You are on page 1of 32

Neural Computing and Applications (2021) 33:4501–4532

https://doi.org/10.1007/s00521-020-05244-4(0123456789().,-volV)(0123456789().
,- volV)

ORIGINAL ARTICLE

A comparative study of ANN and ANFIS models for the prediction


of cement-based mortar materials compressive strength
Danial Jahed Armaghani1 • Panagiotis G. Asteris2

Received: 12 December 2019 / Accepted: 24 July 2020 / Published online: 10 August 2020
 Springer-Verlag London Ltd., part of Springer Nature 2020

Abstract
Despite the extensive use of mortars materials in constructions over the last decades, there is not yet a reliable and robust
method, available in the literature, which can estimate its strength based on its mix parameters. This limitation is due to the
highly nonlinear relation between the mortar’s compressive strength and the mixed components. In this paper, the
application of artificial intelligence techniques toward the prediction of the compressive strength of cement-based mortar
materials with or without metakaolin has been investigated. Specifically, surrogate models (such as artificial neural
network, ANN and adaptive neuro-fuzzy inference system, ANFIS models) have been developed to the prediction of the
compressive strength of mortars trained using experimental data available in the literature. The comparison of the derived
results with the experimental findings demonstrates the ability of both ANN and ANFIS models to approximate the
compressive strength of mortars in a reliable and robust manner. Although ANFIS was able to obtain higher performance
prediction to estimate the compressive strength of mortars compared to ANN model, it was found through the verification
process of some other additional data, the ANFIS model has overfitted the data. Therefore, the developed ANN model has
been introduced as the best predictive technique for solving problem of the compressive strength of mortars. Furthermore,
using the optimum developed model an ambitious attempt to reveal the nature of mortar materials has been made.

Keywords Artificial neural networks  Cement  Compressive strength  Metakaolin  Mortar  Artificial intelligence
techniques  Adaptive neuro-fuzzy inference system

Abbreviations logsig Log-sigmoid transfer function


ANFIS Adoptive neuro-fuzzy inference system MDA Maximum diameter of aggregate
ANNs Artificial neural networks MK/B Metakaolin percentage in relation to total binder
AI Artificial intelligence ML Machine learning
AS Age of the specimen PSO Particle swarm optimization
BNNs Biological neural networks purelin Linear transfer function
BPNNs Back-propagation neural networks SVM Support vector machine
B/S Binder to sand ratio SP Superplasticizer in relation to the total binder
CS Compressive strength Tansig Hyperbolic tangent sigmoid transfer function
DNNs Deep neural networks W/B Water-to-binder ratio (W/B)
FIS Fuzzy inference system
1 Introduction
& Panagiotis G. Asteris
asteris@aspete.gr; panagiotisasteris@gmail.com Mortars are important elements of masonry structures, as
Danial Jahed Armaghani they are the material used to join stone and/or brick ele-
danialjahedarmaghani@duytan.edu.vn ments comprising a masonry. Mortars consist of binder
1 materials, aggregates and, in some cases, additives. The
Institute of Research and Development, Duy Tan University,
Da Nang 550000, Vietnam choice today in relation to contemporary construction is the
2 use of cement as binding material. However, efforts are
Computational Mechanics Laboratory, School of Pedagogical
and Technological Education, 14121 Heraklion, Athens, being made to substitute cement in the mortar mix in the
Greece current constructions, aiming to minimize the

123
4502 Neural Computing and Applications (2021) 33:4501–4532

environmental impacts of the cement industry and in an literature. In the study conducted by Safa et al. [47], the
effort to improve the life cycle assessment of the mortars shear strength values of steel–concrete composite beams
used in construction [1]. A material which has been used in were predicted according to their effective parameters.
this context, with the added effect of improving cement They showed that the developed ANFIS model is capable
mortar characteristics, is metakaolin. Although much of providing high-performance prediction for shear
research has been conducted regarding these materials, no strength values. In addition, they concluded that the con-
tool yet exists which can assist in a quantitative manner in crete compression strength is the most influential factor on
the optimum design of cement-based mortars. This is the shear strength capacity of composite beam. The feasi-
attributed to the fact that the mechanical properties of bility of using the ANFIS technique to estimate the ser-
mortar materials exhibit a strong nonlinear nature derived viceability behavior of near surface mounted strengthened
from the parameters involved in their composition. Due to reinforced concrete beams has been investigated by ud
existing this nonlinear behavior, the development of an Darain et al. [50]. The obtained error values by ANFIS
analytical formula for the prediction of the mechanical indicated that this technique is an excellent alternative for
properties using deterministic methods is considered as a the experimental investigations. In another investigation of
difficult task. ANFIS, Naderpour and Mirrashid [51] tried to estimate the
The successful use of artificial intelligence (AI) and moment capacity values of spiral-reinforced concrete col-
machine learning approaches has been highlighted in the umns. They showed that the ANFIS model obtains a high
literature [2–10]. Artificial neural network (ANN), as a capacity of performance prediction in their investigated
benchmark of the AI techniques, has been developed in field. Based on above discussion, it can be concluded that
past decades as smart meta-modeling techniques, which are the ANN and ANFIS are considered as powerful, appli-
able to solve a vast number of science and engineering cable and proper AI systems that are able to solve almost
problems (e.g., material science). In order to solve the any problem in science and engineering.
problems, according to Alexandridis [11], ANNs have the In this paper, the mechanical characteristics of mortar
ability to approximate critical quantities without having to materials have been examined developing AI techniques.
measure them. Notably, these models can be built using In particular, ANN as a powerful technique was selected
even a limited number of datasets in training stage. In fact, and proposed to estimate the compressive strength of
by proposing these models, the required time as well as cement-based mortars. Additionally, an ANFIS model was
money-consuming experiments will be reduced in many developed for the same purpose. For the developments of
science and engineering issues. Thus far, several studies the mentioned AI techniques, a database consisting of 276
highlighted the successful use of ANNs for estimating the specimens, taken from the literature, was utilized. Namely,
mechanical properties of concrete materials [12–20]. based on this database, two parameters related to sample
Asteris et al. [20] used ANNs to estimate the compressive properties (the age of the specimen, AS and the maximum
strength of self-compacting concrete through a training diameter of aggregate, MDA), as well as four parameters of
process involving as input parameters the eleven parame- synthesis (metakaolin to total binder, MK/B, ratio, water-
ters of synthesis with output parameter the value of the to-binder, W/B, ratio, superplasticizer, SP, binder-to-sand,
compressive strength. Moreover, similar methods such as B/S, ratio), were used as input parameters, while the value
fuzzy logic and genetic algorithms have also been applied of compressive strength, CS was used as output parameter.
for modeling the compressive strength of concrete mate- The optimum developed ANN and ANFIS models have
rials [21–29]. The detailed relevant reports can be found in proven to be very successful, exhibiting a high level of
the literature [30–35]. reliabilities.
Adoptive neuro-fuzzy inference system or ANFIS is a
combination intelligent technique of both fuzzy inference
system (FIS) and ANN. In fact, ANFIS is able to engage 2 Significance of the subject
advantages of these two models into a uniform solution to
solve science and engineering problems. In field of civil Much research has been conducted internationally regard-
engineering, the successful implement of this technique in ing the addition of metakaolin in cement mortars, substi-
various applications, e.g., geotechnical engineering tuting a percentage of cement content in the mortar mix, in
[36–38], hydraulic engineering [39, 40], transportation order to achieve a mortar of enhanced characteristics. Due
engineering [41, 42], construction management [43, 44], to the nonlinear behavior between mixed components and
environmental engineering [45, 46] and more specifically mortar characteristics, it is difficult to predict the CS of a
structural engineering [47–49] has been reported in the mortar mix, and thus, this difficulty leads to the need for

123
Neural Computing and Applications (2021) 33:4501–4532 4503

costly and time-consuming experiments based on empirical weighted links to exchange information (signals) and
calculations of the appropriate mortar synthesis parameters. activation functions to process them [66]. For detailed
To this end, AI techniques, such as ANN and ANFIS, description of the ANN model and its training system and
can contribute as feasible and applicable tools for the implementation, some other references can be found in the
approximation of the mechanical properties of concrete literature [67].
[24, 52–61]. In this study, AI techniques have been
developed for the prediction of the CS of mortars, using six 3.2 ANFIS
parameters, i.e., AS, MDA, MK/B, W/B, SP and B/S as
model inputs. ANFIS is a hybrid model that consists of neural network
and FIS. In general, FIS is widely appreciated because of
its capacity of mapping prior knowledge into some con-
3 Materials and methods straint sets. The created sets can be employed at the net-
work topology level in order to make the search space
3.1 Artificial neural networks optimized [68]. Together with FIS, neural networks with
BP are also integrated into ANFIS in a way to achieve
ANN is an artificial computational system configured from automation in tuning parameters of fuzzy controller [69].
simulating organizational principles of the nervous system ANFIS has the necessary capacity to do monotonic tuning
functions. Having ability of learning automatically from in network using the Takagi–Sugeno (TS) controller in the
the given training patterns, ANN can solve a mapping right-hand side tuning. If ANFIS is generalized and holds
problem by finding close approximation of correlation the following neural characteristics, it will be able to
input data with output data, and such a feature distinct it estimate the CS of cement-based mortars. Such estimation
from other common expert systems [62]. An ANN com- capacity is expected where the triangular, trapezoidal and
puting system is composed of artificial neurons, which play Gaussian membership functions (MFs) [70] have been
the role of fundamental units and resemble parallel process examined concerning the datasets along with differentiable
of a biological brain to find the answer. T-norm rules [71]. Typically, ANFIS employs learning
The ANN network behavior is affected by pattern of algorithms through a certain mechanism that comprises
neurons connections, which determine class of network as two phases: forward pass with the least square error, which
well. Mentioned earlier, train a network in order to improve is known as an offline learning mechanism; then, in the
network performance, is conceivable. In more exact terms, second phase, the gradient descent algorithm is used with
structure and connection weights of network modify iter- BP. A schematic diagram of a simple ANFIS structure for
atively so that the error that refers to every output layer TS-FIS is depicted in Fig. 1. As this figure shows, in
node, becomes minimum. A squared error function shows ANFIS, two inputs and a single output are involved.
by E, calculate the result output error as: In the FIS structure of TS, there are five layers of per-
P  2 ceptron’s or neurons. Within one layer, the perceptron’s or
1X
E¼ tðiÞ  yðiÞ ð1Þ neurons are similar to each other with the identical func-
2 i¼1 tionalities as follow [72, 73]:
where t, y and P stand for target value, produced actual (1) Layer 1 (Fuzzyfing Layer): Neurons in this layer are
value and the number of training patterns, respectively. considered as adaptive nodes that comprise premise
Back-propagation (BP) learning algorithm is a gradient- parameters.
based learning procedure usually used for network learning (2) Layer 2 (Implication Layer): The neurons are labeled
task, particularly regarding multilayer feed-forward nets P and shown by a circle. The output node, then, is
[63, 64]. A twofold procedure that includes both forward formed based on incoming signals. The output node
and backward stages consists of any training period in BP wi indicates the firing strength of a rule.
learning algorithm. Input signals move forwards through (3) Layer 3 (Normalizing Layer): Every neuron in this
the network in forward stage and expel error signal for each layer is a fixed neuron to be identified by a circle and
output layer node. Then, rates of resulting error cross labeled as N. The output is achieved according to the
backward along the network and correct weights and biases ratio of the ith rule’s firing strength over the
of network in subsequent step [65]. ANNs classify as two summation of firing strength of all rules.
functional groups regarding to network architecture: feed- (4) Layer 4 (Defuzzyfing Layer): Neuron that exist
forward and feedback. Among multilayer perceptron is a within this layer are actually adaptive neurons
popular variant of multilayer feed-forward networks, that containing consequence parameters.
its successive layers of processing units (neurons) using

123
4504 Neural Computing and Applications (2021) 33:4501–4532

Fig. 1 Overall structure of


ANFIS model with two input
parameters

(5) Layer 5 (Combining Layer): This layer contains a procedure, but in terms of real time between the production
single neuron adding up all the inputs. of the specimen and its actual measurement; during this
time, it must be stored, and this demands space of specific
requirements and the subsequent cost. This is the reason
3.3 Cement mortar database
that many researchers investigate mortar characteristics up
to 28 or 90 days. In the literature, there is available data of
A prerequisite for the successful function of the AI tech-
CS obtained from specimens of larger ages, however, data
niques is to establish a reliable database which is able to be
at these higher ages are relatively scarce, with the conse-
well-trained. In the case of mortar, this poses an issue due
quent problems incurred on the AI prediction.
to many factors. Mortars as composite materials containing
Thus, the compiling of an adequate experimental data-
binder material and aggregates, while in most cases, nat-
base is achieved through the accumulation of smaller
ural/manufactured additives are utilized. Therefore, mor-
databases acquired by different researchers and available in
tars are a mixture of water with various natural and
the relevant literature. During the compiling of the data-
manufactured raw materials. During the configuration of
base, the reliability of each individual database must be
the database, it is essential to distinguish the necessary mix
examined. In particular, the raw materials used must be
parameters. Furthermore, it is crucial to be accurate
adequately described; the type of cement and metakaolin is
regarding the type of raw materials used, in order to train
of the utmost importance, as different types must be dis-
the system appropriately. For example, the effects of using
cerned due to their different effects on CS. Furthermore, it
the ordinary Portland cement (OPC) on the final CS are
is important to note that the same standards have been
different compared to the use of high alumina cement or
followed during the experimental procedure, in order for
white cement. Hence, the types and amounts of cement in
the results to be meaningfully comparable. An adequate
the mortars will have a deep impact on their CS and sub-
number of specimens must have been tested in order for the
sequently in the published database for AI modeling.
values to be statistically acceptable; a small amount of
Another issue is related to the difficulty of one
tested specimens, regardless of credibility, cannot give a
researcher obtaining a large enough amount of experi-
result that can be considered reliable. When training an AI
mental data capable of adequately training the AI tech-
technique, in addition to the reliability of the database, it is
niques. Mortars are produced, casted into steel molds in
crucial that the values of the input parameters (mortar mix
order to set (dimensions of molds and mortar curing con-
synthesis parameters and specimen age) cover all possible
ditions related on standard used and on mortar type pro-
value ranges of the parameters. It is no exaggeration to
duced) and then removed from molds and preserved in
state that the reliability of the optimum developed AI
specific environmental conditions until the testing date,
techniques is crucially dependent on the reliability of the
when the mortar specimen has reached the desired age. It is
experimental data, thus confirming the famous expression
obvious that the production of a very large amount of
in the field of informatics Garbage In, Garbage Out
specimens is problematic and costly, while CS measure-
(GIGO). Predictive analytics begins with good data; more
ments are time-consuming, not in terms of experimental

123
Neural Computing and Applications (2021) 33:4501–4532 4505

data do not necessarily mean better data. A successful 6.00 mm due to the fact that are available in literature
predictive analytics project requires, first and foremost, are less than ten having as a result their inclusion in
relevant and accurate data. It is obvious from the above that database to reduce the reliability of the proposed
mortars present certain difficulties in predicting their CS. mathematical model,
This complexity of course is the reason that the use of AI • Also from the works used to extract the datasets of the
techniques is intriguing in relation to predicting the final database used herein all the specimens with age greater
characteristics of the mortar after setting and hardening, as to 90 days have also been removed,
no linear approach can be successfully applied. • Only experimental data based only on the six param-
In light of the above discussion, a large database has eters used herein have been used. Namely, these six
been composed. Specifically, the database used herein parameters AS, MDA, MK/B, W/B, SP and B/S have
consists of 276 experimental data sets that have been been used.
obtained from 10 experimental studies available in the • For data which corresponded to the same experiment,
literature (Table 1). Detailed and in-depth the database is with the same mechanical and geometric characteris-
presented in Table 8 of Appendix. tics, however, with different strength values, the
Despite the fact that a large number of databases are average of the measured mechanical strength of the
available in the relevant literature, it was decided to com- different experiments was used as output values for the
pile a new database, in contrast to complementing an respective dataset.
existing database. This decision was made for the follow-
The experimental data selected from the literature were
ing reasons: (1) the existing databases took into account
that of mortars with OPC (cement) as main binding
using different input parameters than the current research,
material and the addition of high-quality metakaolin in
(2) quite frequently different values are given for the same
different percentages in order to ensure the consistency of
experiments (incorrect transliteration of the original
the experimental data. Namely, Vu et al. [74] produced
experimental data), but above all, because (iii) the datasets
cement mortars, using Portland cement, metakaolin, sand
are not distributed to cover the full range of input param-
and superplasticizer. Mortar mixes consisted of 1 part
eter values, but only a certain area of the whole range. It is
binder and 2.75 parts sand by weight. They managed to
worth noting that the number of a database’s datasets is not
measure CS of 144 samples, at four different ages (7 days
enough to ensure its reliability; the distribution of input
to 3 months), with varying water-to-binder ratio
parameter values taken into account, however, plays a
(0.40–0.53), and also with partial replacement of cement
crucial role in this procedure. That is, if the model can
with metakaolin in various percentages (ranging from 0 to
represent and manage the knowledge available through
30% with a 5% step increment). The superplasticizer
experiments for the totality of parameters and, if possible,
(naphthalene sulfonate-based type TM OFT-3) was added
for the whole range of their values.
in varying percentages (0%, 0.5% and 1.4%). The speci-
In light of the above, the database used herein was
mens were cast in steel molds of dimensions 10 cm 9 40
compiled, abiding by the following principles:
cm 9 40 cm. CS was measured in accordance with ASTM
• It has been decided not to include mortars with C109-80 (1983) on six cubic samples 10 cm 9 10 cm 9
maximum diameter of aggregates greater than 10 cm. Courard et al. [75] produced cement mortar

Table 1 Data from 10


Nr. Reference Number of samples Parameters Compressive strength
experimental studies published
in the literature MK SP (MPa)

1 Vu et al. [74] 144 H H 15.45–54.90


2 Courard et al. [75] 15 H 27.40–71.20
3 Parande et al. [76] 20 H 22.00–67.00
4 Sumasree and Sajja [77] 35 H 24.93–35.71
5 Batis et al. [78] 16 H 17.60–69.70
6 Kadri et al. [79] 12 H H 35.21–99.177
7 Mardani-Aghabaglou et al. [80] 6 H 37.27–56.76
8 Potgieter-Vermaak and Potgieter [81] 20 H 23.40–92.80
9 Cizer et al. [82] 5 H 47.94–66.23
10 Al-Chaar et al. [83] 3 H 6.06–10.73
Total 276 6.06–99.17

123
4506 Neural Computing and Applications (2021) 33:4501–4532

specimens, using ordinary Portland cement, metakaolin in constant in the two mixes (with and without metakaolin), at
different percentages (0–20% with a 5% increment) and 0.37 by weight, while the water-to-binder ration was also
sand. The binder-to-sand ratio was constant at 1/3 by kept constant at 0.485 by weight. Superplasticizer was not
weight. The mortars were cast into 4 cm 9 4 cm 9 16 cm added in any mix. CS was measured on cubic samples (50
steel molds and tested at different ages, ranging from cm x 50 cm x 50 cm) in accordance with ASTM C 109 at
3 days to 28 days. CS was measured in accordance with three different ages (7, 28 and 90 days), with values
NBN B12-208 (1969) on six samples (dimensions ranging from 37.27 to 56.76 MPa.
4 cm 9 4 cm 9 4 cm). In the study of Parande et al. [76], Potgieter-Vermaak and Potgieter [81] examined the use
ordinary Portland cement is used with aggregates and of metakaolin as an extender of South African cement.
metakaolin (0–20% with a 5% increment) to produce They produced cement mortars, using a local ordinary
blended MK-cement mortars. The binder-to-sand ratio was Portland cement, local metakaolin, heated at different
constant at 1/3 by weight, and the water binder ration was temperatures and with different activation times, and sand.
constant at 0.40. Cube mortar specimens were produced For the database herein compiled, the mortar mixes where
(10 cm 9 10 cm 9 10 cm) and measured at different ages metakaolin was heated at 750 C were taken into account,
(3 days to 90 days). Sumasree and Sajja [77] studied the as to be in accordance with the metakaolin used by the
CS of OPC/OPC-metakaolin mortars at five different ages, other researchers as well. The mortar mixes were produced
ranging from 3 days to 56 days. Metakaolin replaced with different metakaolin percentages in relation to total
cement in the mix at different percentages, ranging from 0 binder (0–30% with an increment of 10%). The water-to-
to 30% with a 5% increment. Binder to sand remained at binder and the binder-to-sand ratios were kept constant at
0.50 by weight and water/binder ratio remained constant at 0.38 and 0.33 (by weight), respectively. The specimens
0.46. CS measurements were conducted on specimens (dimensions 4 cm 9 4 cm 9 4 cm) were crushed to fail-
4 cm 9 4 cm 9 4 cm. Batis et al. [78] also studied the ure on a Farnell cube press in order to obtain the CS value
effect of metakaolin on the CS of cement mortars, pro- (mean of six specimens). Measurements were conducted at
duced through mixing ordinary Portland cement with sand, 1, 2, 14 and 28 days, yielding CS values ranging from 23.4
while metakaolin was added in various cement substitution to 92.8 MPa.
ratio (0%, 10%, 20%). Water-to-binder remained constant Cizer et al. [82] examined the CS of blended cement-
for all specimens at 0.6, as well as the binder-to-aggregates lime mortars, for different ratios, using Portland cement
ratio, which was kept at 0.33 by weight. CS was measured (CEM I 52.5 N), commercial lime hydrate and lime putty.
at four ages (varying from 1 day to 28 days) and was CS measurements were conducted using standard mortar
performed in accordance with EN 196-1 (1994). The prisms at 3, 7, 28, 60, 90 and 180 days of hardening, in
blended cement metakaolin mortars presented CS values accordance with EN 196-1. Only the control cement mortar
ranging from 17.6 to 69.70 MPa taking into account all was incorporated in the database presented herein. Binder
measured ages. to sand was 1:3 by weight and w/b was 0.45 for the cement
Kadri et al. [79] studied the influence of metakaolin on reference mortar taken into account and CS ranged from
the development of CS in cement mortar systems, produced 47.94 MPa at the smallest age to 64.27 at 180 days, while
with the use of ordinary Portland cement, metakaolin in 28 days and 90 days CS presented even higher values than
two different substitution percentages (0% and 10% of total 180 days, 66.23 MPa and 66.01 MPa, respectively.
binder) and sand. Superplasticizer was added in three dif- Al-Chaar et al. [83] examined the use of natural poz-
ferent percentages in relation to the binder material (1.4%, zolans as a partial substitution of cement in mortars, using
2.3% and 2%), while the water-to-binder ratio was kept Type I Portland cement. The experimental procedure was
constant for all mortar mixes at 0.36. The binder-to-sand conducted in accordance with ASTM C109/C109M-08,
ration (by weight) was also kept constant at 0.5. The and cubic specimens (50 cm 9 50 cm 9 50 cm) were
mortars were studied at 1, 7, 28 and 56 days in relation to used for CS measurements. In the database presented
their CS, while the CS measurements were conducted on herein, only the reference mortar characteristics were
4 cm 9 4 cm 9 4 cm specimens, abiding by EN 196-1 incorporated (that is for 0% cement replacement), as the
(1994). substitution with natural pozzolan results in different
Mardani-Aghabaglou et al. [80] studied the mechanical modification of the cement mortar mixes’ characteristics in
performance of cement mortar mixes, using a CEM I 42.5 relation to substitution with metakaolin, which is a highly
R type cement conforming to EN 197-1 (2011) standard reactive artificial pozzolan. Water-to-binder ratio was 0.48,
and standard sand conforming to EN 196-1 (1994) stan- while binder-to-sand ratio was 0.36. CS measurements
dard. Metakaolin was also added in one mix, substituting were conducted at 7, 38, 90 and 180 days and strength
cement in a percentage of 10% per weight in relation to values ranged from 6.06 MPa at the smallest age to 12.51
total binder materials. The binder-to-sand ration was kept at the highest.

123
Neural Computing and Applications (2021) 33:4501–4532 4507

Based on the above database, each input training vector input variables may result to poor efficiency of the methods
p is of dimension of 1 9 5 and consists of the value of the and to the difficulty in construing the effects of the
age and the values of the four parameters of synthesis, expository variables on the respond. Subsequently, the
namely the percentage of metakaolin in relation to the total correlation coefficients between all possible variables are
binder materials (MK/B, w/w %), the water-to-binder ratio specified and presented in Table 3.
(W/B), calculated as the weight of water divided by the As can be seen in the table, there are not significant
weight of total binder materials (w/w), the superplasticizer correlations between the independent input variables. On
(SP), meaning the addition of superplasticizer in relation to the other hand, in order to develop reliable, robust and
the total binder (%w/w), and the binder-to-sand ratio (B/S), optimum AI models, the correlation coefficients between
meaning the w/w of binder materials to aggregate materi- the input variables (parameters) and the output parameter
als. The corresponding output training vectors are of (CS), last highlighted line in Table 3, are needed to be as
dimension 1 9 1 and consist of the value of the CS of the great as possible. Based on these values, it is clearly shown
mortar specimens. Their mean values together with the that there is a strong relation between the mortar CS and
minimum, maximum values as well standard deviation the input parameters, i.e., AS, W/B and SP. It is worth
(STD) values are listed in Table 2. mentioning that the analyses presented in Table 3 were
Moreover, Figs. 2 and 3 demonstrate both the frequency performed using a statistical software package of SPSS
histograms of the parameters and the value of the CS vs the version 16 [84].
parameters of the mortars. These figures are extremely
useful as it demonstrates the complex and nonlinear 3.4 Sensitivity analysis of the compressive
behavior of the mortar mixes. In fact, this nonlinear strength based on experimental database
behavior is the reason that deterministic methods are not
able to provide a reliable analytical formula for the pre- In general, sensitivity analysis (SA) of a numerical model
diction of mortar CS. In addition, despite the fact that the is a technique used to determine if the output of the model
database used herein is among the largest presented and is affected by changes in the input parameters. This will
used in the relevant literature, Figs. 2 and 3 demonstrate provide feedback as to which input parameters are the most
which parameter value ranges have not been investigated significant, and thus, by removing the insignificant ones,
as yet in a sufficient manner and are in need of further the input space will be reduced and subsequently the
experimental investigation and research. At the same time, complexity of the model as well as the training times
knowing which parameter value ranges are lacking in required for its training will be also reduced. In order to
experimental data also provides the invaluable knowledge identify the effects of model inputs on the output, the SA
of the areas of high uncertainty in relation to the reliability can be conducted on the database. Sometimes, the results
of the proposed numerical model. This particular issue of SA could help the researchers/designers to remove one
shall be further analyzed in the results and discussion or more input parameters from the database to have better
section, where the plans of the authors for future research analyses with higher level of performance prediction. To
shall also be presented. perform the SA, the cosine amplitude method (CAM)
Furthermore, some of the cement metakaolin mortars which was used by many researchers (e.g., [85–87]) was
variables could be dependent on each other. High negative selected and implemented. In CAM, data pairs will be used
or positive values of the correlation coefficient between the to construct a data array, X, as follows:

Table 2 The input and output parameters used in the development of AI models
Variable Symbol Units Category Statistics
Min Average Max STD

Age of specimen AS Days Input 1 34 90 31.03


Max diameter of aggregate MDA mm Input 0.60 2.01 4.75 1.76
Metakaolin percentage in relation to total binder (MK/B) (%w/w) Input 0.00 14.08 30.00 10.03
Water-to-binder ratio W/B w/w Input 0.36 0.46 0.60 0.06
Superplasticizer SP (%w/w) Input 0.00 0.24 2.35 0.53
Binder-to-sand ratio B/S w/w Input 0.33 0.44 0.50 0.08
Compressive strength CS MPa Output 6.06 42.99 99.17 16.47

123
4508 Neural Computing and Applications (2021) 33:4501–4532

Fig. 2 Histograms of the parameters

X ¼ fx1 ; x2 ; x3 ; . . .; xi ; . . .; xn g ð2Þ xi ¼ fxi1 ; xi2 ; xi3 ; . . .; xim g ð3Þ


Variable xi in array, X, is a length vector of m as:

123
Neural Computing and Applications (2021) 33:4501–4532 4509

Fig. 3 Histograms of the parameters

Pm
The relationship between Rij (strength of the relation) k¼1 xik xjk
Rij ¼ pP
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
m 2
Pm 2 ffi ð4Þ
and datasets of Xi and Xj is presented by the following k¼1 xik k¼1 xik
equation:

123
4510 Neural Computing and Applications (2021) 33:4501–4532

Table 3 Correlation matrix of


Variables Input Output parameter
the input variables
AS MDA MK/B W/B SP B/S CS

Input AS 1.00
MDA - 0.34 1.00
MK/B 0.07 - 0.12 1.00
W/B 0.03 - 0.37 0.08 1.00
SP 0.06 - 0.24 - 0.03 - 0.53 1.00
B/S 0.23 - 0.35 0.19 - 0.08 0.37 1.00
Output CS 0.20 0.13 - 0.02 - 0.45 0.37 - 0.11 1.00

P 2
The Rij values between the CS and input parameters are 2 ximes  xipred
i
R ¼1 P 2
ð5Þ
shown in Fig. 4. This analysis reveals that, among the
i ðximes  xÞ
parameters examined, the B/S and the W/B have the   
greatest influence on the CS values, with strength values of var ximes  xipred
VAF ¼ 1   100 ð6Þ
0.91 and 0.90, respectively, followed by the MK/B, the AS, varðximes Þ
the MDA and whereas the SP has the lowest impact on the vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u N
CS. u 1 X 2
RMSE ¼ t xipred  ximes ð7Þ
N i¼1
3.5 Performance evaluation
1X N
MAE ¼ xipred  ximes ð8Þ
The developed models should be evaluated to assess their N i¼1
capability in solving science and engineering problems.
The present study used several performance indices to where, ximes and xipred are the measured and predicted
measure the performance of the developed models. These values, respectively. x represents the average measured
include the coefficient of determination (R2), the values, N is the total No. of datasets. The values of 1,
mean absolute error (MAE), the root mean square error 100%, 0 and 0 are required for performance indices of R2 ,
(RMSE) and the variance account for (VAF). These indices VAF, RMSE and MAE,, respectively, when a perfect pre-
were widely used in the previous investigations for the dictive model is of interest.
performance assessment purpose of the AI techniques Apart from the mentioned performance indices, a new
[88–92]. Their computation formulas are presented as performance index, namely, the a10-index has been
follows: recently introduced and used to recognize the strength of

Fig. 4 Sensitivity analysis of


CS on mortars mix parameters
based on experimental database

123
Neural Computing and Applications (2021) 33:4501–4532 4511

model capability [93–105]. Its computation formula is performance index. Namely, the value of RMSE has been
presented as follows: determined using only the testing datasets which are the
m10 20% of the total datasets, while the other 80% has been
a10index ¼ ð9Þ used for the training of the developed ANN models.
M
Based on the above ranking, the optimum architecture of
where M = the No. of datasets, m10 = the No. of samples the ANN model was the 6-7-12-1 that corresponds to case
with values of rate measured/predicted value (range ANN models with two hidden layers. Furthermore, this
between 0.9 and 1.1). It is important to mention that the optimum ANN model corresponds to the case of using
a10-index with value of one should be obtained for a MinMax normalization technique in the range 0.10-0.90,
perfect predictive model. transfer functions both for first and second hidden layers
the Log-sigmoid transfer function while the linear transfer
function for the output layer as it is presented in Fig. 5.
4 Results and discussion
4.2 Development of optimum ANFIS model
4.1 Development of optimum ANN model
ANFIS is an intelligent neuro-fuzzy approach that is used
In order to determine the optimum architecture of the ANN to model and control unspecified and uncertain systems.
model, a detailed and in-depth investigation has been ANFIS takes into account the input/output data pairs of
contacted. Namely, a plethora of ANN models have been systems. The proposed ANFIS structure is utilized for
trained and developed. The development and training of modeling the predictive system. For modeling purposes,
ANN took place for a series of hidden layers ranging from ANFIS needs to be used at the following steps: determin-
1 to 2 and with a number of neurons ranging from 1 to 30. ing/preparing the input and output values, determining/in-
In addition, each ANN model was trained for various vestigating the fuzzy sets for input variables, specifying
activation functions, as well as for different training algo- fuzzy rules and creating and training the neural network.
rithms. Furthermore, each one of the developed ANN There is a need to develop instrument for the aim of exe-
models was trained though ten different random initial cuting and testing the proposed structure. It is capable of
values of weights and biases through the use of the producing and evaluating fuzzy systems with the help of
parameters which are presented in Table 4, and four dif- the graphical user interface (GUI). This tool contains a FIS
ferent cases of ANN architectures were investigated. editor, a MF editor, a rule editor, the output surface viewer
Specifically, two scenarios were investigated regarding the and the fuzzy inference viewer. The FIS editor shows a
number of hidden layers: with one and two hidden layers general information about FIS. The MF editor shows and
and each one of them with and without normalization edits the MFs that are accompanied with all input and
technique (Table 5). output variables. The rule editor has the capacity of con-
The above procedure resulted in the design and training structing the rule statements in an automatic way through
of 1,965,600 different neural networks models. These clicking upon and choosing an item within each input
developed ANN models were ranked based on RMSE variable box, an item within each output box, as well as a

Table 4 Training parameters of ANN models


Parameter Value

Training Algorithm Levenberg–Marquardt Algorithm


Normalization Minmax in the range 1.00–1.00 and 0.10–0.90
Number of Hidden Layers 1; 2
Number of Neurons per Hidden Layer 1 to 30 by step 1
Control random number generation rand(seed, generator) where generator range from 1 to 10 by step 1
Training Goal 0
Epochs 250
Cost Function MSE; SSE
Transfer Functions Tansig (T); Logsig (L); Purelin (P)
MSE Mean Square Error; SSE Sum Square Error; Tansig (T) Hyperbolic Tangent Sigmoid transfer function; Logsig (L) Log-sigmoid transfer
function; Purelin (P) Linear transfer function

123
4512 Neural Computing and Applications (2021) 33:4501–4532

Table 5 Scenarios of developed


Scenario Hidden layers MinMax Number of models
ANN models
1 2 Normalization technique

I • 10,800
II • • 10,800
III • 976,000
IV • • 976,000
Total 1,965,600

connection item. With the use of the rule viewer, the user accomplished for both the datasets used in the training of
will be capable of interpreting the completely fuzzy the two models, as well as the datasets, which were used
inference process at once. In the ANFIS editor, the GUI for the evaluation of the models’ simulation credibility.
selection panel initializes the FIS training, saves the FIS The comparison between the experimental CS data and the
object and opens up a novel Sugeno scheme for the inter- respective predicted CS data is presented in Fig. 12, for
pretation of FIS. both models. It should be noted that in Fig. 12, only the
In design of MF, although previous investigations data used for the assessment (testing phase) of the two
highlighted the successful use of Gaussian or bell-shaped models were taken into account.
MF in the developed fuzzy models as it provides both Before selecting the optimum model among the two
simplicity and flexibility [106], some other types of MFs models, better capable of simulating the problem in ques-
like triangular and trapezoidal have been applied and per- tion, it should be noted that for the total of all five indices,
formance prediction of the ANFIS model was determined. both models have succeeded in achieving the highest val-
The results demonstrated that Gaussian MF type is the best ues achieved up to now in the literature regarding the
one among all applied MF types, and hence, this MF was simulation of cement-based mortars’ CS. Indicatively, it is
selected in the ANFIS model development. Each input noticed that both prediction models present a deviation of
parameter with three MFs shows the best results for pre- less than ± 10% (a10-index) regarding predicted in rela-
diction of the CS of cement-based mortars. Therefore, a tion to experimental values, which is a highly accept-
total number of 729 fuzzy rules (3 9 3 9 3 9 3 9 3 9 3) able deviation for this type of composite materials
were designed by the system to solve the problem. The (mortars). Specimens corresponding to the same mortar
linguistic variables of low (L), medium (M) and high mix, under uniaxial compressive test in some cases, present
(H) were assigned in modeling process of MFs. Addition- standard deviations even higher than ± 10%; this obser-
ally, the type of output MF was set to be linear. In the vation affirms the good prediction accuracy of the model.
developed ANFIS model, after epoch number 5, there are Based on the above and following common practice, it is
no changes in network performance. Figures 6, 7, 8, 9, 10 derived that, according to the statistical indices, the ANFIS
and 11 show Gaussian MFs of the input parameters model is the optimum among the two, achieving better
designed by the developed ANFIS model in predicting the values for all the performance indices.
CS of cement-based mortars. In addition, Table 6 presents Although it is most common for model evaluation to be
the information concerning the ANFIS that is established in constrained to the assessment of the statistical indices, the
the programming console. It should be noted that like ANN authors feel that such an approach may lead to misleading
model development part, whole datasets were divided into conclusions. More specifically, through the statistical
2 separate sections of train (80% of total datasets) and test evaluation, one can be assured that the ANFIS model
(20% of total datasets) for model development and model predicts better values for the data of the experimental
evaluation, respectively. database, used for the development and training of the
models; however, it is possible that, for values other than
4.3 Assessment of the optimum ANN and ANFIS the exact ones included in the database, the same model
models may not be the optimum one and may in fact provide worse
prediction, even negative CS values, a result void of
It is a common practice for researchers to evaluate com- physical meaning.
putational simulations based mostly on statistic/perfor- At this point, it should be stressed that a mathematical
mance indices. In this direction, the values of the five simulation is proposed to forecast the output parameter
prevailing evaluation indices, presented in detail in the value in a reliable manner, and, for that matter, for input
previous sections, were calculated for both the ANN and parameter combinations and values others than those
the ANFIS models and are presented in Table 7. This was experimentally examined; this means that we do not know

123
Neural Computing and Applications (2021) 33:4501–4532 4513

Compressive Strength

Linear transfer function


Output Layer

Using MinMax [0.10, 0.90] Technique


DeNormalized of Output Parameter

Logt-Sigmoid Transfer Function


b3

11
2

12
3

10
4

9
Hidden Layers

b2

6
1

Logt-Sigmoid Transfer Function


b1

6
3

Using MinMax [0.10, 0.90] Technique


Normalized of Input Parameters
Input Layer

Max diameter of aggregate (MDA)

Binder to Sand ratio (B/S)


Water to Bind ratio (W/B)
Age of specimen (AS)

Superplasticizer (SP)
Metakaolin (MK)

Fig. 5 Architecture of the optimum ANN

a priori the behavior of a material in these, ‘‘new,’’ forecast laws which govern the behavior of the material under
areas. Therefore, it is desired to design and develop a investigation.
model, which, taking advantage of available knowledge, Based on the above, the authors propose that an exper-
based mainly on experimental data, can succeed in imental verification of each model’s prediction capacity
revealing other areas of interest and assist in revealing the takes place, aiming to select the optimum model and to

123
4514 Neural Computing and Applications (2021) 33:4501–4532

Fig. 6 The designed MF for AS

Fig. 7 The designed MF for


MDA

avoid the issue of overfitting the data. It is important to training, but at the same time, for values of input param-
mention that the process presented in the next two sections, eters others than the ones used for the development and
can be used by other researchers or designers as a practical training process, the model may predict extremely unusual
application of the AI techniques developed in this study. values.
With this in mind, in this section, the results of an
4.4 Experimental verification of optimum ANN additional, thorough experimental verification of both
and ANFIS model models are presented. Specifically, both models were
implemented to examine five cases taken from the litera-
As aforementioned in the previous section, among the ture, covering a wide range and combination of input
problems that one may encounter during the mathematical parameter values which affect the CS of cement-based
simulation of datasets, overfitting the problem is extremely mortars.
common. This means that a model may succeed in an In the first to experimental verification cases, the CS in
excellent simulation of data used for its development and relation to metakolin to total binder percentage was

123
Neural Computing and Applications (2021) 33:4501–4532 4515

Fig. 8 The designed MF for


MK/B

Fig. 9 The designed MF for


W/B

examined for the experimental results of Vu et al. [74] overfitted the data. In fact, the overfitting issue is more
(Fig. 13), as well as for the experimental results of Parande intense in the case of simulating the Sumasree and Sajja
et al. [76] (Fig. 14). From Figs. 13 and 14, one can clearly [77] (Fig. 16) experimental data.
see that, regarding the ANFIS model, overfitting has indeed The fifth case corresponds to the experimental results
occurred. In fact, in the case of simulating the experimental presented by Courard et al. [75] (Fig. 17), where, in
datasets of Parande et al. [76] (Fig. 14), the overfitting accordance with the first two cases, CS is examined in
issue is more intense. relation to the metakaolin to total binder percentage, for
In the third and fourth case, the value of CS in relation three different ages of 3, 7 and 28 days. From comparing
to the age of the mortar specimen was examined, taking the simulation curves to the experimental data, in this
into account the experimental data presented by Kadri et al. figure as well, the smoothness and normality of the curves
[79] (Fig. 15), as well as the experimental data of deriving from the ANN model is in direct contrast with the
Sumasree and Sajja [77] (Fig. 16). From these figures as respective, highly irregular curves deriving from the
well, it is obvious that, again, the ANFIS model has ANFIS model.

123
4516 Neural Computing and Applications (2021) 33:4501–4532

Fig. 10 The designed MF for


SP

Fig. 11 The designed MF for


B/S

Table 6 Parameters setting for the developed ANFIS structure Table 7 Summary of prediction capability
Parameter Values/types Datasets Model Performance Indices

Input layer 6 a10-index R2 RMSE MAPE VAF


Output layer 1 Training ANN 1.0000 0.9985 0.6319 0.0123 99.8514
Input membership number 3 ANFIS 1.0000 0.9995 0.3261 0.0040 99.9531
Input membership shape Gaussian Test ANN 0.8913 0.9743 2.7400 0.0563 97.4015
Output membership shape Linear ANFIS 0.9091 0.9909 2.0975 0.0516 99.0893
Optimization method Hybrid algorithm
Stopping epoch number 5
Number of fuzzy rules 729

Based on the above, the conclusion is drawn that


although the ANFIS model achieves higher statistical

123
Neural Computing and Applications (2021) 33:4501–4532 4517

allow the ANN to succeed in an ambitious goal, which is


no other than to be used as a tool to reveal the nature of the
material.
The first case corresponds to a continuation of the
experimental investigation of Courard et al. [75], presented
in the previous section (Fig. 17). In the case under exam-
ination, CS of different mortars designed with different
metakaolin to binder percentages was evaluated experi-
mentally for three different ages: 3, 7 and 28 days. In the
current investigation, and taking into account the excellent
fitting of the predicted values with the experimental values,
12 new curves were added for the case of specimen ages
from 1 day up to 91 days. Based on the derived Fig. 18, the
following conclusions can be drawn:
• Regarding early mortar ages, corresponding up to
7 days age of specimen, increase in the metakaolin to
total binder (MK/B) percentage leads to decrease in CS
• Regarding higher ages (from 7 days forward), the
opposite trend is noticed; increase in the metakaolin
to total binder (MK/B) percentage leads to increase in
CS values
• It is also concluded that, beyond a certain metakaolin to
total binder (MK/B) percentage, (threshold level at
about 15%), the value of CS is either constant,
regardless increase beyond that percentage, or, in some
cases, even decreases, with higher percentages of
metakaolin. However, at the highest age (91 days) an
upward trend is noticed, which means that, ultimately,
at three months, the mortar with the highest percentage
of metakaolin to total binder also presents the highest
value of CS
• This investigation indicates the importance of studying
cement-based mortars beyond the usual 28-days testing,
as CS increases substantially, while trends regarding
materials and their comparison may be reversed
Finally, through implementation of the ANN model, the
Fig. 12 Comparison experimental with predicted values of the CS of mapping of the material through a series of maps was
mortars for the ANN and ANFIS models (case of test datasets) undertaken (Fig. 19) corresponding to different mortar
ages (3, 7, 14, 28, 56 and 90 days) and aiming to reveal the
indices, the best model, more capable of describing the nature of the material. Thus, CS maps may assist in
totality of the multi-dimensional (seven dimensions) space, selecting areas of design, in optimizing mixes or in
which defines the CS of cement-based mortars, is in fact selecting areas of research interest. Despite the excellent
the ANN model. This highlights the demand to evaluate results, the figure in question is presented only indicatively,
computational models, not only based on statistical in order to highlight the capability of computational models
indexes, but on a level of physical meaning as well. to reveal the nature of mortar materials aiming to a holistic
approach during their design. This can be accomplished
4.5 Revealing the nature of mortars materials when a reliable database is used, which covers the full
range of parameter values affecting CS.
In the present section, two areas of the multi-dimensional
space of CS, deriving from the use of the proposed ANN,
are presented, in detail. These areas were selected in
accordance with the presence of adequate data, which can

123
4518 Neural Computing and Applications (2021) 33:4501–4532

Fig. 13 Comparison
experimental data by Vu et al.
[74] with predicted values of CS
of mortars for both ANN and
ANFIS models

Fig. 14 Comparison
experimental data by Parande
et al. [76] with predicted values
of CS of mortars for both ANN
and ANFIS models

5 Limitations The reliability of the AI system for parameters beyond


these limit values is quite restricted. It is also worth noting
It is worth noting that the application field of the proposed that, despite the satisfactory derived results, the proposed
optimum AI system is defined within the range of param- ANN model should be applied with caution. Despite the
eter values used as inputs for its design and training. fact that the database used is the largest used in the relevant
Specifically, the proposed AI system provides reliable literature up to day, the authors consider that this database
forecasts for parameter values ranging within the minimum needs to be embellished with further experimental data. To
and maximum value of the input parameters, as stated in this end, it is within the authors’ future plans to conduct
Table 2.

123
Neural Computing and Applications (2021) 33:4501–4532 4519

Fig. 15 Comparison
experimental data by Kadri
et al. [79] with predicted values
of CS of mortars for both ANN
and ANFIS models

Fig. 16 Comparison
experimental data by Sumasree
and Sajja [77] with predicted
values of CS of mortars for both
ANN and ANFIS models

further experiments related to mortars. In particular, the between 2 and 5 mm are also missing from the
experiments lacking are related to the cases of: available literature,
3. Regarding the addition of metakaolin, more experi-
1. CS at 14 days, a mortar age which, as can be clearly
ments should be conducted for metakaolin 5% mixes,
seen in Fig. 2, is lacking in experiments, and available
4. Regarding W/B experiments should be conducted for
data do not suffice to support the reliability of the
ratio values between 0.41 and 0.45, as well as
model in this area,
0.51–0.60,
2. In relation to the maximum aggregate size parameter
and its influence on CS, experiments for values

123
4520 Neural Computing and Applications (2021) 33:4501–4532

Fig. 17 Comparison
experimental data by Courard
et al. [75] with predicted values
of CS of mortars for both ANN
and ANFIS models

5. Regarding the use of superplasticizer, experiments 6 Conclusions


should be conducted for superplasticizer additions
between 0.50 and 2.00%, and An extensive database with 276 datasets comprising of six
6. Regarding the B/S ratio, experimental data should be model inputs, i.e., AS, MDA, MK/B, W/B, SP and B/S and
obtained for mortars with B/S ratio between 0.40 and one model output, i.e., the CS of cement-based mortars,
0.45. was considered and used to propose AI techniques namely
Finally, it is worth noting that the cement grade used for ANN and ANFIS. Therefore, several strategies have been
each dataset has not be taken into account in the proposed planned and conducted for modeling of ANN and ANFIS
model, which, however, is anticipated to play an important models and these models were evaluated in their categories
role—and thus prove to be a crucial input parameter—in in order to select the best one in each modeling technique.
the development of the CS of cement mortars, especially at Then, the best ANN model and the best ANFIS model were
early ages. compared using R2 , VAF, RMSE, MAE and a10-index, as
This embellishment will allow the ANN to reveal the performance indices. In terms of all performance indices,
combined influence of the different parameters on CS to its the developed ANFIS model showed a better result com-
full extent, thus revealing laws which govern the devel- pared to the developed ANN model in approximating the
opment of CS of mortars. CS of mortars. For instance, RMSE and R2 results of
(0.3261, 0.9995) and (2.0975, 0.9909) were obtained for
training and testing datasets of ANFIS model, respectively,

123
Neural Computing and Applications (2021) 33:4501–4532 4521

Fig. 18 Mortar CS vs
Metakaolin to total Binder (MK/
B) ratio at various ages of the
mortar specimen
(MDA = 2.00 mm,
W/B = 0.50, SP = 0.00 and
B/S = 0.33) based on the
proposed ANN model

whereas these values were as (0.6319, 0.9985), (2.74, ing. In order to figure out this problem, a series of addi-
0.9743) for ANN model which indicate slightly better tional experimental data (collected from the literature)
performance prediction of ANFIS model. However, the were examined to verify the developed ANN and ANFIS
ANFIS model was not selected as the best predictive model models. According to the obtained results of ANN and
in estimating the CS of mortars. This is due to the problem ANFIS models, it was obvious that, the ANFIS model has
of overfitting, which has been occurred in ANFIS model- overfitted the data. Therefore, the conclusion was drawn

123
4522 Neural Computing and Applications (2021) 33:4501–4532

Fig. 19 Maps of Mortar CS vs Metakaolin to total Binder (MK/B) ratio at various ages of the mortar specimen (MDA = 2.00 mm, SP = 0.00 and
B/S = 0.50) based on ANN model

123
Neural Computing and Applications (2021) 33:4501–4532 4523

that although the ANFIS model achieves higher perfor- Chrysoula Karamani, Athanasia Skentou and Ioanna Zoumpoulaki for
mance prediction, the best model, more capable to describe their assistance on the computational implementation of the ANN
models.
the behavior of the CS of cement-based mortars, is in fact
the ANN model and then, it was introduced as the best
Compliance with ethical standards
predictive model in this study. Other researchers and stu-
dents with caution can utilize the results, developed models Conflict of interest The authors confirm that this article content has
and their modeling procedure presented in this study, when no conflict of interest.
the new database with the same model inputs and the same
ranges of the inputs is available.
Appendix
Acknowledgements The authors would like to thank Dr. Liborio
Cavaleri, Prof. of Structural Engineering and Seismic Design at
Dipartimento di Ingegneria Civile, Ambientale, Aerospaziale, dei
See Table 8.
Materiali, University of Palermo, Italy and Dr. Binh Thai Pham, Prof.
at University of Transport Technology, Hanoi, Vietnam, for their
valuable comments and discussions. The authors would also like to
express his acknowledgement to graduate students Maria Douvika,

Table 8 Database
No Sample Input parameters Compressive strength Reference
code
AS MDA MK/B W/B SP B/S Exp Predicted Exp/Pred
ANN ANFIS ANN ANFIS
– – (days) mm (%w/w) (w/w) (%w/w) (w/w) Mpa Mpa Mpa – – –
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14)

1 M1 7 0.60 0.00 0.48 0.00 0.36 16.56 16.96 16.66 0.98 0.99 Vu et al. [74]
2 M2 7 0.60 10.00 0.48 0.00 0.36 18.99 18.47 17.77 1.03 1.07
3 M3 7 0.60 15.00 0.50 0.00 0.36 18.17 17.90 16.63 1.02 1.09
4 M4 7 0.60 20.00 0.51 0.00 0.36 18.13 17.95 17.97 1.01 1.01
5 M5 7 0.60 25.00 0.52 0.00 0.36 16.30 16.56 15.81 0.98 1.03
6 M6 7 0.60 30.00 0.53 0.00 0.36 15.45 14.50 15.43 1.07 1.00
7 M7 7 0.60 0.00 0.53 0.00 0.50 25.50 26.02 23.00 0.98 1.11
8 M8 7 0.60 10.00 0.53 0.00 0.50 25.90 27.36 23.83 0.95 1.09
9 M9 7 0.60 15.00 0.53 0.00 0.50 28.40 27.92 26.15 1.02 1.09
10 M10 7 0.60 20.00 0.53 0.00 0.50 28.20 27.98 29.38 1.01 0.96
11 M11 7 0.60 25.00 0.53 0.00 0.50 27.50 26.99 23.06 1.02 1.19
12 M12 7 0.60 30.00 0.53 0.00 0.50 26.80 24.47 26.80 1.10 1.00
13 M13 7 0.60 0.00 0.50 0.00 0.50 29.30 28.39 29.34 1.03 1.00
14 M14 7 0.60 10.00 0.50 0.00 0.50 30.40 29.64 30.24 1.03 1.01
15 M15 7 0.60 15.00 0.50 0.00 0.50 31.30 30.08 31.32 1.04 1.00
16 M16 7 0.60 20.00 0.50 0.00 0.50 29.40 29.93 29.56 0.98 0.99
17 M17 7 0.60 25.00 0.50 0.00 0.50 29.10 28.59 29.01 1.02 1.00
18 M18 7 0.60 30.00 0.50 0.00 0.50 26.80 25.49 26.82 1.05 1.00
19 M19 7 0.60 0.00 0.47 0.00 0.50 31.50 31.19 31.52 1.01 1.00
20 M20 7 0.60 10.00 0.47 0.00 0.50 31.90 32.44 31.83 0.98 1.00
21 M21 7 0.60 15.00 0.47 0.00 0.50 32.60 32.51 32.63 1.00 1.00
22 M22 7 0.60 20.00 0.47 0.00 0.50 30.40 31.62 30.30 0.96 1.00
23 M23 7 0.60 25.00 0.47 0.00 0.50 30.20 29.09 30.56 1.04 0.99
24 M24 7 0.60 30.00 0.47 0.00 0.50 29.00 24.12 28.75 1.20 1.01
25 M25 7 0.60 0.00 0.44 0.50 0.50 35.40 35.68 35.41 0.99 1.00
26 M26 7 0.60 10.00 0.44 0.50 0.50 38.00 37.07 37.94 1.02 1.00
27 M27 7 0.60 15.00 0.44 0.50 0.50 36.60 37.40 36.74 0.98 1.00

123
4524 Neural Computing and Applications (2021) 33:4501–4532

Table 8 (continued)
No Sample Input parameters Compressive strength Reference
code
AS MDA MK/B W/B SP B/S Exp Predicted Exp/Pred
ANN ANFIS ANN ANFIS
– – (days) mm (%w/w) (w/w) (%w/w) (w/w) Mpa Mpa Mpa – – –
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14)

28 M28 7 0.60 20.00 0.44 0.50 0.50 35.70 36.64 35.53 0.97 1.00
29 M29 7 0.60 25.00 0.44 0.50 0.50 33.80 33.62 33.95 1.01 1.00
30 M30 7 0.60 30.00 0.44 0.50 0.50 31.70 26.79 31.63 1.18 1.00
31 M31 7 0.60 0.00 0.40 1.30 0.50 40.00 39.98 40.03 1.00 1.00
32 M32 7 0.60 10.00 0.40 1.30 0.50 42.40 42.21 42.30 1.00 1.00
33 M33 7 0.60 15.00 0.40 1.30 0.50 41.90 43.07 41.91 0.97 1.00
34 M34 7 0.60 20.00 0.40 1.30 0.50 41.40 41.96 41.45 0.99 1.00
35 M35 7 0.60 25.00 0.40 1.30 0.50 39.60 39.20 39.73 1.01 1.00
36 M36 7 0.60 30.00 0.40 1.30 0.50 35.50 36.61 35.39 0.97 1.00
37 M1 28 0.60 0.00 0.48 0.00 0.36 23.68 23.37 23.68 1.01 1.00
38 M2 28 0.60 10.00 0.48 0.00 0.36 25.19 26.40 25.19 0.95 1.00
39 M3 28 0.60 15.00 0.50 0.00 0.36 26.24 25.05 26.24 1.05 1.00
40 M4 28 0.60 20.00 0.51 0.00 0.36 25.42 25.08 25.42 1.01 1.00
41 M5 28 0.60 25.00 0.52 0.00 0.36 23.25 23.73 23.25 0.98 1.00
42 M6 28 0.60 30.00 0.53 0.00 0.36 22.48 21.93 22.48 1.02 1.00
43 M7 28 0.60 0.00 0.53 0.00 0.50 36.20 36.14 36.20 1.00 1.00
44 M8 28 0.60 10.00 0.53 0.00 0.50 38.50 38.62 38.50 1.00 1.00
45 M9 28 0.60 15.00 0.53 0.00 0.50 40.00 39.62 40.00 1.01 1.00
46 M10 28 0.60 20.00 0.53 0.00 0.50 40.20 39.95 40.19 1.01 1.00
47 M11 28 0.60 25.00 0.53 0.00 0.50 38.10 39.03 38.11 0.98 1.00
48 M12 28 0.60 30.00 0.53 0.00 0.50 35.70 36.26 35.69 0.98 1.00
49 M13 28 0.60 0.00 0.50 0.00 0.50 38.30 38.19 38.19 1.00 1.00
50 M14 28 0.60 10.00 0.50 0.00 0.50 41.30 40.89 41.69 1.01 0.99
51 M15 28 0.60 15.00 0.50 0.00 0.50 42.10 41.82 42.04 1.01 1.00
52 M16 28 0.60 20.00 0.50 0.00 0.50 42.70 41.98 42.32 1.02 1.01
53 M17 28 0.60 25.00 0.50 0.00 0.50 39.90 40.85 40.10 0.98 1.00
54 M18 28 0.60 30.00 0.50 0.00 0.50 38.10 37.79 38.07 1.01 1.00
55 M19 28 0.60 0.00 0.47 0.00 0.50 39.50 39.64 39.45 1.00 1.00
56 M20 28 0.60 10.00 0.47 0.00 0.50 43.60 43.50 43.77 1.00 1.00
57 M21 28 0.60 15.00 0.47 0.00 0.50 44.10 44.55 44.02 0.99 1.00
58 M22 28 0.60 20.00 0.47 0.00 0.50 44.00 44.53 44.24 0.99 0.99
59 M23 28 0.60 25.00 0.47 0.00 0.50 42.80 42.95 41.89 1.00 1.02
60 M24 28 0.60 30.00 0.47 0.00 0.50 38.70 39.00 39.32 0.99 0.98
61 M25 28 0.60 0.00 0.44 0.50 0.50 41.80 41.95 41.77 1.00 1.00
62 M26 28 0.60 10.00 0.44 0.50 0.50 45.90 44.82 46.05 1.02 1.00
63 M27 28 0.60 15.00 0.44 0.50 0.50 46.30 45.93 45.95 1.01 1.01
64 M28 28 0.60 20.00 0.44 0.50 0.50 45.40 46.13 45.82 0.98 0.99
65 M29 28 0.60 25.00 0.44 0.50 0.50 44.70 44.36 44.33 1.01 1.01
66 M30 28 0.60 30.00 0.44 0.50 0.50 40.60 39.14 40.77 1.04 1.00
67 M31 28 0.60 0.00 0.40 1.30 0.50 44.00 44.18 43.92 1.00 1.00
68 M32 28 0.60 10.00 0.40 1.30 0.50 47.20 47.48 47.46 0.99 0.99
69 M33 28 0.60 15.00 0.40 1.30 0.50 48.10 48.45 48.08 0.99 1.00
70 M34 28 0.60 20.00 0.40 1.30 0.50 49.00 48.35 48.88 1.01 1.00
71 M35 28 0.60 25.00 0.40 1.30 0.50 47.70 47.97 47.38 0.99 1.01

123
Neural Computing and Applications (2021) 33:4501–4532 4525

Table 8 (continued)
No Sample Input parameters Compressive strength Reference
code
AS MDA MK/B W/B SP B/S Exp Predicted Exp/Pred
ANN ANFIS ANN ANFIS
– – (days) mm (%w/w) (w/w) (%w/w) (w/w) Mpa Mpa Mpa – – –
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14)

72 M36 28 0.60 30.00 0.40 1.30 0.50 42.10 48.64 42.38 0.87 0.99
73 M1 60 0.60 0.00 0.48 0.00 0.36 28.16 26.92 28.16 1.05 1.00
74 M2 60 0.60 10.00 0.48 0.00 0.36 29.20 29.15 29.20 1.00 1.00
75 M3 60 0.60 15.00 0.50 0.00 0.36 30.94 29.54 30.94 1.05 1.00
76 M4 60 0.60 20.00 0.51 0.00 0.36 29.98 30.07 29.98 1.00 1.00
77 M5 60 0.60 25.00 0.52 0.00 0.36 28.45 28.73 28.45 0.99 1.00
78 M6 60 0.60 30.00 0.53 0.00 0.36 27.97 26.61 27.97 1.05 1.00
79 M7 60 0.60 0.00 0.53 0.00 0.50 41.30 41.37 41.30 1.00 1.00
80 M8 60 0.60 10.00 0.53 0.00 0.50 43.80 43.44 43.80 1.01 1.00
81 M9 60 0.60 15.00 0.53 0.00 0.50 43.40 43.78 43.40 0.99 1.00
82 M10 60 0.60 20.00 0.53 0.00 0.50 44.30 43.46 44.31 1.02 1.00
83 M11 60 0.60 25.00 0.53 0.00 0.50 42.50 42.19 42.49 1.01 1.00
84 M12 60 0.60 30.00 0.53 0.00 0.50 39.60 39.48 39.61 1.00 1.00
85 M13 60 0.60 0.00 0.50 0.00 0.50 44.50 44.13 44.59 1.01 1.00
86 M14 60 0.60 10.00 0.50 0.00 0.50 46.00 45.45 45.67 1.01 1.01
87 M15 60 0.60 15.00 0.50 0.00 0.50 44.70 45.46 44.75 0.98 1.00
88 M16 60 0.60 20.00 0.50 0.00 0.50 45.10 44.94 45.42 1.00 0.99
89 M17 60 0.60 25.00 0.50 0.00 0.50 43.60 43.60 43.44 1.00 1.00
90 M18 60 0.60 30.00 0.50 0.00 0.50 42.10 40.85 42.13 1.03 1.00
91 M19 60 0.60 0.00 0.47 0.00 0.50 46.30 46.11 46.34 1.00 1.00
92 M20 60 0.60 10.00 0.47 0.00 0.50 46.90 47.57 46.75 0.99 1.00
93 M21 60 0.60 15.00 0.47 0.00 0.50 45.20 47.43 45.27 0.95 1.00
94 M22 60 0.60 20.00 0.47 0.00 0.50 47.00 46.64 46.79 1.01 1.00
95 M23 60 0.60 25.00 0.47 0.00 0.50 43.90 44.90 44.67 0.98 0.98
96 M24 60 0.60 30.00 0.47 0.00 0.50 43.60 41.51 43.08 1.05 1.01
97 M25 60 0.60 0.00 0.44 0.50 0.50 46.60 47.05 46.62 0.99 1.00
98 M26 60 0.60 10.00 0.44 0.50 0.50 48.50 48.78 48.38 0.99 1.00
99 M27 60 0.60 15.00 0.44 0.50 0.50 47.80 49.29 48.09 0.97 0.99
100 M28 60 0.60 20.00 0.44 0.50 0.50 48.70 49.13 48.35 0.99 1.01
101 M29 60 0.60 25.00 0.44 0.50 0.50 47.30 47.45 47.61 1.00 0.99
102 M30 60 0.60 30.00 0.44 0.50 0.50 43.90 42.81 43.75 1.03 1.00
103 M31 60 0.60 0.00 0.40 1.30 0.50 48.70 48.02 48.76 1.01 1.00
104 M32 60 0.60 10.00 0.40 1.30 0.50 50.90 51.31 50.68 0.99 1.00
105 M33 60 0.60 15.00 0.40 1.30 0.50 51.10 52.03 51.12 0.98 1.00
106 M34 60 0.60 20.00 0.40 1.30 0.50 52.00 51.86 52.10 1.00 1.00
107 M35 60 0.60 25.00 0.40 1.30 0.50 51.30 51.47 51.57 1.00 0.99
108 M36 60 0.60 30.00 0.40 1.30 0.50 47.50 52.14 47.26 0.91 1.01
109 M1 90 0.60 0.00 0.48 0.00 0.36 29.74 30.43 29.74 0.98 1.00
110 M2 90 0.60 10.00 0.48 0.00 0.36 31.23 31.29 31.23 1.00 1.00
111 M3 90 0.60 15.00 0.50 0.00 0.36 32.10 30.57 32.10 1.05 1.00
112 M4 90 0.60 20.00 0.51 0.00 0.36 32.16 32.11 32.16 1.00 1.00
113 M5 90 0.60 25.00 0.52 0.00 0.36 32.28 31.54 32.28 1.02 1.00
114 M6 90 0.60 30.00 0.53 0.00 0.36 30.40 29.70 30.40 1.02 1.00
115 M7 90 0.60 0.00 0.53 0.00 0.50 43.70 44.47 43.70 0.98 1.00

123
4526 Neural Computing and Applications (2021) 33:4501–4532

Table 8 (continued)
No Sample Input parameters Compressive strength Reference
code
AS MDA MK/B W/B SP B/S Exp Predicted Exp/Pred
ANN ANFIS ANN ANFIS
– – (days) mm (%w/w) (w/w) (%w/w) (w/w) Mpa Mpa Mpa – – –
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14)

116 M8 90 0.60 10.00 0.53 0.00 0.50 44.50 44.96 44.50 0.99 1.00
117 M9 90 0.60 15.00 0.53 0.00 0.50 45.00 44.88 45.00 1.00 1.00
118 M10 90 0.60 20.00 0.53 0.00 0.50 44.20 44.46 44.20 0.99 1.00
119 M11 90 0.60 25.00 0.53 0.00 0.50 43.40 43.42 43.40 1.00 1.00
120 M12 90 0.60 30.00 0.53 0.00 0.50 43.00 41.30 43.00 1.04 1.00
121 M13 90 0.60 0.00 0.50 0.00 0.50 47.00 46.82 46.97 1.00 1.00
122 M14 90 0.60 10.00 0.50 0.00 0.50 46.70 46.68 46.80 1.00 1.00
123 M15 90 0.60 15.00 0.50 0.00 0.50 46.90 46.41 46.88 1.01 1.00
124 M16 90 0.60 20.00 0.50 0.00 0.50 45.30 45.90 45.21 0.99 1.00
125 M17 90 0.60 25.00 0.50 0.00 0.50 45.30 44.84 45.34 1.01 1.00
126 M18 90 0.60 30.00 0.50 0.00 0.50 44.90 42.74 44.89 1.05 1.00
127 M19 90 0.60 0.00 0.47 0.00 0.50 48.00 47.61 47.99 1.01 1.00
128 M20 90 0.60 10.00 0.47 0.00 0.50 47.40 47.58 47.44 1.00 1.00
129 M21 90 0.60 15.00 0.47 0.00 0.50 47.40 47.17 47.38 1.00 1.00
130 M22 90 0.60 20.00 0.47 0.00 0.50 46.80 46.39 46.86 1.01 1.00
131 M23 90 0.60 25.00 0.47 0.00 0.50 45.60 44.94 45.38 1.01 1.00
132 M24 90 0.60 30.00 0.47 0.00 0.50 45.10 42.22 45.25 1.07 1.00
133 M25 90 0.60 0.00 0.44 0.50 0.50 50.50 50.12 50.49 1.01 1.00
134 M26 90 0.60 10.00 0.44 0.50 0.50 51.20 51.13 51.24 1.00 1.00
135 M27 90 0.60 15.00 0.44 0.50 0.50 51.40 51.38 51.32 1.00 1.00
136 M28 90 0.60 20.00 0.44 0.50 0.50 50.60 51.11 50.70 0.99 1.00
137 M29 90 0.60 25.00 0.44 0.50 0.50 50.50 49.63 50.41 1.02 1.00
138 M30 90 0.60 30.00 0.44 0.50 0.50 47.10 45.63 47.14 1.03 1.00
139 M31 90 0.60 0.00 0.40 1.30 0.50 51.40 51.90 51.38 0.99 1.00
140 M32 90 0.60 10.00 0.40 1.30 0.50 54.20 53.77 54.26 1.01 1.00
141 M33 90 0.60 15.00 0.40 1.30 0.50 54.90 54.14 54.89 1.01 1.00
142 M34 90 0.60 20.00 0.40 1.30 0.50 53.80 53.89 53.77 1.00 1.00
143 M35 90 0.60 25.00 0.40 1.30 0.50 53.50 53.53 53.42 1.00 1.00
144 M36 90 0.60 30.00 0.40 1.30 0.50 50.80 54.29 50.87 0.94 1.00
145 N 3 2.00 0.00 0.50 0.00 0.33 34.50 34.13 34.81 1.01 0.99 Courard et al. [75]
146 5MK 3 2.00 5.00 0.50 0.00 0.33 34.40 34.20 33.79 1.01 1.02
147 10MK 3 2.00 10.00 0.50 0.00 0.33 32.40 33.73 32.97 0.96 0.98
148 15MK 3 2.00 15.00 0.50 0.00 0.33 31.70 31.89 31.38 0.99 1.01
149 20MK 3 2.00 20.00 0.50 0.00 0.33 27.40 27.65 27.45 0.99 1.00
150 N 7 2.00 0.00 0.50 0.00 0.33 47.10 44.34 46.76 1.06 1.01
151 5MK 7 2.00 5.00 0.50 0.00 0.33 46.30 46.66 46.96 0.99 0.99
152 10MK 7 2.00 10.00 0.50 0.00 0.33 48.60 48.75 47.98 1.00 1.01
153 15MK 7 2.00 15.00 0.50 0.00 0.33 47.90 49.77 48.25 0.96 0.99
154 20MK 7 2.00 20.00 0.50 0.00 0.33 49.40 48.86 49.35 1.01 1.00
155 N 28 2.00 0.00 0.50 0.00 0.33 57.00 57.63 57.03 0.99 1.00
156 5MK 28 2.00 5.00 0.50 0.00 0.33 65.10 63.53 65.05 1.02 1.00
157 10MK 28 2.00 10.00 0.50 0.00 0.33 70.20 68.81 70.25 1.02 1.00
158 15MK 28 2.00 15.00 0.50 0.00 0.33 71.20 72.12 71.17 0.99 1.00
159 20MK 28 2.00 20.00 0.50 0.00 0.33 68.40 72.61 68.40 0.94 1.00

123
Neural Computing and Applications (2021) 33:4501–4532 4527

Table 8 (continued)
No Sample Input parameters Compressive strength Reference
code
AS MDA MK/B W/B SP B/S Exp Predicted Exp/Pred
ANN ANFIS ANN ANFIS
– – (days) mm (%w/w) (w/w) (%w/w) (w/w) Mpa Mpa Mpa – – –
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14)

160 N 3 4.75 0.00 0.40 0.00 0.33 22.00 24.02 22.51 0.92 0.98 Parande et al. [76]
161 C ? 5MK 3 4.75 5.00 0.40 0.00 0.33 31.50 29.05 30.59 1.08 1.03
162 C ? 10MK 3 4.75 10.00 0.40 0.00 0.33 30.00 30.37 30.69 0.99 0.98
163 C ? 15MK 3 4.75 15.00 0.40 0.00 0.33 29.00 29.17 28.66 0.99 1.01
164 C ? 20MK 3 4.75 20.00 0.40 0.00 0.33 27.00 27.36 27.05 0.99 1.00
165 N 7 4.75 0.00 0.40 0.00 0.33 32.50 32.93 31.95 0.99 1.02
166 C ? 5MK 7 4.75 5.00 0.40 0.00 0.33 40.00 40.51 40.99 0.99 0.98
167 C ? 10MK 7 4.75 10.00 0.40 0.00 0.33 43.50 44.08 42.75 0.99 1.02
168 C ? 15MK 7 4.75 15.00 0.40 0.00 0.33 42.10 44.28 42.46 0.95 0.99
169 C ? 20MK 7 4.75 20.00 0.40 0.00 0.33 44.00 43.32 43.94 1.02 1.00
170 N 28 4.75 0.00 0.40 0.00 0.33 40.00 38.63 40.04 1.04 1.00
171 C ? 5MK 28 4.75 5.00 0.40 0.00 0.33 52.00 49.80 51.93 1.04 1.00
172 C ? 10MK 28 4.75 10.00 0.40 0.00 0.33 56.00 57.49 56.06 0.97 1.00
173 C ? 15MK 28 4.75 15.00 0.40 0.00 0.33 60.00 59.38 59.97 1.01 1.00
174 C ? 20MK 28 4.75 20.00 0.40 0.00 0.33 58.00 58.88 58.00 0.99 1.00
175 N 90 4.75 0.00 0.40 0.00 0.33 48.00 48.11 48.00 1.00 1.00
176 C ? 5MK 90 4.75 5.00 0.40 0.00 0.33 52.00 51.95 52.00 1.00 1.00
177 C ? 10MK 90 4.75 10.00 0.40 0.00 0.33 64.00 63.58 64.00 1.01 1.00
178 C ? 15MK 90 4.75 15.00 0.40 0.00 0.33 67.00 66.94 67.00 1.00 1.00
179 C ? 20MK 90 4.75 20.00 0.40 0.00 0.33 65.00 65.06 65.00 1.00 1.00
180 Control 3 4.75 0.00 0.46 0.00 0.50 24.93 25.80 24.93 0.97 1.00 Sumasree and Sajja [77]
181 Control 7 4.75 0.00 0.46 0.00 0.50 25.21 26.35 25.21 0.96 1.00
182 Control 14 4.75 0.00 0.46 0.00 0.50 25.98 26.85 25.98 0.97 1.00
183 Control 28 4.75 0.00 0.46 0.00 0.50 26.39 26.97 26.39 0.98 1.00
184 Control 56 4.75 0.00 0.46 0.00 0.50 26.97 27.46 23.40 0.98 1.15
185 5MK 3 4.75 5.00 0.46 0.00 0.50 27.21 26.99 29.54 1.01 0.92
186 5MK 7 4.75 5.00 0.46 0.00 0.50 28.01 27.46 29.81 1.02 0.94
187 5MK 14 4.75 5.00 0.46 0.00 0.50 28.87 27.92 27.48 1.03 1.05
188 5MK 28 4.75 5.00 0.46 0.00 0.50 29.36 28.08 28.85 1.05 1.02
189 5MK 56 4.75 5.00 0.46 0.00 0.50 30.11 28.96 30.11 1.04 1.00
190 10MK 3 4.75 10.00 0.46 0.00 0.50 28.67 28.45 29.51 1.01 0.97
191 10MK 7 4.75 10.00 0.46 0.00 0.50 29.12 28.89 29.75 1.01 0.98
192 10MK 14 4.75 10.00 0.46 0.00 0.50 29.97 29.33 30.20 1.02 0.99
193 10MK 28 4.75 10.00 0.46 0.00 0.50 30.26 29.51 30.41 1.03 1.00
194 10MK 56 4.75 10.00 0.46 0.00 0.50 30.99 30.84 27.70 1.00 1.12
195 15MK 3 4.75 15.00 0.46 0.00 0.50 29.93 30.09 25.19 0.99 1.19
196 15MK 7 4.75 15.00 0.46 0.00 0.50 30.36 30.62 29.47 0.99 1.03
197 15MK 14 4.75 15.00 0.46 0.00 0.50 31.01 31.09 31.88 1.00 0.97
198 15MK 28 4.75 15.00 0.46 0.00 0.50 31.90 31.27 32.08 1.02 0.99
199 15MK 56 4.75 15.00 0.46 0.00 0.50 32.37 32.97 33.81 0.98 0.96
200 20MK 3 4.75 20.00 0.46 0.00 0.50 31.19 31.62 28.88 0.99 1.08
201 20MK 7 4.75 20.00 0.46 0.00 0.50 31.93 32.40 29.44 0.99 1.08
202 20MK 14 4.75 20.00 0.46 0.00 0.50 32.42 32.98 30.20 0.98 1.07
203 20MK 28 4.75 20.00 0.46 0.00 0.50 33.11 33.11 30.74 1.00 1.08

123
4528 Neural Computing and Applications (2021) 33:4501–4532

Table 8 (continued)
No Sample Input parameters Compressive strength Reference
code
AS MDA MK/B W/B SP B/S Exp Predicted Exp/Pred
ANN ANFIS ANN ANFIS
– – (days) mm (%w/w) (w/w) (%w/w) (w/w) Mpa Mpa Mpa – – –
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14)

204 20MK 56 4.75 20.00 0.46 0.00 0.50 33.99 34.84 36.55 0.98 0.93
205 25MK 3 4.75 25.00 0.46 0.00 0.50 32.76 32.18 30.48 1.02 1.07
206 25MK 7 4.75 25.00 0.46 0.00 0.50 33.51 33.38 35.42 1.00 0.95
207 25MK 14 4.75 25.00 0.46 0.00 0.50 33.97 34.21 32.86 0.99 1.03
208 25MK 28 4.75 25.00 0.46 0.00 0.50 34.24 34.34 32.58 1.00 1.05
209 25MK 56 4.75 25.00 0.46 0.00 0.50 35.71 35.52 34.71 1.01 1.03
210 30MK 3 4.75 30.00 0.46 0.00 0.50 30.41 29.89 30.39 1.02 1.00
211 30MK 7 4.75 30.00 0.46 0.00 0.50 31.36 31.68 31.62 0.99 0.99
212 30MK 14 4.75 30.00 0.46 0.00 0.50 32.99 32.96 32.99 1.00 1.00
213 30MK 28 4.75 30.00 0.46 0.00 0.50 33.01 33.28 33.01 0.99 1.00
214 30MK 56 4.75 30.00 0.46 0.00 0.50 33.76 33.68 33.76 1.00 1.00
215 PC 1 2.00 0.00 0.60 0.00 0.33 17.60 17.40 18.65 1.01 0.94 Batis et al. [78]
216 PC 2 2.00 0.00 0.60 0.00 0.33 29.40 26.71 28.12 1.10 1.05
217 PC 7 2.00 0.00 0.60 0.00 0.33 44.50 44.37 44.73 1.00 0.99
218 PC 28 2.00 0.00 0.60 0.00 0.33 57.50 58.13 57.50 0.99 1.00
219 MKC-20 1 2.00 20.00 0.60 0.00 0.33 18.80 17.95 18.37 1.05 1.02
220 MKC-20 2 2.00 20.00 0.60 0.00 0.33 29.10 28.89 28.84 1.01 1.01
221 MKC-20 7 2.00 20.00 0.60 0.00 0.33 50.40 50.34 50.19 1.00 1.00
222 MKC-20 28 2.00 20.00 0.60 0.00 0.33 69.70 65.45 67.50 1.06 1.03
223 MK-10 1 2.00 10.00 0.60 0.00 0.33 20.00 20.82 20.20 0.96 0.99
224 MK-10 2 2.00 10.00 0.60 0.00 0.33 32.00 31.41 31.75 1.02 1.01
225 MK-10 7 2.00 10.00 0.60 0.00 0.33 50.70 51.07 50.75 0.99 1.00
226 MK-10 28 2.00 10.00 0.60 0.00 0.33 68.10 67.38 68.10 1.01 1.00
227 MK-20 1 2.00 20.00 0.60 0.00 0.33 18.00 17.95 18.37 1.00 0.98
228 MK-20 2 2.00 20.00 0.60 0.00 0.33 28.50 28.89 28.84 0.99 0.99
229 MK-20 7 2.00 20.00 0.60 0.00 0.33 50.00 50.34 50.19 0.99 1.00
230 MK-20 28 2.00 20.00 0.60 0.00 0.33 65.30 65.45 67.50 1.00 0.97
231 OPC 1 2.00 0.00 0.36 1.40 0.50 40.00 40.94 40.00 0.98 1.00 Kadri et al. [79]
232 OPC 7 2.00 0.00 0.36 1.40 0.50 70.42 70.48 70.42 1.00 1.00
233 OPC 28 2.00 0.00 0.36 1.40 0.50 80.42 80.35 80.42 1.00 1.00
234 OPC 56 2.00 0.00 0.36 1.40 0.50 84.17 82.28 84.17 1.02 1.00
235 MK1 1 2.00 10.00 0.36 2.35 0.50 35.21 35.18 35.21 1.00 1.00
236 MK1 7 2.00 10.00 0.36 2.35 0.50 74.58 74.58 74.58 1.00 1.00
237 MK1 28 2.00 10.00 0.36 2.35 0.50 95.00 94.19 95.00 1.01 1.00
238 MK1 56 2.00 10.00 0.36 2.35 0.50 96.25 96.22 96.25 1.00 1.00
239 MK2 1 2.00 10.00 0.36 2.04 0.50 35.63 35.60 35.63 1.00 1.00
240 MK2 7 2.00 10.00 0.36 2.04 0.50 82.71 78.14 82.71 1.06 1.00
241 MK2 28 2.00 10.00 0.36 2.04 0.50 97.29 97.35 97.29 1.00 1.00
242 MK2 56 2.00 10.00 0.36 2.04 0.50 99.17 99.17 99.17 1.00 1.00

123
Neural Computing and Applications (2021) 33:4501–4532 4529

Table 8 (continued)
No Sample Input parameters Compressive strength Reference
code
AS MDA MK/B W/B SP B/S Exp Predicted Exp/Pred
ANN ANFIS ANN ANFIS
– – (days) mm (%w/ (w/ (%w/ (w/ Mpa Mpa Mpa – – –
w) w) w) w)
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14)

243 Control 7 2.00 0.00 0.49 0.00 0.36 37.27 42.86 37.27 0.87 1.00 Mardani-Aghabaglou et al. [80]
244 Control 28 2.00 0.00 0.49 0.00 0.36 50.05 49.99 50.05 1.00 1.00
245 Control 90 2.00 0.00 0.49 0.00 0.36 56.76 56.69 56.76 1.00 1.00
246 MK10% 7 2.00 10.00 0.49 0.00 0.36 38.22 38.70 38.22 0.99 1.00
247 MK10% 28 2.00 10.00 0.49 0.00 0.36 48.90 48.71 48.90 1.00 1.00
248 MK10% 90 2.00 10.00 0.49 0.00 0.36 54.76 54.94 54.76 1.00 1.00
249 0% 1 4.75 0.00 0.38 0.00 0.33 23.40 34.39 23.47 0.68 1.00 Potgieter-Vermaak and Potgieter
250 0% 2 4.75 0.00 0.38 0.00 0.33 39.40 37.99 39.31 1.04 1.00 [81]
251 0% 7 4.75 0.00 0.38 0.00 0.33 47.40 48.40 47.42 0.98 1.00
252 0% 14 4.75 0.00 0.38 0.00 0.33 48.80 52.14 49.98 0.94 0.98
253 0% 28 4.75 0.00 0.38 0.00 0.33 50.90 51.32 48.96 0.99 1.04
254 10% 1 4.75 10.00 0.38 0.00 0.33 51.30 52.47 49.47 0.98 1.04
255 10% 2 4.75 10.00 0.38 0.00 0.33 60.60 57.48 57.20 1.05 1.06
256 10% 7 4.75 10.00 0.38 0.00 0.33 73.10 74.28 77.64 0.98 0.94
257 10% 14 4.75 10.00 0.38 0.00 0.33 85.90 83.69 82.77 1.03 1.04
258 10% 28 4.75 10.00 0.38 0.00 0.33 92.80 89.03 94.89 1.04 0.98
259 20% 1 4.75 20.00 0.38 0.00 0.33 58.10 56.40 54.90 1.03 1.06
260 20% 2 4.75 20.00 0.38 0.00 0.33 60.60 61.69 58.69 0.98 1.03
261 20% 7 4.75 20.00 0.38 0.00 0.33 78.80 79.84 80.10 0.99 0.98
262 20% 14 4.75 20.00 0.38 0.00 0.33 89.10 89.72 85.24 0.99 1.05
263 20% 28 4.75 20.00 0.38 0.00 0.33 92.50 92.69 90.00 1.00 1.03
264 30% 1 4.75 30.00 0.38 0.00 0.33 59.10 57.65 57.31 1.03 1.03
265 30% 2 4.75 30.00 0.38 0.00 0.33 62.20 62.62 63.08 0.99 0.99
266 30% 7 4.75 30.00 0.38 0.00 0.33 80.30 79.63 80.30 1.01 1.00
267 30% 14 4.75 30.00 0.38 0.00 0.33 85.00 88.01 85.00 0.97 1.00
268 30% 28 4.75 30.00 0.38 0.00 0.33 88.40 88.51 88.40 1.00 1.00
269 Cref 3 2.00 0.00 0.45 0.00 0.33 47.94 47.96 47.94 1.00 1.00 Cizer et al. [82]
270 Cref 7 2.00 0.00 0.45 0.00 0.33 57.20 60.84 57.20 0.94 1.00
271 Cref 28 2.00 0.00 0.45 0.00 0.33 66.23 66.16 66.23 1.00 1.00
272 Cref 60 2.00 0.00 0.45 0.00 0.33 64.60 64.54 64.60 1.00 1.00
273 Cref 90 2.00 0.00 0.45 0.00 0.33 66.01 70.48 66.01 0.94 1.00
274 REF 7 1.25 0.00 0.48 0.00 0.36 6.06 5.98 6.06 1.01 1.00 Al-Chaar et al. [83]
275 REF 28 1.25 0.00 0.48 0.00 0.36 10.49 10.47 11.52 1.00 0.91
276 REF 90 1.25 0.00 0.48 0.00 0.36 10.73 19.20 10.21 0.56 1.05

123
4530 Neural Computing and Applications (2021) 33:4501–4532

References 18. Mashhadban H, Kutanaei SS, Sayarinejad MA (2016) Prediction


and modeling of mechanical properties in fiber reinforced self-
compacting concrete using particle swarm optimization algo-
1. Apostolopoulou M, Douvika MG, Kanellopoulos IN, Mor-
rithm and artificial neural network. Constr Build Mater
opoulou A, Asteris PG (2018) Prediction of compressive
119:277–287
strength of mortars using artificial neural networks. In: 1st
19. Açikgenç M, Ulaş M, Alyamaç KE (2015) Using an artificial
international conference TMM_CH, transdisciplinary multi-
neural network to predict mix compositions of steel fiber-rein-
spectral modeling and cooperation for the preservation of cul-
forced concrete. Arab J Sci Eng 40:407–419
tural heritage, Athens, Greece
20. Asteris PG, Kolovos KG, Douvika MG, Roinos K (2016) Pre-
2. Woźniak M, Połap D (2020) Soft trees with neural components
diction of self-compacting concrete strength using artificial
as image-processing technique for archeological excavations.
neural networks. Eur J Environ Civ Eng 20:102–122
Pers Ubiquit Comput. https://doi.org/10.1007/s00779-019-
21. Baykasoǧlu A, Dereli TU, Taniş S (2004) Prediction of cement
01292-3
strength using soft computing techniques. Cem Concr Res
3. Woźniak M, Połap D (2019) Intelligent home systems for
34:2083–2090
ubiquitous user support by using neural networks and rule based
22. Akkurt S, Tayfur G, Can S (2004) Fuzzy logic model for the
approach. IEEE Trans Indus Inform. https://doi.org/10.1109/TII.
prediction of cement compressive strength. Cem Concr Res
2019.2951089
34:1429–1433
4. Woźniak M, Połap D (2017) Hybrid neuro-heuristic methodol-
23. Özcan F, Atiş CD, Karahan O, Uncuoğlu E, Tanyildizi H (2009)
ogy for simulation and control of dynamic systems over time
Comparison of artificial neural network and fuzzy logic models
interval. Neural Net 93:45–56
for prediction of long-term compressive strength of silica fume
5. Zhou J, Li E, Yang S, Wang M, Shi X, Yao S, Mitri HS (2019)
concrete. Adv Eng Softw 40:856–863
Slope stability prediction for circular mode failure using gradi-
24. Saridemir M (2009) Predicting the compressive strength of
ent boosting machine approach based on an updated database of
mortars containing metakaolin by artificial neural networks and
case histories. Saf Sci 118:505–518
fuzzy logic. Adv Eng Softw 40(9):920–927
6. Zhou J, Li X, Mitri HS (2015) Comparative performance of six
25. Eskandari-Naddaf H, Kazemi R (2017) ANN prediction of
supervised learning methods for the development of models of
cement mortar compressive strength, influence of cement
hard rock pillar stability prediction. Nat Hazards 79(1):291–316
strength class. Constr Build Mater 138:1–11
7. Aghaabbasi M, Shekari ZA, Shah MZ, Olakunle O, Armaghani
26. Oh T-K, Kim J, Lee C, Park S (2017) Nondestructive concrete
DJ, Moeinaddini M (2020) Predicting the use frequency of ride-
strength estimation based on electro-mechanical impedance with
sourcing by off-campus university students through random
artificial neural network. J Adv Concr Technol 15:94–102
forest and Bayesian network techniques. Transport Res A-Pol
27. Khademi F, Akbari M, Jamal SM, Nikoo M (2017) Multiple
136:262–281
linear regression, artificial neural network, and fuzzy logic
8. Armaghani DJ, Asteris PG, Fatemi SA, Hasanipanah M,
prediction of 28 days compressive strength of concrete. Front
Tarinejad R, Rashid ASA, Huynh VV (2020) On the use of
Struct Civ Eng 11:90–99
neuro-swarm system to forecast the pile settlement. Appl Sci
28. Türkmen İ, Bingöl AF, Tortum A, Demirboğa R, Gül R (2017)
10(6):1904
Properties of pumice aggregate concretes at elevated tempera-
9. Jahed Armaghani D, Asteris PG, Askarian B, Hasanipanah M,
tures and comparison with ANN models. Fire Mater 41:142–153
Tarinejad R, Huynh VV (2020) Examining hybrid and single
29. Nikoo M, Zarfam P, Sayahpour H (2015) Determination of
SVM models with different kernels to predict rock brittleness.
compressive strength of concrete using Self Organization Fea-
Sustainability 12(6):2229
ture Map (SOFM). Eng Comput 31:113–121
10. Duan J, Asteris PG, Nguyen H, Bui XN, Moayedi H (2020) A
30. Adeli H (2001) Neural networks in civil engineering:
novel artificial intelligence technique to predict compressive
1989–2000. Comput-Aid Civ Infrastruct Eng 16:126–142
strength of recycled aggregate concrete using ICA-XGBoost
31. Asteris PG, Nikoo M (2019) Artificial Bee colony-based neural
model. Eng Comput. https://doi.org/10.1007/s00366-020-01003-
network for the prediction of the fundamental period of infilled
0
frame structures. Neural Comput Appl 31(9):4837–4847
11. Alexandridis A (2013) Evolving RBF neural networks for
32. Safiuddin M, Raman SN, Salam MA, Jumaat MZ (2016)
adaptive soft-sensor design. Int J Neural Syst 23:1350029
Modeling of compressive strength for self-consolidating high-
12. Dias WPS, Pooliyadda SP (2001) Neural networks for predicting
strength concrete incorporating palm oil fuel ash. Materials
properties of concretes with admixtures. Constr Build Mater
9:396
15:371–379
33. Mansouri I, Kisi O (2015) Prediction of debonding strength for
13. Lee SC (2003) Prediction of concrete strength using artificial
masonry elements retrofitted with FRP composites using neuro
neural networks. Eng Struct 25:849–857
fuzzy and neural network approaches. Compos Part B Eng
14. Topçu IB, Saridemir M (2008) Prediction of compressive
70:247–255
strength of concrete containing fly ash using artificial neural
34. Reddy TCS (2017) Predicting the strength properties of slurry
networks and fuzzy logic. Comput Mater Sci 41:305–311
infiltrated fibrous concrete using artificial neural network. Front
15. Trtnik G, Kavčič F, Turk G (2009) Prediction of concrete
Struct Civ Eng. https://doi.org/10.1007/s11709-017-0445-3
strength using ultrasonic pulse velocity and artificial neural
35. Salehi H, Burgueño R (2018) Emerging artificial intelligence
networks. Ultrasonics 49:53–60
methods in structural engineering. Eng Struct 171:170–189
16. Waszczyszyn Z, Ziemiański L (2001) Neural networks in
36. Zounemat-Kermani M, Beheshti A-A, Ataie-Ashtiani B, Sab-
mechanics of structures and materials—new results and pro-
bagh-Yazdi S-R (2009) Estimation of current-induced scour
spects of applications. Comput Struct 79:2261–2276
depth around pile groups using neural network and adaptive
17. Belalia Douma O, Boukhatem B, Ghrici M, Tagnit-Hamou A
neuro-fuzzy inference system. Appl Soft Comput 9:746–755
(2016) Prediction of properties of self-compacting concrete
37. Cabalar AF, Cevik A, Gokceoglu C (2012) Some applications of
containing fly ash using artificial neural network. Neural Com-
adaptive neuro-fuzzy inference system (ANFIS) in geotechnical
put Appl. https://doi.org/10.1007/s00521-016-2368-7
engineering. Comput Geotech 40:14–33

123
Neural Computing and Applications (2021) 33:4501–4532 4531

38. Mohamad ET, Jahed Armaghani D, Momeni E, Abad SVANK 57. Demir F (2008) Prediction of elastic modulus of normal and
(2014) Prediction of the unconfined compressive strength of soft high strength concrete by artificial neural networks. Constr
rocks: a PSO-based ANN approach. Bull Eng Geol Environ. Build Mater 22(7):1428–1435
https://doi.org/10.1007/s10064-014-0638-0 58. Altun F, Kişi O, Aydin K (2008) Predicting the compressive
39. Soltani F, Kerachian R, Shirangi E (2010) Developing operating strength of steel fiber added lightweight concrete using neural
rules for reservoirs considering the water quality issues: appli- network. Comput Mater Sci 42(2):259–265
cation of ANFIS-based surrogate models. Exp Syst Appl 59. Gazder U, Al-Amoudi OSB, Saad Khan SM, Maslehuddin M
37:6639–6645 (2017) Predicting compressive strength of blended cement
40. Ma XX, Guo HF, Chen X (2007) Water quality evaluation concrete with ANNs. Comput Concr 20(6):627–634
model based on ANFIS and its application. Water Resour Prot 60. Onyari EK, Ikotun BD (2018) Prediction of compressive and
23:12–14 flexural strengths of a modified zeolite additive mortar using
41. Ziari H, Sobhani J, Ayoubinejad J, Hartmann T (2016) Ana- artificial neural network. Constr Build Mater 187:1232–1241
lysing the accuracy of pavement performance models in the 61. Naderpour H, Mirrashid M (2018) An innovative approach for
short and long terms: GMDH and ANFIS methods. Road Mater compressive strength estimation of mortars having calcium
Pave Des 17:619–637 inosilicate minerals. J Build Eng 19:205–215
42. Stojčić M (2018) Application of ANFIS model in road traffic 62. Zurada JM (1992) Introduction to artificial neural systems. West
and transportation: A literature review from 1993 to 2018. Oper St, Paul
Res Eng Sci Theory Appl 1:40–61 63. Armaghani DJ, Raja RSNSB, Faizi K, Rashid ASA (2017)
43. Özel C, Topsakal A (2015) Comparison of ANFIS and ANN for Developing a hybrid PSO–ANN model for estimating the ulti-
estimation of thermal conductivity coefficients of construction mate bearing capacity of rock-socketed piles. Neural Comput
materials. Sci Iran 22:2001–2011 Appl 28(2):391–405
44. Yadollahi MM, Benli A, Demirboga R (2017) Application of 64. Mohamad ET, Armaghani DJ, Momeni E, Yazdavar AH,
adaptive neuro-fuzzy technique and regression models to predict Ebrahimi M (2018) Rock strength estimation: a PSO-based BP
the compressive strength of geopolymer composites. Neural approach. Neural Comput Appl 30(5):1635–1646
Comput Appl 28:1453–1461 65. Mohamad ET, Hajihassani M, Armaghani DJ, Marto A (2012)
45. Abunama T, Othman F, Younes MK (2018) Predicting sanitary Simulation of blasting-induced air overpressure by means of
landfill leachate generation in humid regions using ANFIS artificial neural networks. Int Rev Modell Simul 5:2501–2506
modeling. Environ Monit Assess 190:597 66. Haykin S (1999) Neural networks: a comprehensive foundation.
46. Kebria DY, Ghavami M, Javadi S, Goharimanesh M (2018) Prentice Hall, Upper Saddle River, New Jersey
Combining an experimental study and ANFIS modeling to 67. Dreyfus G (2005) Neural networks: methodology and applica-
predict landfill leachate transport in underlying soil—a case tion. Springer, Berlin
study in north of Iran. Environ Monit Assess 190:26 68. Jang J-S (1993) ANFIS: adaptive-network-based fuzzy inference
47. Safa M, Shariati M, Ibrahim Z et al (2016) Potential of adaptive system. IEEE Trans Syst Man Cybern 23:665–685
neuro fuzzy inference system for evaluating the factors affecting 69. Jang J-S, Sun C-T (1995) Neuro-fuzzy modeling and control.
steel-concrete composite beam’s shear strength. Steel Compos Proc IEEE 83:378–406
Struct 21:679–688 70. Ali OAM, Ali AY, Sumait BS (2015) Comparison between the
48. Jafari F, Badarloo B (2019) Finite Element Analysis and ANFIS effects of different types of membership functions on fuzzy
investigation of seismic behavior of sandwich panels with dif- logic controller performance. Int J 76:76–83
ferent concrete material in two story steel building. Frat ed 71. Walia N, Singh H, Sharma A (2015) ANFIS: adaptive neuro-
Integrità Strutt 13:209–230 fuzzy inference system-a survey. Int J Comput Appl 123:13
49. Mashrei MA, Mahdi AM (2019) An adaptive neuro-fuzzy 72. Armaghani DJ, Hajihassani M, Sohaei H et al (2015) Neuro-
inference model to predict punching shear strength of flat con- fuzzy technique to predict air-overpressure induced by blasting.
crete slabs. Appl Sci 9:809 Arab J Geosci 8:10937–10950. https://doi.org/10.1007/s12517-
50. Darain KM, Shamshirband S, Jumaat MZ, Obaydullah M (2015) 015-1984-3
Adaptive neuro fuzzy prediction of deflection and cracking 73. Momeni E, Armaghani DJ, Fatemi SA, Nazir R (2018) Predic-
behavior of NSM strengthened RC beams. Constr Build Mater tion of bearing capacity of thin-walled foundation: a simulation
98:276–285 approach. Eng Comput 34:319–327
51. Naderpour H, Mirrashid M (2019) Moment capacity estimation 74. Vu DD, Stroeven P, Bui VB (2001) Strength and durability
of spirally reinforced concrete columns using ANFIS. Compl aspects of calcined kaolin-blended Portland cement mortar and
Intell Syst. https://doi.org/10.1007/s40747-019-00118-2 concrete. Cem Concr Compos 23(6):471–478
52. Ince R (2004) Prediction of fracture parameters of concrete by 75. Courard L, Darimont A, Schouterden M, Ferauche F, Willem X,
Artificial Neural Networks. Eng Fract Mech 71(15):2143–2159 Degeimbre R (2003) Durability of mortars modified with
53. Adhikary BB, Mutsuyoshi H (2006) Prediction of shear strength metakaolin. Cem Concr Res 33(9):1473–1479
of steel fiber RC beams using neural networks. Constr Build 76. Parande AK, Ramesh Babu B, AswinKarthik M, Deepak
Mater 20(9):801–811 Kumaar KK, Palaniswamy N (2008) Study on strength and
54. Kewalramani MA, Gupta R (2006) Concrete compressive corrosion performance for steel embedded in metakaolin blen-
strength prediction using ultrasonic pulse velocity through arti- ded concrete/mortar. Constr Build Mater 22(3):127–134
ficial neural networks. Autom Constr 15(3):374–379 77. Sumasree C, Sajja S (2016) Effect of Metakaolin and Cer-
55. Pala M, Özbay E, Öztaş A, Yuce MI (2007) Appraisal of long- afibermix on mechanical and durability properties of mortars. Int
term effects of fly ash and silica fume on compressive strength J Sci Eng Technol 4(3):501–506
of concrete by neural networks. Constr Build Mater 78. Batis G, Pantazopoulou P, Tsivilis S, Badogiannis E (2005) The
21(2):384–394 effect of metakaolin on the corrosion behavior of cement mor-
56. Topçu IB, Saridemir M (2007) Prediction of properties of waste tars. Cem Concr Compos 27(1):125–130
AAC aggregate concrete using artificial neural network. Comput 79. Kadri EH, Kenai S, Ezziane K, Siddique R, De Schutter G
Mater Sci 41(1):117–125 (2011) Influence of metakaolin and silica fume on the heat of

123
4532 Neural Computing and Applications (2021) 33:4501–4532

hydration and compressive strength development of mortar. 94. Asteris PG, Armaghani DJ, Hatzigeorgiou GD, Karayannis CG,
Appl Clay Sci 53(4):704–708 Pilakoutas K (2019) Predicting the shear strength of reinforced
80. Mardani-Aghabaglou A, Sezer Gİ, Ramyar K (2014) Compar- concrete beams using Artificial Neural Networks. Comput Concr
ison of fly ash, silica fume and metakaolin from mechanical 24(5):469–488. https://doi.org/10.12989/cac.2019.24.5.469
properties and durability performance of mortar mixtures view 95. Apostolopoulou M, Armaghani DJ, Bakolas A, Douvika MG,
point. Constr Build Mater 70:17–25 Moropoulou A, Asteris PG (2019) Compressive strength of
81. Potgieter-Vermaak SS, Potgieter JH (2006) Metakaolin as an natural hydraulic lime mortars using soft computing techniques.
extender in South African cement. J Mater Civ Eng Proc Struct Integr 17:914–923
18(4):619–623 96. Asteris PG, Moropoulou A, Skentou AD, Apostolopoulou M,
82. Cizer O, Van Balen K, Van Gemert D, Elsen J (2008) Blended Mohebkhah A, Cavaleri L, Rodrigues H, Varum H (2019)
lime-cement mortars for conservation purposes: microstructure Stochastic vulnerability assessment of masonry structures:
and strength development. In: Structural analysis of historic concepts, modeling and restoration aspects. Appl Sci 9(2):243
construction: preserving safety and significance—proceedings 97. Huang L, Asteris PG, Koopialipoor M, Armaghani DJ, Tahir
of the 6th international conference on structural analysis of MM (2019) Invasive weed optimization technique-based ANN
historic construction, SAHC08, 2, pp 965–972 to the prediction of rock tensile strength. Appl Sci 9:5372
83. Al-Chaar GK, Alkadi M, Asteris PG (2013) Natural pozzolan as 98. Asteris PG, Argyropoulos I, Cavaleri L, Rodrigues H, Varum H,
a partial substitute for cement in concrete. Open Constr Build Thomas J, Lourenço PB (2018) Masonry compressive strength
Technol J 7:33–42 prediction using artificial neural networks. In International
84. SPSS Inc (2007) SPSS for windows (version 16.0). SPSS Inc., conference on transdisciplinary multispectral modeling and
Chicago cooperation for the preservation of cultural heritage, Springer,
85. Khandelwal M, Armaghani DJ, Faradonbeh RS, Ranjith PG, Cham, Switzerland, pp 200–224
Ghoraba S (2016) A new model based on gene expression 99. Asteris PG, Ashrafian A, Rezaie-Balf M (2019) Prediction of the
programming to estimate air flow in a single rock joint. Environ compressive strength of self-compacting concrete using surro-
Earth Sci 75(9):739 gate models. Comput Concr 24(2):137–150
86. Momeni E, Nazir R, Armaghani DJ, Maizir H (2015) Applica- 100. Asteris PG, Nozhati S, Nikoo M, Cavaleri L, Nikoo M (2019)
tion of artificial neural network for predicting shaft and tip Krill herd algorithm-based neural network in structural seismic
resistances of concrete piles. Earth Sci Res J 19(1):85–93 reliability evaluation. Mech Adv Mater Struct
87. Armaghani DJ, Hajihassani M, Sohaei H, Mohamad ET, Marto 26(13):1146–1153
A, Motaghedi H, Moghaddam MR (2015) Neuro-fuzzy tech- 101. Asteris PG, Roussis PC, Douvika MG (2017) Feed-forward
nique to predict air-overpressure induced by blasting. Arab J neural network prediction of the mechanical properties of
Geosci 8(12):10937–10950 sandcrete materials. Sensors 17(6):1344
88. Yang H, Koopialipoor M, Armaghani DJ, Gordan B, Khorami 102. Apostolopoulou M, Asteris PG, Armaghani DJ, Douvika MG,
M, Tahir MM (2019) Intelligent design of retaining wall struc- Lourenço PB, Cavaleri L, Bakolas A, Moropoulou A (2020)
tures under dynamic conditions. Steel Compos Struct Mapping and holistic design of natural hydraulic lime mortars.
31(6):629–640 Cem Concr Res 136:106167. https://doi.org/10.1016/j.cem
89. Harandizadeh H, Armaghani DJ, Khari M (2019) A new conres.2020.106167
development of ANFIS–GMDH optimized by PSO to predict 103. Ly H, Pham BT, Le LM et al (2020) Estimation of axial load-
pile bearing capacity based on experimental datasets. Eng carrying capacity of concrete-filled steel tubes using surrogate
Comput. https://doi.org/10.1007/s00366-019-00849-3 models. Neural Comput Appl. https://doi.org/10.1007/s00521-
90. Chen W, Sarir P, Bui XN, Nguyen H, Tahir MM, Armaghani DJ 020-05214-w
(2019) Neuro-genetic, neuro-imperialism and genetic program- 104. Asteris PG, Mokos VG (2020) Concrete compressive strength
ing models in predicting ultimate bearing capacity of pile. Eng using artificial neural networks. Neural Comput Appl
Comput. https://doi.org/10.1007/s00366-019-00752-x 32:1807–11826. https://doi.org/10.1007/s00521-019-04663-2
91. Xu H, Zhou J, Asteris GP, Jahed Armaghani D, Tahir MM 105. Asteris PG, Apostolopoulou M, Skentou AD, Antonia Mor-
(2019) Supervised machine learning techniques to the prediction opoulou A (2019) Application of artificial neural networks for
of tunnel boring machine penetration rate. Appl Sci 9(18):3715 the prediction of the compressive strength of cement-based
92. Murlidhar BR, Kumar D, Jahed Armaghani D, Mohamad ET, mortars. Comput Concr 24(4):329–345
Roy B, Pham BT (2020) A novel intelligent ELM-BBO tech- 106. Tutmez B, Dag A, Tercan AE, Kaymak U (2007) Lignite
nique for predicting distance of mine blasting-induced flyrock. thickness estimation via adaptive fuzzy-neural network. In:
Nat Resour Res. https://doi.org/10.1007/s11053-020-09676-6 Proceedings of the 20th international mining congress and
93. Armaghani DJ, Hatzigeorgiou GD, Karamani Ch, Skentou A, exhibition of Turkey (IMCET 2007), pp 151–157
Zoumpoulaki I, Asteris PG (2019) Soft computing based tech-
niques for concrete beams shear strength. Proc Struct Integr Publisher’s Note Springer Nature remains neutral with regard to
17:924–933. https://doi.org/10.1016/j.prostr.2019.08.123 jurisdictional claims in published maps and institutional affiliations.

123

You might also like