediction [19,20]. bn and by, weighted biases and f [ ], activation function,
in this case, the sigmoid function f ðzÞ ¼ 1=ð1 þ ez Þ.
3.2. Network architecture and implementation
3.3. Experimental data
An ANN is a collection of simple processing units or
neurons connected through links called connections. The present research used 155 sets of data, majority
The topology or architecture of a three-layer feedfor- of which can be found from test results compiled by
ward neural network may be presented schematically, Zararis and Papadakis [11]. The database includes test
as in Fig. 2. The neural network is represented in the results from 1962 to 1999 for slender beams at various
form of a directed graph, where the nodes represent the strengths of concrete (low and high), steel ratios, shear
neuron or processing unit, the arcs represent the con- span to depth (a/d) ratios, and geometrical sizes
nections with the normal direction of signal flow is (of length and depth). The present study considered
from left to right. The processing units may be grouped test results of beams without web reinforcement under
into layers of input, hidden and output neurons. The concentrated loading (two-point loading or one-point
neural network in the figure consists of five input loading at midspan), and beams with concrete com-
neurons, five hidden neurons and one output neuron. pressive strength (fc0 ) up to 55 MPa (8000 psi) only.
The main tasks of neurons are to receive input from its The data were grouped randomly into two subsets—a
neighboring units which provide incoming activations, training set of 118 data and a testing or validation set
compute an output, and send that output to its of 37 data. The beam parameters available from these
neighbors receiving its output. The strength of the experiments are the concrete compressive strength (fc0 ),
connections among the processing units is provided by beam width (b), beam height (h), effective depth (d),
a set of weights that affect the magnitude of the input shear span to depth ratio (a/d), longitudinal steel ratio
 A.W.C. Oreta / Engineering Structures 26 (2004) 681–691 685

(q), shear force at failure (Vu). Other beam parameters models. It requires the selection of the input para-
can be easily derived (e.g. ratio, b/d and shear stress, meters and output parameters which will restrict the
vu ¼ Vu =bd). number of the input and output nodes of the network.
Table 2a and b show the statistical parameters of the Based on the experimental data, various combinations
two sets of data. The statistical parameters include the of input and output parameters can be considered in
mean, standard deviation, coefficient of variance, mini- developing the ANN models. Four types of archi-
mum value and maximum value. The statistics of both tectures based on input and output parameters were
training and testing sets are in good agreement mean- considered as summarized in Table 4. Models M1A
ing both represent almost the same population. and M1B have four input nodes representing fc0 , b/d,
a/d and q, while models M2A and M2B have five
3.4. Network data preparation input nodes representing fc0 , b, d, a/d and q. The ‘‘A’’
models are shear force (Vu) predictors while the ‘‘B’’
Preprocessing of data by scaling was carried out to models are shear stress (vu ¼ Vu =bd) predictors. A
improve the training of the neural network. To avoid model is labeled based on the number of input nodes,
the slow rate of learning near the end points specifi- hidden layer nodes and output nodes (e.g. M2B-551
cally of the output range due to the property of the sig- has five input nodes, five hidden layer nodes and one
moid function which is asymptotic to values 0 and 1, output node in Fig. 2).
the input and output data were scaled between ANN simulations were conducted for the different
the interval 0.1 and 0.9. The linear scaling equation: ANN architectures with the number of hidden layer
y ¼ ð0:8=DÞ x þ ð0:9  0:8xmax =DÞ was used in this nodes varied using 118 training data. The values of the
study for a variable limited to minimum (xmin) and learning parameter, momentum parameter, noise, num-
maximum (xmax) values given in Table 3 with ber of cycles were applied to a specific network archi-
D ¼ xmax  xmin . tecture and the behavior of the error is observed. In
most of the simulations, the values of 0.05–0.06 for the
3.5. Neural network simulations learning parameter, 0.008–0.01 for the momentum
parameter, 0.01 for noise were used. The stopping cri-
Using the experimental data on beam tests for shear, teria used are 500–4000 for total number of cycles or a
a three-layer feedforward neural network was value of 0.001 for the error tolerance. The different
developed using the backpropagation learning algor- ANN models were compared with respect to the fol-
ithm with momentum. Details of the backpropagation lowing error metrics, MAE, RMSE and R, using the 37
procedure can be found in the literature, e.g. Ref. [21]. test data.
The minimum number of hidden layers is considered, Table 5 presents the values of the error metrics for
and in this case one hidden layer is sufficient to pro- the various ANN models with varying hidden layer
duce an acceptable model. nodes. The comparison of error metrics is given for
Determining the network architecture is one of the both shear stress and shear force. For the shear force
most important tasks in the development of ANN predictors, M1A and M2A models, the shear stress is

Table 2
Statistics of experimental data

fc0 (MPa) b (cm) d (cm) h (cm) a/d q (%) b/d Vu (kN) vu (MPa)

(a) Training data

No. of data 118 118 118 118 118 118 118 118 118
Mean 29.31 17.99 34.41 38.99 3.83 1.83 0.67 74.57 1.23
Standard 6.53 9.69 26.05 28.92 1.50 0.68 0.41 66.01 0.28
deviation
COV 0.22 0.54 0.76 0.74 0.39 0.37 0.61 0.89 0.23
Min. value 10.50 5.00 7.00 8.00 2.35 0.59 0.14 7.30 0.56
Max. value 52.60 61.20 120.00 125.00 9.05 3.10 2.73 358.40 2.11

(b) Test data

No. of data 37 37 37 37 37 37 37 37 37
Mean 29.07 17.07 32.67 36.54 3.64 1.89 0.66 68.72 1.27
Standard 6.03 8.95 23.62 25.35 0.99 0.80 0.40 54.05 0.30
deviation
COV 0.21 0.52 0.72 0.69 0.27 0.42 0.60 0.79 0.24
Min. value 16.20 6.00 10.60 12.10 2.92 0.74 0.14 9.80 0.70
Max. value 41.70 61.10 109.70 122.00 7.00 3.36 2.25 235.00 1.84
686 A.W.C. Oreta / Engineering Structures 26 (2004) 681–691

Table 3 Table 5
Minimum and maximum values for scaling data Error metrics of ANN models

Variable Minimum Maximum ANN Shear stress (MPa) Shear force (kN)
fc0(MPa) 10.0 55.0 Model MAE RMSE R MAE RMSE R
b (cm) 5.0 65.0
M1A-441 1.413 2.837 0.425 28.649 34.244 0.805
d (cm) 7.0 120.0
M1A-451 1.159 2.164 0.353 27.949 32.145 0.807
b/d 0.1 2.8
M1A-461 1.280 2.410 0.346 30.361 36.399 0.802
a/d 2.5 9.5
q (%) 0.0 4.0
M1B-441 0.126 0.172 0.828 8.721 19.516 0.935
Vu (kN) 7.0 380.0
M1B-451 0.125 0.165 0.842 8.428 18.037 0.943
vu (MPa) 0.5 2.5
M1B-461 0.124 0.166 0.842 8.380 17.873 0.944

M2A-441 0.200 0.286 0.584 7.607 10.150 0.984

derived from the model output, Vu as vu ¼ Vu =bd. On M2A-451 0.200 0.315 0.476 7.104 10.110 0.983
the other hand, for the shear stress predictors, M2A M2A-461 0.164 0.235 0.723 6.605 9.490 0.985
and M2B models, the shear force is derived from the M2B-541 0.103 0.140 0.902 4.529 6.086 0.994
model output, vu, as Vu ¼ vu bd. The error metrics M2B-551 0.101 0.137 0.906 4.517 5.983 0.994
based on the predicted and derived values are then M2B-561 0.102 0.137 0.907 4.622 6.059 0.994
computed.
The best results of the M1A models for the shear
force predictions are MAE ¼ 27:949 kN, RMSE ¼ that M1B models yielded better estimates for shear
32:145 kN and a Pearson coefficient R¼ 0:807. These stress. M2B models performed best among the three
errors are relatively large. Moreover, the derived shear sets of ANN models. Comparing M1B and M2B mod-
stress for M1A models are not satisfactory with the els, it can be seen that the MAE and RMSE values for
best error metrics, MAE ¼ 1:159 MPa, RMSE ¼ the predicted shear stress were reduced by about 18%
2:164 MPa and R¼ 0:425. M1B models perform better and the R values significantly improved from about
than the M1A models. The minimum MAE and 0.84 to 0.91 for shear stress and from about 0.94 to
RMSE values for shear stress are 0.124 and 0.166, 0.99 for the derived shear force. A comparison among
respectively—about 10% of the M1A values—and the the four architectures show that the M2B models have
best R value is 0.842, almost twice that of the M1A the most superior performance with respect to the error
models. The R values for the derived shear force are metrics.
Based on the analysis of errors of the different ANN
also satisfactory with values greater than 0.90. The
models, the M2B-551 model had the minimum values
error values for the derived shear stress particularly on
for MAE and RMSE both for shear stress and shear
the R values for M1A models show that deriving the
force and is thus selected for further validation. The
shear stress from the predicted shear force produces
model has to be tested on its performance with respect
very poor predictions. A shear stress predictor model is
to other empirical models and its ability to simulate
more superior than a shear force predictor model. real physical processes is also verified by parametric
M2A models with five input nodes performed better studies.
than the M1A models with the errors drastically
reduced—a reduction of about 75% for the MAE and
RMSE and improved values of R greater than 0.98.
4. Performance of the neural network model
This shows that an ANN model with five input nodes
and the parameters b and d, represented independently 4.1. Connection weights
as in M2A models and not combined as one para-
meter, b/d in M1A models, performed better. Compar- The M2B-551 model (Fig. 2) has five input nodes
ing M2A and M1B models, on the other hand, show representing the concrete compressive strength (fc0 ),
beam width (b), effective depth (d), shear span to depth
ratio (a/d), longitudinal steel ratio (q), five hidden
Table 4
ANN architecture and model parameters
layer nodes and one output node representing the ulti-
mate shear strength (vu ¼ Vu =bd). The connection
Model No. of input Input parameters Output parameter weights for this model are shown in Table 6. Zero
nodes
biases were set in the developing the model for sim-
M1A 4 fc0 , b/d, a/d, q Vu plicity. Using a causal inference procedure by Garson
M1B 4 fc0 , b/d, a/d, q vu [22], the relative importance of the input variables to
M2A 5 fc0 , b, d, a/d, q Vu
the output can be determined. It can be seen that the
M2B 5 fc0 , b, d, a/d, q vu
effective depth of the beam, d, is the most important
 A.W.C. Oreta / Engineering Structures 26 (2004) 681–691 687

Table 6
Connection weights of M2B-551 model

Hidden nodes Input layer Output Layer

fc0 b d a/d q

1 0.841836 0.520036 2.295805 0.740027 0.576550 1.944814

2 0.136782 3.614227 2.695393 0.805690 3.364672 5.030095
3 0.244926 0.276289 0.023932 0.961188 0.461074 0.790688
4 2.127035 1.374138 0.223555 1.701900 1.371270 2.199725
5 1.386961 3.659177 6.639413 0.337491 1.268700 5.171329
Rel. impt. (%) 14.48 21.26 25.20 19.78 19.29

parameter for the model, followed by the beam width, over, the model of Collins and Kuchma is for the pre-
b, and the shear span to depth ratio, a/d. The para- diction of Vc, the shear stress at inclined cracking not
meters b and d are related to the shearing area while at ultimate failure.
the shear span to depth ratio (a/d) is related to the
beam slenderness which influences mode of failure.
4.4. Comparison with ACI design code equations
Using the computed weights in Table 6, the perform-
ance of the M2B-551 model can be validated with Fig. 6 presents the M2B-551 model predictions com-
respect to errors and its ability to capture the inter-
pared with the three ACI equations for the test data.
relationship of the different parameters used in the
model as will be shown in the succeeding sections.

4.2. Model predictions

The predictions of the M2B-551 model as compared

to the experimental values are shown in Fig. 3 for both
training and test data. The Pearson product moment
correlation coefficient R for training data is 0.9125 and
for test data is 0.9058. The M2B-551 model has a mean
percentage error of about 7.6% and a maximum per-
centage error of about 20% for the test data. The histo-
gram showing the distribution of the percentage error
for both training and test data is shown in Fig. 4.
About 76% of the predictions of the ANN model lie
within the 10% error. There are a few large misses
which have an error outside the 10% range.

4.3. Comparison with empirical and theoretical

equations

The M2B-551 model predictions for the test data are

compared with the predictions of the empirical equa-
tions by Collins and Kuch (assuming a maximum
aggregate size of 25 mm), Niwa et al. and the theoreti-
cal equation by Zararis and Papadakis. Fig. 5 presents
the comparison of the predictions of the neural net-
work model with the different models which consider
size effect. The M2B-551 model predictions lie above
and below the target line—the line where the predicted
value is equal to the experimental value. The M2B-551
model with an R equal to 0.9058 compares well with
Niwa et al. (R¼ 0:9283) and Zararis and Papadakis
(R¼ 0:9315). The model by Collins and Kuchma, how-
ever, has an R¼ 0:6891—this low value may be due to Fig. 3. Neural network shear stress predictions (a) training data,
the assumed value of the maximum aggregate. More- (b) test data.
688 A.W.C. Oreta / Engineering Structures 26 (2004) 681–691

Fig. 4. Histogram of percent error of shear stress predictions.

The ACI simplified and detailed equations gave con-

servative predictions in most cases while the ACI
maximum equation overestimate the shear stress at
failure in almost all cases. Fig. 7, on the other hand, Fig. 6. Comparison of shear strength predictions with ACI code.
compares the prediction of the M2B-551 model and the
ACI detailed equation for the shear strength with
respect to the effective depth of the beam for both experimental data especially in the region of larger
training and test data. The ACI predictions at smaller depths and at high strength concrete.
depths are conservative by as much as 50% in some
cases and the predictions at larger depths are greater
than the experimental values overestimating by as 5. Parametric studies on size effect
much as 40% of the experimental values. The M2B-551 One advantage of neural network models is that
model, on the other hand, gave predictions lying at parametric studies can be easily done by simply varying
1:0 20% of experimental values, covering the range
one input parameter and all other input parameters are
from shallow to larger depths of the beam. If the
set to constant values. Through parametric studies, we
neural network model will be used in design, then an
can verify the performance of the M2B-551 model in
appropriate factor of safety may be applied to the pre-
simulating the physical behavior of an RC beam with-
dicted value. The value of the factor of safety, how-
out web reinforcement particularly the size effect with
ever, has still to be evaluated since the neural network
respect to varying parameters. The results of para-
model can still be improved by considering more
metric studies which follow trends that are consistent
with experimental results will be our qualitative evi-
dence that the ANN model has learned to reproduce
the physical process of shear failure of RC beams with-
out web reinforcement.

Fig. 7. Ratio of ANN and ACI predictions to experimental shear

Fig. 5. Comparison of shear strength predictions. stress.
 A.W.C. Oreta / Engineering Structures 26 (2004) 681–691 689

5.1. Size effect and the parameters fc0 and a/d

Can the neural network model capture the size effect

in shear stress at failure? Fig. 8 presents the predicted
shear stress of the M2B-551 model for varying
effective depth, d, for a compressive strength of con-
crete, fc0 ¼ 28 MPa (4000 psi). The shear span to depth
ratio (a/d) was varied at 2.5–5.0. The other input para-
meters were set to constant values with b¼ 15 cm and
q¼ 2:75%. Superimposed in the figure are the data
from Kani [5] for the tests with the following range of
values: fc0 (24–30 MPa), b (15.1–17.7 cm), a/d (2.35–
6.05) and q(2.59–2.87%). The size effect is clearly
demonstrated in the figure where the shear stress Fig. 9. Size effect for a=d¼ 5:0 and varying d and fc0 .
decreases as the depth increases. The decrease in shear
stress is almost 50% from d¼ 20 to d¼ 100 cm. The decreased by about 50% from 0.85 to 0.4 MPa, while
shear stress at failure is not constant. The decreasing the shear stress for q¼ 3:0% decreased only by about
trend of the shear stress with increasing depth is con- 20% from 1.31 to 1.04 MPa. The value of the
sistent with the experimental results by Kani [5]. The shear strength using the ACI simplified equation for fc0
shear capacity also depends on the shear span to depth ¼ 21 MPa (3000 psi) is also shown in the figure as 0.76
ratio—a shorter beam (a=d¼ 2:5) has a larger shear MPa. Observed that the ACI shear strength values
stress than a longer beam (a=d¼ 5:0). at larger depths for the lightly reinforced beams
Fig. 9 present the behavior of the shear stress at fail- (q¼ 0:5% and 1.0%) are greater than the neural net-
ure for three values of compressive strength of concrete work predictions. Size effect is more pronounced in
at a=d¼ 5:0, respectively. The neural network model lightly reinforced beams.
can realistically simulate the effect of the compressive
strength of concrete. The beam with a higher compress- 5.3. Varying compressive strength of concrete
ive strength (e.g. fc0 ¼ 35 MPa) has a larger shear
capacity than a beam with a lower compressive What is the behavior of the shear strength with vary-
strength (e.g. fc0 ¼ 21 or 28 MPa). ing compressive strength of concrete? Fig. 11 shows an
almost linear relationship between the shear stress and
5.2. Size effect and longitudinal steel ratio fc0 with the other parameters constant. The Pearson
correlation coefficient between Vu =bd and fc0 is equal to
Is the size effect influenced by the longitudinal steel 1.0 for all four curves. The shorter beam (a=d¼ 2:5)
ratio? Fig. 10 shows the effect of the amount of longi- with a higher longitudinal steel ratio (q¼ 2:75%) has
tudinal steel on the size effect. It is observed that the the largest shear capacity among the four cases shown
rate of decrease of the shear stress for the lightly rein- in the figure.
forced beams (e.g. q¼ 0:5% and 1.0%) is larger than Is the relationship between the shear strength and
for beams with higher longitudinal steel ratio (e.g. q the compressive strength of concrete influenced by the
¼ 2:0% and 3.0%). The shear stress for q¼ 0:5%

Fig. 8. Size effect for fc0 ¼ 28 MPa and varying d and a/d. Fig. 10. Size effect and longitudinal steel ratio.
690 A.W.C. Oreta / Engineering Structures 26 (2004) 681–691

6. Conclusion

An ANN model with five input nodes, five hidden

layer nodes and one output node was developed for
predicting the shear stress at failure of slender RC
beams without web reinforcements. The M2B-551
model gave predictions with an error of 20% of
experimental values, covering the range from shallow
to larger depths of the beam. Although there is still a
need to improve the accuracy of the predictions, the
present ANN model successfully simulates the effects of
various parameters on the shear capacity of RC beams
without web reinforcement. This shows the versatility
Fig. 11. Predicted shear stress and effect of fc0 , a/d and q.
of ANNs in constructing relationships among multiple
variables of complex physical processes using actual
other parameters? Fig. 12 also shows a similar relation- experimental data for training.
ship which is almost linear between shear stress and Through parametric studies, the behavior of shear
concrete compressive strength, for a=d¼ 5:0. The rate stress, Vu =bd at failure due to various factors is
of increase of the shear strength is influenced by observed. A beam with a higher compressive strength
the depth, d. The shear strength of the RC beam with
of concrete, fc0 , has a larger shear capacity than a beam
d¼ 20 cm increased about 63% from 0.87 (at
fc0 ¼ 10 MPa) to 1.42 MPa (at fc0 ¼ 55 MPa), while the with a lower compressive strength. However, the
shear strength of the RC beam with d¼ 100 cm increase in the shear capacity with respect to fc0 is
increased only by about 16% from 0.43 (at found to be significant in beams with smaller depths.
fc0 ¼ 10 MPa) to 0.50 MPa (at fc0 ¼ 55 MPa. The The size effect is clearly demonstrated where the shear
increase in the shear capacity with respect to fc0 is stress decreases with the increase in depth. The shear
significant in beams with smaller depths. stress at failure is not constant. The decreasing trend of
It is also interesting to note that the Pearson the shear stress with increasing depth is consistent with
correlation coefficients R between Vu =bd and fc0 for the experimental results. The shear capacity also
a=d¼ 5:0 are about 1.0, indicating a linear relationship depends on the shear span to depth ratio—a shorter
for this case. However, the R values decrease with
beam has a larger shear stress than a longer beam. The
increasing depth—R¼ 1:0 for d¼ 20 cm, R¼ 0:99 for
effect of the amount of longitudinal steel on the size
d¼ 50 cm and R¼ 0:94 for d¼ 100 cm for a=d¼ 5:0.
The effective depth of the beam and the shear span to effect is also observed. The rate of decrease of the shear
depth ratio may have an influence on the relationship stress for the lightly reinforced beams is larger than for
between Vu =bd and fc0 . A linear relationship R between beams with higher longitudinal steel ratio.
Vu =bd and fc0 is observed for beams with smaller If an ANN model which has wide application is
depths. desired, there should be an increase of the training
database which is distributed and varied. Only when
we have sufficient number of data that we can develop
ANNs which can completely model the complex inter-
actions among the multiple variables. The present
ANN model can still be improved and extended by
using more experimental data with a wider range of
values—for example, extending the data up to a=d <
2:5 to include deep beams which fail under shear com-
pression and including higher strength concrete. More
refined neural network training algorithms can also be
used to reduce the error of the models. This study
shows the potential of ANNs as a tool to support the
task of civil and structural engineers in the modeling
Fig. 12. Predicted shear stress for a=d¼ 5:0 and varying fc0 and prediction of behavior of engineering and natural
and d. systems.
 A.W.C. Oreta / Engineering Structures 26 (2004) 681–691 691

Acknowledgements [11] Zararis PD, Papadakis GC. Diagonal shear failure and size effect
in RC beams without web reinforcement. J Struct Eng ASCE
This research was supported by the University 2001;127(7):733–42.
[12] Niwa J, Yamada K, Yokozawa K, Okamura H. Revaluation of
Research Coordination Office (URCO) of the De La
the equation for shear strength of reinforced concrete beams
Salle University—Manila, Philippines. without web reinforcement. Concr Lib JSCE 1987;9:65–84.
[13] Nawy EG. Reinforced concrete: a fundamental approach. New
Jersey: Prentice-Hall Inc; 1996. p. 155–207.
References [14] Chuang PH, Goh ATC, Wu X. Modeling the capacity of pin-
ended slender reinforced concrete columns using neural net-
works. J Struct Eng ASCE 1998;124(7):830–8.
[1] Collins MP, Kuchma D. How safe are our large, lightly rein-
[15] Oreta A, Kawashima K. Neural network modeling of confined
forced concrete beams, slabs, and footings? ACI Struct J
compressive strength and strain of circular concrete columns.
1999;96(4):482–90.
J Struct Eng ASCE 2003;129(4):554–61.
[2] Taylor HPJ. Shear strength of large beams. J Struct Div ASCE
[16] Hadi MNS. Neural networks applications in concrete structures.
1972;98(ST11):2473–90.
Comput Struct 2003;81:373–81.
[3] American Concrete Institute. Building code requirements for
[17] Flood I, Muszynski L, Nandy S. Rapid analysis of externally
concrete (ACI 318-95). Detroit: ACI; 1995.
reinforced concrete beams using neural networks. Comput Struct
[4] Kani GNJ. The riddle of shear failure and its solution. ACI J
2001;79:1553–9.
1964;April:441–66.
[18] Sanad A, Saka MP. Predictions of ultimate shar strength of RC
[5] Kani GNJ. How safe are our large RC beams. ACI J
concrete beams using neural networks. J Struct Eng ASCE
1967;March:128–41.
2001;127(7):818–28.
[6] Krefeld WJ, Thurston CW. Contribution of longitudinal steel
[19] Kasperkiewicz J, Racz J, Dubrawski A. HPC strength prediction
to shear resistance of reinforced concrete beams. ACI J
using artificial neural networks. J Comput Civ Eng ASCE
1966;March:325–42.
1995;9(4):279–84.
[7] Krefeld WJ, Thurston CW. Studies of the shear & diagonal
[20] Lee SC. Prediction of concrete strength using artificial neural
tension strength of simply supported RC beams. ACI J
networks. Eng Struct 2003;25:849–57.
1966;April:451–75.
[21] Freeman JA, Skapura DM. Neural networks: algorithms, appli-
[8] Mphonde AG, Frantz GC. Shear tests of high and low strength
cations, and programming techniques. Reading, MA: Addison-
concrete beams without stirrups. ACI J 1984;July–August:350–7.
Wesley; 1991. p. 89–125.
[9] Zsutty T. Beam shear strength prediction by analysis of existing
[22] Garson GD. Interpreting neural network connection weights. AI
data. ACI J 1968;65(11):942–51.
Expert 1991;April:47–51.
[10] Zsutty T. Shear strength prediction for separate categories of
simple beam tests. ACI J 1971;February:138–43.