Construction

and Building

Construction and Building Materials 20 (2006) 769–775

MATERIALS
www.elsevier.com/locate/conbuildmat

Predicting the compressive strength and slump of high

strength concrete using neural network
a,*
Ahmet Öztasß , Murat Pala b, Erdoǧan Özbay b, Erdoǧan Kanca b,
Naci Çaǧlar c, M. Asghar Bhatti d
a
Civil Engineering Department, University of Gaziantep, Gaziantep, Turkey
b
Technical Programs Department, Kilis MYO, University of Gaziantep, Kilis, Turkey
c
Civil Engineering Department, University of Sakarya, Sakarya, Turkey
d
Civil and Environmental Engineering, University of Iowa, Iowa City, IA, USA

Received 11 August 2004; received in revised form 30 September 2004; accepted 31 January 2005
Available online 23 March 2005

Abstract

High Strength Concrete (HSC) is defined as concrete that meets special combination of performance and uniformity require-
ments that cannot be achieved routinely using conventional constituents and normal mixing, placing, and curing procedures.
HSC is a highly complex material, which makes modelling its behavior very difficult task. This paper aimed to show possible appli-
cability of neural networks (NN) to predict the compressive strength and slump of HSC. A NN model is constructed, trained and
tested using the available test data of 187 different concrete mix-designs of HSC gathered from the literature. The data used in NN
model are arranged in a format of seven input parameters that cover the water to binder ratio, water content, fine aggregate ratio, fly
ash content, air entraining agent, superplasticizer and silica fume replacement. The NN model, which performs in Matlab, predicts
the compressive strength and slump values of HSC. The mean absolute percentage error was found to be less then 1,956,208% for
compressive strength and 5,782,223% for slump values and R2 values to be about 99.93% for compressive strength and 99.34% for
slump values for the test set. The results showed that NNs have strong potential as a feasible tool for predicting compressive strength
and slump values.
Ó 2005 Elsevier Ltd. All rights reserved.

Keywords: High strength concrete; Neural networks; Compressive strength; Slump

1. Introduction high strength at 28 days (typically >60 MPa compressive

strength) or a low water-binder ratio (<0.40). In USA,
High strength concretes (HSCs) are used extensively HSC is supposed to be a special mixture, matching spe-
worldwide in construction projects [1,2]. A HSC is de- cific requirements that cannot be achieved on a routine
fined as concrete that meets special combination of per- basis [3]. In Turkey, HSC is considered to be a concrete
formance and uniformity requirements that cannot be having compressive strength >40 MPa at 28 days [4].
achieved routinely by using conventional constituents Several new advanced concretes have been transferred
and normal mixing, placing, and curing procedures. In from laboratory research to practical applications and
Europe, HSC is considered to be a concrete having a they already occupy a noticeable share of the market.
Based on the latest developments in concrete technol-
*
ogy, HSC is characterized by a superior level of proper-
Corresponding author. Present address: Gaziantep Üniversitesi,
_ ßaat Müh. Bölümü, 27310 Gaziantep, Turkey. Tel./fax: 90 342 360
Ins
ties: workability and strength. These advantages could
11 07. provide large-scale cost savings in many construction
E-mail address: [email protected] (A. Öztasß). projects [5].

0950-0618/$ - see front matter Ó 2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.conbuildmat.2005.01.054
770 A. Öztasß et al. / Construction and Building Materials 20 (2006) 769–775

A major difference between conventional concrete allel architectures that solve difficult problems via the
and HSC is the use of chemical and mineral admixtures cooperation of highly interconnected but simple com-
that reduce the water content, thereby reducing porosity puting elements (or artificial neurons). Basically, the
within the hydrated cement paste [6]. The reduction in processing elements of a neural network are similar to
water content to very low value can be achieved by using the neuron in the brain, which consists of many simple
high dosage of chemical admixtures. However, it is computational elements arranged in layers [8]. The basic
undesirable since the effectiveness of chemical admix- strategy for developing a neural network-based model
tures such as superplasticizer (SP) depends on ambient for material behavior is to train a neural network on
temperature, cement chemistry, and fineness of aggre- the results of a series of experiments using that material.
gate. On the other hand, mineral admixtures act as puz- If the experimental results contain the relevant informa-
zolanic materials as well as fine fillers resulting in a tion about the material behavior, then the trained neural
denser and stronger microstructure of the hardened ce- network will contain sufficient information about mate-
ment matrix [7]. The mineral admixtures are generally rialÕs behavior to qualify as a material model. Such a
industrial by-products and their usage can provide ma- trained neural network not only would be able to repro-
jor economic benefits. Thus, judicial use of SP and min- duce the experimental results, but also it would be able
eral admixtures can lead to economical HSC with to approximate the results in other experiments through
enhanced durability. its generalization capability [10].
The actual compressive strength of concrete is un- Yeh [8], Kasperkiewics [11], Lai [12] and Lee [13] ap-
known during the early life of the structure. Also, the plied the NN for predicting properties of conventional
concrete market is generally very competitive and it concrete and high performance concretes. Bai [14] devel-
turns out that concrete companies have only restricted oped neural network models that provide effective pre-
budgets to spend in mix-design, although from this fun- dictive capability in respect of the workability of
damental stage comes a great deal of consequences for concrete incorporating metakaolin (MK) and fly ash
the site operations and for the structure to be built. Fur- (FA). Ji-Zong [15] developed an automatic knowledge-
thermore, in laboratory, to obtain desired concrete acquisition system based on neural networks to design
strength with suitable workability, technical personnel concrete mix. Ji-ZongÕs system consists of three models:
must try several mix proportions. This time consuming the mix-design model, the slump-prediction model, and
procedure increase the wastage of material and cost of the strength-prediction model; the first model is the core
concrete production. of the system with the other two models supporting the
The current empirical equations presented in the core. Guang and Zong [16] proposed a method to pre-
codes and standards for estimating compressive dict 28-day compressive strength of concrete by using
strength are based on tests of concrete without supple- multi-layer feed-forward neural networks. Dias and
mentary cementitious materials. The validity of these Pooliyadda [17] used back propagation neural networks
relationships for concrete with supplementary cementi- to predict the strength and slump of ready mixed con-
tious materials (fly ash, silica fume, superplasticizer, crete and high strength concrete, in which chemical
etc.) should be investigated. The more we know about admixtures and/or mineral additives were used.
the concrete composition versus strength relationship, The aim of this paper is to present a methodology for
the better we can understand the nature of concrete predicting compressive strength of HSC with suitable
and how to optimize the concrete mixture [8]. workability. For this aim, a computer program was
Several researchers have looked into the characteris- developed in Matlab. Using this program, a NN model
tic parameters that affect the compressive strength and with two hidden layers was constructed, trained and
slump values of HSC. Some of these parameters include tested using the available test data of 187 different con-
quality of aggregate, strength of cement, water content crete mix-designs of HSC gathered from the technical
and water-to-cement ratio. The traditional approach literature. The data used in NN model were arranged
used in modeling the effects of these parameters on the in a format of seven input parameters that cover the
compressive strength and slump of concrete starts with water-to-binder ratio, water content, fine aggregate ra-
an assumed form of analytical equation and is followed tio, fly ash content, air entraining agent content, superp-
by a regression analysis using experimental data to lasticizer content and silica fume replacement. The
determine unknown coefficients in the equation [9]. proposed NN model predicts the compressive strength
Unfortunately, rational and easy-to-use equations are and slump value of HSCs.
not yet available in design codes to accurately predict
the compressive strength and slump.
Over the last two decades, a different predicting meth- 2. Neural Network (NN)
od based on NNs has become popular and has been
used by many researchers for a variety of engineering Neural network is a functional abstraction of the
applications [8–16]. NNs are a family of massively par- biological neural structures of the central nervous
 A. Öztasß et al. / Construction and Building Materials 20 (2006) 769–775 771

system [18–20]. It can exhibit a surprising number of Table 1

human brainÕs characteristics, e.g., learning from expe- Range of input parameters in database
rience and generalizing from previous examples to new Input parameters Minimum Maximum
problems. NN can provide meaningful answers even W/B (ratio, %) 18 45
when the data to be processed include errors or are W (kg/m3) 140 165
incomplete, and can process information extremely s/a (ratio, %) 35 52
FA (ratio, %) 0 20
rapidly when applied to solve real world problems AE (kg/m3) 0.036 0.078
[21]. Neurocomputing architectures can be built into SF (ratio, %) 5 25
physical hardware (or machine) or neurosoftware lan- SP (kg/m3) 1.89 36.5
guages (or programs) that can think and act intelli-
gently like human beings. Among various ratio, silica fume replacement ratio and superplasticizer
architectures and paradigms, the back-propagation content.
network is one of the simplest and most applicable All 187 HSC samples collected from the literature
network being used in performing higher level human were evaluated. The ranges of compressive strength
task such as diagnosis, classification, decision-making, and slump values of samples were from 40 to
planning and scheduling [22]. It is one of the most 120 MPa, and from 95 to 260 mm, respectively. There
popular learning (training) algorithms. Accordingly, are two groups of samples. The first group consists of
for a given input pattern, a flow of activation is for- 104 HSC samples and their ingredients are water to
warded from the input layer to the output layer via binder ratio, water content, fly ash replacement ratio,
the hidden layer(s). Then, the errors in the output air-entraining agent ratio, fine aggregate ratio and
are initiated. The neural network based modelling pro- superplasticizer content. The second group consists of
cess involves five main aspects: (a) data acquisition, 83 HSC samples and their ingredients are water to bin-
analysis and problem representation; (b) architecture der ratio, water content, fine aggregate ratio, silica fume
determination; (c) learning process determination; (d) replacement ratio and superplasticizer content. During
training of the networks; and (e) testing of the trained the data evaluation, these two different sample groups
network for generalization evaluation [23]. were combined and a database of 187 records was ob-
The error incurred during the learning can be ex- tained. In combination of these records, we considered
pressed as root-mean-squared (RMS) one and is calcu- the seven input variables in Table 1 for the NN model.
lated [24] using Eq. (1) The table also shows the minimum and maximum values
s ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 X of these parameters used in the network implementation.
1 2 The two output variables were the compressive
RMS ¼  j t j  oj j : ð1Þ
p j strength and slump values. To test the reliabilty of the
NN model, we randomly selected 18 samples as the test
In addition, the absolute fraction of variance (R2), mean set, while the remaining 169 samples were used to train
absolute percentage error (MAPE) and sum of the the network.
squares error (SSE) are calculated[24] using Eqs. (2)–
(4), respectively.
P 2
! 4. Neural network model and parameters
j ðt j  o j Þ
R2 ¼ 1  P 2
; ð2Þ
j ðoj Þ NN model developed in this research has seven neu-
o  t  rons (variables) in the input layer and two neurons in
MAPE ¼  100; ð3Þ the output layer as illustrated in Fig. 1. The input layer
o neurons receive information from the outside environ-
X ment and transmit them to the neurons of the hidden
SSE ¼ ðoj  tj Þ2 ; ð4Þ
j layers without performing any calculation. The hidden
layer neurons then process the incoming information
where t is the target value, o is the output value, and p is
and extract useful features to reconstruct the mapping
the pattern.
from the input space. The neighbouring layers are fully
interconnected by weights. Lastly, the output layer neu-
rons produce the network predictions to the outside
3. Data sets world. There is no general rule for selecting the number
of neurons in a hidden layer. The choice of hidden layer
Experimental data from different sources were used size is problem dependent and does influence quality of
[25,8,13]. Data were assembled for HSCs containing training pattern. Some guidelines for finding the opti-
water to binder ratio, water content, fine aggregate mum number of hidden layer can be found in references
ratio, fly ash replacement ratio, air-entraining agent [26,27]. As far as the hidden layer is concerned, there is
772 A. Öztasß et al. / Construction and Building Materials 20 (2006) 769–775

Input Layer

W/B 1.Hidden Layer

2. Hidden Layer
W Output Layer

s/a Comp. Strength

FA
Slump
AE.

SP

SF

Fig. 1. Proposed NN model.

no general rule for selecting the number of neurons in a through a particular training pattern in which it ad-
hidden layer. The choice of hidden layer size is mainly a justs the weights by a small amount at a time. The
problem and to some extent depends on the number and learning algorithm used in the study is scaled conju-
quality of the training pattern but to find optimum num- gate gradients algorithm (SCGA), activation function
ber of hidden layer, Masters [26] proposed some ap- is sigmoidal function, and number of epochs is
proaches. By considering aforementioned approaches, 10,000. By using a step size scaling mechanism, SCGA
following NN model has been proposed as seen in Fig. 1. avoids a time consuming line search per learning
In this study, a computer program has been iteration.
developed [28,29] and performed under Matlab. The The values of parameters used in this research are as
back-propagation learning algorithm has been used follows:
in feed-forward two hidden-layers. Back propagation
algorithm (BPA), as one of the most famous training  Number of input layer unit = 7.
algorithms for the multi layer perceptron (MLP), is a  Number of hidden layers = 2.
gradient descent technique to minimize the error e  Number of first hidden layer units = 5.

Fig. 2. Performance of training set of 169 HSC mix-design data.

 A. Öztasß et al. / Construction and Building Materials 20 (2006) 769–775 773

 Number of second hidden layer units = 3. the 187 HSC mix-design data. As mentioned earlier
 Number of output layer units = 2. 169 samples were used for training the network and
 Learning algorithm = scaled conjugate gradients other 18 (chosen randomly) were used to test the pre-
algorithm(SCGA). diction. The performance of training set and test set
 Learning cycle = 10,000. can be seen in Figs. 2 and 3. The results in Fig. 2 indi-
cate that the neural network was successful in learning
the relationship between the different input parame-
ters and two outputs (compressive strength and slump
5. Analysis of results value).
The result of testing phase in Fig. 3 shows that the neu-
The NN model developed in this research is used to ral network was capable of generalizing between input
predict the compressive strength and slump values of variables and the output and reasonably good predic-

Fig. 3. Performance of test set of 18 HSC mix-design data.

NN predicted compressive strength versus experimental compressive

strength
Compressive strength (MPa)

140
120
100
80
60
40
20
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
(a) Test set samples
NN Experiment

NN predicted slump values versus experimental slump values

300
250
Slump values

200
(mm)

150
100
50
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
(b) Test set samples
NN Experimental

Fig. 4. Comparison of experimental compressive strength and slump values with NN predicted compressive strength and slump values.
(a): experimental compressive strength values versus NN predicted values. (b): experimental slump values versus NN predicted values.
774 A. Öztasß et al. / Construction and Building Materials 20 (2006) 769–775

Table 2 is a highly complex material that makes modelling its

Statistical values of proposed NN model behaviour a difficult task. NNs have strong potential
Training set Test set as a feasible tool for predicting compressive strength
Compressive Slump Compressive Slump and slump value. Furthermore, to obtain desired HSC
strength strength strength and suitable workability, technical personnel
RMS 0.019331 0.068743 0.016455 0.055425 must try several mix proportions. The proposed NN
SSE 0.061658 0.070237 0.004874 0.055295 model will save time, reduce waste material and decrease
R2 0.998977 0.992576 0.99931 0.993451 the design cost.
MAPE (%) 2.617104 6.58954 1.956208 5.782223
In construction, early determination of HSC strength
and slump values is very important. Normally, determi-
Error distribution of compressive strength values in test set nation of compressive strength takes 28 days but using
5.00 the proposed NN model, the compressive strength and
slump values can be predicted in a short time.
Error (%)

0.00
-5.00
-10.00
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 References
Test set samples
[1] Shah SP. Recent trends in the science and technology of concrete,
Fig. 5. Error distribution of compressive strength in test set. concrete technology, new trends, industrial applications. In:
Proceedings of the international RILEM workshop, London, E
Error distribution of slump values in test set & FN Spon; 1993. p. 1–18.
[2] Bache HH. Densified cement/ultrafine particle based materials. In:
20.00 The second international conference on superplasticizers in
10.00 concrete, Ottawa; 1981.
Error (%)

0.00 [3] Camoes A, Aguiar B, Jalali S. Durability of low cost high

-10.00 performance fly ash concrete. In: International ash symposium,
-20.00 Kentucky, USA; 2003, paper no. 43.
-30.00 [4] Özkul H, Tasdemir MA, Tokyay M, Uyan M. Her Yönüyle Beton
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 Istanbul. TCMB yayınları 1999;33.
Test set samples [5] Sobolev K. The development of a new method for the propor-
tioning of high performance concrete mixtures. Cem Concr Comp
Fig. 6. Error distribution of slump values in test set. 2004;26(7):901–7.
[6] Mittal A, Basu DC. Development of HPC for PC Dome of NPP.
tions. Graphically, comparisons of experimental data in Indian Concrete J 1999;73(3):671–9.
[7] Bharatkumar BH, Naranayan R, Raghuprasad BK, Ramachan-
test set and NN results were given in Fig. 4(a) and (b).
dramurthy DS. Mix proportioning of HSC. Cem Concr Comp
The statistical values for both compressive strength 2001;23:71–80.
and slump values in training and test set as RMS, R2, [8] Yeh IC. Modeling of strength of HPC using ANN. Cem Concr
MAPE and SSE are also given in Table 2. The values Res 1998;28(12):1797–808.
of MAPE are 2.617104% and 6.58954% in training set [9] Mansour MY. Predicting the shear strength of RC beams using
ANN. Eng Struct 2004;26:781–99.
and 1.956208% and 5.782223% in test set for compres-
[10] Jung HC, Jamshid G. Genetic algorithm in structural damage
sive strength and slump, respectively. The R2 values detection. Comput Struct 2001;30(6):1335–53.
are 99.8977% and 99.2576% in training set and [11] Kasperkiewics J, Racz J, Dubrawski A. HPC strength prediction
99.931% and 99.3451% in test set for compressive using ANN. ASCE J Comput Civil Eng 1995;9(4):279–84.
strength and slump, respectively. The SSE and RMS [12] Lai S, Sera M. Concrete strength prediction by means of neural
network. Constr Build Mater 1997;11(2):93–8.
values for compressive strength and slump values can
[13] Lee SC. Prediction of concrete strength using artificial neural
be seen in Table 2. All of the statistical values in Table Networks. Eng Struct 2003;25:849–57.
2 demonstrate that the proposed NN model is suitable [14] Bai J, Wild S, Ware JA, Sabir BB. Using neural networks to
and predicts the compressive strength and slump values predict workability of concrete incorporating metakaolin and fly
very close to experimental values. A small perceptible ash. Adv Eng Softw 2003;34:663–9.
[15] Ji-Zong W, Hong-Guang N, Jin-Jun H. The application of
deviation is observed for the calculated values. The error
automatic acquisition of knowledge to mix design of concrete.
distribution of compressive strength and slump values in Cem Concr Res 1999;29:1875–80.
test set are given in Figs. 5 and 6. [16] Guang NH, Zong WJ. Prediction of compressive strength of
concrete by neural networks. Cem Concr Res 2000;30:1245–50.
[17] Dias WPS, Pooliyadda SP. Neural networks for predicting
properties of concretes with admixtures. Constr Build Mater
6. Conclusion
2001;15:371–9.
[18] Gunaydin HM, Dogan SZ, A neural network approach for early
The NN-based model was developed for predicting cost estimation of structural systems of buildings. Int J Project
the HSC compressive strength and slump values. HSC Manage; 2004 (in press).
 A. Öztasß et al. / Construction and Building Materials 20 (2006) 769–775 775

[19] Arbib MA. The handbook of brain theory and neural net- structural engineering. Edinburgh: Civil-Comp Press; 1993.
works. Cambridge, MA: MIT Press; 1995. p. 61–6.
[20] Anderson JA. An introduction to neural networks. A bradford [24] Sozen A, Arcaklioglu E. Solar potential in Turkey. Applied Energ
book. Cambridge MA: MIT Press; 1995. 2005;80:35–45.
[21] Lippman RP. An introduction to computing with neural nets. In: [25] Lim CH, Yoon YS. Genetic algorithm in mix proportioning of
Artificial neural networks. The computer society theoretical HPC. Cem Concr Res 2004;34(3):409–20.
concepts. Washington; 1988, p. 36–54. [26] Masters T. Practical neural network recipes. McGraw-Hill; 1993.
[22] Sohabhon B, Spethen OO. Application of ANN to forecast [27] Demirbasß AE. Modelling of seismically excited structures using
construction duration of buildings at the predesign stage. Engi ANN, Master Thesis, METU, Ankara; 1998.
Constr Architect Manage 1999;6(2):133–44. [28] Caglar N. Artificial neural networks in dynamic analysis of
[23] Wu X, Lim SY. Prediction maximum scour depth at the buildings. Ph.D. Thesis, Sakarya University; 2002.
spur dikes with adaptive neural networks. In: Neural [29] Pala M. Dynamic soil structure interaction with artificial neural
networks and combinatorial optimization in civil and network, Ph.D. Thesis, Sakarya University; 2004.