Construction and Building Materials 23 (2009) 2214–2219

Contents lists available at ScienceDirect

Construction and Building Materials

journal homepage: www.elsevier.com/locate/conbuildmat

Neural networks for predicting compressive strength of structural light

weight concrete
Marai M. Alshihri a, Ahmed M. Azmy b, Mousa S. El-Bisy a,*
a
Civil Engineering Dept., College of Engineering and Islamic Architecture, Umm Al_Qura University, Makkah, Saudi Arabia
b
Civil Engineering Dept., Higher Technological Institute, Ramadan Tenth City, Egypt

a r t i c l e i n f o a b s t r a c t

Article history: Neural networks procedures provide a reliant analysis in several science and technology fields. Neural
Received 13 April 2008 network is often applied to develop statistical models for intrinsically non-linear systems because neural
Received in revised form 18 November 2008 networks behave the advantages of simulating complex behavior of many problems. In this investigation,
Accepted 1 December 2008
the neural networks (NNs) are used to predict the compressive strength of light weight concrete (LWC)
Available online 4 January 2009
mixtures after 3, 7, 14, and 28 days of curing. Two models namely, feed-forward back propagation (BP)
and cascade correlation (CC), were used. The compressive strength was modeled as a function of eight
Keywords:
variables: sand, water/cement ratio, light weight fine aggregate, light weight coarse aggregate, silica fume
Neural networks
Back propagation
used in solution, silica fume used in addition to cement, superplasticizer, and curing period. It is con-
Cascade correlation cluded that the CC neural network model predicated slightly accurate results and learned very quickly
Compressive strength as compared to the BP procedure. The finding of this study indicated that the neural networks models
Light weight concrete are sufficient tools for estimating the compressive strength of LWC. This undoubtedly will reduce the cost
and save time in this class of problems.
Ó 2008 Elsevier Ltd. All rights reserved.

1. Introduction methods. Mathematical models based on experimental data are

Lightweight aggregates are broadly classified in to two types:
natural (pumice, diatomite, volcanic cinders, etc.) and artificial
(perlite, expanded shale, clay, slate, sintered PFA, etc.). Lightweight
aggregates can be used to produce low density concretes required
for building applications like cladding panels, curtain walls, com-
posite flooring systems, and load-bearing concrete blocks [1,2].
Structural lightweight concrete has its obvious advantages of
higher strength to weight ratio, better tensile strain capacity, lower
coefficient of thermal expansion, and superior heat and sound
insulation characteristics due to air voids in the lightweight aggre-
gate. Also the reduction in the dead weight of the construction
materials, by the use of lightweight aggregate in concrete, could re-
sult in a decrease in cross-section of concrete structural elements
(columns, beams, plates, and foundation). It is also possible to re-
duce steel reinforcement [3,4].
Most of the mathematical models used to study the behavior of
concrete mixes consist of mathematical rules and expressions that
capture relationship between components of concrete mixes. By
the way, using mathematical models to take and describe experi-
ences from experimental data of concrete mixes behaviors are
most reliable, accurate, scientific, and applicable recommended
called ‘‘free models”, and generally are in regression forms. How-
ever, if the problem contains many independent variables, regres-
sion methods cannot be used because of less accuracy and more
assumptions in regression form (linear, non-linear, exponential,
etc.). In the recent years, new modeling techniques such as artifi-
cial neural networks, expert systems as a free model can approxi-
mate non-linear and complex relation due to any phenomena and
trial and error process by learning real record relationship without
any presumptions [5,6].
Free from such limitations, The NN approach (a class of soft
computing techniques), has been used recently in predictive mod-
els for the field of materials engineering [7–19] because of their
specific features such as non-linearity, adaptively (i.e., learning
from inputs parameters), generalization, and model independence
(no priori models needed).
The present study is aimed to evaluate NN models for predic-
tion of 3, 7, 14, and 28 days compressive strength of a LWC
mixtures.
2. Materials and methods
2.1. Data (mix design) collocation

The main objective of this study is developing a neural network model to pre-
* Corresponding author. Tel.: +966 500518353. dict the compressive strength of LWC. For this aim, at first it is needed to prepare
E-mail address: melbisy@gmail.com (M.S. El-Bisy). data and construct data base for training and testing the neural network model.

0950-0618/$ - see front matter Ó 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.conbuildmat.2008.12.003
 M.M. Alshihri et al. / Construction and Building Materials 23 (2009) 2214–2219 2215

The present study covers the use of crushed hollow block as lightweight coarse
aggregate in concrete containing silica fume as supplementary cementations mate-
rials at different levels namely: 0%, 5%, 10%, and 15% as an addition to cement. The
crushed hollow block aggregate was treated by solution of silica fume and calcium
hydroxide (the concentration of silica fume is 10% and 20% and that of calcium
hydroxide is 1%). The performance of lightweight concrete made with crushed hol-
low block as coarse aggregate was studied in terms of compressive strengths for 3,
7, 14, and 28 days. The main variables, experimental program, are described and re-
ported in Table 1.
Mix proportions of concrete were designed to select suitable materials (cement,
fine aggregate, coarse aggregate, water, etc.) and determine the quantities of these
ingredients for meeting the desired compressive strength. The economy and perfor-
mance characteristics of the concrete product, thus depends on the proportioning of
theses ingredients. The procedures adopted for mix proportioning are still empirical
in spite of a considerable amount of work done on the theoretical aspects of mix
proportioning of normal weight and lightweight concretes.
2.2. Material properties
Crushed hollow block was used as lightweight aggregate. In this investigation,
blocks were broken manually. ASTM D-75 and ASTM C-136 and C-29 were used for
sampling, grading, unit weight, and fineness modulus of aggregate. Maximum aggre-
gate size was 10 mm. The fine aggregate was confirmed to ASTM C-33 requirements.
A locally produced ordinary Portland cement was used in this investigation. The
cement content weighed for about 450, 400, and 350 kg/m3, respectively. The water
has been used for mixing and curing of all concrete mixes and specimen’s was clean,
fresh, and free from any impurities.
Silica fume is a fine powder which acts as a microscopic concrete pore filler. It is
based on a chloride free pozzolanic material consisting of over 90% silicon dioxide.
Its addition to the cement mix will yield a concrete especially able to cope with the
middle eastern environment. The physical and chemical properties of the silica
fume used in this investigation are reported in Tables 2 and 3, respectively.
Superplasticizer is a chloride free, superplasticing admixture based on selected
sulphonated naphthalene polymers. It is supplied as a brown solution which in-
stantly disperses in water. Superplasticizer disperses the fine particles in the concrete
mix, enabling the water content of the concrete to perform more effectively. The very
high levels of water reduction possibly allow major increases in strength that is to be
obtained. It was added with dosage 1%, 2%, 3%, and 4% by weight of cement.
The concrete specimens were 150  150  150 mm (6  6  6 in.) cubes to
determine the compressive strength of the above mentioned lightweight aggregate
concrete. All specimens were immersed in water tank for 24 h after casting (after
finishing initial setting time) to complete hydration reaction through different peri-
ods of time. These curing periods of time are 3, 7, 14, and 28 days. At the end of cur-
ing period, the specimens were removed from water tank and placed it to dry and
were left it under the sun rays for at least 6 h and, then tested. Compressive
strength tests were performed by universal testing machine, equipment that mea-
sures the compressive and splitting bending strengths directly according to the
ASTM specifications ASTM C-39.
2.3. Model induction from experimental data
A neural network consists of many simple elements called neurons which are
grouped together in layers. A neuron has many inputs and a single output. Each in-
put has a coefficient, referred to as a weight, assigned to it. A neuron works in the
following way: inputs of the neuron are multiplied by the corresponding weights.
The product is then summed together and applied to a transfer function, to form
the output. This can be expressed using the following equation:
!
X
n
z¼f w i xi þ d ð1Þ
i¼1
where z is the output from neuron; x1, x2, ...., xn are the input values; w1, w2, ...., wn are
the connection weights; d is the bias value; and f is the activation function.
Artificial neural network can be visualized as a set of interconnected neurons
arranged in layers. The input layer contains one neuron for each of the input vari-
ables. In multi layer network, the output of one layer constitutes the input to the
next layer. For example, in the ANN architecture shown in Fig. 1, this is called
feed-forward type of network where computations proceed along the forward
direction only. The neural network has one input layer, one output layer, and two
hidden layers. The output obtained from the output neurons constitutes the net-
work output.
The connection weights and bias values are initially chosen as a random num-
bers and then fixed by the results of training processes. Many alternative training
processes are available namely back propagation (BP) and cascade correlation
(CC) schemes. The goal of any training algorithm is to minimize the mean square
error (MSE) between predicted outputs and observed outputs (in the training data-
set) and maintaining good generality of the networks. The generality (network per-
formance) is assessed by testing a new dataset. A reasonably good learning process
can be achieved by choosing an appropriate network configuration with regard to
the number of hidden layers and their hidden neurons.
2.3.1. Back propagation algorithm
The back propagation learning is an iterated search process which adjusts the
weights from output layer back to input layer in each run until no further improve-
ment in MSE value is found. The BP algorithm calculates the error, and then used to
adjust the weights first in the output layer, and then distributes it backward from
the output to hidden and input nodes (Fig. 2). This is done using the steepest gra-
dient descent principle where the change in weight is directed towards negative
of the error gradient, i.e.
oE
Dwn ¼ aDwn1  g ð2Þ
ow
where w is the weight between any two nodes; Dwn and Dwn1 are the changes in
this weight at n and n  1 iteration; a the momentum factor; and g is the learning
rate.
After training is completed, the final connection weights are kept fixed, and new
input patterns are presented to the network to produce the corresponding output
consistent with the internal representation of the input/output mapping.
2.3.1.1. Topology of the BPNN. Every stage of any NN project requires a little trial and
error to establish a suitable and stable network for the project. Trial and error may
be extended to building several networks, stopping, and testing the network at dif-
ferent stages of learning and initializing the network with different random
weights. Each network must be tested, analyzed and the most appropriate network
must be chosen for a particular project.
Before deciding on the topology of the network, it is important to select the re-
quired number of input and output parameters. The function of the hidden layer
neurons is to detect relationships between network inputs and outputs. There is

Table 1
Mix proportions.

Sand Cement Water Light weight coarse Light weight fine Normal weight fine Silica fume
replacement (%) (kg) (kg) aggregate (kg) aggregate (kg) aggregate (kg)
Used in Used in addition to
solution cement
a. Cement content = 450 kg/m3
0% 29.250 19.013 21.571 33.786 0 1.901 2.925
25% 29.250 19.013 21.571 25.340 12.995 1.901 2.925
50% 29.250 19.013 21.571 16.893 25.989 1.901 2.925
75% 29.250 19.013 21.571 8.447 38.984 1.901 2.925
b. Cement content = 400 kg/m3
0% 26.325 17.112 21.571 33.786 0 1.711 2.632
25% 26.325 17.112 21.571 25.34 12.995 1.711 2.632
50% 26.325 17.112 21.571 16.893 25.989 1.711 2.632
75% 26.325 17.112 21.571 8.447 38.984 1.711 2.632
c. Cement content = 350 kg/m3
0% 23.034 14.972 21.571 33.786 0 1.711 2.303
25% 23.034 14.972 21.571 25.34 12.995 1.711 2.303
50% 23.034 14.972 21.571 16.893 25.989 1.711 2.303
75% 23.034 14.972 21.571 8.447 38.984 1.711 2.303

Mix A: contains first five columns; Mix B: contains first six columns; Mix C contains all seven columns.
2216 M.M. Alshihri et al. / Construction and Building Materials 23 (2009) 2214–2219

Table 2 2.3.2. Cascade correlation algorithm

Physical properties of the silica fume. The cascade correlation (CC) is another network that modifies its own architec-
ture as training progresses. CC network starts with a minimal topology, consisting
Appearance Fine powder only of the required input and output units. This net is trained until no further
Bulk density (kg/m ) 3
300–600 improvement is obtained. The error for each output until is then computed.
Surface area (m2/kg) 18,000–22,000 Next, one hidden unit is added to the net in a two-step process. During the first
Chloride content Nill to BS 5075 step, a candidate unit is connected to each of the input units, but is not connected to
Ignition loss (%) <3% the output units. The weights on the connections from the input units to the candi-
Flammability Non-flammable date unit are adjusted to maximize the correlation between the candidate’s output
and the residual error at the output units. The residual error is the difference be-
tween the target and the computed output, multiplied by the derivative of the out-
put unit’s activation function, i.e. the quantity that would be propagated back from
Table 3
the output units in the back propagation algorithm. When this training is completed,
Chemical compositions of the silica fume.
the weights are frozen and the candidate unit becomes a hidden unit in the net.
Oxides SiO2 Fe2O3 Al2O3 MgO CaO Na2O K2O The second step in which the new unit is added to the net now begins. The new
hidden unit is then connected to the output units, and the weights on the connec-
(%) >90 <1.5 <1.0 <1.0 <1.0 <0.5 <0.5
tions being adjustable. Now all connections to the output units are trained.
A second hidden unit is then added using the same process. However, this unit
receives an input signal from both the input units and the previous hidden unit. All
weights on these connections are adjusted and then frozen. The connections to the
output units are then established and trained. The process of adding a new unit,
training its weights from the input units and the previously added hidden units,
and then freezing the weights, followed by training all connections to the output
units, is continued until the error reaches an acceptable level or the maximum
number of epochs (or hidden units) is reached.
For a more detailed discussion of the learning algorithms, the interested reader
is referred to Hoehfeld and Fahlman [24], Adams and Waugh [25] and Wu [26].

Output 2.4. Pre-processing of data

Layer

Hidden NNs have been shown to be able to process data from a wide variety sources.
Layer (2) They are however, only able to process the data in a certain format. Furthermore
Input the way the data is presented to the network affects the learning of the network.
Layer Hidden Therefore, a certain amount of data processing is required before presenting the
Layer (1) training patterns to the network. In this study a linear scaling was used.
One of the reasons for pre-processing the output data is that a sigmoidal transfer
Fig. 1. Structure of an artificial neural network. function is usually used within the network. Upper and lower limits of output from a
sigmoid transfer function are generally 1 and 0, respectively. Scaling of the inputs to
the range [1, +1] greatly improves the learning speed, as these values fall in the re-
gion of sigmoid transfer function where the output is most sensitive to variations of
desired the input values. It is therefore recommended to normalize the input and output data
x1 ej Output before presenting them to the network. Scaling data can be linear or non-linear,
w1j Error
depending on the distribution of the data. Most common functions are linear and
logarithmic functions. A linear normalization function within the values of 0 to 1 is

x2 S ¼ ðV  V min Þ=ðV max  V min Þ ð3Þ

w2j
z Activation where S is the normalised value of variable; V, Vmin, and Vmax are variable minimum
Neuron
function Output and maximum values, respectively.
j

2.5. Preparation of training and validation data set

Threshold
xj ANNs are data intensive. NNs learn the underlying physics of the system of
wij
interest from the training samples, which are basically the cause-effects samples.
Therefore, the number of training samples significantly influences the network’s
Fig. 2. Neuron weight adjustments. predictive performance [27]. Increasing the number of training samples provides
more information about the shape of the solution surface(s) and thus increases
the potential level of accuracy that can be achieved by the network. Having too
no general rule for selecting the number of neurons in hidden layer. The choice of few data samples will lead to poor generalization by the network. An optimal data
hidden layer size is mainly problem specific and to some extent depends on the set for training would be the one that fully represents the modeling domain and has
number and the quality of training patterns. The number of neurons in a neural net- the minimum number of repetitive samples (i.e. identical inputs with different out-
work must be sufficient for the correct modeling of the problem, but it should be puts) in training. Since nearly 70% of the whole data were randomly chosen for
sufficiently low to ensure generalization [20]. Too large number of hidden neurons model calibration (training) and rest 30% were kept for model validation.
would encourage over-fitting, i.e. modeling the noise in the data or modeling the
data with unnecessarily complex functions. In general, a smooth interpolation be- 2.6. Model assessment
tween training data is sought.
Some authors relate the number of neurons to the number of input and output To examine how close the predicted to the compressive strength of HSC mix-
variables and the number of training patterns [20,21], but these rules cannot be tures, four indices, mean absolute error MAE, root mean square error RMSE, corre-
generalized. Some researches have mentioned that the upper bound for the re- lation coefficient R, and coefficient of efficiency Ef, were employed to evaluate the
quired number of neurons in the hidden layer should be one greater than twice performance of neural networks based on mixtures properties
the number of input units. This rule does not guarantee generalization of the net-
work. To be able to design stable BP network, it would be more appropriate to carry 1X N
MAE ¼ jPi  Oi j ð4Þ
out a parametric study by changing the number of neurons in the hidden layer in N i¼1
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
order to test the stability of the network. u
u1 X N
RMSE ¼ t ðPi  Oi Þ2 ð5Þ
N i¼1
2.3.1.2. Ratio of first to second hidden layer neurons. A single layer with an optimum
number of neurons will be sufficient for modeling many practical problems. It has PN  



i¼1 P i  P Oi  O
been suggested that for continuous functions a single hidden layer with a sufficient R ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi ð6Þ
PN  
 2 PN Oi  O

 2
number of neurons will be suitable whilst a second layer may be needed for discon- i¼1 P i  P i¼1
" #," #
tinuous functions [22]. Maier and Dandy [23] have shown that the ratio of 3:1 be- Xn X n X
n
Ef ¼ ðOi  OÞ  2 ðP i  Oi Þ2  2
ðOi  OÞ ð7Þ
tween first and second hidden layer neurons yields lowest root mean square error
i¼1 i¼1 i¼1
RMSE compared to other combination.
 M.M. Alshihri et al. / Construction and Building Materials 23 (2009) 2214–2219 2217

0.1 50 40
Train Error
Test Actual
0.08 40 Predicted 35

Compressive Strength (MPa)

30
0.06 30
MSE

25
0.04 20

Error (%)
20
0.02 10
15
0 0
0 4 8 12 16 20 24 28 10
Number of hidden neurons
-10 5
Fig. 3. Performance of BP with different number of hidden neurons.
-20 0
1 6 11 16 21 26 31
Data No.
where Oi is the compressive strength of LWC mixtures, Pi is the predicted value, N is
 is the mean value of observations,
the total number of data points in validation, O
Fig. 5. Comparison of predicted and actual values of 28 days compressive strength
and P is the mean value of predictions.
of tested data for BPNN (8–14–6–4) model.

3. Analysis of results and discussion

performance of the BP networks net work with various number
The compressive strength of LWC mixes are greatly influenced of neurons in one hidden layer. In Fig. 3 the performance of BP net-
by several parameters: sand, light weight coarse aggregate, light works with various number of neurons in one hidden layer is mea-
weight fine aggregate, water/cement ratio, and the percentage of sured by the MSE. MSE is the vertical axis of the graph. In this
silica fume used in solution and used in addition to cement. Conse- network, as can be seen from Fig. 3, 14 neurons in the one hidden
quently developing a standard mix design procedure for LWC layer results in a stable and optimum network. The compressive
mixes requires an extensive understanding of the relation between strength prediction efficiencies of almost all cases of a three layer
these parameters and the properties of the resulting matrix. The
input for LWC models has been chosen as sand S, water/cement ra-
tio WC, light weight fine aggregate LFA, light weight coarse aggre- 40
gate LCA, silica fume used in solution SFS, silica fume used in
Predicted Compressive Strength (MPa)

addition to cement SFC, superplasticizer Su, and curing period CP, 35

and the outputs are strength of LWC after 3, 7, 14, and 28 days.
The data were collected from only one source so aggregate quality, 30
cement characteristics, curing conditions, and loading speed were
assumed to be constant. After identifying the parameters that have 25
a significant effect on the compressive strength of LWC mixtures to
develop NN-based model: the input layer with eight neurons and 20
output layer with four neurons. The input layer includes S, WC,
LFA, LCA, SFS, SFC, Su, and CP. The output layer is compressive 15
strength of LWC after 3, 7, 14, and 28 days.
To design a stable BPNN based on the determined input num- 10
ber, the parametric study carried out by changing the number of 10 15 20 25 30 35 40
Actual Compressive Strength (MPa)
neurons in the hidden layers (one hidden layer and two hidden lay-
ers) in order to test the stability of the network. Fig. 3 shows the Fig. 6. Scatter of forecasted and experimental values of 28 days compressive
strength of trained data for BPNN (8–14–6–4) model.

50 40
Error
Actual 40
40 Predicted 35
Predicted Compressive Strength (MPa)
Compressive Strength (MPa)

30 35
30

25 30
Error (%)

20
20
25
10
15
20
0
10

-10 15
5

-20 0 10
1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 10 15 20 25 30 35 40
Data No. Actual Compressive Strength (MPa)

Fig. 4. Comparison of predicted and actual values of 28 days compressive strength Fig. 7. Scatter of forecasted and experimental values of 28 days compressive
of trained data for BPNN (8–14–6–4) model. strength of tested data for BPNN (8–14–6–4) model.
2218 M.M. Alshihri et al. / Construction and Building Materials 23 (2009) 2214–2219
Table 4
Performance of NN models.
ANN model RMSE (MPa) Maximum (%)
a. Training data
BP (8-14-4) 3.875 9.721
BP (8-14-6-4) 2.492 5.45
CC (8-6-4) 2.363 5.057
b. Test data
BP (8-14-4) 3.812 9.132
BP (8-14-6-4) 2.325 5.23
CC (8-6-4) 2.289 5.01
BPNN having different neurons in the hidden layer were found
more than 86%. Overall, it could easily be observed that BPNN hav-
ing architecture 8–14–4 (eight neurons in input and 14 neurons in
first hidden layers, and four neurons in output layer) for one hid-
den layer case produced the best result in this study.
A further test on whether additional second hidden layer could
improve the network performance was carried out. In this test, the
number of 14 neurons in the first hidden layer was fixed and var-
ious number of neurons in the second hidden layer were used.
BPNN having a structure 8–14–6–4 (eight neurons in input and
14 neurons in first hidden layers, six neurons in second hidden
layer and four neurons in output layer) produced optimum result
for the available inputs. Findings of this study were very similar
to that found by Maier and Dandy [23] ‘‘to yield lowest RMSE com-
pared to other combination, a ratio between first and second hid-
den layer nodes should not exceed of 3:1”.
For the developed four-layer BPNN (8-14-6-4) model men-
tioned earlier, the results of trained and tested instances for the
compressive strength prediction are shown in Figs. 4 and 5. The
predicted results are fairly close to the corresponding actual mea-
surements. Figs. 6 and 7 represent the scatter diagram of predicted
and experimental values of compressive strength for trained and
tested instances. The prediction can be seen as fairly close to the
corresponding actual values of compressive strength. For trained
data, it can be observed that a maximum absolute error of 5.45%,
a minimum absolute error of 0.061% and the mean absolute error
of 2.22% were obtained for compressive strength prediction. Also
for tested data, it can be observed that a maximum absolute error
of 5.23%, a minimum absolute error of 0.053% and the mean abso-
lute error of 1.987% were obtained for compressive strength pre-
diction. The correlation coefficients of 0.972 and 0.977 were
obtained for the training and testing data of compressive strength
prediction. The results indicate that the 3, 7, 14, and 28 days com-
pressive strength of light weight concrete (LWC) mixtures can be
predicted using the four-layer BPNN (8–14–6–4) model based on
the sand, light weight coarse aggregate, light weight fine aggregate,
water, cement, and the percentage of silica fume used in solution
and used in addition to cement.
The strength of LWC based on BP scheme was repeated by using
CC scheme of training. This was done to see whether the training
improves by adopting a different algorithm. Results of the testing
of this network are shown in Table 4 and they indicated almost
same performance of the CC training method (maximum error of
5.057%, a minimum error of 0.0341%, and the average error of
1.927%). For trained data, it can be observed that a maximum abso-
lute error of 5.45%, a minimum absolute error of 0.061%, and the
mean absolute error of 2.22% were obtained for compressive
strength prediction. Also for tested data, it can be observed that a
maximum absolute error of 5.01%, a minimum absolute error of
0.042%, and the mean absolute error of 1.797% were obtained for
compressive strength prediction. The correlation coefficients of
0.974 and 0.982 were obtained for the training and testing data
of compressive strength prediction.
MAE (%) Minimum (%) Ef (%) R
4.02 0.15 88.87 0.964
2.22 0.061 91.16 0.972
1.927 0.0341 93.4 0.974
3.92 0.134 88.92 0.975
1.987 0.053 91.165 0.977
1.797 0.042 93.662 0.982
The results indicated that the CC neural net work has several
advantages over the BPNN: it learns very quickly, the network
determines its own size and topology, it retains the structures it
has built even if the training set changes, and it requires no back
propagation of error signals through the connections of the
network.
4. Conclusions
Concrete is a highly non-linear material, so modeling its behav-
ior is a difficult task. The neural network is a good tool to model
non-linearly systems. In this paper, the application of two NN
types namely, back propagation and cascade correlation in the esti-
mation of 3, 7, 14, and 28 days compressive strength of a LWC mix-
tures have been outlined.
The study includes the manipulation of the tested samples at
the laboratory to train and to validate the neural networks. The
conventional models of compressive strength of LWC mixtures
are developed on the basis of data collected. The data set is divided
into two data files: a training file, and a test file. Promising results
were obtained using both the neural networks model but it was
observed that the cascade correlation neural network model per-
form better compared to back propagation net work ((8–14–6–4)
model). Also, the cascade correlation neural network model has
several advantages over the BPNN (8–14–6–4) model such as it
learns very quickly, and the network determines its own size and
topology.
Material engineers may use these models to predict the stability
of LWC mixtures and avoid conducting costly experimental tests
that require specialized equipments and expertise.
References
[1] Short A, Kinniburgh W. Lightweight concrete. 3rd ed. London: Applied Science
Publishers; 1978.
[2] ACI committee 213 R-87. Guide for structural lightweight aggregate concrete.
ACI manual of concrete practice, part 1, American Concrete Institute,
Farmington Hills; 1987.
[3] Topcu IB. Semi-lightweight concretes produced by volcanic slugs. Cem Concr
Res 1997;27(1):15–21.
[4] Al-Khaiat H, Haque MN. Effect of initial curing on early strength and physical
properties of lightweight concrete. Cem Concr Res 1998;28(6):859–66.
[5] Pooliyadda SP, Dais WPS. Neural networks for predicting properties of concrete
with admixtures. Constr Build Mater 2001;15(7):371–9.
[6] Ramezanianpour AA, Davapanah A. Concrete properties estimation and mix
design optimization based on neural networks, world conference on concrete
materials and structures (WCCNS), Kuala Lumpur, Malaysia; 2002.
[7] Kiewicz KJ et al. HPC strength prediction using artificial neural network. J
Comput Civil Eng 1995;9(4).
[8] Sergio L, Mauro S. Concrete strength prediction by means of neural network.
Constr Build Mater 1997;11(2):93–8.
[9] Yeh IC. Modeling of strength of high performance concrete using artificial
neural networks. Cem Concr Res 1998;28(12):1797–808.
[10] Yeh IC. Design of high performance concrete mixture using neural networks
and nonlinear programming. J Comput Civil Eng 1999;13(1):36–42.
[11] Oh JW, Lee IW, Kim JT, Lee GW. Application of neural networks for
proportioning of concrete mixes. ACI Mater J 1999;96(1):61–7.
[12] Ni HG, Wang JZ. Prediction of compressive strength of concrete by neural
networks. Cem Concr Res 2000;30:1245–50.
 M.M. Alshihri et al. / Construction and Building Materials 23 (2009) 2214–2219
[13] Dias WPS, Pooliyadda SP. Neural networks for predicting properties of
concretes with admixtures. Constr Build Mater 2001;15:371–9.
[14] Mostofi D, Samaee A. HPC strength prediction using artificial neural network.
In: Conference proceeding, first international concrete and development
conference, Tehran, IRAN; 2001.
[15] Lee SC. Prediction of concrete strength using artificial neural network. Eng
Struct 2003;25:849–57.
[16] Kim JI, Kim DK, Feng MQ, Yazdani F. Application of neural networks for
estimation of concrete strength. J Mater Civil Eng 2004;16(3):257–64.
[17] Nahhas TM, El-Bisy MS. Neural networks for predicting properties of high
strength concrete. Al-Azhar Univ, Civil Eng Res Mag Cem 2005;27(2):
474–86.
[18] Mirza FAM, El-Bisy MS. Predictions of compressive strength of structural
lightweight concrete using artificial neural networks. In: 5th International
conference of engineering computational technology, Spain; 2006.
[19] Alshihri M, Azmy AM, El-Bisy MS. Application of neural networks in the
prediction of compressive strength of high strength concrete. Al-Azhar Univ,
Civil Eng Res Mag Cem 2007;29(2):573–89.
2219
[20] Swingler K. Applying neural networks a practical guide. New York,
London: Academic Press; 1996.
[21] Rogers JL. Simulating structural analysis with neural networks. ASCE J Comput
Civil Eng 1994;8(2):252–2265.
[22] Goh AT. Some civil engineering applications of neural networks. Proc Inst Civil
Eng Struct Build 1994;104:463–9.
[23] Maier HR, Dandy GC. The effect of internal parameters and geometry on the
performance of back propagation neural networks: an empirical study.
Environ Modell Softw 1998;13(2):193–209.
[24] Hoehfeld M, Fahlman SE. Learning with limited numerical precision using the
cascade correlation learning algorithm. IEEE Trans Neural Networks
1992;3(4):602–11.
[25] Adams A, Waugh S. Function evaluation and the cascade correlation
architecture. Perth (Western Australia): ICNN; 1995. p. 1126, 1129.
[26] Wu JK. Neural networks and simulation methods. New York: Marcel Dekker;
1994.
[27] Flood I, Kartam N. Neural networks in civil engineering I: principles and
understanding. ASCE J Comput Civil Eng 1994;8(2).