materials
Article
Compressive Strength of Steel Fiber-Reinforced Concrete
Employing Supervised Machine Learning Techniques
Yongjian Li 1 , Qizhi Zhang 2, *, Paweł Kamiński 3, * , Ahmed Farouk Deifalla 4 , Muhammad Sufian 5, * ,
Artur Dyczko 6 , Nabil Ben Kahla 7,8 and Miniar Atig 8,9
Citation: Li, Y.; Zhang, Q.; Kamiński,
P.; Deifalla, A.F.; Sufian, M.; Dyczko,
A.; Kahla, N.B.; Atig, M. Compressive
Strength of Steel Fiber-Reinforced
Concrete Employing Supervised
Machine Learning Techniques.
Materials 2022, 15, 4209. https://
doi.org/10.3390/ma15124209
Academic Editors: Dario De
Domenico and Luís Filipe Almeida
Bernardo
Received: 4 May 2022
Accepted: 4 June 2022
Published: 14 June 2022
Publisher’s Note: MDPI stays neutral
with regard to jurisdictional claims in
published maps and institutional affil-
iations.
Copyright: © 2022 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
Materials 2022, 15, 4209. https://doi.org/10.3390/ma15124209
1 James Watt Engineering School, University of Glasgow, Scotland, UK; yongjian4896@outlook.com
2 School of Architectural Engineering, Huanghuai University, Zhumadian 463000, China
3 Faculty of Civil Engineering and Resource Management, AGH University of Science and Technology,
Mickiewicza 30, 30-059 Kraków, Poland
4 Structural Engineering and Construction Management Department, Faculty of Engineering and Technology,
Future University in Egypt, Cairo 11835, Egypt; ahmed.deifalla@fue.edu.eg
5 School of Civil Engineering, Southeast University, Nanjing 210096, China
6 Mineral and Energy Economy Research Institute of the Polish Academy of Sciences, J. Wybickiego 7a,
31-261 Kraków, Poland; arturdyczko@min-pan.krakow.pl
7 Department of Civil Engineering, College of Engineering, King Khalid University, Abha 61421, Saudi Arabia;
nbohlal@kku.edu.sa
8 Laboratory of Systems and Applied Mechanics, Tunisia Polytechnic School, University of Carthage, La Marsa,
Tunis 2078, Tunisia; miniar.atig@gmail.com
9 Department of Civil Engineering, The Higher National Engineering School of Tunis, University of Tunis,
Tunis, Tunisia
* Correspondence: zdjs_zqz@163.com (Q.Z.); pkamin@agh.edu.pl (P.K.); drsufian@seu.edu.cn (M.S.)
Abstract: Recently, research has centered on developing new approaches, such as supervised machine
learning techniques, that can compute the mechanical characteristics of materials without investing
much effort, time, or money in experimentation. To predict the 28-day compressive strength of steel
fiber–reinforced concrete (SFRC), machine learning techniques, i.e., individual and ensemble models,
were considered. For this study, two ensemble approaches (SVR AdaBoost and SVR bagging) and one
individual technique (support vector regression (SVR)) were used. Coefficient of determination (R2 ),
statistical assessment, and k-fold cross validation were carried out to scrutinize the efficiency of each
approach used. In addition, a sensitivity technique was used to assess the influence of parameters
on the prediction results. It was discovered that all of the approaches used performed better in
terms of forecasting the outcomes. The SVR AdaBoost method was the most precise, with R2 = 0.96,
as opposed to SVR bagging and support vector regression, which had R2 values of 0.87 and 0.81,
respectively. Furthermore, based on the lowered error values (MAE = 4.4 MPa, RMSE = 8 MPa),
statistical and k-fold cross validation tests verified the optimum performance of SVR AdaBoost. The
forecast performance of the SVR bagging models, on the other hand, was equally satisfactory. In order
to predict the mechanical characteristics of other construction materials, these ensemble machine
learning approaches can be applied.
Keywords: concrete; steel fiber; steel fiber–reinforced concrete; compressive strength; mechanical
characteristics; construction materials
1. Introduction
The bridging effect of discontinuous fibers in fiber-reinforced concrete (FRC) can
enhance its strength characteristics. Consequently, incorporation of steel fiber (SF) improves
concrete’s compressive strength as well as its toughness and fracture resistance. Concrete’s
durability is improved by adding the proper quantity of steel fiber (0–1.5%) [1]. Different
types of fibers, i.e., both natural and artificial, are utilized for improving the mechanical
characteristics and crack resistance behavior of concrete and cementitious materials [2–9].
https://www.mdpi.com/journal/materials
Materials 2022, 15, 4209 2 of 18

Several researchers have carried out studies and presented models for the mechanical
characteristics of regular concrete based on a variety of data; however, SFRC contains more
factors that require prediction than normal concrete, i.e., the type of fiber, percentage of
volume, and the aspect ratio, and creating suitable predictive models is still relatively new.
As a consequence, the typical linear and nonlinear regression models struggle to evaluate
the compressive strength of SFRC. The difficulty in forecasting the strength of SFRC can be
solved by using machine learning approaches [10–19].
Machine learning techniques are broadly utilized in the computer science and artificial
intelligence domains, and their influence on engineering is undeniable. Machine learning
techniques have gained attention in civil engineering, mainly for their prediction of the
mechanical properties of concrete. These practices can be used to predict outcomes with
a high degree of precision. Many studies have been undertaken using machine learning
approaches to forecast the strength properties of concrete and its structural elements as
well [10–23]. Two machine learning approaches, the individual and ensemble approaches,
were employed by Ahmad et al. to forecast concrete compressive strength [24]. Su et al.
used different machine learning algorithms to estimate the strength of the link between
concrete and fiber-reinforced polymers [25]. Nguyen et al. employed a machine learning
approach to assess the compressive characteristics of geo-polymer concrete [26]. Never-
theless, further advances in this field are required to understand and implement these
techniques in civil engineering through the use of other machine learning methodologies.
The collection of data for the development of models that account for numerous factors
such as type of fiber, volume%, and aspect ratio (i.e., the length and diameter of the fiber) is
not easy, and there are currently no published studies that define an efficient method for
intensity prediction in SFRC. As a result, a machine learning model that predicts the com-
pressive strength of SFRC is developed in this study, and the most appropriate techniques
are addressed via a comparison analysis.
There has been a detailed evaluation of the use of ML approaches to predict concrete’s
mechanical characteristics [27]. In addition, a number of studies have been conducted to pre-
dict the mechanical characteristics of different types of concrete, such as high-performance
concrete (HPC) [28], self-healing concrete [29], recycled aggregate concrete (RCA) [30],
phase change materials–integrated concrete [31], etc. Steel fiber–reinforced concrete is
widely used in civil engineering applications. Therefore, this study considered steel fiber–
reinforced concrete (SFRC) with compressive strengths of 26 MPa to 99 MPa. Additionally,
the importance of the raw materials was not considered in the datasets on compressive
strength in previous studies, and this factor requires evaluation. Therefore, the effects
of the input parameters (raw materials) on the compressive strength are studied using
sensitivity analysis. However, casting specimens in the lab and curing and evaluating
them is a time-consuming and labor-intensive process. Employing innovative approaches
such as machine learning techniques to evaluate the mechanical properties of SFRC can
address such problems while also reducing experimentation expenses. To estimate the
28-day compressive strength of SFRC, both individual (SVM) and EML (AdaBoost and
bagging) approaches were used. The coefficient correlation (R2 ) value was used to validate
the quality of each model. To compare the outcomes of each method, the statistical error,
namely the mean-absolute error (MSE), root-mean-square error (RMSE), and k-fold cross
validation tests were performed. Sensitivity analysis was used to examine the contribution
of the model parameters to the outcome predictions. Scholars working in the civil engineer-
ing field could benefit from this study since it allows them to predict strength attributes
without having to spend time in the lab.

2. Data Description
The information was gathered from 17 publications [32–48] as shown in Table S1. Since
the goal of this study is to develop a basic machine learning model for SFRC and forecast
its mechanical parameters, this dataset was constructed entirely using data from hook-end
steel fiber–concrete. Despite the fact that the dataset had a large number of variables, only
Materials 2022, 15, 4209 3 of 18

those that were shown to be fundamentally influenced were chosen and preprocessed. As
a result, the dataset has a total of t10 features, including input data and output data. The
following 10 factors were considered for the forecasting of SFRC compressive strength,
since they are the most fundamental mechanical characteristics of SFRC, and each of these
variables impact the compressive strength of SFRC.
Water and Cement: The ratio of water to cement (W/C) has a big impact on concrete
strength. According to Abbass et al., when the W/C ratio rises, the compressive strength
drops [44]. Similarly, Reddy et al. performed tests on self-consolidating concrete based on
the W/C ratio and concluded that it had a significant impact [49]. Nili et al. found that the
water–cement ratio was a factor influencing SFRC compressive strength, and it was chosen
as a variable [50].
Sand and Aggregate: The impact of the sand and aggregate ratio (s/a) on the strength
properties of SFRC has been identified as a critical element. Kim et al. observed that
increasing the s/a ratio caused an increase in SFRC’s compressive strength [51]. Chitlange
et al. found a significant variation in compressive strength based on the s/a ratio [52]. As a
result, the sand and aggregate content was chosen as a key variable in the development of
the ML models.
Superplasticizer: A superplasticizer is an additive commonly used in the manufacture
of high-strength concrete. Khan et al. reported that the combination of superplasticizer
and pozzolanic ingredients can enhance concrete’s mechanical qualities [53]. Furthermore,
a superplasticizer is essential because concrete requires a larger water content, which
causes bleeding. Aruntas et al. found that when the superplasticizer concentration rises
to 1.5%, the workability and compressive strength increase [54]. As a result, superplas-
ticizer was chosen as a feature in the ML models to quantify its direct effect on SFRC’s
compressive strength.
Silica fume: Many earlier researchers have shown the significant impact of silica fume
on concrete strength. According to Köksal et al., increasing the silica fume concentration
improves compressive strength [42]. Nili et al. found an enhancement in concrete compres-
sive strength with increased content of silica fume [55]. As a result, the concentration of
silica fume was discovered as an impact factor on the strength properties of SFRC and was
chosen as a characteristic.
Fly ash: It is a popular ingredient that can increase concrete performance in both states
i.e., fresh (workability) and hardened (strength), over time. According to R. Saravana
and A. Sumathi, incorporating fly ash into SFRC enhanced the compressive strength over
time [56]. The impact of fly ash and steel fibers on the strength qualities of pozzolana cement
concrete was explored by Muntadher, A.C. and Srivastava, V. [57]. Ashish, K.S. inspected
the effect of fly ash on concrete durability and found that the concrete’s compressive
strength was lower than predicted at first but steadily increased with time [58]. Using a
machine learning method, Mohammad, M.R. et al. evaluated the mechanical characteristics
of fly ash concrete [59]. Fly ash was chosen as a criterion because of its importance in
concrete characteristics.
Steel Fiber volume, length, and diameter: In prior investigations it was found that steel
fiber volume, length, and diameter had a significant influence on the concrete’s compressive
strength. Yazc et al. found that the fiber volume fraction increases compressive strength
with lower aspect ratios, while there was no significant development found in compressive
strength with higher aspect ratios [1]. Furthermore, the experimental investigation of
Köksal et al. revealed that the strength properties of SFRC improve as a consequence of
utilizing fiber volume fractions of up to 1% [42]. As a result, fiber volume, length, and
diameter must be considered as variables in the ML models.
To obtain the intended result, ML approaches require a range of input variables [60].
The information utilized to estimate SFRC’s compressive strength is obtained from the
literature. Only the findings for 28-day compressive strength were separated from the data
collected for further study. The models contained cement, water, sand, coarse aggregate,
superplasticizer, silica fume, fly ash, steel fiber, fiber length, and fiber diameter as inputs,
 Materials 2022, 15, x FOR PEER REVIEW 4 of 18

Materials 2022, 15, 4209 4 of 18

data collected for further study. The models contained cement, water, sand, coarse aggre-
gate, superplasticizer, silica fume, fly ash, steel fiber, fiber length, and fiber diameter as
inputs, with just one variable, i.e., compressive strength, as a resultant output. This study
withused
just166
one
datavariable,
points fori.e., compressive
the 28-day strength,
compressive asprediction
strength a resultant output.
of SFRC This study used
(mix propor-
166 tions).
data points for theincludes
The database 28-daynormal
compressive
strength strength prediction
and high-strength SFRC ofwith
SFRC (mix proportions).
compressive
Thestrengths
databaseranging
includesfromnormal
26 MPa strength andTable
to 99 MPa. high-strength SFRC
1 displays the with compressive
statistical analysis out- strengths
comesfrom
ranging of the26
input
MPa variables, i.e., mean,
to 99 MPa. standard
Table error, median,
1 displays mode, standard
the statistical devia-
analysis outcomes of
the tion,
input range, minimum,
variables, i.e.,and maximum
mean, values.
standard Furthermore,
error, median,the relative
mode, frequencydeviation,
standard pattern range,
distribution of all input parameters is depicted in Figure 1.
minimum, and maximum values. Furthermore, the relative frequency pattern distribution
of all input
Table parameters
1. Statistical analysisisofdepicted in Figure
input and output 1.
variables.

Sand Coarse Ag- Silica Fly Steel Fiber Fiber Compressive

Cement WaterTable 1.3 Statistical analysis of input and output variables.
Superplas-
(kg/m gregate Fume Ash Fiber Length Dia Strength MPa (28
(kg/m3) (kg/m3) ticizer (%)
) (kg/m3) % % (%) (mm) (mm) Days)
Coarse Silica Steel Fiber Fiber Compressive
Cement Water Sand Superplasticizer Fly
Aggregate Fume Fiber Length Dia Strength MPa
(kg/m3 ) (kg/m3 ) (kg/m3 ) (%) Ash %
Mean 445.8 170.8 783.7 (kg/m3 )
940.7 0.9 6.0 1.4% 0.8 40.5 (%) 0.6 (mm) (mm)
61.3 (28 Days)
Mean Standard
445.8 170.8 783.7 940.7 0.9 6.0 1.4 0.8 40.5 0.6 61.3
8.2 2.4 11.9 19.9 0.1 0.9 0.4 0.0 1.2 0.0 1.7
Standard Error
8.2 2.4 11.9 19.9 0.1 0.9 0.4 0.0 1.2 0.0 1.7
Error Median 400.0 157.8 743.0 1050.5 0.2 0.0 0.0 1.0 35.0 0.6 62.8
Median Mode 400.0400.0 157.8152.0 835.0
743.0 1047.0
1050.5 0.0 0.2 0.0 0.00.0 0.5 0.0 60.0 1.0 0.8 35.0 29.1
0.6 62.8
Mode Standard 400.0 152.0 835.0 1047.0 0.0 0.0 0.0 0.5 60.0 0.8 29.1
105.4 30.7 153.3 256.8 1.8 11.7 5.7 0.6 16.1 0.2 21.6
Deviation
Standard
105.4 30.7 153.3 256.8 1.8 11.7 5.7 0.6 16.1 0.2 21.6
DeviationRange 400.0 137.0 768.0 1170.0 9.0 43.0 30.0 2.0 60.0 0.9 73.1
RangeMinimum 400.0280.0 137.0
133.0 582.0
768.0 0.0
1170.0 0.0 9.0 0.0 0.0
43.0 0.0 30.0 0.0 2.0 0.0 60.0 26.1
0.9 73.1
MinimumMaximum 280.0680.0 133.0
270.0 1350.0
582.0 1170.0
0.0 9.0 0.0 43.0 30.0 0.0 2.0 0.0 60.0 0.0 0.9 0.0 99.2
0.0 26.1
Maximum
Count 680.0166.0 270.0
166.0 1350.0
166.0 166.0
1170.0
166.0 9.0 166.0 166.0
43.0
166.0 30.0 166.0 2.0 166.0 60.0 166.0
0.9 99.2
Count 166.0 166.0 166.0 166.0 166.0 166.0 166.0 166.0 166.0 166.0 166.0

60
70

50
60

40 50
Count

Count

40
30

30
20
20

10
10

0 0
300 400 500 600 700 0 200 400 600 800 1000 1200
3
Cement (kg/m )
Materials 2022, 15, x FOR PEER REVIEW Coarse aggregate (kg/m3) 5 of 18

(a) (b)
80
120
70

100
60

80 50
Count

Count

40
60

30
40
20

20
10

0 0
0 10 20 30 40 120 140 160 180 200 220 240 260 280
Silica fume (%) Water (kg/m3)

(c) (d)

Figure 1. Cont.
120 45

40
100
35

80 30

25
Count

Count

60
20
 20
10

0 0
0 10 20 30 40 120 140 160 180 200 220 240 260 280
Silica fume (%) Water (kg/m3)

Materials 2022, 15, 4209 5 of 18

(c) (d)

120 45

40
100
35

80 30

25
Count

Count
60
20

40 15

10
20
5

0 0
0 1 2 3 4 0.0 0.5 1.0 1.5 2.0
Superplasticizer (%) Vf of the hooked steel fiber (%)

(e) (f)

40
70
35
60
30

50
25
Count

Count

40
20

30
15

20 10

10 5

0 0
600 800 1000 1200 1400 20 40 60 80 100
Sand (kg/m3) Compressive Strength MPa (28 days)

(g) (h)
Figure 1. Relative frequency patterns of input and output factors: (a) cement; (b) coarse aggregate;
Figure(c) 1. Relative
silica frequency
fume; (d) water; (e)patterns of input
superplasticizer; (f) and output
volume factors:
fraction (a)hooked
of steel cement; (b)(g)
fibers; coarse
sand;aggregate;
(c) silica fume;
and (h) (d) water;
compressive (e) superplasticizer; (f) volume fraction of steel hooked fibers; (g) sand; and
strength.
(h) compressive strength.

3. Research Strategy
The ML models were run with Python code using Anaconda software. The Anaconda
navigator is a desktop graphical user interface featured in the Anaconda program that
allows for the running of apps that provide guidance through the Conda packages, channels,
and environments without having to use command-line techniques. It is also a source for
Python and R programming languages for data science and machine learning applications,
with an emphasis on package development and maintenance. This work used three
techniques to estimate the compressive strength of SFRC, i.e., SVR, AdaBoost, and bagging.
Spyder (version: 4.3.5) was chosen from the Anaconda navigator for model execution.
The degree of accuracy was represented by the R2 value of the predicted result from all
models. R2 values typically range from 0 to 1, with a higher number suggesting greater
accuracy between the measured and predicted outcomes. Furthermore, statistical checks,
error evaluation (including MAE, RMSE), and k-fold cross-validation were done to assess
the performance of the models employed in this study. A sensitivity analysis was also
performed to identify the influence of all input variables. Figure 2 illustrates the research
strategy as a flowchart.
 tion. The degree of accuracy was represented by the R value of the predicted result from
all models. R2 values typically range from 0 to 1, with a higher number suggesting greater
accuracy between the measured and predicted outcomes. Furthermore, statistical checks,
error evaluation (including MAE, RMSE), and k-fold cross-validation were done to assess
Materials 2022, 15, 4209
the performance of the models employed in this study. A sensitivity analysis6 ofwas 18
also
performed to identify the influence of all input variables. Figure 2 illustrates the research
strategy as a flowchart.

Figure 2. 2.Research
Figure Researchmethodology inthe
methodology in thecurrent
current study.
study.

4. Resultsand
4. Results andDiscussions
Discussions
4.1. Statistical Analysis Explanation
4.1. Statistical Analysis Explanation
Figure 3 depicts the pattern of statistical analysis for the actual and predicted outcomes
Figurecompressive
of SFRC 3 depicts the pattern
strength of28
after statistical analysis
days utilizing for the
the SVR actual
model. Theand
SVRpredicted
delivers out-
comes of SFRC compressive strength after 28 days utilizing the SVR
results that are within the permitted range and have a minimal difference between model. Thereal
SVR de-
and results
expected results. 2
The R the= 0.81 demonstrates
livers that are within permitted rangethat
andthe model
have performsdifference
a minimal well in terms
between
of calculating
real and expected outcomes. Figure
results. The 4Rshows
2 = 0.81 thedemonstrates
distribution ofthat
investigational
the modeland estimated
performs well in
outcomes, and the SVR model’s errors. In the dispersal, the largest, lowermost,
terms of calculating outcomes. Figure 4 shows the distribution of investigational and es- and average
values of error were 26.02, 0.53, and 7.03 MPa, respectively. It was discovered that 44% of
timated outcomes, and the SVR model’s errors. In the dispersal, the largest, lowermost,
the erroneous values were less than 5 MPa, 28% were between 5 and 10 MPa, and 28% were
and average values of error were 26.02, 0.53, and 7.03 MPa, respectively. It was discovered
higher than 10 MPa. These statistics represent the level of agreement between the projected
that
and44% of the
actual erroneous values were less than 5 MPa, 28% were between 5 and 10 MPa,
results.
and 28% were higher thanthe
Figures 5 and 6 depict 10AdaBoost
MPa. These statistics
model’s represent
outputs. the levelthe
Figure 5 displays ofconnection
agreement be-
tween the projected
between the genuine and
andactual results.
predicted 2
outcomes, with R = 0.96, which is greater than that
of SVR model, implying that the AdaBoost approach performs better than SVR. Figure 6
depicts the AdaBoost model’s distribution of actual and projected values as well as errors.
The distribution’s highest, lowest and average values of the errors were 11.80, 0.005, and
3.10 MPa, respectively. According to the findings, 80% of the erroneous readings were less
than 5 MPa, 12% were in the range of 5–10 MPa, and 8% were more than 10 MPa. The
AdaBoost model can more effectively predict SFRC’s compressive strength based on the R2
and error distribution of the SVM and AdaBoost models.
Materials 2022, 15, x FOR PEER REVIEW 7 of 18

Materials 2022, 15, 4209

x FOR PEER REVIEW 7 of 18
Training data Estimated (MPa) Poly. (Estimated (MPa))
120
y = -0.0037x2 + 1.1857x + 5.7672
Predicted compressive strength (MPa)

R²= 0.812Estimated (MPa)

Training data Poly. (Estimated (MPa))
100120
y = -0.0037x2 + 1.1857x + 5.7672
Predicted compressive strength (MPa)

R²= 0.812
80100

80
60

60
40
40
20
20
0
00 20 40 60 80 100 120
0 20 Experimental
40 60
compressive 80
strength 100
(MPa) 120
Experimental compressive strength (MPa)
Figure 3. Relationship for support vector regression model: experimental and estimated results.
Figure
Figure 3.
3. Relationship
Relationship for
for support
support vector
vector regression model: experimental
regression model: and estimated
experimental and estimated results.
results.

Experimental (MPa) Estimated (MPa) Errors (MPa)

100 Experimental (MPa) Estimated (MPa) Errors (MPa)
100
Compressive stregth (MPa)
Compressive stregth (MPa)

80 80

60 60

40 40

20
20

0
0
0 5 10 15 20 25
0 5 10 Data set 15 20 25
Data set
Figure 4. Experimental and estimated values and error distribution for support vector regression
Figure 4. Experimental and estimated values and error distribution for support vector regression model.
model.
Figure 4. Experimental and estimated values and error distribution for support vector regression
model. Figures 5 and 6 depict the AdaBoost model’s outputs. Figure 5 displays the connec-
tion between the genuine and predicted outcomes, with R2 = 0.96, which is greater than
Figures 5 and 6 depict the AdaBoost model’s outputs. Figure 5 displays the connec-
tion between the genuine and predicted outcomes, with R2 = 0.96, which is greater than
 0.005, and 3.10 MPa, respectively. According to the findings, 80% of the erroneous re
ure 6 depicts the AdaBoost model’s distribution of actual and projected values as well as
ings were less than 5 MPa, 12% were in the range of 5–10 MPa, and 8% were more than
errors. The distribution’s highest, lowest and average values of the errors were 11.80,
MPa. The AdaBoost model can more effectively predict SFRC’s compressive stren
0.005, and 3.10 MPa, respectively. According to the findings, 80% of the erroneous read-
based on the
ings were lessRthan
2 and error distribution of the SVM and AdaBoost models.
5 MPa, 12% were in the range of 5–10 MPa, and 8% were more than 10
Materials 2022, 15, 4209 8 of 18
MPa. The AdaBoost model can more effectively predict SFRC’s compressive strength
based on the R
Training
2 and error distribution of the SVM and AdaBoost models.
data Estimated (MPa) Poly. (Estimated (MPa))
120
yTraining
= -0.002x 2 + 1.1202x + 0.1891
data Estimated (MPa) Poly. (Estimated (MPa))
Predicted compressive strength (MPa)

120 R²= 0.9605

100 y = -0.002x2 + 1.1202x + 0.1891
Predicted compressive strength (MPa)

R²= 0.9605
100
80
80
60
60
40
40

20
20

0
0
0 20 40 60 80 100 120
0 20 40 60 80 100 120
Experimental compressive strength (MPa)
Experimental compressive strength (MPa)
Figure
Figure5.5.Connection betweenexperimental
Connection between experimental and
and estimated
estimated results
results for SVR
for SVR AdaBoost
AdaBoost model.model.
Figure 5. Connection between experimental and estimated results for SVR AdaBoost model.

Experimental (MPa)
Experimental (MPa) Estimated
Estimated (MPa)
(MPa) Errors
Errors (MPa)
(MPa)
100
100
Compressive strength (MPa)
Compressive strength (MPa)

8080

60
60

40
40

20
20

0
0
0 5 10 15 20 25

0 5 10Data set 15 20 25
Data set
Figure 6. Experimental and estimated values and error distribution for SVR AdaBoost model.
Figure 6. Experimental and estimated values and error distribution for SVR AdaBoost model.
Figure 6. Experimental and estimated values and error distribution for SVR AdaBoost model.
The connection of the actual and expected outcomes for the model of SVR bagging
is shown in Figure 7. The SVR bagging model has R2 = 0.87, showing that this model is
Materials 2022, 15, x FOR PEER REVIEW 9 of 18

The connection of the actual and expected outcomes for the model of SVR bagging is
Materials 2022, 15, 4209 9 of 18
shown in Figure 7. The SVR bagging model has R2 = 0.87, showing that this model is more
precise than the SVR but less precise than SVR AdaBoost models. Figure 8 also shows the
SVR bagging distribution for the actual and projected values and errors. The values 16.3,
more precise than the SVR but less precise than SVR AdaBoost models. Figure 8 also shows
0.53, and 6.5 MPa were the greatest, lowest, and average errors, respectively. It was dis-
the SVR bagging distribution for the actual and projected values and errors. The values
covered that 52% of the erroneous values were less than 5 MPa, 28% were between 5 Mpa
16.3, 0.53, and 6.5 MPa were the greatest, lowest, and average errors, respectively. It was
anddiscovered
10 MPa, that and52%
20%ofwere larger than
the erroneous 10 were
values MPa.lessThis investigation
than 5 MPa, 28% weredemonstrated
between 5 Mpa that the
SVR andAdaBoost
10 MPa, andmodel20%had
weresuperior precision
larger than 10 MPa.compared with thedemonstrated
This investigation SVR and SVR AdaBoost
that the
models due to itsmodel
SVR AdaBoost lowerhaderror and higher
superior R2 readings.
precision compared In addition,
with the SVRtheandMLSVRtechniques,
AdaBoost i.e.,
models due to its SVR
lowerbagging,
error andusedhigher 2 readings. In addition, the ML techniques,
SVR AdaBoost and theRsub-models to obtain the best assessments that
i.e., SVR
yielded AdaBoost
nearly perfectand SVR bagging,
outcomes. Wangused
et the sub-models
al. [61] reported to obtain the best assessments
that AdaBoost machine learn-
that yielded nearly perfect outcomes. Wang et al. [61] reported
ing approaches better predict the compressive strength of geopolymer composites. that AdaBoost machine Zhu et
learning approaches better predict the compressive strength of geopolymer composites.
al. [62] used machine learning to forecast the splitting tensile strength (STS) of concrete
Zhu et al. [62] used machine learning to forecast the splitting tensile strength (STS) of
containing recycled aggregate (RA) and revealed that the precision level of the bagging
concrete containing recycled aggregate (RA) and revealed that the precision level of the
model was
bagging modelbetter.was
Ahmad
better.et Ahmad
al. [63] studied
et al. [63]boosting, and the AdaBoost
studied boosting, ML approaches
and the AdaBoost ML
to predict
approaches to predict the compressive strength of high-calcium fly-ash-based geopolymer.Ada-
the compressive strength of high-calcium fly-ash-based geopolymer. The
Boost
Theand bagging
AdaBoost andmodels
baggingyielded better results.
models yielded As a result,
better results. this study
As a result, found
this study thatthat
found the ML
the ML(SVR
methods methods (SVR AdaBoost
AdaBoost and SVR and SVR bagging)
bagging) wereweremoremore accurate
accurate thanthan
thethe individualstrat-
individual
egystrategy in predicting
in predicting outcomes
outcomes (SVR).(SVR).

Training data Estimated (MPa) Poly. (Estimated (MPa))

120
y = 0.001x2 + 0.6997x + 16.945
Predicted compressive strength (MPa)

R²= 0.8772
100

80

60

40

20

0
0 20 40 60 80 100 120
Experimental compressive strength (MPa)
Figure 7. Experimental and estimated outcome relationship for SVR bagging model.
Figure 7. Experimental and estimated outcome relationship for SVR bagging model.
terials 2022, 15, x FOR PEER REVIEW 10 o
Materials 2022, 15, 4209 10 of 18

Experimental (MPa) Estimated (MPa) Errors (MPa)

100
Compressive strength (MPa)

80

60

40

20

0 5 10 15 20 25
Data set
Figure 8. Experimental and estimated values and error distribution for SVR bagging model.
Figure 8. Experimental and estimated values and error distribution for SVR bagging model.
4.2. Cross-Validation Using the K-Fold Scale
4.2. Cross-Validation
To assess the Using the K-Fold
legitimacy Scale during execution, the k-fold cross-validation
of the model
technique was utilized. This technique is regularly employed to check the correctness of the
To assess the legitimacy of the model during execution, the k-fold cross-validati
model in which the data set spread and split into 10 groups [60–65]. One group is used to
technique was
verify the utilized.
model, while This technique
the other is regularly
9 are utilized employed
for training; 70% of the to check
data set was the correctness
utilized
the model in which
in the training the data
process of theset spread
models, andthe
while split into 1030%
remaining groups
of the[60–65].
data wereOne usedgroup
for is us
to verify the
testing andmodel, while
validation the
of the otherIf9the
models. areR2utilized for and
value is high training; 70%i.e.,
the errors, ofthetheMAEdata set w
and in
utilized RMSEthevalues,
trainingare low, then the
process of model is considered
the models, while more
theaccurate.
remainingTo achieve
30% aofdecent
the data w
outcome, the technique must be performed 10 times. The model’s high accuracy is due to
used this
for complete
testing and validation of the models. If the R value is high and the errors, i
2
methodology. In addition, as shown in Table 2, all models were subjected
the MAE and RMSE
to statistical analysis values,
of errors,are low,
namely MSEthenandtheRMSE. model is considered
The response more
of the models to accurate.
the
achieve a decent
estimate outcome,
was evaluated the technique
by statistical analysis,must be performed
employing Equations (1)10and
times. Thefrom
(2), taken model’s hi
the literature
accuracy is due to [66]:
this complete methodology. In addition, as shown in Table 2, all mod
1 n
n i∑
were subjected to statistical analysis MAEof = errors, | x namely
i− x| MSE and RMSE. The (1) response
=1
the models to the estimate was evaluated v by statistical analysis, employing Equations
and (2), taken from the literature [66]:u u

y −y
2
pred re f
∑ 1
t
RMSE = 𝑛 (2)
n
where n = the total number of sampled MAE
data,=x,𝑛yre
∑=|𝑥 𝑖 − 𝑥| values of the data sample,
reference
f
and xi , y pred = model-predicted values. 𝑖=1
Figures 9–11 show the distributions of MAE, RMSE, and R2 for the k-fold cross-
validation of models SVR, SVR AdaBoost, and SVR bagging, respectively. 2 As shown in
Figure 9, the SVR model’s best,RMSE lowest,=and√∑ (𝑦 𝑝𝑟𝑒𝑑 2− 𝑦𝑟𝑒𝑓 )
average R values are 0.80, 0.55, and 0.69,
respectively. The greatest, lowermost, and average R2 values 𝑛 for the AdaBoost model were
0.90, 0.60, and 0.76, respectively, as illustrated in Figure 10. The greatest, least, and R2
where 𝑛 = the
average total
values fornumber
the baggingof model
sampled
weredata, 𝑥, 𝑦and
0.92, 0.61, = reference
𝑟𝑒𝑓 0.79, values
respectively, of the data
as presented sa
ple, and 𝑥𝑖 , 𝑦𝑝𝑟𝑒𝑑 = model-predicted values.
in Figure 11. When the error values for the SVR model were compared, the average MAE

Table 2. Statistical analysis of the approaches used.

Models MAE (MPa) RMSE (MPa) R2


Materials 2022, 15, 4209
were 9.04 and 13.62, respectively. The AdaBoost model had average MAE and RMSE val-
ues of 8.18 and 11.63, respectively, whereas the bagging model had average MAE and
RMSE values of 6.51 and 8.39, respectively. The AdaBoost model with the lowermost error
and the highest R2 value performed the best in terms of predicting outcomes. Table 3
shows the k-fold analysis findings for the utilized models, including the MAE, RMSE,
11 ofand
18
R2 values.

Table 3. K-fold cross-validation results.

and RMSE were 9.04 and 13.62, respectively. The AdaBoost model had average MAE and
RMSE values of 8.18 and SVR11.63, respectively,SVR whereas the bagging modelSVR
AdaBoost hadBagging
average MAE
K-Fold
and RMSE values
MAE of 6.51RMSE and 8.39,
R2 respectively.
MAE The AdaBoost
RMSE R2 model
MAEwith RMSE
the lowermost
R2
error1and the3.02 2 value performed the best in terms of predicting outcomes. Table 3
highest R3.11 0.78 3.62 8.29 0.91 6.79 8.84 0.72
shows2 the k-fold
3.43 analysis
4.10findings
0.75for the utilized4.29
3.43 models, 0.94
including the MAE,
4.81 RMSE,0.82
5.28 and
R2 values.
3 9.49 11.80 0.30 8.24 9.65 0.70 11.15 14.15 0.48
4 2.72 4.73 0.79 3.49
Table 2. Statistical analysis of the approaches used. 4.29 0.94 4.14 5.93 0.87
5 12.69 14.10 0.28 4.58 7.89 0.87 5.93 7.31 0.75
Models MAE (MPa) RMSE (MPa) R 2
6 4.36 6.76 0.77 7.82 8.76 0.79 4.88 6.88 0.83
7
Support 7.73
vector 10.50 0.60 9.40 10.70 0.69 8.72 11.46 0.62
7.0 9.1 0.81
8 regression
2.95 4.07 0.81 4.79 5.83 0.81 4.66 5.45 0.83
SVR AdaBoost 4.4 8.0 0.96
9 3.52 5.52 0.77 3.23 4.90 0.94 4.83 5.53 0.81
SVR bagging 6.2 7.6 0.87
10 9.76 11.17 0.30 2.46 3.49 0.96 5.16 7.46 0.79

Support vector regressor (SVR)

MAE RMSE R2
16 1.00

14
0.80
Values in bars

12

Values in line
10 0.60
8

6 0.40

4
0.20
2

0 0.00
1 2 3 4 5 6 7 8 9 10
K-fold Validation
Materials 2022, 15, x FOR PEER REVIEW 12 of 18
Figure9.9.Support
Figure Supportvector
vectorregression
regressionmodel
modelwith
withK-fold
K-foldcross-validation
cross-validationrepresentation.
representation.

SVR AdaBoost
MAE RMSE R2
16 1.00

14
0.80
12
Values in line
Values in bars

10 0.60
8

6 0.40

4
0.20
2

0 0.00
1 2 3 4 5 6 7 8 9 10
K-fold Validation

Figure10.
Figure 10.K-fold
K-foldcross-validation
cross-validationrepresentation
representationfor
forthe
theAdaBoost
AdaBoostmodel.
model.

SVR Bagging
MAE RMSE R2
16 1.00

14
 0.20
2

0 0.00
1 2 3 4 5 6 7 8 9 10
Materials 2022, 15, 4209 K-fold Validation 12 of 18

Figure 10. K-fold cross-validation representation for the AdaBoost model.

SVR Bagging
MAE RMSE R2
16 1.00

14
0.80
12
Values in bars

Values in line
10 0.60
8

6 0.40

4
0.20
2

0 0.00
1 2 3 4 5 6 7 8 9 10
K-fold Validation

Figure 11. K-fold cross-validation representation for bagging model.

Figure 11. K-fold cross-validation representation for bagging model.

4.3. Sensitivity
Table Analysis
3. K-fold cross-validation results.
The goal of this study was to determine how input factors affect the forecasting of
SVR
SFRC compressive strength.SVR
TheAdaBoost SVR Bagging
input factors showed a significant impact on the pre-
K-Fold 2 2
MAE dicted result
RMSE R [67]. Figure
MAE 12 depicts
RMSEthe impactR of input variables RMSE
MAE R2
on the prediction of
1 3.02 SFRC’s compressive
3.11 0.78 strength.
3.62 The investigation
8.29 0.91 found that
6.79 silica fume
8.84 was the0.72
utmost
2 3.43 essential
4.10 element,
0.75 accounting
3.43 for 21.9%
4.29 of the total,
0.94 along with
4.81 cement (16.2%)
5.28 and
0.82super
3 9.49 plasticizer0.30
11.80 (16.4%). The
8.24other input
9.65factors had
0.70a smaller11.15
impact on 14.15
the prediction0.48of the
4 2.72 4.73
compressive0.79 strength3.49 4.29 coarse aggregate
of SFRC, with 0.94 4.14
accounting for5.93
13%, water0.87
15.2%,
5 12.69 14.10
and sand 0.28 4.58 of input
6%. The number 7.89factors and
0.87data points
5.93 utilized 7.31
in the design0.75
of the
6 4.36 6.76 0.77 7.82 8.76 0.79 4.88 6.88 0.83
models were proportional to the results of the sensitivity analysis. The influence of input
7 7.73 10.50 0.60 9.40 10.70 0.69 8.72 11.46 0.62
8 2.95 variables 0.81
4.07 on the model output was
4.79 checked using
5.83 0.81 Equations4.66 (3) and (4).
5.45As can be0.83
seen in
9 3.52 Figure 12,0.77
5.52 cement, silica
3.23 fume, water
4.90 and superplasticizer
0.94 had the greatest
4.83 5.53 effect on the
0.81
10 9.76 compressive
11.17 0.30 strength.2.46
This is obvious
3.49 as these 0.96are the main
5.16 factors7.46 contributing to the
0.79
strength development of SFRC, with cement and silica fume especially increasing the
compressive
4.3. Sensitivitystrength
Analysisdue to their hydration and pozzolanic activity. Similarly, water and
The goal of this study was to determine how input factors affect the forecasting of
SFRC compressive strength. The input factors showed a significant impact on the predicted
result [67]. Figure 12 depicts the impact of input variables on the prediction of SFRC’s
compressive strength. The investigation found that silica fume was the utmost essential
element, accounting for 21.9% of the total, along with cement (16.2%) and super plasticizer
(16.4%). The other input factors had a smaller impact on the prediction of the compressive
strength of SFRC, with coarse aggregate accounting for 13%, water 15.2%, and sand 6%.
The number of input factors and data points utilized in the design of the models were
proportional to the results of the sensitivity analysis. The influence of input variables on the
model output was checked using Equations (3) and (4). As can be seen in Figure 12, cement,
silica fume, water and superplasticizer had the greatest effect on the compressive strength.
This is obvious as these are the main factors contributing to the strength development of
SFRC, with cement and silica fume especially increasing the compressive strength due to
their hydration and pozzolanic activity. Similarly, water and superplasticizer are also main
factors as a decreased w/c ratio results in increased compressive strength. Similarly, coarse
aggregate, steel fiber volume, and length also contribute to the strength, as explained in
Section 2.
Ni = f max ( xi ) − f min ( xi ) (3)
 superplasticizer are also main factors as a decreased w/c ratio results in increased com-
pressive strength. Similarly, coarse aggregate, steel fiber volume, and length also contrib-
ute to the strength, as explained in Section 2.

𝑁𝑖 = 𝑓𝑚𝑎𝑥 (𝑥𝑖 ) − 𝑓𝑚𝑖𝑛 (𝑥𝑖 ) (3)

Materials 2022, 15, 4209 𝑁𝑖 13 of 18
𝑆𝑖 = (4)
∑𝑛𝑗−𝑖 𝑁𝑗
The highest and lowest projected Ni over the 𝑖 𝑡ℎ output are represented by
outputs
Si = n (4)
𝑓𝑚𝑎𝑥 (𝑥𝑖 ) and 𝑓𝑚𝑖𝑛 (𝑥𝑖 ), respectively. ∑ j−i Nj

25

20
Percentage (%)

15

10

0
Cement Silica fume Water
Sand Superplasticizer Coarse aggregate
Vf Fiber length Fiber diameter

Figure12.
Figure 12.The
Theinput
inputvariable
variablecontribution
contributiontotothe
theforecast.
forecast.Cement,
Cement,water,
water,superplasticizer,
superplasticizer,silica
silica
fume, coarse aggregate, sand, Vf (volume factor), fiber length, and fiber diameter.
fume, coarse aggregate, sand, Vf (volume factor), fiber length, and fiber diameter.

5. Discussion
The highest and lowest projected outputs over the ith output are represented by
f max ( xThis
i ) and f min ( xiwork
research ), respectively.
aimed to determine how ML methods can be utilized to predict
the compressive strength of SFRC. Learning rates and other features that specifically affect
5.ensemble
Discussion approaches can be used as tuning parameters for the model used in the ensem-
This research
ble techniques. work aimed
Boosting to determine
ensemble models (20 how ML methods
each) with 10, 20, can30,
be …, utilized to predict
200 component
the compressive
sub-models were strength of SFRC.
developed Learning
for base rates
learners inand
this other
work,features that specifically
and correlation affect
with high co-
ensemble approaches
efficient values can be used
was utilized as tuning
to identify theparameters
best model.for Thetherandomization
model used in technique
the ensemble
fur-
techniques.
ther revealedBoosting ensemble
the statistical modelsThe
importance. (20 test
each)waswith 10, 20,
carried out 30, . . .permuting
by (i) , 200 component
the data
sub-models were developed for base learners in this work,
set’s activity values repeatedly, (ii) generating models from the permuted values, and correlation with
andhigh
(iii)
coefficient values was utilized to identify the best model. The
comparing the resulting scores to the score of the original model derived from non-ran- randomization technique
further
domized revealed
activitythe statistical
values. If theimportance.
original model Theistest was carried
statistically out by (i)
significant, thepermuting
score fromthethe
data set’s activity
permuted data shouldvalues berepeatedly,
significantly (ii)higher.
generating
Threemodels
machine from the permuted
learning approaches values,
were
and (iii) comparing
considered: the resulting
one individual, scores
i.e., SVR, andtotwotheensemble,
score of the i.e.,original model derived
SVR AdaBoost and SVRfrom
bag-
non-randomized
ging. The prediction performance of each approach was evaluated to see which the
activity values. If the original model is statistically significant, score
approach
from
is thethemostpermuted
precise data
in should be significantly
prediction. When compared, higher. Three
the SVR machine
and SVR learning approaches
bagging models
were considered: one individual, i.e., SVR, and two ensemble,
yielded R2 values of 0.81 and 0.87, respectively. The SVR AdaBoost outcome was more i.e., SVR AdaBoost and
SVR bagging. The prediction performance of each approach was evaluated
exact, with an R2 value of 0.96. To confirm the efficiency of all models, statistical analysis to see which
approach
and the k-foldis thecross-validation
most precise intechnique
prediction. When
were compared,
applied. The modelsthe SVR work and SVRwith
better bagging
min-
2
models yielded R values of 0.81 and 0.87, respectively. The SVR AdaBoost outcome was
imal error levels. The 2ML techniques frequently take advantage of the vulnerable intern
more exact, with an R value of 0.96. To confirm the efficiency of all models, statistical2
by developing sub-models trained on data and maximized to enhance the value of R .
analysis and the k-fold cross-validation technique were applied. The models work better
with minimal error levels. The ML techniques frequently take advantage of the vulnerable
intern by developing sub-models trained on data and maximized to enhance the value of R2 .
Figures 13 and 14 display the range of R2 values for the sub-models, i.e., the SVR AdaBoost
and SVR bagging approaches. The peak, lowermost, and average R2 values of the AdaBoost
sub-models were 0.96, 0.63, and 0.79, respectively. For the bagging sub-models, the largest,
least, and average R2 values were 0.87, 0.44, and 0.68, respectively. These outcomes indicate
that SVR AdaBoost is more accurate than the bagging sub-models. Sensitivity analysis
Materials 2022, 15, x FOR PEER REVIEW 14 of 18

Figures 13
Figures 13 and
and 14
14 display
display the
the range
range ofof R
R22 values
values for
for the
the sub-models,
sub-models, i.e.,
i.e., the
the SVR
SVR AdaBoost
AdaBoost
and SVR bagging approaches. The peak, lowermost, and average R
and SVR bagging approaches. The peak, lowermost, and average R2 values of the Ada-
2 values of the Ada-

Materials 2022, 15, 4209 Boost sub-models

Boost sub-models werewere 0.96,
0.96, 0.63,
0.63, and
and 0.79,
0.79, respectively.
respectively. For
For the
the bagging
bagging sub-models,
sub-models,
14 ofthe
18
the
largest, least,
largest, least, and
and average
average R R2 values
2 values were
were 0.87,
0.87, 0.44,
0.44, and
and 0.68,
0.68, respectively.
respectively. These
These outcomes
outcomes
indicate that SVR AdaBoost is more accurate than the bagging sub-models.
indicate that SVR AdaBoost is more accurate than the bagging sub-models. Sensitivity Sensitivity
analysis
was was also
also carried
analysis was also carried
out out to
to determine
carried out to determine
determine
how each howhow each
inputeach input parameter
parameter
input parameter
affected the affected the
SFRC’sthe
affected SFRC’s
predicted
SFRC’s
predicted
compressive compressive
predicted compressive strength.
strength. The The
sensitivity
strength. sensitivity analysis
analysis looked
The sensitivity looked
at looked
analysis how much at how
each
at how much each
of the
much of the
the
10 input
each of
10 input
factors
10 input factors
influenced influenced the
the expected
factors influenced expected
the outcome. outcome.
expected outcome.

SVR AdaBoost
SVR AdaBoost
1.00
1.00
0.96
0.96
0.90
0.90

0.80
0.80
R2
R2

0.70
0.70

0.60
0.60

0.50
0.50

0.40
0.40
11 22 33 44 55 66 77 88 99 10
10 11
11 12
12 13
13 14
14 15
15 16
16 17
17 18
18 19
19 20
20
Sub models
Sub models
Figure 13.The
The coefficientof
of correlation(R
(R2) valuesof of theSVR
SVR AdaBoostsub-model.
sub-model.
Figure 13. Thecoefficient
Figure13. correlation (R22)) values
coefficient ofcorrelation values ofthe
the SVRAdaBoost
AdaBoost sub-model.

SVR Bagging
SVR Bagging
1.00
1.00

0.90 0.87
0.87
0.90

0.80
0.80
R2
R2

0.70
0.70

0.60
0.60

0.50
0.50

0.40
0.40
11 22 33 44 55 66 77 88 99 10
10 11
11 12
12 13
13 14
14 15
15 16
16 17
17 18
18 19
19 20
20
Sub models
Sub models
Figure 14. SVR bagging sub-model’s coefficient of correlation (R22) values.
Figure14.
Figure 14.SVR
SVRbagging
baggingsub-model’s
sub-model’scoefficient
coefficientof
ofcorrelation
correlation(R
(R2 ))values.
values.

6.Conclusions
6.
6. Conclusions
Conclusions
Thisresearch
This
This researchaimed
research aimedto
aimed topractice
to practiceindividual
practice individualand
individual andensembled
and ensembledML
ensembled MLapproaches
ML approachesto
approaches toestimate
to estimate
estimate
the
the 28-day
the 28-day compressive
28-day compressive strength
compressive strength of
strength of SFRC.
of SFRC. To
SFRC. To predict
Topredict the
predictthe outcomes,
theoutcomes, the
outcomes,the researchers
theresearchers used
researchersused
used
support
support vector
support vector regression
vector regression (SVR),
regression (SVR), AdaBoost,
(SVR), AdaBoost, and
AdaBoost, and bagging
andbagging models.
baggingmodels. The
models. The following
The following are
following are the
are the
the
findings
findings
findingsofof this
ofthis investigation:
thisinvestigation:
investigation:
• Individual approaches were less accurate than EML procedures in forecasting SFRC’s
compressive strength, while the SVR bagging model displayed the highest accuracy.
• The SVR bagging model outperformed the SVR AdaBoost ensembled machine learning
technique in the forecasting of the 28-day compressive strength of SFRC.
Materials 2022, 15, 4209 15 of 18

• The SVR, SVR AdaBoost, and SVR bagging models have coefficient of determination
(R2) values of 0.81, 0.96, and 0.87, respectively. All of the models’ outputs are within
acceptable bounds, with little variance from the exact results.
• The models’ performances were demonstrated by the k-fold cross-validation test and
statistical analysis, which revealed that the SVR bagging model outperformed the
other models investigated in terms of prediction.
• To determine how much the input parameters mattered, a sensitivity analysis was
utilized, and it was discovered that cement, water, silica fume, sand, superplasticizer,
coarse aggregate, Vf, fiber length, and fiber diameter contributed 16.2%, 15.2%, 21.9%,
6%, 16.4%, 13%, 8.7%, 2.6%, and 0.6%, respectively, to the outcome predictions.
• The unique ensemble machine learning algorithms, especially that of the SVR bagging
model, can effectively estimate concrete strength qualities without the requirement for
prolonged casting and testing process.
However, other prediction approaches [68–72], such as density functional theory can
be used to understand atomistic details of crack and structural failure (which can be fed
directly into machine learning approaches instead of just obtaining data from the literature).
These approaches can reveal mechanisms for the strength of structures in a more unbiased
way, which should be studied in future. In addition, this research was also limited to
the prediction of compressive strength at 28 days with 10 input parameters and did not
consider specimen size and the curing age of concrete. Indeed, proper database and testing
must be applied as these are vital elements for engineering applications. This study was
based on a wide range of data sets with nine input variables; however, the database and
more input parameters, such as specimen size and curing age, among others, need to be
generated in future for a better response from the employed models.

Supplementary Materials: The following supporting information can be downloaded at: https:
//www.mdpi.com/article/10.3390/ma15124209/s1, Table S1: Dataset used for prediction.
Author Contributions: Y.L.: Conceptualization, Investigation, Methodology, Formal analysis, Writing-
original draft. Q.Z.: Methodology, Investigation, Validation, Data Curation, Writing—Review and
Editing. P.K.: Software, Visualization, Resources, Funding acquisition, Writing—Review and Editing,
Supervision. A.F.D.: Conceptualization, Investigation, Methodology, Formal analysis, Funding acqui-
sition, Writing—Review and Editing. M.S.: Conceptualization, Investigation, Methodology, Formal
analysis, Writing-original draft, Supervision. A.D.: Validation, Data Curation, Writing—Review and
Editing, Funding acquisition, Project administration. N.B.K.: Methodology, Investigation, Validation,
Data Curation, Writing—Review and Editing. M.A.: Software, Validation, Visualization, Resources,
formal analysis, Writing—Review and Editing. All authors have read and agreed to the published
version of the manuscript.
Funding: The authors acknowledge the contribution and support provided by the individuals and
organizations that helped throughout this research work.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Data is already available in the supplemtray file.
Acknowledgments: The Authors extend their appreciation to the Deanship of Scientific Research at
King Khalid University for funding this work through the Large Groups Project under grant number
RGP. 2/104/43.
Conflicts of Interest: There authors declare no conflict of interest.

References
1. Nili, M.; Azarioon, A.; Danesh, A.; Deihimi, A. Experimental study and modeling of fiber volume effects on frost resistance of
fiber reinforced concrete. Int. J. Civ. Eng. 2018, 16, 263–272. [CrossRef]
2. Khan, M.; Ali, M. Improvement in concrete behavior with fly ash, silica-fume and coconut fibres. Constr. Build. Mater. 2019, 203,
174–187. [CrossRef]
Materials 2022, 15, 4209 16 of 18

3. Li, L.; Khan, M.; Bai, C.; Shi, K. Uniaxial tensile behavior, flexural properties, empirical calculation and microstructure of
multi-scale fiber reinforced cement-based material at elevated temperature. Materials 2021, 14, 1827. [CrossRef]
4. Cao, M.; Mao, Y.; Khan, M.; Si, W.; Shen, S. Different testing methods for assessing the synthetic fiber distribution in cement-based
composites. Constr. Build. Mater. 2018, 184, 128–142. [CrossRef]
5. Khan, M.; Cao, M.; Xie, C.; Ali, M. Hybrid fiber concrete with different basalt fiber length and content. Struct. Concr. 2022, 23,
346–364. [CrossRef]
6. Khan, M.; Cao, M.; Hussain, A.; Chu, S.H. 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]
7. Arshad, S.; Sharif, M.B.; Irfan-ul-Hassan, M.; Khan, M.; Zhang, J.L. Efficiency of Supplementary Cementitious Materials and
Natural Fiber on Mechanical Performance of Concrete. Arab. J. Sci. Eng. 2020, 45, 8577–8589. [CrossRef]
8. Xie, C.; Cao, M.; Guan, J.; Liu, Z.; Khan, M. Improvement of boundary effect model in multi-scale hybrid fibers reinforced
cementitious composite and prediction of its structural failure behavior. Compos. Part B Eng. 2021, 224, 109219. [CrossRef]
9. Thomas, B.S.; Yang, J.; Bahurudeen, A.; Abdalla, J.A.; Hawileh, R.A.; Hamada, H.M.; Nazar, S.; Jittin, V.; Ashish, D.K. Sugarcane
bagasse ash as supplementary cementitious material in concrete—A review. Mater. Today Sustain. 2021, 15, 100086. [CrossRef]
10. Young, B.A.; Hall, A.; Pilon, L.; Gupta, P.; Sant, G. Can the compressive strength of concrete be estimated from knowledge of the
mixture proportions?: New insights from statistical analysis and machine learning methods. Cem. Concr. Res. 2019, 115, 379–388.
[CrossRef]
11. Akande, K.O.; Owolabi, T.O.; Twaha, S.; Olatunji, S.O. Performance Comparison of SVM and ANN in Predicting Compressive
Strength of Concrete. IOSR J. Comput. Eng. 2014, 16, 88–94. [CrossRef]
12. Chou, J.S.; Tsai, C.F.; Pham, A.D.; Lu, Y.H. Machine learning in concrete strength simulations: Multi-nation data analytics. Constr.
Build. Mater. 2014, 73, 771–780. [CrossRef]
13. Duan, J.; Asteris, P.G.; Nguyen, H.; Bui, X.N.; Moayedi, H. A novel artificial intelligence technique to predict compressive strength
of recycled aggregate concrete using ICA-XGBoost model. Eng. Comput. 2021, 37, 3329–3346. [CrossRef]
14. Gupta, S.M. Support Vector Machines based Modelling of Concrete Strength. World Acad. Sci. Eng. Technol. 2007, 36, 305–311.
15. Chou, J.S.; Pham, A.D. Enhanced artificial intelligence for ensemble approach to predicting high performance concrete compressive
strength. Constr. Build. Mater. 2013, 49, 554–563. [CrossRef]
16. Deepa, C.; SathiyaKumari, K.; Sudha, V.P. Prediction of the Compressive Strength of High Performance Concrete Mix using Tree
Based Modeling. Int. J. Comput. Appl. 2010, 6, 18–24. [CrossRef]
17. Erdal, H.I. Two-level and hybrid ensembles of decision trees for high performance concrete compressive strength prediction. Eng.
Appl. Artif. Intell. 2013, 26, 1689–1697. [CrossRef]
18. Nafees, A.; Khan, S.; Javed, M.F.; Alrowais, R.; Mohamed, A.M.; Mohamed, A.; Vatin, N.I. Forecasting the Mechanical Properties
of Plastic Concrete Employing Experimental Data Using Machine Learning Algorithms: DT, MLPNN, SVM, and RF. Polymers
2022, 14, 1583. [CrossRef]
19. Nafees, A.; Javed, M.F.; Khan, S.; Nazir, K.; Farooq, F.; Aslam, F.; Musarat, M.A.; Vatin, N.I. Predictive modeling of mechanical
properties of silica fume-based green concrete using artificial intelligence approaches: MLPNN, ANFIS, and GEP. Materials 2021,
14, 7531. [CrossRef]
20. Khan, M.A.; Aslam, F.; Javed, M.F.; Alabduljabbar, H.; Deifalla, A.F. New prediction models for the compressive strength and
dry-thermal conductivity of bio-composites using novel machine learning algorithms. J. Clean. Prod. 2022, 350, 131364. [CrossRef]
21. Salem, N.M.; Deifalla, A. Evaluation of the Strength of Slab-Column Connections with FRPs Using Machine Learning Algorithms.
Polymers 2022, 14, 1517. [CrossRef] [PubMed]
22. Ebid, A.; Deifalla, A. Using Artificial Intelligence Techniques to Predict Punching Shear Capacity of Lightweight Concrete Slabs.
Materials 2022, 15, 2732. [CrossRef] [PubMed]
23. Dong, W.; Huang, Y.; Lehane, B.; Ma, G. XGBoost algorithm-based prediction of concrete electrical resistivity for structural health
monitoring. Autom. Constr. 2020, 114, 103155. [CrossRef]
24. 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]
25. Su, M.; Zhong, Q.; Peng, H.; Li, S. Selected machine learning approaches for predicting the interfacial bond strength between
FRPs and concrete. Constr. Build. Mater. 2021, 270, 121456. [CrossRef]
26. Nguyen, K.T.; 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]
27. 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]
28. Castelli, M.; Vanneschi, L.; Silva, S. Prediction of high performance concrete strength using Genetic Programming with geometric
semantic genetic operators. Expert Syst. Appl. 2013, 40, 6856–6862. [CrossRef]
29. Suleiman, A.R.; Nehdi, M.L. Modeling self-healing of concrete using hybrid genetic algorithm-artificial neural network. Materials
2017, 10, 135. [CrossRef]
30. Zhang, J.; Huang, Y.; Aslani, F.; Ma, G.; Nener, B. A hybrid intelligent system for designing optimal proportions of recycled
aggregate concrete. J. Clean. Prod. 2020, 273, 122922. [CrossRef]
Materials 2022, 15, 4209 17 of 18

31. Marani, A.; Nehdi, M.L. Machine learning prediction of compressive strength for phase change materials integrated cementitious
composites. Constr. Build. Mater. 2020, 265, 120286. [CrossRef]
32. Soulioti, D.V.; Barkoula, N.M.; Paipetis, A.; Matikas, T.E. Effects of fibre geometry and volume fraction on the flexural behaviour
of steel-fibre reinforced concrete. Strain 2011, 47, e535–e541. [CrossRef]
33. Yoo, D.Y.; Yoon, Y.S.; Banthia, N. Flexural response of steel-fiber-reinforced concrete beams: Effects of strength, fiber content, and
strain-rate. Cem. Concr. Compos. 2015, 64, 84–92. [CrossRef]
34. Jang, S.J.; Yun, H. Do Combined effects of steel fiber and coarse aggregate size on the compressive and flexural toughness of
high-strength concrete. Compos. Struct. 2018, 185, 203–211. [CrossRef]
35. Johnson, D.A.; Pedersen, N.; Jacobsen, C.B. Effect of steel fibers on flexural behaviour of normal and high strength concrete. Int. J.
Civ. Environ. Eng. 2014, 8, 22–26.
36. Dinh, N.H.; Park, S.H.; Choi, K.K. Effect of dispersed micro-fibers on tensile behavior of uncoated carbon textile-reinforced
cementitious mortar after high-temperature exposure. Cem. Concr. Compos. 2021, 118, 103949. [CrossRef]
37. Thomas, J.; Ramaswamy, A. Mechanical Properties of Steel Fiber-Reinforced Concrete. J. Mater. Civ. Eng. 2007, 19, 385–392.
[CrossRef]
38. Sivakumar, A.; Santhanam, M. Mechanical properties of high strength concrete reinforced with metallic and non-metallic fibres.
Cem. Concr. Compos. 2007, 29, 603–608. [CrossRef]
39. Afroughsabet, V.; Ozbakkaloglu, T. Mechanical and durability properties of high-strength concrete containing steel and polypropy-
lene fibers. Constr. Build. Mater. 2015, 94, 73–82. [CrossRef]
40. Atiş, C.D.; Karahan, O. Properties of steel fiber reinforced fly ash concrete. Constr. Build. Mater. 2009, 23, 392–399. [CrossRef]
41. Lee, J.H.; Cho, B.; Choi, E. Flexural capacity of fiber reinforced concrete with a consideration of concrete strength and fiber content.
Constr. Build. Mater. 2017, 138, 222–231. [CrossRef]
42. Köksal, F.; Altun, F.; Yiǧit, I.; Şahin, Y. Combined effect of silica fume and steel fiber on the mechanical properties of high strength
concretes. Constr. Build. Mater. 2008, 22, 1874–1880. [CrossRef]
43. Yoon, E.S.; Park, S.B. An experimental study on the mechanical properties and long-term deformations of high-strength steel fiber
reinforced concrete. J. Korean Soc. Civ. Eng. 2006, 26, 401–409.
44. Abbass, W.; Khan, M.I.; Mourad, S. Evaluation of mechanical properties of steel fiber reinforced concrete with different strengths
of concrete. Constr. Build. Mater. 2018, 168, 556–569. [CrossRef]
45. Yoo, D.Y.; Yoon, Y.S.; Banthia, N. Predicting the post-cracking behavior of normal- and high-strength steel-fiber-reinforced
concrete beams. Constr. Build. Mater. 2015, 93, 477–485. [CrossRef]
46. Lee, H.-H.; Lee, H.-J. Characteristic Strength and Deformation of SFRC Considering Steel Fiber Factor and Volume fraction.
J. Korea Concr. Inst. 2004, 16, 759–766. [CrossRef]
47. Oh, Y.H. Evaluation of Flexural Strength for Normal and High Strength Concrete with Hooked Steel Fibers. J. Korea Concr. Inst.
2008, 20, 531–539. [CrossRef]
48. Song, P.; Hwang, S. Mechanical properties of high-strength steel fiber-reinforced concrete. Constr. Build. Mater. 2004, 18, 669–673.
[CrossRef]
49. Reddy, V.M.; Rao, D.M.V.S. Effect of w/c ratio on workability and mechanical properties of high strength Self Compacting
Concrete (M70 grade). IOSR J. Mech. Civ. Eng. 2014, 11, 15–21. [CrossRef]
50. Nili, M.; Afroughsabet, V. Combined effect of silica fume and steel fibers on the impact resistance and mechanical properties of
concrete. Int. J. Impact Eng. 2010, 37, 879–886. [CrossRef]
51. Kim, J.J.; Kim, D.J.; Kang, S.T.; Lee, J.H. Influence of sand to coarse aggregate ratio on the interfacial bond strength of steel fibers
in concrete for nuclear power plant. Nucl. Eng. Des. 2012, 252, 1–10. [CrossRef]
52. Chitlange, M.R.; Pajgade, P.S. Strength appraisal of artificial sand as fine aggregate in SFRC. J. Eng. Appl. Sci. 2010, 5, 34–38.
53. Khan, M.; Ali, M. Effect of super plasticizer on the properties of medium strength concrete prepared with coconut fiber. Constr.
Build. Mater. 2018, 182, 703–715. [CrossRef]
54. Aruntaş, H.Y.; Cemalgil, S.; Şimşek, O.; Durmuş, G.; Erdal, M. Effects of super plasticizer and curing conditions on properties of
concrete with and without fiber. Mater. Lett. 2008, 62, 3441–3443. [CrossRef]
55. Nili, M.; Afroughsabet, V. Property assessment of steel-fibre reinforced concrete made with silica fume. Constr. Build. Mater. 2012,
28, 664–669. [CrossRef]
56. Saravana, R.M.K.; Sumathi, A. Effect of fly ash in fiber reinforced concrete composites. Jordan J. Civ. Eng. 2017, 11, 30–39.
57. Challoob, M.A.; Srivastava, V.; Materials, A. Effect of Fly Ash and Steel Fibre on Portland Pozzolana Cement Concrete. Int. J. Eng.
Trends Technol. 2013, 5, 144–147.
58. Saha, A.K. Effect of class F fly ash on the durability properties of concrete. Sustain. Environ. Res. 2018, 28, 25–31. [CrossRef]
59. Roshani, M.M.; Kargar, S.H.; Farhangi, V.; Karakouzian, M. Predicting the effect of fly ash on concrete’s mechanical properties by
ann. Sustainability 2021, 13, 1469. [CrossRef]
60. 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]
61. 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 2022, 14, 1074. [CrossRef] [PubMed]
Materials 2022, 15, 4209 18 of 18

62. Zhu, A.Y.; Ahmad, W.; Ahmad, N.I.; Vatin, A.M.; Mohamed, D.F. Predicting the Splitting Tensile Strength of Recycled Aggregate
Concrete Using Individual and Ensemble Machine Learning Approaches. Crystals 2022, 12, 569. [CrossRef]
63. Ahmad, A.; Ahmad, W.; Chaiyasarn, K.; Ostrowski, K.A.; Aslam, F.; Zajdel, P.; Joyklad, P. Prediction of geopolymer concrete
compressive strength using novel machine learning algorithms. Polymers 2021, 13, 3389. [CrossRef] [PubMed]
64. Leinweber, D.J. Stupid data miner tricks: Overfitting the S&P500. J. Investig. 2007, 16, 15–22.
65. Kohavi, R. A Study of Cross-Validation and Bootstrap for Accuracy Estimation and Model Selection. Int. Jt. Conf. Artif. Intell.
1995, 30, 133.
66. Farooq, F.; Ahmed, W.; Akbar, A.; Aslam, F.; Alyousef, R. Predictive modeling for sustainable high-performance concrete from
industrial wastes: A comparison and optimization of models using ensemble learners. J. Clean. Prod. 2021, 292, 126032. [CrossRef]
67. Ahmad, A.; Ostrowski, K.A.; 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]
68. Leung, K.W.K.; Pan, Z.L.; Warner, D.H. Kohn-Sham density functional theory prediction of fracture in silicon carbide under
mixed mode loading. Model. Simul. Mater. Sci. Eng. 2016, 24, 035004. [CrossRef]
69. Leung, K.W.K.; Pan, Z.L.; Warner, D.H. Atomistic-based predictions of crack tip behavior in silicon carbide across a range of
temperatures and strain rates. Acta Mater. 2014, 77, 324–334. [CrossRef]
70. Saroukhani, S.; Nguyen, L.D.; Leung, K.W.K.; Singh, C.V.; Warner, D.H. Harnessing atomistic simulations to predict the rate at
which dislocations overcome obstacles. J. Mech. Phys. Solids 2016, 90, 203–214. [CrossRef]
71. Kiani, S.; Leung, K.W.K.; Radmilovic, V.; Minor, A.M.; Yang, J.M.; Warner, D.H.; Kodambaka, S. Dislocation glide-controlled
room-temperature plasticity in 6H-SiC single crystals. Acta Mater. 2014, 80, 400–406. [CrossRef]
72. Ilawe, N.V.; Zimmerman, J.A.; Wong, B.M. Breaking badly: DFT-D2 gives sizeable errors for tensile strengths in palladium-hydride
solids. J. Chem. Theory Comput. 2015, 11, 5426–5435. [CrossRef] [PubMed]