Asian Journal of Civil Engineering

https://doi.org/10.1007/s42107-023-00891-z

METHODOLOGY

Prediction of compressive strength of glass fiber‑reinforced

self‑compacting concrete interpretable by machine learning
algorithms
Abhilash Gogineni1 · M. K. Diptikanta Rout1 · Kumar Shubham1

Received: 21 August 2023 / Accepted: 31 August 2023

© The Author(s), under exclusive licence to Springer Nature Switzerland AG 2023

Abstract
Self-compacting concrete (SCC) is a versatile construction material known for its ability to consolidate naturally under its
own weight, making it well suited for challenging placements and reinforced sections. Incorporating a minimal amount of
glass fiber into self-compacting concrete augments its benefits during the fresh state and translates to enhanced performance
in the hardened state. In this present study, an experimental investigation has been carried out to examine the effects of
glass fiber of varying proportions 3%, 6%, and 9% with a constant 10% of fly ash as mineral admixture used as partial
replacement of ordinary portland cement (OPC) in SCC. The experimental result reveals that the introduction of 3%
glass fiber has a substantial impact on the compressive strength. To forecast compressive strength, three machine learning
algorithms AdaBoost, XGBoost, and Gradient Boost regressor were employed to interpret the experimental data and predict
the anticipated outcomes of compressive strength. Among the three models, XGBoost shows higher accuracy with R ­ 2 of
0.9918 and 0.9797 in training and testing in comparison to the other two models due to its adeptness in managing complex
interaction relationships. Additionally, a sensitivity analysis was conducted using both Sobol and SHAP methods. The results
revealed that glass fiber and fly ash exerted substantial influence on the compressive strength, as indicated by their greater
impacts identified through both sensitivity analysis techniques. This study significantly contributes to bridging the gaps that
were present in prior research not only enhances our understanding of the factors influencing compressive strength but also
provides a comprehensive approach to sensitivity analysis, thus making a meaningful contribution to the field.

Keywords Self-compacting concrete · Glass fibers · Compressive strength · ML algorithms

Introduction the way concrete is placed and compacted (Gupta & Sid-
dique, 2019). Unlike traditional concrete requiring manual
Concrete stands as the cornerstone of the construction sec- consolidation through vibration, SCC is characterized by
tor, representing a foundational material that has shaped the its ability to flow and fill even complex formwork without
built environment for centuries (Gupta & Siddique, 2020). external assistance. This remarkable property is achieved
Its paramount importance lies in its versatility, durability, by meticulously designing the mixture with a carefully bal-
and adaptability, making it an essential component in nearly anced combination of fine aggregates, cement, water, and
all types of construction projects. As a structural material, high-range water-reducing admixtures. The result is a highly
concrete provides the backbone for buildings, bridges, workable and cohesive mix that not only improves construc-
roads, and infrastructure, ensuring stability and load-bearing tion efficiency by reducing labor-intensive processes but also
capacity (Rout & Jena, 2015). SCC represents a transforma- ensures uniform distribution and optimal compaction within
tive advancement in concrete technology, revolutionizing the formwork, eliminating voids and honeycombing. SCC's
exceptional flowability enhances the quality of concrete ele-
ments, yielding smoother surfaces and improved durability
* M. K. Diptikanta Rout due to reduced likelihood of entrapped air and segregation.
[email protected] Fiber-reinforced self-compacting concrete (FR-SCC) rep-
1
Department of Civil Engineering, National Institute resents an innovative synergy of two groundbreaking tech-
of Technology, Jamshedpur 831014, India nologies, combining the benefits SCC with the enhanced

13
Vol.:(0123456789)
 Asian Journal of Civil Engineering

mechanical properties offered by various types of discrete of AI for modeling concrete properties has emerged as a
fibers (Güneyisi et al., 2019). This specialized concrete var- prominent field of academic research, attracting substantial
iant integrates the remarkable self-leveling and self-com- attention and activity (Paruthi et al., 2022; Rout et al.,
pacting properties of SCC with the tensile strength, crack 2023b; Shahmansouri et al., 2021; Shubham et al., 2023a).
control, and durability improvements that fiber reinforce- A robust and top-tier dataset is of paramount importance,
ment provides (Rout et al., 2021). The addition of fibers, spanning a diverse spectrum of concrete mixtures, curing
such as steel, polypropylene, or glass, strategically dispersed circumstances, and testing variables (Gogineni et al., 2023b;
throughout the mix, contributes to the concrete's ability Kaveh et al., 2023).
to resist cracking, absorb energy, and withstand dynamic This predictive model not only facilitates the optimization
loads (Ahmadi et al., 2017; Sivakumar et al., 2017). This of GFR-SCC compositions for desired strength thresholds
unique combination resulting into glass fiber-reinforced but also provides invaluable insights into the influence of
self-compacting concrete (GFR-SCC) not only addresses different parameters on the concrete's performance. The
the inherent weaknesses of concrete, such as its low tensile interpretability of these algorithms enhances engineers'
strength and susceptibility to cracking, but also provides the understanding of how glass fibers, mix proportions, and
structural integrity and resilience required for diverse appli- other factors contribute to the compressive strength, aiding
cations (Rout et al., 2023a). This dataset must encompass in informed decision-making. Through the amalgamation
a substantial quantity of cases, encompassing an array of of cutting-edge technology and concrete science, this
compositions, aggregates, water–cement ratios, and curing approach not only elevates the precision of compressive
durations, to comprehensively encapsulate the multifaceted strength prediction but also advances the realm of material
factors that contributes to the concrete strength (Biswas optimization and engineering excellence in GFR-SCC
et al., 2021; Kaveh et al., 2008; Varma et al., 2023). By applications.
harnessing the advantages of both self-compacting proper- This study focuses exclusively on addressing the
ties and fiber reinforcement, FR-SCC ushers in a new era of limitations of prior investigations related to the impact of
construction possibilities, offering durability, efficiency, and glass fiber addition on compressive strength through the
versatility in achieving robust and aesthetically appealing implementation of three distinct machine learning models:
concrete structures (Gill & Siddique, 2018). XGBoost, AdaBoost, and Gradient Boost Regressor. This
In recent times, machine learning (ML) models have novel approach pioneers the integration of these advanced
been implemented on the experimental data to forecasting algorithms to explore the intricate relationship between glass
the strength of concrete properties. Most of researchers fiber incorporation and compressive strength, offering a
have been focused on experimental investigation rather fresh perspective that promises enhanced predictive accuracy
than the ML prediction technique (Rout et al., 2023c). and a deeper understanding of the underlying patterns.
Machine learning facilitates real-time adjustments to mix Through this unique combination of specific machine
designs and construction methodologies, minimizing trial learning techniques, the study presents an innovative
and error while reducing material waste (Gogineni et al., and comprehensive solution that enriches the domain's
2023a; Shubham et al., 2023b) With its ability to handle knowledge base and contributes to the advancement of
non-linarites and interactions, machine learning transforms materials science and predictive modeling.
concrete prediction from a static estimation to a dynamic and
adaptive process, contributing to more resilient, efficient, and
sustainable construction practices (Kaveh & Khavaninzadeh,
2023; Rout et al., 2023c). In the realm of predicting intricate Research objectives
concrete properties like strength, deformation, and modulus
of rupture (Chaabene et al., 2020), artificial intelligence (AI) This study utilizes three machine learning algorithms,
emerges as the paramount approach, with statistical and namely, the XGBoost, AdaBoost, and Gradient Boost
machine learning (ML) techniques playing a pivotal role in regressor for estimation of compressive strength of SCC
forecasting compressive strength growth based on concrete's incorporating glass fiber content. In this investigation,
composition (Asteris et al., 2021; Duan et al., 2013). Diverse a dataset with 129 sets of experimental work has been
ML methods, encompassing gradient boost regressor established the models.
(GR), naive Bayes (NB), support vector machines (SVM), The primary contribution of the work is to:
random forest (RF), decision tree (DT), and extremely
randomized trees (ERT), have been extensively explored • Add the design parameters to the GFR-SCC incorporating
to assess concrete strength across various curing ages (Liu, mineral additives in concrete.
2022; Moein et al., 2022; Thai, 2022), often involving both • Understand the relationship between all the variables
linear and non-linear regression equations. The utilization through experimental approach.

13
Asian Journal of Civil Engineering

• Provide models that can accurately predict the CS Aggregates

of sustainable GFR-SCC, which will benefit in
construction sector as well as real-life applications. Two types of aggregates namely natural coarse aggregates
• Identify the most influence parameters on CS of (CA) and fine aggregates (FA) were utilized in this
concrete through sensitivity assessment via global current research. The coarse aggregates having maximum
sensitivity analysis, i.e., Sobol method and local size of 20 mm with a specific gravity of 2.82 was used.
sensitivity analysis, i.e., SHapley Additive exPlanations Similarly, fine aggregates, i.e., sand particles having a
(SHAP) method for individual predictions. specific gravity of 2.60 and fineness modulus of 2.35, were
calculated as per the IS: 383-1970 guidelines (BIS, 1970).
The current research intends to address the gaps in The physical properties of coarse aggregates, such as Los
understanding and to further enhance the application of Angeles abrasion, crushing value, and impact value, were
Glass Fiber-Reinforced in concrete and promoting its measured as 24.75%, 21.25%, and 16.65%, respectively.
adoption for structures requiring a combination of superior
mechanical, predicting performance and self-compacting
behavior. Glass fiber

Glass fiber (GF), a versatile and innovative material, has

significantly transformed industries and technologies
Materials and mix proportions with its remarkable properties. Composed of ultra-fine
strands of glass, this lightweight and durable substance
Cement possesses exceptional strength-to-weight ratio, making it
an ideal choice for reinforcing materials in a wide array of
Commercially local available ordinary portland cement applications. From construction industry, there is a high
(OPC) of 53 grade was used in this investigation with demand of this glass fiber as it resists the tensile strength,
satisfying the requirements of Indian Standard IS: 12269- corrosion resistance, and insulating capabilities in
2013. It was also noted the standard consistency of 30%. concrete. Table 1 represents the specification of collected
The initial setting time of 35 min exceeds the minimum glass fiber, which has enabled the development of stronger,
requirement of 30 min, offering extended workability time. more efficient, and technologically advanced product, used
However, the final setting time of 580 min falls short of in SCC.
the suggested 600-min minimum. The specific gravity
and water absorption were recorded as 30.10 and 1.95%,
respectively. The fineness has evaluated as 280 ­m 2/kg Super‑plasticizer
through the Blaine's air permeability test.
Super-plasticizer (SP) is a high water-reducing chemical
admixture used in the cement mixes to accelerate its
Fly ash workability and flow ability. Commercially available
super-plasticizer namely sulphonated Naphthalene
For this context, fly ash (FLA) is a byproduct of coal formaldehyde (SNF) product having specific weight 1.2
combustion in power plants has used for the current and specific gravity 1.12 was used in concrete mixtures.
research as mineral admixtures. It contains fine particles
that can react with the calcium hydroxide produced
during cement hydration to form additional cementitious
compounds. Therefore, the addition of mineral admixtures
can lead to various benefits, including improved strength, Table 1  Specifications of glass fiber
reduced permeability, increased durability, enhanced Properties Symbol Specifications
workability, and a more sustainable use of industrial
Tensile strength CGF 3000 MPa
byproducts. For this current research, class F type
Modulus of elasticity EGF 75 GPa
of fly ash was used as admixtures mixing with SCC
Melting temperature Tm 1130 °C
incorporating with various fractions of glass fibers.
Specific weight Ws 2.5 g/cc
Specific gravity SG 2.7
Fiber diameter DF 14 micron
Fiber length LF 15 mm

13
 Asian Journal of Civil Engineering

Concrete mix proportions Workflow methodology

A design of mix M30 concrete prepared for the SCC. A For this current research, the methodology developed by
total number of 129 specimens were casted for the com- taking various ML algorithms to estimate the compressive
pressive strength in the concrete mixtures at various curing strength of GFR-SCC. Figure 4 represents the flowchart of
ages. The experimental setup for measuring compressive the recent research into five stages.
strength of concrete is shown in Fig. 1. The water/cement Stage 1: Preparation of Data.
(w/c) ratio is kept constant for all concrete mixes. For con- At the initial stage, the collection of datasets has gath-
crete mixes, workability of 50 mm has maintained for all ered related to various GFR-SCC mixtures from experi-
the mixes as per IRC: 10262-2019 (2019). The details of mental projects. This dataset should encompass a range of
quantities of materials calculated for concrete mixes along mix designs, including different proportions of ingredients
with their mix designations are shown in Table 2. The mixing with SCC. The collected datasets are normalized
J-ring and slump flow test setup and with their specifica- or scaling the input features to ensure that they are on the
tions are shown in Figs. 2, 3, respectively. same scale. This step prevents certain features from domi-
nating the learning process due to their larger magnitudes.

Fig.1  a Experimental setup of measuring compressive strength through cube test and b cylindrical test of SCC

Table 2  Concrete mix Mix notation Cement FLA CA FA GF Water SP

proportions (considered all units
are in kg/m3) C1 540 0 890 765 0 198 0
C2 520 37 855 732 1.5 198 0.5
C3 505 64 837 704 3 198 1
C4 500 112 825 695 4.5 198 1.5
C5 480 156 795 686 6 198 2
C6 465 172 782 655 7.5 198 0
C7 450 202 764 632 9 198 0.5
C8 430 242 752 620 10.5 198 1
C9 380 256 733 600 12 198 1.5
C10 365 280 710 585 13.5 198 2

13
Asian Journal of Civil Engineering

Fig. 2  a Specifications and b J-Ring test setup for SCC in laboratory

Fig. 3  a Specifications and b slump flow test setup for SCC in laboratory

The database consisting 129 samples was then analyzed as prevent overfitting, and the test set evaluates the models'
seven input parameters, which are cement (C), coarse aggre- performance on unseen data.
gates (CA), fine aggregates (FA), water (W), superplasti- Stage 3: Performance of models.
cizers (SP), glass fiber (GF), and fly ash (FLA) as mineral Evaluate their performance metrics, such as mean squared
admixtures and a single output prediction as compressive error (MSE), mean absolute error (MAE), root-mean-
strength (CS). The whole datasets are randomly divided into squared error (RMSE), and coefficient of determination
80% as training and 20% as testing. (R-squared).
Stage 2: Training—Validation-Test models. Stage 4: Sensitivity analysis.
The training set is used to train the machine learning At this stage, sensitivity analysis was used on the relative
models, the validation set helps tune hyperparameters and influence of individual features on the model's output,

13
 Asian Journal of Civil Engineering

Fig. 4  A detailed methodology

flowchart for current study

providing insights into which attributes have the most At the last stages, the results of the performance
significant effects. Sobol and SHAP sensitivity analysis were metrics and model explanations to draw insights about
implemented to access the above seven input attributes on the relationships between input features and compressive
the prediction of the considered three ML models. strength in GFR-SCC. Moreover, the check the interpret
Stage 5: Interpretation of results.

13
Asian Journal of Civil Engineering

changes in different mix parameters that influences in Machine learning algorithms on GFR‑SCC
predicted outcomes.
The application of machine learning algorithms to glass
fiber-reinforced self-compacting concrete (GFR-SCC)
Database description and analysis presents a novel and promising approach for understanding
the intricate relationships between mix parameters and
Database description and analysis refers to the process of the mechanical properties of concrete. Machine learning
providing an overview and understanding of a database's techniques offer the capacity to capture complex non-linear
structure, content, and patterns. In this study, a total number interactions that influence the compressive strength of GFR-
of 129 datasets were collected from the experimental SCC, providing a comprehensive analysis beyond traditional
examination. The description of datasets is shown in Table 3. empirical methods. Three conventional ML algorithms;
AdaBoost regressor, XGBoost regressor, and gradient
Co‑relation matrix Boost regressor, were applied in the context of compressive
strength estimation of GFR-SCC.
A correlation matrix is a table that displays the correlation
coefficients between multiple variables. Correlation coef-
AdaBoost regressor
ficients measure the strength and direction of the linear
relationship between two variables within a range from − 1
The AdaBoost Regressor algorithm is a powerful technique
to + 1, that means − 1 indicates the negative linear rela-
for regression tasks, leveraging an ensemble model com-
tionship and + 1 indicates the positive linear relationship
prising multiple weak regression models. The algorithm's
(Nazeer et al., 2023). The correlation matrix represented
approach involves initializing equal weights for each train-
through a heatmap graphical view that employs color-coded
ing data point, training weak regressors (often shallow deci-
cells to portray the values within a two-dimensional dataset.
sion trees) on the dataset, evaluating their performance, and
This graphical representation offers a concise overview of
updating the weights based on prediction errors. The ensem-
data, facilitating the identification of patterns, trends, and
ble is iteratively built by repeating these steps, assigning
correlations among variables. The color intensity of each
various weights to weak regressors. The final prediction is a
cell reflects the dataset value it represents, with darker or
weighted average of the individual regressor outputs, result-
brighter shades signifying higher or lower values, respec-
ing in accurate predictions for new data (Dadras et al., 2021).
tively. The heatmap's visual cues allow for swift identifica-
This process adapts the weights of weak regressor based on
tion of regions with elevated or diminished values, unveiling
their performance, contributing to the algorithm's overall
the distribution and concentration of data across multiple
predictive capability. The pictorial representation of Ada-
dimensions or categories. The heatmap analysis reveals
boost regressor Methodology flowchart is shown in Fig. 6.
that glass fiber and fly ash exert substantial influence on the
compressive strength. The heatmap employs color gradients,
with darker or brighter shades indicating lower or higher XGBoost regressor
compressive strength levels, respectively, to visually depict
this relationship. Notably distinct color patterns emerge It is a potent machine learning algorithm renowned for
for glass fiber and fly ash, implying a significant impact on its exceptional performance and efficiency across various
the resulting compressive strength values. The correlation applications, including regression, classification, and rank-
between input and output attributes is shown in Fig. 5. ing. Grounded in a gradient boosting framework, XGBoost
enhances predictive models by iteratively incorporating

Table 3  Statistical description Parameters Count Mean Standard deviation Minimum Maximum
of datasets
Cement 129 321.06 83.979 160 460.0
Fly ash 129 146.51 51.57 0 261.0
Coarse aggregates 129 734.44 174.26 382.40 923.0
Fine aggregates 129 810.27 104.69 478.00 1080.0
Glass fiber 129 46.179 71.166 0.00 273.78
Water 129 213.59 42.22 162.0 341.30
Super-plasticizer 129 2.642 3.814 0 11.24
Compressive strength 129 39.10 13.258 10.20 73.60

13
 Asian Journal of Civil Engineering

Fig. 5  Co-relation matrix of the

input and output variables

Fig. 6  Expression tree for Adaboost regressor

13
Asian Journal of Civil Engineering

weak learners (Chen & Guestrin, 2016). The algorithm applications as an effective machine learning technique.
adopts a sequential tree construction approach, with each Belonging to the boosting algorithm family, it stands out for
new tree aimed at rectifying the limitations of its predeces- its prowess in tackling complex regression challenges. The
sors. This iterative process updates the model by calculating approach involves constructing a robust ensemble model that
gradients and employing boosting steps to optimize specific iteratively enhances predictive accuracy through the inte-
loss functions, continuing until predefined stopping criteria gration of weak prediction models, such as decision trees.
are met. Notably, XGBoost has proven highly effective in Each subsequent model focuses on rectifying errors from
predicting compressive strength in diverse contexts. Studies its predecessors, with a loss function minimized through
have demonstrated its efficacy, such as using the XGBoost gradient descent optimization. This iterative methodology
regressor to forecast concrete compressive strength (Khan yields a remarkably precise and reliable regression model.
et al., 2020), or precisely predict the compressive strength of Diverse industries, spanning from banking and healthcare
recycled aggregate concrete in other research. These findings to preventive maintenance, have reaped the benefits of the
underscore the XGBoost regressor's remarkable accuracy Gradient Boosting Regressor's exceptional performance.
in predicting compressive strength, rendering it a valuable Figure 8 represents Methodology flowchart of Gradient
tool for researchers and engineers exploring this domain. regressor algorithm.
The XGBoost Regressor operates on an ensemble boosting Gradient Boosting Regressor update can be represented
approach, amalgamating weak predictive models, often deci- as follows:
sion trees, to craft a more robust overarching model. Figure 7 For a given iteration t, let ht​(x) be the prediction from
shows the pictorial representation of XGboost algorithm. the tth estimator (weak learner), y be the true target value,
XGBoost employs the Second-Order Taylor and ϵ be the loss function. The goal is to update the current
approximation technique for both classification and prediction ht​(x) using the gradient of the loss function ϵ with
regression tasks. This approach involves approximating the respect to the previous predictions
loss function that incorporates output values in the following ( )
manner: ht+1 (x) = ht (x) + 𝜼.∇ht (x) 𝝐 y, ht (x) , (2)
[ ] [ ]
( ) ( ) d ( ) 1 d2 ( ) where
L y, poi + Ov = L y, pi + L y, pi Ov + L y, p i Ov 2 .
dpi 2 dpi 2
• h t​(x) is the prediction from the tth estimator (weak
(1)
learner) for input x.
• y is the true target value for input x.
Gradient boost regressor • ϵ is the loss function that quantifies the difference
between the predicted value and the true target value.
The Gradient Boosting Regressor has gained substan- • ht​(x)​ϵ(y,ht​(x)) is the gradient of the loss function ϵ with
tial attention in both academic research and practical
respect to the current prediction ht​(x).

Fig. 7  Expression tree for

XGboost algorithm

13
 Asian Journal of Civil Engineering

Fig.8  Expression tree for gradi-

ent regressor algorithm

• η is the learning rate, a hyperparameter that controls the ⎛ ∑n � P � ⎞

a 2
step size of each update. ⎜ i=1 fcki − fcki ⎟
(3)
2
R =1−⎜ � �2 ⎟,
⎜ ∑n fcka − fcka ⎟
The above equation represents how the Gradient Boosting ⎝ i=1 i ⎠
Regressor iteratively refines its predictions by minimizing
the gradient of the loss function with respect to the previous where fckPi is predicted compressive strength, and fckai is the
predictions. This leads to a sequence of weak learners, actual compressive strength.
each correcting the errors of the previous ones, ultimately The range of the R2 value spans from 0 to 1, with higher
improving the overall prediction quality. values signifying a superior capacity to explain data
variance through the model. While R2 serves as a measure
Performance indices of models of correlation, it should be complemented with other
performance metrics to comprehensively assess model
In this study, three indices were utilized to evaluate efficiency, as it does not solely indicate model accuracy due
the performance of ML models, namely coefficient of to its reliance on output variable variance. The root-mean-
determination (R2), root-mean-square error (RMSE), and square error (RMSE), a pivotal metric, quantifies the average
mean absolute error (MAE). To effectively assess the discrepancy between predicted and observed values. A lower
performance of machine learning (ML) models and achieve RMSE signifies heightened model accuracy, rendering it
a quantitative assessment, the establishment of evaluation invaluable for evaluating and comparing predictive models.
parameters becomes pivotal in guiding the selection of the Equation (4) embodies the mathematical relation of RMSE
optimal approach. A spectrum of well-established techniques √ ( )
√
exists for evaluating the efficacy of predictive modeling, each √1 ∑ n
( p )
RMSE = √ fcki − Tfckia 2
. (4)
carrying its inherent strengths and limitations. In this study, n i=1
two specific quantitative evaluation metrics, the coefficient
of determination (R2) and mean square error (MSE), have Conversely, the mean square error (MSE) reflects the
been employed to meticulously evaluate the performance of proximity of data points to the regression line. Functioning
the implemented models. The R2 metric functions as a gauge as a measure of risk, it pertains to the expected value of
of the independent variable's proficiency in forecasting the squared error loss. MSE, calculated by averaging the squared
outcomes of the dependent variable. Essentially, it signifies errors from the data, is representative of the dispersion of
the extent to which data variance can be elucidated by the data points around the mean. Larger MSE values indicate
predictive model. The mathematical expression for the extensive dispersion, while smaller values correlate with
R2 value is encapsulated in Eq. (3), providing a robust diminished errors, thereby enhancing the model’s suitability
representation of its quantitative significance within the for prediction. Equation (5) outlines the computation of
context of this study MSE in predicted model

13
Asian Journal of Civil Engineering

1 ∑| p
n
| dynamics of a model's output, unveiling how each feature's
MSE = | fck − fckai |. (5) presence influences predictions both independently and in
n i=1 | i |
conjunction with other features. By considering all possible
In conclusion, the incorporation of the R2 and RMSE permutations of feature contributions, SHAP values yield
metrics in this study contributes to a thorough evaluation a comprehensive understanding of feature importance
of model performance. While R2 emphasizes variance and interactions. This approach not only enhances model
explanation and correlation, RMSE assesses the average interpretability but also aids in identifying critical features,
predictive accuracy, thereby collectively providing a understanding complex relationships, and making informed
comprehensive appraisal of the model’s capabilities. decisions in various domains.

Sensitivity analysis of experimental Results and discussion

database
AdaBoost regressor
Sobol’s sensitivity analysis
Figure 9 represents the scatter plots of observed and pre-
dicted compressive strength using Ada Booster regressor and
Sobol indices serve as quantitative measures of sensitivity,
its performance was evaluated on a 70:30 split of training
revealing the proportion of overall variation in a model's
and testing data, resulting in R-squared values of 0.9664 for
output attributed to specific input variables or groups
the training set and 0.9491 for the testing set. When con-
of inputs Parhi & Patro, 2023). Leveraging the Sobol
trasted with the XGBoost and gradient boost regressor mod-
approach, the complete output variance is dissected into
els, the AdaBoost regressor's performance appeared compar-
distinct components, encompassing the primary influences
atively low. Both the XGBoost and gradient boost regressor
of individual input variables, interplays between inputs,
models outperformed the AdaBoost model in terms of pre-
and residual variance. To compute Sobol indices, an
dictive accuracy, with the AdaBoost model demonstrating
appropriate sampling method like Monte Carlo sampling is
lower predictive capability across the training and testing
employed to generate a set of samples from input variables.
datasets.
Subsequently, the output variance is dissected based on
the model's evaluations for each input value combination.
XGBoost regressor
First-order Sobol indices quantify the isolated contributions
of individual inputs, while total-order indices encapsulate
Figure 10 represents the scatter plot illustrates the relation-
the unique impacts of each input variable alongside its
ship between observed and predicted compressive strength
interactions with others. This holistic perspective provided
using the XGBoost algorithm. The dataset was divided
by total-order indices enhances our comprehension of input
into training (70%) and testing (30%) subsets. The model
variable impact, rendering Sobol indices a powerful tool
achieved high performance, with R-squared values of 0.9918
for sensitivity analysis. Any given integral function f(x) is
for the training set and 0.9797 for the testing set, indicating
thought to be decomposable using this technique into the
that the model explains 99.18% and 97.97% of the respec-
sum of ­2n orthogonal terms
tive variance. These strong R-squared values suggest that
∑
n
∑
n
( ) the XGBoost model effectively captures the underlying pat-
f (x) = f0 + fi1 … is xi1 , … , xis . (6) terns in the data, yielding accurate predictions of compres-
sive strength.
s=1 i1 <⋯<is

Gradient boost regressor

SHAP sensitivity analysis
The gradient regressor model's performance was assessed
SHAP (SHapley Additive exPlanations) sensitivity analysis
on both the training and testing datasets, resulting in
is a robust method that offers valuable insights into the
R-squared values of 0.9711 and 0.9698, respectively.
feature importance and contribution of input variables
These values reflect the model's capability to explain
to a model's predictions (Mai et al., 2023) It operates on
around 97.11% of the variance in the training data and
the foundation of Shapley values from cooperative game
96.98% in the testing data, indicating its effectiveness in
theory, quantifying the individual impact of each feature
capturing underlying relationships and providing accu-
on predictions by attributing credits or contributions.
rate predictions for the target variable. The data were
SHAP values provide a unified framework to elucidate the
split into training and testing subsets using a 70:30 ratio.

13
 Asian Journal of Civil Engineering

AdaBoost Regressor_Training
AdaBoost Regressor_Testing
80 70

Predicted Compressive Strength(Mpa)

70 y = 0.8573x + 5.4214 60 y = 0.8439x + 6.2479
Predicted Compressive Strength(Mpa)

R² = 0.9664 R² = 0.9491
60
50

50
40
40
30
30
20
20
10
10

0
0
0 20 40 60 80
0 20 40 60 80
Actual Compressive Stremgth(Mpa) Actual Compressive Strength(Mpa)

Fig. 9  Scatter plot for training and testing of AdaBoost regressor algorithm

XGboost regressor_Training XGboost Regressor_Testing

80
70
y = 0.9674x + 1.4076
Predicted Compressive Strength(Mpa)

70 y = 0.9286x + 2.0534
Predicted Compressive Strength(Mpa)

R² = 0.9918 60 R² = 0.9797
60
50
50
40
40
30
30
20
20
10
10

0
0 0 20 40 60 80
0 20 40 60 80 Actual Compressive Strength(Mpa)
Actual Compressive Strength(Mpa)

Fig. 10  Scatter plot for training and testing of XGboost Algorithm

When compared to the AdaBoost regressor, the gradient Sobol sensitivity analysis
boost regressor demonstrated superior performance. In
contrast, when compared to the XGBoost model, the gra- Figure 12 illustrates the sensitivity analysis of various
dient boost regressor exhibited comparatively lower per- parameters using the Sobol method. This analysis provides
formance levels. Figure 11 represents the scatter plots of insights into how changes in different parameters impact
observed and predicted compressive strength of concrete the model's output or prediction. The Sobol method helps
using Gradient boost regressor algorithm. to identify which parameters have the most significant influ-
ence on the results, aiding in understanding the relative

13
Asian Journal of Civil Engineering

Gradient Regressor_Training Gradient Regressor_Testing

80 70
y = 0.868x + 5.4213

Predicted Compressive Strength(Mpa)

y = 0.8953x + 4.1686
Predicted Compressive Strength(Mpa)
70 60 R² = 0.9698
R² = 0.9711
60
50
50
40
40
30
30
20
20
10
10

0 0
0 20 40 60 80 0 20 40 60 80
Actual Compressive Strength(Mpa) Actual Compressive Strength(Mpa)

Fig. 11  Scatter plot of gradient regressor algorithm during training and testing periods

Fig. 12  Sensitivity analysis

through Sobol’s index

importance of each parameter in the context of the mod- influence the overall compressive strength of the concrete
el's performance. Utilizing the Sobol method, a sensitiv- mixture, leading to their heightened sensitivity in the pre-
ity analysis was conducted, revealing that glass fiber and diction model.
fly ash exhibit higher sensitivity compared to other param-
eters in predicting compressive strength. This indicates that SHAP value theory
variations in these two parameters have a more pronounced
impact on the model's output. Glass fibers, in particular, In the context of this research, a sensitivity analysis was
could be more sensitive due to their role in enhancing the conducted through the application of SHAP (SHapley Addi-
mechanical properties of concrete. As an additive, glass fib- tive exPlanations). This analytical approach allowed for the
ers reinforce the matrix, effectively reducing the formation examination of the distinct contributions of various input
of cracks and increasing tensile strength. Consequently, parameters to the predictive outcomes of the model, with
small changes in glass fiber content could significantly a specific focus on the target variable, "Strength." Through

13
 Asian Journal of Civil Engineering

SHAP values, each feature was attributed a significance to the testing phase, the models' performance underwent a
score, revealing insightful findings. Notably, Cement shift. XGBoost demonstrated its proficiency with a value
emerged with a substantial positive effect on the predicted of 2 for R2 and VAF, and 1 for the other metrics. Gradient
strength (SHAP value: 2.74), signifying its pivotal role. Regressor displayed balanced performance, with a higher
Similarly, Coarse Aggregate and Fine Aggregate demon- emphasis on R2 and VAF (both with values of 2), and lower
strated favorable contributions, each yielding SHAP val- emphasis on the remaining metrics (all assigned a value of
ues of 2.07 and 1.80, respectively. Conversely, Glass Fiber 1). Meanwhile, AdaBoost continued its strong performance,
(SHAP value: 0.67) and Water (SHAP value: 0.79) exhibited focusing primarily on R2 and VAF (both with values of 3),
comparatively modest impacts. Super-plasticizer took prec- while other metrics received less emphasis (with a value of
edence as the most influential feature (SHAP value: 3.06), 1). Examining the overall model performance, the compara-
significantly shaping predictions. Furthermore, Fly ash and tive radar diagrams revealed the culmination of the mod-
Water showcased intermediate positive impacts, as reflected els' abilities. XGBoost's overall performance showcased a
by SHAP values of 0.79. This rigorous analysis underscores blend of strengths and slight weaknesses, with R2 and VAF
the utility of SHAP in unraveling intricate feature-outcome assigned values of 1 and 2, and the remaining metrics with
dynamics, enriching the understanding of model behav- values of 2 and 3. Gradient Regressor maintained a balanced
ior, and enabling informed decision-making across diverse approach, giving equal importance to most metrics, result-
domains. Figure 13 represents sensitivity of different mate- ing in values of 2 and 3 across the board. AdaBoost's overall
rial used in estimating of compressive strength. performance remained consistent with its strong predictive
capabilities, emphasizing R2 and VAF (both with values of
3), while other metrics retained a value of 1. In essence,
Radar diagram these radar diagrams provided an at-a-glance view of the
models' performance across various metrics and different
XGBoost, Gradient Regressor, and AdaBoost models were phases. They illuminated the models' comparative strengths,
evaluated across three crucial phases: training, testing, and allowing for easy evaluation of their suitability for specific
overall performance. The radar diagrams generated using tasks based on the prioritized metrics. The distinct patterns
Plotly's Scatterpolar provided an insightful representa- and shapes formed by the filled polygons highlighted the
tion of each model's strengths and weaknesses in different unique performance characteristics of each model, aiding
aspects. During the training phase, XGBoost exhibited a decision-making processes in model selection and refine-
consistent performance across all metrics, with each met- ment. Figure 14 represents pictorial representation of Radar
ric assigned a value of 1. Similarly, the Gradient Regressor diagrams during training, testing, and overall score analysis
model displayed uniformity in performance, as indicated by for simple visualization.
the metric values of 2 across the board. On the other hand,
AdaBoost exhibited higher performance levels during train-
ing, with all metrics assigned a value of 3. This suggests
that AdaBoost had superior predictive capabilities during
the training phase compared to the other models. Moving

Fig. 13  Pictorial representation

of SHAP sensitivity parameters

13
Asian Journal of Civil Engineering

Fig. 14  Radar diagram for score analysis of all the models

Taylor diagram showcases an STD of 1.05 and an R value of 0.9834, while

the Gradient Boost Regressor model sustains an STD of 0.97
The Taylor diagram serves as an invaluable visualization and an R value of 0.9742. By providing a comprehensive
tool for assessing the predictive accuracy of machine learn- view of model’s accuracy and correlation, the Taylor dia-
ing models in the context of predicting the compressive gram effectively aids in identifying the models that closely
strength of GFR-SCC (Kumar et al., 2023). This analysis align with observed data. This graphical representation facil-
encompasses three distinct scenarios: training, testing, and itates the interpretation of the intricate interplay between
the overall model performance. The data-driven insights model predictions and real-world outcomes, thereby guiding
obtained from these diagrams are as follows: For the training the selection and optimization of machine learning algo-
phase, the Taylor diagram indicates that the XGBoost model rithms for the accurate prediction of GFR-SCC compressive
yields a standard deviation (STD) of 0.98, coupled with a strength. Figure 15 represents Graphical representation of
correlation coefficient (R) of 0.9958. This reflects a strong Taylor's stability number during training, testing, and overall
agreement between the model predictions and the observed evaluation phases provides insight into the model's stability
data. Similarly, the AdaBoost model exhibits an STD of across different stages.
0.79 and an R value of 0.9854, while the Gradient Regressor
model shows an STD of 1 and an R value of 0.983. Moving
on to the testing phase, the XGBoost model maintains its Conclusions
promising performance with an STD of 0.99 and an R value
of 0.9897. The AdaBoost model continues to exhibit a com- The concluding remarks can be points out from the recent
petitive edge, showcasing an STD of 1.05 and an R value study are as follows:
of 0.984. Meanwhile, the Gradient Regressor model main-
tains an STD of 0.97 and an R value of 0.9742. In terms of 1. It can be concluded that the incorporation of a 3%
the overall model assessment, combining both training and addition of glass fiber yields the highest levels of
testing data, the XGBoost model demonstrates an STD of compressive strength. This outcome underscores
1, coupled with an R value of 0.9897. The AdaBoost model the potential of glass fiber as a valuable additive for

13
 Asian Journal of Civil Engineering

Fig. 15  Taylor diagram during training, testing, and overall evaluation phases

enhancing the mechanical properties of self-compacting 3. In terms of performance indicators, such as R2 values,
concrete. MSE, and RMSE, the XGBoost algorithm outperformed
2. Among the three models, the XGBoost method proved then Adaboost regressor and Gradient Boost Regressor.
highly effective in predicting concrete compressive The high R 2 scores observed for both the training
strength when compared with the AdaBoost and and testing datasets indicated its accurate forecasting
Gradient Regressor machine learning techniques. The ability for future outcomes. Additionally, the XGBoost
XGBoost consistently demonstrated superior precision, algorithm exhibited reduced values of MSE and RMSE,
efficiency, and generalization abilities, setting it apart underscoring its precision and reliability in prediction
from the AdaBoost and Gradient Regressor approaches. tasks.

13
Asian Journal of Civil Engineering

4. Sensitivity analysis was performed using both the conference on knowledge discovery and data mining. ACM.
Sobol method and the SHAP method. Notably, both https://​doi.​org/​10.​1145/​29396​72293​9785
Dadras, A. K. A., Javadi, E. S. M., & Malek, N. G. (2021). Machine
algorithms consistently indicated that glass fiber holds learning regression approaches for predicting the ultimate
greater sensitivity in predicting compressive strength. buckling load of variable-stiffness composite. Acta Mechanica,
This convergence of results from two distinct sensitivity 232(3), 921–931. https://​doi.​org/​10.​1007/​s00707-​020-​02878-2
analysis techniques emphasizes the significant influence Duan, Z. H., Kou, S. C., & Poon, C. S. (2013). Using artificial neu-
ral networks for predicting the elastic modulus of recycled
of glass fiber on the compressive strength prediction aggregate concrete. Construction and Building Materials, 44,
process. 524–532. https://​doi.​org/​10.​1016/j.​conbu​ildmat.​2013.​02.​064
5. Radar diagrams provided an intuitive representation Gill, A. S., & Siddique, R. (2018). Durability properties of self-com-
of each model's strengths and weaknesses in relation pacting concrete incorporating metakaolin and rice husk ash.
Construction and Building Materials, 176, 323–332. https://​doi.​
to various performance metrics. The distinctive shapes org/​10.​1016/j.​conbu​ildmat.​2018.​05.​054
and patterns formed by the metrics' values within the Gogineni, A., Panday, I. K., & Kumar, P. (2023a). Predictive mod-
radar diagrams highlighted the models' individual elling of concrete compressive strength incorporating GGBS
characteristics, enabling easy identification of their and alkali using a machine-learning approach. Asian Journal of
Civil Engineering. https://​doi.​org/​10.​1155/​2022/​58022​17
suitability for specific tasks and aiding in the decision- Gogineni, A., Rout, M. K. D., & Shubham, K. (2023b). Evaluating
making process for model selection and improvement. machine learning algorithms for predicting compressive strength
of concrete with mineral admixture using long short-term memory
(LSTM) technique. Asian Journal of Civil Engineering. https://​
doi.​org/​10.​1007/​s42107-​023-​00885-x
Author contributions MKDKR: writing, original drafting, and Güneyisi, E., Atewi, Y. R., & Hasan, M. F. (2019). Fresh and rheologi-
conceptualization. AG: model development, drafting, and validation. cal properties of glass fiber reinforced self-compacting concrete
KS: editing, methodology, coding, and formal analysis. with nanosilica and fly ash blended. Construction and Building
Materials, 211, 349–362. https://​doi.​org/​10.​1016/j.​conbu​ildmat.​
Funding No funding was received for conducting this study. 2019.​03.​087
Gupta, N., & Siddique, R. (2019). Strength and micro-structural prop-
Data availability Some or all data, models, or codes generated or erties of self-compacting concrete incorporating copper slag. Con-
used during the study are available from the corresponding author by struction and Building Materials, 224, 894–908. https://​doi.​org/​
request. 10.​1016/j.​conbu​ildmat.​2019.​07.​105
Gupta, N., & Siddique, R. (2020). Durability characteristics of self-
compacting concrete made with copper slag. Construction and
Declarations Building Materials, 247, 118580. https://​doi.​org/​10.​1016/j.​conbu​
ildmat.​2020.​118580
Conflict of interest The authors declare that no potential conflict of IS:10262:2019, I. (2019). Concrete Mix Proportioning. Bureau of
interest in the subject matter. Indian Standards, 1–44
Kaveh, A., Eskandari, A., & Movasat, M. (2023). Buckling resistance
prediction of high-strength steel columns using metaheuristic-
trained artificial neural networks. Structures, 56, 104853. https://​
doi.​org/​10.​1016/j.​istruc.​2023.​07.​043
References Kaveh, A., Gholipour, Y., & Rahami, H. (2008). Optimal design of
transmission towers using genetic algorithm and neural networks.
Ahmadi, M., Farzin, S., Hassani, A., & Motamedi, M. (2017). Mechan- International Journal of Space Structures, 23(1), 1–19. https://​
ical properties of the concrete containing recycled fibers and doi.​org/​10.​1260/​02663​51087​85342​073
aggregates. Construction and Building Materials, 144, 392–398. Kaveh, A., & Khavaninzadeh, N. (2023). Efficient training of two
https://​doi.​org/​10.​1016/j.​conbu​ildmat.​2017.​03.​215 ANNs using four meta-heuristic algorithms for predicting the
Asteris, P. G., Skentou, A. D., Bardhan, A., Samui, P., & Lourenço, FRP strength. Structures, 52, 256–272. https://​doi.​org/​10.​1016/j.​
P. B. (2021). Soft computing techniques for the prediction of istruc.​2023.​03.​178
concrete compressive strength using non-destructive tests. Con- Khan, F. A., et al. (2020). Predicting the compressive strength of recy-
struction and Building Materials., 303, 124450. https://​doi.​org/​ cled aggregate concrete using machine learning. Construction and
10.​1016/j.​conbu​ildmat.​2021.​124450 Building Materials, 239, 117859.
BIS. (1970). IS 383: 1970 Specification for Coarse and Fine Aggre- Kumar, R., Rai, B., & Samui, P. (2023). A comparative study of predic-
gates From Natural Sources for Concrete. Indian Standards, 1–24. tion of compressive strength of ultra-high performance concrete
Biswas, R., Bardhan, A., Samui, P., Rai, B., Nayak, S., & Armaghani, using soft computing technique. Structural Concrete. https://​doi.​
D. J. (2021). Efficient soft computing techniques for the predic- org/​10.​1002/​suco.​20220​0850
tion of compressive strength of geopolymer concrete. Computers Liu, Y. (2022). High-performance concrete strength prediction based
and Concrete., 28(2), 221–232. https://d​ oi.o​ rg/1​ 0.1​ 2989/c​ ac.2​ 021.​ on machine learning. Computational Intelligence and Neurosci-
28.2.​221 ence. https://​doi.​org/​10.​1155/​2022/​58022​17
Chaabene, W. B., Flah, M., & Nehdi, M. L. (2020). Machine learning Mai, H.-V.T., Nguyen, M. H., & Ly, H.-B. (2023). Development of
prediction of mechanical properties of concrete: critical review. machine learning methods to predict the compressive strength of
Construction and Building Materials, 260, 119889. https://​doi.​ fiber-reinforced self-compacting concrete and sensitivity analysis.
org/​10.​1016/j.​conbu​ildmat.​2020.​119889 Construction and Building Materials., 367, 130339. https://​doi.​
Chen, T., & Guestrin, C. (2016). XGBoost a scalable tree boosting org/​10.​1016/j.​conbu​ildmat.​2023.​130339
system. In proceedings of the 22nd ACM SIGKDD international

13
 Asian Journal of Civil Engineering

Moein, M. M., Saradar, A., Rahmati, K., Mousavinejad, S. H. G., Bris- Asian Journal of Civil Engineering. https://​doi.​org/​10.​1007/​
tow, J., Aramali, V., & Karakouzian, M. (2022). Predictive models s42107-​023-​00828-6
for concrete properties using machine learning and deep learning Shahmansouri, A. A., Yazdani, M., Ghanbari, S., Bengar, H. A., Jafari,
approaches: a review. Journal of Building Engineering. https://​ A., & Ghatte, H. F. (2021). Artificial neural network model to pre-
doi.​org/​10.​1016/j.​jobe.​2022.​105444 dict the compressive strength of eco-friendly geopolymer concrete
Nazeer, M., Kapoor, K., & Singh, S. P. (2023). Strength, durability and incorporating silica fume and natural zeolite. Journal of Cleaner
microstructural investigations on pervious concrete made with fly Production, 279, 123697. https://​doi.​org/​10.​1016/j.​jclep​ro.​2020.​
ash and silica fume as supplementary cement. Journal of Building 123697
Engineering. https://​doi.​org/​10.​1016/j.​jobe.​2023.​106275 Shubham, K., Metya, S., & Sinha, A. K. (2023a). Surrogate model-
Parhi, S. K., & Patro, S. K. (2023). Prediction of compressive strength based prediction of settlement in foundation over cavity for reli-
of geopolymer concrete using a hybrid ensemble of grey wolf ability analysis. Transportation Infrastructure Geotechnology.
optimized machine learning estimators. Journal of Building Engi- https://​doi.​org/​10.​1007/​s40515-​023-​00329-8
neering, 71, 106521. https://​doi.​org/​10.​1016/j.​jobe.​2023.​106521 Shubham, K., Rout, M. D., & Sinha, A. K. (2023b). Efficient com-
Paruthi, S., Husain, A., Alam, P., Khan, A. H., Hasan, M. A., & Mag- pressive strength prediction of concrete incorporating industrial
bool, H. M. (2022). A review on material mix proportion and wastes using deep neural network. Asian Journal of Civil Engi-
strength influence parameters of geopolymer concrete: application neering. https://​doi.​org/​10.​1007/​s42107-​023-​00726-x
of ANN model for GPC strength prediction. Construction and Sivakumar, V. R., Kavitha, O. R., Arulraj, G. P., & Srisanthi, V. G.
Building Materials, 356, 129253. https://​doi.​org/​10.​1016/j.​conbu​ (2017). An experimental study on combined effects of glass fiber
ildmat.​2022.​129253 and metakaolin on the rheological, mechanical, and durability
Rout, M. D. K., & Jena, R. R. (2015). Investigation on the development properties of self-compacting concrete. Applied Clay Science,
of light weight concrete with sintered fly ash aggregate and acti- 147, 123–127. https://​doi.​org/​10.​1016/j.​clay.​2017.​07.​015
vated fly ash in blended cement. International Journal of Engi- Thai, H. T. (2022). April). Machine learning for structural engineering:
neering Research & Technology, 4(4), 25–28. a state-of-the-art review. Structures, 38, 448–491. https://​doi.​org/​
Rout, M. K. D., Biswas, S., & Sinha, A. K. (2021). Mechanical and 10.​1016/j.​istruc.​2022.​02.​003
durability properties of alccofine used in reclaimed asphalt con- Varma, B. V., Prasad, E. V., & Singha, S. (2023). Study on predicting
crete pavements (RACP). In S. S. Biswas, S. Metya, & S. Kumar compressive strength of concrete using supervised machine learn-
(Eds.), Advances in sustainable construction materials (pp. 131– ing techniques. Asian Journal of Civil Engineering. https://​doi.​
142). Springer. org/​10.​1007/​s42107-​023-​00662-w
Rout, M. D., Biswas, S., Shubham, K., & Sinha, A. K. (2023a). A
systematic review on performance of reclaimed asphalt pavement Publisher's Note Springer Nature remains neutral with regard to
(RAP) as sustainable material in rigid pavement construction: jurisdictional claims in published maps and institutional affiliations.
Current status to future perspective. Journal of Building Engi-
neering, 76, 107253. Springer Nature or its licensor (e.g. a society or other partner) holds
Rout, M. K. D., Sahdeo, S. K., Biswas, S., Roy, K., & Sinha, A. K. exclusive rights to this article under a publishing agreement with the
(2023b). Feasibility study of reclaimed asphalt pavements (RAP) author(s) or other rightsholder(s); author self-archiving of the accepted
as recycled aggregates used in rigid pavement construction. Mate- manuscript version of this article is solely governed by the terms of
rials, 16, 1504. https://​doi.​org/​10.​3390/​ma160​41504 such publishing agreement and applicable law.
Rout, M. K. D., Shubham, K., Biswas, S., & Kumar, A. (2023c).
An integrated evaluation of waste materials containing recy-
cled asphaltfine aggregates using central composite design.

13