&RPSXWHUV DQG 6WUXFWXUHV   

Contents lists available at ScienceDirect

Computers and Structures

journal homepage: [Link]/locate/compstruc

Stacked-based machine learning to predict the uniaxial compressive

strength of concrete materials
Abdelrahman Kamal Hamed a,* , Mohamed Kamel Elshaarawy a , Mostafa M. Alsaadawi a,b
a
Civil Engineering Department, Faculty of Engineering, Horus University-Egypt, New Damietta 34517, Egypt
b
Structural Engineering Department, Faculty of Engineering, Mansoura University, Mansoura 35516, Egypt

A R T I C L E I N F O A B S T R A C T

Keywords: Compressive strength is a key factor in the design and durability of concrete structures. Accurate prediction of
Concrete compressive strength helps optimize material use and reduce construction costs. This study proposes a novel
Compressive strength stacked model for predicting compressive strength, integrating three base models with linear regression. The
Stacking
base models include Artificial Neural Networks, Random Forest, and Extreme Gradient Boosting, while the
Machine learning
Shapley analysis
stacked model uses Linear Regression as the metamodel. A dataset of 1,030 concrete mix samples covering eight
Prediction critical input parameters, including cement, blast furnace slag, coarse aggregates, fine aggregates, fly ash, water,
superplasticizer, and curing days, was used for training and evaluation. The dataset was split into training (80%),
validation (10%), and testing (10%) subsets. The models were trained independently, and their predictions were
used to develop the stacked model. Among the base models, the Extreme Gradient Boosting model achieved the
highest accuracy, with an R2 of 0.947 during testing. However, the stacked model outperformed it, attaining an
R2 of 0.953 in the testing phase. Shapley additive explanations analysis identified curing duration as the most
influential factor in compressive strength prediction. A user-friendly graphical interface was developed to
facilitate efficient prediction of compressive strength in concrete structures.

component [3]. PC is a major driver of CO2 emissions, accounting for

1. Introduction approximately 7 % of the total global emissions [4]. The production
process of PC involves the calcination of limestone (calcium carbonate),
The increasing release of greenhouse-gases like carbon dioxide, which releases carbon dioxide. Notably, the calcination process, where
methane, and nitrous oxide into the atmosphere, primarily driven by calcium oxide (CaO) is produced, is responsible for a significant portion
human activities such as burning fossil fuels, deforestation, and indus­ of these emissions. As urbanization and infrastructure development
trial processes, is causing profound changes to the Earth’s environment. continue to expand, the demand for concrete and consequently PC has
This rise leads to widespread climate disruptions, including shifts in surged. Presently, the annual production of PCs stands at 4,000 million
weather patterns, more frequent extreme weather events, and severe tons, with forecasts predicting this figure will rise to 6,000 million tons
impacts on ecosystems [1]. A particularly visible consequence is the by 2060 [5]. This trend exacerbates environmental concerns and high­
reduction of polar ice stores in the Arctic and Antarctic regions, where lights the need for sustainable alternatives. Addressing this critical issue
ice caps and glaciers are melting at an accelerating rate. This reduction requires the development of effective substitutes that can satisfy the
poses significant ecological challenges, including rising sea levels and increasing demand while drastically lowering energy consumption and
disrupted ecosystems, which threaten biodiversity and the livelihoods of CO2 emissions [6]. One promising approach is incorporating waste and
millions of people worldwide. Among the contributors to these emis­ recycled materials into concrete production. This method not only ad­
sions, the building industry, a key sector within the global economy, dresses the demand for construction materials but also contributes to
emerges as the largest contributor to greenhouse-gas emissions, ac­ reducing environmental hazards and conserving natural resources. By
counting for nearly 50 % of the total worldwide emissions. This sub­ diverting industrial waste from landfills and utilizing it in concrete, we
stantial impact underscores the industry’s pivotal role in global can mitigate the ecological footprint of construction activities [7,8].
environmental concerns [2]. This is primarily due to the extensive use of Various industrial by-products have been recognized as suitable al­
concrete, which relies heavily on Portland cement (PC) as a key ternatives to PC in concrete production. Some of the most promising

* Corresponding author.
E-mail addresses: akamal@[Link] (A.K. Hamed), melshaarawy@[Link] (M.K. Elshaarawy).

[Link]
Received 29 September 2024; Accepted 2 January 2025
$YDLODEOH RQOLQH  -DQXDU\ 
‹  (OVHYLHU /WG $OO ULJKWV DUH UHVHUYHG LQFOXGLQJ WKRVH IRU WH[W DQG GDWD PLQLQJ $, WUDLQLQJ DQG VLPLODU WHFKQRORJLHV
A.K. Hamed et al. &RPSXWHUV DQG 6WUXFWXUHV   

Nomenclature LR Linear Regression

LSC Lightweight Self-compacting Concrete
AI Artificial Intelligence LSTM Long Short-Term Memory
ANFIS Adaptive Neuro-Fuzzy Inference System LWC Lightweight Concrete
ANN Artificial Neural Network LightGBM Light Gradient Boosting Machine
AdaBoost Adaptive Boosting MAE Mean Absolute Error
BP Back Propagation MAPE Mean Absolute Percentage Error
BMs Base Models MBE Mean Bias Error
BO Bayesian Optimization ML Machine Learning
BiLSTM Bidirectional Long Short-Term Memory MLPNN Multilayer Perceptron Neural Network
C Cement MSSA-DE Multi-objective Metaheuristic Differential Evolution
CC Correlation Coefficient OPC Ordinary Portland Cement
CNN Convolutional neural network PC Portland Cement
CPI Cumulative Prediction Error Index PDP Partial Dependence Plot
CS Compressive Strength PDPs-1D One-Dimensional Partial Dependence Plot
CV Cross-validation RAC Recycled Aggregate Concrete
CaO Calcium Oxide RF Random Forest
Cagg Coarse aggregate RMSD Root Mean Square Difference
cov Covariance RMSE Root Mean Square Error
DL Deep Learning RNN Recurrent Neural Networks
DT Decision Tree R2 Determination coefficient
FA Fly Ash r Pearson correlation coefficient
Fagg Fine aggregate SCC Self-Compacting Concrete
GA Genetic Algorithm SD Standard Deviation
GB Gradient Boosting SHAP Shapley Additive exPlanations
GBM Gradient Boosted Machine SM Stacked Model
GBR Gradient Boosting Regressor SP Superplasticizer
GBRT Gradient Boosted Regression Tree SVM Support Vector Machines
GEP GeneExpression Programming TR Training stage
GRU Gated Recurrent Units TS Testing stage
GUI Graphical User Interface UHPC Ultra-High Performance Concrete
HPC High Performance Concrete VAF Variance Accounted For
HSC High Strength Concrete VD Validation stage
IQR Interquartile Range W Water
KDE Kernel Density Estimation XGB Extreme Gradient Boosting

options include ground granulated blast-furnace slag [9], granite pow­ binder ratio can lead to higher compressive strength but may reduce
der [10], and fly ash [11]. These materials are frequently produced in workability, while aggregates’ size and grading can impact the finished
significant amounts during industrial activities and often present chal­ product’s density and durability. The type of binder, whether it is a
lenges related to disposal. By incorporating these materials into con­ traditional PC or an alternative material, also plays a crucial role in
crete, we can not only improve the properties of the concrete, such as its determining the concrete’s performance [20].
durability and strength but also significantly reduce its environmental Traditionally, the evaluation of concrete’s compressive strength has
impact. For instance, replacing traditional PCs with alternative mate­ depended on laboratory testing techniques, comprising crushing con­
rials like ground granulated blast-furnace slag, granite powder, and fly crete samples with predetermined dimensions following a designated
ash can significantly reduce greenhouse-gas emissions, potentially curing time. This method is universally accepted due to its ability to
reducing up to 80 % [12]. This significant reduction highlights the po­ provide reliable measurements of a material’s load-bearing capacity
tential of these materials to contribute to more sustainable construction [21]. Despite its widespread use, this approach is increasingly viewed as
practices [13]. inefficient. The process is both time-intensive and costly, as it requires
Concrete remains the most widely used construction material glob­ significant labor and specialized equipment. Moreover, it involves
ally, surpassed only by water consumption [14–17]. Its widespread destructive testing, which does not allow for further use of the tested
availability, malleability upon cooling, and strength to support loads samples, thereby adding to material waste and cost. To overcome the
after hardening are reasons for its widespread acclaim. These properties limitations of traditional testing methods, some researchers have
make concrete an essential material for various construction applica­ explored using empirical regression techniques to predict the CS of
tions. The performance of concrete, particularly its compressive concrete based on its mix proportions [22–24]. These techniques involve
strength, is critical for the safety and durability of structures [18]. creating mathematical models that estimate strength outcomes from the
Compressive strength (CS) refers to the material’s ability to withstand known quantities of the mixture’s ingredients, such as cement, water,
loads that compress or reduce its size. This property is influenced by aggregates, and admixtures. However, developing accurate regression
various factors, including the water-to-binder ratio, the size and quality models is challenging due to the intricate and nonlinear relationships
of aggregates, and the type of binder used [19]. The heterogeneous among the various components of concrete. Factors such as the type and
nature of concrete, which comprises a mix of binder, aggregates, water, fineness of cement, the water-to-cement ratio, the size and quality of
and admixtures, adds to the complexity of predicting its compressive aggregates, and the presence of additives can all significantly influence
strength. Each component and how they are combined can significantly the strength, making it difficult to predict outcomes precisely [25].
affect the final properties of the concrete. For example, a lower water-to- In recent years, advancements in artificial intelligence (AI) and


A.K. Hamed et al. &RPSXWHUV DQG 6WUXFWXUHV   

machine learning (ML) have revolutionized various industries, including and consistency in concrete production. These models are particularly
construction. AI, particularly through the use of ML models such as effective in identifying the optimal balance of materials that achieve
Artificial-Neural-Networks (ANN), Support-Vector-Machines (SVM), desired strength properties while minimizing costs and environmental
Gene-Expression-Programming (GEP), and Deep Learning (DL), has impact. The use of ML in this context improves the efficiency of concrete
demonstrated remarkable capabilities in handling complex data sets and design and supports the development of more sustainable building
extracting meaningful patterns [26–30]. These models have been practices. ML models play a pivotal role in promoting the sustainability
applied to a wide range of problems, providing more efficient and ac­ of concrete mix designs. ML can help reduce the environmental footprint
curate solutions compared to traditional statistical methods. Applica­ of concrete production by optimising material combinations. For
tions aimed at predicting concrete mechanical properties emphasize instance, using supplementary cementitious materials like fly ash can
various methods, challenges, and their practical implications, under­ lower carbon emissions associated with cement production, significantly
scoring the expanding role of AI in engineering [31,32]. Techniques contributing to global CO2 emissions. Additionally, optimizing the mix
have also been utilized to optimize concrete mix design, enhancing the design can reduce material waste and improve concrete structures’
reliability and efficiency of construction processes [33]. In structural overall lifespan and durability, leading to long-term sustainability
engineering, ML has been widely used to predict the behavior of struc­ benefits.
tural elements [34–38], detect damage [39,40], and calculate buckling The application of ML in structural engineering has gained signifi­
loads [41,42]. Many studies have employed ML techniques to estimate cant attention, particularly in predicting the concrete’s CS. Various
the axial load capacity of composite columns [43–48]. studies have demonstrated the potential of ML models in this domain.
Recent reviews also highlight the role of ML in predicting the For instance, Chithra et al. [53] successfully utilized Artificial Neural
behavior of concrete under extreme conditions, such as elevated tem­ Networks (ANN) to predict the CS of high-performance concrete (HPC)
peratures [49]. Additionally, heuristic algorithms, combined with semi- containing silica nanoparticles and copper slag, showcasing the ANN
empirical formulas, have been successfully employed for compressive model’s ability to handle complex material properties. Similarly,
strength estimation, offering more transparent relationships between Nguyen et al. [54] developed four different ML models to forecast the
input variables and outcomes [50]. Furthermore, hybrid ML approaches compressive and tensile strength of HPC, with their findings empha­
integrating ANN and particle swarm optimization (PSO) have been sizing the superior accuracy of Gradient-Boosting-Regressor (GBR) and
applied in various studies, such as predicting soil thermal conductivity Extreme-Gradient-Boosting (XGB) models in capturing intricate data
and improving concrete property forecasts’ accuracy [51]. Moreover, patterns. Kumar et al. [55], focusing on lightweight concrete (LWC),
innovative techniques have been used to predict the unconfined employed SVM models based on data from 120 experimental sets,
compressive strength of granite, employing only two non-destructive highlighting SVM’s effectiveness in evaluating LWC compressive
testing indexes, demonstrating the potential of ANN models in diverse strength across diverse datasets. In another study, Ashrafian [56]
geological applications [52]. applied heuristic regression methods to analyze the strength and ultra­
In the context of concrete technology, ML models can be trained on sonic pulse velocity of fibrous concrete, providing valuable insights into
large datasets of past concrete mix compositions and their correspond­ the material’s behavior under various conditions. Using a comprehen­
ing compressive strengths, allowing for the creation of predictive models sive investigation of eight input factors crucial to the model’s accuracy,
that can forecast the strength of new mixtures with high accuracy. Zhang et al. [57] employed the Random Forest (RF) approach to forecast
Machine learning offers several advantages over traditional empirical the uniaxial CS of lightweight self-compacting concrete (LSC).
regression methods, particularly in handling the complexities and non- Furthermore, the incorporation of nanotechnology, particularly
linearities inherent in concrete mixture design. ML models can learn Nanosilica, has also been explored to enhance the compressive strength
from vast amounts of data and identify subtle patterns that might not be of concrete, with various models assessing the influence of mix pro­
evident through conventional methods. This ability to model complex portions and curing conditions on the material’s strength [58]. Aslam
relationships makes ML particularly useful for predicting the CS of et al. [59] developed a GEP model to predict the compressive strength of
concrete, including variations with alternative materials like fly ash. high-strength concrete (HSC), leveraging a comprehensive dataset of
Furthermore, ML models can adapt and improve over time as more data 357 points to illustrate GEP’s robustness in modeling complex material
becomes available, continually refining their predictions. characteristics. Collectively, these studies underscore the diverse ap­
Fly ash, a by-product of coal combustion has emerged as a valuable plications and effectiveness of various ML models in predicting concrete
alternative material in concrete production due to its environmental and properties, highlighting the significant advancements in this field. The
performance benefits. Incorporating fly ash into concrete not only helps overview of these ML models and approaches used in past literature is
in waste management by reducing the need for landfill space but also further detailed in Table 1, providing a comprehensive understanding of
improves the mechanical properties of concrete, such as its durability the advancements in this field.
and resistance to chemical attack. Using ML models to predict the CS of
fly ash-based concrete is essential for optimizing mix designs. Accurate 2. Research significance
predictions ensure that the concrete meets necessary performance
standards while also maximizing the use of sustainable materials. This The novelty of this research resides in its thorough methodology for
approach helps in achieving a balance between structural integrity and using machine learning models to predict CS and developing an inno­
environmental sustainability. Various factors affect the CS of concrete, vative stacked linear regression model (SM-LR), which provides sub­
such as the water-to-binder ratio, the size and quality of aggregates, the stantial enhancements in accuracy and practical usability compared to
type of binder employed, the curing period, and the distinct properties of conventional techniques. This model integrated three ML models: ANN,
alternative materials like fly ash. The water-to-binder ratio, for instance, RF, and XGB, as the base models. These models were chosen based on
directly affects the porosity of the concrete, which in turn influences its their strong predictive capabilities, which, combined with stacking, can
strength and durability. Similarly, aggregates’ size and grading can effectively overcome the weaknesses of each model. These selected
impact the mix’s workability and CS. Understanding these factors and models align better with the dataset’s size and complexity, avoiding the
their interactions is critical for designing concrete mixtures that meet need for extensive data, which deep learning models require. They strike
specific strength and performance criteria. a balance between predictive performance and the ability to interpret
Implementing ML in analysing large datasets from past concrete how inputs affect outputs, which is crucial for applications in structural
mixtures allows for developing models that accurately predict concrete engineering where model insights guide decisions. Given the available
characteristics. By examining historical data, ML models can optimize computational resources, these models offer a practical compromise
the proportions of materials in new mixtures, enhancing quality control between accuracy and computational demands. The literature review


A.K. Hamed et al. &RPSXWHUV DQG 6WUXFWXUHV   

Table 1 3. A sensitivity analysis was performed using Shapley-Additive-

Distinct ML models applied in previous literature. exPlanations (SHAP) to identify the primary variables that impact
Concrete type Dataset Model Performance index of Ref. CS prediction, providing useful insights for enhancing accuracy.
the best model 4. Conducting PDP analysis to illustrate the functional relationship
HPC 471 MLPNN, DT, RF, R2 = 0.89, error less [60] between specific input parameters and model predictions.
and GBRT than 10 MPa. 5. This work highlights the importance of blending theoretical model
HPC 1030 GBRT R2 = 0.92, MSE = [25] development with practical usability by developing an intuitive
22.09 MPa2, RMSE = Graphical User Interface (GUI), which helps to close the gap between
4.7 MPa.
HSC 225 FA and RF R2 = 0.9747 (Training), [61]
sophisticated computational predictions and their effective applica­
0.8753 (test); RMSE = tion in real-world scenarios.
3.6 MPa (Training), 6.6
MPa (test). 3. Research objectives
HPC 1030 XGB, RF, and R2 = 0.95, RMSE = [62]
SVR 0.05, MAE = 0.15.
HPC 270 GEP, DT, and R2 = 0.911 [63] This study extends prior research by utilizing ML models to predict
Bagging the CS of concrete within a data range (2 to 82 MPa). The primary aim
UHPC 78 ANN R2 = 0.9986, MAPE = [64] was to assess the effectiveness of various base ML models in predicting
0.0026 %, RMSE = CS and then enhance their accuracy by stacking techniques. The
4.6564 %.
HPC 1761 SVM, ECSO- R2 = 0.9418, SCC = [65]
research methodology, illustrated in Fig. 1, involved collecting and
SVM, ANN, 0.9369, MAPE = analyzing data from 1030 datasets, including parameters such as cement
ANFIS, ELM, M5, 14.013, MAE = 3.9797, and aggregates, with the properties examined through histograms and
and NLSR RRMSE = 1.0463. heatmaps. This study utilized three ML models to predict the CS of
SCC 7 ANFIS R = 95.79 % (Average [66]
concrete. The models’ effectiveness was assessed by comparing the
of all testing datasets).
HPC 1133 MFA-ANN R = 0.95. [67] actual values with the predicted outcomes, employing various regres­
Pozzolanic 310 GEP and ANN R2 (Training) = 0.94, R2 [68] sion metrics. To ensure reliable results, k-fold cross-validation was used
concrete (Testing) = 0.88. throughout the evaluation process. Additionally, an in-depth sensitivity
OPC and HPC 2817 ANN and ANFIS R2 = 0.9675, RMSE = [69] analysis was performed to explore how different variables affect model
4.6877 MPa, MAE =
3.5349 MPa, MBE =
accuracy. The findings were compared with prior studies to pinpoint the
0.3829 MPa. most accurate model. This methodology presents a more efficient and
HSC 357 GEP R2 = 0.914, RRMSE = [59] economically viable alternative to conventional experimental proced­
0.031, RSE = 0.023, ures, offering deeper insights into the factors influencing concrete CS
MAE = 0.595, RMSE =
prediction.
1.62.
UHPC 626 XGB, LightGBM, R2 = 0.996 (Training), [70]
hybrid XGB- 0.963 (Testing). 4. Methodology
LightGBM
RHA-FA 138 ANN, XGB, GBM R2 = 0.9933 (Train), [71] 4.1. Database collection
concrete 0.8404 (Test), RMSE =
0.6755 (Train), 3.2776
(Test), MAPE = 9.3646 Using information gathered from previous studies or data collected
(Test) in the lab, researchers are able to develop predictive models. This study
UHPC 810 AB, RF, GB R2 = 0.9913 (Train), [72] conducted a review of previous academic literature, including notable
0.9804 (Test).
works by Song et al. [19,60] and Yeh [78], to compile a dataset
Recycled 105 MSSA-DE R2 = 0.9906 (Train), [73]
aggregate 0.9604 (Test). comprising 1,030 data points pertinent to the CS of concrete. The pri­
concrete mary features used as input variables in the analysis included cement
(RAC) (C), blast furnace slag (Slag), fly ash (FA), water (W), superplasticizer
HPC 2171 BiLSTM, CNN, R2 = 0.990 (Training), [74] (SP), coarse aggregate (Cagg), fine aggregate (Fagg), and the curing
GRU, LSTM R2 = 0.961 (Testing).
duration (Age). These factors were crucial in determining the final
HPC 1030 ANN, MARS-L, R2 = 0.8894, RMSE = [75]
MARS-C, GPR, 0.0701, MAE = 0.0490, compressive strength outcome.
MPMR, and MAPE = 16.2757, a-20
HENSM. index = 78.6408. 4.2. Descriptive statistics
Lightweight 449 MARS, M5P, R2 = 0.966, RMSE = [76]
Structural ELM, and GEP 4.549 MPa, RAE =
Concrete 2.484 MPa. Table 2 offers a summary of the collected data’s quantitative fea­
Recycled 209 ICA-XGBoost, R2 = 0.983, MAPE = [77] tures, detailing each of its features in detail. The rows provide detailed
Aggregate ICA-ANN, ICA- 0.031, MAE = 1.147, statistical data for each variable, while each column represents a distinct
Concrete SVR, and ICA- RMSE = 1.479, a10- variable.
(RAC) ANFIS index = 1.000.
The Kernel Density Estimation (KDE) plots are shown in Fig. 2. It
provides a comprehensive overview of the distribution characteristics
confirms their successful application in similar engineering contexts, for the dataset’s nine distinct variables (X1 through X8 and Y). Key
validating their effectiveness and appropriateness for the study’s ob­ statistical indicators included in these plots are the mean, standard de­
jectives. The following are the authors’ specific contributions: viation, skewness, and kurtosis, which correspond to the quantitative
data shown in Table 2. The mean provides the average value of the data,
1. To introduce a stacked model (SM) development approach that in­ representing the central tendency. The standard deviation measures the
tegrates three base ML models: ANN, RF, and XGB, with a linear dispersion or spread of the data around the mean. Skewness quantifies
regression (LR) meta-model and to assess their effectiveness in pre­ the asymmetry of the data distribution: positive skewness indicates a
dicting concrete’s CS. longer right tail, while negative skewness points to a longer left tail.
2. Utilizing Bayesian optimization for hyperparameter tuning signifi­ Kurtosis measures the “tailedness” of the distribution: high kurtosis (>3)
cantly enhances the performance of the base models. indicates heavy tails and peakiness, while low kurtosis (<3) suggests
lighter tails and a flatter distribution.


A.K. Hamed et al. &RPSXWHUV DQG 6WUXFWXUHV   

Fig. 1. Methodology flowchart.

Table 2
Summary of the statistical information for the dataset.
Variables → Statistic ↓ C (kg m− 3) Slag (kg m− 3) FA (kg m− 3) W (kg m− 3) SP (kg m− 3) Cagg (kg m− 3) Fagg (kg m− 3) Age (days) CS (MPa)

​ X1 X2 X3 X4 X5 X6 X7 X8 Y
Mean 281.17 73.90 54.19 181.57 6.20 972.92 773.58 45.66 35.82
Standard Deviation 104.5 86.3 64.0 21.4 6.0 77.8 80.2 63.2 16.7
Minimum 102.0 0.0 0.0 121.8 0.6 801.0 594.0 1.0 2.33
Median 272.9 22.0 0.0 185.0 6.4 968.0 779.5 28.0 34.4
Maximum 540.0 359.4 200.1 247.0 32.2 1145.0 992.6 365.0 82.6
Count 1030 1030 1030 1030 1030 1030 1030 1030 1030

For instance, the input X1, with a mean of 281.17 and a standard 5.97, exhibits a multi-modal distribution with substantial skewness
deviation of 104.51, shows a relatively normal distribution with slight (0.91) and kurtosis (1.41), indicating complex underlying data struc­
positive skewness (0.51) and moderate kurtosis (− 0.32), suggesting a tures. The input X6, with a mean of 972.92 and a standard deviation of
concentration of data points around the mean with a few outliers. The 77.58, displays a near-normal distribution, albeit with slight positive
input X2, having a mean of 78.90 and a standard deviation of 86.28, skewness (0.44) and kurtosis (− 0.60). The input X7′s mean of 773.58
reveals more pronounced positive skewness (0.83) and kurtosis (− 0.41), and standard deviation of 20.18, along with a high kurtosis value
highlighting a higher frequency of lower values and the presence of (− 0.10) and slight negative skewness (− 0.25), suggest a sharp peak
significant outliers. The input X3, with a mean of 54.19 and a standard around the mean. The input X8, characterized by a mean of 45.66 and a
deviation of 64.00, demonstrates a bimodal distribution indicative of standard deviation of 63.17, presents a complex distribution with high
two dominant data clusters, also reflected in its negative kurtosis skewness (2.77) and kurtosis (7.17), indicating a high frequency of
(− 1.33). Similarly, the input X4′s mean of 181.57 and standard devia­ lower values and the presence of significant outliers. Lastly, the response
tion of 21.86, coupled with near-zero skewness (0.17) and kurtosis output (Y), with a mean of 35.82 and a standard deviation of 16.71,
(0.12), point to a distribution that is fairly symmetrical but slightly exhibits a relatively normal distribution with moderate positive skew­
leptokurtic. The input X5, with a mean of 6.20 and a standard deviation ness (0.31) and kurtosis (− 0.31), suggesting a well-centred dataset with


A.K. Hamed et al. &RPSXWHUV DQG 6WUXFWXUHV   

Fig. 2. KDE plots of input and output variables.


A.K. Hamed et al. &RPSXWHUV DQG 6WUXFWXUHV   

some outliers. form linear solid relationships with one another. Since there are no
In summary, most of the variables (Slag, FA, SP, and Age) display a uncorrelated features, we can conclude that all eight input parameters
noticeable positive skew, indicating a higher concentration of lower help predict concrete CS.
values with fewer higher values. Conversely, the variables W, Cagg,
Fagg, and CS show more evenly distributed data, with central tendencies 4.4. Hixbin graphs
being more apparent.
Fig. 4 presents hexbin plots revealing important relationships be­
4.3. Correlation analysis tween the dependent variable (CS of concrete) and the various input
variables. For cement vs. CS, the hexbin plot shows a strong positive
The objective of this analysis is to find and determine the best pre­ correlation. As cement content increases, the CS of concrete also in­
diction model by dissecting inter-variable. It is essential to comprehend creases. The data points show a high density in the middle to upper
these relationships in order to fully appreciate the connections between ranges of cement. The concentrated clustering around the higher cement
the desired strength factor and its dependent features. The Pearson values suggests that cement significantly and consistently impacts
correlation coefficient (r) is the most popular metric for this approach increasing CS. Higher cement content correlates strongly with increased
since it provides insight into these connections [79,80]. It is defined in CS, with the densest areas appearing in the higher range of cement
Eq. (1) as the ratio of the product of the standard deviations of two content. The plot shows a moderate positive correlation for Blast
variables (x, y) to their covariance (cov). Furnace Slag vs. CS. While there is a noticeable upward trend, the data
∑n points are more dispersed compared to cement. The highest concen­
cov (x, y) i=1 (xi − x)(yi − y) tration of data points occurs in the middle range of Slag, implying that
r= = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
̅√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ (1)
σx σy ∑n ∑ n moderate amounts of Slag contribute most significantly to CS. This
(xi − x)2 (yi − y)2 suggests that while Slag positively impacts CS, its effect is most notice­
i=1 i=1
able at intermediate levels.
The average value of two variables, x and y, is denoted as x‾ and y‾, The fly ash (FA) plot shows a weak positive correlation with CS. The
respectively, while the number of datasets is denoted by (n). A heatmap data points are scattered across the plot with a lower density in the
showing the interdependence of all variables is shown in Fig. 3. higher ranges of fly ash. The density appears to be more concentrated in
According to the heatmap, the variables C, SP, and Age exhibit the the lower to middle ranges of FA, which suggests that while fly ash may
strongest positive correlations, with r-values of 0.498, 0.366, and 0.329, improve CS slightly, its effect is less pronounced and more variable,
respectively. These values indicate that these variables are the most particularly at higher levels. Meanwhile, the hexbin plot of water vs. CS
influential. This suggested that It becomes increasingly apparent that the reveals a negative correlation. As water content increases, CS decreases.
addition of cement, superplasticizer, and curing time significantly im­ The data points are most concentrated in the lower to middle ranges of
pacts the compressive strength of concrete as the amount of these ad­ Water, with high data density around lower water content values. This
ditives increases. Given that the most pronounced negative correlation supports the idea that lower water content correlates with higher CS,
between X4 and X5 (− 0.657) is found between the two variables, it whereas higher water content leads to weaker concrete, as indicated by
would appear that water and superplasticizers have a relationship that is the concentration of points in the lower ranges. The Superplasticizer
opposing to one another. In addition, the coefficient of correlation CS of (SP) versus CS plot reveals a moderate positive correlation. The data
concrete and water have a strong negative correlation, with an r-value of points show a concentration in the middle to upper ranges of SP. This
− 0.290. There is only a weak relationship between the remaining var­ suggests that higher amounts of SP contribute to improved CS, likely due
iables and the concrete’s CS, which suggests that the variables do not to better workability without sacrificing strength.

Fig. 3. Heatmap correlation matrix for all studied parameters.


A.K. Hamed et al. &RPSXWHUV DQG 6WUXFWXUHV   

Fig. 4. Hexbin graphs of the relationship between input variables and CS of concrete.

The coarse aggregate (Cagg) plot shows a moderate positive corre­ 4.6. Description of base models
lation with CS. The data points are concentrated around the middle to
higher ranges of Cagg. This suggests that a larger proportion of Cagg 4.6.1. ANN model
tends to improve CS, likely due to better packing and structural integrity The concept of Artificial Neural Networks (ANNs), inspired by the
of the concrete mix. The upper-middle-range of Cagg shows the densest neural networks in the human brain, was first developed by McCulloch
data points, emphasizing its role in enhancing strength. The hexbin plot and Pitts in 1943 [81]. ANNs are adept at modeling linear and nonlinear
for fine aggregate (Fagg) shows a weak positive correlation with CS. The systems without the constraints often associated with traditional sta­
data points are scattered widely, with low density at higher ranges of tistical methods. Typically, an ANN comprises three layers: an input
Fagg. The concentration of points in the lower-middle ranges suggests layer, one or more hidden layers, and an output layer. The training
that Fagg has a minor effect on CS, especially in comparison to other process of an ANN involves adjusting connection weights according to a
components such as cement or coarse aggregate. Lastly, the curing learning rule derived from an input–output dataset, with the goal of
duration (Age) plot shows a clear positive correlation with CS, as ex­ minimizing the difference between predicted outputs and actual values
pected. The data points form a tight upward trend, with a high con­ [82]. This training process includes two main phases: feed-forward and
centration in the higher ranges of (Age), particularly in the upper- back-propagation. During the feed-forward phase, data flows from the
middle to higher ranges. This indicates that longer curing durations input layer through the hidden layers to the output layer. In the back-
are directly associated with increased CS, confirming that prolonged propagation phase, weights are adjusted from the output layer back to
curing allows concrete to reach its full potential strength. These insights the input layer based on the overall error in the network. To further
suggest that optimizing cement and curing duration, along with mod­ reduce this error, an appropriate back-propagation algorithm is
erate levels of slag and superplasticizer, can significantly enhance CS, employed, with the Levenberg-Marquardt technique being the most
while excessive fly ash and fine aggregate may offer diminishing returns. efficient and commonly used method for this purpose. Fig. 5 shows that
the eight variables were used as inputs to the ANN model, and the output
4.5. Data normalization variable Y, which represents the concrete CS, was set as the outcome.

When input data is of varying sizes, some ML models might not 4.6.2. RF model
perform as well as they could. The significantly varying magnitudes of The Random Forest (RF) model, initially introduced by Ho [83] and
the various input features are shown in Table 2, where the cement further developed by Breiman [84], is an advanced ensemble learning
ranges from 102 to 540 kg/m3, and the superplasticizers range from 0 % method that utilizes multiple decision trees to enhance prediction ac­
to 32.2 %. To fix this, we scale all the input variables to the same value curacy. This technique applies bootstrap resampling to create various
using data normalization or rescaling. The max–min mapping function is instances from the original dataset, with each instance used to construct
used in this process, as shown in Eq. (2). an individual decision tree [85]. Predictions from these trees are then
averaged to produce the final output. RF improves on the limitations of
X − Xmin
Xn = (2) traditional decision trees, such as susceptibility to overfitting, by
Xmax − Xmin
incorporating randomness in the selection of parameters during the tree
Every input variable in this equation has a minimum (Xmin) and construction. This not only makes the model more robust by combining
maximum (Xmax) value, and Xn represents the normalized data. The well-established trees but also balances the weaknesses and strengths
original dataset that is being rescaled is denoted by X. The main across different classes. The model structures condition hierarchically
advantage of data rescaling is that it can speed up computations and from the root to the terminal nodes, with decision-making processes at
make the ML prediction model more accurate and stable. each node and a simple regression model at each terminal node to make
predictions. Pruning is applied to reduce complexity and enhance the
model’s generalization capabilities. Fig. 6 shows the RF model’s


A.K. Hamed et al. &RPSXWHUV DQG 6WUXFWXUHV   

Fig. 5. Schematic representation of the ANN model development.

second-order gradients to accelerate convergence. Additionally, XGB is

proficient at handling missing data while making split decisions and
incorporates stochastic gradient descent to enhance model diversity and
further mitigate overfitting. Moreover, XGB includes system-level im­
provements, including cache optimization and parallel processing,
which contribute to its computational efficiency and its ability to
manage large datasets effectively. The architecture of the XGB model is
depicted in Fig. 7.

Fig. 6. Architecture of RF model [86].

architecture.

4.6.3. XGB model

The XGB model, renowned for its implementation of gradient-
boosted decision trees, is specifically engineered to achieve excep­
tional speed and performance. The algorithm is characterized by its
exceptional flexibility and versatility, making it particularly adept at
managing sparse data and efficiently performing both regression and
classification tasks. Its scalability and capacity to generate very
competitive predictive models have enabled its widespread use in many
ML competitions [85]. This study employs XGB due to its beneficial
features, including regularization to mitigate overfitting and the use of
Fig. 7. Architecture of XGB model [87].


A.K. Hamed et al. &RPSXWHUV DQG 6WUXFWXUHV   

4.7. Hyperparameters tuning Both the training and validation datasets were used to develop and test
the BMs. The SM, on the other hand, was built using validation-stage BM
In this study, Bayesian Optimization (BO) is employed to optimize predictions. Then, the remaining 10 % of the database was used to test
hyperparameter settings, offering a more efficient alternative to tradi­ its performance. This way, we can test the model thoroughly on data it
tional methods like Grid Search (GS) and Random Search (RS) [88]. has never seen before while still training it on a statistically valid subset
Unlike GS, which exhaustively searches the entire hyperparameter of the data. However, when comparing and evaluating the adopted
space, BO models the objective function’s prior distribution and refines models, two popular approaches are quantitative and visual methods
the search to identify the optimal configuration. The BO process begins [92].
with defining the hyperparameter ranges and initial distributions for Visual methods, such as scatter plots, violin boxplots, and Taylor
each model, aiming to maximize performance within these predefined diagrams, provide quick insights into the accuracy of various statistical
bounds. By reducing unnecessary trials, BO accelerates the identification measures, including maximum, minimum, median, and quartiles
of the most effective hyperparameter combinations, significantly [93,94]. These visual tools offer rapid, engaging, and informative, easily
enhancing tuning efficiency [89]. Critical hyperparameters, such as the interpretable comparisons. However, they may not provide detailed
number of estimators, learning rate, maximum tree depth, and regula­ information about overall model performance [95,96]. Seven quanti­
rization parameters, were optimized using BO. This systematic approach tative metrics were employed to provide detailed numerical insights into
minimizes trial-and-error, ensuring that the models achieve optimal model performance. These performance metrics equations and their
performance. The selected configurations were validated on the test ideal values are listed in Table 3.
dataset to confirm their predictive efficacy. The A20 index serves as a crucial metric for assessing the reliability
Furthermore, this research integrates k-fold cross-validation (CV) to and accuracy of predictive models [97,98]. The metric quantifies the
assess model robustness and mitigate overfitting. In particular, a 10-fold percentage of predictions that lie within a 20 % error margin of the
CV is utilized, where the dataset is partitioned into ten subsets, with actual values, providing a straightforward assessment of the model’s
each serving as the test set while the remaining nine are used for performance. The A20 index is calculated as follows.
training. This method, known for balancing computational efficiency m20
and the bias-variance trade-off, was combined with BO for hyper­ A20 = (3)
n
parameter optimization (BO + 10CV). The use of a 10-fold CV ensures
accurate performance estimation while maintaining manageable where m20 represents the number of predictions where the predicted
computational demands. value is within 20 % of the actual value. A lower A20 value indicates
better predictive reliability, as it suggests that a greater proportion of the
4.8. Description of the stacked linear regression model model’s predictions are close to the actual values.

Stacking is an ensemble learning technique designed to enhance the 4.10. Interpretation of studied features
predictive accuracy of machine learning models by combining multiple
individual models [90,91]. The fundamental idea is to train several base To evaluate and interpret machine learning models at both broad
models (BMs), often referred to as level-0 models, on the same dataset and detailed levels, researchers employ the SHapley Additive exPlana­
and then use another model, known as the meta-model or level-1 model, tions (SHAP) method, which is grounded in cooperative game theory
to aggregate their predictions into a final prediction. In a stacking principles [99]. This method is used to assess the comparative influence
ensemble, multiple BMs are trained independently on the training data. of input variables on the prediction process, providing a sophisticated
These models make predictions on the validation dataset, generating tool within the field of explainable AI. SHAP facilitates the clarification
several forecasts. These predictions form a new dataset, where each of the intricate relationships between input variables and estimates
column represents the predictions from a different base model. The made by models, offering critical insights into which features most
meta-model is then trained on this new dataset to learn how to combine significantly influence the predictions and how they do so [100].
the base model predictions into a final output. On the other hand, the current study examined a Partial Dependence
When linear regression is used as the meta-model in a stacking Plot (PDP) to illustrate the functional link between a limited set of input
ensemble, the process starts with training the base models: ANN, RF, and parameters and the model’s predictions. It illustrates the extent to which
XGB on the training dataset. These BMs make predictions on the vali­
dation dataset, creating a new dataset composed of these predictions. A Table 3
linear regression model is trained on this new dataset, using the actual Equations of performance metrics and their ideal values.
target values from the validation data as the response variable. The Metric Equation Ideal
linear regression model learns to weigh the contributions of each base Value
model’s predictions to produce a final, combined prediction.
Determination R2 = ​ 1
Stacking with a linear regression meta-model can significantly coefficient (R2) ∑n ( )2 ∑n ( )2
yi − ymean − i=1 yi − ŷi
improve model performance by leveraging the strengths of various BMs. i=1
∑n ( )2
yi − ymean
This method allows for flexibility in model choice and can capture √i=1̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
Root Mean Square Error ∑n ( )2 ​ 0
diverse patterns in the data, making the final model more robust and i=1 yi − y ̂i
(RMSE) RMSE =
n
accurate. Additionally, by combining multiple models, stacking reduces ∑ n ⃒⃒ ⃒
Mean Absolute Error yi − ŷi ⃒ ​ 0
the risk of overfitting, as the meta-model balances the strengths and (MAE) MAE = i=1
⃒ n ⃒
weaknesses of each base model. By employing stacking with linear Mean Absolute ∑ ⃒⃒yi − ŷi ⃒⃒ ​ 0
regression, we achieve a more accurate and robust predictive model that Percentage Error
MAPE =
⃒ yi ⃒
× 100
benefits from the diverse insights provided by the base models. This (MAPE) n
∑( )
Mean Bias Error (MBE) ŷi − yi 0
approach is particularly effective in complex datasets where no single
​
MBE =
n
model performs optimally. Cumulative Prediction
∑ ⃒⃒
ŷi − yi ⃒
⃒
​ 0
Error Index (CPI) CPI = ∑
yi
( )
4.9. Evaluation methods of adopted models Variance Accounted For Var yi − ŷi ​ 1
VAF =1 − ( )
(VAF) Var yi
The dataset was carefully divided into three separate sets: training, Where n is the dataset number yi and ŷi are actual and predicted ith values,
respectively.
validation, and testing, with respective weights of 80 %, 10 %, and 10 %.


A.K. Hamed et al. &RPSXWHUV DQG 6WUXFWXUHV   

the predictions are influenced by the values of these specific input pa­ The RF model, optimized for this study, consists of 313 estimators,
rameters [101]. Moreover, PDPs highlight the impact of each parameter with each tree having a maximum depth of 23. The model uses a min­
on the predicted outcomes generated by a machine learning algorithm. imum samples split of 2.0 and a minimum samples leaf of 1.0, ensuring
PDPs consider all instances within the dataset as a global method, that each split and leaf has a sufficient number of samples to avoid
revealing the overall relationship between a feature and the forecasted overfitting. The XGB model is fine-tuned with 493 estimators and a
outcome [102]. These plots offer valuable insights into the relative learning rate of 0.067, balancing the speed of learning and convergence.
contribution of each input variable to the predicted outcome. Further­ The model’s maximum depth is set to 4.0, allowing it to capture in­
more, one-dimensional PDPs (PDPs-1D) depict the relationship between teractions between features. Additionally, a column sample with tree
the expected result and a single input parameter. value of 0.724 and a sub-sample value of 0.40 are used to prevent
overfitting by adding randomness to the model training process. These
4.11. Graphical user interface (GUI) optimized hyperparameters enhance the performance and accuracy of
the respective ML models, enabling them to deliver robust and reliable
A user-friendly GUI has been developed to make the predictive predictions.
model for estimating CS easily accessible. This interface, built using
Python’s Tkinter library [103], leverages its simplicity and standard 5.2. Quantitative assessment of BMs
distribution for creating interactive applications [104]. The develop­
ment process involves setting up the Python environment, with the 5.2.1. ANN model
design focused on user guidance and ease of use. Key components Fig. 8 presents a scatter plot comparing actual versus predicted
include input fields (TextEntry widgets) for user data entry, a Button values using an ANN model. The plot includes data points for both the
widget to trigger predictions, and a Label or Text widget to display the training (TR) stage, represented by blue hexagons, and the validation
predicted CS. The predictive model is seamlessly integrated into the GUI (VD) stage, represented by orange hexagons. These points are distrib­
using the pickle library to load and apply the model based on user in­ uted around the equality line (0 % error), which denotes a scenario
puts. To facilitate accessibility, collaboration, and distribution, the GUI where predicted values perfectly match actual values.
is hosted on GitHub, providing version control and easy sharing. In the TR stage, the performance metric of the R2-value is 0.972,
indicating a strong linear relationship between the actual and predicted
5. Results and discussions values. The error metrics are as follows: RMSE of 2.776 MPa, MAE of
1.169 MPa, MAPE of 4.23 %, and MBE of 0.194 MPa, suggesting that the
5.1. Optimal hyperparameters of developed BMs model exhibits good prediction accuracy with a slight overestimation
bias. Additionally, the CPI is 0.033, and the VAF is 0.973, further
Hyperparameters play a crucial role in the performance and effi­ highlighting the model’s robust performance. In the VD stage, the R2-
ciency of machine learning models. Unlike model parameters, which are value decreases from 0.972 to 0.913, which still indicates a strong
learned from the data during training, hyperparameters are set prior to linear relationship. However, the RMSE increases to 5.270 MPa and the
the training process and guide the learning process. The selection of MAE to 3.625 MPa, suggesting that the model’s predictions are less
appropriate hyperparameters can significantly influence the model’s accurate in the validation set compared to the training set. The MAPE is
ability to generalize and make accurate predictions on unseen data. For 13.60 %, and the MBE is 0.107 MPa, indicating a reduced prediction
the ANN, the optimal configuration includes three hidden layers, each bias. The CPI is 0.103, and the VAF is 0.913, reflecting good model
with 20 neurons. Both the hidden layers and the output layer use the performance on unseen data.
sigmoid activation function, which helps achieve non-linearity and en­ Additionally, Fig. 8 includes two dashed lines representing ± 10 %
sures the model can capture complex patterns in the data. deviation from the equality line, which helps visually assess the model’s

Fig. 8. Scatter plot comparing the actual and predicted CS-values based on the ANN model.


A.K. Hamed et al. &RPSXWHUV DQG 6WUXFWXUHV   

prediction accuracy. The majority of data points fall within these and 0.155 MPa in RMSE and MAE, respectively. This suggests that the
bounds, indicating that most predictions are within 10 % of the actual RF model had better prediction accuracy. In the VD stage, the RF model
values. This visualization underscores the model’s effectiveness in pre­ also shows a lower RMSE by 0.92 MPa compared to the ANN model and
dicting values that closely match the actual data with a relatively small a slightly lower MAE by 0.245 MPa. This indicates that the RF model has
margin of error. Overall, the ANN model demonstrates strong perfor­ a lower overall error, while the ANN model may be more susceptible to
mance on the training set and maintains good prediction accuracy on the larger individual errors. Finally, both the RF and ANN models have
validation set. The performance metrics indicate a robust linear rela­ similar MAPE values during the TR stage (RF = 4.60 % and ANN = 4.23
tionship and low error rates, with most predictions falling within a ± 10 %). However, when tested with unseen data in the VD stage, the RF
% deviation from the actual values. model shows a lower MAPE, with a difference of 1.86 % compared to the
ANN model.
5.2.2. RF model In summary, while both RF and ANN models perform well, the RF
Fig. 9 displays a scatter plot comparing actual versus predicted model performs slightly better in capturing variance and minimizing
values using the RF model. This plot shares the same characteristics as errors, particularly in the VD stage. The ANN model, though robust,
the previous plot for the ANN model. During the TR stage, the R2-value shows marginally higher errors and lower correlation metrics, especially
is 0.985, indicating an extremely strong correlation between the actual in the VD stage.
and predicted values. The error metrics are RMSE = 2.049 MPa and MAE
= 1.324 MPa, demonstrating the model’s high prediction accuracy. 5.2.3. XGB model
Additionally, the MAPE and MBE values are 4.60 % and − 0.077 MPa, Fig. 10 presents a scatter plot comparing actual versus predicted
respectively, indicating a slight underestimation bias. The CPI is 0.037, values using the XGB model. This plot shares the same features as the
and the VAF is 0.985, further underscoring the model’s strong previous ones for the ANN and RF models. During the TR stage, the R2-
performance. value indicates an exceptionally strong linear relationship between the
In the VD stage, the R2 value is 0.932, slightly lower than in the TR actual and predicted values, demonstrating that the model’s predictions
stage. The RMSE and MAE values are 4.350 MPa and 3.118 MPa, closely align with the actual values. The RMSE (2.002 MPa), MAE
respectively, suggesting that the RF model’s prediction errors have (1.384 MPa), and MAPE (5.05 %) are relatively low, suggesting that the
increased compared to the TR stage. The CPI is 0.088, and the VAF is XGB model exhibits a low average error. However, the errors increase
0.932, reflecting solid performance on unseen data. Moreover, most slightly in the VD stage compared to the TR stage, indicating possible
predictions fall within ± 10 % deviations from the equality line, high­ overfitting. The CPI is 0.038, and the VAF is 0.986, further highlighting
lighting the model’s effectiveness in accurately predicting values with a the model’s strong performance.
relatively small margin of error. In the VD stage, the R2-value shows a minimal decrease from 0.986
Several observations can be made when comparing the RF model to 0.953, indicating a good fit. The RMSE and MAE values of 3.859 MPa
with the previous ANN model. Both models exhibit strong linear re­ and 2.715 MPa, respectively, suggest that the XGB model’s prediction
lationships between actual and predicted values, with high correlation errors have slightly increased compared to the TR stage. The MAPE was
coefficients and R2 values in both the TR and VD stages. However, the RF 9.58 %, and the MBE was − 0.070 MPa, indicating a slight prediction
model shows slightly higher correlations during the TR and VD stages, bias. The CPI is 0.077, and the VAF is 0.953, reflecting good model
with an increase in R2 of 1.30 % and 1.90 %, respectively, compared to performance on unseen data. Additionally, most data points fall within
the ANN model. This indicates that the RF model captures the variance the ± 10 % bounds, indicating that the majority of predictions are
in the TR stage slightly better than the ANN model. Regarding the error within 10 % of the actual values. This visualization underscores the
metrics, the RF model has lower RMSE and MAE values than the ANN model’s effectiveness in predicting values that closely match the actual
model in the TR stage, with the ANN model being higher by 0.727 MPa data with a relatively small margin of error. Several observations can be

Fig. 9. Scatter plot comparing the actual and predicted CS-values based on the RF model.


A.K. Hamed et al. &RPSXWHUV DQG 6WUXFWXUHV   

Fig. 10. Scatter plot comparing the actual and predicted CS-values based on the XGB model.

made when comparing the XGB model with the previously discussed 5. The MBE for all models indicates minimal prediction bias, with the
ANN and RF models: XGB model displaying the least bias. The CPI and VAF metrics
consistently indicate strong model performance across all models,
1. All three models demonstrate strong linear relationships between with the RF and XGB models slightly outperforming the ANN model.
actual and predicted values, with R2-values greater than 0.91 during
both the TR and VD stages. In summary, the XGB model demonstrates slightly better overall
2. The XGB model, followed by RF, has slightly higher R2 values than performance than the others. The RF model performs comparably, with
the ANN model, suggesting a better capture of variance in both the substantial prediction accuracy and minimal bias. However, The ANN
TR and VD stages. model shows a larger deviation, suggesting it may be more prone to
3. The RF and XGB models exhibit similar RMSE and MAE values overfitting than the other two models.
during the TR stage, which are lower than those of the ANN model,
indicating better prediction accuracy. In the VD stage, the XGB
model shows lower RMSE than both the ANN and RF models, indi­ 5.3. Visual assessment of BMs
cating superior performance in these error metrics.
4. The MAPE values for the RF and XGB models are approximately the 5.3.1. Violin plots
same during the TR stage and are lower than those for the ANN The violin plots provide a visual comparison of the distribution of
model. In the VD stage, the XGB model exhibits lower MAPE than predicted values from various models against the actual values. A model
both the ANN and RF models, suggesting that the XGB model may whose predicted values closely match the shape and spread of the actual
generalize better in terms of MAPE. values’ violin plot is considered more accurate. These plots also help
identify significant deviations or biases in the model predictions. Fig. 11

Fig. 11. Violin plots comparing the predicted CS-values from the BMs during the (a) TR and (b) VD stages.


A.K. Hamed et al. &RPSXWHUV DQG 6WUXFWXUHV   

features two violin plots comparing the predicted CS-values from stages. The statistical results, including SD and CC, are derived from
different BMs during the TR and VD stages. The CS values are plotted these diagrams.
along the y-axis, while the x-axis represents various models. Each violin For the ANN model, the SDs are 16.94 MPa in the TR stage and 19.15
plot displays the kernel density estimation of the data distribution, with MPa in the VD stage, with corresponding CCs of 0.986 and 0.967. The RF
wider sections indicating a higher density of data points. The central model exhibits SDs of 16.10 MPa in the TR stage and 16.70 MPa in the
white dot signifies the median value, while the thick band encompassing VD stage, with CCs of 0.993 and 0.972, respectively. The XGB model
the upper and lower values denotes the interquartile range (IQR). shows SDs of 16.50 MPa in the TR stage and 17.35 MPa in the VD stage,
In the TR stage (Fig. 11a), the RF and XGB models are identified as with CCs of 0.993 and 0.976. The results clearly indicate that the RF and
the top-performing models, exhibiting centered medians of 35.35 and XGB models outperform the others in terms of SD and CC values. The
35.45, respectively, which closely align with the actual CS of 35.20 MPa. XGB model, in particular, has the highest CC and the smallest centered
The plots exhibit symmetry relative to the actual data, suggesting min­ RMSD, demonstrating its superior ability to capture the variance and
imal variance. The interquartile ranges are 24.13 to 45.30 MPa for the trends in the data accurately. The RF model, while slightly trailing the
RF model and 24.32 to 45.19 MPa for the XGB model, both closely XGB model, still performs strongly with a high CC and low centered
aligning with the actual CS range of 23.72 to 46.22 MPa. The compact RMSD. The ANN model, however, exhibits a higher centered RMSD and
distribution indicates that the RF and XGB models exhibit a strong fit to slightly lower CC than the XGB and RF models, suggesting that although
the training data. The ANN model demonstrates slightly broader yet still it captures the general trends, it is less precise in matching the actual
well-structured distributions. The data exhibited a median of 35.60 and data distribution and has a larger prediction error. Thus, the visual as­
an IQR from 23.78 to 46.14 MPa, showing moderate variance alongside sessments conducted highlight the efficacy of the XGB and RF models in
reliable performance. accurately and consistently predicting CS-values.
During the VD stage (Fig. 11b), the XGB and RF models exhibit su­
perior performance, characterized by narrow and symmetric distribu­ 5.3.3. Rank analysis
tions that effectively generalize from the training data, with XGB A rank analysis was conducted to evaluate the overall performance of
showing a slight advantage over RF. The XGB model exhibits a median of the developed base models using the performance metrics presented in
32.34 MPa with an IQR spanning from 22.11 to 47.21 MPa. The RF Table 3. In this analysis, a model attaining an overall rank value of 1
model shows a median of 32.09 MPa and an IQR from 21.75 to 47.28 signifies the highest level of performance, whereas a rank of 3 denotes
MPa. Both models demonstrate low variance relative to the actual the lowest level of performance. The sum of the individual ratings across
CS values, which have a median of 31.37 MPa and an IQR from 19.77 to various metrics is used to determine the overall ranking of each model.
46.39 MPa. The ANN model demonstrates greater variance, presenting a The model exhibiting the greatest cumulative rank is regarded as the
median of 33.05 MPa and an interquartile range from 19.13 to 47.59 least efficient, while the model exhibiting the smallest cumulative rank
MPa, indicating reduced generalization capability. is defined as the most efficient. An analysis of the rankings for both the
TR and TS stages is presented in Table 4.
5.3.2. Taylor diagrams During the TR stage, the XGB model emerges as the most exceptional
The second visual assessment involved using Taylor diagrams, which performer, achieving a total score of 13 by consistently ranking first
graphically compare the statistical properties of different models with across all seven performance metrics. The RF model follows closely with
the actual data. These diagrams feature axes for standard deviation (SD) a total score of 14, ranking second overall and demonstrating strong
and correlation coefficient (CC), with the actual data represented as a performance. The ANN model secures the third spot with a score of 15.
reference point. The radial distance from the origin indicates the SD, the All models developed in the TR stage exhibited robust predictive capa­
angle represents the CC, and the distance from the actual point reflects bilities, as evidenced by their close scores. In the TS stage, the XGB
the centered Root Mean Square Difference (RMSD) between the model model emerges as the most effective, achieving the lowest total score of
predictions and the actual values. Fig. 12 displays the Taylor diagrams 8, indicating top performance across all metrics. This consistency from
comparing predicted CS-values from the BMs during the TR and VD the training to the testing stages underscores its robustness and

Fig. 12. Taylor diagrams comparing predicted CS-values from the BMs during the (a) TR and (b) VD stages.


A.K. Hamed et al. &RPSXWHUV DQG 6WUXFWXUHV   

Table 4
Rank analysis for the developed BMs.
Model Stage R2 RMSE (MPa) MAE (MPa) MAPE MBE CPI VAF Total Score Overall Rank

ANN TR 3 3 1 1 3 1 3 15 36
VD 3 3 3 3 3 3 3 21
RF TR 2 2 2 2 2 2 2 14 27
VD 2 2 2 2 1 2 2 13
XGB TR 1 1 3 3 1 3 1 13 21
VD 1 1 1 1 2 1 1 8

reliability. The RF model follows in second place with a score of 13, 5.4. Prediction of concrete CS via SM-LR
demonstrating dependable performance in real-world testing scenarios.
The ANN model remains the least effective in the testing phase, with a The prediction of concrete CS using a Stacked Model-based Linear
score of 21, consistent with its training performance, highlighting sig­ Regression (SM-LR) involves integrating the outputs of three distinct
nificant challenges in predictive accuracy. Overall, the XGB model dis­ base models: ANN, RF, and XGB. Through a systematic trial-and-error
tinguishes itself among three BMs, achieving an impressive overall score process, a linear equation was derived that optimally combines the
of 21. The RF model holds the second rank with a score of 27. Lastly, the predictions from these individual models into a cohesive ensemble
ANN model consistently ranks lowest, with notably high scores of 36, model. The equation for the SM-LR model is as follows:
indicating the need for further improvement in its predictive
CS(SM− LR) = − 0.7122 + 0.1877 CS(ANN) + 0.4506 CS(RF) + 0.3814 CS(XGB)
capabilities.
(4)
In this equation, the coefficients associated with each BM indicate

Fig. 13. Scatter plots during the TS stage for (a) ANN, (b) RF, (c) XGB, and (d) SM-LR models.


A.K. Hamed et al. &RPSXWHUV DQG 6WUXFWXUHV   

their relative contributions to the final prediction. The ANN model 5.5.2. Visual assessment
contributes 0.1877 to the overall prediction, RF provides the most Fig. 14a presents violin plots comparing the predicted CS-values
substantial weight of 0.4506, and XGB adds a contribution of 0.3814. from the base models and the SM-LR during the TS stage. The distri­
The constant term (− 0.7122) adjusts the predictions for any inherent bution of the SM-LR model closely matches the actual values, with the
biases in the model. SM-LR exhibiting a median of 33.72 and an IQR from 24.58 to 44.95. In
contrast, the actual testing dataset shows a median of 33.21 and an IQR
from 24.92 to 45.85, suggesting robust predictive capability relative to
5.5. Comparison between the BMs and SM-LR all BMs employed. The XGB and RF models follow the SM-LR, exhibiting
distributions that closely resemble the actual testing values. The CGB
5.5.1. Quantitative assessment model has a median of 34.4 and an IQR spanning from 24.81 to 43.68,
Fig. 13 presents scatter plots illustrating the performance of four whereas the RF model has a median of 33.47 and an IQR range from
models (ANN, RF, XGB, and SM-LR) in predicting CS values during the 24.65 to 44.84. Unexpectedly, the ANN model demonstrates a reduced
testing (TS) stage. Each plot compares actual CS values with predicted deviation in the median and IQR when compared during the TR and VD
values, with individual data points representing specific predictions. stages, with a median of 33.73 and an IQR spanning from 24.41 to
The plots include a dashed black line representing perfect predictions 45.65; however, it remains inferior to the XGB and RF models in
(equality line) and dashed red and green lines indicating ± 10 % devi­ ranking.
ation from the equality line. The models’ performance was evaluated Fig. 14b displays Taylor diagrams comparing the predicted CS-values
using the R2 and RMSE metrics displayed on each plot. among the base models and SM-LR in the TS stage. The ANN model
The ANN model (Fig. 13a) demonstrates the lowest performance shows an SD of 16.10 MPa and a CC of 0.946, while the RF model has an
among the four models, with an R2-value of 0.884 and an RMSE of SD of 14.20 MPa and a CC of 0.955. The XGB model exhibits an SD of
5.203 MPa. While many data points fall within the deviation lines, sig­ 14.60 MPa and a CC of 0.972, and the SM-LR model demonstrates an SD
nificant deviations suggest that the ANN model’s prediction accuracy of 14.80 MPa and a CC of 0.984. These results indicate that the SM-LR
could be improved. In contrast, the RF model (Fig. 13b) shows a higher model performs exceptionally well, with a high CC and low centered
R2-value of 0.923, indicating better predictive performance than the RMSD, closely followed by the XGB model. The RF model also shows
ANN model, and an RMSE of 4.229 MPa, reflecting a lower error rate strong performance but has a slightly lower correlation coefficient and a
and improved prediction accuracy. A more significant proportion of the higher centered RMSD than the SM-LR and XGB models. The ANN
RF model’s data points lie within the deviation lines, underscoring its model, however, shows the lowest CC and the highest centered RMSD
solid performance in accurately predicting CS values. The XGB model among the four models, suggesting it is less precise in matching the
(Fig. 13c) achieves an R2-value of 0.947, indicating a very high corre­ actual data distribution and has a larger prediction error compared to
lation between actual and predicted values. The model’s RMSE is 3.524 the top-performing models.
MPa, signifying a lower error rate and superior prediction accuracy In summary, quantitative analysis and visual plots confirm that the
compared to both the RF and ANN models. Data points for the XGB SM-LR model outperforms the other models by accurately capturing the
model are closely clustered around the equality line, highlighting its statistical properties of the data and minimizing prediction errors. The
robust performance. The SM-LR model (Fig. 13c) outperforms the other XGB model closely follows the SM-LR model in performance, while the
models, with the highest R2-value of 0.953 and the lowest RMSE of RF model demonstrates good accuracy but slightly higher errors.
3.315 MPa, suggesting the best fit between actual and predicted values, Although still effective, the ANN model shows comparatively lower
as well as the smallest prediction errors. Most data points for the SM-LR accuracy and higher error rates.
model are tightly clustered around the equality line, underscoring its
exceptional predictive accuracy. 5.5.3. K-fold CV analysis
Additionally, the A20 results for the tested models indicate signifi­ A more accurate assessment of model efficacy is achieved through k-
cant differences in their predictive performance, allowing for an fold cross-validation (CV), which reduces the likelihood of overfitting to
assessment of the reliability of the adopted models. The XGB model specific subsets of the data. By mitigating the risk associated with a
recorded the lowest A20 index at 10.07, indicating it provides the most single train-test split, k-fold CV provides a more robust evaluation of
accurate and reliable predictions compared to the other models. The SM- model performance [105]. The results indicate that the models exhibit
LR model exhibits a low A20 value of 10.08, indicating its reliability is both high accuracy and reliability. Fig. 15 illustrates a comparison of the
nearly comparable to that of XGB. The two models are the most accurate developed models utilizing 10-fold across 10-fold CV using five perfor­
and reliable for the TS dataset. The RF model exhibited moderate per­ mance metrics: R2, RMSE, MAE, MAPE, and MBE. For the R2 metric
formance, achieving an A20 index of 13.07. Although this surpasses XGB (Fig. 15a), all models show high R2 values across all folds, with SM-LR
and SM-LR, it still reflects a satisfactory level of prediction accuracy, and XGB consistently achieving values close to 1.0, indicating strong
albeit with somewhat increased variability in the predictions. The ANN model performance. The RF model also performs well, while the ANN
model exhibited the highest A20 index of 13.48, indicating it was the model shows slightly more variability in R2-values. For the RMSE metric
least reliable among the four models assessed. This indicates that ANN (Fig. 15b), the SM-LR model exhibits the lowest RMSE values, high­
exhibited the highest proportion of predictions exceeding the 20 % error lighting its accuracy in minimizing significant errors. The XGB and RF
margin, demonstrating reduced consistency and precision relative to the models follow, with the RF showing slightly higher RMSE in some folds.
other models. The A20 index results indicate that XGB and SM-LR yield The ANN model again shows slightly higher RMSE than the others,
the most accurate predictions with minimal error, whereas ANN and RF indicating it might be less robust to large errors.
exhibit greater error rates, with ANN being the least dependable. The SM-LR model consistently has lower MAE and RMSE values
According to the aforementioned quantitative evaluation, the SM-LR across most folds (Fig. 15c). The XGB, ANN, and RF models follow
model demonstrates the best performance during the TS dataset, fol­ closely, with the ANN sometimes outperforming others in specific folds.
lowed closely by the XGB model, which also exhibits high predictive Lastly, MBE values closer to zero suggest unbiased predictions for MBE
accuracy. The RF model performs well but with slightly lower accuracy values. Accordingly, as shown in Fig. 15d, the XGB and SM-LR models
than the XGB model. The ANN model, while still effective, shows the have MBEs closer to zero across most folds, indicating minimal bias. The
lowest accuracy among the four models. These results highlight the ANN model shows more variability, suggesting it might have some
effectiveness of the stacking method used in the SM-LR model, which systematic bias depending on the fold. The RF model also displays some
enhances prediction accuracy and solid model performance when bias, though it varies significantly across folds. Overall, the SM-LR
applied to unseen data. model demonstrates the best performance across most metrics, with


A.K. Hamed et al. &RPSXWHUV DQG 6WUXFWXUHV   

Fig. 14. Visual evaluations for predicted CS-values between the BMs and SM in the TS stage by (a) Violin boxplot and (b) Taylor diagram.

consistently low RMSE, MAPE, MAE, and MBE values, as well as high R2 variable and from − 45 to 50 for the Slag variable. In contrast, the
values, indicating superior predictive power and minimal bias. The XGB Superplasticizer (SP) exhibits a less pronounced effect, as indicated by
and RF models also show good performance, particularly in R2 and MAE, SHAP values that cluster near zero, with a range from − 25 to 20,
and can be considered reliable alternatives if computational efficiency implying a minimal influence on CS values. The coarse aggregate (Cagg)
or simplicity is a concern. Finally, the ANN model tends to show slightly and fly ash (FA) exhibit neutral to weak effects on the model’s pre­
worse performance in terms of error metrics but still provides a dictions, as indicated by SHAP values clustering near zero, which im­
reasonably high R2, suggesting it can still capture a significant portion of plies a minimal influence on compressive strength within the analyzed
the variance in the data. range of values. Lastly, water (W) exhibits a less pronounced effect on
the model’s predictions, as indicated by SHAP values that are predom­
5.6. Feature importance and interpretability inantly clustered around zero. This suggests that although they can
impact concrete strength, their influence is less significant in compari­
5.6.1. SHAP analysis son to other variables.
The purpose of this sensitivity analysis was to assess the impact of In addition, Fig. 16b indicates that curing age is the primary variable
individual input features on the output variable in the prediction of affecting the CS of concrete, with cement content and fine aggregates
concrete compressive strength. The SHAP feature importance method following as secondary and third factors, respectively. The curing age
was utilized to analyze the impact of each input variable during the directly influences the hardening process of concrete, whereas the
testing stage based on the most effective predictive base ML model (i.e., cement content is essential for bonding the mix. Consequently, these two
XGB model). Fig. 16 presents the results, involving both dot and sum­ factors are pivotal in predicting strength. Slag and fly ash, while sig­
mary bar plots that offer significant insights into feature importance. nificant, provide a more moderate influence, generally contributing to
The interpretations of the SHAP analysis are as follows: the enhancement of long-term strength and durability. Coarse aggre­
gates contribute significantly, albeit to a lesser extent, while the impact
• Inputs are ordered by importance, with the most important inputs at of water and superplasticizers is minimal. The low ranking of water
the top. indicates that other factors, such as cement content, mediate its influ­
• The position of the dots along the x-axis shows the impact of the ence. The limited impact of superplasticizers highlights their function in
input on the concrete’s CS value. The positive values indicate a enhancing workability rather than directly affecting strength. The pri­
positive impact, while the negative values indicate a negative effect. mary factors influencing concrete strength in this model are curing age,
• Inputs with longer bars exert a greater influence on the output of the cement content, and aggregate properties.
model.
• Color coding facilitates the comprehension of the input’s value by 5.6.2. PDP analysis
assigning blue to low values and red to high values. Fig. 17 illustrates the PDPs for eight features that have been carried
out with the XGB model. The plots illustrate the marginal effect of each
Fig. 16a illustrates a comprehensive SHAP dot plot, with each dot feature on the model’s predictions, demonstrating the changes in pre­
denoting a SHAP value corresponding to a particular instance within the dicted output as the value of each feature changes, with all other fea­
dataset. The figure demonstrates that the curing age (Age) exerts the tures kept at unchanged values. The plot indicates a distinct positive
most significant influence on the concrete CS, with values ranging from correlation between the feature value of input cement (C) and the
− 100 to 250. The cement content (C) shows a comparable trend to model’s prediction. The predicted values of CS exhibit a steady increase
curing age, serving as the second critical input, which highlights its with the rise in cement content. The data indicates that the cement
advantageous impact on the enhancement of concrete strength. The variable serves as a significant positive predictor within the XGB model.
SHAP values exhibit a range from − 75 to 100, where elevated values The plot for the blast furnace slag (Slag) displays a relatively flat line,
signify enhanced concrete strength. The inputs fine aggregate (Fagg) indicating that variations in the Slag have minimal to no impact on the
and blast furnace slag (Slag) exhibit a moderate influence, as indicated predicted concrete CS values. In a similar manner, the plot for the input
by SHAP values that are less dispersed compared to those of curing age fly ash (FA) shows a flat line, further indicating that input FA, similar to
and cement content. The SHAP values range from − 60 to 70 for the Fagg Slag, has a minimal impact on the predicted CS values. The analysis of


A.K. Hamed et al. &RPSXWHUV DQG 6WUXFWXUHV   

Fig. 15. Radar plots derived from BO + 10 folds CV across performance metrics: (a) R2 , (b) RMSE, (c) MAE, (d) MAPE, and (e) MBE.

the water content (W) input indicates a negative correlation with the critical. The plot of fine aggregate (Fagg) displays fluctuations without a
predicted values. As the water amount rises, the predicted values of CS noticeable trend, indicating a weak or inconsistent correlation with the
decrease, which may indicate a risk factor or an adverse variable within predicted CS values of concrete. The data indicates that the impact of
the model. The plot of the superplasticizer (SP) shows minor fluctuations fine aggregate on the model’s predictions is neither strong nor consis­
while generally trending upwards, suggesting a weak positive correla­ tent, making it a less reliable predictor. The plot of the curing duration
tion with the predicted CS values. The influence of the SP on the pre­ (Age) presents a pronounced increase initially, followed by a stabiliza­
dictions is positive, although not as significant as that of the cement tion phase. This demonstrates that, at the beginning, extending the
content in the concrete mixture. curing duration resulted in notable enhancements in the concrete CS.
The plot for the coarse aggregate (Cagg) demonstrates a relatively However, beyond a specific threshold, additional increases yield only
flat trend with minimal fluctuations, suggesting that Cagg exerts negli­ marginal improvements in predicted CS values.
gible influence on the CS values. In a manner comparable to the Slag and
FA variables, this indicates that the coarse aggregate does not play a
significant role in the model’s predictions and may be regarded as less


A.K. Hamed et al. &RPSXWHUV DQG 6WUXFWXUHV   

Fig. 16. SHAP charts: (a) summary dot plot and (b) feature importance bar plot based on the XGB model.

5.7. Comparison with previous studies model in this study outperformed similar models from other studies
with comparable dataset sizes, particularly in terms of the R2-value.
The developed SM-LR was evaluated against recent studies using ML
and deep learning (DL) approaches to predict concrete CS. As shown in
Fig. 18, a comparative analysis with previous research on CS prediction 5.8. Developed an interactive GUI
is presented. Liu [106] achieved an exceptionally high accuracy with an
R2-value of 0.999 using the XGB model. However, it should be noted that This section highlights significant advancements that effectively
this result was obtained with a relatively small dataset of 60 samples, address the practical requirements of engineers and designers for the
which may not generalize well to larger datasets. Liang et al. [107] and efficient utilization of machine learning models. The complexities of
Hongwei et al. [19] conducted similar studies using smaller datasets of database compilation, model training, and testing have historically
190 and 98 samples, respectively. Their best models achieved high R2- impeded the incorporation of machine learning into everyday design
values (0.947 and 0.950, respectively) but had higher RMSE and MAE tasks. To overcome these challenges, an innovative approach was
values. Satish et al. [108] developed an XGB model with better perfor­ developed: a Python-based web application featuring a model with
mance on a moderately large dataset of 633 samples, achieving an R2- optimized hyperparameters accessible via a user-friendly GUI. This
value of 0.950, RMSE of 3.06 MPa, and MAE of 2.13 MPa. GUI was developed specifically for predicting the CS of concrete, as
In this research, the SM-LR model achieved an R2-value of 0.953, an illustrated in Fig. 19. The GUI provides a streamlined layout where users
RMSE of 3.315 MPa, and an MAE of 2.493 MPa, demonstrating excep­ can input values for various variables, with the predicted CS dynami­
tional performance on a large dataset of 1030 samples. Additionally, cally displayed based on these inputs. To encourage broad access and
Feng et al. [109] developed an Adaptive Boosting (AdaBoost) model on support collaborative enhancements, the GUI has been hosted on
the same dataset, achieving an R2 of 0.940, RMSE of 1.93 MPa, and MAE GitHub, making it available for use and further development by the
of 1.43 MPa. Elhishi et al. [110] reported an R2 of 0.910, RMSE of 4.37 community. This initiative democratizes the use of advanced predictive
MPa, and MAE of 3.04 MPa, indicating a lower correlation and higher models and invites contributions to refine and adapt the tool to meet
error metrics compared to the SM-LR model. In summary, the SM-LR various specific needs within civil engineering. The GUI can be accessed
at the following URL: [Link]


A.K. Hamed et al. &RPSXWHUV DQG 6WUXFWXUHV   

Fig. 17. PDP analysis results for the studied variables.

To assess the generalizability and reliability of the developed GUI majority of data points for the predicted CS-values are closely aligned
tool for practical applications, it will validate its performance on an with the equality line and the deviation lines of ± 10 %, thereby
independent dataset, different from the one utilized in the current study underscoring the accuracy and precision of the GUI tool. The proximity
for concrete CS prediction. The efficacy of the developed GUI tool will be to the equality line signifies that the tool’s predictions closely match the
verified by utilizing it on external experimental samples and comparing actual results, and the minor deviations within the ± 10 % range un­
the predicted CS values with the actual experimental outcomes. To this derscore its reliability and consistency. In conclusion, the performance
end, it employed the dataset from the research conducted by Song et al. of the GUI tool on the independent dataset confirms its superior pre­
[60], used as a reliable external benchmark to evaluate the predictive dictive capabilities. This affirms the tool’s dependability and applica­
accuracy and robustness of the developed GUI tool. bility for practical use, offering it a vital tool for predicting concrete
Fig. 20 presents a comparative plot of the predicted CS values compressive strength across various scenarios.
(derived from the GUI tool) and the actual CS values (obtained from
Song et al.). The developed GUI tool exhibits excellent predictive ac­ 6. Conclusions
curacy on the independent dataset. The findings indicate an R2 value of
0.977 and an RMSE of 2.210 MPa, signifying a strong correlation be­ In this research, an extensive analysis of 1030 datasets on CS was
tween the predicted and actual compressive strength values, along with conducted, utilizing information from previous journal papers. The
minimal prediction errors. This indicates that the tool offers a highly main aim of this study was to compare the performance of a Stacked
dependable predictive model with exceptional fit. Moreover, the Model based on Linear Regression (SM-LR) with base models (BMs):


A.K. Hamed et al. &RPSXWHUV DQG 6WUXFWXUHV   

Fig. 18. Comparative analysis for prediction of the concrete CS.

Fig. 19. GUI screenshot for calculating the CS-values of concrete.

ANN, RF, and XGB, including eight input parameters. The dataset is 3. SHAP and PDP analyses indicated that the most critical factor was
divided into three separate sets: training, validation, and testing. Both the age of the concrete, which consistently improved predictions as it
training and validation datasets were used to develop and test the BMs. increased. The second most important input was the cement content,
While the SM-LR was built using the developed BMs predictions in the which had a threshold effect that significantly influenced predictions
validation stage, then the last testing dataset was used to evaluate its up to a certain level. Conversely, factors such as blast-furnace slag
performance. Comparative analysis was conducted to assess the pre­ and coarse aggregates exhibited moderate effects, whereas fly ash,
dictive accuracy of both types of models. The study yielded the following superplasticizer, and fine aggregates had the least influence.
main findings: 4. Rank analysis showed that among the BMs, the XGB stood out, fol­
lowed by the RF model, which demonstrated good accuracy but had
1. Among the base models, the XGB model demonstrated superior slightly higher errors. The ANN models ranked lowest, with lower
predictive accuracy during the testing stage, with an R2 value of correlation and higher errors.
0.947 and an RMSE of 3.524 MPa. 5. A user-friendly GUI was developed to make the predictive models
2. The SM-LR model outperformed all base models, achieving the accessible for real-world applications. The GUI, available on GitHub,
highest R2 of 0.953 and the lowest RMSE of 3.315 MPa, which un­ allows engineers to input concrete mix parameters and receive
derscores its superior predictive capabilities.


A.K. Hamed et al. &RPSXWHUV DQG 6WUXFWXUHV   

Fig. 20. Validation of developed GUI tool against actual experimental results from Song et al. [60].

predicted CS values in real-time, enhancing the usability of these Funding

models in practical design scenarios.
This research did not receive funding from specific agencies in the
7. Limitations and future work public, commercial, or not-for-profit divisions.

While the developed model demonstrates strong predictive perfor­

mance, several limitations should be noted. The model’s effectiveness Declaration of competing interest
heavily relies on the quality and representativeness of the input data.
Since the dataset used in this study does not include key factors such as The authors declare that they have no known competing financial
temperature, humidity, and curing methods, its ability to generalize to interests or personal relationships that could have appeared to influence
other environments or experimental setups may be limited. Further­ the work reported in this paper.
more, the model does not account for all variables that may influence
concrete CS, such as specific mix proportions and external environ­ Acknowledgments
mental conditions, potentially affecting its accuracy in certain real-
world scenarios. Additionally, while the GUI is designed for simplicity, The authors would like to thank the journal editor and the anony­
users must still have an understanding of the input parameters and their mous reviewers for their editing and comments.
relevance to make the most effective use of the tool.
Future work should focus on expanding the dataset to incorporate a Data availability
broader range of concrete compositions, environmental conditions, and
laboratory variables, including those that might affect the curing pro­ Data will be made available on request.
cess. Moreover, integrating advanced deep learning models could
enhance predictive accuracy by identifying more complex relationships
References
in the data. The GUI could also be improved by adding interactive fea­
tures, guidance, and tutorials to help users with varying levels of [1] Hassan HS, Abdel-Gawwad HA, Vásquez-García SR, Israde-Alcántara I, Flores-
expertise navigate the input process. Finally, validating the model using Ramirez N, Rico JL, et al. Cleaner production of one-part white geopolymer
real-world data from diverse applications and environmental conditions cement using pre-treated wood biomass ash and diatomite. J Clean Prod 2019;
209:1420–8. [Link]
will be crucial to ensure its robustness, generalizability, and practical [2] Liu G, Yang H, Fu Y, Mao C, Xu P, Hong J, et al. Cyber-physical system-based real-
applicability. time monitoring and visualization of greenhouse gas emissions of prefabricated
construction. J Clean Prod 2020;246:119059. [Link]
jclepro.2019.119059.
CRediT authorship contribution statement [3] Elshaarawy MK, Eltarabily MG. Machine Learning Models for Predicting Water
Quality Index: Optimization and Performance Analysis for El Moghra, Egypt.
Abdelrahman Kamal Hamed: Funding acquisition, Formal anal­ Water Supply 2024. [Link]
[4] Benhelal E, Shamsaei E, Rashid MI. Novel modifications in a conventional clinker
ysis, Data curation, Conceptualization. Mohamed Kamel Elshaarawy: making process for sustainable cement production. J Clean Prod 2019;221:
Software, Resources, Project administration, Methodology, Investiga­ 389–97. [Link]
tion. Mostafa M. Alsaadawi: Writing – review & editing, Writing – [5] Samimi K, Kamali-Bernard S, Akbar Maghsoudi A, Maghsoudi M, Siad H.
Influence of pumice and zeolite on compressive strength, transport properties and
original draft, Visualization, Validation.
resistance to chloride penetration of high strength self-compacting concretes.
Constr Build Mater 2017;151:292–311. [Link]
conbuildmat.2017.06.071.
[6] Taji I, Ghorbani S, de Brito J, Tam VWY, Sharifi S, Davoodi A, et al. Application of
statistical analysis to evaluate the corrosion resistance of steel rebars embedded


A.K. Hamed et al. &RPSXWHUV DQG 6WUXFWXUHV   

in concrete with marble and granite waste dust. J Clean Prod 2019;210:837–46. [30] Shaban WM, Daef KS. Performance of eco-friendly concrete: a safe direction to
[Link] sustainable cities development. Smart Constr Sustain Cities 2023;1:13. https://
[7] Barcelo L, Kline J, Walenta G, Gartner E. Cement and carbon emissions. Mater [Link]/10.1007/s44268-023-00015-1.
Struct 2014;47:1055–65. [Link] [31] W. Ben Chaabene, M. Flah, M.L. Nehdi, Machine learning prediction of
[8] Elshaarawy MK, Elkiki M, Selim T, Eltarabily MG. Hydraulic comparison of mechanical properties of concrete: Critical review, Constr. Build. Mater. 260
different types of lining for irrigation canals using computational fluid dynamic (2020) 119889. doi: 10.1016/[Link].2020.119889.
models, [Link]. Thesis. Civil Eng Depart Facul Eng, Port Said Univ 2024. https:// [32] Li Z, Yoon J, Zhang R, Rajabipour F, Srubar III WV, Dabo I, et al. Machine
[Link]/10.13140/RG.2.2.21927.97441. learning in concrete science: applications, challenges, and best practices. Npj
[9] Czarnecki S, Shariq M, Nikoo M, Sadowski Ł. An intelligent model for the Comput Mater 2022;8:127. [Link]
prediction of the compressive strength of cementitious composites with ground [33] Ziolkowski P, Niedostatkiewicz M. Machine learning techniques in concrete mix
granulated blast furnace slag based on ultrasonic pulse velocity measurements. design. Materials (Basel) 2019;12:1256. [Link]
Measurement 2021;172:108951. [Link] [34] Isleem HF, Qiong T, Alsaadawi MM, Elshaarawy MK, Mansour DM, Abdullah F,
measurement.2020.108951. et al. Numerical and machine learning modeling of GFRP confined concrete-steel
[10] A. Chajec, Granite Powder vs. Fly Ash for the Sustainable Production of Air-Cured hollow elliptical columns. Sci Rep 2024;14:18647. [Link]
Cementitious Mortars, Materials (Basel). 14 (2021) 1208. doi: 10.3390/ s41598-024-68360-4.
ma14051208. [35] Ali L, Isleem HF, Bahrami A, Jha I, Zou G, Kumar R, et al. Integrated behavioural
[11] Shubbar AA, Jafer H, Dulaimi A, Hashim K, Atherton W, Sadique M. The analysis of FRP-confined circular columns using FEM and machine learning.
development of a low carbon binder produced from the ternary blending of Compos Part C Open Access 2024;13:100444. [Link]
cement, ground granulated blast furnace slag and high calcium fly ash: An jcomc.2024.100444.
experimental and statistical approach. Constr Build Mater 2018;187:1051–60. [36] P.t. m., h.m.n. s.,. Predicting Stress and Strain of FRP-Confined Square/
[Link] Rectangular Columns Using Artificial Neural Networks. J Compos Constr 2014;
[12] Phul AA, Memon MJ, Shah SNR, Sandhu AR. GGBS And Fly Ash Effects on 18:4014019. [Link]
Compressive Strength by Partial Replacement of Cement Concrete. Civ Eng J [37] Naser MZ, Kodur VK. Explainable machine learning using real, synthetic and
2019;5:913–21. [Link] augmented fire tests to predict fire resistance and spalling of RC columns. Eng
[13] M.G. Eltarabily, M.K. Elshaarawy, Risk Assessment of Potential Groundwater Struct 2022;253:113824. [Link]
Contamination by Agricultural Drainage Water in the Central Valley Watershed, [38] Naser MZ, Kodur V, Thai H-T, Hawileh R, Abdalla J, Degtyarev VV. StructuresNet
California, USA, in: Handb. Environ. Chem., Springer Berlin Heidelberg, Berlin, and FireNet: Benchmarking databases and machine learning algorithms in
Heidelberg, 2023: pp. 1–40. doi: 10.1007/698_2023_1051. structural and fire engineering domains. J Build Eng 2021;44:102977. https://
[14] Elrefaei AEMM, Alsaadawi MM, Wagdy W. Characteristics of high-strength [Link]/10.1016/[Link].2021.102977.
concrete reinforced with steel fibers recovered from waste tires. Key Eng Mater [39] Niu Y, Wang W, Su Y, Jia F, Long X. Plastic damage prediction of concrete under
2023;945:145–56. [Link] compression based on deep learning. Acta Mechanica 2024;235:255–66. https://
[15] Wangler T, Roussel N, Bos FP, Salet TAM, Flatt RJ. Digital concrete: a review. [Link]/10.1007/s00707-023-03743-8.
Cem Concr Res 2019;123:105780. [Link] [40] Jeyasehar CA, Sumangala K. Damage assessment of prestressed concrete beams
cemconres.2019.105780. using artificial neural network (ANN) approach. Comput Struct 2006;84:
[16] Alsaadawi MM, Amin M, Tahwia AM. Thermal, mechanical and microstructural 1709–18. [Link]
properties of sustainable concrete incorporating Phase change materials. Constr [41] Waszczyszyn Z, Bartczak M. Neural prediction of buckling loads of cylindrical
Build Mater 2022;356:129300. [Link] shells with geometrical imperfections. Int J Non Linear Mech 2002;37:763–75.
conbuildmat.2022.129300. [Link]
[17] A.E. Elrefaei, M. Alsaadawi, M.M. Elshafiey, M. Abdolwahab, A.F. Oan, [42] Degtyarev VV, Tsavdaridis KD. Buckling and ultimate load prediction models for
Performance Evaluation of Ultra High Performance Concrete Manufactured with perforated steel beams using machine learning algorithms. J Build Eng 2022;51:
Recycled Steel Fiber, in: 2024: pp. 3–13. doi: 10.4028/p-dWhX1H. 104316. [Link]
[18] Khaloo AR, Dehestani M, Rahmatabadi P. Mechanical properties of concrete [43] Wang C, Chan T-M. Machine learning (ML) based models for predicting the
containing a high volume of tire–rubber particles. Waste Manag 2008;28: ultimate strength of rectangular concrete-filled steel tube (CFST) columns under
2472–82. [Link] eccentric loading. Eng Struct 2023;276:115392. [Link]
[19] Song H, Ahmad A, Farooq F, Ostrowski KA, Maślak M, Czarnecki S, et al. engstruct.2022.115392.
Predicting the compressive strength of concrete with fly ash admixture using [44] Zhou X-G, Hou C, Feng W-Q. Optimized data-driven machine learning models for
machine learning algorithms. Constr Build Mater 2021;308:125021. [Link] axial strength prediction of rectangular CFST columns. Structures 2023;47:
org/10.1016/[Link].2021.125021. 760–80. [Link]
[20] Cotsovos DM, Pavlović MN. Numerical investigation of concrete subjected to [45] Lyu F, Fan X, Ding F, Chen Z. Prediction of the axial compressive strength of
compressive impact loading. Part 2: Parametric investigation of factors affecting circular concrete-filled steel tube columns using sine cosine algorithm-support
behaviour at high loading rates. Comput Struct 2008;86:164–80. [Link] vector regression. Compos Struct 2021;273:114282. [Link]
10.1016/[Link].2007.05.015. compstruct.2021.114282.
[21] Li M, Hao H, Shi Y, Hao Y. Specimen shape and size effects on the concrete [46] Zarringol M, Patel VI, Liang QQ. Artificial neural network model for strength
compressive strength under static and dynamic tests. Constr Build Mater 2018; predictions of CFST columns strengthened with CFRP. Eng Struct 2023;281:
161:84–93. [Link] 115784. [Link]
[22] Bhanja S, Sengupta B. Investigations on the compressive strength of silica fume [47] Ma L, Zhou C, Lee D, Zhang J. Prediction of axial compressive capacity of CFRP-
concrete using statistical methods. Cem Concr Res 2002;32:1391–4. [Link] confined concrete-filled steel tubular short columns based on XGBoost algorithm.
org/10.1016/S0008-8846(02)00787-1. Eng Struct 2022;260:114239. [Link]
[23] Bharatkumar B, Narayanan R, Raghuprasad B, Ramachandramurthy D. Mix [48] Xu C, Zhang Y, Isleem HF, Qiu D, Zhang Y, Alsaadawi MM, et al. Numerical and
proportioning of high performance concrete. Cem Concr Compos 2001;23:71–80. machine learning models for concentrically and eccentrically loaded CFST
[Link] columns confined with FRP wraps. Struct Concr 2024. [Link]
[24] Eltarabily MG, Elshaarawy MK, Elkiki M, Selim T. Computational fluid dynamics suco.202400541.
and artificial neural networks for modelling lined irrigation canals with low- [49] Alkayem NF, Shen L, Mayya A, Asteris PG, Fu R, Di Luzio G, et al. Prediction of
density polyethylene and cement concrete liners. Irrig Drain 2024;73:910–27. concrete and FRC properties at high temperature using machine and deep
[Link] learning: A review of recent advances and future perspectives. J Build Eng 2024;
[25] Li D, Tang Z, Kang Q, Zhang X, Li Y. Machine Learning-Based Method for 83:108369. [Link]
Predicting Compressive Strength of Concrete. Processes 2023;11:390. [Link] [50] Nguyen N-H, Vo TP, Lee S, Asteris PG. Heuristic algorithm-based semi-empirical
org/10.3390/pr11020390. formulas for estimating the compressive strength of the normal and high
[26] Eltarabily MG, Elshaarawy MK, Elkiki M, Selim T. Modeling surface water and performance concrete. Constr Build Mater 2021;304:124467. [Link]
groundwater interactions for seepage losses estimation from unlined and lined 10.1016/[Link].2021.124467.
canals. Water Sci 2023;37:315–28. [Link] [51] Kardani N, Bardhan A, Samui P, Nazem M, Asteris PG, Zhou A. Predicting the
23570008.2023.2248734. thermal conductivity of soils using integrated approach of ANN and PSO with
[27] Zhou G, Moayedi H, Bahiraei M, Lyu Z. Employing artificial bee colony and adaptive and time-varying acceleration coefficients. Int J Therm Sci 2022;173:
particle swarm techniques for optimizing a neural network in prediction of 107427. [Link]
heating and cooling loads of residential buildings. J Clean Prod 2020;254: [52] A.D.S. and P.G.A. Danial J. Armaghani, Anna Mamou, Chrysanthos Maraveas,
120082. [Link] Panayiotis C. Roussis, Vassilis G. Siorikis, Predicting the unconfined compressive
[28] Das P, Kashem A, Hasan I, Islam M. A comparative study of machine learning strength of granite using only two non-destructive test indexes, Geomech. Eng. 25
models for construction costs prediction with natural gradient boosting algorithm (2021). doi: 10.12989/gae.2021.25.4.317.
and SHAP analysis, Asian. J Civ Eng 2024;25:3301–16. [Link] [53] Chithra S, Kumar SRRS, Chinnaraju K, Alfin Ashmita F. A comparative study on
s42107-023-00980-z. the compressive strength prediction models for High Performance Concrete
[29] Shaban WM, Elbaz K, Yang J, Thomas BS, Shen X, Li L, et al. Effect of pozzolan containing nano silica and copper slag using regression analysis and Artificial
slurries on recycled aggregate concrete: Mechanical and durability performance. Neural Networks. Constr Build Mater 2016;114:528–35. [Link]
Constr Build Mater 2021;276:121940. [Link] 10.1016/[Link].2016.03.214.
conbuildmat.2020.121940. [54] Nguyen H, Vu T, Vo TP, Thai H-T. Efficient machine learning models for
prediction of concrete strengths. Constr Build Mater 2021;266:120950. https://
[Link]/10.1016/[Link].2020.120950.


A.K. Hamed et al. &RPSXWHUV DQG 6WUXFWXUHV   

[55] Kumar A, Arora HC, Kapoor NR, Mohammed MA, Kumar K, Majumdar A, et al. ICA-XGBoost model. Eng Comput 2021;37:3329–46. [Link]
Compressive Strength Prediction of Lightweight Concrete: Machine Learning s00366-020-01003-0.
Models. Sustainability 2022;14:2404. [Link] [78] Yeh I-C. Concrete Compressive Strength. UCI Machine Learning Repository 2007.
[56] A. Ashrafian, M.J. Taheri Amiri, M. Rezaie-Balf, T. Ozbakkaloglu, O. Lotfi-Omran, [Link]
Prediction of compressive strength and ultrasonic pulse velocity of fiber [79] Selim T, Elshaarawy MK, Elkiki M, Eltarabily MG. Estimating seepage losses from
reinforced concrete incorporating nano silica using heuristic regression methods, lined irrigation canals using nonlinear regression and artificial neural network
Constr. Build. Mater. 190 (2018) 479–494. doi: 10.1016/j. models. Appl Water Sci 2024;14:90. [Link]
conbuildmat.2018.09.047. 02142-1.
[57] Zhang J, Ma G, Huang Y, Sun J, Aslani F, Nener B. Modelling uniaxial [80] Williams B, Halloin C, Löbel W, Finklea F, Lipke E, Zweigerdt R, et al. Data-driven
compressive strength of lightweight self-compacting concrete using random forest model development for cardiomyocyte production experimental failure
regression. Constr Build Mater 2019;210:713–9. [Link] prediction, in. Comput Aided Chem Eng, Elsevier 2020:1639–44.
conbuildmat.2019.03.189. [81] McCulloch WS, Pitts W. A logical calculus of the ideas immanent in nervous
[58] Ali R, Muayad M, Mohammed AS, Asteris PG. Analysis and prediction of the effect activity. Bull Math Biophys 1943;5:115–33. [Link]
of Nanosilica on the compressive strength of concrete with different mix BF02478259.
proportions and specimen sizes using various numerical approaches. Struct Concr [82] Elshaarawy MK, Hamed AK. Stacked ensemble model for optimized prediction of
2023;24:4161–84. [Link] triangular side orifice discharge coefficient. Eng Optim 2024:1–31. [Link]
[59] Aslam F, Farooq F, Amin MN, Khan K, Waheed A, Akbar A, et al. Applications of org/10.1080/0305215X.2024.2397431.
Gene Expression Programming for Estimating Compressive Strength of High- [83] T.K. Ho, Random decision forests, in: Proc. 3rd Int. Conf. Doc. Anal. Recognit.,
Strength Concrete. Adv Civ Eng 2020;2020:1–23. [Link] IEEE, 1995: pp. 278–282.
8850535. [84] Breiman L. Random forests. Mach Learn 2001;45:5–32.
[60] Song Y, Zhao J, Ostrowski KA, Javed MF, Ahmad A, Khan MI, et al. Prediction of [85] Eltarabily MG, Abd-Elhamid HF, Zeleňáková M, Elshaarawy MK, Elkiki M,
Compressive Strength of Fly-Ash-Based Concrete Using Ensemble and Non- Selim T. Predicting seepage losses from lined irrigation canals using machine
Ensemble Supervised Machine-Learning Approaches. Appl Sci 2021;12:361. learning models. Front Water 2023;5. [Link]
[Link] frwa.2023.1287357.
[61] Huang J, Sabri MMS, Ulrikh DV, Ahmad M, Alsaffar KAM. Predicting the [86] Tian W, Isleem HF, Hamed AK, Elshaarawy MK. Enhancing discharge prediction
Compressive Strength of the Cement-Fly Ash–Slag Ternary Concrete Using the over Type-A piano key weirs: An innovative machine learning approach. Flow
Firefly Algorithm (FA) and Random Forest (RF) Hybrid Machine-Learning Meas Instrum 2024;100:102732. [Link]
Method. Materials (Basel) 2022;15:4193. [Link] flowmeasinst.2024.102732.
[62] M. Wang, J. Kang, W. Liu, J. Su, M. Li, Research on prediction of compressive [87] Elshaarawy MK, Hamed AK. Machine Learning and Interactive GUI for Estimating
strength of fly ash and slag mixed concrete based on machine learning, PLoS One. Roller Length of Hydraulic Jumps. Appl: Neural Comput; 2024.
17 (2022) e0279293. doi: 10.1371/[Link].0279293. [88] Isleem HF, Elshaarawy MK, Hamed AK. Analysis of Flow Dynamics and Energy
[63] Ahmad A, Farooq F, Niewiadomski P, Ostrowski K, Akbar A, Aslam F, et al. Dissipation in Piano Key and Labyrinth Weirs Using Computational Fluid
Prediction of Compressive Strength of Fly Ash Based Concrete Using Individual Dynamics, in. Comput Fluid Dyn - Anal Simulations, Appl [Working Title],
and Ensemble Algorithm. Materials (Basel) 2021;14:794. [Link] IntechOpen 2024. [Link]
10.3390/ma14040794. [89] Eltarabily MG, Selim T, Elshaarawy MK, Mourad MH. Numerical and
[64] J. Zhang, Y. Zhao, Prediction of Compressive Strength of Ultra-High Performance Experimental Modeling of Geotextile Soil Reinforcement for Optimizing
Concrete (UHPC) Containing Supplementary Cementitious Materials, in: 2017 Settlement and Stability of Loaded Slopes of Irrigation Canals. Environ Earth Sci
Int. Conf. Smart Grid Electr. Autom., IEEE, 2017: pp. 522–525. doi: 10.1109/ 2024.
ICSGEA.2017.150. [90] Breiman L. Stacked regressions. Mach Learn 1996;24:49–64. [Link]
[65] Yu Y, Li W, Li J, Nguyen TN. A novel optimised self-learning method for 10.1007/bf00117832.
compressive strength prediction of high performance concrete. Constr Build [91] Wolpert DH. Stacked generalization. Neural Netw 1992;5:241–59. [Link]
Mater 2018;184:229–47. [Link] org/10.1016/S0893-6080(05)80023-1.
[66] Al-Mughanam T, Aldhyani THH, Alsubari B, Al-Yaari M. Modeling of Compressive [92] Selim T, Hamed AK, Elkiki M, Eltarabily MG. Numerical investigation of flow
Strength of Sustainable Self-Compacting Concrete Incorporating Treated Palm Oil characteristics and energy dissipation over piano key and trapezoidal labyrinth
Fuel Ash Using Artificial Neural Network. Sustainability 2020;12:9322. https:// weirs under free-flow conditions. Model Earth Syst Environ 2023. [Link]
[Link]/10.3390/su12229322. org/10.1007/s40808-023-01844-w.
[67] Bui D-K, Nguyen T, Chou J-S, Nguyen-Xuan H, Ngo TD. A modified firefly [93] Elshaarawy MK, Hamed AK, Hamed S. Regression-Based Models for Predicting
algorithm-artificial neural network expert system for predicting compressive and Discharge Coefficient of Triangular Side Orifice. J Eng Res 2023;7:224–31. https
tensile strength of high-performance concrete. Constr Build Mater 2018;180: ://[Link]/erjeng/vol7/iss5/31.
320–33. [Link] [94] Elshaarawy MK, Hamed AK. Predicting discharge coefficient of triangular side
[68] Al-Hashem MN, Amin MN, Raheel M, Khan K, Alkadhim HA, Imran M, et al. orifice using ANN and GEP models. Water Sci 2024;38:1–20. [Link]
Predicting the Compressive Strength of Concrete Containing Fly Ash and Rice 10.1080/23570008.2023.2290301.
Husk Ash Using ANN and GEP Models. Materials (Basel) 2022;15:7713. https:// [95] Eltarabily MG, Hamed AK, Elkiki M, Selim T. Hydraulic assessment of different
[Link]/10.3390/ma15217713. types of piano key weirs. ISH J Hydraul Eng 2024:1–24. [Link]
[69] Golafshani EM, Behnood A, Arashpour M. Predicting the compressive strength of 09715010.2024.2415938.
normal and High-Performance Concretes using ANN and ANFIS hybridized with [96] Hamed AK, Eltarabily MG, Elkiki M, Selim T. Optimum Hydraulic Design of Weirs
Grey Wolf Optimizer. Constr Build Mater 2020;232:117266. [Link] for Downstream Zone Energy Dissipation. Civil Engineering Department, Faculty
10.1016/[Link].2019.117266. of Engineering: Port Said University; 2024. [Link]
[70] P. Das, A. Kashem, Hybrid machine learning approach to prediction of the RG.2.2.18564.54409 [[Link]. Thesis].
compressive and flexural strengths of UHPC and parametric analysis with shapley [97] Asteris PG, Koopialipoor M, Armaghani DJ, Kotsonis EA, Lourenço PB. Prediction
additive explanations, Case Stud. Constr. Mater. 20 (2024) e02723. doi: 10.1016/ of cement-based mortars compressive strength using machine learning
[Link].2023.e02723. techniques. Neural Comput Appl 2021;33:13089–121. [Link]
[71] R. Karim, M.H. Islam, S.D. Datta, A. Kashem, Synergistic effects of supplementary s00521-021-06004-8.
cementitious materials and compressive strength prediction of concrete using [98] Asteris PG, Karoglou M, Skentou AD, Vasconcelos G, He M, Bakolas A, et al.
machine learning algorithms with SHAP and PDP analyses, Case Stud. Constr. Predicting uniaxial compressive strength of rocks using ANN models:
Mater. 20 (2024) e02828. doi: 10.1016/[Link].2023.e02828. Incorporating porosity, compressional wave velocity, and schmidt hammer data.
[72] A. Kashem, R. Karim, S.C. Malo, P. Das, S.D. Datta, M. Alharthai, Hybrid data- Ultrasonics 2024;141:107347. [Link]
driven approaches to predicting the compressive strength of ultra-high- [99] Elazab MA, Kabeel AE, El-Said EMS, Dahab HA, Hamed AK, Alsaadawi MM, et al.
performance concrete using SHAP and PDP analyses, Case Stud. Constr. Mater. 20 Exergoeconomic assessment of a multi-section solar distiller coupled with solar
(2024) e02991. doi: 10.1016/[Link].2024.e02991. air heater: Optimization and economic viability. Desalin Water Treat 2024;319:
[73] Shaban WM, Elbaz K, Yang J, Shen S-L. A multi-objective optimization algorithm 100535. [Link]
for forecasting the compressive strength of RAC with pozzolanic materials. [100] Vakharia V, Gupta VK, Kankar PK. A comparison of feature ranking techniques for
J Clean Prod 2021;327:129355. [Link] fault diagnosis of ball bearing. Soft Comput 2016;20:1601–19. [Link]
[74] Islam N, Kashem A, Das P, Ali MN, Paul S. Prediction of high-performance 10.1007/s00500-015-1608-6.
concrete compressive strength using deep learning techniques, Asian. J Civ Eng [101] Elshaarawy MK, Elmasry NH, Selim T, Elkiki M, Eltarabily MG. Determining
2024;25:327–41. [Link] Seepage Loss Predictions in Lined Canals Through Optimizing Advanced Gradient
[75] Asteris PG, Skentou AD, Bardhan A, Samui P, Pilakoutas K. Predicting concrete Boosting Techniques. Water Conserv Sci Eng 2024. [Link]
compressive strength using hybrid ensembling of surrogate machine learning s41101-024-00306-3.
models. Cem Concr Res 2021;145:106449. [Link] [102] Kabeel AE, Elazab MA, El Hadi Attia M, Elshaarawy MK, Hamed AK,
cemconres.2021.106449. Alsaadawi MM, et al. Exploring the potential of conical solar stills: Design
[76] Ashrafian A, Panahi E, Salehi S, Karoglou M, Asteris PG. Mapping the strength of optimization and enhanced performance overview. Desalin Water Treat 2024;
agro-ecological lightweight concrete containing oil palm by-product using 320:100642. [Link]
artificial intelligence techniques. Structures 2023;48:1209–29. [Link] [103] Elshaarawy MK, Alsaadawi MM, Hamed AK. Machine Learning and Interactive
10.1016/[Link].2022.12.108. GUI for Concrete Compressive Strength Prediction. Sci Rep 2024. [Link]
[77] Duan J, Asteris PG, Nguyen H, Bui X-N, Moayedi H. A novel artificial intelligence 10.1038/s41598-024-66957-3.
technique to predict compressive strength of recycled aggregate concrete using


A.K. Hamed et al. &RPSXWHUV DQG 6WUXFWXUHV   

[104] Lundh F. An introduction to tkinter, URL Www. Pythonware Com/Library/ Eng Softw 2023;185:103532. [Link]
Tkinter/Introduction/Index Htm 1999;539:540. advengsoft.2023.103532.
[105] Benzaamia A, Ghrici M, Rebouh R, Zygouris N, Asteris PG. Predicting the shear [108] Paudel S, Pudasaini A, Shrestha RK, Kharel E. Compressive strength of concrete
strength of rectangular RC beams strengthened with externally-bonded FRP material using machine learning techniques. Clean Eng Technol 2023;15:100661.
composites using constrained monotonic neural networks. Eng Struct 2024;313: [Link]
118192. [Link] [109] Feng D-C, Liu Z-T, Wang X-D, Chen Y, Chang J-Q, Wei D-F, et al. Machine
[106] Liu Y. High-Performance Concrete Strength Prediction Based on Machine learning-based compressive strength prediction for concrete: An adaptive
Learning. Comput Intell Neurosci 2022;2022:1–7. [Link] boosting approach. Constr Build Mater 2020;230:117000. [Link]
2022/5802217. 10.1016/[Link].2019.117000.
[107] Liang W, Yin W, Zhong Y, Tao Q, Li K, Zhu Z, et al. Mixed artificial intelligence [110] Elhishi S, Elashry AM, El-Metwally S. Unboxing machine learning models for
models for compressive strength prediction and analysis of fly ash concrete. Adv concrete strength prediction using XAI. Sci Rep 2023;13:19892. [Link]
10.1038/s41598-023-47169-7.