Modelling
Modelling
net/publication/371237340
CITATIONS READS
2 87
2 authors:
All content following this page was uploaded by Sankar Boomibalan on 19 June 2023.
1. Introduction
Concrete consumption is rapidly increasing worldwide and raising concerns over the continuous extraction of raw materials.
Use of alternative sustainable materials such as fly ash, blast-furnace slag, silica fume and metakaolin as a partial replacement
for cement can have environmental benefits such as reduced CO2 emission during cement production and also can increase
the service life of concrete structures [1,2]. Modern concrete essentially includes Supplementary cementitious materials (SCMs)
in the mix composition, as it improves the concrete’s early strength development capacity and durability towards aggressive
environments [3–5].
∗ Corresponding author.
E-mail addresses: [email protected] (B. Sankar), [email protected] (P. Ramadoss).
https://doi.org/10.1016/j.rico.2023.100241
Received 23 December 2022; Received in revised form 17 May 2023; Accepted 27 May 2023
Available online 1 June 2023
2666-7207/© 2023 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license
(http://creativecommons.org/licenses/by/4.0/).
B. Sankar and P. Ramadoss Results in Control and Optimization 12 (2023) 100241
Recently, there has been a significant increase in demand for such High-performance concrete (HPC) concerning sustainability.
Moreover, the design process of HPC is extremely complicated when compared with ordinary concrete. The precise prediction of
the mechanical properties of HPC for optimal proportioning of ingredients is given paramount importance in construction sectors;
to maintain proper workability, durability and strength [6]. Compressive strength is an exceptional mechanical property that can
pretty much suffice the quality of the cement concrete. Hence, various researchers have attempted mixture proportioning based on
empirical models and regression analysis with compressive strength as the target output. Accurate prediction of strength properties of
concrete becomes almost impossible due to the large number of parameters involved and their complex non-linear interactions [7].
The majority of pioneering strength models were single variable formulas based on water-to-cement ratio only [8–10], this was
because of a logical theory that for a particular w/c ratio, the varying paste volume has a negligible effect in terms of strength
when mixed with same volume fractions of aggregate. However, in later years, a considerable number of experimental results
contradicted this general belief stating that binder content and SCM content also have a considerable effect on the mechanical
properties of concrete [11–19]. Also, the strength of concrete is dependent on its time of maturity, and time is the most crucial
factor in any construction project for moving forward in levels of construction phases. Consequently, SCMs are added in order for
the concrete to gain early strength for quick removal of formwork. In doing so, the constructions can be at high risk of catastrophe
due to inaccurate predictions in the strength gain of in-situ concrete. This confirms the necessity to develop accurate empirical
models to predict the strength gain of concrete [20,21].
Advancements in soft computing techniques have found their way into the field of civil engineering in the last decade via
optimization and prediction models. Some of the widely popular techniques used are Artificial Neural Networks (ANN) [38–41],
Support vector machines (SVM) [16,42], Particle swarm optimization (PSO) [43,44], KNN [42,45] Genetic expression programming
(GEP) [46], M5P Model tree [47–49], and Multivariate adaptive regression splines (MARS) [46,48]. Even though some of these
techniques have higher accuracy in prediction, they are not formula based nor easily comprehensible. The lack of simplistic models
has made it necessary to look for more reliable approaches for the mixture design of HPC. Multiscale statistical approach and
response surface method are one such approach that provide representational models and have advantages over black-box models
such as ANN. Mohammed et al. [50] developed multiscale models including LR, NLR, MLR, M5P tree and ANN to predict compressive
strength of fly ash-based SCC and observed that MLR and M5P models perform well in terms of RMSE and MAE values. Kumar [51]
used RSM to optimize lightweight concrete containing micro fine waste to obtain better hardening properties. Khudhair et al. [52]
formulated RSM model to predict compressive strength of HPC with water reducers and setting accelerators.
In recent years, there has been a steady increase in the utilization of calcined clay as SCM in concrete. MK is one such form
of calcined clay obtained from the calcination of kaolin clay through heat treatment at 600–800 ◦ C. MK is utilized for different
performance requirements based on its fineness [53]. MK is reported to have high pozzolanic reactivity associated with increased
rate of strength gain [54]. A noticeable increase in compressive strength, splitting tensile strength, modulus of elasticity and flexural
strength was observed for MK replacement level of 0%–20%, a further increase showed negative effect on the mentioned properties
of concrete [55,56]. Kadri et al. [57] reported escalated heat of hydration and high early strength development due to MK addition.
Jaya et al. [58] investigated the effect of MK addition in geopolymer concrete and noticed refinement in pore structure with enhanced
thermal and mechanical properties. Earlier studies have attempted to predict the compressive strength of concrete containing MK
using soft computing techniques such as ANN and GEP [46]. Ahmed et al. [59] used statistical techniques such as MLR, NLR and
M5P tree to develop prediction models for predicting the compressive strength of high-strength concrete containing MK.
Also, the influence of compressive strength ranges of concrete on the performance of prediction models were focused on some
of the studies [60]. It is very well known that composition of materials varies with different strength ranges of concrete. As the
compressive strength of concrete is in direct relation with other important properties such as permeability and durability of concrete
[61]; it is important to check the validity of the prediction models for various strength ranges. Emad et al. [62] in their study, used
metamodel techniques such as NLR, pure quadratic and ANN to predict the compressive strength of Ultra-high performance fiber
reinforced concrete (UHPFRC), and the models so developed were validated with experimental results of studies containing Normal
strength concrete (NSC). Ibrahim et al. [63] developed prediction models such as LR, NLR, pure quadratic and full quadratic model
to predict the compressive strength of concrete containing recycled aggregate for a wide range of strength ranging between 8.4 to
78.4 MPa.
In most cases, the targeted compressive strength of a concrete mix has influence from factors such as w/b ratio, cement content,
SCM content, porosity/air content in the paste, properties of aggregate, aggregate content and curing time [63]. Nath and Sarker [64]
conducted studies on geopolymer concrete containing GGBS and FA and reported that binder content, GGBS content, and type of
alkaline solution as influential parameters. Al-Alaily and Hassan [65] in their study used RSM technique to optimize the compressive
strength and chloride permeability of MK concrete with binder content, w/b ratio and MK content as independent variables. Popovics
and Ujhelyi [66] in their study observed that the strength of hardened concrete depends on total porosity and air content in the
cement paste. Besides, the failure in hardened concrete occurs more often in the Interfacial transition zone between paste and
aggregate, therefore De britto et al. [67] argues that the type and quality of aggregate do affect the strength of concrete, which is
also relative to the quality of the cement paste bonded with the aggregate. Hence, majority of the past studies agree on the fact that
the cement paste controls a substantial part of the strength of hardened concrete.
In this study, the w/c ratio, cement content, metakaolin content and SiO2 /Al2 O3 are taken as independent influencing factors
to predict the 28-day compressive strength of HPC. For this purpose, 152 experimental datasets containing mentioned input and
output variables were collected from published literature. Detailed information on the collected data is mentioned in Table 1. The
aim of the present study is to develop meaningful and reliable models to predict the compressive strength of HPC. Also, a critical
comparison of all the developed models is established.
2
B. Sankar and P. Ramadoss Results in Control and Optimization 12 (2023) 100241
Table 1
Summary of the collected data from literature.
References Binder Water/Binder Metakaolin Cement Metakaolin Specimen type, size 28-day compressive
(kg/m3 ) w/b (%) (mm) strength (MPa)
SiO2 Al2 O3 SiO2 Al2 O3
(%) (%) (%) (%)
[3] 500 0.34 0–30 21.25 4.95 52.1 44.7 Cube 75.3–78.7
(100 × 100 × 100)
[16] 400 0.35–0.45 0–20 21.32 3.83 69–76.7 17.4– Cube 45–67
24.46 (100 × 100 × 100)
[17] 426 0.38 0–15 – – 48.45 44.17 Cylinder 47.8–69.7
(100 × 200)
[22] 500 0.3 0–20 21 5.9 53.2 43.9 Cube 96.5–116.8
(100 × 100 × 100)
[23] 435 0.4 0–15 21.25 4.33 52 46 Cube 48.4-61.5
(100 × 100 × 100)
[18] 435–564 0.25–0.35 0–25 – – 61.47 17.55 Cube 41–72.7
(100 × 100 × 100)
[24] 450 0.4 0–10 19.3 5.62 52.68 36.34 Cube 62-67.1
(150 × 150 × 150)
[12] 547–589 0.28–0.35 0–10 20.07 4.49 53 48.8 Cylinder 88.1–112.7
(100 × 200)
[25] 450 0.28 0–15 20.69 4.72 51.6 41.3 – 87–103.5
[26] 512 0.32 0–50 – – 56.26 43.93 Cylinder 49.85-63.06
(100 × 200)
[19] 400–460 0.4–0.32 0–15 31.38 3.9 96.88 0.26 Cube 44.52–55.07
(100 × 100 × 100)
[15] 462 0.35 0–15 21 5.2 51.2 45.3 Cylinder 46.7–74.26
(100 × 200)
[27] 425–500 0.28–0.35 0–15 20.99 6.19 57.4 35.26 Cube 83.5–103.6
(100 × 100 × 100)
[28] 385 0.4 0–10 20.1 4.51 51.34 41.95 Cylinder 36.4–39.9
(100 × 200)
[29] 400 0.35–0.4 0–15 21.32 3.83 74.3 17.8 Cube 33–64.6
(100 × 100 × 100)
[30] 550 0.25–0.3 0–20 19.8 4.2 52.9 41.9 Cube 104-129.7
(100 × 100 × 100)
[31] 590 0.33 0–15 – – – – Cube 67–74.3
(100 × 100 × 100)
[32] 450 0.3 0–20 19.6 7.3 53.2 43.9 Cube 104–125.9
(100 × 100 × 100)
[33] 450 0.4 0–10 – – 73.45 18.04 Cube 57.5–62.9
(100 × 100 × 100)
[34] 435 0.45 0–20 21.46 5.55 52.1 42.8 Cube 31.2–38.5
(100 × 100 × 100)
[35] 550 0.3 0–15 21.07 4.61 59.93 27.04 80.3–87
[36] 450 0.28 0–15 20.69 4.72 51.6 41.3 Cube 86.7–104
(100 × 100 × 100)
[37] 600 0.25 0–15 19.37 3.92 57.55 39.45 Cube 85–102
(100 × 100 × 100)
2. Methodology
In order to develop robust prediction models to predict the 28-day compressive strength of HPC containing metakaolin, most
appropriate literatures were collected that contained the necessary input variables chosen for the study. The collection of literature
was limited to high-strength concrete and high-performance concrete containing metakaolin. The highly influential input parameters
selected from the literature were B-binder content (kg/m3 ), w/b-water to binder ratio, MK-metakaolin content (%) and SiO2 /Al2 O3 -
silicon dioxide to aluminium oxide ratio. The datasets were organized and randomly split into training set and testing set. The input
and output variables were analysed using descriptive statistical measures. Five different models were developed using different
statistical techniques namely LR, MLR, and RSM from the training datasets, followed by validation using the testing dataset and
3
B. Sankar and P. Ramadoss Results in Control and Optimization 12 (2023) 100241
then checked for performance criterion such as R, MAE, RMSE and SI. Finally, a sensitivity analysis was conducted to determine
the most influential parameter among the input variables that govern the target output variable. The flowchart shown in Fig. 1
illustrates the methodology used in the present study.
3. Dataset
Collection of reliable datasets is an essential part in a modelling study. The quality and number of datasets can affect the precision
of a predicting model. A total of 152 concrete mixture details were collected with 28-day compressive strength of concrete as the
target output variable. Majority of the specimen sizes from the experimental datasets were 100 × 100 × 100 mm cubes. Hence,
the remaining compressive strength data with different specimen sizes were converted to 100 × 100 × 100 mm cubes using the
conversion factor provided by Elwell and Fu [68], in order to provide a meaningful comparison between the output variables. The
final collection of datasets was randomly split into training and testing datasets. The splitting ratio used in this study for training
and testing datasets is 7:3 (107:45).
4
B. Sankar and P. Ramadoss Results in Control and Optimization 12 (2023) 100241
4. Statistical assessment
In this section, descriptive statistical evaluation was performed for the selected parameters from the collected dataset. The
evaluation criteria included are Standard Deviation (SD), Variance (VAR), Interquartile range (IQR), Skewness (SKEW) and Kurtosis
(KUR). Standard deviation and variance represent the spread of the data in a given dataset, the interquartile range is the difference
between upper half and lower half of the distributed data from the median, skewness is the level of distortion of symmetry in the
distribution and kurtosis indicates the sharpness of the central distribution relative to the tail distribution. The histograms of the
input variables (B, w/b, MK, SiO2 /Al2 O3 ) and output variable (CS) are shown in Fig. 2. The descriptive statistics of input variables
and output variable are presented in Table 2. The findings of the statistical evaluation of all the parameters are discussed below.
5
B. Sankar and P. Ramadoss Results in Control and Optimization 12 (2023) 100241
Table 2
Descriptive statistics of parameters.
Parameters N Mean Standard Variance Minimum Q1 Median Q3 Maximum Interquartile Skewness Kurtosis
deviation (SD) (VAR) range (IQR) (SKEW) (KUR)
Input variables
Binder 152 475.34 58.75 3451.5 385 435 458.5 512 600 77 0.44 −0.89
(kg/m3 )
Water/Binder 152 0.33 0.054 0.0029 0.25 0.3 0.341 0.395 0.45 0.095 0.29 −0.69
(W/B)
Metakaolin 152 9.73 8.284 68.617 0 5 10 15 50 10 1.24 3.5
(%)
SiO2 /Al2 O3 152 3.93 2.158 4.658 1.097 2.713 3.503 4.57 12.28 1.857 2.07 4.98
Output variable
CS (MPa) 152 70.63 24.06 579 31.2 51.35 64 87 129.7 35.65 0.66 −0.43
The amount of cement content without replacing metakaolin is referred as binder content. Majority of the research works from
which the datasets were collected included ordinary Portland cement as binder material with specific gravity ranging between
3.05–3.2, which is in compliance with ASTM C 150 standards. The quantity of binder content in the collected 152 datasets ranges
between 435 kg/m3 to 600 kg/m3 , with a mean of 475.34 kg/m3 . Other statistical parameters, including SD, VAR, IQR, SKEW and
KUR, were 58.75 kg/m3 , 3451.51, 77 kg/m3 , 0.44 and −0.89. A positive skew (right-skew) indicates left-side clustering of data
from the mean. While a negative kurtosis indicates a squished distribution with fat tails. The histogram of binder content is shown
in Fig. 2.
The w/b ratio in the dataset ranges between 0.25 and 0.45, respectively. Based on the statistical investigation, the mean and
standard deviation was found to be 0.338 and 0.054. Other parameters such as VAR, IQR, SKEW and KUR were 0.0029, 0.095, 0.29
and −0.69. A bigger IQR range of w/b distribution (nearly 70%) indicates numerous clustering over a large number of w/b values.
The skew value indicates a slight distortion in symmetry due to left-side clustering. The w/b distribution has a negative kurtosis
with a moderately squished distribution. The histogram of w/b ratio is shown in Fig. 2.
Metakaolin content in the collected literature ranges between 0% and 50%, respectively. The statistical parameters such as
mean, SD, VAR, IQR, SKEW and KUR were 9.73%, 8.284%, 68.617, 10%, 1.24 and 3.5. The IQR of MK has a smaller range of 20%,
since most of the collected literature contained MK range between 0 to 20% only. Skewness indicates a highly distorted data with
poor symmetry due to left-side clustering. Since the majority of MK data falls within the range of 0 and 20%, the distribution was
somewhat sharp with positive kurtosis. The histogram of MK is shown in Fig. 2.
From the collected dataset, the SiO2 and Al2 O3 percentages for the cement were found in the range between 18 to 31.38% and
3 to 7.3%, whereas for metakaolin, it was found to be 50 to 75% and 17 to 48.8%. The total SiO2 and Al2 O3 percentages present
in the cementitious matrix (cement & metakaolin combined) can be calculated from Eqs. (1) and (2).
100 − 𝑀𝐾 𝑀𝐾
SiO2 = 𝐶𝑒𝑚𝑒𝑛𝑡SiO2 × + 𝑀𝑒𝑡𝑎𝑘𝑎𝑜𝑙𝑖𝑛SiO2 × (1)
100 100
100 − 𝑀𝐾 𝑀𝐾
Al2 O3 = 𝐶𝑒𝑚𝑒𝑛𝑡Al2 O3 × + 𝑀𝑒𝑡𝑎𝑘𝑎𝑜𝑙𝑖𝑛Al2 O3 × (2)
100 100
where,
SiO2 and Al2 O3 represents the total percentage of silicon dioxide and aluminium oxide present in the cementitious matrix.
𝐶𝑒𝑚𝑒𝑛𝑡SiO2 , 𝐶𝑒𝑚𝑒𝑛𝑡Al2 O3 and 𝑀𝑒𝑡𝑎𝑘𝑎𝑜𝑙𝑖𝑛SiO2 , 𝑀𝑒𝑡𝑎𝑘𝑎𝑜𝑙𝑖𝑛Al2 O3 represents the percentages of SiO2 and Al2 O3 present in cement and
metakaolin respectively, and MK is the percentage of metakaolin.
From the statistical analysis of SiO2 /Al2 O3 ratio from the datasets, the parameters such as mean, SD, VAR, IQR, SKEW and KUR were
found to be 3.93, 2.158, 4.658, 1.857, 2.07 and 4.98. A higher kurtosis with sharp canter and flattened tail distribution indicates
the clustering of data within a short range. The histogram of SiO2 /Al2 O3 from the dataset is shown in Fig. 2.
6
B. Sankar and P. Ramadoss Results in Control and Optimization 12 (2023) 100241
The collected dataset contained the compressive strength of metakaolin concrete within the HPC range. The maximum and
minimum value for CS was 31.2 and 129.7 MPa respectively. After performing statistical investigation, the mean, SD, VAR, IQR,
SKEW and KUR were found to be 70.63, 24.06, 579, 35.65, 0.66 and −0.43. Clustering was observed over a large range of data,
with distribution having fat tails and blunted centres. A slight distortion towards the left side of the distribution was observed. The
histogram of CS from the dataset is shown in Fig. 2.
5. Modelling
In this section, models were developed using different statistical techniques such as LR, MLR, NLR and RSM using the training
datasets. The models thus developed were compared to determine the best predictor of the output variable (CS). The validation of
the proposed model was done using the testing dataset. Below is a detailed summary of the modelling techniques used in this study.
Linear regression (LR) is the most common statistical technique to predict the response of an output variable with a single or
several input variables [69,70]. The mathematical expression for the linear regression model is illustrated in Eq. (3).
∑
𝑛
𝑌 = 𝛽0 + 𝛽𝑖 𝑋𝑖 (3)
𝑖=1
where Y and X is the output and input parameter, 𝛽0 is a constant and 𝛽i are coefficients that vary from i = 1,2,3. . . n. The modified
equation of the linear model after inputting the parameters used in the study can be expressed as shown in Eq. (4).
where CS, B, w/b, MK and SiO2 /Al2 O3 are 28-day compressive strength (MPa), binder content (kg/m3 ),
metakaolin content (%) and
silicon dioxide to aluminium oxide ratio. The coefficients 𝛽0 , 𝛽1 , 𝛽2 , 𝛽3 and 𝛽4 are model parameters obtained via the least squares
analysis method.
Multi-logistic regression (MLR) is generally used to predict the outcome of the dependent variable having more than two
categorical stages. In the present study, compressive strength of concrete is the target dependent variable having multiple levels
of outcome influenced by ‘n’ number of input variables. Hence, MLR is best suited to process the difference between independent
variables and predictor variable. The general form of MLR is given in Eq. (5).
𝛽 𝛽 𝛽
𝑌 = 𝛽1 𝑋1 2 ∗ 𝑋2 3 … 𝑋𝑛 𝑛 (5)
where Y and X are the output and input parameters. The modified form of Eq. (5) using the input parameters used in the present
study is given in Eq. (6). In Eq. (6), CS, B, w/b, MK and SiO2 /Al2 O3 represents 28-day compressive strength (MPa), binder content
(kg/m3 ), metakaolin content (%) and silicon dioxide to aluminium oxide ratio. 𝛽1,2...n are model parameters obtained via the least
squares analysis method. However, the above model has a limitation and can only be used for MK content >0%.
A pure linear relation cannot always exist between input variables and output variables, in most cases, the model needs to be
modified based on the nonlinearity of the variables [71,72]. Hence in this study, in addition to linear and multi-logistic regression,
a nonlinear regression model is also developed to determine the interrelation between input and output variables. The below Eq. (7)
provides the general expression for a nonlinear regression model. The modified form of Eq. (7) using the input parameters from the
present study is expressed in Eq. (8).
[ ] [ ]
𝛽 𝛽 𝛽𝑛 𝛽 𝛽 𝛽𝑛−1 𝛽
𝑌 = 𝛽1 𝑋1 2 ∗ 𝑋2 3 … 𝑋𝑛−1 + 𝛽𝑛1 𝑋1 𝑛2 ∗ 𝑋2 𝑛3 … 𝑋𝑛−1 × 𝑋𝑛 𝑛 (7)
𝐶𝑆 = 𝛽1 (𝐵)𝛽2 (𝑤∕𝑏)𝛽3 (SiO2 ∕Al2 O3 )𝛽4 + 𝛽5 (𝐵)𝛽6 (𝑤∕𝑏)𝛽7 (SiO2 ∕Al2 O3 )𝛽8 (𝑀𝐾)𝛽9 (8)
where Y and X are the output and input parameters. In Eq. (8), CS, B, w/b, MK and SiO2 /Al2 O3 represent 28-day compressive
strength (MPa), binder content (kg/m3 ), metakaolin content (%) and silicon dioxide to aluminium oxide ratio. 𝛽1,2...n are model
parameters obtained via the least squares analysis method.
7
B. Sankar and P. Ramadoss Results in Control and Optimization 12 (2023) 100241
Response surface methodology (RSM) is a statistical calibrating technique that estimates the interaction between categorical
factors (input parameters) and response factors (output parameters) in order to provide an ideal mathematical model [73]. The
idea behind RSM is, the input variables form the vertices of a cuboid, and the points with similar responses are connected using a
contour line, and the combination of such contour lines from all the interactions predict the target output variable. In most cases,
a second-order polynomial (Quadratic) function is performed for satisfactory prediction of compressive strength. However, in this
study, a quadratic (QRSM) and a higher-order polynomial function (HORSM) is fitted to correlate the independent factors with
the response for assessment in terms of accuracy and reliability. This method is more suitable when there are more number of
independent variables and can also be effectively utilized with a limited number of datasets.
In the training dataset, binder content (kg/m3 ), w/b ratio, MK content (%) and SiO2 /Al2 O3 are the categorical factors with multi-
levels and 28-day compressive strength is assigned as the response. The RSM model can be represented using the mathematical
expression given in Eq. (9).
∑
𝑛 ∑
𝑛 ∑
𝑛
( )
𝑌 = 𝛽0 + 𝛽𝑖 𝑋𝑖 + 𝛽𝑖𝑖 𝑋𝑖2 + 𝛽𝑖𝑗 𝑋𝑖 𝑋𝑗 … + 𝑒 𝑋𝑖 , 𝑋𝑗 , … , 𝑋𝑛 (9)
𝑖=1 𝑖=1 𝑖<𝑗
where Xi and Y denote independent factors and response, 𝛽0 is the intercept, 𝛽i , 𝛽ii and 𝛽ij are model coefficients that characterize
the linear, squared and interaction effect of the model, respectively. A commercially available software with statistical design and
analysis tool pack was used to develop the RSM models [74].
Statistical error metrics are necessary to determine the robustness and accuracy of a prediction model. In this section, the different
statistical evaluation criteria used to determine the reliability of models using the actual and predicted values of compressive
strength are coefficient of correlation (r), Mean absolute error (MAE), Root mean square error (RMSE) and Scatter index (SI) given
in Eqs. (10)–(13).
∑𝑛
𝑖=1 (𝐶𝑆𝑎𝑐𝑡,𝑖 − 𝐶𝑆𝑎𝑐𝑡 )(𝐶𝑆𝑝𝑟𝑒,𝑖 − 𝐶𝑆𝑝𝑟𝑒 )
𝑅= √ √ (10)
∑𝑛 2
∑𝑛 2
𝑖=1 (𝐶𝑆 𝑎𝑐𝑡,𝑖 − 𝐶𝑆𝑎𝑐𝑡 ) 𝑖=𝑛 (𝐶𝑆𝑝𝑟𝑒,𝑖 − 𝐶𝑆𝑝𝑟𝑒 )
√
√
√1 ∑ 𝑛
𝑅𝑀𝑆𝐸 = √ (𝐶𝑆𝑎𝑐𝑡,𝑖 − 𝐶𝑆𝑝𝑟𝑒,𝑖 )2 (11)
𝑁 𝑖=1
1 ∑|
𝑛
|
𝑀𝐴𝐸 = |(𝐶𝑆𝑎𝑐𝑡,𝑖 − 𝐶𝑆𝑝𝑟𝑒,𝑖 )| (12)
𝑁 𝑖=1 | |
𝑅𝑀𝑆𝐸
𝑆𝐼 = (13)
𝐶𝑆
In the above equations, 𝐶𝑆𝑎𝑐𝑡 and 𝐶𝑆𝑝𝑟𝑒 represents the actual and predicted compressive strength, 𝐶𝑆𝑎𝑐𝑡 and 𝐶𝑆𝑝𝑟𝑒 represents an
average of actual and predicted compressive strength, N is the total number of datasets. The values of MAE, RMSE and SI should be
close to zero, whereas, for R, the value should be close to 1. In the case of SI, a value of <0.1, 0.1–0.2, 0.2–0.3 and >0.3 indicates
the excellent, good, fair and poor performance of the models.
Sensitivity analysis is carried out to determine the extent of influence of a single input variable on the performance of target
output variable. For this purpose, every time an input variable is removed from the training set while developing the model and
checked for performance metrics such as R, RMSE and MAE [75]. The lower the performance of the model, higher is the influence of
the removed parameter. Although such method of analysing the sensitivity of parameters is consistent, in this study cosine amplitude
method is employed. The mathematical expression for the cosine amplitude method is given in Eq. (14)
∑𝑛
𝑖,𝑘=1 (𝑋𝑖𝑘 × 𝑌𝑖𝑘 )
𝑅𝑖𝑗 = √ (14)
∑𝑛 ∑𝑛
𝑖,𝑘=1 𝑋𝑖𝑘 𝑖,𝑘=1 𝑌𝑖𝑘
In the above equation, Xi and Yi are input and output parameter, n and k represents the numbers and levels of input and output
variables in the dataset. The value of 𝑅𝑖𝑗 will vary between 0 to 1. A magnitude of 𝑅𝑖𝑗 equal to 0 corresponds to zero influence of
the input variable on the target variable. However, a magnitude of 1 or close to 1 indicates a strong influence of the variable to
attain the target variable.
8
B. Sankar and P. Ramadoss Results in Control and Optimization 12 (2023) 100241
LR model was developed using 107 training datasets and was validated with 45 testing datasets. The weight of the input
parameters determines the influence of each parameter to attain the target compressive strength. The coefficients of the model
were arrived by reducing the sum of squares of the errors using the least square method. The proposed LR model for predicting the
28-day compressive strength of metakaolin incorporated HPC from the current study is presented in Eq. (15).
𝐶𝑆 = 73.708 + 0.0747 (𝐵) − 95.225 (𝑤∕𝑏) − 0.3705 (𝑀𝐾) − 1.9795 (SiO2 ∕Al2 O3 ) (15)
From the equation, it is evident that w/b has a greater influence over the prediction of compressive strength. This matches with
the experimental results from the previous literature [76–78]. Mohammed et al. [79] proposed linear regression model to predict
the compressive strength of high-volume fly ash concrete using factors such as w/b ratio, cutting time and fly ash content and
observed similar results with w/b ratio having the highest effect on the compressive strength of concrete. The comparison between
the actual compressive strength of HPC containing metakaolin and the predicted compressive strength values from the LR model
for the training and testing datasets are illustrated in Fig. 3(a). From the figure, it is noticeable that the predicted values are more
scattered, and the trendline is more inclined away from the line of zero variation.
The proposed MLR model for predicting the 28-day compressive strength of metakaolin incorporated HPC is presented in Eq. (16).
The model parameters derived using the least square method indicates that w/b and B are the most influential input variables
affecting the target compressive strength and are also in agreement with the earlier studies [80–82]. Similarly, in a study containing
concrete with steel slag aggregate, the major factor seen affecting the compressive strength of concrete from the MLR model
developed was w/b ratio, cement content, curing time and steel slag aggregate content [48]. The comparison between the actual
compressive strength and the predicted compressive strength values from the MLR model for the training and testing datasets, are
illustrated in Fig. 3(b). From the analysis of the figure, a slight convergence with the clustering of data points was observed. The
more inclined trendline, away from the line of zero variation, indicates the poor performance of the model for higher compressive
strength values.
The proposed equation for the NLR model, developed using training datasets from the current study, is provided below
The model parameters for the NLR model given in Eq. (17) are determined using the least square method. The most influential
parameters affecting the target compressive strength were found to be B and w/b. The NLR model indicates that an increase in
binder content and a decrease in the w/b ratio will result in an increment of compressive strength; the same has been noticed
from past studies [59,83,84]. From the model, it was observed that the effect of SiO2 /Al2 O3 is very well pronounced and it
highly influences the compressive strength of concrete. A similar observation was made by Abdalla and Mohammed [85] in
their study containing cement kiln dust, where an increase in SiO2 % increased the long-term compressive strength of concrete.
A higher accuracy was observed in the NLR model when compared with LR and MLR models [86], which is in agreement with the
observations from the present study. The comparison between the actual compressive strength from the experimental datasets and
the predicted compressive strength values from the NLR model for the training and testing datasets are illustrated in Fig. 3(c). The
noticeable deductions from the figure are, more converged data points and less inclined trendline when compared to LR and MLR
models.
The QRSM and HORSM models developed using the RSM technique is expressed in the following form given in Eqs. (18) and
(19). The developed models account for linear, square, cubic and interaction effects of input variables on the response variable. The
surface plot of the interactions of input variables from QRSM model is shown in Fig. 4.
9
B. Sankar and P. Ramadoss Results in Control and Optimization 12 (2023) 100241
Fig. 3. Relationship between actual and predicted compressive strength for training and testing datasets (a) LR (b) MLR (c) NLR (d) QRSM (e) HORSM.
10
B. Sankar and P. Ramadoss Results in Control and Optimization 12 (2023) 100241
Table 3
Coefficients of QRSM and HORSM model.
Factors QRSM HORSM
Coefficient P-value Coefficient P-value
Intercept 706.6788 <0.0001 −7487.278 <0.0001
A-Binder −0.96713 0.0015 7.27207 0.053
B-W/C −2309.15 <0.0001 1.08E+05 0.0651
C-MK 10.4161 0.0298 −584.4878 0.9345
D-SiO2 /Al2 O3 −12.8256 0.0021 −3521.712 0.6692
AB 1.61178 0.001 −215.4152 <0.0001
AC −0.01425 0.0007 1.70961 0.0261
AD 0.000291 <0.0001 10.41243 0.9992
BC −8.99108 0.0005 1183.584 0.0479
BD 30.98148 <0.0001 7009.5218 0.6287
CD 0.051488 0.0079 1.11334 0.0163
A2 0.000651 0.0068 0.008804 <0.0001
B2 1812.056 0.0206 −2.25E+05 0.0002
C2 −0.023 <0.0001 −0.547368 <0.0001
D2 −0.13101 0.0166 −13.02174 0.0387
ABC – – −1.78561 0.0075
ABD – – −10.9045 0.0686
ACD – – 0.002431 0.5354
BCD – – −8.31573 0.9805
A2 B – – 0.103947 0.4051
A2 C – – −0.001226 0.6592
A2 D – – −0.007291 0.01
AB2 – – 252.41538 0.0452
AC2 – – 0.000813 0.2337
AD2 – – 0.003874 0.1225
B2 C – – −531.589 0.0369
B2 D – – −3411.001 0.4482
BC2 – – 0.182529 0.1095
BD2 – – 36.53844 0.7529
C2 D – – 0.017149 0.5421
CD2 – – 0.09571 0.0587
A3 – – −6.24E−06 0.2467
B3 – – 1.24E+05 0.0036
C3 – – 0.000453 0.6538
D3 – – −0.236407 0.0462
Model fit summary
Source Sequential P-value Adjusted R2 Predicted R2
Quadratic <0.0001 0.886 0.828 Suggested
Cubic 0.0002 0.972 0.443 Suggested
+ 𝛽44 𝐷2 + 𝛽123 𝐴𝐵𝐶 + 𝛽124 𝐴𝐵𝐷 + 𝛽134 𝐴𝐶𝐷 + 𝛽234 𝐵𝐶𝐷 + 𝛽112 𝐴2 𝐵 + 𝛽113 𝐴2 𝐶 + 𝛽114 𝐴2 𝐷 + 𝛽122 𝐴𝐵 2 + 𝛽133 𝐴𝐶 2
+ 𝛽144 𝐴𝐷2 + 𝛽223 𝐵 2 𝐶 + 𝛽224 𝐵 2 𝐷 + 𝛽233 𝐵𝐶 2 + 𝛽244 𝐵𝐷2 + 𝛽334 𝐶 2 𝐷 + 𝛽344 𝐶𝐷2 + 𝛽111 𝐴3 + 𝛽222 𝐵 3 + 𝛽333 𝐶 3 + 𝛽444 𝐷3 (19)
The coefficients of the equations are provided in Table 3. The fit summary of the models suggests that both QRSM and HORSM has
higher accuracy in predicting the compressive strength of concrete. Even though the adjusted R2 of HORSM is greater than QRSM,
the predicted R2 is radically low for the HORSM model, which indicates the ineffectiveness of the HORSM model in predicting the
compressive strength for lower and upper bound values of the input variables. The relationship between measured and predicted
compressive strength values of full quadratic model from an earlier study [63] showed 35%–40% residual error, whereas the
quadratic RSM model from the present study showed a maximum of 20% error. The charts plotted between actual compressive
strength and predicted compressive strength from the HORSM model for training and testing datasets display more scattered data
points with residual error greater than ±20%, nevertheless the trendline is closer to the line of zero variation, shown in Fig. 3(e).
For QRSM model, the predictive accuracy is greater than 80%, with a maximum number of data points lying within ±20% error
line, shown in Fig. 3(d). In comparison to all the models developed in the present study, QRSM is more reliable in predicting the
28-day compressive strength of metakaolin incorporated HPC.
The statistical error metrics such as R, RMSE and MAE for all the developed models were compared using boxplot diagrams
shown in Figs. 5–7. The boxplot of each model contains the details of the median, interquartile range and whiskers of the error
range. In the figures, a and b represent boxplots of training and testing datasets. For the training datasets, the MLR and LR models
showed the lowest R values of 0.67 and 0.76, whereas QRSM and HORSM showed the highest R values of 0.89 and 0.86. For the
testing datasets, QRSM, HORSM and NLR models showed a higher correlation of 0.88, 0.90 and 0.92, shown in Fig. 5. The RMSE
11
B. Sankar and P. Ramadoss Results in Control and Optimization 12 (2023) 100241
Fig. 4. Surface plot of the interactions of input variables from QRSM model.
of developed models for the training and testing datasets are shown in Fig. 6. From the figure, it is seen that the nonlinear models
and RSM models exhibit lower RMSE values. The mean RMSE of testing datasets for LR, MLR, NLR, HORSM and QRSM models
were 14.3 MPa, 12.6 MPa, 10.4 MPa, 8.4 MPa and 6.5 MPa. A similar trend was observed for MAE of testing datasets with mean
MAE values for LR, MLR, NLR, HORSM and QRSM models were 12.3 MPa, 9.2 MPa, 7.6 MPa, 7 MPa and 5.2 MPa, shown in Fig. 7.
Apart from the comparison between models from the present study, existing models from previously published studies were used to
validate the accuracy of the QRSM model. The models used for validation are the M5 Prime tree model from Ahmed et al. [59] and
12
B. Sankar and P. Ramadoss Results in Control and Optimization 12 (2023) 100241
Fig. 5. Boxplot of Coefficient of correlation (R) (a) Training datasets (b) Testing datasets.
Fig. 6. Boxplot of Root mean squared error (RMSE) Training datasets (b) Testing datasets.
RSM model from Al-Alaily and Hassan [65], both are strength prediction models of high-strength concrete containing metakaolin.
Fig. 8 provides a more detailed illustration of improvement in the performance of QRSM model over existing models. A comparative
analysis of the prediction of experimental results from Dvorkin et al. [87] is presented in Table 4. From the table, it is seen that
QRSM model has the least relative error with respect to the experimental results compared to all other models. A comparison of
scatter index (SI) plot for the developed models is shown in Fig. 10. The results of SI value declare LR, MLR and NLR as fair models
with values falling between 0.3 and 0.2, whereas HORSM and QRSM as good models to predict compressive strength with values
in the range of 0.1–0.2. Based on the results of different performance metrics used in the study, the QRSM model shows an overall
excellent performance in compressive strength prediction.
Also, the effectiveness of the prediction models must be evaluated for different strength ranges of concrete [88]. For this purpose, 105
data containing compressive strength values of Normal strength concrete (NSC) with strength range of 19–55 MPa and Ultra-high
performance concrete (UHPC) with strength range of 101–158 Mpa (Table 5) were collected from past literature. The relationship
between measured and predicted compressive strength values of NSC and UHPC are shown in Fig. 9. From the figure, it is seen that
the developed models perform well in the strength ranges of NSC and UHPC. The predicted values were well within the range of
the 20% error line for QRSM model. Maximum error was observed in the UHPC strength ranges for LR and MLR models.
Sensitivity analysis was done based on the cosine amplitude method, and the results are presented in Fig. 11. Binder content
and w/b are the crucial input parameters affecting the prediction of compressive strength. The results are in compliance with the
weightage of the input parameters from the developed models and from the previous studies. 𝑅𝑖𝑗 values of B, w/b and SiO2 /Al2 O3
were 0.98, 0.93 and 0.91, which are closer to 1, indicating strong influence of the respective parameters in predicting the compressive
strength.
Limitation for future work
The number of datasets utilized in the present study was rather small. Training the prediction models using a larger dataset will
increase the reliability and accuracy of the models. In future research, experimental studies on HPC containing MK can be used to
13
B. Sankar and P. Ramadoss Results in Control and Optimization 12 (2023) 100241
Fig. 7. Boxplot of Mean absolute error (MAE) Training datasets (b) Testing datasets.
Fig. 8. Improvement in the performance of QRSM model from the present study over M5P and RSM models from previous studies [59,65].
Table 4
Comparison of experimental values from Dvorkin et al. [87] and corresponding prediction values using different models.
S No. Input variables Models from the present study Existing models Experimental 28-day
compressive
strength (MPa)
Binder w/b MK (%) SiO2∕Al O
2 3 LR MLR NLR HORSM QRSM RSM [65] M5P [59]
(kg/m3 )
1 383 0.399 17.49 2.32 53.215 58.323 52.573 69.425 62.552 84.009 68.982 59
2 428 0.400 5.14 3.35 59.110 53.151 57.399 58.517 55.736 68.980 68.151 51.8
3 468 0.340 17.52 2.32 65.249 67.799 64.735 71.515 78.997 66.502 63.971 85
4 523 0.340 5.16 3.34 71.845 61.744 70.648 86.639 81.004 57.117 67.724 76.6
5 383 0.339 17.49 2.32 58.934 64.189 54.995 90.886 85.870 54.002 59.617 74.8
6 428 0.339 5.14 3.35 64.895 58.566 60.077 92.829 69.330 66.241 62.656 71
7 450 0.400 11.11 2.72 57.750 58.195 59.863 72.691 61.137 86.416 70.812 62
8 450 0.340 11.11 2.72 65.464 64.033 62.614 86.419 77.651 67.946 62.972 80.3
9 495 0.370 11.11 2.72 65.999 62.574 66.122 76.694 71.229 63.025 71.778 73
10 405 0.370 11.11 2.72 59.208 59.152 56.119 83.978 65.867 76.767 62.948 65.9
11 425 0.369 17.64 2.31 59.175 62.898 58.477 75.822 70.195 61.168 65.886 74.2
12 475 0.371 5.26 3.33 61.374 57.311 63.803 73.695 67.466 77.799 65.982 65.2
13 450 0.371 11.11 2.72 62.501 60.819 61.117 69.384 68.978 70.741 65.975 70.4
14 450 0.371 11.11 2.72 62.501 60.819 61.117 69.384 68.978 70.741 66.451 71
15 523 0.400 5.16 3.34 66.200 56.180 67.580 91.252 66.974 84.214 70.087 53.6
14
B. Sankar and P. Ramadoss Results in Control and Optimization 12 (2023) 100241
Table 5
Properties of concrete mixes collected from the literature.
Normal strength concrete (NSC) Ultra-high performance concrete (UHPC)
References Water/binder ratio Compressive strength (MPa) References Water/binder ratio Compressive strength (MPa)
[89] 0.44 42.2–54.9 [90] 0.15 138.8
[91] 0.49–0.66 29–48 [92] 0.15 121.59
[93] 0.45–0.5 34.4–41.8 [94] 0.16–0.26 110–156
[95] 0.4–0.42 24–46.3 [96] 0.13 112–125.4
[97] 0.5–0.9 20–42.7 [98] 0.21 101.2–120.4
[99] 0.5–0.8 19.5–51.2 [100] 0.12–0.14 140.4-145.26
[101] 0.16–0.2 103.1–158
Ranged between 0.4–0.9 19.5–54.9 Ranged between 0.12–0.26 101.2–158
Fig. 9. Comparison between measured and predicted compressive strength values of NSC and UHPC using the developed models.
support and validate the prediction models. The authors intend to use other statistical techniques such as ANN, SVM and M5 tree
to examine the accuracy of models by comparing with the present study. Also, some important factors may have been overlooked
while extracting the data. In future research, the effect of curing time, aggregate ratio and superplasticizer content must be included
as model parameters to verify whether any improvement can be observed in the precision of proposed models.
8. Conclusion
The investigations on modelling techniques to predict the compressive strength of HPC using the previously acquired data can
be used to arrive at the following conclusions.
1. A total of 152 datasets were collected from previous studies containing different mixture proportion specifications such as
w/b ratio, cement content, MK percentage as a partial replacement for OPC and SiO2 /Al2 O3 ratio in the cementitious material
were used as input parameters.
15
B. Sankar and P. Ramadoss Results in Control and Optimization 12 (2023) 100241
2. Statistical techniques such as LR, MLR, NLR and RSM were used for a novel formulation of prediction models to predict the
28-day compressive strength of HPC containing MK.
3. A comparison between LR, MLR and NLR models shows that the NLR model has higher accuracy in predicting the compressive
strength of concrete. This indicates the non-linearity of the relation between the input and output variables.
4. The performance criteria indicate that the QRSM model performed better than the rest of the models developed in the present
study. Statistical metrics such as R, RMSE, MAE and SI of the QRSM model for the testing datasets were 0.92, 6.5 MPa, 5.2
MPa and 0.08 respectively.
5. For the QRSM model, the predicted compressive strength values for training and testing datasets were well within the range
of ±20% error. The order that was decided upon for the provided models with greater efficiency and best suitability was
QRSM > HORSM > NLR > LR > MLR.
6. The effectiveness of the proposed models was evaluated in different strength ranges such as NSC and UHPC. From the results,
it is seen that the quadratic model using the RSM technique is more reliable in predicting wide ranges of compressive strength
of concrete with less residual error.
7. The results of the sensitivity analysis infer that the most effective parameters that have higher impact on the compressive
strength of HPC were binder content and w/b ratio, which are in agreement with the model parameters of the current study
and from previous studies.
8. The obtained result indicates that the effect of SiO2 /Al2 O3 ratio in the cementitious material is more significant than
presumed, and had higher impact on predicting the compressive strength of HPC.
9. The collective analysis deduces that the developed models can be effectively utilized for the 28-days compressive strength
prediction of HPC; the higher accuracy and predictive capability of the QRSM model can be utilized in field practices for
16
B. Sankar and P. Ramadoss Results in Control and Optimization 12 (2023) 100241
the preliminary design of mixes. The statistical methods used to develop the models can be extended over a wide variety of
concrete mixes containing different supplementary cementitious materials.
The authors declare that they have no known competing financial interests or personal relationships that could have appeared
to influence the work reported in this paper.
Data availability
Acknowledgement
The authors would like to express their thanks to All India Council for Technical Education (AICTE) for providing financial
assistance towards the research work.
References
[1] Mehta PK. Pozzolanic and cementitious by-products in concrete - another look. In: ACI spec publ 114. 1989, Published online.
[2] Mohan Malhotra V, Zhang MH, Read PH, Ryell J. Long-term mechanical properties and durability characteristics of high-strength/high-performance
concrete incorporating supplementary cementing materials under outdoor exposure conditions. ACI Struct J 2000;97(5):518–25. http://dx.doi.org/10.
14359/9284.
[3] AzariJafari H, Taheri Amiri MJ, Ashrafian A, Rasekh H, Barforooshi MJ, Berenjian J. Ternary blended cement: An eco-friendly alternative to improve
resistivity of high-performance self-consolidating concrete against elevated temperature. J Clean Prod 2019;223:575–86. http://dx.doi.org/10.1016/j.
jclepro.2019.03.054.
[4] Sankar B, Ramadoss P. Review on fiber hybridization in ternary blended high-performance concrete. Mater Today: Proc 2021. http://dx.doi.org/10.1016/
j.matpr.2021.01.366.
[5] Sankar B, Ramadoss P. Mechanical and durability properties of high strength concrete incorporating different combinations of supplementary cementitious
materials: A review. 2022, http://dx.doi.org/10.1007/978-981-16-4321-7_45.
[6] Sobhani J, Najimi M, Pourkhorshidi AR, 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(5):709–18. http://dx.doi.org/10.1016/j.conbuildmat.2009.10.037.
[7] Gao W, Karbasi M, Derakhsh AM, Jalili A. Development of a novel soft-computing framework for the simulation aims: a case study. Eng Comput 2019.
http://dx.doi.org/10.1007/s00366-018-0601-y, Published online.
[8] Abrams DA. Design of concrete mixtures. In: Bull 1, struct mater res lab lewis institute, Chicago. 1918, Published online https://www.ptonline.com/
articles/how-to-get-better-mfi-results.
[9] Walker S, Bloem DL. Effects of aggregate size on properties of concrete. ACI J Proc 1960. http://dx.doi.org/10.14359/8021, Published online.
[10] Ujhelyi JE. Water demand of concrete mixtures. Period Polytech Civ Eng 1997. Published online.
[11] Baghabra Al-Amoudi OS, Al-Kutti WA, Ahmad S, Maslehuddin M. Correlation between compressive strength and certain durability indices of plain and
blended cement concretes. Cem Concr Compos 2009;31(9). http://dx.doi.org/10.1016/j.cemconcomp.2009.05.005.
[12] Akcay B, Tasdemir MA. Performance evaluation of silica fume and metakaolin with identical finenesses in self compacting and fiber reinforced concretes.
Constr Build Mater 2018;185:436–44. http://dx.doi.org/10.1016/j.conbuildmat.2018.07.061.
[13] 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. http://dx.doi.org/10.1016/j.conbuildmat.2007.04.017, Published online.
[14] Hooton RD. Influence of silica fume replacement of cement on physical properties and resistance to sulfate attack, freezing and thawing, and alkali-silica
reactivity. ACI Mater J 1993. http://dx.doi.org/10.14359/4009, Published online.
[15] Ding JT, Li Z. Effects of metakaolin and silica fume on properties of concrete. ACI Mater J 2002;99(4):393–8. http://dx.doi.org/10.14359/12222.
[16] Safarzadegan Gilan S, Bahrami Jovein H, Ramezanianpour AA. Hybrid support vector regression - particle swarm optimization for prediction of compressive
strength and RCPT of concretes containing metakaolin. Constr Build Mater 2012;34:321–9. http://dx.doi.org/10.1016/j.conbuildmat.2012.02.038.
[17] Qian X, Li Z. The relationships between stress and strain for high-performance concrete with metakaolin. Cem Concr Res 2001;31(11):1607–11.
http://dx.doi.org/10.1016/S0008-8846(01)00612-3.
[18] Dubey M, Deo SV, Ramtekkar G. Evaluating properties of high performance concrete containing metakaolin as cementitious material. Int Rev Appl Sci
Eng 2022;13(3):257–62. http://dx.doi.org/10.1556/1848.2021.00359.
[19] Dhinakaran G, Thilgavathi S, Venkataramana J. Compressive strength and chloride resistance of metakaolin concrete. KSCE J Civ Eng 2012;16(7):1209–17.
http://dx.doi.org/10.1007/s12205-012-1235-z.
[20] Perumal R. Performance and modeling of high-performance steel fiber reinforced concrete under impact loads. Comput Concr 2014. http://dx.doi.org/
10.12989/cac.2014.13.2.255, Published online.
[21] Sankar B, Ramadoss P. Experimental and statistical investigations on alccofine based ternary blended high-performance concrete. Int J Eng Trans B Appl
2022;35(8):1629–40. http://dx.doi.org/10.5829/IJE.2022.35.08B.19.
[22] Poon CS, Kou SC, L. Lam. Compressive strength, chloride diffusivity and pore structure of high performance metakaolin and silica fume concrete. Constr
Build Mater 2006;20(10):858–65. http://dx.doi.org/10.1016/j.conbuildmat.2005.07.001.
[23] Gowram I, Beulah M, Sudhir MR, Mohan MK, Jain D. Efficacy of natural zeolite and metakaolin as partial alternatives to cement in fresh and hardened
high strength concrete. Adv Mater Sci Eng 2021;2021. http://dx.doi.org/10.1155/2021/4090389.
[24] Güneyisi E, Gesoǧlu M, Algin Z, Mermerdaş K. Optimization of concrete mixture with hybrid blends of metakaolin and fly ash using response surface
method. Compos Part B Eng 2014. http://dx.doi.org/10.1016/j.compositesb.2014.01.017, Published online.
[25] Brooks JJ, Johari MAM. E € ect of metakaolin on creep and shrinkage of concrete. 23 (2001) 0-7.
[26] Sharaky IA, Ghoneim SSM, Abdel Aziz BH, Emara M. Experimental and theoretical study on the compressive strength of the high strength concrete
incorporating steel fiber and metakaolin. Structures 2021;31(2020):57–67. http://dx.doi.org/10.1016/j.istruc.2021.01.061.
[27] Razak HA, Wong HS. Strength estimation model for high-strength concrete incorporating metakaolin and silica fume. Cem Concr Res 2005;35(4):688–95.
http://dx.doi.org/10.1016/j.cemconres.2004.05.040.
17
B. Sankar and P. Ramadoss Results in Control and Optimization 12 (2023) 100241
[28] Zhang MH, Malhotra VM. Characteristics of a thermally activated alumino-silicate pozzolanic material and its use in concrete. Cem Concr Res
1995;25(8):1713–25. http://dx.doi.org/10.1016/0008-8846(95)00167-0.
[29] Ramezanianpour AA, Bahrami Jovein H. Influence of metakaolin as supplementary cementing material on strength and durability of concretes. Constr
Build Mater 2012;30:470–9. http://dx.doi.org/10.1016/j.conbuildmat.2011.12.050.
[30] Ženíšek M, Vlach T, Laiblová L. Dosage of metakaolin in high performance concrete. Key Eng Mater 2017;722 KEM:311–5. http://dx.doi.org/10.4028/
www.scientific.net/KEM.722.311.
[31] Bhaskar M, Rao PS, Swami BL. Strength properties of fibre reinforced high performance concrete blended with metakolin and condensed silica fume.
IOSR J Mech Civ Eng 2016;16(053):99–105. http://dx.doi.org/10.9790/1684-160530199105.
[32] Nadeem A, Memon SA, Lo TY. The performance of fly ash and metakaolin concrete at elevated temperatures. Constr Build Mater 2014;62:67–76.
http://dx.doi.org/10.1016/j.conbuildmat.2014.02.073.
[33] Badogiannis E, Kakali G, Dimopoulou G, Chaniotakis E, Tsivilis S. Metakaolin as a main cement constituent. Exploitation of poor greek kaolins. Cem
Concr Compos 2005;27(2):197–203. http://dx.doi.org/10.1016/j.cemconcomp.2004.02.007.
[34] Madandoust R, Mousavi SY. Fresh and hardened properties of self-compacting concrete containing metakaolin. Constr Build Mater 2012;35:752–60.
http://dx.doi.org/10.1016/j.conbuildmat.2012.04.109.
[35] Yudong W, Xiaochun F. Experimental research on physical and mechanical properties of steel fiber high-strength concrete. Adv Mater Res
2011;168–170:1061–4. http://dx.doi.org/10.4028/www.scientific.net/AMR.168-170.1061.
[36] Megat Johari MA, Brooks JJ, Kabir S, Rivard P. Influence of supplementary cementitious materials on engineering properties of high strength concrete.
Constr Build Mater 2011;25(5):2639–48. http://dx.doi.org/10.1016/j.conbuildmat.2010.12.013.
[37] Shui Z, Zeng J, Liao Y, Leng Z. Influence of metakaolin on strength and microstructure of high-strength concrete. Key Eng Mater 2012;509:33–9.
http://dx.doi.org/10.4028/www.scientific.net/KEM.509.33.
[38] Mohammed A, Burhan L, Ghafor K, Sarwar W, Mahmood W. Artificial neural network (ANN), M5P-tree, and regression analyses to predict the early
age compression strength of concrete modified with DBC-21 and VK-98 polymers. Neural Comput Appl 2021;33(13):7851–73. http://dx.doi.org/10.1007/
s00521-020-05525-y.
[39] Ahmed HU, Mohammed AS, Mohammed AA. Proposing several model techniques including ANN and M5P-tree to predict the compressive strength of
geopolymer concretes incorporated with nano-silica. Environ Sci Pollut Res 2022;29(47):71232–56. http://dx.doi.org/10.1007/s11356-022-20863-1.
[40] Barkhordari MS, Armaghani DJ, Mohammed AS, Ulrikh DV. Data-driven compressive strength prediction of fly ash concrete using ensemble learner
algorithms. Buildings 2022;12(2). http://dx.doi.org/10.3390/buildings12020132.
[41] Moradi MJ, Khaleghi M, Salimi J, Farhangi V, Ramezanianpour AM. Predicting the compressive strength of concrete containing metakaolin with different
properties using ANN. Meas J Int Meas Confed 2021. http://dx.doi.org/10.1016/j.measurement.2021.109790, Published online.
[42] Liao S, Li G, Li J, Jiang D, Jiang G, Sun Y. Multi-object intergroup gesture recognition combined with fusion feature and KNN algorithm. 38 (2020)
2725-2735. http://dx.doi.org/10.3233/JIFS-179558.
[43] Mohammed A, Kurda R, Armaghani DJ, Hasanipanah M. Prediction of compressive strength of concrete modified with fly ash: Applications of neuro-swarm
and neuro-imperialism models. Adv Concr Constr 2021;11(5):489–512. http://dx.doi.org/10.12989/acc.2021.11.5.489.
[44] Yu Mingchao, Li Gongfa, Du Jiang GJ, Zeng Fei, Zhao Haoyi, Chen D. Application of PSO-RBF neural network in gesture recognition of continuous
surface EMG signals. J Intell Fuzzy Systems 2020. http://dx.doi.org/10.3233/JIFS-179535, Published online.
[45] Sun Y, Chen D, Weng Y, et al. Gesture recognition algorithm based on multi-scale feature fusion in RGB-d images. IET Image Process 2022;17:1280–90.
http://dx.doi.org/10.1049/ipr2.12712.
[46] Asteris PG, Lourenço PB, Roussis PC, et al. Revealing the nature of metakaolin-based concrete materials using artificial intelligence techniques. Constr
Build Mater 2022;322(2021):126500. http://dx.doi.org/10.1016/j.conbuildmat.2022.126500.
[47] Emad W, Mohammed ASalih, Kurda R, et al. Prediction of concrete materials compressive strength using surrogate models. Structures
2022;46(November):1243–67. http://dx.doi.org/10.1016/j.istruc.2022.11.002.
[48] Piro NS, Mohammed AS, Hamad SM. Evaluate and predict the resist electric current and compressive strength of concrete modified with GGBS and
steelmaking slag using mathematical models. J Sustain Metall 2023;9(1):194–215. http://dx.doi.org/10.1007/s40831-022-00631-8.
[49] Shah HA, Nehdi ML, Khan MI, et al. Predicting compressive and splitting tensile strengths of silica fume concrete using M5P model tree algorithm.
Materials (Basel) 2022;15(15):1–17. http://dx.doi.org/10.3390/ma15155436.
[50] Mohammed A, Rafiq S, Sihag P, Kurda R, Mahmood W. Soft computing techniques: Systematic multiscale models to predict the compressive strength of
HVFA concrete based on mix proportions and curing times. J Build Eng 2021. http://dx.doi.org/10.1016/j.jobe.2020.101851, Published online.
[51] Kumar R. Modified mix design and statistical modelling of lightweight concrete with high volume micro fines waste additive via the box-behnken design
approach. Cem Concr Compos 2020. http://dx.doi.org/10.1016/j.cemconcomp.2020.103706, Published online.
[52] Khudhair MH, Al-Anweh AM, Nomaan MH, et al. Response surface modeling of compressive strength of highperformance concrete formulated by a high
water reducing and setting accelerating superplasticizer. Box-behnken experimental design. J Chem Technol Metall 2019;54(1):135–44.
[53] Rashad AM. Metakaolin as cementitious material: History, scours, production and composition-a comprehensive overview. Constr Build Mater 2013.
http://dx.doi.org/10.1016/j.conbuildmat.2012.12.001, Published online.
[54] Akalin O, Akay KU, Sennaroglu B. Self-consolidating high-strength concrete optimization by mixture design method. ACI Mater J 2010. http://dx.doi.
org/10.14359/51663861, Published online.
[55] Paiva H, Velosa A, Cachim P, Ferreira VM. Effect of metakaolin dispersion on the fresh and hardened state properties of concrete. Cem Concr Res 2012.
http://dx.doi.org/10.1016/j.cemconres.2012.01.005, Published online.
[56] Hassan AAA, Lachemi M, Hossain KMA. Effect of metakaolin and silica fume on the durability of self-consolidating concrete. Cem Concr Compos 2012.
http://dx.doi.org/10.1016/j.cemconcomp.2012.02.013, Published online.
[57] Kadri EH, Kenai S, Ezziane K, Siddique R, Schutter GDe. Influence of metakaolin and silica fume on the heat of hydration and compressive strength
development of mortar. Appl Clay Sci 2011. http://dx.doi.org/10.1016/j.clay.2011.06.008, Published online.
[58] Jaya NA, Yun-Ming L, Cheng-Yong H, Abdullah MMAB, Hussin K. Correlation between pore structure, compressive strength and thermal conductivity of
porous metakaolin geopolymer. Constr Build Mater 2020. http://dx.doi.org/10.1016/j.conbuildmat.2020.118641, Published online 2020..
[59] Unis H, Abdalla AA, Mohammed AS, Mohammed AA. Mathematical modeling techniques to predict the compressive strength of high-strength concrete
incorporated metakaolin with multiple mix proportions. Clean Mater 2022;5(April):100132. http://dx.doi.org/10.1016/j.clema.2022.100132.
[60] Abdalla A, Mohammed AS. Hybrid MARS-, MEP-, and ANN-based prediction for modeling the compressive strength of cement mortar with various sand
size and clay mineral metakaolin content. Arch Civ Mech Eng 2022;22(4):1–13. http://dx.doi.org/10.1007/s43452-022-00519-0.
[61] Ali B, Kurda R, Herki B, et al. Effect of varying steel fiber content on strength and permeability characteristics of high strength concrete with micro
silica. Materials (Basel) 2020;13(24):1–17. http://dx.doi.org/10.3390/ma13245739.
[62] Emad W, Mohammed AS, Bras A, et al. Metamodel techniques to estimate the compressive strength of UHPFRC using various mix proportions and a
high range of curing temperatures. Constr Build Mater 2022;349(May):128737. http://dx.doi.org/10.1016/j.conbuildmat.2022.128737.
[63] Ibrahim AK, Dhahir HY, Mohammed AS, Omar HA, Sedo AH. The effectiveness of surrogate models in predicting the long-term behavior of varying
compressive strength ranges of recycled concrete aggregate for a variety of shapes and sizes of specimens. Arch Civ Mech Eng 2023;23(1):1–13.
http://dx.doi.org/10.1007/s43452-022-00595-2.
18
B. Sankar and P. Ramadoss Results in Control and Optimization 12 (2023) 100241
[64] Nath P, P.K. Sarker. Effect of GGBFS on setting, workability and early strength properties of fly ash geopolymer concrete cured in ambient condition.
Constr Build Mater 2014. http://dx.doi.org/10.1016/j.conbuildmat.2014.05.080, Published online.
[65] Al-Alaily HS, Hassan AAA. Refined statistical modeling for chloride permeability and strength of concrete containing metakaolin. Constr Build Mater
2016. http://dx.doi.org/10.1016/j.conbuildmat.2016.03.187, Published online.
[66] Popovics S, Ujhelyi J. Contribution to the concrete strength versus water-cement ratio relationship. J Mater Civ Eng 2008;20(7):459–63. http:
//dx.doi.org/10.1061/(asce)0899-1561(2008)20:7(459).
[67] de Brito J, Kurda R, da Silva PR. Can we truly predict the compressive strength of concrete without knowing the properties of aggregates? Appl Sci
2018;8(7):1–21. http://dx.doi.org/10.3390/app8071095.
[68] Elwell DJ, Fu G. Compression testing of concrete: cylinders vs. cube. Spec Rep 1995;119:21, Published online.
[69] Golafshani EM, Behnood A. Application of soft computing methods for predicting the elastic modulus of recycled aggregate concrete. J Clean Prod 2018.
http://dx.doi.org/10.1016/j.jclepro.2017.11.186, Published online.
[70] Demircan E, Harendra S, Vipulanandan C. Artificial neural network and nonlinear models for gelling time and maximum curing temperature rise in
polymer grouts. J Mater Civ Eng 2011. http://dx.doi.org/10.1061/(asce)mt.1943-5533.0000172, Published online.
[71] Mohammed A, Rafiq S, Mahmood W, et al. Artificial neural network and NLR techniques to predict the rheological properties and compression strength
of cement past modified with nanoclay. Ain Shams Eng J 2021;12(2):1313–28. http://dx.doi.org/10.1016/j.asej.2020.07.033.
[72] Mohammed A, Mahmood W, Tga Ghafor K. Rheological properties with maximum shear stress and compressive strength of cement-based grout modified
with polycarboxylate polymers. Constr Build Mater 2020. http://dx.doi.org/10.1016/j.conbuildmat.2019.117534, Published online.
[73] Ma H, Sun Z, Ma G. Research on compressive strength of manufactured sand concrete based on response surface methodology (RSM). Appl Sci 2022;12(7).
http://dx.doi.org/10.3390/app12073506.
[74] Design-expert. 9th. Stat-Ease software; 2012, Published online.
[75] Salih A, Rafiq S, Sihag P, Ghafor K, Mahmood W, Sarwar W. Systematic multiscale models to predict the effect of high-volume fly ash on the maximum
compression stress of cement-based mortar at various water/cement ratios and curing times. Meas J Int Meas Confed 2021. http://dx.doi.org/10.1016/j.
measurement.2020.108819, Published online.
[76] Dinakar P, Sahoo PK, Sriram G. Effect of metakaolin content on the properties of high strength concrete. Int J Concr Struct Mater 2013. http:
//dx.doi.org/10.1007/s40069-013-0045-0, Published online.
[77] Nuruddin MF, Khan SU, Shafiq N, Ayub T. Strength development of high-strength ductile concrete incorporating metakaolin and PVA fibers. Sci World
J 2014. http://dx.doi.org/10.1155/2014/387259, Published online.
[78] Raj RR, Pillai EBP, Santhakumar AR. Evaluation and mix design for ternary blended high strength concrete. Procedia Eng 2013;51(NUiCONE 2012):65–74.
http://dx.doi.org/10.1016/j.proeng.2013.01.012.
[79] Salih A, Rafiq S, Sihag P, Ghafor K, Mahmood W, Sarwar W. Systematic multiscale models to predict the effect of high-volume fly ash on the
maximum compression stress of cement-based mortar at various water/cement ratios and curing times. Meas J Int Meas Confed 2021;171:108819.
http://dx.doi.org/10.1016/j.measurement.2020.108819.
[80] Abouhussien AA, Hassan AAA. Application of statistical analysis for mixture design of high-strength self-consolidating concrete containing metakaolin. J
Mater Civ Eng 2014. http://dx.doi.org/10.1061/(asce)mt.1943-5533.0000944, Published online.
[81] Zeng JJ, Fan ZH, Chen L. Influence of metakaolin on strength and microstructure of steam-cured high-strength concrete. Adv Mater Res 2014.
http://dx.doi.org/10.4028/www.scientific.net/AMR.838-841.42.
[82] Shen P, Lu L, Chen W, Wang F, Hu S. Efficiency of metakaolin in steam cured high strength concrete. Constr Build Mater 2017. http://dx.doi.org/10.
1016/j.conbuildmat.2017.07.006, Published online.
[83] Liu W, Tan H, Ni C, Chen Z, Luo T, Yu L. Effect of silica fume and fly ash on compressive strength and weight loss of high strength concrete material
in sulfuric and acetic acid attack. In: Key engineering materials. 2017, http://dx.doi.org/10.4028/www.scientific.net/KEM.748.301.
[84] Jaturapitakkul C, Kiattikomol K, Sata V, Leekeeratikul T. Use of ground coarse fly ash as a replacement of condensed silica fume in producing high-strength
concrete. Cem Concr Res 2004. http://dx.doi.org/10.1016/S0008-8846(03)00150-9, Published online.
[85] Abdalla AA, Mohammed ASalih. Theoretical models to evaluate the effect of SiO2 and CaO contents on the long-term compressive strength of cement
mortar modified with cement kiln dust (CKD). Arch Civ Mech Eng 2022;22(3):1–15. http://dx.doi.org/10.1007/s43452-022-00418-4.
[86] Ahmed A, Abubakr P, Salih Mohammed A. Efficient models to evaluate the effect of C3S, C2S, C3A, and C4AF contents on the long-term compressive
strength of cement paste. Structures 2023;47(2022):1459–75. http://dx.doi.org/10.1016/j.istruc.2022.11.114.
[87] Dvorkin L, Bezusyak A, Lushnikova N, Ribakov Y. Using mathematical modeling for design of self compacting high strength concrete with metakaolin
admixture. Constr Build Mater 2012;37:851–64. http://dx.doi.org/10.1016/j.conbuildmat.2012.04.019.
[88] Ali R, Muayad M, Mohammed AS, Asteris PG. Analysis and prediction of the effect of nanosilica on the compressive strength of concrete with different
mix proportions and specimen sizes using various numerical approaches. Struct Concr 2022(July):1–24. http://dx.doi.org/10.1002/suco.202200718.
[89] Khedr SA, Idriss AF. Resistance of silica-fume concrete to corrosion-related damage. J Mater Civ Eng 1995;7(2):102–7. http://dx.doi.org/10.1061/(asce)
0899-1561(1995)7:2(102).
[90] Hassan AMT, Mahmud GH, Mohammed AS, Jones SW. The influence of normal curing temperature on the compressive strength development and flexural
tensile behaviour of UHPFRC with vipulanandan model quantification. Structures 2021;30(2020):949–59. http://dx.doi.org/10.1016/j.istruc.2021.01.063.
[91] Alexander MG, Magee BJ. Durability performance of concrete containing condensed silica fume. Cem Concr Res 1999;29(6):917–22. http://dx.doi.org/
10.1016/S0008-8846(99)00064-2.
[92] Yang SL, Millard SG, Soutsos MN, Barnett SJ, Le TT. Influence of aggregate and curing regime on the mechanical properties of ultra-high performance
fibre reinforced concrete (UHPFRC). Constr Build Mater 2009;23(6):2291–8. http://dx.doi.org/10.1016/j.conbuildmat.2008.11.012.
[93] Zain MFM, Safiuddin M, Mahmud H. Development of high performance concrete using silica fume at relatively high water-binder ratios. Cem Concr Res
2000;30(9):1501–5. http://dx.doi.org/10.1016/S0008-8846(00)00359-8.
[94] Corinaldesi V, Moriconi G. Mechanical and thermal evaluation of ultra high performance fiber reinforced concretes for engineering applications. Constr
Build Mater 2012;26(1):289–94. http://dx.doi.org/10.1016/j.conbuildmat.2011.06.023.
[95] Zhang MH, Malhotra VM, Bouzoubaa N. Mechanical properties and durability of concrete made with high-volume fly ash blended cements using a coarse
fly ash. Cem Concr Res 2001;31:1393–402.
[96] Yazici H. The effect of curing conditions on compressive strength of ultra high strength concrete with high volume mineral admixtures. Build Environ
2007;42(5):2083–9. http://dx.doi.org/10.1016/j.buildenv.2006.03.013.
[97] Oner A, Akyuz S, Yildiz R. An experimental study on strength development of concrete containing fly ash and optimum usage of fly ash in concrete.
Cem Concr Res 2005;35(6):1165–71. http://dx.doi.org/10.1016/j.cemconres.2004.09.031.
[98] Yu R, Spiesz P, Brouwers HJH. Mix design and properties assessment of ultra-high performance fibre reinforced concrete (UHPFRC). Cem Concr Res
2014;56:29–39. http://dx.doi.org/10.1016/j.cemconres.2013.11.002.
[99] Dotto JMR, De Abreu AG, Molin DCCDal, Müller IL. Influence of silica fume addition on concretes physical properties and on corrosion behaviour of
reinforcement bars. Cem Concr Compos 2004;26(1):31–9. http://dx.doi.org/10.1016/S0958-9465(02)00120-8.
[100] Mohammed BH, Sherwani AFH, Faraj RH, Qadir HH, Younis KH. Mechanical properties and ductility behavior of ultra-high performance fiber reinforced
concretes: Effect of low water-to-binder ratios and micro glass fibers. Ain Shams Eng J 2021;12(2):1557–67. http://dx.doi.org/10.1016/j.asej.2020.11.008.
[101] Alsalman A, Dang CN, Martí-Vargas JR, Hale WMicah. Mixture-proportioning of economical UHPC mixtures. J Build Eng 2020;27. http://dx.doi.org/10.
1016/j.jobe.2019.100970.
19