polymers

Article
Compressive Strength Estimation of Geopolymer Composites
through Novel Computational Approaches
Muhammad Nasir Amin 1, * , Kaffayatullah Khan 1 , Waqas Ahmad 2 , Muhammad Faisal Javed 2 ,
Hisham Jahangir Qureshi 1 , Muhammad Umair Saleem 3 , Muhammad Ghulam Qadir 4
and Muhammad Iftikhar Faraz 5

1 Department of Civil and Environmental Engineering, College of Engineering, King Faisal University,
Al-Ahsa 31982, Saudi Arabia; [email protected] (K.K.); [email protected] (H.J.Q.)
2 Department of Civil Engineering, COMSATS University Islamabad, Abbottabad 22060, Pakistan;
[email protected] (W.A.); [email protected] (M.F.J.)
3 Service Stream Limited Co., Chatswood, NSW 206, Australia; [email protected]
4 Department of Environmental Sciences, COMSATS University Islamabad, Abbottabad 22060, Pakistan;
[email protected]
5 Department of Mechanical Engineering, College of Engineering, King Faisal University,
Al-Ahsa 31982, Saudi Arabia; [email protected]
* Correspondence: [email protected]; Tel.: +966-13-589-5431; Fax: +966-13-581-7068

Abstract: The application of artificial intelligence approaches like machine learning (ML) to forecast
material properties is an effective strategy to reduce multiple trials during experimentation. This
study performed ML modeling on 481 mixes of geopolymer concrete with nine input variables,
including curing time, curing temperature, specimen age, alkali/fly ash ratio, Na2 SiO3 /NaOH ratio,
NaOH molarity, aggregate volume, superplasticizer, and water, with CS as the output variable. Four
types of ML models were employed to anticipate the compressive strength of geopolymer concrete,
Citation: Amin, M.N.; Khan, K.;
and their performance was compared to find out the most accurate ML model. Two individual ML
Ahmad, W.; Javed, M.F.; Qureshi, H.J.;
techniques, support vector machine and multi-layer perceptron neural network, and two ensembled
Saleem, M.U.; Qadir, M.G.; Faraz, M.I.
ML methods, AdaBoost regressor and random forest, were employed to achieve the study’s aims. The
Compressive Strength Estimation of
performance of all models was confirmed using statistical analysis, k-fold evaluation, and correlation
Geopolymer Composites through
Novel Computational Approaches.
coefficient (R2 ). Moreover, the divergence of the estimated outcomes from those of the experimental
Polymers 2022, 14, 2128. https:// results was noted to check the accuracy of the models. It was discovered that ensembled ML models
doi.org/10.3390/polym14102128 estimated the compressive strength of the geopolymer concrete with higher precision than individual
ML models, with random forest having the highest accuracy. Using these computational strategies
Academic Editors: Wei-Hao Lee,
will accelerate the application of construction materials by decreasing the experimental efforts.
Yung-Ching Ding and Kae-Long Lin

Received: 29 March 2022 Keywords: machine learning; geopolymer concrete; artificial intelligence; prediction models;
Accepted: 18 May 2022 compressive strength
Published: 23 May 2022

Publisher’s Note: MDPI stays neutral

with regard to jurisdictional claims in
published maps and institutional affil- 1. Introduction
iations. Construction is a vital component of any economy [1]. The building sector generates
large amounts of waste and emits considerable amounts of greenhouse gases (GHGs) into the
environment [2,3]. Cement-based materials such as concrete are the primary building mate-
rials utilized in the construction industry worldwide [4–7]. It is now well accepted that the
Copyright: © 2022 by the authors.
manufacture of cement leads to the emission of significant amounts of GHGs that contribute to
Licensee MDPI, Basel, Switzerland.
global warming, as well as the use of significant amounts of raw materials [8–10]. It has been
This article is an open access article
calculated that around two tons of raw materials (shale and limestone) are consumed in the
distributed under the terms and
manufacturing of one ton of cement, and approximately one ton of carbon dioxide (CO2 ) and
conditions of the Creative Commons
Attribution (CC BY) license (https://
nitrogen oxide (NO) pollutants are released [11]. With over two billion tons of GHGs emitted
creativecommons.org/licenses/by/
yearly as a result of cement manufacture, cement production accounts for approximately 6%
4.0/).
of global anthropogenic GHG emissions [12–14]. The extensive use of natural raw materials

Polymers 2022, 14, 2128. https://doi.org/10.3390/polym14102128 https://www.mdpi.com/journal/polymers

Polymers 2022, 14, 2128 2 of 21

in the manufacture of cement has also resulted in the overexploitation of natural resource
reserves, resulting in a degradation of the aesthetics of the environment and the modification
of ecosystems [15,16]. Apart from the substantial GHG emissions associated with cement
manufacture, the process is extremely energy demanding [17,18]. Recent urbanization, partic-
ularly in developing nations, has exacerbated the negative environmental effect of cement
manufacturing [19]. As a result, it is critical that sustainable alternatives to cement be utilized
in building applications in order to preserve the environment’s sustainability [20,21]. Numer-
ous waste products created by various sectors can be utilized as sustainable substitutes for the
traditional resources used in the cement manufacturing process. As a result, the utilization
of such wastes in the manufacturing of a sustainable alternative to cement would result in a
considerable decrease in GHG emissions, the cost of raw materials, and the use of natural raw
resources connected with cement [22]. Materials that have been activated with alkali, such as
geopolymers, may be preferred to conventional cement concrete [23–25].
Davidovits was the first to propose geopolymers consisting of semi-crystalline three-
dimensional aluminosilicate materials in 1979 [26]. These geopolymers may be manufac-
tured using a variety of source materials, including fly ash, metakaolin, ground granu-
lated blast furnace slag, and rice husk bark ash [27–30]. Since then, scientists have paid
close attention to geopolymers due to their unique combination of superior mechanical
performance, chemical and fire resistance, low CO2 emissions, and low energy consump-
tion [31,32]. These features are intimately connected to the chemical interactions between
aluminosilicate and alkali-polysialate [33]. The use of geopolymer concrete (GeoPC) in
place of conventional cement concrete results in an embodied carbon reduction of up to
80%, depending on the precursor and activator utilized [34]. GeoPC is mostly composed of
waste materials from various industrial and agricultural activities. GeoPC may be consid-
ered more ecologically friendly and an efficient method of managing enormous amounts
of waste created by industries [35–37]. The utilization of locally accessible materials as
precursors, such as laterite soil, can help increase the sustainability of GeoPCs [11]. Thus, by
utilizing geopolymers as a sustainable alternative to cement, GHG emissions, raw material
consumption, and waste management costs would be significantly reduced [38–40].
The practice of developing models for forecasting the strength of concrete is ongoing
in order to reduce unnecessary test repetitions and material waste. There are several
prominent models for modeling concrete properties, such as best fit curves (based on
regression analysis). However, due to the nonlinear behavior of concrete [41], regression
models generated using this technique may not accurately represent the underlying nature
of the material. Additionally, regression methods may understate the effect of constituent
materials in concrete [42]. Artificial intelligence techniques such as machine learning (ML)
are some of the more contemporary modeling techniques that have been used in the area
of civil engineering. These approaches use input parameters to model responses, and
the output models are validated by experimentation. For construction applications, ML
algorithms estimate concrete strength [43–47], bituminous mixture performance [48], and
concrete durability [49–51].
This study focuses on the application of ML techniques to forecast the compressive
strength (C-S) of GeoPC. Four distinct ML techniques were used, including support vector
machine (SVM), multi-layer perceptron neural network (MLPNN), AdaBoost regressor
(AR), and random forest (RF) to anticipate the C-S of GeoPC. The effectiveness of all
techniques was evaluated by applying statistical tests and correlation coefficients (R2 ).
Furthermore, k-fold analysis and error distributions were used to determine the validity of
each technique. SVM and MLPNN are individual ML techniques, while AR and RF are
ensemble ML methods [52]. This study is interesting in that it predicts the C-S of GeoPC
utilizing both individual and ensemble ML techniques. However, experimental studies
require considerable human effort, the cost for experimentation, and time for material
collection, sample casting, curing, and testing. The application of novel methods, such as
ML, in the construction field to anticipate material characteristics will decrease the aforesaid
issues by obviating the need for experimental work. ML methods need a data set, which
Polymers 2022, 14, 2128 3 of 21

may be collected from the past studies since a considerable amount of investigation has
been undertaken to determine material characteristics, and the data set might be utilized
for training the ML models and forecasting the material properties. The purpose of this
work is to ascertain the top appropriate ML method for the C-S estimation of GeoPC based
on the results estimation and the effect of input variables on ML model performance.

2. Data Description
ML methods need a diverse range of input parameters to acquire the desired out-
come [53]. The C-S of GeoPC was forecasted utilizing data obtained from past studies
(see Table S1, Supplementary Materials). The data set was arbitrarily selected from the
past studies to avoid biased images. This analysis obtained only a C-S-based data set to
run the models. The precursor material and activation solution were the same for all data
samples, i.e., fly ash and Na2 SiO3 –NaOH solution, respectively. Nine input parameters
were employed to run the models, including curing temperature, curing time, specimen
age, alkali/fly ash ratio, Na2 SiO3 /NaOH ratio, NaOH molarity, aggregate volume, super-
plasticizer, and water, with C-S as the output variable. In the present research, a data set
of 481 points was utilized for the outcome prediction using ML methods. The quantity of
input parameters and data sets have a considerable impact on the technique’s results [54].
According to prior research, a minimum of 300 data points and eight input variables can
result in increased precision for ML models [55,56]. As a result, the data set acquired for this
research is optimal for the ML model’s performance. Table 1 lists the descriptive statistical
analysis of all input variables. The mode, median, and mean values correspond to central
propensity, while the standard deviation, minimum, and maximum values correspond to
irregularity. Figure 1 depicts the dispersion of input parameters utilized in the research in
terms of their relative incidence. It illustrates the overall number of observations linked to
each value or sequence of values.

Table 1. Results of the descriptive statistical analysis of input parameters.

Curing Age of NaOH

Parameter Temperature Curing Specimen Alkali/Fly Na2 SiO3 /NaOH Molarity Aggregate Superplasticizer
Time (h) Ash Ratio Ratio Volume (%) (%) Water (%)
(◦ C) (days) (M)
Mean 68.94 27.46 21.53 0.43 2.25 11.89 59.98 1.93 53.56
Median 70.00 24.00 7.00 0.40 2.50 12.00 70.00 1.55 55.90
Mode 60.00 24.00 7.00 0.35 2.50 10.00 70.00 2.00 55.90
Standard
25.19 13.24 45.33 0.11 0.53 2.73 28.97 2.41 3.82
Deviation
Range 100.00 92.00 539.00 0.70 3.60 12.00 80.00 11.30 18.90
Minimum 20.00 4.00 1.00 0.30 0.40 8.00 0.00 0.00 45.10
Maximum 120.00 96.00 540.00 1.00 4.00 20.00 80.00 11.30 64.00

3. Machine Learning Methods Employed

Individual ML approaches (SVM and MLPNN), as well as ensemble ML methods
(AR and RF), were employed to ascertain the goals of this research with Python codes
through the Anaconda Navigator software. Spyder (version 4.3.5) was selected to run
the SVM, MLPNN, AR, and RF techniques. These ML methods are typically employed
to forecast the required results on the basis of input factors. These methods, amongst
other aspects, are able to estimate the temperature influence, the strength characteristics,
and the material’s durability [57,58]. The R2 value for the expected outcome indicates
the performance/validity of ML methods. The R2 is a statistic that is used to estimate
the degree of variation in a response variable specified by a model. In other words, it
quantifies the model’s fit to the data. A value close to zero suggests that fitting the mean is
similar to fitting the model, whereas a value near one indicates that the date and model are
virtually perfectly suited [59]. The data are split: 20% for testing and 80% for training the
ML models. The sub-segments underneath describe the ML approaches used in this study.
Furthermore, k-fold evaluation, statistical checks, and error measurements (root mean
square error (RMSE) and mean absolute error (MAE)) is performed on all ML methods to
Polymers 2022, 14, 2128 4 of 21

Polymers 2022, 14, x

validate them. In addition, sensitivity analysis (SA) is carried out to find out the influence
of every input variable on the results anticipation. The flow diagram in Figure 2 describes
the research technique followed in the present study.

120 450

400
100
350

80 300

250
Count

Count
60
200

40 150

100
20
50

0 0
20 40 60 80 100 120 140 0 20 40 60 80 100
Curing Temperature °C Curing Time (h)

Curing temperature Curing time

500 250

400 200

300 150
Count

Count

200 100

100 50

0 0
0 100 200 300 400 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Age of specimen (days) Alkali/fly ash ratio

Age of specimen Alkali/fly ash ratio

350 160

300 140

120
250
100
200
Count

Count

80
150
60
100
40

50 20

0 0
0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 8 10 12 14 16 18 20 22
Na2SiO3/NaOH ratio NaOH molarity

Na2SiO3/NaOH ratio NaOH molarity

Figure 1. Cont.
 Polymers 2022, 14, x

Polymers 2022, 14, 2128 5 of 21

160 200

140 180
160
120
140
100 120

Count
Count

80 100

60 80
60
40
40
20 20
0 0
0 20 40 60 80 100 0 2 4 6 8 10 12 14
Aggregate volume (%) Superplasticizer (%)

Aggregate volume Superplasticizer

300

250

200
Count

150

100

50

0
45 50 55 60 65 70
Water (%)

Polymers 2022, 14, x Water 6 of 24

Figure 1. Relative
Figure frequencyfrequency
1. Relative dispersal of inputs parameters.
dispersal of inputs parameters.

3. Machine Learning Methods Employed

Individual ML approaches (SVM and MLPNN), as well as ensemble ML me
(AR and RF), were employed to ascertain the goals of this research with Python
through the Anaconda Navigator software. Spyder (version 4.3.5) was selected to r
SVM, MLPNN, AR, and RF techniques. These ML methods are typically emplo
forecast the required results on the basis of input factors. These methods, amongst
aspects, are able to estimate the temperature influence, the strength characteristic
the material’s durability [57,58]. The R2 value for the expected outcome indicat
performance/validity of ML methods. The R2 is a statistic that is used to estima
degree of variation in a response variable specified by a model. In other wo
quantifies the model’s fit to the data. A value close to zero suggests that fitting the
is similar to fitting the model, whereas a value near one indicates that the date and
are virtually perfectly suited [59]. The data are split: 20% for testing and 80% for tr
the ML models. The sub-segments underneath describe the ML approaches used
study. Furthermore, k-fold evaluation, statistical checks, and error measurements
mean square error (RMSE) and mean absolute error (MAE)) is performed on a
methods to validate them. In addition, sensitivity analysis (SA) is carried out to fin
Figure the influence of for
every input variable on the results anticipation. The flow diagr
Figure2.2.Sequence
Sequenceemployed
employed thethe
for study.
study.
Figure 2 describes the research technique followed in the present study.
3.1. Support Vector Machine
SVM is an individual ML technique that is used to evaluate data for classification and
regression. An SVM technique is a way of describing the samples as points in space that
have been plotted in such a way that the patterns of the unique classifications are
separated by a distinct vector (line/plane) with the greatest possible separation.
 Figure 2. Sequence employed for the study.
Polymers 2022, 14, 2128 6 of 21
3.1. Support Vector Machine
SVM is an individual ML technique that is used to evaluate data for classification a
3.1. Support
regression. An Vector
SVM Machine
technique is a way of describing the samples as points in space th
have been plotted in suchMLa technique
SVM is an individual way that thatthe
is used to evaluate
patterns of data
the for classification
unique and
classifications a
regression. An SVM technique is a way of describing the samples as points in space that
separated by a distinct vector (line/plane) with the greatest possible
have been plotted in such a way that the patterns of the unique classifications are separated
separatio
Additional casesvector
by a distinct are then superimposed
(line/plane) on thatpossible
with the greatest same space and categorized
separation. Additional cases according
whichareside
thenof the vector on
superimposed they
thatlie
sameon,space
as illustrated in Figure
and categorized 3. to
according Figure
which 4sideillustrates
of t
the vector they lie on, as illustrated in Figure 3. Figure 4 illustrates the
procedure for the SVM model. This model is employed to assess the material’s streng procedure for the
sinceSVM model. This model is employed to assess the material’s strength, since it takes into
it takes into account the combined influence of various components. T
account the combined influence of various components. The optimization approach is used
optimization approach
to ascertain is used
the parameters to SVM
of the ascertain
model.the parameters of the SVM model.

Polymers 2022, 14, x 7 of 24

FigureFigure
3. Support vector
3. Support machine
vector modelmapping.
machine model mapping. Reprinted/adapted
Reprinted/adapted with permission
with permission from [60]. from [60

Data collection
80% 20%

Training data Crossover Testing data

Save mode
SVM
parameters

Poor Results Good Final SVM

comparison model

Figure
Figure 4. Sequence
4. Sequence ofofsupport
supportvector
vector machine
machinemodeling
modelingprocess.
process.
3.2. Multi-Layer Perceptron Neural Network
3.2. Multi-Layer Perceptron Neural Network
An artificial neural network (ANN) is a collection of connected nodes that are em-
An artificial
ployed neural
to represent network
and solve issues(ANN) is a complicated
that involve collection of connected
interactions nodes
among that are
causal
employed to reactions.
events and represent MLPNN
and solve issues
is one that
of the involve
highly complicated
efficient ANN methodsinteractions among
for estima-
causal events and reactions. MLPNN is one of the highly efficient ANN methods for
estimation and modeling. MLPNN has been chosen as the standard method in numerous
studies [61,62]. Due to MLPNN’s excellent universal approximation capabilities, it has
been commonly utilized to describe nonlinear and complicated phenomena in the actual
 3.2. Multi-Layer Perceptron Neural Network
An artificial neural network (ANN) is a collection of connected nodes that ar
employed to represent and solve issues that involve complicated interactions amon
causal events and reactions. MLPNN is one of the highly efficient ANN methods fo
Polymers 2022, 14, 2128 7 of 21
estimation and modeling. MLPNN has been chosen as the standard method in numerou
studies [61,62]. Due to MLPNN’s excellent universal approximation capabilities, it ha
been commonly utilized to describe nonlinear and complicated phenomena in the actu
tion and modeling. MLPNN has been chosen as the standard method in numerous
world [63–65]. The MLPNN is a feed-forward technique that comprises a single inpu
studies [61,62]. Due to MLPNN’s excellent universal approximation capabilities, it has
layer,
been one or more
commonly utilizedhidden layers,
to describe and a and
nonlinear single output layer
complicated [66], asinshown
phenomena in Figure
the actual
world [63–65]. The MLPNN is a feed-forward technique that comprises a single input layer, source
Usually, the number of nodes in the input layer is determined by the data
specified
one or morefactor,
hidden while
layers, the
and number of hidden
a single output neurons
layer [66], is measured
as shown in Figure 5.using a particula
Usually,
training
the numberdata set. The
of nodes hidden
in the layers
input layer are utilizedby
is determined forthe
computing,
data source’swhereas
specifiedthe output laye
factor,
while the for
is used number of hidden
modeling. neurons
Every nodeisinmeasured usinglayer
the hidden a particular
shouldtraining datatoset.
be linked all The
nodes in th
hidden
input layers
layer are
andutilized
then to forall
computing,
nodes inwhereas the output
the output layer.layer
The isMLPNN
used for training
modeling.operatio
Every
mightnode in the
split intohidden
two layer
stepsshould be linked
via these to all nodes
connections: in the and
ahead inputback,
layer and then tothe back
utilizing
all nodes in the output layer. The MLPNN training operation might split into two steps via
propagation technique [63].
these connections: ahead and back, utilizing the back-propagation technique [63].

Figure 5. Sequence of a multi-layer perceptron neural network modeling process. Reprinted/adapted

with permission from [67].

3.3. AdaBoost Regressor

The AR method is the most common ensemble ML technique in the boosting class.
The AR algorithm is unique in that it uses the primary training data to develop a weak
learner, and then alters its dispersion of training data based on the projection performance
of the weak learner in the subsequent turn of weak learner training. It is important to
mention that in the subsequent phase, the training models with lower estimation accuracy
from the former phase will receive greater consideration. Following that, the weak learners
are combined with a strong learner using a range of weights to create the final pattern [68].
The AR running process is divided into four stages, including data collecting, developing a
strong learner, analyzing or confirming the learner, and applying the learner to engineering
problems. The second phase is critical to the AR method. As stated before, it is composed
of two elements, i.e., a structure for incorporating weak learners into a stronger one and
a regression learning algorithm for generating the weak learner from the training data.
The SVM technique is employed to construct the weak learner, and the weak learners are
combined using the average of the weighted weak learners. The flow diagram for this
approach is depicted in Figure 6.
 pattern [68]. The AR running process is divided into four stages, including data collecting,
developing a strong learner, analyzing or confirming the learner, and applying the learner
to engineering problems. The second phase is critical to the AR method. As stated before,
it is composed of two elements, i.e., a structure for incorporating weak learners into a
stronger one and a regression learning algorithm for generating the weak learner from the
training data. The SVM technique is employed to construct the weak learner, and the weak
Polymers 2022, 14, 2128 8 of 21
learners are combined using the average of the weighted weak learners. The flow diagram
for this approach is depicted in Figure 6.

80% 20%
Training set Data collection Testing set

Weak learner weights

DT Validation
Weight distribution D1 Weak learner G1 W1

Update the sample weights

DT
Weight distribution D2 Weak learner G2 W2 𝛴 Strong learner

Update the sample weights

Weight distribution Dn Weak learner Gn Wn Resultant model

Figure 6. Sequence
Figure ofof
6. Sequence AdaBoost
AdaBoost regressor modeling
regressor modeling process
process [69]. [69].

3.4. Random Forest

3.4. Random Forest
The The random
random split
split selection
selection techniqueisisused
technique usedtotodeploy
deployRFRFononbagging
bagging DTsDTs [70].
[70]. Figure 7
Figure 7 schematically depicts the modeling method of the RF technique.
schematically depicts the modeling method of the RF technique. Each tree in the Each tree in the forest is
forest is produced from an aimlessly selected training set, and every split within a tree is
produced from an aimlessly selected training set, and every split within a tree is constructed
constructed from an erratically chosen subgroup of input parameters, developing a forest
from[71].
an This
erratically
elementchosen subgroup
of uncertainty of input
increases parameters,
the tree’s variety. Thedeveloping a made
entire forest is forestup[71]. This
element of uncertainty increases the tree’s variety. The entire forest
of completely mature binary trees. The RF approach has proven to be an extremely is made up of completely
mature binarytool
powerful trees.
forThe RF approachclassification
general-purpose has proven toandberegression.
an extremely
When powerful
the numbertool for
of general-
purpose classification
variables surpassesand the regression. When the number
number of observations, of variables
the approach, whichsurpasses
aggregates thethe
number of
predictions of numerous randomized DTs, demonstrates increased
observations, the approach, which aggregates the predictions of numerous randomized DTs, precision.
Polymers 2022, 14, x Furthermore,
demonstrates it is adjustable
increased precision.to Furthermore,
both large-scaleitand ad hoc learning
is adjustable to bothtasks, returning
9 of 24
large-scale and ad hoc
measures with varying degrees of significance [72].
learning tasks, returning measures with varying degrees of significance [72].

Figure 7.7.Sequence
Figure of random
Sequence forest modeling
of random process. Reprinted/adapted
forest modeling with permissionwith
process. Reprinted/adapted frompermission
[70].
from [70].
4. Results and Discussions
4.1. Support Vector Machine Model
The outcomes of the SVM model for the C-S of GeoPC are displayed in Figures 8 and
9. The correlation among the experimental and forecasted results is shown in Figure 8.
The SVM method generated results with a lower degree of accuracy and a marginal
Polymers 2022, 14, 2128 9 of 21

4. Results and Discussions

4.1. Support Vector Machine Model
The outcomes of the SVM model for the C-S of GeoPC are displayed in Figures 8 and 9.
The correlation among the experimental and forecasted results is shown in Figure 8. The
SVM method generated results with a lower degree of accuracy and a marginal difference
amongst the experimental and forecasted results. The R2 of 0.78 confirms that the SVM model
has a lower degree of accuracy in anticipating the C-S of GeoPC. Figure 9 demonstrates the
dispersion of experimental, anticipated, and error values for the SVM model for testing data
alone, which is 20% of the overall data set. The analysis of the experimental and estimated
values discovered that the divergence of outcomes (error) was in the limit of 0.00 to 47.0 MPa,
with an average of 7.72 MPa. Moreover, for 8 mixes, the divergence from the experimental
results was lower than 1 MPa; for 17 mixes, the divergence was between 1 and 3 MPa; for
21 mixes, the divergence was between 3 and 6 MPa; and for 45 mixes, the variance was greater
than
Polymers 2022, 14, x 6 MPa. This indicated a higher deviation from the projected findings for the
10 ofSVM
24 model
Polymers 2022, compared
14, x to the experimental results. Thus, the SVM technique is less accurate in anticipating
10 of 24

the C-S of GeoPC.

70
70
y = 0.7721x + 8.5851
60 y2= 0.7721x + 8.5851
60 R 2= 0.7762
R = 0.7762
50
(MPa)

50
(MPa)

40
Estimated

40
Estimated

30
30
20
20
10
10
0
00 10 20 30 40 50 60 70 80
0 10 20 30 40 50 60 70 80
Experimental (MPa)
Experimental (MPa)
Figure 8. Relationship between experimental and estimated results for the support vector machine
Figure 8. Relationship
Figure 8.between
model. experimental
Relationship and estimated
between experimental results
and estimated forfor
results thethesupport vectormachine
support vector machine model.
model.

Experimental Estimated Error

80 Experimental Estimated Error
80
70
(MPa)

70
60
(MPa)

60
50
strength

50
strength

40
40
30
Compressive

30
Compressive

20
20
10
10
0
0
-10
-10 0 40 20 60 80 100
0 40 20 60 80 100
Data set (mix number)
Data set (mix number)
Figure 9. Distribution
Figure 9. of experimental,
Distribution estimated,
of experimental, and error
estimated, values
and error forforthe
values thesupport vector
support vector machine model.
machine
model.
Figure 9. Distribution of experimental, estimated, and error values for the support vector machine
model.
4.2. Multi-Layer Perceptron Neural Network Model
4.2. Multi-Layer Perceptron Neural Network Model
Polymers 2022, 14, 2128 10 of 21

4.2. Multi-Layer Perceptron Neural Network Model

Polymers 2022, 14, x
Figures 10 and 11 illustrate a contrast of the MLPNN model’s experimental and antici-
11 of 24
pated results. Figure 10 exemplifies the relationship among experimental and projected
outcomes, with an R2 of 0.81 suggesting that the MLPNN model is more specific than
the SVM model in Figures 10 and 11the
estimating illustrate
GeoPC a contrast of the MLPNN
C-S. Figure model’s experimental
11 illustrates and
the distribution of ex-
anticipated results. Figure 10 exemplifies the relationship among experimental and
perimental, estimated, and error values for the MLPNN model. The variation
projected outcomes, with an R2 of 0.81 suggesting that the MLPNN model is more specific
between
experimental than
andtheestimated
SVM modelvalues wasthe
in estimating found
GeoPCtoC-S.
beFigure
between 0.06 and
11 illustrates 22.77 MPa,
the distribution of with an
average of 5.86 MPa. Additionally,
experimental, estimated, andthe errorvariation from
values for the MLPNNthe model.
experimental results
The variation betweenwas lower
experimental and estimated values was found to be between 0.06 and 22.77 MPa, with an
than 1 MPa for 10 mixes, between 1 and 3 MPa for 25 mixes, between 3 and 6 MPa for
average of 5.86 MPa. Additionally, the variation from the experimental results was lower
23 mixes, andthan
greater
1 MPa than 6 MPa
for 10 mixes, for 39
between mixes.
1 and 3 MPaThis
for 25 also
mixes,indicates
between 3 anda greater
6 MPa fordivergence
23
of the MLPNN model’s
mixes, predicted
and greater than 6 MPa outcomes when
for 39 mixes. compared
This also to thedivergence
indicates a greater experimentalof the results.
Therefore, theMLPNN
MLPNN model’s predicted is
technique outcomes whenaccurate
also less comparedatto predicting
the experimental results. C-S, but
GeoPC’s
Therefore, the MLPNN technique is also less accurate at predicting GeoPC’s C-S, but
slightly more slightly
accuratemorethan thethan
accurate SVM the model.
SVM model.

70

y = 0.6896x + 13.372
60
R2 = 0.8098
50
Estimated (MPa)

40

30

20

10

0
0 10 20 30 40 50 60 70 80
Experimental (MPa)
Polymers 2022, 14, x 12 of 24
Figure 10. Relationship between between
Figure 10. Relationship experimental andand
experimental estimated results
estimated results for for the multi-layer
the multi-layer perceptron
perceptron
neural
neural network model. network model.

Experimental Estimated Error

80
70
Compressive strength (MPa)

60
50
40
30
20
10
0
-10
0 20 40 60 80 100
Data set (mix number)
Figure 11. Distribution
Figure 11. of experimental,
Distribution estimated,
of experimental, and error
estimated, values
and error valuesfor
forthe multi-layer
the multi-layer perceptron
perceptron
neural network model.
neural network model.

4.3. AdaBoost Regressor Model

A comparable illustration of the AR model results is depicted in Figures 12 and 13.
Figure 12 indicates the relationship among the experimental and anticipated results. The
AR method produced outcomes with a higher degree of exactness and a minimal
divergence amongst the experimental and projected results. The R2 of 0.89 indicates that
Polymers 2022, 14, 2128 11 of 21

4.3. AdaBoost Regressor Model

A comparable illustration of the AR model results is depicted in Figures 12 and 13.
Figure 12 indicates the relationship among the experimental and anticipated results. The
AR method produced outcomes with a higher degree of exactness and a minimal divergence
amongst the experimental and projected results. The R2 of 0.89 indicates that the AR model
is reasonably precise at predicting the C-S of GeoPC. The dispersal of the experimental,
anticipated, and error readings for the BR model are shown in Figure 13. The difference
(error) between the experimental and estimated values ranged from 0.00 to 22.80 MPa,
with a mean of 4.03 MPa. Furthermore, for 18 mixes, the variation from the experimental
outcomes was lower than 1 MPa; for 26 mixes, it was between 1 and 3 MPa; for 32 mixes, it
was between 3 and 6 MPa; and for only 21 mixes it was larger than 6 MPa. When compared
tox the experimental data, the AR model’s outcomes showed minimal divergence and
Polymers 2022, 14, 13 of higher
24

precision,
Polymers 2022, 14, x because this technique uses the training data to build a weak learner13and of 24then
trains it by altering the dispersal of the training data until it forms a strong learner.
80
80 y = 0.7919x + 9.7379
70
y 2==0.7919x
R 0.8923 + 9.7379
70
60 2
R = 0.8923
(MPa)

60
(MPa)

50
Estimated

50
40
Estimated

40
30
30
20
20
10
10
0
0 0 10 20 30 40 50 60 70
0 10 20 Experimental
30 40
(MPa) 50 60 70
Experimental (MPa)
Figure 12. Relationship between experimental and estimated results for the AdaBoost regressor
model.
Figure 12. Relationship
Figure 12.between experimental
Relationship and estimated
between experimental results for
and estimated the for
results AdaBoost regressor
the AdaBoost model.
regressor
model.
Experimental Estimated Error
80 Experimental Estimated Error
80
70
(MPa)

70
60
(MPa)

60
50
strength

50
strength

40
40
30
Compressive

30
Compressive

20
20
10
10
0
0
-10
-10 0 20 40 60 80 100
0 20 40 set (mix 60
Data number) 80 100

Figure 13. Distribution Data set (mix number)

Figure 13.ofDistribution
experimental, estimated,
of experimental, and errorand
estimated, values
errorfor the AdaBoost
values regressor
for the AdaBoost model.
regressor
model.
Figure 13. Distribution of experimental, estimated, and error values for the AdaBoost regressor
model.
4.4. Random Forest Model
4.4. Random Forest Model
Polymers 2022, 14, 2128 12 of 21

Polymers 2022, 14, x 14 of 24

4.4. Random Forest Model
Figures 14 and 15 present a similar representation of the RF model’s results. In
Figure 14, an R2 value of 14
Figures 0.95
andspecifies
15 presentthat the representation
a similar RF model performs with the
of the RF model’s highest
results. In Figure
precision compared 2 value of 0.95 specifies that the RF model performs with the highest precision
14, antoRthe other models employed in this study. Figure 15 exemplifies the
compared to the
scattering of experimental, other models
projected, andemployed in thisfor
error values study.
theFigure 15 exemplifies
RF model. the scattering
The variation
of experimental, projected, and error values for the RF model. The variation (error)
(error) between the experimental and estimated values was found to be between 0.05 and
between the experimental and estimated values was found to be between 0.05 and 14.99
14.99 MPa, with an average of 2.34 MPa. In addition, the variation from the experimental
MPa, with an average of 2.34 MPa. In addition, the variation from the experimental
outcomes was lower thanwas
outcomes 1 MPa
lowerfor 381mixes,
than MPa forbetween
38 mixes,1between
and 3 MPa1 andfor 29 mixes,
3 MPa between
for 29 mixes, between
3 and 6 MPa for 324 mixes,
and 6 MPaand greater
for 24 mixes,than 6 MPathan
and greater for only
6 MPa6formixes.
only 6This indicates
mixes. a smaller
This indicates a smaller
variation between the experimental
variation and predicted
between the experimental andoutcomes. Therefore,
predicted outcomes. the RFthe
Therefore, technique
RF technique
is more suitable,isdemonstrating the highest precision
more suitable, demonstrating the highestin estimating
precision the C-S the
in estimating of GeoPC.
C-S of GeoPC.

80
y = 0.9306x + 3.2466
70
R2 = 0.9521
60

50
Estimated (MPa)

40

30

20

10

0
0 10 20 30 40 50 60 70
Polymers 2022, 14, x Experimental (MPa) 15 of 24

Figure 14. Relationship between experimental and estimated outcomes for the random forest model.
Figure 14. Relationship between experimental and estimated outcomes for the random forest model.

Experimental Estimated Error

80
70
Comoressive strength (MPa)

60
50
40
30
20
10
0
-10
-20
0 20 40 60 80 100
Data set (mix number)
Figure 15. Distribution ofDistribution
Figure 15. experimental, estimated, and
of experimental, error and
estimated, values
errorfor the random
values forestforest
for the random model.model.

5. Model’s Validation
K-fold and statistical approaches were employed to validate the performance of all
models. Typically, the k-fold analysis method is carried out to find out the model’s
validity [73], during which related data are arbitrarily dispersed and split into 10 groups.
Polymers 2022, 14, 2128 13 of 21

5. Model’s Validation
K-fold and statistical approaches were employed to validate the performance of
all models. Typically, the k-fold analysis method is carried out to find out the model’s
validity [73], during which related data are arbitrarily dispersed and split into 10 groups.
Nine groups will be utilized for training the models and one will be used for validation. The
lower error values (MAE and RMSE) and the higher R2 values suggest the higher precision
of a model [69]. Moreover, the process must be repeated 10 times to obtain a suitable
decision. This broad endeavor provides the notable precision of a model. Moreover, as
displayed in Table 2, each ML method was statistically assessed based on errors (MAE and
RMSE). These evaluations also supported the ensemble ML model’s greater precision in
comparison to the individual techniques, owing to its lower error readings. The projecting
accuracy of the models was ascertained statistically through Equations (1) and (2), taken
from previous work [55,74,75].
n
1
MAE =
n ∑ | Pi − Ti| (1)
i =1
s
( Pi − Ti )2
RMSE = ∑ n
(2)

where n = number of data points, Ti = experimental values, and Pi = predicted values.

Table 2. Statistical checks of the ML methods used in the present study.

Model MAE RMSE

Support vector machine 6.720 8.145
Multi-layer perceptron neural network 5.864 7.492
AdaBoost regressor 4.027 5.543
Random forest 2.338 3.394

To evaluate the k-fold analysis results, the R2 , MAE, and RMSE were calculated, and
the resulting values for the SVM, MLPNN, AR, and RF techniques are summarized in
Table 3. To compare the MAE values for all of the models from the k-fold analysis, Figure 16
was generated. The MAE values for the SVM model were in the range of 6.72 to 14.26 MPa,
with an average of 10.53 MPa. The same values for the MLPNN model were between 5.86
and 13.79 MPa, with an average of 9.39 MPa. Additionally, for the AR method, these values
were between 4.03 and 11.94 MPa, with an average of 8.20 MPa. The MAE values for the RF
model were in the range of 2.34 to 11.10 MPa, with an average of 6.90 MPa. This analysis
validated the higher accuracy of ensemble ML models, with the RF model having the
lowest error/deviation from the experimental results. This was further confirmed by the
results of RMSE, as depicted in Figure 17. The average RMSE value for the SVM, MLPNN,
AR, and RF models was 13.29, 11.08, 9.91, and 7.97, respectively. The results of R2 from
the k-fold analysis were compared and are presented in Figure 18. It was determined that
the RF model has higher R2 values with an average of 0.71, compared to the other models,
which yielded an average R2 of 0.42, 0.49, and 0.62 for the SVM, MLPNN, and AR models,
respectively. The RF model with smaller deviations from the experimental results and
higher R2 values outperformed the other models in estimating the C-S of GeoPC. Hence,
this analysis suggests the use of an RF model for this purpose.
Polymers 2022, 14, 2128 14 of 21

Table 3. Results of the k-fold process.

SVM MLPNN AR RF
K-Fold
MAE RMSE R2 MAE RMSE R2 MAE RMSE R2 MAE RMSE R2
1 14.26 18.02 0.67 9.10 12.19 0.25 7.56 11.05 0.22 8.10 9.34 0.91
2 9.09 11.53 0.57 8.81 11.62 0.81 6.23 6.04 0.79 7.48 9.36 0.32
3 13.15 15.48 0.58 6.46 9.53 0.34 11.09 14.76 0.21 2.34 13.46 0.84
4 10.96 17.65 0.78 13.79 19.49 0.26 10.21 10.06 0.89 11.10 3.39 0.39
5 7.56 11.24 0.26 5.86 9.43 0.72 7.46 8.46 0.58 7.49 9.13 0.76
6 6.72 9.16 0.13 7.67 8.40 0.21 4.74 5.54 0.81 4.26 4.81 0.80
7 11.94 13.45 2022, 14,0.60
Polymers x 11.21 7.49 0.63 8.08 8.87 0.50 7.81 10.31 0.60 17 of 24
8 9.94 13.45 2022, 14,
Polymers 0.18
x 12.60 8.92 0.40 11.94 12.23 0.64 8.65 4.11 0.84 17 of 24
9 14.23 15.82 0.32 10.05 15.75 0.58 10.62 15.35 0.85 9.38 12.10 0.95
10 7.43 8.15 0.14 8.34 8.03 0.69 4.03 6.69 0.71 2.44 3.72 0.66

SVM MLPNN AdaBoost RF

16 SVM MLPNN AdaBoost RF
16
14
14
(MPa)

12
(MPa)

12
10
error

10
error

8
absolute

8
absolute

6
6
4
Mean

4
Mean

2
2
0
0 1 2 3 4 5 6 7 8 9 10
1 2 3 4 5 6 7 8 9 10
Fold
Fold
Figure 16. Comparison of mean absolute error for all models from the k-fold analysis.
Figure 16. Comparison of mean absolute error for all models from the k-fold analysis.
Figure 16. Comparison of mean absolute error for all models from the k-fold analysis.

SVM MLPNN AdaBoost AdaBoost

20 SVM MLPNN AdaBoost AdaBoost
20
18
18
(MPa)

16
(MPa)

16
14
error

14
12
error

12
square

10
square

10
8
8
mean

6
mean

6
4
Root

4
Root

2
2
0
0 1 2 3 4 5 6 7 8 9 10
1 2 3 4 5 6 7 8 9 10
Fold
Fold
Figure 17. Comparison of root mean square error for all models from the k-fold analysis.
Figure 17. Comparison of root mean square error for all models from the k-fold analysis.
Figure 17. Comparison of root mean square error for all models from the k-fold analysis.
Polymers 2022, 14, 2128 Polymers 2022, 14, x 15 of 21 18 of 24

SVM MLPNN AdaBoost RF

1.0
0.9
0.8

Correlation coefficient (R2) 0.7

0.6
0.5
0.4
0.3
0.2
0.1
0.0
1 2 3 4 5 6 7 8 9 10
Fold
Figure 18. Comparison of the correlation coefficient for all models from the k-fold analysis.
Figure 18. Comparison of the correlation coefficient for all models from the k-fold analysis.
6. Sensitivity Analysis
6. Sensitivity Analysis
The intent of this evaluation is to find out the impact of input variables on GeoPC’s
The intent of this evaluation is to find out the impact of input variables on GeoPC’s
C-S prediction. The anticipated result is considerably influenced by the input factors [76].
C-S prediction. The anticipated result is considerably influenced by the input factors [76].
Figure 19 illuminates
Figurethe impact of each
19 illuminates inputofvariable
the impact each inputon variable
the C-S onforecast
the C-Sofforecast
GeoPC. of The
GeoPC. The
analysis revealed that curing time, curing temperature, and age of specimen
analysis revealed that curing time, curing temperature, and age of specimen were were thethe most
most important important
constituents that influence the ML model’s performance in estimating
constituents that influence the ML model’s performance in estimating the C-S
the C-S of GeoPC, accounting
of GeoPC, for 22.5%,
accounting 20.1%,20.1%,
for 22.5%, and 18.5%, respectively.
and 18.5%, TheThe
respectively. remaining
remaining input
input variables, variables,
includingincluding
superplasticizer, NaOH molarity,
superplasticizer, water, alkali/fly
NaOH molarity, ash ratio,
water, alkali/fly ash ratio,
Na
Na2 SiO3 /NaOH ratio, and aggregate volume, had a contribution of 12.5%, 9.4%, 4.8%, 4.8%,
2SiO 3/NaOH ratio, and aggregate volume, had a contribution of 12.5%, 9.4%,
4.2%, respectively.
4.2%, 4.1%, and 3.9%, 4.1%, and 3.9%, SArespectively. SA revealedbetween
revealed relationships relationships betweenofthe
the quantity quantity of
input
factors and the data points used to build the ML models. The impact of input parametersof input
input factors and the data points used to build the ML models. The impact
on the ML model’sparameters on the
results was ML model’susing
ascertained results was ascertained
Equations (3) andusing
(4). Equations (3) and (4).
𝑁 = 𝑓 𝑥 − 𝑓 𝑥 (3)
Ni = f max ( xi ) − f min ( xi ) (3)
𝑁
𝑆 = (4)
Ni ∑ 𝑁
Si = n (4)
where 𝑓 𝑥 and 𝑓 ∑ j−the
𝑥 are i Njhighest and lowest of the projected outcome over
the 𝑖 output, respectively. The Si is the attained impact percentage for the specific input
where f max ( xi ) and f min ( xi ) are the highest and lowest of the projected outcome over
parameter.
the ith output, respectively. The Si is the attained impact percentage for the specific
input parameter.

7. Discussions
The objective of this study was to add to the body of knowledge concerning the appli-
cation of contemporary methods for evaluating the C-S of GeoPC. This kind of exploration
will benefit the building industry by facilitating the progress of rapid and cost-efficient
material property prediction tools. By encouraging eco-responsive construction through
these measures, the adoption and usage of GeoPC in the building sector will be hastened.
Since GeoPC might be manufactured from waste constituents, including aluminosilicates,
its usage in the building sector has a variety of benefits, including reduced energy con-
sumption, waste reduction, natural resources protection, reduced CO2 emissions, better
material properties, and green construction materials [27].
Polymers 2022, 14, 2128 Polymers 2022, 14, x 16 of 21 19 of

25.0

Percentage contribution (%)

20.0

15.0

10.0

5.0

0.0

Figure 19. Impact Figure

of input
19.factors
Impacton
ofthe model’s
input factorsprediction.
on the model’s prediction.

This research validates how ML techniques can be utilized to foresee the C-S of
7. Discussions
GeoPC. Four ML methods were employed: two individual (SVM and MLPNN) and two
The objective of this study was to add to the body of knowledge concerning t
ensembled (AR and RF). All ML methods were assessed for precision to determine which
application of contemporary methods for evaluating the C-S of GeoPC. This kind
is the most effective model. The RF model generated a more accurate result with an R2
exploration will benefit the building industry by facilitating the 2 progress of rapid a
of 0.95, compared to the AR, MLPNN, and SVM models, which yielded R of 0.89, 0.81,
cost-efficient material property prediction tools. By encouraging eco-respons
and 0.78, respectively. Furthermore, all models’ performance was confirmed by k-fold
construction through these measures, the adoption and usage of GeoPC in the buildi
and statistical analysis techniques.
sector will TheSince
be hastened. fewerGeoPC
errors might
in thebe model, the morefrom
manufactured precise it is.
waste constituen
However, establishing
includingand aluminosilicates,
suggesting the ideal ML method
its usage in thefor forecasting
building sectoroutcomes across of benef
has a variety
a number of areasincluding
is challenging,
reduced since any model’s
energy performance
consumption, is highly dependent
waste reduction, on the protectio
natural resources
input parametersreduced
and data set utilized to execute the algorithm. Ensembled ML
CO2 emissions, better material properties, and green construction materials [2methods
frequently make use This of the weak learner
research validates by how
building sub-modelscan
ML techniques that
bemay be trained
utilized on the C-S
to foresee
data and tweaked to maximize the R 2 value.
GeoPC. Four ML methods were employed: two individual (SVM and MLPNN) and tw
The dispersal of R2 values
ensembled for the
(AR and RF).AR Alland RF sub-models
ML methods is represented
were assessed in Figure
for precision 20.
to determine wh
The lowest, average, and maximum R 2 values for AR sub-models were 0.811, 0.864, and 0.892,
is the most effective model. The RF model generated a more accurate result with an R2
respectively. The0.95,
lowest, average, andAR, maximum 2 values for RF sub-models were 20.938,
compared to the MLPNN,Rand SVM models, which yielded R of 0.89, 0.81, a
0.947, and 0.952, respectively. These figures
0.78, respectively. Furthermore,indicateallthe superior
models’ exactness ofwas
performance the RF method by
confirmed in k-fold a
comparison to the statistical analysis techniques.
AR in estimating the C-S of GeoPC. The fewer Other errors in the have
researchers model, alsothe more precise it
observed
that the AR and RF However,
models areestablishing
more accurateand in suggesting
predictingthe ideal ML
outcomes method Feng
[68,77,78]. for forecasting
et al. [68] outcom
observed that theacross
AR model a number of areasindividual
outperformed is challenging,
models, since any model’s
including ANN and performance
SVM, in is high
terms of R2 and error
dependent
values. on the input
Likewise, Farooq parameters
et al. [78]and data the
assessed set accuracy
utilized ofto RF
execute the algorith
with that
Ensembled
of the decision tree, ML methods
gene expression frequentlyand
programming, make use of neural
artificial the weak learnermethods
network by buildingandsub-mod
found that the RFthat mayhad
be trained
a greateronprecision
data andthan tweaked to maximize
with antheR2Rofvalue.
2
model the others, 0.96.
In addition, an The dispersal of Rout values for the AR theand RF of
sub-models is represented
variable onin Figure
2
SA was carried to identify effect each input
The lowest, average, and maximum R 2 values for AR sub-models were 0.811, 0.864, a
GPC’s anticipated C-S. The model’s effectiveness may be influenced by the input variables
and the size of the0.892,
datarespectively.
set. The SAThe lowest, average,
established the degree andtomaximum
which each R2 values
of the for
nine RFinputs
sub-models we
0.938, 0.947, and 0.952, respectively. These figures indicate the superior exactness of t
influenced the projected outcome. Curing time, curing temperature, and age of specimen
RF method in comparison to the AR in estimating the C-S of GeoPC. Other research
were found to be the three most highly crucial input factors. However, there are several
have also observed that the AR and RF models are more accurate in predicting outcom
other parameters involved in the manufacture of GeoPC that affect the C-S, such as the
chemical composition of the precursors and the superplasticizer, which may be used as
input parameters in future ML-based modeling to study their impact.
 GPC’s anticipated C-S. The model’s effectiveness may be influenced by the input variables
and the size of the data set. The SA established the degree to which each of the nine inputs
influenced the projected outcome. Curing time, curing temperature, and age of specimen
were found to be the three most highly crucial input factors. However, there are several
other parameters involved in the manufacture of GeoPC that affect the C-S, such as the
Polymers 2022, 14, 2128 chemical composition of the precursors and the superplasticizer, which may 17 of
be21used as
input parameters in future ML-based modeling to study their impact.

AdaBoost Random forest

1.00

0.95

Correlation Coefficient (R2) 0.90

0.85

0.80

0.75

0.70
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Sub-models
Figure 20. The correlation
Figure 20.coefficient of ensemble
The correlation coefficientmachine learning
of ensemble sub-models.
machine learning sub-models.

8. Conclusions 8. Conclusions
The purpose of this
Thestudy wasoftothis
purpose apply both
study ensemble
was to applyand individual
both ensemble machine learning
and individual machine
(ML) methods tolearning
estimate themethods
(ML) compressive strength
to estimate (C-S) of geopolymer
the compressive concrete
strength (C-S) (GeoPC).
of geopolymer concrete
Two individual approaches—support
(GeoPC). Two individualvector machine (SCM)
approaches—support and multi-layer
vector machine (SCM) perceptron
and multi-layer
perceptron neural network
neural network (MLPNN)—were employed (MLPNN)—were employedand
to forecast outcomes, to forecast outcomes,ML
two ensemble and two
approaches wereensemble ML approaches
used, namely, AdaBoost wereregressor
used, namely, AdaBoost
(AR) regressor
and random (AR) and
forest random
(RF). This forest
(RF).
study reached the This study
following reached the following findings:
findings:
1. Ensemble ML methods (AR and RF) outperformed individual ML techniques (SVM
1. Ensemble ML methods (AR and RF) outperformed individual ML techniques (SVM
and MLPNN) in forecasting the C-S of GeoPC, with the RF model performing with the
and MLPNN) in forecasting the C-S of GeoPC, with the2RF model performing with
highest accuracy. The correlation coefficients (R ) were 0.95, 0.89, 0.81, and 0.78 for RF,
the highest accuracy. The correlation
AR, MLPNN, coefficients
and SVM models, (R2 ) were 0.95, 0.89, 0.81, and 0.78
respectively.
for RF, AR,2.MLPNN, and SVMofmodels,
The comparison respectively.
experimental and anticipated results verified the AR and RF
2. The comparison of experimental
models’ andasanticipated
superior accuracy, the projectedresults verifiedless
values deviated thefrom
ARtheand RF
experimental
models’ superior accuracy,
values. On theasother
the projected values deviated
hand, the MLPNN and SVMless from
model the experimental
results deviated more from
values. On thethe experimental
other hand, theresults,
MLPNN making
andthem
SVMless suitable
model for predicting
results deviatedthe moreC-S from
of GeoPC.
the experimental results, making them less suitable for predicting the C-S of GeoPC.
3. Statistical analysis and k-fold evaluation were used to validate the model performance.
These evaluations validated the RF model’s superior accuracy. The ensembled models’
decreased deviation (MAE and RMSE) and higher R2 values supported their increased
accuracy over individual models.
4. Sensitivity analysis discovered that curing time, curing temperature, and specimen
age were the most significant elements influencing the ML model’s performance in
predicting GeoPC’s C-S, accounting for 22.5%, 20.1%, and 18.5%, respectively. The
other input variables, including superplasticizer, NaOH molarity, water, alkali/fly ash
ratio, Na2 SiO3 /NaOH ratio, and aggregate volume, contributed 12.5%, 9.4%, 4.8%,
4.2%, 4.1%, and 3.9%, respectively.
5. This kind of study will benefit the construction industry by allowing for the progress
of rapid and cost-efficient strategies for estimating the strength of materials. Moreover,
by applying these methods to encourage eco-responsive construction, the acceptance
and usage of GeoPC in the building sector will be enhanced.
Polymers 2022, 14, 2128 18 of 21

Supplementary Materials: The following supporting information can be downloaded at: https:
//www.mdpi.com/article/10.3390/polym14102128/s1, Table S1: Data used for modeling.
Author Contributions: Conceptualization, M.N.A. and K.K.; data curation, W.A.; Formal Analysis,
W.A. and M.F.J.; Funding acquisition, M.N.A., K.K., M.G.Q. and M.I.F.; Investigation, W.A. and M.F.J.;
Methodology, W.A., K.K. and H.J.Q.; Project administration, M.N.A.; Resources, M.N.A., K.K., H.J.Q.
and M.G.Q.; Supervision, M.N.A.; Validation, M.F.J. and M.U.S.; Visualization, K.K., H.J.Q., M.U.S.,
M.G.Q. and M.I.F.; Writing—original draft, M.N.A., W.A. and M.F.J.; Writing—review and editing,
M.N.A. All authors have read and agreed to the published version of the manuscript.
Funding: This research was funded by the Deanship of Scientific Research, Vice Presidency for Grad-
uate Studies and Scientific Research, King Faisal University, Saudi Arabia (Project No. AN000500).
The APC was funded by the same project.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: The data used in this research have been properly cited and reported
in the main text.
Acknowledgments: The authors acknowledge the Deanship of Scientific Research, Vice Presidency
for Graduate Studies and Scientific Research, King Faisal University, Saudi Arabia (Project No.
AN000500). The authors extend their appreciation to the financial support that has made this
study possible.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Hawa, A.; Tonnayopas, D.; Prachasaree, W.; Taneerananon, P. Development and Performance Evaluation of Very High Early
Strength Geopolymer for Rapid Road Repair. Adv. Mater. Sci. Eng. 2013, 2013, 1–9. [CrossRef]
2. Tayeh, B.A.; Alsaffar, D.; Alyousef, R. The Utilization of Recycled Aggregate in High Performance Concrete: A Review. J. Mater.
Res. Technol. 2020, 9, 8469–8481. [CrossRef]
3. Ahmad, W.; Ahmad, A.; Ostrowski, K.A.; Aslam, F.; Joyklad, P.; Zajdel, P. Sustainable approach of using sugarcane bagasse ash in
cement-based composites: A systematic review. Case Stud. Constr. Mater. 2021, 15, e00698. [CrossRef]
4. Xie, C.; Cao, M.; Si, W.; Khan, M. Experimental evaluation on fiber distribution characteristics and mechanical properties of
calcium carbonate whisker modified hybrid fibers reinforced cementitious composites. Constr. Build. Mater. 2020, 265, 120292.
[CrossRef]
5. Khan, M.; Cao, M.; Xie, C.; Ali, M. Efficiency of basalt fiber length and content on mechanical and microstructural properties of
hybrid fiber concrete. Fatigue Fract. Eng. Mater. Struct. 2021, 44, 2135–2152. [CrossRef]
6. Khan, M.; Cao, M.; Ali, M. Cracking behaviour and constitutive modelling of hybrid fibre reinforced concrete. J. Build. Eng. 2020,
30, 101272. [CrossRef]
7. Ahmad, W.; Farooq, S.H.; Usman, M.; Khan, M.; Ahmad, A.; Aslam, F.; Al Yousef, R.; Al Abduljabbar, H.; Sufian, M. Effect of
Coconut Fiber Length and Content on Properties of High Strength Concrete. Materials 2020, 13, 1075. [CrossRef]
8. Khan, M.; Ali, M. Improvement in concrete behavior with fly ash, silica-fume and coconut fibres. Constr. Build. Mater. 2019, 203,
174–187. [CrossRef]
9. Khan, M.; Cao, M.; Hussain, A.; Chu, S. Effect of silica-fume content on performance of CaCO3 whisker and basalt fiber at matrix
interface in cement-based composites. Constr. Build. Mater. 2021, 300, 124046. [CrossRef]
10. Khan, M.; Rehman, A.; Ali, M. Efficiency of silica-fume content in plain and natural fiber reinforced concrete for concrete road.
Constr. Build. Mater. 2020, 244, 118382. [CrossRef]
11. Almutairi, A.L.; Tayeh, B.A.; Adesina, A.; Isleem, H.F.; Zeyad, A.M. Potential applications of geopolymer concrete in construction:
A review. Case Stud. Constr. Mater. 2021, 15, e00733. [CrossRef]
12. Amran, Y.M.; Alyousef, R.; Alabduljabbar, H.; El-Zeadani, M. Clean production and properties of geopolymer concrete; A review.
J. Clean. Prod. 2020, 251, 119679. [CrossRef]
13. Hamada, H.; Tayeh, B.; Yahaya, F.; Muthusamy, K.; Al-Attar, A. Effects of nano-palm oil fuel ash and nano-eggshell powder on
concrete. Constr. Build. Mater. 2020, 261, 119790. [CrossRef]
14. Arafa, M.; Tayeh, B.A.; Alqedra, M.; Shihada, S.; Hanoona, H. Investigating the effect of sulfate attack on compressive strength of
recycled aggregate concrete. J. Eng. Res. Technol. 2017, 4, 4.
15. Hamada, H.M.; Thomas, B.S.; Tayeh, B.; Yahaya, F.M.; Muthusamy, K.; Yang, J. Use of oil palm shell as an aggregate in cement
concrete: A review. Constr. Build. Mater. 2020, 265, 120357. [CrossRef]
16. Tayeh, B.A.; Hasaniyah, M.W.; Zeyad, A.M.; Awad, M.M.; Alaskar, A.; Mohamed, A.M.; Alyousef, R. Durability and mechanical
properties of seashell partially-replaced cement. J. Build. Eng. 2020, 31, 101328. [CrossRef]
Polymers 2022, 14, 2128 19 of 21

17. Li, X.; Qin, D.; Hu, Y.; Ahmad, W.; Ahmad, A.; Aslam, F.; Joyklad, P. A systematic review of waste materials in cement-based
composites for construction applications. J. Build. Eng. 2021, 45, 103447. [CrossRef]
18. Hassan, A.; Arif, M.; Shariq, M. A review of properties and behaviour of reinforced geopolymer concrete structural elements- A
clean technology option for sustainable development. J. Clean. Prod. 2020, 245, 11876. [CrossRef]
19. Ahmad, W.; Ahmad, A.; Ostrowski, K.A.; Aslam, F.; Joyklad, P. A scientometric review of waste material utilization in concrete
for sustainable construction. Case Stud. Constr. Mater. 2021, 15, e00683. [CrossRef]
20. Hamada, H.M.; Alya’a, A.; Yahaya, F.M.; Muthusamy, K.; Tayeh, B.A.; Humada, A.M. Effect of high-volume ultrafine palm oil
fuel ash on the engineering and transport properties of concrete. Case Stud. Constr. Mater. 2020, 12, e00318. [CrossRef]
21. Farooq, F.; Jin, X.; Javed, M.F.; Akbar, A.; Shah, M.I.; Aslam, F.; Alyousef, R. Geopolymer concrete as sustainable material: A state
of the art review. Constr. Build. Mater. 2021, 306, 124762. [CrossRef]
22. Das, S.K.; Mishra, J.; Singh, S.K.; Mustakim, S.M.; Patel, A.; Das, S.K.; Behera, U. Characterization and utilization of rice husk ash
(RHA) in fly ash—Blast furnace slag based geopolymer concrete for sustainable future. Mater. Today Proc. 2020, 33, 5162–5167.
[CrossRef]
23. Cleetus, A.; Shibu, R.; Sreehari, P.M.; Paul, V.K.; Jacob, B. Analysis and Study of the Effect of Ggbfs on Concrete Structures. Int.
Res. J. Eng. Technol. (IRJET) 2018, 5, 3033–3037.
24. Meesala, C.R.; Verma, N.K.; Kumar, S. Critical review on fly-ash based geopolymer concrete. Struct. Concr. 2020, 21, 1013–1028.
[CrossRef]
25. Huseien, G.F.; Shah, K.W.; Sam, A.R.M. Sustainability of nanomaterials based self-healing concrete: An all-inclusive insight. J.
Build. Eng. 2019, 23, 155–171. [CrossRef]
26. Davidovits, J.; Cordi, S.A. Synthesis of new high temperature geo-polymers for reinforced plastics/composites. Spe Pactec. 1979,
79, 151–154.
27. Yang, H.; Liu, L.; Yang, W.; Liu, H.; Ahmad, W.; Ahmad, A.; Aslam, F.; Joyklad, P. A comprehensive overview of geopolymer
composites: A bibliometric analysis and literature review. Case Stud. Constr. Mater. 2022, 16, e00830. [CrossRef]
28. Nguyen, K.; Ahn, N.; Le, T.A.; Lee, K. Theoretical and experimental study on mechanical properties and flexural strength of fly
ash-geopolymer concrete. Constr. Build. Mater. 2016, 106, 65–77. [CrossRef]
29. Nguyen, K.T.; Le, T.A.; Lee, J.; Lee, D.; Lee, K. Investigation on properties of geopolymer mortar using preheated materials and
thermogenetic admixtures. Constr. Build. Mater. 2017, 130, 146–155. [CrossRef]
30. Nazari, A.; Bagheri, A.; Riahi, S. Properties of geopolymer with seeded fly ash and rice husk bark ash. Mater. Sci. Eng. A 2011,
528, 7395–7401. [CrossRef]
31. Nguyen, K.; Lee, Y.H.; Lee, J.; Ahn, N. Acid Resistance and Curing Properties for Green Fly Ash-geopolymer Concrete. J. Asian
Arch. Build. Eng. 2013, 12, 317–322. [CrossRef]
32. Saxena, S.; Kumar, M.; Singh, N. Fire Resistant Properties of Alumino Silicate Geopolymer cement Mortars. Mater. Today Proc.
2017, 4, 5605–5612. [CrossRef]
33. Yadollahi, M.M.; Benli, A.; Demirboğa, R. Prediction of compressive strength of geopolymer composites using an artificial neural
network. Mater. Res. Innov. 2015, 19, 453–458. [CrossRef]
34. Tayeh, B.A.; Zeyad, A.M.; Agwa, I.S.; Amin, M. Effect of elevated temperatures on mechanical properties of lightweight
geo-polymer concrete. Case Stud. Constr. Mater. 2021, 15, e00673.
35. Zhang, X.; Bai, C.; Qiao, Y.; Wang, X.; Jia, D.; Li, H.; Colombo, P. Porous geopolymer composites: A review. Compos. Part A Appl.
Sci. Manuf. 2021, 150, 106629. [CrossRef]
36. Alhawat, M.; Ashour, A.; Yildirim, G.; Aldemir, A.; Sahmaran, M. Properties of geopolymers sourced from construction and
demolition waste: A review. J. Build. Eng. 2022, 50, 104104. [CrossRef]
37. D’Angelo, G.; Fumo, M.; Merino, M.; Capasso, I.; Campanile, A.; Iucolano, F.; Caputo, D.; Liguori, B. Crushed Bricks: Demolition
Waste as a Sustainable Raw Material for Geopolymers. Sustainability 2021, 13, 7572. [CrossRef]
38. Yang, K.-H.; Song, J.-K.; Song, K.-I. Assessment of CO2 reduction of alkali-activated concrete. J. Clean. Prod. 2013, 39, 265–272.
[CrossRef]
39. Ren, B.; Zhao, Y.; Bai, H.; Kang, S.; Zhang, T.; Song, S. Eco-friendly geopolymer prepared from solid wastes: A critical review.
Chemosphere 2020, 267, 128900. [CrossRef]
40. Mohajerani, A.; Suter, D.; Jeffrey-Bailey, T.; Song, T.; Arulrajah, A.; Horpibulsuk, S.; Law, D. Recycling waste materials in
geo-polymer concrete. Clean Technol. Environ. Policy 2019, 21, 493–515. [CrossRef]
41. Awoyera, P.O. Nonlinear finite element analysis of steel fibre-reinforced concrete beam under static loading. J. Eng. Sci. Technol.
2016, 11, 1669–1677.
42. Sadrmomtazi, A.; Sobhani, J.; Mirgozar, M. Modeling compressive strength of EPS lightweight concrete using regression, neural
network and ANFIS. Constr. Build. Mater. 2013, 42, 205–216. [CrossRef]
43. Topçu, I.B.; Sarıdemir, M. Prediction of compressive strength of concrete containing fly ash using artificial neural networks and
fuzzy logic. Comput. Mater. Sci. 2008, 41, 305–311. [CrossRef]
44. Öztaş, A.; Pala, M.; Özbay, E.; Kanca, E.; Çaglar, N.; Bhatti, M.A. Predicting the compressive strength and slump of high strength
concrete using neural network. Constr. Build. Mater. 2006, 20, 769–775. [CrossRef]
45. Sarıdemir, M. Predicting the compressive strength of mortars containing metakaolin by artificial neural networks and fuzzy logic.
Adv. Eng. Softw. 2009, 40, 920–927. [CrossRef]
Polymers 2022, 14, 2128 20 of 21

46. Ni, H.-G.; Wang, J.-Z. Prediction of compressive strength of concrete by neural networks. Cem. Concr. Res. 2000, 30, 1245–1250.
[CrossRef]
47. Sobhani, J.; Najimi, M.; Pourkhorshidi, A.R.; Parhizkar, T. Prediction of the compressive strength of no-slump concrete: A
comparative study of regression, neural network and ANFIS models. Constr. Build. Mater. 2010, 24, 709–718. [CrossRef]
48. Shafabakhsh, G.; Ani, O.J.; Talebsafa, M. Artificial neural network modeling (ANN) for predicting rutting performance of
nano-modified hot-mix asphalt mixtures containing steel slag aggregates. Constr. Build. Mater. 2015, 85, 136–143. [CrossRef]
49. Hodhod, O.; Ahmed, H. Modeling the corrosion initiation time of slag concrete using the artificial neural network. HBRC J. 2014,
10, 231–234. [CrossRef]
50. Carmichael, R.P. Elationships between Young’s Modulus, Compressive Strength, Poisson’s Ratio, and Time for Early Age Concrete;
Swarthmore College: Pensylvania, PA, USA, 2009.
51. Bal, L.; Buyle-Bodin, F. Artificial neural network for predicting drying shrinkage of concrete. Constr. Build. Mater. 2013, 38,
248–254. [CrossRef]
52. Ben Chaabene, W.; Flah, M.; Nehdi, M.L. Machine learning prediction of mechanical properties of concrete: Critical review. Constr.
Build. Mater. 2020, 260, 119889. [CrossRef]
53. Sufian, M.; Ullah, S.; Ostrowski, K.; Ahmad, A.; Zia, A.; Śliwa-Wieczorek, K.; Siddiq, M.; Awan, A. An Experimental and Empirical
Study on the Use of Waste Marble Powder in Construction Material. Materials 2021, 14, 3829. [CrossRef] [PubMed]
54. Ahmad, A.; Farooq, F.; Niewiadomski, P.; Ostrowski, K.; Akbar, A.; Aslam, F.; Alyousef, R. Prediction of Compressive Strength of
Fly Ash Based Concrete Using Individual and Ensemble Algorithm. Materials 2021, 14, 794. [CrossRef] [PubMed]
55. Nguyen, K.; Nguyen, Q.D.; Le, T.A.; Shin, J.; Lee, K. Analyzing the compressive strength of green fly ash based geopolymer
concrete using experiment and machine learning approaches. Constr. Build. Mater. 2020, 247, 118581. [CrossRef]
56. Deepa, C.; Sathiyakumari, K.; Sudha, V. Prediction of the Compressive Strength of High Performance Concrete Mix using Tree
Based Modeling. Int. J. Comput. Appl. 2010, 6, 18–24. [CrossRef]
57. Song, Y.-Y.; Ying, L.U. Decision tree methods: Applications for classification and prediction. Shanghai Arch. Psychiatry 2015, 27,
130.
58. Hillebrand, E.; Medeiros, M. The Benefits of Bagging for Forecast Models of Realized Volatility. Econ. Rev. 2010, 29, 571–593.
[CrossRef]
59. Ahmad, A.; Ahmad, W.; Aslam, F.; Joyklad, P. Compressive strength prediction of fly ash-based geopolymer concrete via
ad-vanced machine learning techniques. Case Stud. Constr. Mater. 2022, 16, e00840.
60. Ling, H.; Qian, C.X.; Kang, W.C.; Liang, C.Y.; Chen, H.C. Machine and K-Fold cross validation to predict compressive strength of
concrete in marine environment. Constr. Build. Mater. 2019, 206, 355–363. [CrossRef]
61. Bui, K.-T.T.; Tien Bui, D.; Zou, J.; Van Doan, C.; Revhaug, I. A novel hybrid artificial intelligent approach based on neural fuzzy
inference model and particle swarm optimization for horizontal displacement modeling of hydropower dam. Neural Comput.
Appl. 2018, 29, 1495–1506. [CrossRef]
62. Bui, D.T.; Bui, K.T.T.; Bui, Q.T.; Van Doan, C.; Hoang, N.D. Hybrid Intelligent Model Based on Least Squares Support Vector Regression
and Artificial Bee Colony Optimization for Time-Series Modeling and Forecasting Horizontal Displacement of Hydro-Power Dam. Handbook
of Neural Computation; Elsevier: Amsterdam, The Netherlands, 2017; pp. 279–293.
63. Bui, D.T.; Tuan, T.A.; Klempe, H.; Pradhan, B.; Revhaug, I. Spatial prediction models for shallow landslide hazards: A comparative
assessment of the efficacy of support vector machines, artificial neural networks, kernel logistic regression, and logistic model
tree. Landslides 2016, 13, 361–378. [CrossRef]
64. Pham, B.T.; Tien Bui, D.; Prakash, I.; Dholakia, M.B. Hybrid integration of Multilayer Perceptron Neural Networks and machine
learning ensembles for landslide susceptibility assessment at Himalayan area (India) using GIS. Catena 2017, 149, 52–63. [CrossRef]
65. Sadowski, J.; Hoła, J.; Czarnecki, S.; Wang, D. Pull-off adhesion prediction of variable thick overlay to the substrate. Autom.
Constr. 2018, 85, 10–23. [CrossRef]
66. Kavzoglu, T.; Mather, P.M. The use of backpropagating artificial neural networks in land cover classification. Int. J. Remote Sens.
2003, 24, 4907–4938. [CrossRef]
67. Pham, B.T.; Nguyen, M.D.; Bui, K.T.T.; Prakash, I.; Chapi, K.; Bui, D.T. A novel artificial intelligence approach based on Multilayer
Perceptron Neural Network and Biogeography-based Optimization for predicting coefficient of consolidation of soil. CATENA
2019, 173, 302–311. [CrossRef]
68. Feng, D.-C.; Liu, Z.-T.; Wang, X.-D.; Chen, Y.; Chang, J.-Q.; Wei, D.-F.; Jiang, Z.-M. Machine learning-based compressive strength
prediction for concrete: An adaptive boosting approach. Constr. Build. Mater. 2019, 230, 117000. [CrossRef]
69. Wang, Q.; Ahmad, W.; Ahmad, A.; Aslam, F.; Mohamed, A.; Vatin, N.I. Application of Soft Computing Techniques to Predict the
Strength of Geopolymer Composites. Polymers 2020, 14, 1074. [CrossRef]
70. Han, Q.; Gui, C.; Xu, J.; Lacidogna, G. A generalized method to predict the compressive strength of high-performance concrete by
improved random forest algorithm. Constr. Build. Mater. 2019, 226, 734–742. [CrossRef]
71. Grömping, U. Variable Importance Assessment in Regression: Linear Regression versus Random Forest. Am. Stat. 2009, 63,
308–319. [CrossRef]
72. Xu, Y.; Ahmad, W.; Ahmad, A.; Ostrowski, K.A.; Dudek, M.; Aslam, F.; Joyklad, P. Computation of High-Performance Concrete
Compressive Strength Using Standalone and Ensembled Machine Learning Techniques. Materials 2021, 14, 7034. [CrossRef]
Polymers 2022, 14, 2128 21 of 21

73. Ahmad, A.; Chaiyasarn, K.; Farooq, F.; Ahmad, W.; Suparp, S.; Aslam, F. Compressive Strength Prediction via Gene Expression
Programming (GEP) and Artificial Neural Network (ANN) for Concrete Containing RCA. Buildings 2021, 11, 324. [CrossRef]
74. Naseri, H.; Jahanbakhsh, H.; Hosseini, P.; Nejad, F.M. Designing sustainable concrete mixture by developing a new machine
learning technique. J. Clean. Prod. 2020, 258, 120578. [CrossRef]
75. Prayogo, D.; Cheng, M.-Y.; Wu, Y.-W.; Tran, D.-H. Combining machine learning models via adaptive ensemble weighting for
prediction of shear capacity of reinforced-concrete deep beams. Eng. Comput. 2019, 36, 1135–1153. [CrossRef]
76. Ahmad, A.; Ostrowski, K.; Maślak, M.; Farooq, F.; Mehmood, I.; Nafees, A. Comparative Study of Supervised Machine Learning
Algorithms for Predicting the Compressive Strength of Concrete at High Temperature. Materials 2021, 14, 4222. [CrossRef]
77. Feng, D.-C.; Liu, Z.-T.; Wang, X.-D.; Jiang, Z.-M.; Liang, S.-X. Failure mode classification and bearing capacity prediction for
reinforced concrete columns based on ensemble machine learning algorithm. Adv. Eng. Inform. 2020, 45, 101126. [CrossRef]
78. Farooq, F.; Amin, M.N.; Khan, K.; Sadiq, M.R.; Javed, M.F.F.; Aslam, F.; Alyousef, R. A Comparative Study of Random Forest and
Genetic Engineering Programming for the Prediction of Compressive Strength of High Strength Concrete (HSC). Appl. Sci. 2020,
10, 7330. [CrossRef]