Construction and Building Materials 94 (2015) 371–379

Contents lists available at ScienceDirect

Construction and Building Materials

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

Explicit formulation of drying and autogenous shrinkage of concretes

with binary and ternary blends of silica fume and fly ash
Kasım Mermerdasß ⇑, Mohamed Moafak Arbili
Hasan Kalyoncu University, Faculty of Engineering, Civil Engineering Department, 27410 Gaziantep, Turkey

h i g h l i g h t s g r a p h i c a l a b s t r a c t

 Explicit formulation of shrinkage of

concretes with mineral admixtures
was aimed.
 Experimental data from literature
were collected to derive an analytical
model.
 The utilized soft computing technique
is artificial neural networks.
 The proposed models provided high
accuracy prediction performance.
 An experimental study was
conducted to verify the proposed
model.

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

Article history: Shrinkage is generally considered as an important durability property of hardened concrete. During the
Received 25 April 2015 drying process, free and absorbed water is lost from the concrete. When the drying shrinkage is
Received in revised form 1 June 2015 restrained, cracks can occur depending on the internal stresses in the concrete. The ingress of deleterious
Accepted 12 July 2015
materials through these cracks can cause decrease in the compressive strength and the durability of con-
Available online 16 July 2015
crete. In the first stage of the study, a prediction model through the most popular soft computing method
called neural network (NN) was derived. The data set used for training and testing of the prediction
Keywords:
model covers the experimental data presented in the literature. In the second stage of the study, the find-
Shrinkage
Modeling
ings of an experimental study on drying shrinkage behavior of concretes incorporated with silica fume
Prediction (SF) and fly ash (FA) were reported. Free shrinkage strain measurements as well as corresponding weight
Experimental validation loss were measured over 40 days of drying. The obtained experimental results were also used for the val-
Mineral admixtures idation of the proposed prediction models. The highest amount of mineral admixture resulted in high
shrinkage strain development. Moreover, the proposed NN model also accurately predicted the values
obtained from experimental study.
Ó 2015 Elsevier Ltd. All rights reserved.

1. Introduction

⇑ Corresponding author. Concrete is the most widely used construction material in all
E-mail address: kasim.mermerdas@hku.edu.tr (K. Mermerdasß). over the world because of its serviceability performance. The more

http://dx.doi.org/10.1016/j.conbuildmat.2015.07.074
0950-0618/Ó 2015 Elsevier Ltd. All rights reserved.
372 K. Mermerdasß, M.M. Arbili / Construction and Building Materials 94 (2015) 371–379
durable the concrete is, the longer the service life of the structure
is. In some aspects shrinkage is of concern when it relates to dura-
bility of concrete structure. Excessive shrinkage may cause con-
crete cracking, even structural failure. Thus, cracking may lead to
increased corrosion rate of steel reinforcement in concrete struc-
ture [1]. In the view of global sustainable development, therefore,
researchers start to make use of blending of two or three supple-
mentary cementing materials (SCMs) to optimize durability and
cost of the concrete. The industrial by-products used as SCMs, such
as fly ash and silica fume, have become more efficient admixtures
to diminish the shrinkage effects and increase the durability of
concrete. Usage of SCMs could substantially reduce the final cost
of concrete mixtures since these materials are quite cheaper in
comparison to Portland cement [2,3].
The problems encountered in the field of engineering are gener-
ally unstructured and imprecise influenced by intuitions and past
experiences of a designer [4]. Complexity to mathematically model
real world problems has compelled the human civilization to
search for nature inspired computing tools. The evolution of such
computing tools revolves around the information processing char-
acteristics of biological systems. In contrast to conventional com-
puting, these tools are rather ‘‘soft’’ as they lack the exactness
and therefore placed under the umbrella of a multidisciplinary
field called soft computing. Soft computing is an emerging collec-
tion of methodologies, which aim to exploit tolerance for impreci-
sion, uncertainty and partial truth to achieve robustness,
tractability and total low cost [5].
Soft Computing tools exploit the reasoning, intuition, con-
sciousness, wisdom and adaptability to changing environments
possessed by human beings for developing computing paradigms
like fuzzy logic (FL), neural networks (NN) and genetic algorithms
(GA). The integration of these techniques into the computing envi-
ronment has given impetus to the development of intelligent and
wiser machines possessing logical and intuitive information pro-
cessing capabilities equivalent to human beings. These techniques
whether complementing each other or working on their own, are
able to model complex or unknown relationships which are either
nonlinear or noisy. Soft computing techniques have a self-adapting
characteristic paving a way for development of automated design
systems. A synergistic partnership exploiting the strengths of these
individual techniques can be harnessed for developing
hybrid-computing tools [5]. In order to handle complex nonlinear
relationships between various inputs and outputs, soft computing
techniques are used to derive mathematical models obtained from
soft computing methods such as neural network, genetic
programming.
In the recent years, soft-computing techniques have been
widely utilized to predict some critical properties of concrete
and/or reinforced concrete [6–21]. For example, Vanluchene and
Sun [6] presented an introduction to neural network by using
back-propagation algorithm to solve three different structural
engineering problems related to pattern recognition, decision mak-
ing and problems that have numerically complex solutions. In the
study of Ozcan et al. [10], compressive strength prediction was
done by using NN and fuzzy logic. Shahin et al. [12] used neural
networks for predicting settlement of shallow foundations on
cohesion less soils. They compared predictive ability of NN with
three of the most commonly used traditional methods. Tokar and
Johnson [18] used ANN to forecast daily runoff as a function of
daily precipitation, temperature and snowmelt. Lee and Han [21]
developed efficient neural network models for generation of artifi-
cial earthquakes and response spectra.
Other than above mentioned studies, soft computing based
explicit formulation of some design parameters of structural steel
members and mechanical behavior of post installed anchors have
also been reported in recent years [22–26]. This can be considered
as an indication on how large the field of application of soft com-
puting based modeling to handle the engineering problems.
In this study an explicit formulation of prediction model from
NN was presented. For this purpose, experimental data were col-
lected from the available test results presented in the previous
studies. The prediction parameters were selected from mixture
constituents of concrete and drying period. An experimental vali-
dation study was also conducted to evaluate the performance of
the prediction model. In the study, SF and FA were used as a
replacement for Portland cement (PC), ranging from 0% to 15% by
weight, to evaluate its efficiency upon the concrete properties.
Four different concrete mixtures with w/b ratio of 0.45 were
designed. The focus of the experimental study is to evaluate the
effectiveness of FA and SF on drying shrinkage of the concretes
as well as assessment of the prediction capability of the proposed
NN model.
2. Neural networks (NN)
An artificial neural network (NN) is an information-processing
paradigm that is inspired by the way biological nervous systems,
such as the brain, process information. The key element of this
paradigm is the novel structure of the information processing sys-
tem. It is composed of a large number of highly interconnected
processing elements (neurons) working in unison to solve specific
problems. NNs, like people, learn by example. An NN is configured
for a specific application, such as pattern recognition or data clas-
sification, through a learning process. Learning in biological sys-
tems involves adjustments to the synaptic connections that exist
between the neurons.
The training of NNs by back propagation have three stages [27]:
(i) the feed forward of the input training pattern, (ii) the calcula-
tion and back propagation of the associated error, and (iii) the
adjustment of the weights. This process can be used with a number
of different optimization strategies. The error between the output
of the network and the target value is propagated backward during
the backward pass and used to update the weights of the previous
layers as shown in Fig. 1 [28].
In this study, neural network fitting tool (nftool) provided as a
soft-computing tool in MatlabV.R2012a was utilized to perform
neural network modeling. In fitting problems, a neural network
may be used to map between a data set of numeric inputs and a
set of numeric targets. The nftool helps create and train a network,
and evaluate its performance using mean square error and regres-
sion analysis.
A two-layer feed-forward network with sigmoid hidden neu-
rons and linear output neurons can fit multi-dimensional mapping
problems arbitrarily well, given consistent data and enough neu-
rons in its hidden layer. The network was trained with
Levenberg–Marquardt back propagation algorithm.
An artificial neuron consists of three main components namely
weights, bias, and an activation function. Each neuron receives
inputs I1, I2, . . ., In attached with a weight wi which shows the con-
nection strength for that input for each connection. Each input is
then multiplied by the corresponding weight of the neuron con-
nection. A bias can be defined as a type of connection weight with
a constant nonzero value added to the summation of weighted
inputs, as given in Eq. (1). Generalized algebraic matrix operation
was also given in Eq. (2). to clarify the mathematical operations
in an artificial neuron.
X
n
U k ¼ Biask þ wk;j  Ij ð1Þ
j¼1
 K. Mermerdasß, M.M. Arbili / Construction and Building Materials 94 (2015) 371–379 373

Start

Select the first NN algorithm

Train and test the NN with the first input variable

Reference=Test result of NN with the first input variable

i=1

Add one exogenous variable into the NN

Train and test the new NN

No Yes
If test result
>= Reference

Variable does not stay in Variable stays in the

the model model

Reference=Test result of NN

i=i+1

If i >= no. of No
variables available

Train and test the NN with all the available variables

Choose another NN algorithm or architecture parameter

Identify the overall best model

End

Fig. 1. Forward strategy for selecting NN architecture and model [28].

2 3 2 3 2 3 2 3
w11 w12 : : : w1n I1 Bias1 U1 of the NN. Moreover, the obtained results are also in the normal-
6w
6 21 : : 77 6I 7
6 27
6 Bias 7
6 27
6U 7
6 27
ized form. Therefore, considering the Eq. (3) and the normalization
6 7 6 7 6 7 6 7 coefficients a and b for outputs, de-normalization process is
6 : 7 6:7 6 : 7 6 : 7
6
6 : : 77 6 7 6 7 6 7 applied and the results are monitored.
Uk ¼ 6 : 7 6 7 þ6 7 ¼6 7
6 : : : 7 6:7 6 : 7 6 : 7
6 7 6 7 6 7 6 7 bnormalized ¼ ab þ b ð3Þ
6 7 6 7 6 7 6 7
4 : : : 5 4:5 4 : 5 4 : 5
where b is the actual input parameter or output values given in
wm1 : : : : wmn In n1 Biasm m1 Um m1 Table 1. bnormalized is the normalized value of input parameters or
mn
ð2Þ outputs ranging between [1, 1]. a and b are normalization coeffi-
cients given in the following equations (Eqs. (4) and (5)).
Since nftool uses the normalized values in the range of [1, 1],
the input parameters were normalized by means of Eq. (3) in order 2
a¼ ð4Þ
to get the prediction results after execution of the training process bmax  bmin
374 K. Mermerdasß, M.M. Arbili / Construction and Building Materials 94 (2015) 371–379

Table 1
Summary of experimental database.

Data source Input Output

(586 data samples) X1 X2 X3 X4 X5 X6 X7 X8 Y
w/b SF FA cement Agg/b. fc Mpc Type of Shrinkage Drying period Shrinkage
@28 days
Zhang et al. [29] 0.27–0.35 0–50 0 446–498 3.38–3.70 57.33–86.94 0–1 1–98 34–282
Wongkeo et al. [30] 0.49 0–42 0–269 269–538 2.64–2.75 29.05–69.05 1 7–91 93–1100
Yoo et al. [31] 0.3 0–88 0–175 408–583 2.68–2.57 54.8–69.8 0 1–49 39–400
Khatib et al. [32] 0.36 0 0–400 100–500 3.25–3.5 11–72.58 1 2–56 5–432
Khatri and Sirivivatnanon [33] 0.34–0.36 0–46 0–100 282–425 4.15–4.30 65–94.99 1 7–400 267–895

bmax þ bmin data output variable curve. Higher R coefficients indicate a model
b¼ ð5Þ
bmax  bmin with better output approximation capability.
P
where bmax and bmin are the maximum and minimum actual values ðmi  m0 Þðpi  p0 Þ
R ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð6Þ
of either inputs or outputs. ðmi  m0 Þ2 ðpi  p0 Þ2

3. Description of the database used for derivation of the models
The proposed formulations of S for shrinkage were derived
using a set of 586 experimental data available in the technical lit-
erature [29–33] for training and testing the proposed models.
Table 1 summarizes the selected experimental data. In detail,
the generated models for shrinkage following input parameters:
w/b (water/binder), SF (silica fume) content in kg/m3, FA (fly ash)
content in kg/m3, C (cement) content in kg/m3, aggregate/binder
ratio, fc (compressive strength) in MPa, type of shrinkage for drying
shrinkage 1, for autogenous shrinkage 0, and dry time in days.
All data samples were put in an order to establish a consistent
sequence of the inputs to be used for derivation of the models.
Thus, eight inputs parameters were utilized for development of
prediction models. The data set was randomly divided into two
parts to obtain training and testing databases. MatlabV.R2012a
software was used for derivation of the NN based mathematical
model.
For clarity sake, the correlation is evaluated by means of the
correlation coefficient ‘‘R’’ (Eq. (6)), which describes the fit of the
models’ output variable approximation curve to the actual test
where m0 and p0 are mean values of measured (mi) and predicted (pi)
values, respectively.
4. Proposed NN model
A neural network architecture, as shown in Fig. 2, was adopted
to develop the NN model. That means there are eight nodes in the
input layer, corresponding to eight factors from I1 to I8, 20 nodes in
the hidden layer, and one in the output layer corresponding to the
shrinkage. It should be noted that all numeric variables were nor-
malized to a range of [1, 1] before being introduced to the NN.
Therefore, one must enter the normalized values in the mathemat-
ical operations given for NN model. Normalization of the data is
achieved according to the mathematical operations given in Eqs.
3–5 It should also be noted that the final result obtained from
Eq. (7) is also in the normalized form, which needs to be
de-normalized according to Eq. (3) and normalization coefficients
given in Table 2.
X
m
Shrinkage ¼ Biasoutput layer þ LWk f ðU k Þ ð7Þ
k¼1

Fig. 2. Architecture of neural network.

 K. Mermerdasß, M.M. Arbili / Construction and Building Materials 94 (2015) 371–379 375

where Biasoutput layer = 1.7249 and f(x) (Hyperbolic tangent) is the
activation function given in Eq. (8), LWk is layer weight matrix Uk
numerical value of neurons
Calculation of U is shown in Eq. (9) LWk matrix is also given in
Eq. (10)
5. Comparison of the proposed model with experimental data
In order to compare the prediction of the proposed models with
experimental shrinkage values available in database, the Figs. 5–9
were plotted. Fig. 5 includes the experimental and predicted auto-

2
f ðxÞ ¼ 1 ð8Þ
1  e2x
2 3 2 3 2 3
0:1908 1:803 0:0873 0:7623 0:1331 1:4559 0:3369 1:5765 2:5334 U1
6 7 6 7 6 7
6 1:336 0:257 1:9303 0:796 0:851 2:4924 0:016 1:2352 7 6 2:0022 7 6 U2 7
6 7 6 7 6 7
6 0:7413
6 0:3955 0:921 0:7256 0:2407 1:2586 2:3354 0:9309 7
7
6 1:4367 7
6 7
6 U3 7
6 7
6 7 6 7 6 7
6 0:7555 0:3286 0:2206 1:2888 0:8885 1:0273 0:2604 2:3917 7 6 2:16 7 6 U4 7
6 7 6 7 6 7
6 0:5732 0:208 1:5959 1:44 0:1226 1:4484 0:4694 7 6 7 6 7
6 3:2405 7 6 3:7156 7 6 U5 7
6 7 6 7 6 7
6 1:2354 0:8284 0:5011 0:2014 1:9002 0:8489 0:9686 0:2244 7 6 1:5463 7 6 U6 7
6 7 2 3 6 7 6 7
6 0:3063
6 1:8275 0:2715 0:5259 1:3442 2:2712 0:367 0:6669 7
7 w=b 6 2:3974 7
6 7
6 U7 7
6 7
6 7 6 SF 7 6 7 6 7
6 0:1888 0:2276 0:4856 0:737 0:3709 0:06 5:3211 8:2235 7 6 7 6 3:9 7 6 U8 7
6 7 6 7 6 7 6 7
6 0:5043
6 0:4569 1:1168 1:4441 0:2978 0:3478 0:7296 2:9012 7
7 6 FA 7 6 2:4121 7
6 7
6U 7
6 97
6 7 6 7 6 7 6 7
6 1:0541 0:5796 0:5104 1:9961 0:4117 0:7198 1:5012 0:1507 7 6 C 7 6 1:1096 7 6 U 10 7
7 6 7 7 7
U¼6  6 7 þ 6 ¼ 6
6 1:1694
6 0:1551 0:1716 0:7359 0:232 7:6509 1:8679 0:2992 7
7 6 Agg=b 7
6 7
6 1:0209 7
6 7
6U 7
6 11 7
6 7 6 7 6 7 6 7
6 0:067
6 1:6273 2:3283 0:6925 0:3623 1:2884 0:1942 0:1286 7
7 6 fc 7 6 0:4657 7
6 7
6 U 12 7
6 7
6 0:5885 7 6 7 6 7 6 7
6 1:4898 1:2218 0:7913 1:448 0:3074 0:3418 0:7812 7 4 Shr:type 5 6 0:4179 7 6 U 13 7
6 7 6 7 6 7
6 1:4778
6 2:7926 0:4601 0:4953 2:559 4:16 0:0882 0:3506 7
7 Time 81 6 6 1:161 7
7 6 U 14 7
6 7
6 0:3467
6 0:2172 0:3685 1:542 1:2205 3:8501 1:0123 1:027 7
7
6 0:2536 7
6 7
6U 7
6 15 7
6 7 6 7 6 7
6 3:1121
6 1:1517 0:8458 0:3275 0:0566 8:3806 1:8627 0:583 7
7
6 1:1396 7
6 7
6 U 16 7
6 7
6 7 6 7 6 7
6 1:0581 1:3466 0:2873 1:3621 3:031 2:0197 1:9895 1:0801 7 6 2:5899 7 6 U 17 7
6 7 6 7 6 7
6 0:2626
6 0:6351 0:6929 1:748 0:069 6:0709 1:4018 0:2991 7
7
6 2:2622 7
6 7
6 U 18 7
6 7
6 7 6 7 6 7
4 0:8306 0:506 0:4179 1:1953 0:2531 1:1383 0:7668 0:2324 5 4 2:1491 5 4 U 19 5
0:3459 1:0081 1:0881 0:4976 1:1249 3:0761 0:8297 2:8866 208 4:3735 201 U 20 201
ð9Þ

2 3
1:6671 genous shrinkage values, while the other figures contain drying
6 7 shrinkage values.
6 1:3796 7
6 7 Observing Fig. 5 it can be seen that prediction performance of
6 2:892 7
6 7 the proposed model for autogenous shrinkage values between 0
6 7
6 0:65194 7
6 7 and 100 microstrain is slightly misleading. However, the model
6 2:5559 7
6 7 performed well for the other intervals. For the autogenous shrink-
6 7
6 1:0008 7 age values higher than 100 microstrain the model demonstrated
6 7
6 0:57559 7 almost prefect estimation except a few underestimated values.
6 7
6 7 For drying shrinkage values, the model had both overestimated
6 3:9264 7
6 7 and underestimated results between experimental values of 0–300
6 1:3711 7
6 7 microstrain (Fig. 6). However, the tendency is very close to the
6 7
6 1:5766 7 actual values at this interval. As the experimental drying shrinkage
Layer weights ¼ 6 7 ð10Þ
6 2:5045 7 values were increased the prediction capability of the proposed
6 7
6 7 model became better (Figs. 7–9). Especially for the drying shrink-
6 3:1582 7
6 7 age values of 900–1200 microstrain all of the predicted values
6 0:93087 7
6 7
6 7 had about 5–10% divergence from the original values (Fig. 9).
6 1:2166 7
6 7
6 0:93796 7
6 7 6. The limitations of the proposed model
6 7
6 2:2957 7
6 7
6 7 The limitations of the proposed model can be considered as
6 1:7001 7
6 7 follows;
6 5:1574 7
6 7
6 7
4 1:7862 5 (i) Although the range of the inputs presented in this study cov-
0:55315 201 ers wide ranges of intervals they have to be within the max-
imum and minimum input values (bmin < Input data < bmax)
The obtained results from the NN model are also plotted in used for derivation of the model.
Figs. 3 and 4 yielding 0.993 and 0.954 correlation coefficients for (ii) The input entered to the model cannot be used as they are.
training and testing data sets, respectively. The estimated results The normalization of the data should be conducted before
have close tendency to the experimental values. introduction to the model.
376 K. Mermerdasß, M.M. Arbili / Construction and Building Materials 94 (2015) 371–379

Table 2
Normalization coefficients.

Normalization parameter w/b ratio SF FA C Agg/b fc⁄ (Mpa) @28 days Type of Shrinkage Dry time (days) Shrinkage strain
bmax 0.49 88 400 583 4.468 112.8 1 400 1100
bmin 0.27 0 0 100 2.57 11 0 1 5.4
a 9.296699 0.022727 0.005 0.004141 1.060519 0.019646 2 0.0050125 0.001827
b 3.51011 1 1 1.41408 3.72861 1.21611 1 1.0050125 1.00987

(iii) The proposed model presented is not suitable for hand cal- 7.2. Materials
culation. Nevertheless, the proposed explicit formulation
can be transferred to a computer by means of a suitable CEM I 42.5 R type Portland cement having specific gravity of
computer language before beneficiation. 3.14 and Blaine fineness of 327 m2/kg was utilized for preparing
the concrete specimens used in determination of compressive
strength and dry shrinkage.
7. Experimental validation of the proposed model The fly ash (FA) used in this research was a class F type accord-
ing to ASTM C 618 [34] (2002) and obtained from
7.1. Introduction Yumurtalik-Sugozu thermal power plant in the form of
Commercial grade. It had a specific gravity of 2.25 and the Blaine
In this stage, an experimental study was conducted determine fineness of 287 m2/kg.
the compressive strength, and drying shrinkage properties of four A commercial grade silica fume (SF) obtained from Norway was
mixes containing mineral admixtures. The hardened concretes utilized in this study. It had a specific gravity of 2.2 and the specific
tested for the compressive strength, and drying shrinkage accom- surface area (Nitrogen BET Surface Area) of 21080 m2/kg.
panied by the water loss for a drying period of 40 days. The aim
of the experimental study is to verify the proposed prediction
model presented above.
600
NN prediction
1200 Experimetal shrinkage
Autogenouse shrinkage
Predicted shrinkage (microstrain)

y = 0.982x + 8.919
(microstrain)

1000 400
R² = 0.993

800

600 200

400

200
0
0 50 100 150 200
0 Test no
0 200 400 600 800 1000 1200
Experimental shrinkage (microstrain) Fig. 5. Comparison of experimental autogenous shrinkage values with those
predicted by the proposed NN model.
Fig. 3. Predicted shrinkage values from NN vs. experimental data for training.

600
1200
NN prediction
Predicted shrinkage (microstrain)

Drying hrinkage (microstrain)

1000 Experimetal shrinkage

y = 0.990x + 4.798
800 R² = 0.990
300
600

400

200

0
0 0 50 100 150 200
0 200 400 600 800 1000 1200 Test no
Experimental shrinkage (microstrain)
Fig. 6. Comparison of experimental drying shrinkage values between 0 and 300
Fig. 4. Predicted shrinkage values from NN vs. experimental data for testing. microstrain with those predicted by the proposed NN model.
 K. Mermerdasß, M.M. Arbili / Construction and Building Materials 94 (2015) 371–379 377

Fine aggregate and coarse aggregates used for production of A sulphonated naphthalene formaldehyde superplasticizer (SP)
concrete is the mixture of crushed stone with specific gravities of with a specific gravity of 1.19 was used in all mixtures.
2.65, 2.66 respectively. In addition, the grading of the aggregates
was kept constant for concrete production. 7.3. Mix proportions

900 The mixtures were designed at 0.45 water/binder ratios (w/b).

NN prediction In the production of the concretes the w/b ratio was 0.45 and the
Experimetal shrinkage total cementitious materials content was 400 kg/m3. In codifica-
Drying shrinkage (microstrain)

tion of concretes the abbreviations F and S were used to indicate

600 replacement levels of fly ash and silica fume, respectively. As
shown in Table 3 four concrete mixes were designed.
The mixture S0F0 in Table 4 was designated as the control mix-
ture which included only ordinary Portland cement as the binder
while the remaining mixtures incorporated binary (PC + FA,
300 PC + SF) ternary (PC + FA + SF) cementitious blends in which a pro-
portion of Portland cement was replaced with the mineral admix-
tures. The maximum replacement levels for both FA and SF were
10% and 15% by weight of the total binder, respectively. When
0 preparing binary and ternary mixtures, FA and SF were replaced
0 20 40 60 80 100 by cement according to the specified replacement levels shown
Test no in Table 3.

Fig. 7. Comparison of experimental drying shrinkage values between 300 and 600
microstrain with those predicted by the proposed NN model.
7.4. Specimen preparation and curing

All concretes were mixed in accordance with ASTM C192 [35]

1,200 standard in a power driven rotating pan mixer with a 50 l capacity.
NN prediction All samples were poured into the steel moulds in two layers, each
Experimetal shrinkage of which being vibrated for a couple of seconds. After casting the
Drying shrinkage (microstrain)

moulded specimens were protected with a plastic sheet and left

in the casting room for 24 h. Thereafter, the samples of compres-
900
sive strength were demolded and cured in water until the testing
ages.

7.5. Test methods

600
7.5.1. Compressive strength
For compressive strength measurement of concretes,
150  150  150 mm cubes were tested according to ASTM C39
300 [36] by means of a 3000 kN capacity testing machine. The test
0 20 40 60 80 100 was performed on the test specimens at the ages 28 days to mon-
Test no itor the compressive strength development. The compressive
strength was computed from average of three specimens at each
Fig. 8. Comparison of experimental drying shrinkage values between 600 and 900 testing age.
microstrain with those predicted by the proposed NN model.

Table 3
1,200 Designation and composition properties of mixes.

No. Mix Cement (CEM I 42.5 R type) Fly ash Silica fume
designation (%) (%) (%)
Drying shrinkage (microstrain)

1 Plain F0S0 100 0 0

2 Binary F10S0 90 10 0
900 3 Binary F0S15 85 0 15

NN prediction Table 4
Mix proportions for concrete (kg/m3).
600 Experimetal shrinkage
Mix description SF0FA0 SF0FA10 SF15FA0 SF15FA10
Cement 400 360 340 300
FA 0 40 0 40
SF 0 0 60 60
Water 180 180 180 180
300 Fine aggregate 970.8 962.9 957.3 949.8
0 10 20 30 40 Coarse aggregate (Medium 812.7 806.0 801.3 795.1
Test no only)
Superplasticiser 4.4 3.2 8.0 6.4
Fig. 9. Comparison of experimental drying shrinkage values between 900 and 1200 Fresh unit weight (kg/m3) 2367.9 2352.1 2346.7 2331.3
microstrain with those predicted by NN and GEP.
378 K. Mermerdasß, M.M. Arbili / Construction and Building Materials 94 (2015) 371–379

7.5.2. Drying shrinkage and weight loss in SF content. While added FA to concrete mixes, compressive
Free shrinkage test specimens having a dimension of strength systematically decreases.
70  70  280 mm for each mixture were cured for 24 h at 20 °C Fig. 14 shows the tendencies of the shrinkage values obtained
and 100% relative humidity and then were demoulded. After that, from experimental study and proposed prediction models.
the specimens were exposed to drying in a humidity cabinet at
23 ± 2 °C and 50 ± 5% relative humidity, as per ASTM C157 [37]
3.5
for about 40 days. The length change was measured by means of
a dial gage extensometer with a 200 mm gage length. The shape
of the shrinkage specimens as well as the location of the reference 3.0
pins are shown in Fig. 10. Measurements were carried out every
24 h for the first 3 weeks and then 3 times a week. At the same 2.5

Weight loss (%)

time, weight loss measurements were also performed on the same
specimens. Variations in the free shrinkage strain and the weight 2.0
loss were monitored during the 41-day drying period (at
23 ± 2 °C and 50 ± 5% relative humidity) and the average of three 1.5
prism specimens were used for each property. SF0FA0
1.0 SF0FA10
7.6. Discussion of experimental results SF15FA0
0.5 SF15FA10
Free shrinkage strain developments of the concretes are
depicted in Fig. 11; it can be observed that Mix containing 0.0
SF15FA10 showed higher shrinkage than the other mixtures. The 0 10 20 30 40
highest shrinkage in SF15FA10 mix in the age of 40 days is found Drying period (days)
515 microstrain. The lowest shrinkage in SF0FA10 mix at the age
40 days is found to be lower than those in control mixture. Fig. 12. Weight loss of concrete.
Weight losses of the concretes for 40 days of drying period are
illustrated in Fig. 12. The maximum weight loss of 3.15% was
observed in SF15FA10 concrete while the minimum was observed
at control concrete as 2.2%.
Fig. 13 shows the compressive strength values of concrete. The
maximum value observed in FA0SF15. The figure indicated that
there was an increase in compressive strength with the increase

Fig. 13. Compressive strength of concrete.

Fig. 10. Free shrinkage specimens.

Experimental NN model
600
700
SF0FA0 SF0FA10 SF15FA0 SF15FA10
Drying shrinkage (microstrain)

500 600
Shrinkage (microstrain)

500
400
400
300
300

200 SF0FA0 200

SF0FA10
100 SF15FA0 100
SF15FA10
0
0 0 20 40 60 80 100
0 10 20 30 40
Test No
Drying period (days)
Fig. 14. Comparison between the proposed model and experimental drying
Fig. 11. Shrinkage of concretes over 40 days of drying period. shrinkage values.
 K. Mermerdasß, M.M. Arbili / Construction and Building Materials 94 (2015) 371–379
Although the model showed similar trends to that of experimental
study, the best performance seemed to be obtained for SF0FA10
concrete. The proposed NN model illustrated almost prefect estima-
tion capability for all four types of concrete. This proves that the
proposed NN model can be reliably utilized for prediction purpose.
8. Conclusions
Based on the mathematical modeling and experimental results
reported above, the following conclusions can be drawn:
 Numerical modeling of shrinkage of concrete containing min-
eral admixtures was conducted using neural network (NN). To
this aim, available experimental data presented in the existing
literature were used to derive the prediction model. In order
to evaluate the efficiency and advantages, the performance of
the proposed model was compared to that provided by the col-
lected data in the previous studies. The comparison of the pre-
dicted values with the collected data referred that the NN
models provides excellent prediction performance, especially
for the actual shrinkage values of higher than 200 lstrain and
300 lstrain respectively for autogenous and drying shrinkage.
In short, the accuracy of the proposed model is found to be good
enough to be utilized for prediction purposes.
 The prediction model for shrinkage estimation of the concretes
produced with fly ash and silica fume can efficiently be con-
structed using NN. The constructed NN model showed a good
performance on both training and testing data sets. The correla-
tion coefficients of 0.993 and 0.990 were obtained for training
and testing data sets, respectively.
 Experimental study indicated that utilization of mineral admix-
tures affected the shrinkage behaviors of concretes significantly.
The highest shrinkage strain development was observed to be
515 lstrain for SF15FA10 concrete while SF0FA10 concrete
demonstrated the lowest value of 393 lstrain at 40 days. It
could be due to the fact that FA has low pozzolanic reactivity,
and hence autogenous shrinkage at early ages is low. Control
concrete (0% FA, 0% SF) and SF15FA0 concrete indicated almost
similar behavior in shrinkage strain development.
 The highest compressive strength value at the end of 28 days of
curing was observed for FA0SF15 concrete as 86.3 MPa while
the lowest compressive strength of 76.5 MPa was obtained for
FA10SF0 concrete. The improvement of concrete was due to
high pozzolanic reaction of SF and its micro filling effect.
 The comparison of the shrinkage value obtained from the pro-
posed models with the observed experimental results of this
study proved that NN model can reliably be benefited as a pre-
diction tool. Almost perfect estimation performance was
observed from the concretes with mineral admixtures
(SF0FA10, SF15FA0, SF15FA10).
References
[1] M. Tia, Y. Liu, B. Haranki, Y.M. Su, Modulus of elasticity, creep and shrinkage of
concrete. Final Report, U.F. Project No. 00059678, Gainesville: Department of
Civil & Coastal Engineering, College of Engineering, 2009, 203p.
[2] E.H. Yang, Y. Yang, V.C. Li, Use of high volumes of fly ash to improve ECC
mechanical properties and material greenness, ACI Mater. J. 104 (6) (2007)
620–628.
[3] X. Wang, F. Bektas, P. Taylor, K. Wang, P. Tikalsky, Drying shrinkage behavior of
mortars made with ternary blends, J. Transport. Res. Board (2012) 52–59.
[4] V. Chandwani, V. Agrawal, R. Nagar, Applications of soft computing in civil
engineering: a review, Int. J. Comput. Appl. 81 (2013). 0975–888.
[5] D.K. Chaturvedi, Soft computing techniques and its applications in electrical
engineering, Stud. Comput. Intel. 103 (2008) 1–10.
[6] R.D. Vanluchene, R. Sun, Neural networks in structural engineering, Comput.-
Aid. Civil Infrastruct. Eng. 5 (3) (1990) 207–215.
_ Topçu, A.R. Boğa, F.O. Hocaoğlu, Modeling corrosion currents of reinforced
[7] I.B.
concrete using ANN, Automat. Constr. 18 (2) (2009) 145–152.
379
[8] C.H. Lim, Y.S. Yoon, J.H. Kim, Genetic algorithm in mix proportioning of high
performance concrete, Cem. Concr. Res. 34 (3) (2004) 409–420.
[9] E.M.R. Fairbairn, M.M. Silvoso, R.D.T. Filho, J.L.D. Alves, N.F.F. Ebecken,
Optimization of mass concrete construction using genetic algorithms,
Comput. Struct. 82 (2–3) (2004) 281–299.
[10] F. Özcan, C. Atis, O. Karahan, E. Uncuoglu, H. Tanyildizi, Comparison of artificial
neural network and fuzzy logic models for prediction of long term
compressive strength of silica fume concrete, Adv. Eng. Softw. 40 (2009)
856–863.
_
[11] G. Inan, A.B. Göktepe, K. Ramyar, A. Sezer, Prediction of sulfate expansion of PC
mortar using adaptive neuro-fuzzy methodology, Build. Environ. 42 (3) (2007)
1264–1269.
[12] M.A. Shahin, H.R. Maier, M.B. Jaksa, Predicting settlement of Shallow
Foundations using Neural Networks, J. Geotech. Geoenviron. Eng. 128 (9)
(2002) 785–793.
_
[13] I.B. Topçu, M. Sarıdemir, Prediction of properties of waste AAC aggregate
concrete using artificial neural network, Comput. Mater. Sci. 41 (1) (2007)
117–125.
_
[14] I.B. Topçu, M. Sarıdemir, Prediction of compressive strength of concrete
containing fly ash using artificial neural networks and fuzzy logic, Comput.
Mater. Sci. 41 (3) (2008) 305–311.
[15] B.B. Adhikary, H. Mutsuyoshi, Prediction of shear strength of steel fiber RC
beams using neural networks, Constr. Build. Mater. 20 (9) (2006) 801–811.
[16] K. Mermerdasß, E. Güneyisi, M. Gesoğlu, T. Özturan, Experimental evaluation
and modeling of drying shrinkage behavior of metakaolin and calcined kaolin
blended concretes, Constr. Build Mater. 43 (2013) 337–347.
[17] Z.H. Duan, S.C. Kou, C.S. Poon, Prediction of compressive strength of recycled
aggregate concrete using artificial neural networks, Constr. Build Mater. 40
(2013) 1200–1206.
[18] A.S. Tokar, P.A. Johnson, Rainfall-runoff modeling using artificial neural
networks, J. Hydrol. Eng. 4 (3) (1999) 232–239.
[19] A.F. Ashour, L.F. Alvarez, V.V. Toropov, Empirical modeling of shear strength RC
deep beams by genetic programming, Comput. Struct. 81 (5) (2003) 331–338.
[20] M. Sarıdemir, Effect of specimen size and shape on compressive strength of
concrete containing fly ash: application of genetic programming for design,
Mater. Des. 56 (2014) 297–304.
[21] S.C. Lee, S.W. Han, Neural-network-based models for generating artificial
earthquakes and response spectra, Comput. Struct. 80 (2002) 1627–1638.
[22] M. D’Aniello, E. Mete-Güneyisi, R. Landolfo, K. Mermerdasß, Predictive models
of the flexural overstrength factor for steel thin-walled circular hollow section
beams, Thin-Wall. Struct. 94 (2015) 67–78.
[23] E. Mete-Güneyisi, M. Gesoğlu, E. Güneyisi, K. Mermerdasß, Assessment of shear
capacity of adhesive anchors for structures using neural network based model,
Mater. Struct. (2015), http://dx.doi.org/10.1617/s11527-015-0558-x.
[24] E. Mete-Güneyisi, M. D’Aniello, R. Landolfo, K. Mermerdasß, Prediction of the
flexural overstrength factor for steel beams using artificial neural network,
Steel Compos. Struct. 17 (2014) 215–236.
[25] M. D’Aniello, E. Mete-Güneyisi, R. Landolfo, K. Mermerdasß, Analytical
prediction of available rotation capacity of cold-formed rectangular and
square hollow section beams, Thin-Wall. Struct. 77 (2014) 141–152.
[26] M. Gesoğlu, E. Mete-Güneyisi, E. Güneyisi, E. Yılmaz, K. Mermerdasß, Modeling
and analysis of the shear capacity of adhesive anchors post-installed into
uncracked concrete, Compos. B Eng. 60 (2014) 716–724.
[27] J. Schalkoff, Artificial Neural Networks, McGraw-Hill, New York, 1997.
[28] M.Z. Susac, N. Sarlija, M. Bensic, Selecting neural network architecture for
investment profitability predictions, J. Inform. Org. Sci. 29 (2) (2005) 83–95.
[29] M.H. Zhang, C.T. Tam, M.P. Leow, Effect of water-to-cementitious materials
ratio and silica fume on the autogenous shrinkage of concrete, Cem. Concr. Res.
33 (2003) 1687–1694.
[30] W. Wongkeo, P. Thongsanitgarn, A. Chaipanich, Compressive Strength of
Binary and Ternary Blended Cement Mortars Containing Fly Ash and Silica
Fume Under Autoclaved Curing TIChE International Conference 2011, Hatyai,
Songkhla, Thailand, 2012
[31] S. Yoo, S. Kwon, S. Jung, Analysis technique for autogenous shrinkage in high
performance concrete with mineral and chemical admixtures, Constr. Build.
Mater. 34 (2012) 1–10.
[32] J. Khatib, Performance of self-compacting concrete containing fly ash, Constr.
Build. Mater. 22 (2008) 1963–1971.
[33] R. Khatri, V. Sirivivatnanon, Effect of different supplementary cementitious
materials on mechanical properties of high performance concrete, Cement
Concr. Res. 25 (1995) 209–220.
[34] ASTM C 618. Standard Specification for Coal Fly Ash and Raw or Calcined
Natural Pozzolan for Use in Concrete. American Society for Testing and
Materials. Annual Book of ASTM Standards, vol. 04.02. West Conshohocken,
PA: ASTM, 2002.
[35] ASTM C192, Standard Practice for Making and Curing Concrete Test Specimens
in the Laboratory. American Society for Testing and Materials. Annual Book of
ASTM Standards, vol. 04.02. West Conshohocken, PA: ASTM, 2014.
[36] ASTM C39, Standard Test Method for Compressive Strength of Cylindrical
Concrete Specimens. American Society for Testing and Materials. Annual Book
of ASTM Standards, vol. 04.02. West Conshohocken, PA: ASTM, 2012.
[37] ASTM C157, Standard Test Method for Length Change of Hardened Hydraulic-
Cement Mortar and Concrete. American Society for Testing and Materials.
Annual Book of ASTM Standards, vol. 04.02. West Conshohocken, PA: ASTM,
2008.