You are on page 1of 13

Construction and Building Materials 148 (2017) 757–769

Contents lists available at ScienceDirect

Construction and Building Materials


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

Model for mix design of brick aggregate concrete based on neural


network modelling
Tanja Kalman Šipoš a,⇑, Ivana Miličević b, Rafat Siddique c
a
Department for Technical Mechanics, Faculty of Civil Engineering Osijek, J.J. Strossmayer University of Osijek, Vladimira Preloga 3, Osijek, Croatia
b
Department for Materials and Structures, Faculty of Civil Engineering Osijek, J.J. Strossmayer University of Osijek, Vladimira Preloga 3, Osijek, Croatia
c
Department of Civil Engineering, Thapar University, Patiala (Punjab), India

h i g h l i g h t s

 Exploring the possibility of predicting RBAC compressive strength with neural network modelling.
 Investigation of concrete components variation effects on the concrete compressive strength.
 Development of mix design model for concrete made of recycled brick aggregate.

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

Article history: This article proposes an optimized quantitative model for proportioning concrete mixtures based on
Received 20 February 2017 cement content, water-cement ratio and percentage of recycled aggregate replacement according to pref-
Received in revised form 8 May 2017 fered recycled brick aggregate concrete (RBAC) compressive strength. A database compiled from 147
Accepted 9 May 2017
experimental tests of RBAC compressive strength was processed by neural network modelling to achieve
a reliable prediction, which was investigated by three-fold validation. The performance of the represen-
tative neural network model was verified by parametric analysis with a brief review of the influence of
Keywords:
each RBAC component. The focus of the main results is enhancement of the neural network modelling
Neural network
Recycled aggregate
results and consequently new interpretation and conceptualisation for theoretical advancement and
Compressive strength practical applied research on RBAC concrete content.
Crushed brick Ó 2017 Elsevier Ltd. All rights reserved.
Mix design

1. Introduction crushed bricks can be used as natural aggregates with substitution


up to 15% without any loss in compressive strength of concrete;
Use of recycled materials as replacement for natural aggregate however, for 30% natural aggregate substitution, there is a reduc-
in concrete is of paramount importance due to the environmen- tion of concrete properties (up to 20%, depending on the type of
tally friendly aspect of their re-use. Crushed clay bricks and clay brick).
roof tiles as alternative aggregates have particular significance as Other authors [3,4] have found that crushed brick aggregate
their use in this fashion can considerably reduce the problem of concrete has a relatively lower strength at early age than normal
waste storage as well as help in the preservation of natural aggre- aggregate concrete. They attributed this characteristic to the
gates [1]. Therefore, recycled brick aggregate concrete (RBAC) is an higher water absorption of crushed brick aggregate compared to
interesting, environmentally friendly material with unknown gravel, which was used as the control aggregate. The compressive
properties that need to be investigated. strength is lower for crushed brick concrete; the higher the rate of
Numerous studies have been performed to evaluate the poten- substitution normal aggregate with brick, the lower the compres-
tial applications of crushed brick as an aggregate. Debieb and Kenai sive strength. The use of bricks as concrete aggregate caused a
[1] used coarse and fine crushed bricks and reported a reduction in 40% reduction in compressive strength [5,6]. Concrete produced
compressive strength of RBAC concrete from 20% to 30%, depend- with these aggregates did not perform as well as concrete pro-
ing of the degree of substitution. Cachim [2] concluded that duced with natural aggregates in terms of strength. Strength prop-
erties of specimens with crushed brick aggregate are lower
⇑ Corresponding author. compared to the same properties for the reference mixtures with
E-mail addresses: tkalman@gfos.hr (T. Kalman Šipoš), ivana.milicevic@gfos.hr natural aggregate. Results of experimental investigations in [7]
(I. Miličević), siddique_66@yahoo.com (R. Siddique).

http://dx.doi.org/10.1016/j.conbuildmat.2017.05.111
0950-0618/Ó 2017 Elsevier Ltd. All rights reserved.
758 T. Kalman Šipoš et al. / Construction and Building Materials 148 (2017) 757–769

revealed lower compressive strength of concrete with recycled  To examine the parametric analysis of and investigate how con-
brick aggregate. These results are expected due to the lower grain crete components’ variation influences concrete compressive
hardness of a crushed brick/tile aggregate compared to a river strength;
aggregate. However, the concrete still has sufficient strength to  To develop a model for defining concrete content made of recy-
make it suitable for some applications, with the added benefit that cled brick aggregate for required concrete compressive
density values are much lower, making it suitable in situations strength.
where self-weight is a problem [3]. Poon and Chan [8,9] observed
that incorporation of 20% fine crushed brick aggregate decreased 2. Experimental database
the compressive strength and modulus of elasticity of concrete
by 18%. Aliabdo [10] indicated that the highest reduction in com- Regardless of the numerous experimental studies on recycled
pressive strength due to the presence of recycled aggregate is asso- aggregate concrete, there is no algorithm for defining concrete
ciated with the highest cement content at a comparable compressive strength. Furthermore, experimental tests are quite
replacement level. expensive: testing one mixture costs approximately €6125 [25].
The main inference from previous studies is that using crushed To resolve these issues, we attempted to establish the relationship
brick aggregates as alternative reduces concrete compresive between concrete compressive strength of RBAC and general con-
strength. These relationships between components and concrete crete components by collecting test data.
properties cannot be demonstrated with a mathematical formula, The collected experimental database contains data from 147
and this knowledge is imperative for optimizing the quantities of published tests (Table 1). Database parameters were selected
components used in the manufacture of RBAC. In this aspect, according to all available data samples. However, some parameters
numerical modelling for such relationships was accomplished by were excluded as they were not available for the entire database
designing an artificial neural network model (ANN) of concrete (admixture). Several of the latest research studies on RBAC con-
compressive strength. crete used a large experimental database with no modification to
ANN modelling has been widely used for prediction of concrete the replacement ratio of brick aggregate [26–28]. Therefore, these
properties with natural aggregates [11–21]. Nevertheless, few data are not applicable for the present study.
studies focused on modelling the compresive strength of concrete The concrete mixture components were the input data param-
containing recycled aggregate. Topçu and Sarıdemir [22] com- eters: cement, w/c ratio (w/c), crushed tile ratio (CT), crushed brick
pleted a trial that used ANN to predict the compressive and split- ratio (CB), and natural aggregate (NA) ratio. Physical and mechan-
ting tensile strengths of recycled aggregate concrete (RAC) ical properties of crushed bricks and tiles from various studies are
containing silica fume. Duan [23] demonstrated the possible appli- shown in [1–10]. The main conclusion that can be drawn regarding
cability of ANNs to predict the compressive strength of RAC the individual effect of specified parameters is that the type or
obtained from old buildings or pavements containing small quality of clay aggregate (i.e., brick or tile) used in this investiga-
amounts of soft soils, natural stones, clay bricks, and other impuri- tion can be neglected.
ties like paper, wood, glass, tiles, and metals. Dantas [24] used ANN Total aggregate ratio was 100% according to the following
models to predict compressive strength at 3, 7, 28, and 91 days of equation:
concrete containing construction and demolition waste (i.e., mor-
tar, concrete, red ceramic, and other recycled materials). CT½% þ CB½% þ NA½% ¼ 100½% ð1Þ
The literature review, as discussed above, reveals that no Furthermore, each aggregate partition was divided into two
research has been undertaken on the use of ANN to predict the sub-parts: fine (0–4 mm) and coarse aggregate (4–16 mm). In
compressive strength of concrete made only of recycled brick accordance with this division, the experimental database has
aggregate. eight input parameters. For evaluation of the performance of recy-
So far, research involving ANN models has mainly aspired to cled brick aggregate, it is necessary to estimate the compressive
gain insight into the possibility of predicting as well as elucidating strength of concrete. According to the availability and uniformity
the relationship between concrete components and concrete prop- of data in the experimental database, compressive strength at the
erties, with the main emphasis on performance measures. The age of 28 days was used as the output parameter. The general dis-
intention of this paper is not to characterize the statistical mea- tribution of input and output data is presented in Table 2, and the
sures of neural network prediction, but rather to provide a repre- complete database with all data is provided in Appendix. Mini-
sentative model for RBAC compressive strength from the mum and maximum values of input parameters in the experi-
perspective of concrete content based on parametric analysis. mental database were chosen based on the literature review
Consequently, the main issues of the study can be stated as: and results of experimental investigations by the authors of this
paper [1–10]. Maximum w/c ratio is 1,08 based on the research
 To explore the possibility of predicting RBAC compressive [1], in which has been decided to carry out the research project
strength; at a constant workability (slump between 60 and 70 mm), and

Table 1
Experimental database of RBAC: list of authors and samples (see Appendix for complete database).

Author Year No. of samples Aggregate type Refs.


Miličević 2011 62 Crushed brick and tile [7]
Debieb & Kenai 2008 12 Recycled brick and limestone [1]
Khalaf & DeVenny 2004 9 Crushed brick [3]
Rühl & Atkinson 1999 2 Recycled brick [4]
Khatib 2005 5 Recycled brick [6]
Cachim 2009 10 Recycled brick [2]
Poon et all -I 2007 3 Recycled brick and tile [8]
Poon et all -II 2007 4 Recycled brick and tile [9]
Topçu & Canbaz 2007 18 Crushed brick [5]
Alibdo et all 2014 22 Crushed clay brick [10]
T. Kalman Šipoš et al. / Construction and Building Materials 148 (2017) 757–769 759

Table 2
The range of input and output parameters in the experimental database.

Parameter Min Max Average


Input Cement (kg) 250 526.76 374.17
w/c ratio 0.4 1.08 0.54
CT 0–4 (%) 0 100 16.92
CT 4–16 (%) 0 100 16.78
CB 0–4 (%) 0 100 23.35
CB 4–16 (%) 0 100 26.48
NA 0–4 (%) 0 100 59.73
NA 4–16 (%) 0 100 56.73
Output Fc28 (MPa) 8.7 80.5 29.81

the water content has been adapted. Coarse aggregates did not 3.1. Number of hidden neurons
absorb any mixing water because the mix design method used
is based on a coarse aggregate in a saturated surface dry condi- The random selection of a number of hidden neurons (Nh) may
tion; hence, they were soaked in water for 24 h before use. How- cause either overfitting or underfitting of the predicting models
ever, fine brick aggregates could not be fully soaked and absorbed [32]. This concern arises when the network corresponds to the data
part of the mixing water; higher water content was thus required so closely that the ability to generalise over the test data is lost.
for mixes containing these fine brick aggregates as natural sand The number of hidden neurons has a strong influence on the
replacement. All other authors used the effective water/cement stability of the neural network, which is estimated by error (mini-
ratio defined as the amount of water available to react with the mal error reflects better stability). An excessive number of hidden
cement of the mixture. neurons will cause overfitting where neural networks overestimate
the complexity of the target problem. In this sense, determination
of the proper number of hidden neurons to prevent overfitting is
3. Neural network modelling critical for problem prediction with the capability for steady gener-
alisation with the lowest possible deviation in prediction. Accord-
ANNs are massive parallel architectures that can determine ingly, Nh must be delimited within a reasonable range.
answers to demanding problems with the participation of simple The number of hidden neurons was determined using recom-
mutually dependent processing elements (artificial neurons), as mendations of several researchers that have proposed various
is stated in [29]. An ANN has powerful aspects of learning and data
processing and thus can be an effective tool for engineering
applications. Table 3
The multi-layer backpropagation network is the most popular Empirical criteria for the number of hidden neurons Nh.
ANN paradigm according to [30] and is regularly used for efficient No Method Nh Summarized in
generalisation competence. ANNs have the ability to perform with
1. Neville (1986) 0.75 Ni [34]
a good amount of generalisation from the patterns on which they
2. Neville (1986) 2 Ni + 1 [34]
are trained [31]. Training incorporates processing neural networks 3. Hecht-Nielsen (1987) 2 Ni [33]
with a set of known input-output data. Backpropagation neural 4. Hush (1989) 3 Ni [33]
networks generally consist of layers with neurons: an input layer, 5. Popovics (1990) (Ni + No)/2 [34]
6. Gallant (1993) 2 Ni [34]
an output layer, and one or more hidden layers (Fig. 1). The learn-
7. Wang (1994) 2 Ni/3 [33]
ing process is continued in the output layer, where the error 8. Masters (1994) (Ni + No)1/2 [33]
between the network outputs and desired outputs is calculated 9. Li (1995) ((1 + 8 Ni)1/2  1)/2 [32]
and then propagated back to the network with updated weights 10. Tamura (1997) Ni + 1 [32]
for the direction in which the performance function decreases 11. Lai (1997) Ni [34]
12. Nagendra (1998) Ni + No [34]
most rapidly [14]. The entire training process is repeated for the
13. Zhang (2003) 2Ni/n + 1 [32]
number of epochs until the desired accuracy in network output 14. Shibata (2009) (Ni No)1/2 [32]
is accomplished. When the network is trained, after validation, 15. Sheela (2013) (4 N2i + 3)/(N2i  8) [32]
which is used as stopping method, it is used for testing against data
Ni – number of input neurons; No – number of output neurons.
that the network has never before seen.


X1 w1j
w2j
X2 Oj

Xi
wij Σwij i
(net)j

win Output
Xn Sum Log-sigmoid
function activation
function
Input

Fig. 1. Architecture of applied neural network.


760 T. Kalman Šipoš et al. / Construction and Building Materials 148 (2017) 757–769

34 ensure that it covers the range of input parameters. Furthermore,


all data were scaled in the range [0,1] with the following equation:
30 max USV
SV ¼ ð2Þ
max  min
26
where SV is scaled value, USV is unscaled value, and max and min
22 are maximum and minimum values of the observed data.
The performance of an ANN model primarily depends on the
network architecture and parameter settings. The values of param-
Nh

18
eters used in the neural network model are given in Table 4.
In this study, the Matlab ANN toolbox [38] was used for ANN
14
applications. All the networks were trained using the Levenberg-
Marquardt algorithm with ‘log-sigmoid’ transfer functions
10
between the first (input) and second (hidden) layers and the ‘lin-
ear’ transfer function between the second and third layers (output).
6
A three-fold procedure was used for neural network evaluation:
performance measures, performance evaluation, and success ratio
2 (Fig. 4).
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Five statistical performance measures were used to evaluate the
Empirical criteria effectiveness of each network and its ability to make accurate pre-
dictions. The performance measures used for the accuracy of the
Fig. 2. Range of number of hidden neurons according to Ni = 8, No = 1.
proposed learning models are presented in Table 5. Lower values
of MAE, RMSE, and MAPE and higher values of R and E above
methodologies. These empirical criteria for the number of hidden 0.90 illustrate good efficiency and predictability of the model.
neurons in the first hidden layer are presented in Table 3. There Furthermore, qualitative validity of neural networks is evalu-
are eight input and one output parameters, according to Table 2. ated by the success ratio SR [39] given by the following equation:
The considered empirical expressions resulted in a wide variety NBp
of values (3–33). The most frequent Nh numbers used for ANN SR ¼  100% ð3Þ
D
modelling were 3, 5, and 9 (Fig. 2). Subsequently, the neural net-
work results were evaluated to determine the number of neurons     X
y0i  y0i  n
that will provide satisfactory results. pi ¼ 1   100% ; jpi j ¼ 1    100% ; p ¼ 1n jpi j
yi yi i¼1

ð4Þ
3.2. Neural network architecture and evaluation procedure
where NBp denotes the number of data entries corresponding to the
The ANN model has eight neurons in the input layer and one range of relative error (i.e., |p|  Bp for the assumed restrained error
neuron in the output layer, as demonstrated in Fig. 3. The number Bp); D = L, V, T, P – the number of training (L), validating (V) testing
of neurons in the hidden layer is defined in the previous section. To (T) and all (P) data entries, respectively; and p = relative error.
ensure good generalisation under ANN processing, it is compulsory
to divide the experimental data in three subdivisions: training, val-
Table 4
idation, and testing.
The values of neural network parameters used in the ANN model.
For database partitioning, Looney [35] recommends 25% for
testing, whereas Swingler [36] proposes 20% and Nelson and Illing- Parameters ANN
worth [37] suggest 20 to 30%. Therefore, in the present work, the Number of input layer units 8
training data set comprises 88 data entries (60%), and the remain- Number of hidden layer 1
Number of hidden layer units 3, 5, 9
ing data entries (59) are divided between the validation and testing
Number of output layer units 1
sets. To test the reliability of the neural network model, 37 samples Learning rate 0.01
were randomly selected as the test set and 22 samples as the val- Performance goal 0
idation set. The division process was carried out randomly between Maximum number of epoch 10000
the three sets, and each dataset was statistically examined to

CT 0-4

CT 4-16
-4
-16
-4

-16

Fig. 3. Structure of ANN network: single hidden layer with five hidden neurons.
T. Kalman Šipoš et al. / Construction and Building Materials 148 (2017) 757–769 761

3.3. Evaluation of ANN prediction models

To facilitate easier comparison between the neural network


predicted data and starting experimental data, regression plots
for analysis with three, five, and nine hidden neurons are presented
(Figs. 5–7). According to the plots, the difference between the
results of training and validation sets is negligible, considering
the fact that all three neural network models exhibit similar corre-
lation coefficients.
As shown in Fig. 7 and Table 6, analysis with nine hidden neu-
rons indicates agreement between the experimental and predicted
data, according to all statistical performance measures: lower val-
ues of MAE, RMSE, and MAPE is commonly recognized as good pre-
diction of performed models. On the other hand, higher values of R
and E imply that the obtained data corresponds well to the exper-
imental data.
From the viewpoint of success ratio, neural network processing
performed with nine neurons (Fig. 8) indicates that 80% of all pre-
dicted data correspond to the relative error equal to or smaller
than 10%, with values of 14% and 21% for five and three hidden
neurons, respectively. Prediction error for 100% of the data had a
relative error smaller than 17% for best neural network model of
Fig. 4. Flow chart for optimizing the neural network model. nine hidden neurons.
The attempt to provide a detailed evaluation included a review
of performance evaluation (Fig. 9). The data training was stopped
In general, the success ratio represents what percentage of data when the MSE in the validation set began to increase, which signi-
SR (%) is used for the prediction, with the error not greater than p fies that the ANN generalisation stopped increasing. The training
(%). curves in Fig. 9 demonstrate the training procedure. The validation
Performance evaluation is characterized by plots of mean and test curves are complementary. In all three cases, the test
squared error (MSE) vs. epoch for the training, validation, and test curve did not increase significantly before the validation curve
performances of the data. In principle, the error is reduced after increased, so it is unreasonable that some overfitting occurred.
more epochs of training but may begin to increase on the valida- However, it should be emphasized that the lowest MSE is for the
tion data set as the network starts overfitting the training data. network with nine hidden neurons.
To facilitate the best ANN, performance is taken from the epoch In summary, the results of conducted analysis indicate that the
with the lowest validation error, after which there is no more proposed neural network models can be used for prediction of con-
increase in MSE error. crete compressive strength with sufficient accuracy.

Table 5
Statistical performance measures.

Statistical parameter Equation Statistical parameter Equation


Pn 0
Pn
Mean absolute error (MAE) MAE ¼ 1n i¼1 jy  y j
(5) Coefficient of correlation (R) ððyy Þðy0  y0 ÞÞ (8)
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi R ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pn i¼1 Pn 0 0
Root mean squared error (RMSE) P (6)
2
Þ 
ðyy ðy 

RMSE ¼ 1n i¼1 ðy0  yÞ2
n i¼1 i¼1

P 0 Pn
Mean absolute percentage error (MAPE) MAPE ¼ 1n ni¼1 j yy y j
(7) Nash-Sutcliffe efficiency (E)
E ¼ 1  Pi¼1
ðyy0 Þ2 (9)
n
 2 ðyyÞ
i¼1

y0 presents the predicted value of compressive strength; y is the experimental value; and n is the number of data samples, y0 is the mean predicted value; y is the mean
experimental value.

90 90 90
Predicted data (MPa)

Predicted data (MPa)

Predicted data (MPa)

60 60 60

30 30 30

0 0 0
0 30 60 90 0 30 60 90 0 30 60 90
a) b) c)
Experimental data (MPa) Experimental data (MPa) Experimental data (MPa)
Fig. 5. Comparison of experimental and predicted data of RBAC compressive strength (MPa) for: a) training, b) validation, and c) testing sets including three hidden nodes.
762 T. Kalman Šipoš et al. / Construction and Building Materials 148 (2017) 757–769

90 90 90

Predicted data (MPa)


Predicted data (MPa)

Predicted data (MPa)


60 60 60

30 30 30

0 0 0
0 30 60 90 0 30 60 90 0 30 60 90
a) Experimental data (MPa) b) Experimental data (MPa) c) Experimental data (MPa)
Fig. 6. Comparison of experimental and predicted data of RBAC compressive strength (MPa) for: a) training, b) validation, and c) testing sets including five hidden nodes.

90 90 90
Predicted data (MPa)

Predicted data (MPa)

Predicted data (MPa)


60 60 60

30 30 30

0 0 0
0 30 60 90 0 30 60 90 0 30 60 90
a) Experimental data (MPa) b) c)
Experimental data (MPa) Experimental data (MPa)
Fig. 7. Comparison of experimental and predicted data of RBAC compressive strength (MPa) for: a) training, b) validation, and c) testing sets including nine hidden nodes.

Table 6
Statistical performance of ANN models.

ANN model Statistical errors


Neural network label Learning algorithm No. of hidden nodes MAE (kg/m3) RMSE (kg/m3) MAPE (%) R E
LM_3 Levenberg-Marquardt 3 3.709 0.426 13.157 0.929 0.859
LM_5 5 3.265 0.481 11.633 0.949 0.897
LM_9 9 2.189 0.558 7.844 0.976 0.953

100 100 100

80 80 80
Succes ratio SR (%)

Succes ratio SR(%)


Succes ratio SR(%)

60 60 60

40 40 40

20 20 20

0 0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30 0 5 10 15 20 25 30
p (%) p (%) p (%)
(a) (b) (c)
Fig. 8. Success ratio for neural network predictions according to: a) three hidden nodes, b) five hidden nodes, and c) nine hidden nodes.
T. Kalman Šipoš et al. / Construction and Building Materials 148 (2017) 757–769 763

Fig. 9. Performance evaluation for: a) three hidden nodes, b) five hidden nodes, and c) nine hidden nodes.

Relying on these results, further analysis was conducted to test 4. Parametric analysis based on ANN model results
the accuracy of prediction with completely new input data to
ensure that the reference model of the neural networks provides In this section, results of analysis of concrete content parame-
reliable results for parametric analysis on the influence and varia- ters, based on the reliable ANN model, are provided, together with
tions of the components of RBAC concrete. a brief review of the influence of each parameter on RBAC com-
pressive strength.

3.4. Validation of ANN prediction models

To verify that a given prediction model is suitable for providing 4.1. Effect of clay aggregate type (tile or brick) and aggregate size (fine
reliable results, completely new data, which were not used in the or coarse) content
evaluation, were used for out of sample validation.
Input data of three samples from [7] are presented in Table 7. A simulation of aggregate type influence was carried with a
The results in Table 8 clearly demonstrate that for all three sam- concrete mixture with 350 kg/m3 cement and w/c ratio of 0.4,
ples, values predicted with ANN models are strongly consistent where the content of aggregate is varied by the replacement of nat-
with the experimentally obtained values. ural aggregates with recycled aggregates in increments of 20% to
It can be seen that the LM_9 (nine hidden neurons) model has the total replacement of natural aggregates (Fig. 10.).
the smallest MAE, RMSE, and MAPE values and highest R and E The results indicate that both aggregate types (CB and CT)
(0.999) values. It is clear that the modelling results are exception- behave similarly. In particular, the distribution of recycled aggre-
ally correct; therefore, there is no doubt regarding the accuracy of gate size content is almost the same: utilization of fine aggregate
the prediction performance of the ANN model with nine hidden with a 20% increment for every step cause a decrease in RBAC com-
neurons, which was applied for parametric analysis. pressive strength of an average of 4% of the initial value. However,

Table 7
Input data for validation of ANN prediction models.

Sample Input
Cement (kg) w/c CT 0–4 CT 4–16 CB 0–4 CB 4–16 NA 0–4 NA 4–16
C1 400 0.5 0 0 0 25 100 75
C2 400 0.5 0 0 50 25 50 75
C3 400 0.5 0 50 0 25 100 25

Table 8
Statistical performance for validation of ANN prediction models.

ANN label Sample CS _test1 CS_NN2 (MPa) Abs.error Rel.error (%) MAE (kg/m3) RMSE (kg/m3) MAPE (%) R E
(MPa) (MPa)
LM_3 C1 39.36 44.18 -4.82 12.24 3.82 1.95 11.98 0.999 0.988
C2 25.53 29.3 -3.77 14.76
C3 31.95 34.8 -2.85 8.92
LM_5 C1 39.36 42.3 -2.94 7.46 2.43 1.56 7.41 0.999 0.993
C2 25.53 27.02 -1.49 5.83
C3 31.95 34.8 -2.85 8.92
LM_9 C1 39.36 38.12 1.24 3.15 1.21 0.878 3.717 0.999 0.998
C2 25.53 27.19 -1.66 6.50
C3 31.95 30.22 1.73 5.41
1
CS_test – experimentally obtained compressive strength;
2
CS_NN – compressive strength predicted with neural network model.
764 T. Kalman Šipoš et al. / Construction and Building Materials 148 (2017) 757–769

3 = 3 =
40 40
Compressive strength (MPa)

Compressive strength (MPa)


35 35

30 30

25 25

20 20
0 20 40 60 80 100 0 20 40 60 80 100
% Recycled aggregate: clay tile % Recycled aggregate: clay brick
Fig. 10. Effect of clay aggregate type and aggregate size.

total replacement of natural aggregates would result in 19% less RBAC depends on the type and composition of CB as aggregate. By
compressive strength. using fine and coarse CB aggregate, the required cement content
Another interesting point is the effect of coarse aggregate appli- can be even higher than the aforementioned 20% [41]. This conclusion
cation: regardless of aggregate type, RBAC compressive strength can be adjusted according to the analysis with 100% replacement of
has much lower values, especially for percentages equal to or natural aggregate with RBAC. From the analysis presented in Fig. 11,
higher than 50% (for an increment of 20% of aggregate replace- it can be concluded that for compressive strength of 33 MPa with nat-
ment, the reduction is 10%; however, the use of 100% coarse recy- ural aggregates (0% replacement) and cement amount of 350 kg/m3,
cled aggregate causes a decrease in compressive strength by 33%). the same compressive strength can be achieved with total replace-
On the other hand, fine aggregate, regardless of whether the ment of 100% coarse aggregate and an increase of cement content of
aggregate is brick or tiles, shows a lower decrease in the value of approximately 420 kg/m3, which confirms the aforementioned
compressive strength. In the case of total replacement (100%), claim. However, for 100% fine aggregate replacement, increase of
compressive strength is decreased by 18%. According to the con- cement content is only 9% (350 kg/m3 to 380 kg/m3 for 33 MPa).
ducted analysis, fine aggregate had a higher compressive strength, To facilitate the direct influence of cement content on compres-
followed by mixtures containing coarse aggregate at the same sive strength of RBAC, different amounts cement in the concrete
replacement level. This can be attributed to the pozzolanic reaction mixture were explored. Note that analysis was limited to values
between silica and alumina that exists in the very fine portion of of 350–450 kg/m3 in increments of 50 kg/m3. Qualitative results
crushed clay aggregate and the product of cement hydration. [40] (Fig. 11) indicate that 50 kg/m3 of cement causes an increase in
It should be noted that the conducted analysis implies that there compressive strength up to 30%, regardless of the aggregate size
is no significant impact of aggregate type so it will be disregarded in used (fine or coarse).
further analysis. However, aggregate size parameter (fine or coarse)
will be included in further analysis for more accurate evaluation of 4.3. Effect of water-cement ratio
its effect. Based on this investigation and previously conducted test-
ing in [40], compressive strength that nearly equal to the compres- According to [42], differences in concrete strength for a given
sive strength of common concrete with natural aggregate may be water/cement (w/c) ratio can result from:
achieved in concrete with CB as the aggregate, taking into account
the 45% limit for coarse and 50% limit for fine particle fraction in  changes in the aggregate size, grading, surface texture, shape,
substitution of natural aggregate with CB aggregate. strength, or stiffness;
 differences in types and sources of cementing materials;
4.2. Effect of cement amount  entrained-air content;
 the presence of admixtures;
The cement content in RBAC can be up to 20% higher than its con-  the length of curing time.
tent in ordinary concrete with natural aggregate. Cement content in

70 70
Compressive strength (MPa)

Compressive strength (MPa)

60 60

50 50

40 40

30 30

20 20
0 20 40 60 80 100 0 20 40 60 80 100
% Recycled coarse aggregate: clay brick % Recycled fine aggregate: clay brick

Fig. 11. Effect of cement amount.


T. Kalman Šipoš et al. / Construction and Building Materials 148 (2017) 757–769 765

3 3

Compressive strength (MPa)

Compressive strength (MPa)


40 40

30 30

20 20

10 10
0 20 40 60 80 100 0 20 40 60 80 100
% Recycled coarse aggregate: clay brick
% Recycled fine aggregate: clay brick

Fig. 12. Effect of water-cement ratio.

Fig. 13. Model for compressive strength according to RBAC content for cement in the amount of 350 kg.

In this investigation, the w/c ratio varied from 0.4 to 0.6 with not control, the w/c ratio should be selected on the basis of con-
increments of 0.1. Modification of the w/c ratio is associated with a crete compressive strength. In such cases, the w/c ratio and mix-
change of the compressive strength value in the following way: an ture proportions for the required strength should be based on
increase from 0.4 to 0.5 reduces the compressive strength to an aver- adequate field data or trial mixtures made with actual job materi-
age of 9%, whereas next increment of 0.5 to 0.6 produces a an average als to determine the relationship between the ratio and strength.
reduction of 15% (Fig. 12). The strength of the cementitious paste The aim of the designer should always be to produce concrete mix-
binder in concrete depends on the quality and quantity of the react- tures of optimum strength at minimum cement content and
ing paste components and the degree to which the hydration reac- acceptable workability.
tion has progressed [43]. Concrete becomes stronger with time as Any mix design procedure provides a first approximation of
long as there is moisture and a favourable temperature. The strength the proportions and must be checked by trial batches. Consider-
at any particular age is a function of the original w/c ratio and the ing all conclusions from previous subsections, the contribution
degree to which the cementitious materials have hydrated. The of concrete components to RBAC compressive strength implies
importance of prompt and thorough curing is easily recognized. that their quantifying effects must be included in the process
A systematic decrease in compressive strength can be observed of defining a concrete mixture for required compressive
by increasing the w/c ratio due to the extra amount of added water strength. To facilitate easier application of all previous conclu-
in mixtures containing RBA in addition to the inferior strength of sions, a new model for defining concrete content is proposed.
CBA compared with natural aggregates [40], which is confirmed The model shown in Fig. 13 was produced by using 312 points
in this analysis. obtained by the reliable ANN model, processed with regression
with a minimum value of the square of the correlation coeffi-
4.4. Relation of compressive strength and proportioning RBAC mixtures cient of 0.85.
Once the w/c ratio is established and the workability required
Since most of the properties of hardened concrete depend pri- for the specific design is determined, the rest involves simple
marily upon the quality of the cement paste, the first step in pro- manipulation with diagrams and tables based on large numbers
portioning a concrete mixture is the selection of the appropriate of trial mixes. The diagram shown on Fig. 13 allow an estimation
w/c ratio for the strength and durability needed. Grading and nat- of the required proportions of RB aggregate, water, and cement
ure of aggregate particles are two characteristics that have an for the chosen compressive strength and permits predetermination
important influence on proportioning concrete mixtures because of small unrepresentative batches. Accordingly, for a w/c ratio of
they affect the workability of the fresh concrete. Within the normal 0.5 and replacement of 50% of coarse clay brick, expected compres-
range of strengths used in concrete construction, the compressive sive strength of RBAC is approximately 29 MPa. However, if we
strength is inversely related to the w/c ratio. When durability does want to achieve a specified RBAC compressive strength of
766 T. Kalman Šipoš et al. / Construction and Building Materials 148 (2017) 757–769

32 MPa for cement content of 350 kg and w/c ratio of 0.5, it is nec- aggregates; an increase in cement amount is directly related
essary to replace 25% of coarse clay brick according to the sug- to the replacement ratio of RBA;
gested mix model. An extrapolation is recommended if required.  in contrast, the w/c ratio must be decreased to preserve the
same compressive strength because of the negative effect of
5. Conclusion water in mixtures containing RBA, in addition to the inferior
strength of CBA compared with natural aggregates;
In this study, we tested ANNs as an alternative to classic exper-  the replacement ratio of recycled clay aggregate is the most
imental tests for simulating material properties in concrete with important parameter as it directly affects RBAC compressive
crushed brick and roof tile aggregate. strength by a decrease of up to 33%.
The results of this research indicate that compressive strength
of RBAC, presented with eight input parameters, can be reliably Compressive strength that is nearly equal to that of concrete
predicted with neural networks, according to three-fold evaluation with natural aggregate can be achieved in concrete with CBA, tak-
and out of sample validation. ing into account the limits for coarse and fine particle fraction in
The contribution of concrete components to RBAC compressive the substitution of natural aggregate with CB aggregate at approx-
strength was verified by parametric analysis, which implied that imately 20% and 40%, respectively.
their quantifying effects must be included in the process of defin- The focus of the current research is enhancement of neural net-
ing a concrete mixture for required compressive strength. work models results and consequently, a new interpretation and
The beneficial results of this study include the following conceptualisation for theoretical advancement and practical
conclusions: applied research for RBAC content. To facilitate easier application
of all previous conclusions, a new model for defining concrete con-
 the type of clay aggregate (brick or tile) can be neglected tent is proposed.
because indicated results prove that both aggregate types In regard to the large number of experimental investigations
behave similarly; that have evaluated only a small range of replacement ratios (most
 aggregate size (fine or coarse) has a significant influence on RBAC of the studies have looked at 50% and 100%), results from this study
compressive strength: coarse aggregate would cause a decrease prove that additional experimental research is necessary for
in compressive strength value twice as high as fine aggregate; replacement ratio of 0–50% with smaller increments. Subse-
 cement content must be increased up to 20% to achieve the quently, the proposed model for design mix can be confirmed or
same compressive strength of RBAC as concrete with natural improved.

Appendix
Data sources.

Author Year Sample Cement (kg) w/c CT1 0–4 CT 4–16 CB2 0–4 CB 4–16 NA3 0–4 NA 4–16 Fc28 (MPa)
Miličević [7] 2011 1 400 0.5 0 0 0 25 100 75 10.95
2 400 0.5 0 0 50 25 50 75 23.5
3 400 0.5 0 50 0 25 100 25 8.7
4 400 0.5 0 50 50 25 50 25 16.6
5 400 0.5 50 0 0 25 50 75 22.4
6 400 0.5 50 0 50 25 0 75 16.84
7 400 0.5 50 50 0 25 50 25 18.8
8 400 0.5 50 50 50 25 0 25 9.6
9 300 0.5 25 25 0 0 75 75 15.5
10 300 0.5 25 25 0 50 75 25 15.27
11 300 0.5 25 25 50 0 25 75 20
12 300 0.5 25 25 50 50 25 25 10.84
13 500 0.5 25 25 0 0 75 75 54.5
14 500 0.5 25 25 0 50 75 25 28.4
15 500 0.5 25 25 50 0 25 75 45.2
16 500 0.5 25 25 50 50 25 25 25.4
17 400 0.4 25 0 25 0 50 100 61.75
18 400 0.4 25 0 25 50 50 50 23.07
19 400 0.4 25 50 25 0 50 50 23.32
20 400 0.4 25 50 25 50 50 0 14.83
21 400 0.6 25 0 25 0 50 100 26
22 400 0.6 25 0 25 50 50 50 21.13
23 400 0.6 25 50 25 0 50 50 27.53
24 400 0.6 25 50 25 50 50 0 16.74
25 300 0.4 0 25 25 25 75 50 46.43
26 300 0.4 50 25 25 25 25 50 21.33
27 300 0.6 0 25 25 25 75 50 17.25
28 300 0.6 50 25 25 25 25 50 13.05
29 500 0.4 0 25 25 25 75 50 41.33
30 500 0.4 50 25 25 25 25 50 46
31 500 0.6 0 25 25 25 75 50 43.75
32 500 0.6 50 25 25 25 25 50 33.15
33 400 0.5 0 25 25 0 75 75 9.97
34 400 0.5 0 25 25 50 75 25 24.43
35 400 0.5 50 25 25 0 25 75 34.8
T. Kalman Šipoš et al. / Construction and Building Materials 148 (2017) 757–769 767

Appendix (continued)

Author Year Sample Cement (kg) w/c CT1 0–4 CT 4–16 CB2 0–4 CB 4–16 NA3 0–4 NA 4–16 Fc28 (MPa)
36 400 0.5 50 25 25 50 25 25 42.24
37 400 0.5 0 25 25 0 75 75 37.6
38 400 0.5 0 25 25 50 75 25 22.4
39 400 0.5 50 25 25 0 25 75 32
40 400 0.5 50 25 25 50 25 25 16.82
41 300 0.5 25 0 25 25 50 75 21
42 300 0.5 25 50 25 25 50 25 23.97
43 300 0.5 25 0 25 25 50 75 18.04
44 300 0.5 25 50 25 25 50 25 9.04
45 500 0.5 25 0 25 25 50 75 29.8
46 500 0.5 25 50 25 25 50 25 26.13
47 500 0.5 25 0 25 25 50 75 32.13
48 500 0.5 25 50 25 25 50 25 40.3
49 400 0.4 25 25 0 25 75 50 43.8
50 400 0.4 25 25 50 25 25 50 44.8
51 400 0.4 25 25 0 25 75 50 33
52 400 0.4 25 25 50 25 25 50 15.34
53 400 0.6 25 25 0 25 75 50 20.65
54 400 0.6 25 25 50 25 25 50 25.16
55 400 0.6 25 25 0 25 75 50 47.75
56 400 0.6 25 25 50 25 25 50 27.97
57 400 0.5 25 25 25 25 50 50 19.28
58 400 0.5 25 25 25 25 50 50 28.05
59 400 0.5 25 25 25 25 50 50 20.7
60 400 0.5 25 25 25 25 50 50 23.08
61 400 0.5 25 25 25 25 50 50 22.53
Author Year Sample Cement (kg) w/c CT 0–4 CT 4–16 CB 0–4 CB 4–16 NA 0–4 NA 4–16 Fc28 (MPa)
Miličević [7] 2011 62 400 0.5 25 25 25 25 50 50 26.15
Debieb & Kenai [1] 2008 C0/0 350 0.61 0 0 0 0 100 100 30.62
C0/25 350 0.69 0 0 25 0 75 100 28.21
C0/50 350 0.77 0 0 50 0 50 100 26.73
C0/75 350 0.85 0 0 75 0 25 100 25.82
C0/100 350 0.93 0 0 100 0 0 100 22.35
C50/50 350 0.75 0 0 50 50 50 50 21.51
C100/100 350 0.89 0 0 100 100 0 0 18.26
C+100/100 350 0.86 0 0 100 100 0 0 21.23
C75/25 350 0.66 0 0 25 75 75 25 21.25
C100/50 350 0.72 0 0 50 100 50 0 20.73
C25/75 350 0.85 0 0 75 25 25 75 23.13
C50/100 350 1.08 0 0 50 100 50 0 19.56
Khalaf & DeVenny [3] 2004 M1 G 350.35 0.55 0 0 0 0 100 100 45.7
M1 O 329.47 0.55 0 0 0 100 100 0 37.6
M1 O+ 345.95 0.55 0 0 0 100 100 0 46.7
M2 G 498.64 0.4 0 0 0 0 100 100 66.8
M2 O 453.13 0.4 0 0 0 100 100 0 53.8
M2 O+ 480.63 0.4 0 0 0 100 100 0 66.7
M3 G 418.55 0.43 0 0 0 0 100 100 42.7
M3 O 384.27 0.43 0 0 0 100 100 0 38.8
M3 O+ 402.35 0.43 0 0 0 100 100 0 44.2
Rühl & Atkinson [4] 1999 NZ 320 0.55 0 0 0 0 100 100 35.39
ZI 320 0.55 0 0 0 100 100 0 32.99
Khatib [6] 2005 control 325 0.5 0 0 0 0 100 100 46.7
CB 25 319 0.5 0 0 25 0 75 100 39.2
CB 50 314 0.5 0 0 50 0 50 100 37.7
CB 75 307 0.5 0 0 75 0 25 100 36.1
CB 100 303 0.5 0 0 100 0 0 100 33.2
Cachim [2] 2009 NN 45 400 0.45 0 0 0 0 100 100 36.2
NA 45 400 0.45 0 0 0 15 100 85 32.1
NB 45 400 0.45 0 0 0 15 100 85 38.5
AA 45 400 0.45 0 0 0 30 100 70 27.6
BB 45 400 0.45 0 0 0 30 100 70 32.3
NN 50 400 0.5 0 0 0 0 100 100 30.5
NA 50 400 0.5 0 0 0 15 100 85 29.4
NB 50 400 0.5 0 0 0 15 100 85 32.3
AA 50 400 0.5 0 0 0 30 100 70 24.5
BB 50 400 0.5 0 0 0 30 100 70 29
Poon et all [8] 2007 Mix 1 410 0.55 0 0 0 0 100 100 53.8
Mix 2 410 0.55 0 0 20 0 80 100 47.2
Mix 3 410 0.55 20 0 0 0 80 100 45.5
Poon et all [9] 2007 control 526.76 0.41 0 0 100 100 0 0 80.5
10 T 489.76 0.49 10 10 0 0 90 90 65.6
5T5B 491.49 0.49 5 5 5 5 90 90 62.4
Topcu & Canbaz [5] 2007 4B4G2T 493.98 0.49 2 2 8 8 90 90 66.2
K300_0 300 0.63 0 0 0 0 100 100 21.63

(continued on next page)


768 T. Kalman Šipoš et al. / Construction and Building Materials 148 (2017) 757–769

Appendix (continued)

Author Year Sample Cement (kg) w/c CT1 0–4 CT 4–16 CB2 0–4 CB 4–16 NA3 0–4 NA 4–16 Fc28 (MPa)
K300_50 300 0.63 50 50 0 0 50 50 15
K300_100 300 0.63 100 100 0 0 0 0 14.23
K350_0 350 0.54 0 0 0 0 100 100 27.15
K350_50 350 0.54 50 50 0 0 50 50 27.95
K350_100 350 0.54 100 100 0 0 0 0 19.81
K400_0 400 0.48 0 0 0 0 100 100 36.5
K400_50 400 0.48 50 50 0 0 50 50 30.12
K400_100 400 0.48 100 100 0 0 0 0 22.16
I300_0 300 0.63 0 0 0 0 100 100 26.32
I300_50 300 0.63 50 50 0 0 50 50 25.71
I300_100 300 0.63 100 100 0 0 0 0 15.13
I350_0 350 0.54 0 0 0 0 100 100 36.12
I350_50 350 0.54 50 50 0 0 50 50 27.36
I350_100 350 0.54 100 100 0 0 0 0 21.71
I400_0 400 0.48 0 0 0 0 100 100 36.23
I400_50 400 0.48 50 50 0 0 50 50 34.12
I400_100 400 0.48 100 100 0 0 0 0 31.65
Alibdo et all [10] 2014 I_1 350 0.5 0 0 0 0 100 100 33.6
Aliabdo et all [10] 2014 I_2 350 0.5 0 0 25 0 75 100 36.5
I_3 350 0.5 0 0 50 0 50 100 34.6
I_4 350 0.5 0 0 75 0 25 100 32.1
I_5 350 0.5 0 0 100 0 0 100 27.6
I_6 350 0.5 0 0 0 25 100 75 34.2
I_7 350 0.5 0 0 0 50 100 50 33.6
I_8 350 0.5 0 0 0 75 100 25 25.8
I_9 350 0.5 0 0 0 100 100 0 22.3
I_10 350 0.5 0 0 50 50 50 50 29.6
I_11 350 0.5 0 0 100 100 0 0 23.8
II_14 250 0.7 0 0 0 0 100 100 22.5
II_15 250 0.7 0 0 25 0 75 100 23.4
II_16 250 0.7 0 0 50 0 50 100 21.9
II_17 250 0.7 0 0 75 0 25 100 23.1
II_18 250 0.7 0 0 100 0 0 100 15.9
II_19 250 0.7 0 0 0 25 100 75 22.3
II_20 250 0.7 0 0 0 50 100 50 22.1
II_21 250 0.7 0 0 0 75 100 25 18.6
II_22 250 0.7 0 0 0 100 100 0 16.2
II_23 250 0.7 0 0 50 50 50 50 22.3
II_24 250 0.7 0 0 100 100 0 0 15.5
1
CT – crushed tile;
2
CB – crushed brick;
3
NA – natural aggregate.

References [11] H.-G. Ni, J.-Z. Wang, Prediction of compressive strength of concrete by neural
networks, Cem. Concr. Res. 30 (8) (2000) 1245–1250, http://dx.doi.org/
10.1016/S0008-8846(00)00345-8.
[1] F. Debieb, S. Kenai, The use of coarse and fine crushed bricks as aggregate in
[12] S.C. Lee, Prediction of concrete strength using artificial neural networks, Eng.
concrete, Constr. Build. Mater. 22 (5) (2008) 886–893, http://dx.doi.org/
Struct. 25 (7) (2003) 849–857, http://dx.doi.org/10.1016/S0141-0296(03)
10.1016/j.conbuildmat.2006.12.013.
00004-X.
[2] P.B. Cachim, Mechanical properties of brick aggregate concrete, Constr. Build.
[13] R. Ince, Prediction of fracture parameters of concrete by artificial neural
Mater. 23 (3) (2009) 1292–1297, http://dx.doi.org/10.1016/
networks, Eng. Fract. Mech. 71 (15) (2004) 2143–2159, http://dx.doi.org/
j.conbuildmat.2008.07.023.
10.1016/j.engfracmech.2003.12.004.
[3] I. Miličević, Fire resistance of concrete with crushed bricks and roof tiles PhD
[14] M.A. Kewalramani, R. Gupta, Concrete compressive strength prediction using
Thesis, Faculty of Civil Engineering Osijek, University J.J. Strossmayer Osijek,
ultrasonic pulse velocity through artificial neural networks, Autom. Constr. 15
2014. https://urn.nsk.hr/urn:nbn:hr:133:213795.
(3) (2006) 374–379, http://dx.doi.org/10.1016/j.autcon.2005.07.003.
[4] F.M. Khalaf, A.S. DeVenny, Recycling of demolished masonry rubble as coarse
[15] M. Pala, E. Özbay, A. Öztas, M.I. Yuce, Appraisal of long-term effects of fly ash
aggregate in concrete: review, J. Mater. Civ. Eng. 16 (2004) 331–340, http://dx.
and silica fume on compressive strength of concrete by neural networks,
doi.org/10.1061/(ASCE)0899-1561(2004) 16:4(331).
Constr. Build. Mater. 21 (2) (2007) 384–394, http://dx.doi.org/10.1016/
[5] M. Rühl, G. Atkinson, The influence of recycled aggregate concrete on stress-
j.conbuildmat.2005.08.009.
strain relation of concrete, Darmstadt Concr. 26 (14) (1999) 36–52.
[16] I.-C. Yeh, Modeling slump flow of concrete using second-order regressions and
[6] J.M. Khatib, Properties of concrete incorporating fine recycled aggregate, Cem.
artificial neural networks, Cement Concr. Compos. 29 (6) (2007) 474–480,
Concr. Res. 35 (2005) 763–769, http://dx.doi.org/10.1016/j.
http://dx.doi.org/10.1016/j.cemconcomp.2007.02.001.
cemconres.2004.06.017.
[17] G. Trtnik, F. Kavčič, G. Turk, Prediction of concrete strength using ultrasonic
[7] C.-S. Poon, D. Chan, The use of recycled aggregate in concrete in Hong Kong,
pulse velocity and artificial neural networks, Ultrasonics 49 (1) (2009) 53–60,
Resour. Conserv. Recycl. 50 (3) (2007) 293–305, http://dx.doi.org/10.1016/j.
http://dx.doi.org/10.1016/j.ultras.2008.05.001.
resconrec.2006.06.005.
[18] F. Özcan, C.D. Atis, O. Karahan, E. Uncuoğlu, H. Tanyildizi, Comparison of
[8] C.-S. Poon, D. Chan, Effects of contaminants on the properties of concrete
artificial neural network and fuzzy logic models for prediction of long-term
paving blocks prepared with recycled concrete aggregates, Constr. Build.
compressive strength of silica fume concrete, Adv. Eng. Softw. 40 (9) (2009)
Mater. 21 (1) (2007) 164–175, http://dx.doi.org/10.1016/
856–863, http://dx.doi.org/10.1016/j.advengsoft.2009.01.005.
j.conbuildmat.2005.06.031.
[19] M.M. Alshihri, A.M. Azmy, M.S. El-Bisy, Neural networks for predicting
[9] I.B. Topçu, M. Canbaz, Utilization of crushed tile as aggregate in concrete, Iran.
compressive strength of structural light weight concrete, Constr. Build.
J. Sci. Technol. Trans. B Eng. 31 (5) (2007) 561–565.
Mater. 23 (6) (2009) 2214–2219, http://dx.doi.org/10.1016/
[10] A.A. Aliabdo et al., Utilization of crushed clay brick in concrete industry,
j.conbuildmat.2008.12.003.
Alexandria Eng. J. 53 (1) (2014) 151–168, http://dx.doi.org/10.1016/j.
aej.2013.12.003.
T. Kalman Šipoš et al. / Construction and Building Materials 148 (2017) 757–769 769

[20] C. Bilim, C.D. Atis, H. Tanyildizi, O. Karahan, Predicting the compressive [31] M. Bilgehan, P. Turgut, The use of neural networks in concrete compressive
strength of ground granulated blast furnace slag concrete using artificial strength estimation, Comput. Concr. 7 (3) (2010) 271–283.
neural network, Adv. Eng. Softw. 40 (5) (2009) 334–340, http://dx.doi.org/ [32] K. Gnana Sheela, S.N. Deepa, Review on methods to fix number of hidden
10.1016/j.advengsoft.2008.05.005. neurons in neural networks, Math. Prob. Eng. (2013), http://dx.doi.org/
[21] J. Sobhani, M. Najimi, A.R. Pourkhorshidi, T. Parhizkar, Prediction of the 10.1155/2013/425740. Article ID 425740, 11 pages.
compressive strength of no-slump concrete: a comparative study of [33] H. Sonmez, C. Gokceoglu, H.A. Nefeslioglu, A. Kayabasi, Estimation of rock
regression, neural network and ANFIS models, Constr. Build. Mater. 24 (5) modulus: for intact rocks with an artificial neural network and for rock masses
(2010) 709–718, http://dx.doi.org/10.1016/j.conbuildmat.2009.10.037. with a new empirical equation, Int. J. Rock Mech. Min. Sci. 43 (2) (2006) 224–
[22] I.B. Topçu, M. Sarıdemir, Prediction of mechanical properties of recycled 235, http://dx.doi.org/10.1016/j.ijrmms.2005.06.007.
aggregate concretes containing silica fume using artificial neural networks and [34] M. Özturan, B. Kutlu, T. Özturan, Comparison of concrete strength prediction
fuzzy logic, Comput. Mater. Sci. 42 (1) (2008) 74–82, http://dx.doi.org/ techniques with artificial neural network approach, Build. Res. J. 56 (1) (2008)
10.1016/j.commatsci.2007.06.011. 23–36.
[23] Z.H. Duan, S.C. Kou, C.S. Poon, Prediction of compressive strength of recycled [35] C. Looney, Advances in feedforward neural networks: demystifying knowledge
aggregate concrete using artificial neural networks, Constr. Build. Mater. 40 acquiring black boxes, IEEE Trans. Knowl. Data Eng. 8 (2) (1996) 211–226.
(2013) 1200–1206, http://dx.doi.org/10.1016/j.conbuildmat.2012.04.063. [36] K. Swingler, Applying neural networks: a practical adaptive weights,
[24] A.T.A. Dantas, M. Batista Leite, K. De Jesus Nagahama, Prediction of Proceedings of an IEEE International Guide, Academic Press, New York, 1996.
compressive strength of concrete containing construction and demolition [37] M. Nelson, W.T. Illingworth, A Practical Guide To Neural Nets, Addison-Wesley,
waste using artificial neural networks, Constr. Build. Mater. 38 (2013) 717– Reading, MA, 1990.
722, http://dx.doi.org/10.1016/j.conbuildmat.2012.09.026. [38] M.H. Beale, M.T. Hagan, H.B. Demuth, Matlab: Neural networks toolbox user
[25] I. Miličević, N. Štirmer, D. Bjegović, Optimizing the Concrete mixture made guide, Mathworks (2015).
with recycled aggregate using experiment design, recent advances in fluid [39] K. Kuźniar, Z. Waszczyszyn, Neural networks for the simulation and
mechanics and heat and mass transfer, in: Proc. of the 9th IASME/WSEAS Int. identification analysis of buildings subjected to paraseismic excitations, in:
Conf. on Fluid Mechanics and Aerodynamics, FMA’11, Proc. of the 9th IASME/ N. Lagaros, Y. Tsompanakis (Eds.), Intelligent Computational Paradigms in
WSEAS Int. Conf. HTE’11, 2011, pp. 110–115. Earthquake Engineering, Idea Group, Hershey, PA, USA, 2007 (Chapter XVI).
[26] T.U. Mohammed et al., Effects of maximum size of brick aggregate on [40] D.M. Sadek, Physico-mechanical properties of solid cement bricks containing
properties of concrete, Constr. Build. Mater. 134 (2017) 713–726, http://dx.doi. recycled aggregates, J. Adv. Res. 3 (3) (2012) 253–260, http://dx.doi.org/
org/10.1016/j.conbuildmat.2016.12.164. 10.1016/j.jare.2011.08.001.
[27] T.U. Mohammed et al., Recycling of brick aggregate concrete as coarse [41] I. Miličević, N. Štirmer, I. Banjad Pečur, Residual mechanical properties of
aggregate, J. Mater. Civ. Eng. 27 (7) (2014), http://dx.doi.org/10.1061/(ASCE) concrete made with crushed clay bricks and roof tiles aggregate after exposure
MT.1943-5533.0001043. to high temperatures, Materials 9 (4) (2016) 295-1–295-5, http://dx.doi.org/
[28] A.E.B. Cabral et al., Performance estimation for concretes made with recycled 10.3390/ma9040295.
aggregates of construction and demolition waste of some brazilian cities, [42] T.C. Hansen, Recycling of demolished concrete and masonry, RILEM Rep. 6,
Mater. Res. 15 (6) (2012) 1037–1046, http://dx.doi.org/10.1590/S1516- E&FN Spon, London, 1992.
14392012005000119. [43] S.H. Kosmatka, B. Kerkhoff, W.C. Panarese, Design and Control of Concrete
[29] S. Haykin, Neural Networks: A Comprehensive Foundation, second ed., Mixutes, EB001, 14th edition., Portland Cement Association, Skokie, Illinois,
Pretince Hall, 1999. USA, 2008.
[30] R. Lippman, An introduction to computing with neural nets, IEEE ASSP Mag. 4
(2) (1987) 4–22.